SCIENTIFIC STUDIES OF READING, 11(A), 357-383 Copyright 2007, Lawrence Erlbaum Associates, Inc.

Reading Ability: Lexical Quality to Comprehension

Charles Perfetti
University of Pittsburgh

The lexical quality hypothesis (LQH) claims that variation in the quality of word representations has consequences for reading skill, including comprehension. High lexical quality includes well-specified and partly redundant representations of form (orthography and phonology) and flexible representations of meaning, allowing for rapid and reliable meaning retrieval. Low-quality representations lead to specific word-related problems in comprehension. Six lines of research on adult readers demonstrate some of the implications of the LQH. First, large-scale correlational results show the general interdependence of comprehension and lexical skill while identifying disassociations that allow focus on comprehension-specific skill. Second, word-level semantic processing studies show comprehension skill differences in the time course of form-meaning confusions. Studies of rare vocabulary learning using event-related potentials (ERPs) show that, third, skilled comprehenders learn new words more effectively and show stronger ERP indicators for memory of the word learning event and, fourth, suggest skill differences in the stability of orthographic representations. Fifth, ERP markers show comprehension skill differences in meaning processing of ordinary words. Finally, in text reading, ERP results demonstrate momentary difficulties for low-skill comprehenders in integrating a word with the prior text. The studies provide evidence that word-level knowledge has consequences for word meaning processes in comprehension. In reading, the singular recurring cognitive activity is the identification of words. From this follow two other, related observations about reading: Comprehension depends on successful word reading. Skill differences in comprehension can arise from skill differences in word reading. These simple observations form the core of a theory of comprehension skill published over 20 years ago (Perfetti, 1985). Verbal efficiency theory claimed that Correspondence should be sent to Charles Perfetti, LRDC, University of Pittsburgh, Pittsburgh, PA 15260. E-mail: Perfetti@pitt.edu

358

PERFETTI

word identification, the rapid retrieval of a word's phonology and meaning, was a limiting factor in comprehension. I referred to these cognitive events of word identification as "retrievals" because they operated on information about a word stored in a reader's orthographically addressable memory. But at the heart of word identification were the phonological procedures that allowed a word (or a nonword) to be decoded, whether or not meaning was also retrieved. The theory assumed the ability to decode nonwords was the hallmark of basic alphabetic reading skill. In fact, phonology was important enough in this account that it had redundant participation. Phonology was both stored as part of the word (and thus retrieved during identification) and generated by connections among subword units that were part of the word. This conceptualization was explicit in the Restricted Interactive Model, which focused on the development with experience of specific and redundant sublexical components suggested in Perfetti (1992). In the theory, the link from word-level reading to comprehension was through the assumption that comprehension included higher level processes that required cognitive resources (working memory), for example, integrative processes, inferences, syntactic repairs. Word identification, and certainly the sublexical processes that produce it, were candidates for low-resource or automatic processes (LaBerge & Samuels, 1974) that could preserve processing resources for higher level comprehension. Automatic, resource-cheap word-level processesverbal efficiency were assumed to support comprehension. Children who have this efficiency would be able to achieve high levels of comprehension, and children with inefficient word-level processes would have problems with comprehension. The research showing correlations between children's decoding skill and comprehension was consistent with this account. Furthermore, there is no reason to suppose that this relationship disappears for older readers (Shankweiler et al., 1999). This general account continues to seem correct to me. However, I think its emphasis on completely general processesdecoding, phonological processes, retrieval, memory, automaticityalthough theoretically consistent, seemed to leave knowledge out of the picture. Skilled reading was about efficient processing mechanisms and less skilled reading was about these same mechanisms executed inefficiently. This description seemed to predict that becoming faster at word identification leads to better comprehension. Inefficient readers can indeed become more efficient (Breznitz & Share, 1992), and improving individual word reading speed may increase fluency (Martin-Chang & Levy, 2005) and, under some circumstances, comprehension (Breznitz & Share, 1992; Tan & Nicholson, 1997). However, increasing decoding speed by itself has not always increased comprehension (Fleisher, Jenkins, & Pany, 1979; Perfetti, 1985). Overall, although the hypothesis that training word-reading speed raises comprehension has some research support, it is not the primary practical implication of the general idea that comprehension depends on efficient word reading.

LEXICAL QUALITY

359

Efficiency is not the same as speed. Efficiency is a ratio of outcome to effort, with time as a proxy for effort. So although processing descriptions make a coherent framework for efficiency, they leave out the basic nature and source of the word reading outcomes on which efficiency depends. These outcomes are word identities that momentarily represent form and meaning components that are the basic elements of comprehension. On this description, the thing to understand is not speed but rather the ability to retrieve word identities that provide the meanings the reader needs in a given context. This source of this ability is the knowledge a reader has about words, specific lexical representations.

LEXICAL QUALITY Underlying efficient processes are knowledge components; knowledge about word forms (grammatical class, spellings and pronunciations) and meanings. Add effective practice (reading experience) of these knowledge components, and the result is efficiency: the rapid, low-resource retrieval of a word identity. Lexical quality (LQ) refers to the extent to which the reader's knowledge of a given word represents the word's form and meaning constituents and knowledge of word use that combines meaning with pragmatic features. Thus the vocabulary of a given language includes, for a given reader, words of widely varying LQ, from rare words never encountered to frequently encountered and well-known words. Likewise, individual readers differ in the average LQ of their words. This reader variability is not just about the size of vocabulary, although it includes this; it is about the representation of words, the stable and less stable knowledge the reader has about the word's form and meaning. Of course, the question becomes what is "quality," a word that could evoke suspicion without some definition. Quality is the extent to which a mental representation of a word specifies its form and meaning components in a way that is both precise and flexible. The precision is needed because "pretty and petty" and "knight and night" are not the same. The flexibility is needed because the meanings of "roaming charge" and "a fee charged by a mobile phone service for calls initiated or received outside a contracted service area" are the same. Both precision and flexibility are needed to understand and pronounce record in "You need a record of the transaction" and "They can't record the conversation." These simple examples are just the tip of the iceberg of form-meaning complexities. LQ provides a means for safe passage through them. Earlier chapters (Perfetti & Hart, 2001, 2002) contain additional examples and theoretical discussions of LQ. One way to become more specific about LQ is to identify the features that we hypothesize to distinguish higher quality from lower quality representations. Table 1 does this. It identifies five features of lexical representation that distinguish high and low quality and shows three (there may be more) hypothesized consequences of

360

PERFETTI

TABLE 1 Properties and Consequences of Lexical Quality Representational Properties of Lexicon Orthography

High Quality

Low Quality Not fully specified; some letters are variables Less stable because of variable word-specific phonology and/ or grapheme-phoneme phonology Incomplete range of form class uses; less stable morpho-syntax More context bound; fewer relevant meaning dimensions to discriminate among related words Orthographic, phonological, and semantic constituents are less tightly bound

Fully specified; letters are constants Redundant word-specific Phonology phonology and context-sensitive grapheme-phoneme phonology Grammar All grammatical classes of the word represented; morpho-syntactic inflections represented Meaning More generalized, less context-bound; fuller range of meaning dimensions to discriminate among words in same semantic field. Constituent binding Orthographic, phonological, and semantic constituents are Possible processing consequences tightly bound during reading Stability Higher; word identity is reliably retrieved from an orthographic or phonological input Word identity constituents are activated and retrieved in synchrony as a word identity

Synchronicity

Meaning integration Higher; word identities available for building comprehension

Lower; word identity is sometimes not retrieved from an orthographic or phonological input Word constituents may be activated and retrieved asynchronously; (e.g. labored decoding; activation of incorrect meanings from partial input) Lower; comprehension processes that operate over word identities at risk

these quality features for reading processes. The representational features are the four constituents of word identityorthography, phonology, morpho-syntax, and meaning. The fifth representation feature is constituent binding, the degree to which the first four features are bound together (especially the first three; the grammatical feature might be considered to be implemented by a grammatical process that operates on the lexeme). Bindings are connections that secure coherence among the constituents, the orthographic, phonological, and semantic representations, which together are the word's identity. The binding feature is not independent but rather a

LEXICAL QUALITY

361

consequence of the orthographic, phonological and semantic constituents becoming well specified in association with another constituent. The consequences of high quality in sublexical and lexical knowledge are also shown in Table 1. These are hypothesized consequences that are subject to empirical testing, and some have at least indirect evidence. For example, the hypothesis that low LQ can lead to the asynchronous activation of word constituents is consistent with results of Breznitz and Misra (2003), who found that ERP indicators of orthography and phonology for low-skill reading were more asynchronous than those of skilled readers. More generally, reading words in context is affected by LQ, providing the important hypothesized link between LQ and comprehension (Perfetti & Hart, 2001). The general description of this link is that local processes of integrating word meanings within and across sentence boundaries are affected by the LQ of words that are identified as part of the comprehension process. One final observation concerns the difference between a spoken and written word. My focus is on reading, and for that orthography is part of LQ. However, a lexical analysis can be applied to just spoken language with a focus on phonological representations and meaning. WORD PROCESSING, WORD LEARNING, AND COMPREHENSION STUDIES With this background on the general nature of the LQ hypothesis (LQH) and its links to the process-oriented account of verbal efficiency, I turn now to a review of some studies of reading that bear on the LQH. Although these are studies of adult reading, I believe their conclusions apply also to children's reading. The Structure of Lexical and Comprehension Skill in Reading Lexical knowledge and comprehension should be associated, and they are. Positive correlations between word-processing measures of various kinds and reading comprehension assessments are well established in both children (Perfetti, 1985) and adults (Haenggi & Perfetti, 1994). More interesting is the fact that this correlation is generally in the moderate range, leaving plenty of room for disassociation between the two. We've maintained a large database of college students for whom we take various reading and reading-related measures. In general, we are interested in knowing the associates of reading comprehension skill and in the outcomes of experiments that related specific reading processes to these measures. Perfetti and Hart (2002) reported some results of factor analysis on a sample of 445 individuals from this database. They concluded that skilled readers' knowledge of spelling, phonology, and decoding could be represented reasonably well by a single word form factor with a second factor reflecting meaning and comprehension. However, less skilled readers, in addition to a meaning factor, required two form

362

PERFETTI

factorsone loaded with more phonological tasks and the other with more orthographic tasks, suggesting less coherence of word identities. Since then two dissertations have assessed large samples from this database, based on partly different tasks, one by Hart (2005) and a later one by Landi (2005). Both Hart and Landi were interested in the disassociation of word-level skill from comprehension skill, as assessed by the Nelson-Denny comprehension test. Hart analyzed the scores of 792 students, concluding that, among those who could be most confidently classified, 18% were below the median in comprehension but at or above the median on lexical measures; 64% showed the more typical association pattern of lexical and comprehension scores both high or both low. In her experiments with a carefully defined subset of this sample, Hart asked whether certain aspects of learning an artificial language might depend more on first-language lexical knowledge compared with first-language comprehension skill. Her results showed a complex pattern, with both comprehension and lexical knowledge predicting various measures of performance with the novel language during learning and in postlearning transfer. However, lexical knowledge, more than comprehension, predicted the learning of this artificial language, including its novel orthography and its decoding mappings. It also predicted resistance from interference by homophones that were planted in the new language, replicating a result reported in Perfetti and Hart (2001) for English. Landi's study was carried out 2 years later on a different sample of 799 students and used an overlapping but partly distinct set of tasks. Landi's factor analysis of five tasks yielded a comprehension component and a lexical component, which she then used to weight normalize individual participant scores on each test. Figure 1 shows a scatter plot of these scores. In this normalized analysis, 23% were below the median on the comprehension component but above the median on the lexical component; the reverse pattern, high comprehension but low lexical components, was observed for 9% of the sample. This asymmetry is the pattern one would expect (but might not find based on median splits of test scores). Lexical knowledge is not sufficient for comprehension, so the low-comprehension/high lexical pattern is more prominent than the high-comprehension/low-lexical pattern. The approximately 20% (18% in Hart's sample, 23% in Landi's sample) of college students whose comprehension levels undershoot the level expected by lexical skills identifies a group for whom hypotheses about other sources of comprehension problems can be meaningfully tested. The idea that lexical processes are not sufficient for comprehension should not be controversial, although a careless reading of verbal efficiency theory might have led some to believe that the theory assumed there was nothing to comprehension beyond efficient word reading. Observations to the contrary have long been in the literature, and at least a few of them seem to have controlled adequately for word-level skills. Studies by Oakhill and colleagues (Cain & Oakhill, 1999;

LEXICAL QUALITY 363

Lexical Component
-1

-2 "3

- 2 - 1

12

Comprehension

Component

FIGURE 1 Scatter plot of normalized component scores from principal components analysis of reading test scores from a sample of 799 college students, based on Landi's (2005) dissertation. In this procedure, normalized Z scores for each test for each participant were multiplied by the factor score for that test determined from the Principle Components Analysis. Thus, the plot is a factor-weighted composite of five tests (decoding, spelling, vocabulary, comprehension, and the Author Recognition Test) that weighted differentially on the two components.

Oakhill, Cain, & Bryant, 2003) seem to show that some children have trouble drawing inferences during comprehension, despite having good decoding skills. More generally, comprehension problems can arise from general language comprehension problems even when word decoding appears to be adequate (Stodhard & Hulme, 1996). (For reviews, see Nation, 2006, and Perfetti, Landi, & Oakhill, 2005). Another candidate for comprehension problems, closer to the idea of LQ, is the hypothesis that children with adequate decoding and phonological-level skills can have word-level semantic problems that affect comprehension. This semantic deficit hypothesis (Nation & Snowling, 1998,1999), which has some evidence in studies of children, allows for several interpretations of semantic deficits, including a problem with semantic categories. For now, the Landi and Hart studies extend to the adult population the observation that decoding is not sufficient for comprehension.

Studies of Form-Meaning Confusions One way to study the effects of LQ on comprehension is to experimentally create threats to quality. This is the basis for experiments reported in Perfetti and Hart (2001), which I only briefly review here. The key idea is that for a word like wails,

364

PERFETTI

the quality of its identity, which entails its spelling, pronunciation, and meaning, is threatened by the existence of whales, which shares its pronunciation. So how do readers of higher and level skill handle this kind of threat? Our hypothesis was that better readers, defined by comprehension assessment, have higher quality representations, so given wail, they should retrieve wail associates, not whale associates, even if there is momentary activation of both the presented word and its homophone (Gernsbacher & Faust, 1991). In the meaning task reported in Perfetti and Hart (2001), readers decided whether two words presented in succession were semantically related. Some trials contained homophones of words that would have been related, as when wails was followed by dolphins. Skilled comprehenders showed faster-meaning decisions for both control pairs and homophones, and they showed less homophone confusion when the presented (confusable) form was the one of higher frequency {whales-cries rather than wails-dolphins). However, they did show confusions when presented with a form of low frequency, and this effect occurred at shorter latencies than it did for less skilled comprehenders. So night-armor did not produce confusion (in the form of longer decision times) for skilled comprehenders, although it did for less skilled comprehenders. However, knight-evening produced interference for skilled comprehenders, and this emerged very rapidly, at a Stimulus Onset Asynchrony of 150 msec, whereas for less skilled comprehenders the interference effect did not emerge until 450 msec. Thus, skilled comprehenders show less interference based on form, and when they show interference, it occurs within 150 ms of exposure to the homophone, suggesting an early activation of word phonology. This frequency effect in homophone confusions appears to be dependent on word experience. In a study reported in Hart and Perfetti (in press), readers were provided with experience on the member of a homophone pair that was judged to be less familiar, so as to make it more familiar than its mate. For example, in the pair night-knight, knight was rated lower in familiarity, so it was the one participants experienced in training; similarly, in hair-hare, hare was rated lower, so it was the training word. The result of the training, which consisted of visual exposures to the word associated with meaning, was the reversal of the frequency effect in homophone interference. In terms of the examples, training on knight caused semantic decisions on knight-evening to produce less interference than semantic decisions on night-armor. So form-based confusions depend on the relative frequencies of competing forms, which in turn depend on reading experience. A highly frequent form is relatively protected from interference, because it retrieves its meaning and pronunciation rapidly as a stable, unique word identity. An infrequent form is less protected because it is more likely to retrieve an unstable identity based on shared phonology with the more frequent form. The application of this to reading skill is that LQ depends on experience with words. A skilled comprehender has

LEXICAL QUALITY

365

had more experience with a given word than has a less skilled reader, and this has important implications.1 The relationship between word frequency estimates and various word-processing tasks is nonlinear, generally logarithmic. In low frequency ranges, a given difference in frequency between two words may have a large effect on measures of speed of processing; in high frequency ranges, that same frequency difference has a smaller effect. Skill differences in word-reading experiments are usually greater for low-frequency words than high-frequency words. This may reflect the importance of some minimum number of exposures for a word to be identified with low effort. If we assume that a given word has been read more frequently by a skilled reader than a less skilled reader, then it follows that the skill differences we observe in processing that word reflect this frequency of experience difference. However, the difference in exposures seems to have an effect only for low-frequency words, consistent with assumption that it is the low-frequency range where increments in frequency are most important. Although recency effects can be disguised as frequency effects, this does not matter for the skill conclusion. More reading leads, statistically, to more frequent and more recent encounters and both may have this nonlinear effect on word-reading efficiency. Although this statistical perspective is important, it does not mean that all experiences with words are equal. In fact, in the next section, I review a study that suggests that skilled comprehenders make better use of their experiences with words they are trying to learn.

Learning the Meanings of New Words Given the implication that LQ is acquired through effective experience with words, we might be able to observe the acquisition of LQ during word learning. Furthermore, by comparing the learning of readers who differ in comprehension skill, we can examine the link between LQ and comprehension in a situation that controls the word experiences. In this research, we have used both behavioral and ERP measures. Perfetti, Wlotko, and Hart (2005) taught the meanings of very rare words to undergraduates and then tested the effects of this learning in a simple meaning judgment task while recording electroencephalograms (EEGs). Examples of the words taught include the following: gloaming, flexion, clement, ibex, agog, bastion, tiglon, and quisling. In a pretest lexical decision task, our rare words were judged to be real words only at a rate of 8% on average. To assure that the words we would
'Gemsbacher and Faust (1991; also Gernsbacher, 1990) explained less skilled readers' problems in meaning processing as due to problems in suppressing irrelevant meanings that are activated by a word. Differences between their mechanism-based account and the knowledge-based account of the LQH are discussed further in Perfetti and Hart (2001).

366

PERFETTI

train and then test were unknown, the to-be-learned words were individually tailored for each participant according to the pretest. Following simple association instruction (50 min in which the rare words were paired with brief definition-like paraphrases), the trained words, untrained rare words, and familiar words were presented for meaning judgments. In the meaning judgment task, the first word appeared for 1 sec and then disappeared, replaced by the second word that was related in meaning on 50% of trials. For example, gloaming followed by twilight, should get a yes response. EEGs were recorded continuously during these judgments, so we obtained ERP indicators associated with viewing the first word (gloaming) and its meaning mate (twilight), Figure 2 shows the behavioral results obtained during the posttraining meaning judgments task. The thing to notice is that skilled comprehenders were correct significantly more often than less skilled comprehenders in meaning judgments made to the rare words that we taught them but not to either untrained rare words or familiar words (also not trained). The lack of a skill difference for untrained rare words shows that overall knowledge of rare words was not different across the two groups. Instead, the conclusion is that the higher comprehenders actually learned the new words better. The results of the ERP analysis, which are shown in Figure 3 for the group of skilled comprehenders, add to this picture. Plotted are the grand average waveforms at one electrode (the central reference electrode) for each condition for meaning-related trials. The conditions show a similai- pattern for the first 200 msec or so, reflecting visual orthographic processes shared by all words. The first point of separation occurs at 200 msec, where trained words separate from both untrained and familiar words in a negative going shift. This reflects an early "notice"

FIGURE 2 Comprehension skill differences in word meaning judgments following learning of rare words. Skilled comprehenders showed higher accuracy in judging meaning relations for trained rare words but not for untrained rare words or familiar (known) words. Based on Perfetti, Wlotko. and Hart (2005).

LEXICAL QUALITY

367

Cz (129)

FIGURE 3 An event-related potential (ERP) record for skilled comprehenders during a meaning judgment. The grand average waveform is shown for the reference electrode (Cz). The onset of the first word is exemplified for the trained rare word gloaming. The onset of the second word is exemplified for the related word twilight. The three curves represent ERP records for trained rare words (darkest line), untrained familiar (known) words (intermediate darkness), and untrained rare words (lightest line). Two significant effects of training are visible, at 200 msec and about 550 msec, the latter representing word-level episodic memory for the trained word. About 400 msec after the onset of a related word, a reduction in the N400 is observed for trained and familiar words. For the trained words, less skilled comprehenders (not shown) show weaker episodic effects at 550 msec and weaker N400 meaning effects for the second related word. See Perfetti, Wlotko, & Hart (2005) for fuller skill comparisons.

of words that had been recently viewed in training, based on preidentity sublexical patterns. A second separation at around 550 msec, which does mark word identity, further distinguishes trained words from the other two classes, now in a positive going shift. This shift is the same in key respects (distribution and polarity) as the P600 that is observed in memory studies when a previously viewed item is presented (Curran, 1999; Rugg, 1995). This ERP shift marks recognition of the episodic memory laid down by the training event: In effect, the brain responds to this word as familiar because the word has been part of the previous hour's training. A word that has gained its familiarity through experiences prior to the experiment (the familiar words) shows no such effect. The implication of this is that we have identified a marker of familiarity-based learning that is expressed when a reader views a word.

368

PERFETTI

So now the question is whether this word-level episodic memory effect is observed equally in our skilled and less skilled comprehenders. The answer is no. Less skilled readers showed the same Figure 3 pattern of ERP shifts during the meaning judgments. However, the key marker of episodic memory at 550 msec (the P600 training effect) was significantly reduced in amplitude for the less skilled comprehenders. Thus, on average, a word that had been learned just prior to the experiment made less of an impression on the less skilled readers. It is interesting that the first effect of training, the 200-msec negativity for trained words, was not different for the two skill groups. If our interpretation of these two components is correctthat the 200 msec is based on sublexical familiarity whereas the 550-msec effect is based on lexical identitythen we conclude that all learners respond to the distinctive letter pattern of a recently trained word, but they differ in recognizing the word episodefor example, this is that word gloaming that I just experienced a few minutes ago. Recognizing a "gloaming" word-episode may involve retrieving its meaning, but it may not. The more direct test of a meaning process is in the response to the second word, twilight. If the participant has learned the meaning of gloaming, then the second word, twilight, which is closely related in meaning, should produce a reduced N400. The N400 is a signature for semantic congruence, a large negative going shift when a word is incongruent with its preceding context. When a word is congruent with its preceding context, the N400 is reduced. This N400 reduction is what happened when gloaming was followed by twilightprovided the meaning of gloaming was learned. As shown in Figure 3, an N400 appeared for the untrained words, reflecting the fact that if a participant had not experienced gloaming in training, there was no particular congruence provided by the word twilight. However, if gloaming had been learned, then the N400 should be reduced, and this reduction, as well as one for familiar words, is visible in Figure 3. Once again, the ERP effect was different for skilled and less skilled comprehenders. Although both groups showed a reduction of the N400 when the second word was related in meaning to the first word, this N400 for less skilled comprehenders was significantly less reduced for trained words compared with the reduction for skilled readers. We must conclude that after less than 50 min spent learning the meanings of 60 rare words whose meanings were unknown prior to the study, skilled comprehenders made more effective use of the learning period. Especially interesting is the fact that this difference was seen in ERPs recorded during the brief period in which the word was being viewed, prior to the appearance of a second word. The N400 on the second word reflected the stronger learning of the meaning by skilled comprehenders; the episodic marker on the first word indicated a stronger association between the word and its training. One might argue with the interpretation that these are learning effects as opposed to subtle experience effects. Although we chose words uhat were individually

LEXICAL QUALITY

369

tailored to be unknown for a given participant, perhaps that procedure underestimated very slight familiarity differences that favored the more skilled and more experienced readers. This seems unlikely. For one thing, we did not find differences in the untrained words, which were from the same population of pretested rare words. If skilled readers had some unmeasured familiarity with the trained words prior to the study, then they should have the same unmeasured familiarity for the untrained words. But the behavioral and ERP results both say they did not. Furthermore, the interpretation that we have a difference in the ability to learn word meanings is consistent with the results of Hart's (2005) dissertation. Readers high in comprehension skill showed better meaning decision performance than low-skill comprehenders on words learned in her artificial language. Acquiring Lexical Form Stability As indicated in Table 1, one of the features of LQ is a stable lexical representation. Stability occurs as letters and phoneme constituents become specifiable as constants (fully specified) rather than variables in the word representation. High-quality representations are fully specified (Perfetti, 1992). In a study that used rare-word learning to examine form stability during learning (Yang & Perfetti, 2006), skilled and less skilled comprehenders learned the meanings of 42 rare words over four mini-training "sessions" in the course of a single day. After each session, participants also made lexical decisions, choosing whether a given letter string was the correct form of a word they were learning. The foils varied systematically in their orthographic and phonological overlap with the correct form. This manipulation allowed tests of the learners' form stability. If learners acquire well-specified representations of the word they were learning, similar forms should produce less interference. Table 2 illustrates what the learners were up against. In the midst of trying to learn the meaning of the new, very rare word, hebetude, learners encounter either one of the spellings shown in Table 2 or the correct form of the word hebetude, deciding whether the presented form is the correct form.2 Orthographically similar foils (two left columns of Table 2) had high spelling overlap with the target, and some of these also had very high phonemic overlap. Control foils shared only an initial letter with the target. An unusual feature of this study is the use of multisyllabic words, which required some flexibility in creating foils. As Table 2 shows, the orthographic foils for hebetude all share both graphemes and phonemes for the first of the three syllables. (This means substantial phonemic overlap even among low overlap foils.) High phonemic overlap in this case was carried by
2 According to World Wide Words (Quinion, July 2007), an August 2001 column in The Washington Post observed that "Too many Americans slouch toward a terminal funk of hebetude and sloth." It was also the "word of the day" on Dictionary.com on January 24, 2004.

370

PERFETTI

TABLE 2 Foils for the Rare Word Hebetude

High Orthographic Overlap Higher Phonological Overlap hebitude hebatude hebutude hebotude Lower Phonological Overlap hebetide hebetade hebetede hebetode Control: First Letter Only hodilane harisade higifore hudufise

identical first and third syllables (left column). The four high overlap foils differed only on the vowel letter of the second unstressed syllable. Because the syllable is unstressed, vowels can migrate toward a minimal vowel. One could conceivably pronounce all the second syllable vowels as unstressed uh. Of interest are results of both form and meaning learning. The meaning part of the study involved viewing the word and hearing a spoken definition, which could be repeated at the learner's option, followed by a test of meaning in which the word was spoken. We explained to participants that learning the meanings of rare words like hebetude was the goal. Each of the four sessions of meaning learning was followed by the form task (lexical decision). The meaning results are shown in Figure 4. Skilled comprehenders learned more than less skilled comprehenders from the first session, and their advantage remained constant over the four sessions. This result converges with that of

Meaning Accuracy Each Session

100% jr ------------ -----------------95% -90% 85% -80% 75%

S1 S4 S2 S3 S4 S1 S2 Lskilled S3 Skilled

Skill and Session

FIGURE 4 Increase in rare word learning over four learning "sessions" (all within one day). Skilled comprehenders learned more during the first session and maintained this slight advantage.

LEXICAL QUALITY

371

Perfetti, Wlotko, & Hart (2005) in finding that skilled comprehenders learn the meanings of new words more effectively than less skilled comprehenders. Form learning, which is shown in Figure 5, was fairly good overall, with false alarms to foils ranging around 10%. Nevertheless, skilled comprehenders were more accurate at rejecting similar foils and selecting the correct form. After one session of learning, less skilled comprehenders chose a foil on 20% of trials. Figure 5 does not distinguish among foil types, but there were differences: The foils with high phonemic overlap attracted the highest percentage of false alarms, and foils with less phonemic overlap but high orthographic overlap were next. Control foils, which shared only the initial letter with the target, attracted few false alarms. However, there was no interaction of the type of foil with reader group. Both skilled and less skilled comprehenders made more errors to foils when both orthographic and phonemic overlap was high. Thus, the conclusion is that skilled comprehenders learn not only new word meanings more effectively but also new forms more effectively. LQ implies coherence between form and meaning components, and skill in reading is associated with higher LQ right from the beginning of learning. The idea of stability is more specific than effective form learning. It implies that the word's representation comes to comprise spelling and pronunciation patterns that are identical on successive observations. One way to examine stability in this

100%

FIGURE 5 Form accuracy (lexical decisions) over four meaning learning sessions. Hit rates for real rare words (e.g., hebetude) compared with false alarms to similar forms (hebitude, hebetade, etc.). Less skilled comprehenders show less word-form accuracy.

372

PERFETTI

sense is shown in Figure 6, which plots performance on successive lexical choice trials. The performance plotted is conditional for sessions beyond the first: Given a correct choice in one test session, the probability of being correct the next time. Thus choosing hebetude at one test but then hebitude on the next would contribute negatively to this measure, which is one of stabilitychoosing the same form the next time. Figure 6 shows again the slightly better performance of skilled compre-henders right from the beginning, that is, from the first session. Stability (conditional probabilities of correct responses) shows a small but consistent difference on the next two sessions, and the difference becomes largest on the final session, when skilled comprehenders produced 96% stability compared with 86% for less skilled comprehenders. An informal characterization of this result is that the less skilled group, on average, shows a small but noticeable instability even after four sessions of learning. This is evidence in favor of the stability implication of LQ. Less Skilled Comprehenders' Semantic Processing of Ordinary Words The two preceding sections have concluded that problems in learning word meanings and in learning word forms both associate with comprehension skill. In the
Given correct on trial n, probability of correct on

LS(23) : HS(16)

0.98 0.96 0.94 0.92 0.9 0.88 0.86 0.84 0.82 0.8 0.78

FIGURE 6 An indicator of form stability during word learning. The conditional probability of a correct decision on rare words given a correct response on the preceding trial over successive lexical decision sessions. Less skilled comprehenders show slightly less word form stability by this measure.

LEXICAL QUALITY

373

Studies of Form-Meaning Confusions section, we concluded that problems in meaning processes that were associated with form were more likely for less skilled readers. In this section we consider semantic processing differences for ordinary words already known to the reader. The hypothesis that semantic deficits may explain comprehension problems has been developed and tested for children (Nation & Snowling, 1998, 1999). The basic idea is that such children have achieved the basics of phonological decoding but are limited in the meaning processes that are linked to word identification. Landi (2005; Landi & Perfetti, 2007) extended the range of this hypothesis to include adults, linking behavioral and ERP measures to the assessments of specific lexical and comprehension skill. In effect, Landi's experiments targeted the group of readers of Figure 1 who are in the lower right quadrantbelow the diagonal and to the right of the z-axis midpoint: readers of high lexical skill but low comprehension skill. Landi tested the hypothesis that such readers are less able to effectively use semantic category information. Participants made meaning decisions for word pairs that were related associatively and categorically or categorically only. For example, brother-sister, dog-cat, pillow-sleep were pairs both associatively and categorically related; green-pink, banana-tomato, kite-balloon were pairs related only categorically. Landi's hypothesis was based on Nation and Snowling's (1999) conclusion that children's comprehension problems reflected failures to represent semantic category relations. Thus, less skilled comprehenders would be comparable to skilled comprehenders in semantic processes when associative relations could facilitate detection of semantic relations, but they would be less successful than skilled comprehenders when they had to rely only on categorical relations. To provide more informative comparisons, participants also made semantic judgments on pairs of pictures (controlling for word reading) and homophone decisions, for example, boar-bore, chantschance, tacks-tax (assessing phonological processing). The key results can be summarized as follows: Accuracy was generally high for all tasks and not different between the two groups. However, times for correct decisions were faster for the skilled group than the less skilled group across all tasks. Thus, skilled comprehenders showed faster semantic processing whether the stimuli were words or pictures and faster word judgments whether the decision was based on meaning or pronunciation. For word meanings, the decision times to semantic category pairs were slower than to associative pairs. This confirms the assumption that detecting category relations would require more processing than detecting relations between associations. Correct decisions based on category relations took an average of 68 msec longer than those based on associative relations for the skilled group and 139 msec longer for the less skilled group. These group differences, although in the direction implied by Landi's extension of Nation and Snowling's (1999) hypothesis, were not reliable.

374

PERFETTI

TABLE 3 Summary Skill Pattern in Landi ERP Results Comprehension Task Phonological decisions Semantic picture decisions word decisions- -Categorical Semantic Semantic word decisions Note. -Associative Skill Differences in Decision Times ? Yes Yes Yes Yes Comprehension Skill Differences in ERP Measures? No No Yes, but small (P200, N400) Yes, large (P200, N400)

Source: Landi and Perfetti (2007). ERP = event-related potential.

Whereas the decisions times showed very general processing-speed differences, the ERP measures showed differences that were restricted to the semantic word decisions, and these results present a slightly different picture. Both groups showed a reliable reduction in the N400 for related trials for both category and associative pairs, although in each case the reduction shown by skilled compre-henders was somewhat larger. Moreover, the high-skill group, but not the low-skill group, showed an additional N400 reduction for the associatively related pairs relative to the category pairs. Thus, contrary to the expectation that less skilled comprehenders would have specific problems with category relations, the ERPs suggest that semantic category processes were comparable in the two groups, but the ERPs for associative relations were not. (The tendency for less skilled comprehenders to take relatively longer than skilled comprehenders on category decisions may be relevant; however such a difference could be interpreted as an additional checking process well beyond the more automatic semantic process that is reflected in the N400.)3 Table 3 shows a summary of skill differences across tasks. Taken together, they suggest detailed evidence for semantic-processing differences between the skilled and less skilled comprehenders. The ERP evidence suggested that skilled comprehenders had a stronger meaning congruence response when words were categorically related, and this response was even stronger when the words were also associatively related. The less skilled group showed a congruence response also, but this response was not strengthened by associative relations. Notice that if all we had were behavioral data, our conclusion would be different, and perhaps misleading. We would conclude that there is no semantic processing skill difference
The ERP results also produced evidence for early semantic effects at 200 msec that were more consistent across electrode sites and trial conditions for skilled comprehenders. They also showed 200-msec effects for the phonological task, comparing homophones versus nonhomophones that were comparable across the two groups.
3

LEXICAL QUALITY

375

specific to language, because of differences in picture decision speed. And we would have concluded that these groups differed either on phonological processes (because of the phonological decision speed differences) or, more likely, on nonspecific processing-speed differences. The ERP data are telling us more directly about the brain's response to words, and they seem to say that there are specific skill differences related to word-meaning processes. Lexical Processing During Text Comprehension My final example returns to the key link between lexical processes and comprehension. This link is most direct at the level of short runs of text, a sentence or two, where one can observe word processing "on-line" as part of text reading. The focal issue is the processes that integrate the word currently being read with the ongoing representation of the text. This integration is a central connecting event between word identification and text comprehension. The skill question is whether differences in globally assessed comprehension arise at this local (i.e., one word at a time) processing. Do skilled comprehenders integrate words immediately into the text? Do less skilled comprehenders? The LQH (and also verbal efficiency theory) predicts word-text integration problems for less skilled readers. Once again, our studies on this question use ERPs to examine the N400 semantic congruence indicator. The N400 has been found to vary with demands on sentence and text-level integration (van Berkum, Hagoort, & Brown, 1999), and it is that fact that we exploited in two parallel studies, one with skilled comprehenders (Yang, Perfetti, & Schmalhofer, 2007) and one with less skilled comprehenders (Yang, Perfetti, & Schmalhofer, 2005). The approach of the studies is illustrated in the following text: After being dropped from the plane, the bomb hit the ground and exploded. The explosion was .... ERPs were acquired as each word was read. The underlining (not visible to participants) marks the target word (explosion) for our analysis. When the reader comes to explosion at the beginning of the second sentence, it should be relatively easy to integrate the word with an understanding of the text based on the first sentence. That is, we can assume that the reader's text memory includes the event-proposition [bomb exploded], among other propositions from the first sentence, and the word explosion is integrated with this event-proposition; that is, it is taken as coreferential with the proposition. This condition is termed "explicit" because there are explicit coreferential phrases (exploded, explosion). A second condition, the paraphrased condition, was of special interest from a semantic processing point of view. Instead of the first sentence referring to

376

PERFETTI

"exploded," it contains the paraphrase "the bomb blew up." For a semantic process that links words through their stored meanings (or through one that generates context-sensitive referential meanings from words), integration is possible in this condition. However, it might take a bit more processing work, as a reader searches for the meaning of explosion and then links it to the meaning of "blew up" in the context of the first sentence. A third condition, the inference condition, allowed the critical word to be integrated readily only if the reader had made an appropriate forward (predictive) inference in the first sentence. In the example, the first sentence would say only that the "the bomb hit the ground." The reader may predictively infer that the bomb exploded, in which case encountering the word explosion in the next sentence would be easily integrated. However, notice that such a predictive inference in the first sentence is not necessary for comprehension. Finally a baseline condition, the unrelated condition, provided an initial sentence that contained no possible antecedent for the word explosion. To control for specific words, it contained plane, bomb, and dropped, but, as can be seen in Table 4, these words were not in propositions that suggested bomb dropping. This condition is a baseline, because one expects an N400 when the word explosion is read in the second sentence. The question is, compared with the N400 produced in the baseline condition, which other conditions produce reduction of the N400? When meaning integration is relatively easy, such a reduction is expected. Table 4 shows both examples of texts and the pattern of skill results for these four conditions. In viewing the table, it is important to keep in mind that yes and no cannot do justice to the full data. The pattern as shown is an interpretation based on
TABLE 4 Example of Materials for Word-to-Text Integration and Pattern of ERP Results From Yang, Perfetti, and Schmalhofer (2006, 2007) ERP (N400) Effect Condition Explicit Sample Passage After being dropped from the plane, the bomb hit the ground and exploded. The explosion was quickly reported to the commander. After being dropped from the plane, the bomb hit the ground and blew up. The explosion was quickly reported to the commander. After being dropped from the plane, the bomb hit the ground. The explosion was quickly reported to the commander. Once the bomb was stored safely on the ground, the plane dropped off its passengers and left. The explosion was quickly reported to the commander. Skilled Yes Less Skilled Yes

Paraphrased

Yes (P300)

No

Inference

No

Yes

Unrelated groups

Baseline. Both show large N400.

LEXICAL QUALITY

377

a number of analyses reported in Yang et al. (2005,2007). The inference condition, for example, needs some hedging. Less skilled readers showed a significant reduction of the N400 in this condition, whereas skilled readers did not. However, the skilled comprehenders' N400 appeared to be intermediate between baseline and the two conditions (explicit and paraphrased) that produced significant N400 reduction, suggesting that skilled readers make inferences some times but not generally, just as one would expect for inferences that are not required to maintain coherence. (See Perfetti, Yang, & Schmalhofer, in press, for further discussion.) To focus on what is clear and consistent in the pattern of Table 4, there is an important difference between skilled and less skilled comprehenders in the N400 data. Skilled readers showed that they easily integrated a word with the prior text through a paraphrase process. For example, when they read explosion, they link it to the event described in the preceding sentence as "blew up." Paraphrase relations, although they have not been studied much in comprehension research, are important for semantic linking processes. In comprehension, words need to link with referentially specified mental representations, not merely words. Paraphrases are context dependent"blew up" is not always a paraphrase of "exploded"; it can also be a paraphrase of "enlarged" as in "To see more detail, we blew up the photograph." Paraphrastic semantic relations require a lexical representation that allows a flexible range of meanings that can be fit to contexts. Thus, paraphrase is an especially important process because it is at the "interface" of lexical knowledge and comprehension. In evaluating the overall pattern of evidence in these studies, we concluded that less skilled comprehenders showed "sluggish" word integration processes (Yang et al., 2005). This evidence exceeds what we are reviewing here and included topographic maps (ERP data displayed over the scalp and over time) that show a dramatic difference between the two groups. Skilled readers' maps showed an early rise of positivities in explicit and paraphrase conditions. Less skilled comprehenders' maps showed widely distributed negativities even at 400 msec, before shifting to positive at points beyond 500 msec. Thus, we suggested their comprehension was "sluggish"slow and not always successful in integrating a given word with the understanding of the text. This is just the kind of effortful and inefficient comprehension process the LQH predicts for readers with low-quality word knowledge.

CONCLUSION: CONSEQUENCES OF LQ There are some generalizations to emphasize concerning the consequences of lexical processing (and lexical knowledge). An obvious one is that LQ determines the accuracy and fluency of word identification. Less obvious consequences include the following. LQ influences readers' resistance to form confusions, the ability to

378

PERFETTI

learn the meanings of new words, the retrieval of meanings of learned words, the stability of form representations, and the integration of words with text representations. Especially interesting is the consequence of verbal efficiency, which is demonstrated in the studies of Yang et al. (2005, 2007) reviewed previously. Most of the studies I reviewed here assessed reading comprehension only, and the links to lexical processing are in experimental manipulations rather than independent assessments. In these cases, we depend on the general correlation between comprehension and word processing to support the conclusion that our comprehension groups also differed in some aspect of LQ. One study that assessed lexical processing and comprehension independently (Landi & Perfetti, 2007) observed specific word-level semantic differences between comprehension groups that differed on comprehension only and not word decoding. Another (Hart, 2005) showed specific language-learning effects for high-LQ participants. It remains an important goal of the LQH to show more specific consequences of the various components of lexical knowledge. The Causes of Low Ability in Reading There are other theories of low ability in reading, and it is useful to place the LQH in the context they provide. Phonological deficits, naming deficits, and semantic deficits are three deficit-based hypotheses at the word and subword level that have been important. Let's consider them one at a time. Phonological processing indeed has a central place in the acquisition of reading skill. The LQH has its roots in one idea about comprehensionwords matter for comprehension (verbal efficiency; Perfetti, 1985)and one idea about word knowledge (the acquisition of effective word representations depends on precision and redundancy in sublexical constituents; Perfetti, 1992). It is the second of these ideas for which phonology is relevant. Procedures for phonological decoding have an important role in establishing high-quality word representations. Indeed, these procedures may be the primary means for establishing word-specific orthographic representations (Share, 1995, 1999). The difference in explanation for reading problems is a matter of developmental level and theoretical emphasis. The phonological deficit hypothesis applies with full force to young children. Their problems in reading words are observable and so is their problem with phonology, and because the latter can explain the former, the phonological hypothesis is privileged. One idea about phonological deficits is that they arise from low-quality phonological representations. The general form of this idea has been part of various explanations about phonological deficits (e.g., Katz, 1986; Snowling, 1995; Snowl-ing, Wagtendonk, & Stafford, 1988). Elbro (1996, 1998) gave it a specific form in the phonological distinctness hypothesis. Elbro argued that the distinctiveness of phonological representations is critical for distinguishing phonologically similar words from each other and suggested that dyslexics had incomplete, nondistinct

LEXICAL QUALITY

379

phonological representations. Although the application of this hypothesis is generally to children, it is important to note that adult variation is word-reading skill is associated with phonological performance in spelling (Dietrich & Brady, 2001). Phonological quality thus is a specific dimension of LQ that may be significant beyond its application to dyslexic children. With development, the payoff for phonological processingand the cost of problems with phonological processingis word knowledge. Thus, although phonological processing may have placed limits on LQ, the focus shifts away from phonological procedures and directly to word knowledge itself as a limiting factor in reading. On this understanding, phonological processes at higher (adult) levels of skill have not lost their importance, but they no longer have the status of direct cause. In the case of naming speed and the corresponding naming deficit, the situation is not so clear. There is neither compatibility nor interesting contrast between LQ and rapid naming if rapid naming is completely general rather than about words. It is possible that rapid naming of words is a by-product of LQ, just as verbal efficiency is. The contrast between rapid naming and verbal efficiency has been about a distinctive role for word reading. Word retrieval can be only as fast as the limit set by general symbol retrieval (Perfetti, 1985). Our earlier research tended to show that among children who were garden variety low-skill readers, slow word-retrieval speed is not explained solely by number or picture retrieval. However, there appear to be ample demonstrations that some very low-skill readers are generally slower at nonreading naming tasks (Wolfe, Bowers, & Biddle, 2000), although some meta-analyses have raised questions about the robustness of rapid naming effects across studies (Swanson, Trainin, Necoechea, & Hammill, 2003). In any case, LQ remains a separate concept to explain general variation in reading and is not reducible to naming speed. The semantic deficit hypothesis, like the phonological hypothesis, is a more specific cousin of LQ. LQ entails both semantic and phonological components. It inherits the weakness of low disconfirmation risk that comes with very general propositions. By encompassing multiple components, the LQH seems to have verification in any study that shows skill-related consequences of any one of these. However, this problem is not intrinsic to the LQH but rather reflects the limitations generally seen in individual differences research. For example, a specific prediction of the LQH is that high coherence among lexical constituents has consequences for reading, but this prediction has not been tested. This is partly because it implies a difficult separation of spelling, pronunciation, and meaning knowledge for specific words and readers so that the only difference is the correlation among these components. Although this separation has not been done, it could be, at least in principle. Furthermore, notice this same problem limits both the semantic and phonological deficit hypotheses, which must be able to convincingly show that in a given group of readers, only one of these deficits is causal. Although the research approximates this standard, it does not assess skill and knowledge on an individual word-by-word basis but rather by

380

PERFETTI

controlling one variable through a global assessment (e.g., phonological decoding) so that it can measure the targeted variable experimentally. In addition to these word-level hypotheses, there are several hypotheses about comprehension-specific deficits that target processes above the word level (e.g., Oakhill et al., 2003). The possibilities are wide ranging (e.g., inferences, comprehension monitoring) and generally involve processes that operate on the outcomes of word-level processes. These higher level accounts might be complementary to the LQH insofar as word meanings can be considered the interface between word identification and comprehension. However, to make this possibility meaningful, one would have to show that word-meaning retrievals are responsible for inferences or for coherence checks or whatever comprehension process is at issue. Alternatively, one might argue that working memory is the link and that low-quality word representations and, for example, inference processes compete for processing resources. A difference among theses various approaches (those at both the low and the high levels) and the LQH is that the phonological, naming, and semantic deficit hypotheses (and also the inference hypothesis) seem to be about mechanisms that are not functioning properly. In contrast, the LQH is about knowledge that has not been acquired or practiced to a high-enough level. In the LQH, processes that do not operate effectively or efficiently arise from knowledge representations. Because knowledge and practice with this knowledge accumulate with age and experience, the application of the LQH is very broad, relevant for reading by adults as well as children. To summarize, the LQH implies that variation in the quality of lexical representations, including both form and meaning knowledge, lead to variation in reading skill, including comprehension. For some readers, the quality problem may be in the semantic constituents of words; for most readers, the problem cuts across meaning, orthographic, and phonological knowledge. The consequences of LQ can be seen in processing speed at the lexical level, and, especially important, in comprehension. There is no on-off deficit in this characterization. LQ is graded across words for a given individual and across individuals for a given word. The source of LQ variation must arise through literacy and language experiences, although effective use of these experiences is likely to be influenced by biology as well as culture. These experiences include, among other things, learning to decode printed words, practice in reading and writing, and engagement with concepts and their language forms.

ACKNOWLEDGMENTS This article is based on the author's Distinguished Research Award address to the Society for the Scientific Study of Reading, Vancouver, July 2006. Some of the

LEXICAL QUALITY

381

studies reviewed in this article were supported by the Institute of Education Sciences (IES), U.S. Department of Education, through Grant R305G020006 to the University of Pittsburgh. IES asks authors to note that any opinions expressed are those of the authors and do not represent views of the U.S. Department of Education. REFERENCES
Breznitz, Z., & Misra, A. (2003). Speed of processing of the visual-orthographic and auditory-phonological systems in adult dyslexics: The contribution of "asynchrony" to word recognition deficits. Brain and Language, 85, 486-502. Breznitz, Z., & Share, D. (1992). Effects of accelerated reading rate on memory for text. Journal of Educational Psychology, 84, 193-199. Cain, K., & Oakhill, J. V. (1999). Inference ability and its relation to comprehension failure in young children. Reading and Writing, 11, 489-503. Curran, T. (1999). The electrophysiology of incidental and intentional retrieval: ERP old/new effects in lexical decision and recognition memory. Neuropsychologica, 37, 771-785. Dietrich, J. A., & Brady, S. A. (2001). Phonological representations of adult poor readers: An investigation of specificity and stability. Applied Psycholinguistics, 22, 383^-18. Elbro, C. (1996). Early linguistic abilities and reading development: A review and a hypothesis. Reading and Writing, 8, 453^-85. Elbro, C. (1998). When reading is "readn" or somethn. Distinctness of phonological representations of lexical items in normal and disabled readers. Scandinavian Journal of Psychology, 39, 149-153. Fleisher, L. S., Jenkins, J. R., &Pany, D. (1979). Effects on poor readers'comprehension of training in rapid decoding. Reading Research Quarterly, 15, 30-48. Gernsbacher, M. A. (1990). Language comprehension as structure building. Hillsdale, NJ: Lawrence Erlbaum Associates. Gernsbacher, M. A., & Faust, M. (1991). The mechanism of suppression: A component of general comprehension skill. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17, 245-262. Haenggi, D., & Perfetti, C.A. (1994). Processing components of college level reading comprehension. Discourse Processes, 17, 83-104. Hart, L. A. (2005). A training study using an artificial orthography: Effects of reading experience, lexical quality, and text comprehension in LI and L2. Unpublished doctoral dissertation, University of Pittsburgh. Hart, L., & Perfetti, C. A. (in press). Learning words in Zekkish: Implications for understanding lexical representation. In E. Grigorenko & A. Naples (Eds.), Single word reading. Mahwah, NJ: Lawrence Erlbaum Associates. Katz, R. B. (1986). Phonological deficiencies in children with reading disability: Evidence from an object-naming task. Cognition, 22, 225-257. LaBerge, D., & Samuels, S. J. (1974). Toward a theory of automatic information processing in reading. Cognitive Psychology, 6, 293-323. Landi, N. (2005). Behavioral and electrophysiological investigations of the relationship between semantic processing and reading comprehension. Brain and Language, 102, 30^-5. Landi, N., & Perfetti, C. A. (2007). An electrophysiological investigation of semantic and phonological processing in skilled and less-skilled comprehenders. Brain and Language. Martin-Chang, S. L., & Levy, B. A. (2005). Fluency transfer: Differential gains in reading speed and accuracy following isolated word and context training. Reading and Writing, 18, 343-376.

382

PERFETTI

Nation, K. (2006). Children's reading comprehension difficulties. In M. J. Snowling & C. Hulme (Eds.), The science of reading: A handbook (pp. 227-247). Oxford, UK: Blackwell. Nation, K., & Snowling, M. J. (1998). Semantic processing and the development of word recognition skills: Evidence from children with reading comprehension difficulties. Journal of Memory and Language, 39, 85-101. Nation, K & Snowling, M. J. (1999). Developmental differences in sensitivity to semantic relations among good and poor comprehenders: evidence from semantic priming. Cognition. 70, B1-B13. Oakhill, J. V., Cain, K., & Bryant, P. (2003). The dissociation of word reading and text comprehension: Evidence from component skills. Language and Cognitive Processes, 18, 443-468. Perfetti, C. A. (1985). Reading ability. New York: Oxford University Press. Perfetti, C. A. (1992). The representation problem in reading acquisition. In P. B. Gough, L. C. Ehri, & R. Treiman (Eds.), Reading acquisition (pp. 145-174). Hillsdale, NJ: LawrenceErlbaum Associates. Perfetti, C. A, & Hart, L. (2001). The lexical bases of comprehension skill. In D. Gorfien (Ed.), On the consequences of meaning selection (pp. 67-86). Washington, DC: American Psychological Association. Perfetti, C. A., & Hart, L. (2002). The lexical quality hypothesis. In L. Vehoeven. C. Elbro, & P. Reitsma (Eds.), Precursors of functional literacy (pp. 189-213). Amsterdam: John Benjamins. Perfetti, C. A., Landi, N., & Oakhill, J. (2005). The acquisition of reading comprehension skill. In M. J. Snowling & C. Hulme (Eds.), The science of reading: A handbook (pp. 227-247). Oxford, UK: Blackwell. Perfetti, C. A., Wlotko, E. W., & Hart, L. A. (2005). Word learning and individual differences in word learning reflected in Event-Related Potentials. Journal of Experimental Psychology: Learning Memory and Cognition, 31, 1281-1292. Perfetti, C. A., Yang, C-L., & Schmaihofer, F. (in press). Word-to-text integration processes: Neuro-cognitive indicators and reading skill differences. Applied Cognition. Quinon, M. (July 2007). World Wide Words. www.WorldWideWords.org. Rugg, M. C. (1995). Event-related potential studies of human memory. In M. S. Gazzaniga (Ed.), The cognitive neurosciences (pp. 789-801). Cambridge, MA: MIT Press. Shankweiler, D., Katz, L., Stuebing, K., Fletcher, J. M., Brady, S., Fowler, A., Dreyer, L. G., Marchione, K. E., Shaywitz, S. E., & Shaywitz, B. A. (1999). Comprehension and decoding: Patterns of association in children with reading difficulties. Scientific Studies of Reading, 3, 69-94. Share, D. L. (1995). Phonological recoding and self-teaching: Sine qua non of reading acquisition. Cognition, 55, 151-218. Share, D. L. (1999). Phonological recoding and orthographic learning: A direct test of the self-teaching hypothesis. Journal of Experimental Child Psychology, 72, 95-129. Snowling, M. J. (1995). Phonological processing and developmental dyslexia. Journal of Research in Reading, 18, 132-138. Snowling, M. J., Wagtendonk, B.. & Stafford, C. (1988). Object naming deficits in developmental dyslexia. Journal of Research in Reading, 11, 67-85. Stothard, S., & Hulme, C. (1996). A comparison of reading comprehension and decoding difficulties in children. In C. Cornoldi & J. Oakhill (Eds.), Reading comprehension difficulties: Processes and interventions. Mahwah, NJ: Lawrence Erlbaum. Swanson, H. L., Trainin, G., Necoechea, D. M., & Hammill, D. D. (2003). Rapid naming, phonological awareness, and reading. A meta-analysis of the correlational evidence. Review of Educational Research, 73, 407-444. Tan, A., & Nicholson, T. (1997). Flashcards revisited: Training poor readers to read words faster improves their comprehension of text. Journal of Educational Psychology, 89, 276-288. van Berkum, J. J. A., Hagoort. P.. & Brown. C. M. (1999). Semantic integration in sentences and discourse: Evidence from the N400. Journal of Cognitive Neuroscience, 11, 657-671. Wolfe, M., Bowers, P. G., & Biddle, K. (2000). Naming-speed processes, timing, and reading. Journal of Learning Disabilities, 33, 387^107.

LEXICAL QUALITY

383

Yang, C-L., & Perfetti, C. A. (2006, July). Reading skill and the acquisition of high quality representations for new words. Paper presented at the annual meeting of the Society for the Scientific Study of Reading, Vancouver. Yang, C. L., Perfetti, C. A., & Schmalhofer, F (2005). Less skilled comprehenders' ERPs show sluggish word-to-text integration processes. Written Language & Literacy, 8, 233-257. Yang, C-L., Perfetti, C. A., & Schmalhofer, F. (2007). Event-related potential indicators of text integration across sentence boundaries. Journal of Experimental Psychology: Learning, Memory, and Cognition, 33, 55-89.

Journal of Experimental Psychology: Learning. Memory, and Cognition 2005, Vol. 31, No. 6, 1281-1292

Copyright 2005 by the American Psychological Association 0278-7393/05/S12.00 DOl: 10.1037/0278-7393.31.6.1281

Word Learning and Individual Differences in Word Learning Reflected in Event-Related Potentials
Charles A. Perfetti, Edward W. Wlotko, and Lesley A. Hart
University of Pittsburgh Adults learned the meanings of rare words (e.g., gloaming) and then made meaning judgments on pairs of words. The 1st word was a trained rare word, an untrained rare word, or an untrained familiar word. Event-related potentials distinguished trained rare words from both untrained rare and familiar words, first at 140 ms and again at 400-600 ms after onset of the 1st word. These results may point to an episodic memory effect. The 2nd word produced an N400 that distinguished trained and familiar word pairs that were related in meaning from unrelated word pairs. Skilled comprehenders learned more words than less skilled comprehenders and showed a stronger episodic memory effect at 400-600 ms on the 1st word and a stronger N400 effect on the 2nd word. These results suggest that superior word learning among skilled comprehenders may arise from a stronger episodic trace that includes orthographic and meaning information and illustrate, how an episodic theory of word identification can explain reading skill. ' Keywords: word learning, ERP study, N400, and vocabulary learning

Adult English speakers, know the meanings of thousands of words and are vaguely familiar with many more, on one estimate learning about 3,000 new words a year from the beginning of literacy (Nagy & Herman, 1987). Nearly all college students know the meanings of even many low-frequency nouns such as rubble, flint, and abstention, all of which have printed word frequencies of less than 5 per million words of text in some word counts (e.g., Kucera & Francis, 1967). However, most students do not know the meaning of gloaming, ibex, and agog, rare words that fail to occur in some of these word counts (e.g., Kucera & Francis, 1967). However, just as many people learn the meaning of abstention from some reading or spoken language experience, some will also encounter gloaming, ibex, or some other word they do not know, and perhaps add its form and something about its meaning to their mental lexicon. Our interest here is in examining the consequences of learning a new word for subsequent encounters with the word. Of course, if the meaning of the word has been learned, we should observe that the learner can recognize the word and understand its meaning. However, beyond this behavioral outcome, the process of reading the word, the time course of its identification, and meaning retrieval processes also should be affected. Recordings of event-related potentials (ERPs) may expose the consequences of learning in a word-processing task.

Charles A. Perfetti, Edward W. Wlotko, and Lesley A. Hart, Learning Research and Development Center, University of Pittsburgh. This research was supported by U.S. Department of Education Grant R305G020006. We gratefully acknowledge the important contributions of Dayne Grove and Chin-Lung Yang to data analysis and the helpful comments of Erik Reichle and Gwen Frishkoff on a draft of the article. Correspondence concerning this article should be addressed to Charles Perfetti, Learning Research and Development Center, University of Pittsburgh, 3939 O'Hara Street, Pittsburgh, PA 15260. E-mail: perfetti@pitt.edu 1281

Beyond using ERPs to expose the consequences..of new word ._ learning, we examine a corollary question about individual differences in reading comprehension skill. Comprehension skill among children and adults is supported by their knowledge of words, including, according to the Lexical Quality Hypothesis (Perfetti & Hart, 2001), the precision of the reader's representation of orthography, phonology, and meaning, as well as the sheer number of known words. Skill in reading comprehension, to the extent that it has a word knowledge component, may also support the ability to learn the meanings of new words. Skilled readers may be better able to take advantage of word training events by remembering a new association between an orthographic form and a meaning. If so, we may observe the consequences of differential learning in an ERP component that reflects memory for recently learned words. In examining these issues, we exploit two well-established ERP facts. One is that ERPs reveal the differences between "old" and "new" words in recognition memory experiments (Curran, 1999; Rugg & Doyle, 1992; Rugg & Nagy, 1989). A previously encountered word produces a late positive-going wave (peaking at around 600 ms) following the onset of the word compared with a word not previously encountered in the experiment ("new word"). This late positive wave (or P600 component) is thus a marker of an episodic memory trace (see Rugg, 1995 for a review). Thus, if we have recently taught the meaning of a word to a participant, presenting this word should evoke a P600 compared with a word that was not taught. The second fact is that an ERP component, the N400, reflects the meaning congruence between a word and its previous context. In sentence contexts (Kutas & Hillyard, 1980) and in single-word semantic priming contexts (Nobre & McCarthy, 1994), a word that is incongruent with its context produces a negative-going wave peaking at about 400 ms after the onset of the word, whereas a congruent word produces a reduced N400. Thus, if we test whether a participant has learned the meaning of a word that was taught by presenting the taught word followed by a word that could be related in meaning, we expect to see a reduced N400 in the related case, compared with the unrelated case.

1282

PERFETTI, WLOTKO, AND HART did not appear in the Kucera & Francis (1967) corpus of over a million words, 51 relatively common words (a Kucera-Francis frequency rating of 40 words per million or greater), and 64 pseudowords, that is, legally spelled and pronounceable nonwords. They were instructed to mark only the letter strings that they were sure were real words in English. From the results of this task, a stimulus list was constructed for each participant by randomly selecting for training 60 of the rare words that the participant failed to mark as words. The remaining rare words that were not marked as words by the participant became the set of untrained rare words. Those relatively common words that the participant did mark as words became the set of familiar words. Thus, this procedure resulted in three classes of words, two classes of rare wordstrained and untrainedand a class of familiar words. Notice that the procedure produced a different set of randomly selected trained words for each participant. Training Participants studied the 60 rare training words for 45 min. The experimenter presented them with flashcards containing words on the front and their definitions on the back. For example, gloaming was defined as "the twilight period before dark" and clowder was "a collection or group of cats." Participants were instructed that they would be given 45 min to learn the 60 words and that they should become as familiar with the words and their definitions as they could, with the understanding mat they might not be able to learn all of them. Posttraining Semantic Decisions After the training period, participants performed a semantic decision task while electroencephalograms (EEGs) were recorded. On each trial, a word was selected randomly from the total set of words (trained, untrained, and familiar) and presented for 1,000 ms. The offset of this word was followed immediately by a second word, a meaning probe, which remained visible while the participant responded with a button press to indicate whether it was related in meaning to the first word. Each trial was preceded by a fixation cross for 400 to 550 ms. (Variability.w.as..added.to reduce any. impact of nonstimulus-related time-locked electrical activity.) On half of the trials, the probe word was semantically related to the first word; on the other half, the two words were unrelated. Semantically related pairs were created by experimenter judgment. Many trained words were paired with a meaning probe that had occurred as part of the definition (e.g., clowder-cats), but others were paired with a word that had not occurred as part of the training definition (e.g., gloaming-evening). Semantically unrelated pairs were created by shuffling the word pairs in each individual participant's stimulus list. Each stimulus list contained the pairing of a stimulus word with both a related and an unrelated probe word. The order of stimulus words as well as the order of related and unrelated probe words was randomized for each participant. Participants were instructed to press the "1" key with their right index finger if the two words in a trial were related in meaning and to press die "2" key with their right middle finger if the two words were not related in meaning. The meaning probe word was removed from the screen when a response was made or after 2,000 ms elapsed, whichever came first. Accuracy feedback (correct or incorrect response) was presented after each trial in me form of a stylized smiling face for correct answers and a frowning face for incorrect answers. The feedback image remained in view for 800 ms prior to the onset of the next trial. Recordings Scalp potentials were recorded from 128 sites with Electrical Geodesic, Inc.'s (EGI's) Geodesic Sensor Net with Ag/AgCl electrodes. The potentials were recorded with a sampling rate of 500 Hz and a hardware bandpass filter of 0.1 to 200 Hz. Impedances generally were kept below a

In the study reported here, college students first learned the meanings of very rare English words (e.g., gloaming). We controlled the selection of rare words for each participant individually such that the rare words were unfamiliar to a given participant prior to the experiment. The learning phase of the study was very simplethe presentation of the rare word on one side of a card and a brief definition on the opposite side. Following learning, participants made meaning judgments about pairs of words, presented one at a time while we recorded the ERPs. In the meaning judgment task, the first word was from one of three categories: the rare words that we had just taught to this participant, rare words that we did not teach to the participant, and familiar, medium-frequency words that the participant had not seen during the experimental session. Each word was followed by a word that was either related or unrelated in meaning, and the participant pressed a "yes" or "no" button to indicate their decision about whether the two words were related in meaning. For example, gloaming followed by twilight would require a "yes" response. We hypothesized that following word learning, we would see ERP evidence, during a meaning judgment task, that the participant had become familiar with the words we taught. The evidence for this would come from an indicator of episodic memory, specifically a late positivity (P600) that would show differences between trained words and both untrained rare words and familiar (but not recently viewed) words. We hypothesized also that during the reading of a second word, we would see evidence for a meaning match for trained rare and familiar words, compared with untrained rare words, in a reduced N400. Finally, we sought to test the hypotheses that skilled comprehenders would show more effective learning of rare words and that ERP differences would reflect stronger recognition and meaning match components. In addition to these targeted tests, we assumed an exploratory approach to other ERP results, for example, whether we could observe a meaning- response as- distinct from an episodic memory response. If so, some ERP component should separate both familiar words and trained rare words from untrained rare words. Method ' Participants

Twenty-four undergraduates from the University of Pittsburgh Psychology Department participant pool provided data for the experiment. Data from 6 additional participants were not analyzed because of hardware malfunction, and data from 2 participants were rejected due to excessive recording artifacts. Participants were invited from a larger pool whose individuals had previously completed a variety of reading-related tasks. Twelve were skilled readers and 12 were less skilled, as determined by performance on the Nelson-Denny comprehension test. The two skill groups were drawn from pools that included the top 20% of those tested (skilled comprehenders) and the bottom 20% of those tested (less skilled comprehenders). All skilled comprehenders exceeded an accuracy of 64% on the Nelson-Denny; no less skilled comprehender exceeded an accuracy of 50%. Procedure Word Selection Participants first completed a paper-and-pencil word detection task. They viewed a list of 250 letter strings, consisting of 135 rare words that

WORD LEARNING AND EVENT-RELATED POTENTIALS threshold of 40 kfi. A digital low-pass elliptical filter of 30 Hz was applied to the recordings. The ERPs were stimulus-locked averages consisting of a 100-ms baseline and a 1,800-ms epoch defined by the presentation of the stimulus word (Word 1 = 1,000 ms) plus the meaning probe (Word 2 = 800 ms). Bad channels were removed from the recordings and replaced by spherical spline interpolation with data from the remaining channels. Trials containing eye movement, eyeblink, and channel artifacts were rejected and not used in analysis. ERPs were transformed using the average reference. Finally, the ERP segments were corrected relative to a 100-ms baseline. Following rejection of trials with artifacts, two thirds of the participants had 30 trials per condition, and all had at least 20 trials. Results Behavioral Results Tables 1 and 2 show the accuracy and decision time results for the meaning judgment task. The results showed higher accuracies for familiar words (87.3%, SD = 0.02%) and trained rare words (83.7%, SD = 0.01%) compared with untrained rare words (56.4%, SD = 0.10%). For accuracy, an analysis of variance (ANOVA) showed a main effect of word type, F(2, 44) = 331.00, p < .01, and relatedness, F(l, 22) = 7.58, p = .01, as well as a Word Type X Relatedness interaction, F(2, 44) = 50.70, p < .01. The interaction showed a "no" response bias for the untrained rare words only: a tendency to judge an untrained word and its meaning probe as"~unrelated: There was also a'Word Type X Skill interaction, F(2, 44) = 6.72, p < .03, which reflects the fact that skilled comprehenders were more accurate than less skilled comprehend-ers for trained rare words (about 10% difference) but not for untrained rare words (<1% difference) or for familiar words (about 1% difference). As shown in Table 2, response times varied between 700 ms and 900 ms, depending on word type, relatedness, and correctness of response. Responses to untrained rare words were slower than to trained and familiar words, responses to related words were faster than responses to unrelated words, and correct responses were faster than incorrect responses. An analysis of meaning decision times showed a main effect of word type, F(2, 42) = 8.17, p < .01; relatedness, F(l, 21) = 10.14, p = <.01, and response accuracy, F(l, 21) = 57.41, p < .01. However, an interaction showed that the difference between correct and incorrect responses was present for trained and familiar words only, with untrained rare words showing no effect, F(2, 42) = 20.34, p < .01. Table 1 Behavioral Results: Percent Accuracy Condition Familiar
Related Unrelated M :ained rare Related Unrelated M ntrained rare Related Unrelated M 87.6 88.3 88.0 89.4 87.8 88.6 40.4 69.6 55.0 87.1 86.3 86.7 81.6 76.0 78.8 39.6 69.0 54.3 87.4 87.3 87.4 85.5 83.7 81.9 40.0 69.3 54.7

1283

Event-Related Potential Results To take advantage of the high-density recordings across the full epoch, we carried out a temporal principal components analysis (PCA) on the ERPs. The logic of PCA is to use the full set of electrodes and time points to determine the intercorrelations of ERP shifts over time, allowing an overall data-driven view, from which factors emerge for further statistical testing. The temporal variables consisted of 900 time points across two words (1,800 ms at 2-ms samples), with the data consisting of the recording from each electrode in each condition for each participant. Thus, this PCA analyzes an epoch that consists of two words, allowing us to observe not only the ERP effects on the first wordthe trained rare, untrained rare, and untrained familiar wordsbut also on the second word, which participants used to make a decision. The PCA used a correlation matrix and a promax rotation (K = 4).1 Because the promax rotation does not assume that the factors are orthogonal, it has some advantages for temporally correlated ERP data (Dien & Frishkoff, 2005; Dien, 1998). The results of the PCA showed 6 factors with eigenvalues greater than 10, which we retained for further analysis. To these, we added Factor 11, which corresponded to 400 ms after the onset of the probe word, a time point for which we hypothesized a meaning congruence indicator (N400). Together, these factors accounted for 93.3% of the total variance. The factor scores for these seven factors, shown in Figure 1, were analyzed with repeated-measures ANOVA. Fifteen scalp locations were chosen for analysis based on the 10/10 system, a set of 10 locations commonly reported in ERP research, plus 5 additional locations. To each of these 15 recording locations (3 central, 3 parietal, 4 temporal, and 2 occipital), we added those channels immediately adjacent. This creates a cluster scheme in which the data from a cluster are the average of between 4 and 8 electrodes, with most clusters having 7 electrodes. The original waveforms for the 15 single electrodes are shown in Figure 2, and the clusters are shown in Figure 3. We based our clusters on standard 10/10 locations in order to have our results more easily compared with

Skilled

Less skilled

PCA involves a decision about the form of the association matrix that calculates relationships between pairs of variables by associating their data points. Matrices are usually of one of two types: covariance and correlation. A significant methodological literature has developed around the consequences of choosing one over the other. A recent analysis concluded that misallocation of variance is more likely with correlation than covariance matrices (Dien & Frishkoff, 2005), whereas other analyses conclude that differences are negligible for most ERP data (e.g., Chapman & McCrary, 1995; Van Boxtel, 1998). We used a correlation matrix, which, because it normalizes variances across variables, allows all variables equal weight in determining the factor structure. Although this can increase the chances that noise will influence the factor solution, it also can help to detect small but theoretically important variables in the PCA, something that seems advantageous in the relatively unexamined word learning question we are studying here. To test the robustness of our conclusions over these matrix alternatives, we also carried out a covariance-based PCA. Its factor solutions converged on those we report here, with only minor differences on the second word that do not affect our interpretation.

1284

PERFETTI, WLOTKO, AND HART Table 2 Behavioral Results: Semantic Decision Times (in Milliseconds) Skilled Less skilled M Correct Incorrect M M

Condition Familiar Related Unrelated M Trained rare Related Unrelated M Untrained rare Related Unrelated M Overall M

Correct

Incorrect

730 763 747 700 771 736 891 866 879 787

897 917 907 909 878 894 873 904 889 896

814 840 827 805 825 815 882 885 884 842

702 743 723 709 758 734 840 811 889 768

778 886 832 823 837 830 804 856 826 830

740 815 778 766 798 782 822 834 828 796

111 828 801 786 812 799 852 860 857 819

any effects in the previous literature while still taking advantage of the benefits of clustering.2 For the ANOVAs, the 15 clusters organized factors corresponding to hemisphere (left, right, midline) and lobe (frontal, central, parietal, two temporal locations, and occipital). Because temporal and occipital locations do not have a "midline," we performed separate ANOVAs for midline and lateral locations. We tested the lateral locations using a Hemisphere (2) X Lobe (6) X Word Type (3) X Relatedness (3) X Skill (2, between subjects) ANOVA. The midline locations were tested with a Lobe (3) X Word Type (3) X Relatedness (2) X Skill (2, between subjects) ANOVA. For factors that did not show sphericity across factor levels, the Greenhouse-Geisser correction was used. Our analysis spans a time periodacross the-presentation of two words with the factors identified by the PCA. Four factors were associated with the time period of Word 1, the stimulus word: at 500 ms (Factor 2), 234 ms (Factor 4), at 140 ms (Factor 5), and 64 ms (Factor 6) after the onset of Word 1. Three factors peak within the time period of Word 2, the probe word: one peaks at 372 ms (Factor 11) after the onset of Word 2; another (Factor 1) is a broad "slow wave" factor that begins to rise sharply at approximately

300 ms after the onset of Word 2 and peaks at 626 ms after the onset. Factor 3 is a two-word factor, rising from about 600 ms after the onset of the first word and peaking at 32 ms after the onset of Word 2. In reporting the analyses of these factors below, we group them according to whether their peaks occurred during the first word or the second word and refer to them by the latencies of their peaks. Because two factors (peaks at 64 and 234 ms during Word 1) showed no effects of experimental conditions, we do not report their analysis below. The 64-ms peak appears to reflect exogenous factors associated with the onset of the first word. The 234 ms factor rises again in a less pronounced form at the same latency following the onset of Word 2, suggesting a word-onset related process. Its timing and topography (positive going.in.frontal sites, negative in posterior) are similar to ERP components that have been interpreted as graphic processing (Liu & Perfetti, 2003; Liu, Perfetti, & Hart, 2003) and graphic-phonological coding (Barnea & Breznitz, 1998) as well as more general attention (Hackley, Woldorff, & Hillyard, 1990), feature detection (Luck & Hillyard, 1994), and short-term storage (Chapman, McCrary, & Chapman, 1978) processes.

Word 1
Two Words: Promax PCA yes/no 140 ms. The 140-ms factor showed a main effect of word type in the midline ANOVA, F(2, 44) = 3.97, p = .03. Trained words were more negative than untrained words and familiar words. Pairwise tests showed that trained rare words were distinguished from both familiar (p .03) and untrained rare words (p = .01), which did not differ. The greater negativity at 140 ms for trained words, although not pronounced, can be seen in all frontal and central electrodes as well as left temporal and parietal sites. The 2 ERP clustering does not yet have a standardized procedure and may not be appropriate for applications to traditional low-density electrode recordings. However, for high-density recordings, clustering, or regional averaging, of electrodes has advantages. It reduces the noise associated with individual electrodes while allowing a large number of electrodes to be used (Dien & Santuzzi, 2005).

Time(ms) Figure 1. Principal components analysis (PCA) factors retained for analysis. Factors 1, 2, 3, 5, and 11 showed significant effects of experimental variables. Factors 4 and 6 did not show effects of experimental variables, instead reflecting general task effects.

WORD LEARNING AND EVENT-RELATED POTENTIALS

1285

P7 (59)

P8 (92)

Familiar Trained Rare Untrained Rare

Ol (72) 02 (77)

Figure 2.

Grand average waveforms for 15 electrodes.

1286

PERFETTI, WLOTKO, AND HART

Cluster Locations 10/10 System F3 Fz F4 C3 Cz C4 'T7 T8 P3 Poz P4 P7 P8 01 02 EGI 128 25 11 124 37 VREF 105 46 109 53 68 87 59 92 72 77

Figure 3. Electrode clusters. Fifteen electrodes from the 10-10 were defined as cluster centers. EGI Electrical Geodesic, Inc. sensor; VREF = voltage-reference electrode. general topography for this early time window is a negativity at frontal and central sites, with a positivity at parietal sites, midline lobe effect, F(2, 44) = 6.96, p < .01, Greenhouse-Geisser e = .717. The lateral analysis also showed this pattern, bilaterally negative at frontal and central locations and positive at parietal, posterior temporal, and occipital locations, lobe effect F(5, 110) = 11.25, Greenhouse-Geisser s = .264. 500 ms. In this time period, trained words were more positive than other word types in posterior locations. This trained-word factor appears to correspond to the old-new memory effect (P600 or late positive complex [LPC]) that is found for recently viewed and recognized items. Furthermore, this trained-word factor was more pronounced for skilled comprehenders than less skilled com-prehenders. These conclusions are supported by Word Type X Lobe interaction, F(10, 15) = 4.73, p < .01, Greenhouse-Geisser s = .43, and a Word Type X Lobe X Skill interaction, F(10, 15) = 3.76, p < .01, Greenhouse-Geisser s = .786 in the lateral ANOVA; a Word Type X Lobe interaction also occurred in the midline ANOVA, F(4, 88) = 5.24, p < .01, Greenhouse-Geisser e = .703. Figure 4 shows the Word Type X Lobe X Skill interaction and the waveforms for the Poz cluster that reflects the interaction. The more pronounced positivity is visible in this central parietal cluster as well as in individual parietal electrodes (and Cz) of Figure 2.3 32 ms After Onset of Word 2. The factor that spans the first and second word appears to be sensitive to a separation of all three word types. This factor rises from 6Q0 ms of the first word, spilling over into the presentation of the probe word. Word Type X Lobe interactions were present for bom the lateral ANOVA, F(10, 220) = 7.71, p < .01, Greenhouse-Geisser e = .36 and midline ANOVA, F(4, 88) = 22.75, p < .01, Greenhouse-Geisser s = .333. The midline analysis showed the frontal sites to be more positive for familiar words than for rare words (both trained or untrained), whereas the central and parietal sites distinguished both familiar and trained rare words (more negative) from untrained rare words. The lateral analysis also showed that familiar and trained words (more negative) were distinguished from untrained words (less negative) in the central sites, whereas trained rare words were less positive than familiar and untrained rare words in frontal sites. However, this pattern was modified by a Word Type X Lobe X Hemisphere interaction in the lateral ANOVA, F(10, 220) = 3.59, p < .01, Greenhouse-Geisser s = .443.-Trained words were positive only in left frontal sites, whereas
3

In addition to the analysis based on the two-word epoch reported here, we carried out an analysis separately for each word, 1,000 ms for Word 1 and 1,000 ms for Word 2. The trained-word effect actually is seen even more clearly in the separate word analysis. In general, however, the two-word and separate word analyses showed very similar patterns, except that only in the two-word epoch analysis can one see a factor that overlaps the offset of the first word and the onset of the second word.

WORD LEARNING AND EVENT-RELATED POTENTIALS

1287

Word type x Lobe x Skill

0.8 0.6 0.4 <u S
Familiar *Trained Rare untrainea nare T

0.2 u , u 0 u -0.2 to
-0.4 -0.6

1
^4
;

T
^NS -

T 4-^/1-t.

k\

Cr^J ]
-L

y
5

CJ

'S CM

.& u O

to

a U

'B P
H

LessS killed

Skilled

Poz - Less Skilled

Poz -High Skilled -Familiar -Trained Rare -Untrained Rare

Time (ms)

Time (ms)

Figure 4. Word Type X Lobe X Skill interaction for Word 1, 500-ms factor (top) and Poz clusters for skilled and less skilled readers showing the interaction 500-600 ms after the onset of Word 1 (bottom).

untrained words were bilaterally positive, with familiar words showing an intermediate pattern, bilateral but more positive in left than right sites. This pattern can be seen in the waveforms in Figure 2. Finally, both the midline and lateral analyses showed a Word Type X Skill interaction, respectively, F(2, 44) = 5.57, p < .01, Greenhouse-Geisser e = .935, and F(2, 44) = 4.74, p = .02, Greenhouse-Geisser e = .920. The distinction among the three word types was more pronounced for skilled comprehenders. As can be seen in the midline interaction shown in Figure 5, skilled comprehenders were sensitive to differences among all three word types in this time window, whereas less skilled comprehenders showed less sensitivity, especially to the difference between familiar and trained words.

Word 2 372 ms. Because the task required a semantic decision, we hypothesized an N400 relatedness effect as the participant read the second word; accordingly, we tested the 372 ms factor, although its eigenvalue was less than 10. The hypothesis predicts relatedness effects according to whether Word 2 was related or unrelated in meaning to Word 1. The lateral analysis showed a significant relatedness effect, F(l, 22) = 4.67, p = .04, and significant interactions of Word Type X Relatedness X Hemisphere, F(2, 44) = 3.97, p = .03, and Word Type X Lobe X Hemisphere X Skill, F(10, 220) = 3.12, p = .02, Greenhouse-Geisser e = .41. The midline ANOVA showed interactions of Word Type X Re-

Temp r 3/ o a 4 l Temp r 5/ o a 6 l ita l

.& uO

Temp r 3/ o a 4 l Temp r 5/ o a 6 l ita l

;ta l

-0.8

it a l

1288

PERFETTI, WLOTKO, AND HART Word type x Skill

Less Skilled Figure 5.

Skilled

Word Type X Skill, two-word factor, with peak at 32 ms after onset of Word 2. Type X Lobe, lateral F(10, 220) = 5.02, p < .01, Greenhouse-Geisser s = .415, and midline F(4, 88) = 5.70, p < .01, Greenhouse-Geisser e = .618. The interaction corresponds to the late separation of the word types, especially in frontal regions, reduced in posterior locations. The general pattern is that familiar words are most positive and untrained rare words are most negative, as can be seen in the waveforms of Figure 2. This pattern may reflect meaning retrieval or verification processes that are stronger for the familiar words than for rare words, especially untrained rare words. Discussion Our results demonstrate that ERP measures can be used as indicators of word learning. When people learned the meaning equivalence of a rare word such as gloaming, the consequences of this learning were observable when the learners made meaning judgments on the word. Accuracy of meaning judgments was about 84% for 60 words following 45 min of training, comparable to the accuracy on medium-frequency words already familiar to the learners. More interesting is that the effects of learning were observed in ERP records as well as in behavioral measures. The effect of training was seen in a late positive shift that we identify as an episodic training effect, similar to an old-new P600 observed in recognition memory. Trained words showed the effect, whereas untrained rare words and familiar words did not. We interpret this effect as an episodic memory indicator, that is, that

latedness, F(2, 44) = 6.74, p < .01, and Word X Lobe, F(4, 88) = 8.76, p < .01. (See Figures 6 and 7.) The interactions confirmed that an N400 effect, which was larger in the right than the left recording locations, was present for familiar words and trained words, but not untrained words. The interactions also reflected a larger N400 effect for skilled comprehenders than for less skilled comprehenders. As can be seen in Figure 6, skilled comprehenders showed a stroiiger relatedness effect, especially in midcentral and parietal sites. Figure 7 shows that the word type patterns for the two skill groups were similar in right hemisphere sites (although frontal sites were more positive for skilled comprehenders); however, in the left hemisphere temporal (T3) cluster, less skilled comprehenders showed no separation of word types, whereas for skilled comprehenders, familiar and trained words were separated clearly from untrained rare words. Figure 8 shows the basic N400 effect in the average waveforms from the right parietal (P4) cluster, where the reduced negativity for related words can be seen for familiar and trained words. Notice the lack of an effect for untrained rare words, for which one expects their relatedness to be undetected. Slow wave factor. The "slow wave" factor, a typical component in PCAs for ERP data, can include noncognitive time-locked factors as well as cognitive factors. In these data, this factor reflected a clear cognitive component in both ANOVAs. A word type effect appeared in both the lateral, F(2, 44) = 9.32, p < .01, Greenhouse-Geisser s = .780, and midline analyses, F(2, 44) = 17.330, p < .01. Both analyses also showed an interaction of Word

WORD LEARNING AND EVENT-RELATED POTENTIALS

1289

Relatedness x Lobe x Skill

Figure 6.

Relatedness X Lobe X Skill interaction for Word 2 at 372 ms.

participants were recognizing the trained words as recently experienced during training. However, the behavioral data indicate that more than mere recognition occurred, because participants were nearly as accurate on the rare trained words as on familiar words. Moreover, trained rare words separated from both untrained and familiar words even prior to this late positive shift, at around 140 ms after the onset of Word 1. This time window is rather early to be interpreted in the same manner as the late positivity old-new effect. However, it might reflect a process in which visual attention is drawn to features of a word that has been recently viewed, a slightly different form of an old-new effect reflecting something less than full orthographic analysis. Although word-related components, including a sensitivity to word frequency, have been observed around 140 ms in a study by Sereno, Rayner, and Posner (1998), the topography in mat study differed from what we found here. Other studies have found early word-processing components at around 170 ms (Bentin, Mouchetant-Rostaing, Giard, Echallier, & Pernier, 1999), but again the topography is different. Furthermore, what is distinctive in the present study is the sensitivity of an early component to the episodic status of the word, not its frequency or orthographya component not dependent on the word itself but on its recent exposure history. Neither the early nor the later separation of trained words from other words can be taken as an indicator of learning word meanings. For a learning indicator, we have, in addition to behavioral results, ERP data from the N400, which was observed during the presentation of the second word, the meaning probe. In a semantic judgment task, an unrelated probe word is expected to produce a

larger N400 than a related probe word, because the unrelated probe is semantically incongruent with the first word. In the present study, unrelated probes for both familiar and trained words produced a large negative deflection in the N400 compared with related probes. In contrast, the unrelated probes to untrained rare words produced no N400 effect, because participants did not know the meanings of the untrained words. This gives us further evidence that participants learned something about the meaning of the trained words that allowed a congruence effect to be observed on a following meaning probe. We also found several interesting differences between skilled and less skilled comprehenders, evidenced in both behavioral and ERP data. First, skilled comprehenders were reliably more accurate than less skilled comprehenders in meaning judgments on recently trained words. This result may reflect slightly better learning by the skilled readers. On this interpretation, skilled comprehenders, who also have larger vocabularies, are better able to use their word knowledge to add new words to their vocabularies or are simply better at learning new associations or retaining specific episodic information. Although one might consider other explanations, for example, some familiarity for the rare words for skilled comprehenders, we point out again that the words used for training were individually selected as words that a given participant did not recognize as a real word. Furthermore, no skill effects were observed for the equally unfarmliar rare words that were not trained. These facts suggest that the differences between skill groups emerged during encounters with the word during learning that are reflected at testing.

1290

PERFETTI, WLOTKO, AND HART

Word type x Lobe x Hemisphere x Skill

0.8 0.6 0.4 O 0.2

Fatmhar

1
W-0.2 -0.4

t 1 1E
---a
43

. ---- \

* Trained Rare * Untrained Rare

V
in

_ _ _

-0.6
T3

-0.8

a r> s

!P

a S

-*

.
rr

TS

a
.ts

a
fi

^f

yi

empora l emp o

emp o emp ra o l

emp ra o l

te,

Left

o O

a.

O.

empo

'S 'H
H

a,

O Left

# o

ft.

c U

fl

Right HEM Skilled

Right Hi ,M Less kille 1 E c

Figure 7.

Word Type X Lobe X Hemisphere X Skill interaction for word 2 at 372 ms. HEM = hemisphere.

The 10% accuracy differences between skilled and less skilled comprehenders (and the absence of differences for untrained and familiar words) are mirrored by ERP differences that help explain them. A general result is that ERPs of less skilled readers showed less sensitiyity to the differences among the three word types. More specifically, skilled comprehenders showed a larger episodic memory effect, the P600 that distinguished trained rare words from other words. This suggests directly that the training produced a stronger memory trace of the word for skilled comprehenders. A related difference in a P600 effect was reported in a study of recognition memory by Riisseler, Probst, Johannes, and Miinte (2003), who found that a P600 old-new effect was obtained for normal adult readers (more positive for old words in a left parietal electrode) but not adult dyslexics. It is interesting to note tiiat although the distinction between trained words and other words was visible both very early (140 ms) and later (500 ms), comprehension skill was associated only with the later component. As we suggested above, the earlier component may reflect a general episodic effect that depends not on orthographically based word identification but on some visual attention factor. Comprehension skill may be less relevant to this level of processing, compared with a word-form-based episodic effect that occurs later. It is an open question whether this later episodic trace includes only the orthographic form of the trained word as presented or also the meaning that was associated with it

during training. It is possible that both the word form and its associated meaning are part of the episode that is reflected in this later component. Participants knew they would make a meaning judgment. Recalling an associated meaning of the first word would help with that task. However, in research on recognition memory, results suggest that an intention to retrieve information is not necessary for the P600 (Curran, 1999; Paller & Kutas, 1992). Thus, although participants may have been either automatically or intentionally trying to retrieve meaning information associated with the trained word during this 400-600 ms time window, there is no basis to conclude that they were. It is important to note that our results represent a case in which ERP data help constrain the interpretation of a behavioral result. The skill difference in the late positive shift suggests that skill differences observed in accuracy reflect the strength of the familiarity that resulted from training. However, familiarity is not the end of the story. Skilled comprehenders also showed a larger N400 effect during the presentation of the second word, reflecting a stronger congruence when the second word matched the meaning of the first word. This suggests that skilled comprehenders achieved a better learning of the meaning of the trained word, allowing a related word to show a congruence response. A few other studies have focused on the N400 as capable of distinguishing reading skill. For example, Coch and Holcomb (2003) reported that first-grade children of high reading skill but not low reading

empora l emp o

13
B

cS

^ .& '6 'd d.

fi

J3

'5 3

-3-

4l

r S" i

WORD LEARNING AND EVENT-RELATED POTENTIALS P4 Cluster-From the Onset of Word 2 Familiar Related 800 Familiar Unrelated

1291

Ttme(ms)

P4 Cluster-From the Onset of Word 2

Trained Rare Related Trained Rare Unrelated

Time(ms)

P4 CInster-From the Onset of Word 2

- Untrained Rare Related - Untrained Rare Unrelated

Figure 8. P4 cluster showing word type and relatedness. From 372 ms following the onset of Word 2, there is a separation of related from unrelated words for familiar and trained rare words only. skill showed N400 responses in passive viewing of words and pseudowords. More related to the present result, Rubin and Johnson (2002) reported that learning-disabled adults showed a longer N400 latency to words in context than did non-disabled readers.

However, the specific point for our N400 results is the role of the N400 as an indicator of learning success and the conclusion that skilled comprehenders are more successful at learning new word meanings. We add some methodological observations to our discussion. First, we note the value of PCA in a study of ERPs, even in the absence of source analysis. Because we had specific hypotheses about a late-latency positive component during the first word (P600) and a midlatency negative component, we could have tested for these components in t tests on specific electrodes. However, the PCA allowed a more data-driven approach that could expose other task-related shifts, and indeed it appears to have done so. The PCA revealed a late positive shift that, beyond any non-cognitive components it might have contained, differentiated among word types, suggesting a meaning verification process for words whose meanings were either previously known or recently learned. PCA also exposed a component that overlaps the end of the first word's presentation and the start of the second word's presentation. Notably, the effects that are traditionally reported as components (N400, P600) on the basis of the waveforms of selected electrodes were visible in the PCA as well, allowing follow-up tests of waveforms. Second, we note the potential value of treating two successive stimuli as a single recording epoch. When the interest is in processes that are distributed over two words, the single epoch can expose both word-specific components and shared components, while accurately reflecting the temporal dynamics that go with asking people about relations between successive words. Finally, we return to the general question of learning word meanings and the role of ERPs in studying this question. We do not assume that native language word learning usually involves the kind of associative training used in our study. (However, there is a similarity to classroom procedures for second-language learning.) We conclude that there are individual differences in the ability to learn the meanings of new words, as evidenced by the behavioral results. Adults who were higher in comprehension skill showed better learning of 60 words from 45 min of training than did those lower in comprehension skill, as evidenced in their performance on a single-word meaning probe following training. The ERP evidence adds to these behavioral results by showing that skilled comprehenders were more sensitive to whether a word had been in the training set. Because words to be trained were chosen so as to be unknown to individual participants, it is likely that this difference in sensitivity reflected a difference in learning that allowed skilled comprehenders to establish stronger episodic traces for trained words. The source of this stronger episodic trace has at least two possibilities. One is that skilled comprehenders were better at learning the meanings they were taught. Thus, presented with gloaming, they retrieved the episodic trace that established the association "gloaming means twilight." A related possibility is that skilled comprehenders were better able to encode (and thus recognize) the orthographic word form, gloaming. Retrieving whatever was learned about the meaning of a word depends on recognizing the form of the word. Skilled comprehenders generally know more about word forms (orthography and pronunciations) than do less skilled comprehenders (Bell & Perfetti, 1994; Perfetti & Hart, 2001). Because our study was about word meanings, one might suppose that the observed skill differences reflected the ability to learn

1292

PERFETTI, WLOTKO, AND HART Kucera, H., & Francis, W. N. (1967). Computational analysis of present-day American English. Providence, RI: Brown University Press. Kutas, M., & Hillyard, S. A. (1980). Event-related brain potentials to semantically inappropriate and surprisingly large words. Biological Psychology, 11(2), 99-116. Liu, Y., & Perfetti, C. A. (2003). The time course of brain activity in reading English and Chinese: An ERP study of Chinese bilinguals. Journal of Human Brain Mapping, 18(3), 167-175. Liu, Y., Perfetti, C. A., & Hart, L. (2003). ERP evidence for the time course of graphic, phonological, and semantic information in Chinese meaning and pronunciation decisions. Journal of Experimental Psychology: Learning, Memory, and Cognition, 29, 1231-1247. Luck, S. J., & Hillyard, S. A. (1994). Electrophysiological correlates of feature analysis during visual search. Psychophysiology, 31, 291-308. Nagy, W. E., & Herman, P. A. (1987). Breadth and depth of vocabulary knowledge: Implications for acquisition and instruction. In M. Mc-Keown & M. E. Curtis (Eds.), The nature of vocabulary acquisition (pp. 19-36). Hillsdale NJ: Erlbaum. Nation, K., & Snowling, M. J. (1999). Developmental differences in sensitivity to semantic relations among good and poor comprehenders: Evidence from semantic priming. Cognition, 70, B1-B13. Nobre, A. C, & McCarthy, G. (1994). Language-related ERPs: Scalp distributions and modulation by word-type and semantic priming. Journal of Cognitive Neuroscience, 6, 233-255. Paller, K. A., & Kutas, M. (1992). Brain potentials during memory retrieval provide neurophysiological support for the distinction between conscious recollection and priming. Journal of Cognitive Neuroscience, 4, 375-391. Perfetti, C. A., & Hart, L. (2001). The lexical bases of comprehension skill. In D. Gorfien (Ed.), On the consequences of meaning selection (pp. 67-86). Washington, DC: American Psychological Association. Reichle, E. D., & Perfetti, C. A. (2003). Morphology in word identification: A word-experience model that accounts for morpheme frequency effects. Scientific Studies of Reading, 7(1), 219-238. Rubin, S. S., & Johnson, C. M. (2002). Lexical access in college students with learning disabilities: An electrophysiological and performance-based investigation. Journal of Learning Disabilities, 35(4), 258-267. Rugg, M. C. (1995). Event-related potential studies of human memory. In M. S. Gazzaniga (Ed.), The cognitive neurosciences (pp. 789-801). Cambridge, MA: MIT Press. Rugg, M. D., & Doyle, M. C. (1992). Event-related potentials and recognition memory for low-frequency and high-frequency words. Journal of Cognitive Neuroscience, 4, 69-79. Rugg, M. D., & Nagy, M. E. (1989). Event-related potentials and recognition for words. Electroencephalography and Clinical Neurophysiology, 72, 395-406. Riisseler, J., Probst, S., Johannes, S., & Miinte, T. F. (2003). Recognition memory for high- and low-frequency words in adult normal and dyslexic readers: An event-related brain potential study. Journal of Clinical and Experimental Neuropsychology, 25, 815-829. Sereno, S., Rayner, K., & Posner, M. (1998). Establishing a time-line of word recognition: Evidence from eye movements and event-related potentials. NeuroReport, 9, 2195-2200. Van Boxtel, G. J. M. (1998). Computational and statistical methods for analyzing event-related potential data. Behavior Research Methods, Instruments, & Computers, 30(1), 87-102. Received October 14, 2004 Revision received March 7, 2005 Accepted May 31, 2005

or remember word meanings. Certainly, knowledge about word meanings, as distinct from word forms, could be an independent contributor to reading skill (Nation & Snowling, 1999). However, it is important to keep in mind that the learning of meanings is seldom just about meanings but also about the connections between forms and meanings. Word form knowledge and its connection to meaning is the core of the lexical processing, and weakness in this knowledge will negatively affect word-level comprehension. Our results add the finding that among adults, more highly skilled comprehenders are better at learning new meanings and more sensitive to the episodic status of a word. This result carries an interesting implication: An episodic word processa memory for a word experiencecan produce differences in semantic knowledge. Such a conclusion is compatible with an episodic theory of word identification of the sort proposed by Reichle and Perfetti (2003), in which the development of a mental lexicon is the result of functional encodings of word episodes that lay down form and meaning relations. The key idea is that effective experiences with wordsthe multiple encounters with a word that lead to an abstracted representation of form and meaningis what creates reading skill. The present results can be taken to suggest the plausibility of this proposal in accounting for differences in reading comprehension. References Barnea, A., & Breznitz, Z. (1998). Phonological and orthographic processing of Hebrew words: Electrophysiological aspects. Journal of Genetic Psychology, 159(4), 492-504. Bell, L. C, & Perfetti, C. A. (1994). Reading skill: Some adult comparisons. Journal of Educational Psychology, 86, 244-255. Bentin, S., Mouchetant-Rostaing, Y., Giard, M. H., Echallier, J. F., & Pernier, J. (1999). ERP manifestations of processing printed words at different psycholinguistic levels: Time course and scalp distribution. Journal of Cognitive Neuroscience, 11, 235-260. Chapman, R. M., & McCrary, J. W. (1995). EP component identification and measurement by principal components analysis. Brain and Cognition, 27, 288-310. , Chapman, R. M., McCrary, J. W., & Chapman, J. A. (1978). Short-term memory: The storage component of human brain responses predicts recall. Science, 202(4373), 1211-1214. Coch, D., & Holcomb, P. J. (2003). The N400 in beginning readers. Developmental Psychology, 43, 146-166. Curran, T. (1999). The electrophysiology of incidental and intentional retrieval: ERP old/new effects in lexical decision and recognition memory. Neuropsychologica, 37, 771-785. Dien, J. (1998). Addressing misallocation of variance in principal components analysis of evoked potentials. Brain Topography, 11(1), 43-55. Dien, J., & Frishkoff, G. A. (2005). Principal components analysis of event-related potential datasets. In T. Handy (Ed.), Event-related potentials: A methods handbook (pp. 189-207). Cambridge, MA: MIT Press. Dien, J., & Santuzzi, A. M. (2005). Application of repeated measures ANOVA to high-density ERP datasets: A review and tutorial. In T. Handy (Ed.), Event-related potentials: A methods handbook (pp. 57-82). Cambridge, MA: MIT Press. Hackley, S. A., Woldorff, M., & Hillyard, S. A. (1990). Cross-modal selective attention effects on retinal, myogenic, brainstem and cerebral evoked potentials. Psychophysiology, 27, 195-208.

Tokowicz, N.. A Perfetti, C. A. (2005). Introduction to section II: Comprehension. In J. F. Kroll & A. M. B. de Sroot (Eds.), Handbook of bilingualism: Psycholinguistic approaches (pp. 173-177). New York: Oxford University Press.

Introduction to Part
Comprehension

he goal for many aspiring bilinguals is successful communication in their second language (L2), including becoming able to comprehend spoken and written messages. Language comprehension is complex from a scientific point of view. For a competent speaker, however, comprehension is "a piece of cake." All the lexical, semantic, syntactic, and textual processes that compete for attention and memory resources (and fill up diagrams in models of comprehension) are executed with ease and without notice. Such may not be the case for the learner of an L2 or perhaps even for the moderately skilled bilingual. On the other hand, for the skilled bilingual, as for the monolingual, the machinery of comprehension may be so skillfully engaged that only the most clever of experimental designs can expose any confusion or difficulty. For the bilingual, which factors influence the success of L2 comprehension? The chapters in this part provide an overview of the research on some of the levels of bilingual comprehension. To introduce the issues seen in the study of bilingual comprehension, we first outline a general framework for comprehension processes. This framework reflects a body of research largely undertaken without the slightest notice that some comprehenders might be able to engage more than one language. That is,-it reflects the consensus view of comprehension from the perspective of research in monolingual contexts. We then describe some of the relevant research conducted on bilinguals for each aspect of comprehension. Our review is far from comprehensive, focusing on only a few major issues in each area and pointing to the chapters of this section for more detailed reviews.

a platform to view problems of bilingual comprehension. Reviews of spoken language comprehension and written language comprehension by Cutler and Clifton (1999) and Perfetti (1999), respectively, provide frameworks for the key component processes of comprehension and their interrelationships. Here, we simply provide an outlined description of some of the key components. Ignoring the physical properties of speech and print, no small matter, simplifies the problem of comprehension so that what we need to account for is merely the following: 1. Word identification: How words are identified such that their context-appropriate meanings are selected. 2. Parsing: How words and morphemes are configured into phrasal units that govern interpretation. 3. Semantic-syntactic representations: How the meanings of words and the grammar of the language combine to provide the meaning of clauses and sentences. 4. Text representation: How the meanings of clauses and sentences are integrated into a coherent representation of an extended discourse. 5. Understanding: How all the above function to yield actual comprehension, a more-or-less veridical representation of a token discourse. Finally, in all of these processes, there can be individual differences that produce variability in comprehension skill. These processes can serve as a starting point for the study of bilingual comprehension. Of course, an L2 brings added complexity to an already rather complex problem. This may be part of the reason for an unevenness seen in the extent to which

The Processes of Comprehension The component processes of language comprehension and the ways they are interconnected provide

173

174

Comprehension

component processes and their relationships have been addressed in bilingual research. In particular, there is much more to say about bilingual word-level processes than higher-level comprehension processes.

Word Identification
Word identification entails lexical access through phonological and printed inputs. It is axiomatic that these inputs are linguistically specific. One hears a word with Dutch phonology or with French phonology, and a comprehender with the required language skill identifies the word accordingly. From this point, however, the details become interesting. Two of the chapters in this part relate to the study of bilingual word identification. Although it seems intuitively reasonable to skilled bilinguals that they can effectively "turn off" or attenuate one of their languages, the research by now suggests that this seldom happens. Perhaps one language can be "turned down," but not quite turned off. As Dijkstra (chapter 9) demonstrates, bottom-up factors such as stimulus list composition and task demands make a difference for bilingual word recognition. Furthermore, top-down information, such as the knowledge that only one of your languages is needed for a given task is not sufficient and can be overridden by the bottom-up information (see also MacWhinney, chapter 3). A classic question is whether word form information for the two languages is stored together or separately. Given the above results and others, we may conclude that word form information is most likely stored in a shared way (or at least in a way that allows sufficient cross talk between the two languages; see Francis, chapter 12). As mentioned, task demands will influence whether there appears to be selective or nonselective access of word forms in the two languages. The critical issue of how words are recognized by bilinguals recently has received much attention because of the precision available in mathematical models. Thomas and Van Heuven (chapter 10) provide a review of the two major types of computational models used in this area, localist and distributed models. Their review includes a summary of the issues that have been tackled with models; these issues include neighborhood effects, priming, and homograph/cognate effects. Although we are far from a complete model of bilingual comprehension, progress in computational modeling comes from models designed for specific problems rather than for general

purposes. Bilingual word recognition has made great advances in the recent past as a result of the available models. Thomas and Van Heuven suggest that joining localist and distributed models will further our understanding of bilingual comprehension. Beyond the representational details of models, however, is the value of building competing models that address the same problems. This competition exposes basic assumptions about language processes that can be hidden when each model addresses a different problem.

Parsing
Listeners and readers must do something with the words they hear and see to construct messages. Building phrasal units from strings of words and connecting these units with each other in the way allowed by the grammar of the language is a large part of this process. How to explain parsing in the first language (LI) has proved to be difficult and contentious. How do comprehenders decide, on a word-by-word basis, how to attach a word to the current representation of a sentence? Theories that stress basic principles of simplicity and theories that stress more complex multiple constraints offer rather different solutions to this question. In the case of an L2, the question becomes even more difficult. The grammar of the L2 is not as well represented as that of the LI in most cases. So, how does a learner of a second language go about deciding how to attach a word to a current sentence representation? Frenck-Mestre (chapter 13) reviews some of the recent research on bilingual parsing. In particular, she considers the evidence that bilinguals use information from their LI to process their L2. Thus, a person's LI can indicate which particular syntactic structures will be difficult to comprehend in L2. A similar conclusion was reached by Fender (2003), who showed that Japanese and Arabic speakers of English as a second language have opposite difficulties in processing English as a result of different native language structures. The dominance of LI syntactic structures in L2 comprehension was also evident in research by Tokowicz and MacWhinney (2002), who showed that native English speakers learning Spanish had difficulty rejecting Spanish sentences with grammatical errors when the word-by-word translation mapped directly to an acceptable English structure. Also, Tokowicz and MacWhinney (in press) found that these learners showed brain responses (measured by event-related potentials) that indicated more

Introduction to Part II 175 sensitivity to grammatical violations in their L2 (Spanish) when the constructions were formed similarly, rather than differently, in LI and L2. This was true despite the participants' inability to distinguish grammatically acceptable and unacceptable sentences overtly. Finally, evidence shows that non-proficient bilinguals initially comprehend L2 through an LI lens. McDonald (1987) showed that English learners of Dutch declined in their use of word order (a valid English cue) and increased in their use of case inflection (a valid Dutch cue) to comprehend L2 sentences as their Dutch competence increased. and Tokowicz, Kroll, De Groot, & Van Hell, 2002, for more information about the consequences of imprecise meaning overlap across languages). In answer to the question of whether cognates are stored in a special way relative to noncognates, Sanchez-Casas and Garcia-Albea (chapter 11) conclude that there is preliminary evidence to support a special status for cognate representations. They argue that cognates are treated as morphologically related words within a language and demonstrate that they follow the same priming pattern as such words. Interestingly, Francis (chapter 12) provides evidence that translation equivalents in general are not treated as within-language synonyms. Another factor that has been shown to influence meaning representation is age of acquisition (AoA). Izura and Ellis (2002, 2004) showed that regardless of LI AoA, L2 words learned earlier are processed more rapidly than L-2- words learned later. This pattern has been observed in several tasks, including translation recognition, lexical decision, and object naming. Thus, the age at which an L2 word is learned has an impact on the word form-to-meaning connection that is the foundation of L2 comprehension.

Semantic-Syntactic Representations
Representing meaning is central to comprehension at all levels. Word identification brings access to word meanings and their associated concepts, and parsing builds groupings of words and morphemes into phrasal units that provide both reference and semantic relationships. The result of these word identification and syntactic processes is a representation of meaning at the clausal and sentence levels. This meaning representation, corresponding to a proposition in theories of comprehension (Kinstsch, 1988), can be considered the basic unit of relational meaning in a text, spoken or written. It is our impression that there is little in bilingual research that corresponds fully to this level of analysis, although several chapters in this section focus on parts of it. For example, how words are represented in the memory of a bilingual has been a major question. Are words from the two languages stored separately in their own language or connected together by their meaning similarity? Do translation equivalents activate identical meaning representations? Are cognate translations stored differently from noncognate translations? Each of these issues is addressed in this section.. The basic answer to the first of these questions is, well, it depends. A single pool of semantic features most likely comprises the meanings of translation equivalents. Whether translation equivalents activate exactly the same meaning may depend on the manner in which L2 was learned (e.g., in the classroom or abroad; see De Groot, 1992). However, as always, there are caveats. Generally, it seems that the differences in meaning are few and far between. For the most part, translations are just that, words that have the same meaning across languages (see Guasch, 2001; Sanchez-Casas, Suarez-Buratti, & Igoa, 1992; Tokowicz, 2000;

Text Representation and Integration (and Understanding)

Text representation and integration is an area that has received relatively little attention in the psycho-linguistic literature on bilingualism and is not represented in the chapters in this part. This is true also for the level of real understanding (fifth in our list of comprehension processes), so we comment on these two together. We suspect that the neglect results from the natural focus on word- and, to a lesser extent, syntactic-level processes that are the building blocks of comprehension. In the long run, we would expect to see increased attention at least to the consequences for text representation of the lexical and syntactic processes that have been studied. Presumably, a parsing problem in reading a sentence in L2 must lead to one of two consequencesa breakdown in comprehension such that both the current sentence and subsequent sentences are misunderstood or a reflective repair that slows the comprehension process, but keeps the representation coherent. Both of these outcomes place comprehension at risk. Similarly, at the word level, does it matter "downstream" in the representation of sentence and clause meaning that a word read in L2 has also activated an LI word representation for a few

176

Comprehension

milliseconds? Moreover, does sustained reading or listening to an L2 text build up some protection from this word-level interference? Beyond these basic questions about how text-level processes might interact with lexical and parsing processes is the application of text comprehension research tools to bilingual processing. For example, computational models of comprehension (e.g., Kintsch, 1988; Van den Broek, Young, Tzeng, & Linderholm, 1999) can be sensitive to limitations in working memory, readers' knowledge and goals, and other factors that would apply to L2 comprehension as well as LI.

Individual Differences
Comprehension processes in LI show wide-ranging individual differences in adults and children; these differences arise from such components as we reviewed above, plus others (Perfetti, 1999). Similarly, there are many individual difference that are likely to affect how one learns and processes an L2, and some of these appear to lie in LI abilities. Michael and Gollan (chapter 19), in part III on language production and control, provide an overview of research on the effects of LI processing skill (e.g., working memory capacity and suppression) on L2 processing. Furthermore, motivational factors can also have an impact on an individual's success in L2 learning and, ultimately, comprehension. With recent applications of neuroimaging and electrophysiological techniques to the study of language processing, such as functional magnetic resonance imaging, positron emission tomography, and event-related potentials, we have even more methods to study bilingual comprehension. Having these added techniques, along with the advances in mathematical modeling, will undoubtedly enhance the already-rich picture of what happens during bilingual language processing. These advances will allow researchers to pose questions other than those already asked. The converging evidence from this set of increasingly diverse methods is likely to encourage the development of models of bilingual comprehension that are more complete and, at the same time, better capture the implications for general models of language comprehension that in the past have focused on monolingual experience alone. References Cutler, A., & Clifton, C. E. (1999). Comprehending spoken language: A blueprint of the

listener. In C. M. Brown & P. Hagoort (Eds.), The neurocognition of language (pp. 123-166). Oxford, England: Oxford University Press. De Groot, A. M. B. (1992). Determinants of word translation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 1001-1018. Fender, M. (2003). English word recognition and word integration skills of native Arabic- and Japanese-speaking learners of English as a second language. Applied Psycholinguistics, 24, 289-315. Guasch, M. (2001). Forma y significado en el procesamiento lexico de bilingues del castellano y del Catalan. Unpublished master's thesis, Universitat Rovira i Virgili, Tarragona, Spain. Izura, C. &C Ellis, A. W. (2002). Age of acquisition effects in word recognition and production in first and second languages. Psicologica, 23, 245-281. Izura, C. & Ellis, A. W. (2004). Age of acquisition effects in translation judgment tasks. Journal of Memory and Language, 50, 165-181. Kintsch, W. (1988). The role of knowledge in discourse processing: A construction-integration model. Psychological Review, 95, 163-182. McDonald, J. L. (1987). Sentence interpretation in bilingual speakers of English and Dutch. Applied Psycholinguistics, 8, 379-A14. Perfetti, C. A. (1999). Comprehending written language: A blueprint of the reader. In C. Brown 8t P. Hagoort (Eds.), The neurocognition of language (pp. 167-208). Oxford, England: Oxford University Press. Sanchez-Casas, R., Suarez-Buratti, B., & Igoa, J. M. (1992, September). Are bilingual lexical representations interconnected. Paper presented at the Fifth Conference of the European Society for Cognitive Psychology, Paris. Tokowicz, N. (2000). Meaning representation within and across languages. Unpublished doctoral dissertation, The Pennsylvania State University, University Park. Tokowicz, N., Kroll, J. F., De Groot, A. M. B., &C Van Hell, J. G. (2002). Number-of-translation norms for Dutch-English translation pairs: A new tool for examining language production. Behavior Research Methods, Instruments, and Computers, 34, 435-451. Tokowicz, N., &c MacWhinney, B. (2002, April). Judging grammatical acceptability in L2: Competing grammatical systems in the second language learner. Paper presented at the Forty-Seventh Annual Meeting of the International Linguistic Association, Toronto, Canada. Tokowicz, N., &C MacWhinney, B. (in press). Implicit and explicit measures of sensitivity to

Introduction to Part II 177

violations in second language grammar: An event-related potential investigation [Special issue]. Studies in Second Language Learning. Van den Broek, P., Young, M., Tzeng, Y., & Linderholm, T. (1999). The landscape model

of reading: Inferences and the on-line construction of a memory representation. In H. van Oostendorp &c S. R. Goldman (Eds.), The construction of mental representations during reading (pp. 71-98). Mahwah, NJ: Erlbaum.

Perfetti, C. A., & Hart, L (2001). The lexical bases of comprehension skill. In D. S. Gorfien (Ed.), C/7 fAe consequences of meaning selection: Perspectives on resolving lexical ambiguity {pp. 67-86). Washington, DC: American Psychological Association.

5 _______________

_______

The Lexical Basis of Comprehension Skill

Charles A. Perfetti and Lesley Hart
How shall we understand the source of individual variation in comprehension skill? What is it that skilled comprehenders do during reading that less-skilled readers do not do? Or vice versa: What do less-skilled readers do that skilled readers do not? In this chapter, we revisit this question from a long-range perspective coupled with some new research. Fifteen years ago, Perfetti (1985) suggested an answer to this question roughly as follows: Although skilled comprehenders differ from less-skilled comprehenders in a number of ways, at least some of these differences arise from a foundational processing factor: the effectiveness of basic word identification skill. The bases of this claim lay in a program of research that consistently found that when either children or adults were separated by their scores on a reading comprehension test, they sorted themselves also on their speed of written word and pseudoword identification. Of course, even with a long list of studies with the same result, the causal connection between the lexical process variability and the comprehension variability could not be established. This problem, we believe, was not just the usual conundrum that blocks causal inferences from correlational data but was intrinsically part of how skill is acquired in reading: Lexical skills allow comprehension, comprehension allows reading practice, reading practice strengthens lexical skills, and so on. The recurrent nature of component interactions in the process assured an intimate connection between lexical and comprehension skills. Nevertheless, the natural privilege in the genesis of the recurrent accumulation of skill belonged to the lexical processes. Although it was possible to imagine being able to read words without comprehending the texts that contain them, it was more difficult to imagine being able to comprehend the texts without reading the words. So we summarized the relationship in the verbal efficiency theory. The core idea was that manynot allproblems in comprehension arose from ineffective lower level processes needed for the identification of words. Rapid and perhaps modular word identification was important for a comprehension system of limited capacity. The question we address here is, roughly, does some part of this idea remain tenable? And what more specific instantiation of the general idea is now possible? In what follows, we first review the background of verbal efficiency theory and then illustrate the concept of lexical quality and the lexical quality
67

68

PERFETTI AND HART

hypothesis, the central ideas that link lower and higher level reading processes. Then we summarize experiments on simple meaning judgments, whose outcomes, including some nonobvious ones, are predicted by this hypothesis.

Verbal Efficiency and Lexical Knowledge In its earlier formulation, verbal efficiency was elaborated in two compatible ways. In the first, its source was lexical processing efficiency, and in the second, it was functional working memory (Perfetti, 1985, pp. 112-115). Both of these hypotheses have received ample attention in research, although distinguishing them clearly has not seemed to be a high priority. Research on adult comprehension focusing on memory differences (e.g., Carpenter, Miyake, & Just, 1994) has tended to ignore lexical factors. The solution to the question of whether there are multiple low-level causes or a single one remains a difficult question. If it is to be a single mechanism, it needs to account for two facts: (a) the pervasive association of lexical skills with comprehension skills in adults as well as children and (b) the association (only slightly less pervasive) of listening and reading comprehension skill. The obvious problem here is that it is not straightforward to convert a fundamental problem in written word identification into a cause of listening comprehension problems. So the unparsimonious two-factor theory may be required: (a) limitations in word identification efficiency and (b) limitations in functional working memory (perhaps with a phonological component). In discussing the single- versus two-mechanism alternatives, Perfetti (1985) offered a one-factor conjecture based on linguistic code manipulation. The basic idea was that, in any modality, efficiency is the rapid retrieval, from inactive memory, of codes that are part of a stored linguistic symbol. And "to the extent retrieval is effortful and the retrieved codes low in quality, the processing is inefficient" (p. 118). Whether by spoken language or by written language, a low-quality code retrieved with effort would jeopardize comprehension processes that depend on a high-quality representation. However, the additional linguistic step in readingmapping an orthographic form onto a lexical representationcould make the problem more noticeable in reading than in listening. Lexical Quality The question of code quality, in this account, becomes central. The original suggestion was that the retrieval of a lexical representation is high in quality "to the extent that it contains both semantic and phonetic information sufficient to recover its memory location.... [This quality must be retained] long enough for subsequent processes to perform their work. Thus a 'name' without meaning and a meaning without a 'name' are both low quality" (Perfetti, 1985, p. 118). This idea was developed further in a theory of reading acquisition by reference to lexical specificity and redundancy (Perfetti, 1992). A lexical representation has high quality to the extent that it has a fully specified orthographic representation (a spelling) and redundant phonological representations (one from spoken language and one recoverable from orthographic-to-phonological mappings). If a lexical representation is specific and redundant, its retrieval is more

LEXICAL BASIS OF COMPREHENSION SKILL

69

likely to be coherent and reliable. By coherent, we mean that the constituents are available synchronously at retrieval, giving the impression of a unitary word perception event. (The contrast is constituent asynchrony, a fragmented appearance of constituents and parts of constituents, as when an effortful speech recoding occurs followed by meaning retrieval; see Perfetti, 1985, p. 114.) The consequence of reliability is that multiple encounters with a given word tend to produce a common core representation consisting of a nexus of orthographic phonological and semantic information. To put it approximately, these defining features of high quality allow the reader to get exactly the word that is printed rather than parts of it that may also be parts of other words. Confusion about word meaning and word form is minimized by high-quality representations. One should be able to detect variation in quality in a number of ways: For example, inconsistency in attempts to spell and hesitant or effortful retrieval of pronunciations or meanings reflect low-quality representations. A reliable, coherent, high-quality representation is retrieved easily and consistently. Figure 5.1 represents a high-quality representation by indicating a tightly bonded set of word constituents: the orthographic (OR), phonological (PH), and semantic (SE) specifications of the word. The identification of the word is the retrieval of these constituents. It is useful to consider the implications of constituent as opposed, to say component or levela common reference in hierarchical models. A constituent in a linguistic or algebraic representation is not merely a part of a larger whole; it is a necessary or defining symbol, a variable. Thus, a sentence is a noun phrase plus verb phrase (NP + VP), major constituents that can assume specific values. Analogously, a word is an unordered triple of PH, OR, and SE. The specification of each is subject to constraints from a relevant system. Those involving phonology are of special interest in word identification but are beyond our purpose here. (See Berent & Shimron, 1997, for an example of constraints on phonology that may be relevant for reading.) Similarly, SE constituents reflect basic conceptual and grammatical constraints. One consequence of the constituent idea is that it encourages the view that the word identity is both unitary and compositional. Thus, the lexical representation of the word cat is the (unitary) linguist object such that it has spelling C-A-T, phonology ce, and meaning (whatever it is that cat means). In general terms, all words are triples of {PH, OR, SE} specification. The SE constituent, however, is incomplete in the cat example, because it ignores grammatical information. To avoid a fourth constituent, we can stipulate that SE is actually a collection of meaning and grammatical information. Thus, it includes the lemma, the abstract grammatical frame that appears necessary to understand the production of speech (Bock & Levelt, 1994), and a conceptual structure that links to the lemma. We can now state the key idea about individual differences: People vary in the quality of their lexical representations. Any representation that does not specify the value of one of its constituents is low quality. Consider the following examples: 1. Presented with the word incarcerate, the reader can pronounce it accurately and knows it has some negative meaning but is not sure what that meaning is.

70

PERFETTI AND HART

Figure 5.1. A high-quality representation for the word gate. Lexical representations (for a literate mind) are triples of orthographic (OR), phonological (PH), and semantic-syntactic (SE) constituents. Quality is the extent to which each constituent is fully specified (constants instead of variables for OR and PH form constituents) and conceptually and syntactically differentiated (for SE constituents). Here quality is indicated more crudely as concatenation links between constituents. These links are not to be understood as activation links, but as bonding links. Thus, gate is the word such that its OR, PH, and SE constituents are specified as given in the diagram and mutually bonded into a unitary lexical object. 2. Presented with incarcerate, the reader stumbles on its pronunciation, producing something like in-cark-rate. 3. Presented with incarcerate, the reader can pronounce it and indicate that it means something like "to confine in prison." When attempting to speak the word to produce a message about someone going to jail, however, the reader sometimes produces incarcerate and sometimes something more like incarsate. 4. A reader can perform all the tasks failed variably in 1, 2, and 3 above but can spell the word correctly only on some attempts. The examples indicate several unreliable representations of the word {incarcerate [inkarsreyt], (verb trans; put in jail)}. They will have differential consequences in different tasks. Case 4, for example, is one familiar to many individuals of high literacythe feeling of semantic and phonological competence coupled with a spelling block. Case 1 may represent a skilled reader of limited experience who can "decode" a word that he or she does not really know, although the reader has heard it pronounced and thus has some phonological representation.

LEXICAL BASIS OF COMPREHENSION SKILL

71

One thing should be made clear, given these examples. Illustrating the basic idea with a low-frequency word generalizes it. Even skilled readers have low-quality representations for some words. What skilled readers have are foundational resources to help retrieve impoverished representations and add information (about spelling, pronunciation, or meaning) to those representations. The gain in this analysis is the focus on words: It is not individuals, but word representations, that vary in quality. A skilled reader, in addition to having foundational resources (decoding, spelling, and grammatical skills), is one who has many high-quality word representations. A less-skilled reader is one who has fewer high-quality word representations. Examining the genesis of the number of quality word representations is another matter. Certainly, basic and explicit phonological knowledge supports early progress in reading, and subsequent practice in reading, listening, speaking, and writing promotes more high-quality word representations. Whatever the source of variation in any set of individuals, the key idea is that there is a continuum of lexical knowledge, and this continuum is bisected when researchers refer to skilled and less-skilled readers. As a general methodological point, progress in understanding lexical knowledge and especially in relating it to comprehension skill requires more individual (readerand word-specific) assessment than is typical in research. For example, to detect unreliability of word representations, one may need more than one sample of easy-to-give recognition tests for spelling or multiple-choice vocabulary tests. Fortunately, the kind of assessment we have in mind can be crudely approximated by such tests as a starting point. Homophones and Lexical Quality Words with multiple meanings on single forms are interesting because they risk lexical quality. Quality is at risk because there is a lack of one-to-one mapping among one or more form and meaning constituents. There are three possibilities: 1. Ambiguity: single OR form, single PH form, multiple meanings. This is the most general case and has been the staple of research addressing meaning selection process. Example: count (1. enumeration; 2. feudal title). 2. Homophony: two OR forms, one PH form. This has been the staple of research on phonological processes and has been relatively neglected in research on meaning selection. Example: seed, cede. 3. Homography: one OR form, two PH forms. There are few cases in English, and indeed these are rare across writing systems. This rarity is consistent with the universal privilege gained by phonology as a reliable mapping for orthography (Perfetti & Tan, 1998). Example: bass (1. voice or instrument of the lower pitch; 2. type of fish). One can capture the possibilities of form-meaning divergence more simply by defining a two-way classification of phonological and orthographic sameness.

72

PERFETTI AND HART

Ambiguity is then same phonology and same orthography. But that leads to infelicitous phrasing, such as homographic homophones, nonhomographic homophones, and nonhomophonic homographs. So, we prefer the three-way classification above. To illustrate how homophones place stress on lexical representations, Figure 5.2 expands the representations of Figure 5.1 by adding the word gait. Although each orthographic specification is associated with a unique meaning specification (at least for our purposes), a single phonological specification links to two spellings and two meanings.1 These are not low-quality representations because they are specified as fully as they can be. Thus, they can be considered the limiting case: high-quality word knowledge characteristic of a skilled reader. There is potential for confusion here. If one encounters the spoken form [geyt], both, gait and gate might be activated, along with their different meanings. Context will select the right one, and the confusion will be of no real consequence for comprehension. Less obviously, in reading, confusion also can result even though the spelling uniquely identifies the word. Thus gait, despite being beyond dispute as to its identity, can cause its homophonic partner gate to be activated, leading to momentary activation. So the case for a skilled reader is that the written presentation of homophones can lead to the activation (retrieval) of two words. Although the representations are as good as they can be, the momentary retrieval of these representations can yield confusion, even for skilled readers as Gernsbacher and Faust (1991a) have shown. If the potential for confusion exists for the skilled reader, it is even stronger for the less-skilled reader. According to the lexical quality assumption, a low-quality representation could be observed for gait, for gate, or for both. Imagine a reader whose skill is low enough that the word gait is completely unknown. No spelling, no meaning. This reader has no confusion when gate is presented. Or, more carefully, the only problem with, gate is due to its representation quality independent of gait. It should be no different from a control word, one that is not a homophone. However, a more typical case might be a reader with an unreliable low-quality representation for gait. Now there is the increased potential of confusion beyond what one might see from a nonhomophone. This is because the representation for gait, no matter how impoverished, can be partially activated by a spelling, a pronunciation, or a meaning when gate is presented. Frequency must be taken into account. Not only is word frequency an important determinant of word processing, but it is specifically implicated in disambiguating processes. Reading the more frequent member of a pair of homophones may occur without leading to access of the less-frequent member of the pair, a finding that has long been observed in research with ordinary semanti-cally ambiguous words (Duffy, Morris, & Rayner, 1988; Hogaboam & Perfetti, 1978). According to the lexical quality hypothesis, this frequency result is a consequence of the benefits of experience for the higher frequency meaning,
J One important feature of meaning is not represented in Figure 5.2 or in our discussion to this point. Meaning is relatively indeterminate at the lexical level. Semantic specification is no more than referential pointers or category indicators. Meanings, in the usual sense, are context-dependent in a way that orthographic and phonological forms are not.

LEXICAL BASIS OF COMPREHENSION SKILL

73

Figure 5.2. Representations for gate and gait as lexical processing units. The bonding links of Figure 5.1 are replaced by activation links (nondirectional), with stronger links represented by thicker lines. Thus, the high-frequency word gate has stronger internal activation links than the lower frequency word gait. The orthographic (OR) and phonological (PH) links are stronger in both than other links. Note the two words do not share a constituent. The idea is that each word in an autonomous lexicon strives to maintain a unique identity. The similarity relation between the two words arises from connections between their PH constituents rather than from some shared PH component. making it a more stable part of the word, compared with the lower frequency meaning. Moreover, if we consider word representation quality to be distributed across individuals the way we have suggested, then high-frequency words may be less a source of confusion for skilled readers than for less-skilled readers. This is because a word that is high frequency according to a corpus count may have rather different functional frequency characteristics for skilled and less-skilled readers, who differ substantially by college age in the amount of reading they have done. Thus, a word that is functionally high frequency for a skilled reader may be functionally a low-frequency word for a less-skilled reader. We illustrate this point in Figure 5.3.

Functional Identifiability and Context

The basic idea illustrated in Figure 5.3 is functional identifiability. A word is identifiable to the extent a reader has had sufficient experience with the word so as to allow its constituents to be fully specified and linked. Familiarity (functional frequency) can be considered a proxy for identifiability. For a given reader,

74

PERFETTI AND FIART

Activation Functions for Iirelewant V\fcrcJ

40

Oao

-M*e 94llec l!

Q10
<

0
0 50 150 480 1000. 2D00 3000 4000

TIIVE(9C^
Figure 5.3. Hypothetical lexical activation functions. For a given word, the function for a skilled reader rises rapidly and deactivates accordingly. For a less-skilled reader, the activation and deactivation occur more slowly. Equivalently, given a reader, the more rapid function is for a higher frequency word and the delayed function is for a lower frequency words. SOA = stimulus onset asynchrony.

a functionally high-frequency word has a more rapid activation functionreaching an identification threshold more quicklythan a low-frequency word. And for a given word, a skilled reader will show a more rapid activation function than a less-skilled reader. Thus, words and readers are interchangeable in the identifiability functions. There is an empirical basis for this interchangeability assumption. Perfetti and Roth (1981) summarized data from several conditions across experiments on the relationship between word naming in context and reading skill. Across conditions, there was variability in the constraint provided by context, word frequency word degrading (varied to examine bottom-up stimulus effects), and reading skill. Then, all these critical variables were ignored so as to examine the relationship between the time to identify a word in context as a function of the time to identify the same word in isolation. The result was a strong linear relationship between the time to identify the word in isolation and the time to identify it in context. In other words, one does not need to know the frequency of the word, the skill of the reader, or the visibility of the stimulus. To predict the time to identify the word in context, one needs to know only the time it takes to identify the word in isolation. This general function accounts for certain facts about individual differences. For

LEXICAL BASIS OF COMPREHENSION SKILL

75

example, the fact that identification times of less-skilled readers are aided more by context than are skilled readers is because their identification times are slower for the same words in isolation (Perfetti, Goldman, & Hogaboam, 1979; Stanovich & West, 1981; West & Stanovich, 1978). The theoretical implication of these observations is that the identifiability of words in isolation can be a basis of predictions about comprehension of text segments that contain these words. It also means that we cannot expect variables defined only at the reader level or only at the objective text level (e.g., word frequency) to be complete. The functional identifiability is critical, and this can vary for the same word across readers and for the same reader across words. We have, then, the basis for some nonobvious predictions about homophone confusions. The relationship between reading skill and homophone confusion effects should have different time courses and should show different frequency effects. First, because a skilled reader has higher quality representations for more words than does a less-skilled reader, the mutual activation of homophones should occur more quickly than for the less-skilled reader. Because both members of a homophone pair (e.g., gate and gait) are more identifiable for the skilled reader, activation will spread more quickly from the one to the other. This can lead to a homophone confusion more quickly for a skilled reader than for a less-skilled reader. However, this same assumptionthat the skilled reader has better quality representations of homophones as for all wordsleads to the assumption that this confusion will be short lived. The presented word will quickly have more activation than its unpresented homophone mate. For the less-skilled reader, confusion should build more slowly (because of a lower functional identifiability of both homophones) and release more slowly. This description referring to confusion and release from confusion is more general (and more theoretically neutral) than one referring to suppression (Gernsbacher & Faust, 1991a). We leave open the possible operation and failure of a suppression mechanism, although it is not required by our analysis of the process of homophone meaning selection. The effect of word frequency is also implicated by this analysis. A skilled reader, whose knowledge of the high-frequency member of the pair (e.g., gate) is of very high quality, should show little interference from a presentation of that high-frequency word. For the low-frequency member of the pair {e.g., gait), the skilled reader's representation is of lower quality; so the presentation of gait allows confusion activation to spread to the better known gate before comprehension of gait is complete. Compare this with a less-skilled reader, who, by assumption, has lower quality representations for both the higher and lower frequency member of a homophone pair. The presentation of a high-frequency member of the pair will now allow confusion. Because its functional identifiability is lower than it is for the high skilled reader, this allows activation of gate to spread to the incomplete representation of gait. For this to occur, we need to assume that gait is partially represented, not entirely absent. The Lexical Quality Hypothesis To summarize, we have reintroduced a theoretical framework that has organized observations about individual differences in reading comprehension. In

76

PERFETTI AND HART

its earlier form, the theory focused on the efficiency of lexical processes as a causal component in comprehension variability. In this form of the theory, word-level efficiency allows processing resources to be directed to comprehension, especially the encoding and integration of propositions. In its subsequent and current elaborations, efficiency of word identification remains an important part of the explanation. However, it is not to be understood as fundamentally about speed of processing or even resource allocation but rather about the quality of lexical representationsdetailed knowledge about word forms and meanings. High-qualuy representations are what drive rapid processing. More importantly, they are responsible for automaticity (or at least efficiency) of word identification, which is what allows processing resources to be devoted to higher level comprehension. The quality of lexical representations can be defined reasonably well in terms of the full specification of word constituent information: triples of phonology, orthography, and semantics (and syntax). Assessments of individual differences have seldom been thorough enough to test the^role of lexical quality against alternative hypotheses of comprehension differences. However, it is possible to approximate some of these assessments with traditional assessments of lexical processing, which is what we have done in the study summarized below. And it is possible to draw out the implications of the lexical quahty hypothesis for processing homophones, which put additional pressure on lexical representations to be of high quahty. Homophone processing should vary among readers of varying skill in systematic ways that reflect lexical quality.

Experimental Support for the Hypothesis A test of the implications of the lexical quality hypothesis comes from two studies of meaning decisions. The main study uses simple word-pair decisions, in which only word form is informative, providing the main support for the lexical quality hypothesis. A second study examines decisions in sentence contexts. In both studies, our samples are drawn from college students across a wide range of comprehension skill, assessed by a time-limited version of the NelsonDenny Reading Comprehension Test. We refer to higher scoring students as "skilled readers" and lower scoring students as "less-skilled readers." Meaning Decisions In the meaning decision experiment, the key questions center on readers'judgments that two words are related in meaning. When presented with the word king followed by the word royalty, the participant's decision should be "yes"; for evening followed by royalty, the decision should be "no." According to the lexical quality hypothesis, skilled readers should be faster than less-skilled readers in both the "yes" and the "no" judgments. This is because skilled readers have higher quality lexical representations, which make their critical constituents available more coherently and reliably, with less confusion. Consider now the word night followed by the word royalty. The decision is "no" as it is for evening-royalty. But night is a homophone, and we expect po-

LEXICAL BASIS OF COMPREHENSION SKILL

77

tential confusion to arise because its phonology is shared with knight. So it might take longer to make a decision. To complete the picture, of course, we want to know about knight-royalty, night-dark, kingdark, and evening-dark a full comparison of "yes" and "no" meaning judgments for homophones and nonhomophone controls. Let us consider for a moment what we are not doing in the meaning decision task. We are not assessing sentence comprehension, a task used by Gernsbacher and Faust (1991a) to examine skill differences in homophone processing. In their study, participants viewed sentences, such as He had lots of patients, and then decided whether a probe word was related to the sentence meaning. The probe word calm should be given a "no" response here, but confusion can arise from the homophone patience. Thus, readers were slower to reject calm following a sentence with a homophone than following a sentence with a nonhomophone control word {He had lots of students). Gernsbacher and Faust (1991a, Experiment 3) found that both skilled readers and less-skilled readers were slowed in decisions to the homophone sentence at the shorter of two intervals between sentence and probe (450-500 ms between sentence final-word onset and probe onset). At a longer interval (1,350 to 1,400 ms between sentence final-word onset and probe onset), only less-skilled readers showed longer decision times, relative to control sentences. We think our word meaning decision task is tapping some of the same processing as the sentence task. However, it requires attention to word form only, making for a more direct test of a hypothesis based on lexical knowledge. In fact, it demands that the decision be made only on the word form. It allows no additional confusion to arise from sentence comprehension processes beyond the word level.2 What results do we expect on the basis of the lexical quality hypothesis? First, less-skilled readers should show slower decisions across the board. This means for control words as well as homophones, one should see the basic cost of lower quality representations. Second, recall that we suggested two predictions concerning homophones from the lexical quality hypothesis: (a) Skilled readers should show homophone confusions sooner than less-skilled readers, and (b) skilled readers should show confusions mainly for the less-frequent member of a homophone pair, whereas less-skilled readers should also show confusions for the more frequent member of the pair. Each of these results received strong support in the experiment. First, consider the time course. The time between the onset of the first word (the target) and the onset of the second (the probe) was varied at three values: 150-, 450-, and 2,000-ms stimulus onset asynchrony (SOA). The first word always disappeared prior to the presentation of the probe. Thus, we can also describe the presentation sequence in terms of an initial (target) duration plus an interstimulus interval (ISI): 150-ms SOA, target duration = 100 ms, ISI = 50 ms; 450-ms SOA, target duration = 350 ms, ISI = 100 ms; and 2,000-ms
3 Because less-skilled readers, by definition, are less good at comprehension, there is some risk of their performance in a sentence task to reflect comprehension strategies in interaction with word knowledge. In particular, if less-skilled readers use context more to support reading, as the evidence suggests, they may try to use the sentence to support word identification, even when the sentence is intended to be unhelpful for this purpose.

78

PERFETTI AND HART

1000
900

-Homophone
-o-

Control

800 01 700

600 150 2000 450 2000 150 450

Less-Skilled

SOA (ms)

More-Skilled

Figure 5.4. Decision times for homophone and control words across three stimulus onset asynchronies (SOAs) for skilled and less-skilled readers. Skilled readers showed reliable homophone interference at 150 ms only; less-skilled readers showed reliable interference at 450 ms only. RT = reaction time.

SOA, target duration = 350 ms, ISI = 1,650 ms. (The second word, the probe, remained viewable until the participant responded.) The 450-ms SOA approximates the shorter of the two ISI intervals used by Gernsbacher and Faust (1991a) to allow a comparison. Figure 5.4 shows the main results for two groups of participants defined by their scores on time-limited version of the Nelson-Denny Reading Comprehension Test. In defining reader skill groups, we first administered the comprehension test to 300 undergraduates. For the experiment, the high-skill group were 44 participants in the top third of this distribution; the low-skill group were 38 participants with scores in the bottom third of the distribution. We carried out the experiment over a fuller range of comprehension skill that included 119 participants. But we excluded participants in the middle of this distribution to focus on the more highly contrastive groups of skilled and less-skilled readers, as is common in individual-differences research. However, our assumption is that we are dealing with a continuum of reading skill, and for other purposes, we can include the whole set of 119 participants.

LEXICAL BASIS OF COMPREHENSION SKILL

79

Figure 5.5. Decision time difference scores (homophones-controls) for high-frequency words. Skilled readers show no reliable homophone interference effect at any stimulus onset asynchrony (SOA). Less-skilled readers show a reliable homophone effect at 450 ms only. Two main things can be seen in Figure 5.5. First, as predicted, less-skilled readers are slower than skilled readers on control words. Although this result may seem obvious enough, it is important theoretically. According to the lexical quality hypothesis, the general lexicon of skilled readers contains more high-quality representations than that of the less-skilled readers. All other lexically based differences, including those from homophones, arise from this fact. And indeed homophone differences are seen. In relation to control words, skilled readers showed longer decision times at the shortest SOA of 150 ms. The homophone effect disappeared by 450 ms and did not return. The less-skilled readers showed no homophone confusion (beyond their confusion for control words) at 150 ms. However, they did at 450 ms. By 2,000 ms, they showed release from confusion. Note that there is a small and unreliable difference remaining at 2,000 ms for the less-skilled readers. As is usually the case, however, nonsignificant differences can be traced to individual reader and word differences that affect variance estimates. Because such differences are the heart of individual-differences research, we return to these in the next section. For now, the general point is that rapid confusion leads to rapid release and delayed confusion leads to delayed release. The rapid confusion-release pattern is characteristic of the skilled readers and the delayed confusion-release pattern is characteristic of the less-skilled readers. A highly general acti-vation-deactivation function is sufficient to explain this pattern. As a word is activated, its identification occurs at a sufficiently high level of activation. Along

80

PERFETTI AND HART

the way, activation can spread to other words that have formal or semantic links to it. Deactivation of nonidentified words occurs naturally in this process.3 No additional mechanism such as suppression failure (Gernsbacher, 1990; Gernsbacher & Faust, 1991a) is needed. Less-skilled comprehenders take longer to show confusion and, accordingly, take longer to show release from it. A second prediction concerns word frequency. Figure 5.5 shows the results for high-frequency words and Figure 5.6 shows the results for low-frequency words. In both Figures 5.5 and 5.6, difference scores between decision times to homophones and controls are shown. A confusion effect is represented by a score above the zero baseline. Figure 5.5 shows that homophone confusion for high-frequency words occurred at the middle SOA (450 ms) for less-skilled readers. (We again see a nonsignificant difference at 2,000 ms.) No reliable homophone confusion occurred for skilled readers at any SOA. For low-frequency words, a different pattern can be seen in Figure 5.6. Less skilled readers show no reliable homophone confusion, whereas skilled readers show confusion at the shortest SOA, 150 ms, which is maintained at 450 ms before releasing. Although the interference effect is small and unreliable for less skilled readers, the effect is above zero (especially at 450 ms). We think this effect is small because less skilled readers' problems with low-frequency words are very general, and simply processing any less-familiar word absorbs the effect that otherwise can be seen for homophones. Less-skilled readers' decision times for low-frequency control words across the three SOAs was around 900 ms for the shortest and longest SOAs and about 830 ms at 450-ms SOA, where one sees the largest difference between controls and homophones in Figure 5.6. To put these results in terms of our gate I gait example, skilled readers are confused by gait but not gate. Less-skilled readers are confused by gate; they are probably confused by gait as well; but they are also confused by stride, a low-frequency control word, and thus have little opportunity to show a confusion that is specific to homophones. These results not only are consistent with the assumptions of the lexical quality hypothesis but also provide evidence for specific nonobvious predictions concerning the time course of homophone confusion and differential patterns for high and low frequency. The temporal pattern has two interesting elements. One is the more rapid confusion and more rapid release from confusion shown by skilled readers, as predicted. A second is that the less-skilled readers not only showed a slower confusion, but they generally did show release from confusion by 2 s. Gernsbacher and Faust (1991a) found that less-skilled readers showed confusion at a shorter interval, about 1350-1400 ms after the final word of a sentence. It appears that confusion may not be a permanent affliction for many less-skilled readers, although we qualify this conclusion below.
3 The process that produces the confusion effect in meaning decisions is not entirely clear. In particular, the probe word should reinitiate activation of the presented word and possible reactivation of its homophone. Obviously this "backward priming" possibility applies equally to probes following sentences. Although our discussion has not explicitly taken this process into account, the same activation and deactivation concepts will apply, but in a more complex way. We are developing computational models that can stimulate meaning decision data, including homophone interference effects, in both "no" and "yes" responses.

LEXICAL BASIS OF COMPREHENSION SKILL

81

80

JL

40

o c

I 0Q
-40 150 2000 450 2000 SOA (ms) 150 450

Less-Skilled

More-Skilled

Figure 5.6. Decision time difference scores (homophones-controls) for low-frequency words. Skilled readers show reliable homophone interference effect at both 150 and 450 ms. Less-skilled readers show no reliable homophone interference. SOA = stimulus onset asynchrony. Lexical Knowledge and Meaning Decisions If it is true that the sources of the comprehension skill differences lie in lexical knowledge, we may be able" to find further evidence. In the course of carrying out this study, we obtained assessments of comprehension skill and lexical knowledge for some 445 students. Although this larger group provides confirming evidence for the interrelationships among various lexical and comprehension skills, we focus on the assessments for just the 119 participants who made the timed meaning decision

judgments. Can we predict meaning decision times from some of these assessments? Our assessments were crude judged by the standards we have suggested in this chapter. Instead of thorough assessments about the specific words used in the meaning decision task, we have general tests of spelling, decoding, and vocabulary that serve as rough indicators of an individual's knowledge of phonological, orthographic, and semantic constituents of a general sample of words. The tested words were not the words of the decision task, and their assessment did not usually meet the standard (e.g., actual spelling and actual meaning analysis) we think is appropriate for careful assessments of lexical quality. Instead they are generally multiple-choice tests that assess the student's ability to select, for example, a correct spelling or meaning. An exception is that our test of phonological decoding requires the participant to decode a pseudoword.

82

PERFETTI AND HAKT

In brief, our tests were as follows: 1. A multiple-choice, timed vocabulary test. 2. A phoneme manipulation test, in which participants repeated a word, deleting a phoneme. For example, stay without the It/. 3. A multiple-choice spelling test, in which the participants chose from five alternatives. 4. A homophone discrimination test, in which participants chose which of two homophone fit into a sentence. For example, The deep Arctic Ocean is inhabited by ___________ (a) wails (b) whales. 5. A pseudoword-decoding test, based on a computerized adaptation of the Word Attack Subtest of the Woodcock-Johnson Psychoed-ucational Battery. The test orders items in difficulty from very easy (e.g., hap) to more difficult (e.g.,phigh). Items were presented one at a time. 6. A word identification test, based on a computerized adaptation of the Word Identification Subtest of the Woodcock-Johnson Psychoeducational Battery. Items were presented one at a time. Accuracy and naming times were collected on all but the first two tasks. Only accuracy was measured for the phoneme manipulation test, and only the total number correct was measured on the vocabulary test. The general results for reading skill are what would be expected. Skilled readers' performance reliably exceeded that of less-skilled readers on all measures. This merely confirms what has been found before: Comprehension skill, even in adults, is associated with lexical skill. The lexical quality hypothesis leads to some more specific predictions, however. For example, the hypothesis predicts more rapid homophone confusion as a function of identification skill, and indeed skilled readers showed early homophone confusion at 150 ms. But what about less-skilled readers? A key idea in the lexical quality hypothesis is that there is a continuum of skill and a continuum of functional word identifi-ability. Do some less-skilled readers also show early homophone confusion? The answer is yes, and it tends to be those who have higher decoding skill, as measured by the assessment of pseudoword reading. This correlation, along with a handful of others that were reliable, is shown in Table 5.1. The correlations in Table 5.1 are quire modest, but it must be emphasized that what is being correlated is a difference score: performance on homophones minus performance on controls. Difference scores show modest correlations when the major factor in reading skillthe abihty to identify wordshas already been removed. In effect, one sees only the residual effect of homophone confusion. Nonetheless, rapid homophone confusiona homophone effect at 150 ms is positively correlated with pseudoword decoding speed. Equally interesting is the fact the continuation of confusion at the longest SOA is negatively correlated with pseudoword decoding speed. Thus, there are some possibly interesting individual differences rather than mere "error" in the tendency for less-skilled readers to show a nonreliable homophone-control difference at the longest SOA (see Figure 5.4). Although some participants indeed showed release from

LEXICAL BASIS OF COMPREHENSION SKILL

83

Table 5.1. Correlations of Lexical Variables With Homophone Interference (Homophone-Control Reaction Times) at 150- and 2,000-ms SOA for Less-Skilled Readers Reading group and variable Less-skilled readers Homophone discrimination HF Homophone discrimination LF Pseudoword reading Comprehension3 Skilled readers Homophone discrimination HF Homophone discrimination LF 150-ms 2,000-ms -.36 .37 -.35 .31 .30

.38 .36

Note: All correlations, p < .05. SOA = stimulus onset asynchrony; HF = high frequency; LF = low frequency. aAssessed by the Nelson-Denny Reading Comprehension Test.

confusion by 2,000 ms, some did not. And the correlations tell us that those who failed to show release from confusion were those slower at decoding. We see a similar pattern for the homophone-spelling test. A failure to show release from homophone confusion was associated with slower and more error-filled performance on a task that required participants to choose the correct form of two homophones, given the meaning of one. Overall, the correlations add support to the hypothesis that lexical quality is responsible for homophone interference. Not only are the patterns of experimental effectsthe temporal patterns and the frequency patternspredicted by the hypothesis, the correlations further suggest that even the "error" in the data is systematically related to measures of lexical processing that are related to the lexical quality assumptions in sensible ways.

Homophones in Context Finally, we return to a brief discussion of context. In the meaning decision experiment, we found specific patterns of homophone interference that distinguished skilled and less-skilled readers. The results support the lexical quality hypothesis. In the larger sample factor analysis, we obtain evidence for the lexical processing differences as a function of reading skill, further suggesting that specific assumptions of the lexical quality hypothesis might be reasonable. If all this is more or less correct, it becomes of interest to examine homophone interference when the constraint of sentence meaning reinforces the evidence in the word form that it must be this word rather than that wordthat it must he gait rather than gate. If top-down comprehension strategies are available to less-skilled readers (Perfetti, 1985; Stanovich, 1980), we may see a reduction of interference. For this test, we carried out a second meaning decision experiment with a new sample of participants defined as in the first experiment. SOAs were varied as in the first experiment. The only difference was that a constraining sen-

84

PERFETTI AND HART

tence was presented, ending in a homophone and followed by a test word after a varying SOA. For example, Because of his leg injury, the man walked with an unusual gait. The key comparison is between a decision to a word related to the meaning of the homophone mate (fence) and a control word. The design was complete and fully counterbalanced, replicating the first experiment. The results showed several interesting outcomes. For one, although the two groups of readers were very accurate in making these decisions, skilled readers were faster, not just for homophones but also for control words. This skill difference was a main effect that did not interact with SOA. Both groups were faster to make correct decisions at 450-ms SOA than at either the shorter (150-ms) or longer (2,000-ms) SOA. Most interesting is the result that neither group showed any effects of homophones measured in the difference between controls and homophones plotted in Figure 5.7. The confusion we observed when the word itself was the only source of information is now eliminated, both for less-skilled readers and for skilled readers.4 We must keep in mind that we have a different group of readers here than in the first experiment. Nevertheless, it is striking to see the disappearance of the confusion effect when a sentence context is available. This is certainly consistent with the findings that even less-skilled readers use context. Its implications, when coupled with the results of the lexical experiment, are clear: Homophone confusion depends on comprehension skill. However, the effect of comprehension skill is mediated solely by lexical factors and is not visible in sentence comprehension itself. This seeming paradox is resolved as follows: The comprehension task here was targeted to the single final word. In effect, the semantic component of the word is reinforced sufficiently to override any unstable orthography and phonology. Less-skilled readers may be more dependent on meaning (because of weakness elsewhere in their lexical representations), thus providing more cues to the meaning is very helpful. More directly persuasive, however, is the clear fact that this group does not comprehend as well as the skilled group. The groups were defined this way. Thus, their general comprehension problem has a clear lexical basis. It can be overcome only in certain tasks that are not typical of ordinary comprehension. There is nothing in this account to suggest that less-skilled readers have a problem in suppression. In terms of the structure building framework (Gernsbacher, 1990), one could say that our context effects demonstrate the enhancement of structures is something that less-skilled readers can do well. The verbal efficiency theory makes the same prediction, based not on enhancement but on basic lexical identification. The slower the process (because of lexical quality), the more room for context facilitation. Moreover, this idea generalizes to account for the facts of homophone confusion when there is no context.

"The slight effect shown by skilled readers at the shortest SOA is not reliable, although it might be interesting. Recall that skilled readers showed this short SOA effect for the word-only experiment. At least some skilled readers may have a brief activation to the phonology of the presented word that is sufficient to interfere with a probe that is only 150 ms later. Indeed, if this difference were reliable, we would be tempted to note its parallel in the research on processing ambiguous words in context. Does the encounter of a word with more than one meaning briefly activate all its meanings, even when spelling disambiguates the word?

LEXICAL BASIS OP COMPREHENSION SKILL

85

80

J,
o c
0)

40

a
5
-40

150 2000

450

2000 SOA (ms)

150

450 More-Skilled

Less-Skilled

Figure 5.7. Constraining sentence context experiment: Decision time differences between controls and homophones for skilled and less-skilled readers. No effect of homophones was found for either group. SOA = stimulus onset asynchrony. When the reader is solely dependent on word form, weaknesses in word form are revealed. This fact results" in the pattern of experimental results we observed: slower activation and deactivation by the less-skilled reader and confusion restricted to more familiar word forms. Thus, we do not need to add a suppression mechanism to the explanation. To do so, one would have to say that less-skilled readers have low-quality word

representations and faulty suppression mechanisms. Of course, that might be the case. Another possibility is that suppression is nothing more than a prolonged period of competing activations. But what would be the source of this prolonged competition? A defective additional mechanism that suppresses the unwanted competitor? Or the single-word identification problem that has caused the competition in the first place? It is possible that suppression will continue to be a necessary mechanism to understand related phenomena in nonlinguistic tasks and even in other language comprehension tasks (Gernsbacher, 1990). But in this one case, words with partially identical forms, it is not necessary. If we are right, then we need to reemphasize the lexical foundations of reading comprehension. Although there are many other component skills in comprehension and an enormous contribution from conceptual knowledge, word identification is the central recurring event of reading. A lot depends on the lexical knowledge that drives it.

86

PERFETTI AND HART

Conclusion
We have argued that the quality of lexical representations is a critical factor in reading and that lexical quality can explain individual differences in simple reading comprehension tasks. Variation in lexical qualitythe extent to which individual word forms and meanings are represented reliably and coherently produces variation in identifying a specific word from among words that share orthographic, phonological, and semantic information. Experiments can expose lexical quality with words that share forms or meaning, thus putting pressure on lexical quality. Our experiments used homophones, which share phonology, to study the simple comprehension of word pairs with a focus on the time course of confusion effects. The important findings, consistent with the lexical quality hypothesis, were that skilled comprehenders showed a more rapid build up of confusion and a more rapid release from this confusion compared with less-skilled readers. Furthermore, skilled readers, but not less-skilled readers, avoided confusions for the high-frequency member of a homophone pair. A second experiment found no confusions for either'skilled or less-skilled comprehenders when the homophone was in a sentence context consistent with its meaning. The results demonstrate that variations in lexical quality produce variation in simple comprehension performance that correlates with overall reading comprehension skill. Accordingly, they provide new and partly nonobvious evidence for a specific lexical knowledge explanation of individual differences in comprehension.

References Berent, I. & Shimron, J. (1997). The representation of Hebrew words: Evidence from the obligatory contour principle. Cognition, 64, 39-72. Bock, K., & Levelt, W. (1994 ). Language production: grammatical encoding. In M. A. Gemsbacher (Ed.), Handbook ofpsycholinguistics (pp. 945-984). San Diego: Academic Press. Carpenter, P. A., Miyake, A., & Just, M. A. (1994). Working memory constraints in comprehension: Evidence from individual differences, aphasia, and aging. In M. A. Gemsbacher (Ed.), Handbook of psycholinguistics (pp. 1075-1122). San Diego: Academic Press. Duffy, S. A., Morris, R. K., & Rayner, K. (1988). Lexical ambiguity and fixation times in reading. Journal of Memory and Language, 27, 429-446. Gemsbacher, M. A. (1990). Language comprehension as structure building. Hillsdale, NJ: Erlbaum. Gemsbacher, M. A., & Faust, M. E. (1991). The mechanism of suppression: A component of general comprehension skill. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17, 245-262. Hogaboam, T. W., & Perfetti, C. A. (1978). Reading skill and the role of verbal experience in decoding. Journal of Educational Psychology, 70, 717-729. Perfetti, C. A. (1985). Reading ability. New York: Oxford University Press. Perfetti, C. A. (1992). The representation problem in reading acquisition. In P. B. Gough, L. C. Ehri, & R. Treiman (Eds.), Reading acquisition (pp. 145-174). Hillsdale, NJ: Perfetti, C. A., & Roth, S. F. (1981). Some of the interactive processes in reading and their role in reading skill. In A. M. Lesgold & C. A. Perfetti (Eds.), Interactive processes in reading (pp. 269-297). Hillsdale, NJ: Erlbaum. Perfetti, C. A., Goldman, S. R., & Hogaboam, T. W. (1979). Reading skill and the identification of words in discourse context. Memory & Cognition, 7, 273-282. Perfetti, C. A., & Tan, L. H. (1998). The time-course of graphic, phonological, and semantic activation in Chinese character identification. Journal of Experimental Psychology: Learning, Memory, and Cognition, 24, 1-18. Stanovich, K. E. (1980). Toward an interactive-compensatory model of individual differences in the development of reading fluency. Reading Research Quarterly, 16, 32-71. Stanovich, K. E., & West, R. F. (1981). The effect of sentence context on on-going word recognition: Tests of a two-process theory. Journal of Experimental Psychology: Human Perception and Performance, 7, 658-672. West, R. F., & Stanovich, K. E. (1978). Automatic contextual facilitation in readers of three ages. Child Development, 49,111-121.

SCIENTIFIC STUDIES OF READING, 4(2), 101-132 Copyright 2000, Lawrence Erlbaum Associates, Inc.

The Association of Rapid Temporal Perception With Orthographic and Phonological Processing in Children and Adults With Reading Impairment
James R. Booth
Department of Communication Sciences and Disorders Northwestern University

Charles A. Perfetti
Department of Psychology University of Pittsburgh

Brian MacWhinney and Sean B. Hunt

Department of Psychology Carnegie Mellon University

Adults and children with reading impairment (N= 67) were administered a rapid auditory task, a rapid visual task, and a battery of orthographic and phonological tasks. Our results support a differential development model of reading disability that argues that deficits in rapid auditory ability in children are primarily associated with problems in phonological processing, whereas deficits in rapid visual ability in children are primarily associated with problems in orthographic processing (Farmer & Klein, 1995). In contrast to the children, the adults showed a strong relation between rapid auditory ability and both orthographic and phonological processing. These results suggest that continued deficits in auditory ability may have a pervasive and negative impact on word processing in general. In addition, adults did not exhibit a relation between rapid visual ability and orthographic-processing problems. Orthographic-processing deficits may Requests for reprints should be sent to James R. Booth, Department of Communication Sciences and Disorders, Northwestern University, 2299 North Campus Drive, Evanston, IL 60208-3560. E-mail: j -booth@nwu.edu

102

BOOTH ETAL.

result from a reading delay condition that can be overcome with increased reading exposure (Harm & Seidenberg, 1999). Many correlational and training studies have shown that later reading achievement is predicted best by earlier phonological knowledge, as measured by phoneme segmentation, phoneme deletion, and phoneme blending (Wagner, Torgesen, & Rashotte, 1994). In fact, attempts at phonological decoding during reading may be the main learning mechanism that brings about knowledge of whole word spellings (Jorm & Share, 1983). Indeed, some have suggested that phonological decoding from letters to sounds may be a "self-teaching device" that functions to establish orthographic representations and, hence, an autonomous orthographic lexicon (Share & Stanovich, 1995). Thus, phonological processes not only have a causal role in later word decoding ability (Wagner et al., 1994), but may also serve to increase the representational properties of words as they are actually read. Taken together, this research suggests that deficits in phonological processing may be the marker characteristic of reading disability. However, there is also clear evidence that some children with reading problems have difficulty in learning the orthographic spelling patterns of English (Castles & Coltheart, 1993). Children with reading problems are not a homogeneous group, but rather a heterogeneous group including individuals with distinct behavioral deficits that may result from different etiologies.

DIFFERENTIAL DEVELOPMENT MODEL OF READING DISABILITY

There is debate in the literature over what underlies individual differences in orthographic and phonological processing. Some investigators claim that reading and oral-language disabilities in children are caused by a low-level deficit in rapid temporal perception. In fact, studies show that reading disabilities are associated with the inability to represent rapidly presented visual stimuli (Eden, Stein, Wood, & Wood, 1995) and that oral-language disabilities are associated with the inability to represent rapidly presented auditory stimuli (Tallal et al, 1996). Tallal (1980) and colleagues argued that children with oral language impairment suffer from a problem in processing rapid temporal changes in speech. This deficit may result in the inability to form accurate representations of phonemes that may eventually result in a phonemic awareness deficit and then reading disability. Within the field of reading, a transient system deficit in the magnocellular stream was implicated in readers with impairments (Lovegrove, Martin, & Slaghuis, 1986). The magnocellular system is important for processing stimuli that change rapidly over time, such as moving objects, whereas the parvocellular system is important for

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

103

processing location, color, and orientation. The parvocellular system is clearly involved in reading because letter and word recognition require the analysis of location and orientation, but individuals with dyslexia appear to not have abnormalities in the parvocellular system. Indeed, physiological studies of deceased individuals with dyslexia have revealed abnormalities only in magnocellular neurons in the lateral and medial geniculate nucleus, which are involved in visual and auditory processing, respectively (Galaburda, Menard, & Rosen, 1994; Livingstone, Rosen, Drislane, & Galaburda, 1991). Imaging studies also show less activity in V5/MT when individuals with dyslexia are processing moving stimuli (Eden et al., 1996), and the amount of activity in V5/MT (extrastriate visual area) is related to reading rate in individuals with dyslexia (Demb, Boynton, & Heeger, 1998). This visual deficit may result in increased "persistence" of stimuli in the visual system, and therefore, the reader with impairment may not be able to effectively inhibit past visual word forms to accurately perceive new visual word forms when reading text. This persistence disrupts normal reading acquisition and may slow the acquisition of low-frequency exception words. A recent model of reading disabilities (Farmer & Klein, 1995) suggested that rapid visual perception deficits may lead to orthographic problems, whereas rapid auditory perception deficits may lead to phonological problems. Figure 1 displays the differential development model of reading disability that suggests deficits in visual temporal processing should be primarily related to orthographic problems in children, whereas auditory ability deficits should be primarily related to phonological problems in children and adults. These orthographic and phonological deficits in children with reading disabilities may correspond to the two types of dyslexia discussed in the developmental literature (Castles & Coltheart, 1993; Manis, Seidenberg, Doi, McBride-Chang, & Petersen, 1995; Stanovich, Siegel, & Gottardo, 1997). Individuals with surface dyslexia have deficits in orthographic processing, such as in naming exception words and in spelling tasks. Individuals withphonological dyslexia have deficits in phonological processing, such as in naming nonwords and in making phonetic judgments. Note that these subtypes of children tend to have relative deficits in orthographic or phonological processing. Overall, individuals with dyslexia are impaired at exception word and nonword reading relative to readers without impairment. There are at least three potential levels of perceptual deficits in reading disability (Farmer & Klein, 1995). The first level is the detection of a single stimulus. Detection requires the determination of the presence or absence of a stimulus. The literature provides little evidence for auditory or visual deficits in this very low level perceptual ability in readers with impairment (Blackwell, Mclntyre, & Murray, 1983; Tallal, 1980). The lack of deficits in the detection task is crucial to the hypothesis that readers with impairment have a temporal-processing deficit because the detection task does not require temporal perception. The second level is the individuation of two stimuli. Individuation requires the determination of the existence of two stimuli with a very short interstimulus interval (ISI). The litera-

104

BOOTH ETAL.

Orthographic System 'Delay' Visual Temporal Processing

Deficit

Orthographic

Auditory

Phonemic System 'Disability'

Reading Disabled Phonological

Fl G U R E1 A differential development model of reading disability that suggests deficits in rapid visual temporal processing are primarily related to orthographic problems, whereas rapid auditory ability deficits are primarily related to phonemic problems. The narrow arrows in the visual-orthographic pathway represent that this is a reading delay condition that can diminish in adulthood. The wide arrows in the auditory-phonemic pathway represent that these problems often persist into adulthood. The arrows in the figure are bidirectional because we assume that higher levels of processing can influence lower levels of processing and vice versa. The phonemic system is connected to the orthographic system by a bidirectional arrow to represent that persistent deficits in phonological processing may negatively impact orthographic processing. This model is based loosely on ideas in Farmer and Klein (1995) and Harm and Seidenberg (1999).

ture shows some evidence for visual deficits, but little evidence for auditory deficits, in readers with impairment in the individuation tasks (McCroskey & Kidder, 1980; Slaghuis & Lovegrove, 1985). However, more recent studies have questioned the relation between dyslexia and individuation ability as measured by visual flicker and spatial frequency tasks (Victor, Conte, Burton, & Nass, 1993; Walther-Mueller, 1995). The third level is the temporal ordering of two or more stimuli. This requires the reader to report the order of stimuli presented in a sequence. For example, in one auditory temporal-ordering task, the participant is required to report aloud the sequence of three high or low pure tones. Temporal-ordering tasks are more demanding than the previous two levels because these tasks require an accurate representation of stimuli and short-term memory processes. The literature provides strong evidence for auditory deficits in temporal ordering and some evidence for visual deficits in readers with impairment (Brannon & Williams, 1988; Tallal, 1980). When taken together, this research suggests that reading disability should be related to deficits in visual individuation and both visual and auditory temporal ordering.

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

105

Most of the aforementioned studies have investigated nonspeech stimuli, such as pure tones. There has been some controversy in the literature as to whether speech represents a specialized area of auditory processing (Liberman, 1982; Liberman & Mattingly, 198 5). A review of the literature concluded that readers with impairment have deficits in detection, individuation, and temporal-ordering tasks when speech stimuli are used (McBride-Chang, 1995). As discussed earlier, the research on nonspeech stimuli suggests limited deficits in detection and individuation, but the research on speech stimuli seems to suggest deficits at these levels. One of the most studied individuation tasks in individuals with dyslexia is categorical speech perception. These studies show that deficits in categorical perception may be weak (Werker & Tees, 1987) and apply only to a subset of individuals with dyslexiafor example, those who have low phonological awareness (Manis et al., 1997). It maybe that readers with impairment have deficits in accurately representing phonetic factors (articulatory features), but they are not impaired at perceiving rapid spectral changes (Mody, Studdert-Kennedy, & Brady, 1997).

THIS STUDY The general approach of this study was to examine the relation between individual differences in perceptual ability and reading skill within a sample of children with reading impairment (Experiment 1) and adults with reading impairment (Experiment 2). This study examined the association of both rapid visual and rapid auditory ability with both orthographic and phonological processing. No previous study has included an examination of the relation among all of these component processes (Reed, 1989). However, all of these component processes must be examined to provide a proper test of the differential development model of reading disability presented in Figure 1. A failure to obtain the predicted correlations of sensory-processing deficits with orthographic and phonological deficits would invalidate the differential development model. However, the correlational nature of this study makes it impossible to reach any conclusions regarding causal mechanisms. Indeed, it may be difficult to determine causality because perceptual and word-processing components are likely to be interactive (McClelland & Rumelhart, 1981). Computational modeling as well as animal studies clearly show that there is an interaction between high- and low-level processing. Studies with male rats have shown that cortical lesions result in fewer large and more small neurons in the medial geniculate nucleus and an accompanying deficit in fast auditory temporal processing (Herman, Galaburda, Fitch, Carter, & Rosen, 1997). This study employed naming and priming tasks to measure orthographic and phonological processing. We realize that all reading measures require both orthographic and phonological knowledge; however, these components of processing may be differentially weighted in particular tasks. We chose exception word nam-

106

BOOTH ET AL.

ing as a measure of orthographic processing and nonword naming as a measure of phonological processing because these have been used in previous studies on subtypes of reading disorders (Castles & Coltheart, 1993). Exception word reading requires knowledge of spelling patterns that violate statistical regularities in English, whereas nonword reading requires generalization of knowledge about phonological structure to unfamiliar nonwords. The priming task used the brief duration identification paradigm (Perfetti, Bell, & Delaney, 1988). This task allows an assessment of the magnitude of orthographic and phonological priming. This paradigm requires the participant to write down a very briefly presented target (e.g., 60 ms) displayed after a briefly presented (60 ms) orthographic or phonological prime. In a previous investigation of readers without impairment, Booth, Perfetti, and MacWhinney (1999) reported that good readers (second through sixth graders) exhibited more phonological and orthographic priming than did poor readers. Because these stimuli were displayed for a duration (less than 60 ms) that was brief enough to prevent complete processing, this suggests that good readers activate orthographic and phonological information more quickly and automatically than do poor readers. This study established the utility of the brief duration identification paradigm for examining reading skill differences in orthographic and phonological priming.

EXPERIMENT 1 Experiment 1 examined the relation of visual and auditory perceptual ability to orthographic and phonological processing in a population of children with reading impairment. We expected these low-ability children with reading impairment to have difficulty in processing phonological as well as orthographic representations. Apart from this overall deficit, we expected rapid visual ability to be uniquely associated with orthographic-processing deficits and rapid auditory ability to be uniquely associated with phonological-processing deficits. This pattern of relations should be reflected in the naming and possibly in the priming tasks. These findings would be consistent with the differential development model presented in Figure 1.

Method Participants Participants were 35 children (M= 15.2 years, range = 11-18 years) from an educational program for children with specific learning disabilities in the Pittsburgh metropolitan area. Only children with at least a second-grade proficiency in read-

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

107

ing, based on their scores on the Word Identification subtest (Woodcock, 1987) were chosen. The participants could not complete the experimental tasks without this level of proficiency in reading. All participants had been administered the Wechsler Intelligence Scale for Children (Wechsler, 1991) within the last 5 years. The mean Full IQ was 80 (range = 65-109)4 children had an IQ score below 70, and 10 children had an IQ score between 70 and 80. All children also had existing neuropsychological examinations from registered clinical psychologists. The most common Diagnostic and Statistical Manual of Mental Disorders (4th ed. [DSM-IV]; American Psychiatric Association, 1994) diagnoses were Cognitive Disorder (43%), Borderline Intellectual Functioning (39%), Reading Disorder (37%), Mathematics Disorder (19%), Writing Disorder (21%), and Attention Deficit Hyperactivity Disorder (24%). No children had diagnosed behavior or emotional problems according to school records. In addition, no children had motoric problems that interfered with their ability to complete the experimental protocol. We realize that there is controversy over how to define the population that has reading impairments. Some researchers argue that there must be a significant discrepancy between reading achievement and IQ, whereas others argue that there need not be a discrepancy for a child to have a learning disability in reading (Shaywitz, Fletcher, & Shaywitz, 1994; Stanovich & Siegel, 1994). There is some research suggesting that children with reading impairment who are with and without an IQ discrepancy behave in similar ways (Shaywitz, Fletcher, Holahan, & Shaywitz, 1992; Siegel, 1992). In our view, the inclusion of children who have low IQ scores and who lack the IQ discrepancy makes our findings more general in that the results of the study appear to apply to a larger population of children and adults. However, the inclusion of many children with low IQ scores may make it difficult to compare the results of this study to other studies of children with reading disorders who have IQs in the normal range. Nevertheless, to make sure that the results of our study were not caused by age or IQ differences, we examined the relation of auditory and visual perceptual ability with orthographic and phonological processing after partialling for age and IQ.

Materials and Procedure

All materials administered to the participants were presented on identical 15-in. MultiScan Macintosh monitors controlled by a Macintosh PowerPC computer. All tasks described hereafter were presented with PsyScope (Cohen, MacWhinney, Flatt, & Provost, 1993), and responses were recorded with the CMU button box.

Rapid visual. The participant was asked to fixate on a cross in the middle of the computer screen about 50 cm away. After the participant pressed a button on a

108

BOOTH ETAL.

button box, four to nine black dots flashed one after the other at the same place on a white computer screen. The end of the series was marked by a number line (3 to 10) presented on the bottom of the screen. After the participant saw the number line, he or she was asked to report orally the number of dots. The experimenter entered the number reported by the participant into the computer. Each dot was . 167 x . 167 in. and was presented for 40 ms. The interstimulus intervals were 200,300, or 400 ms. There were 9 practice trials and 72 test trials, so this meant that there were 24 test trials at each interstimulus interval. This rapid visual task was used because performance on a similar version was shown to be related to reading problems (Eden et al., 1995). We considered the rapid visual task to be primarily an individuation task because it required the participant to detect gaps between the same visual stimuli (Farmer & Klein, 1995). However, this task may have also had other cognitive components that contributed to performance. For example, some children may have used a counting strategy in which they retrieved number names and internally articulated them. At the very least, this task had a short-term memory component because it required participants to remember the number of dots presented and then to report that number to the experimenter. However, we argue that if the memory component of this task was responsible for the effects discussed later, then there is no reason to predict why rapid visual ability should be associated only with visual orthographic and not with visual phonological processing.

Rapid auditory. The participant was asked to listen to a series of two orthree tones. The participant controlled the presentation of each trial with a button box. The end of each trial was marked by an "enter order" prompt on the computer screen. The participant was then asked to report aloud the sequence of tones to the experimenter, who then entered the sequence into the computer (e.g., HLH for high, low, high). The pure tones were digitized for presentation by computer through headphones. The tones were either 100 or 300 Hz, with a 75- or 150-ms duration and an interstimulus interval of 100 or 300 ms. All tones faded in and out for 20 ms. All possible combinations of frequencies, duration, and interstimulus intervals were presented randomly. No feedback was given during the test session (64 trials), although during the practice session (12 trials) the experimenter provided feedback until the participant could accurately discriminate between the high and low tones. The practice session contained the easier two-tone sequences with long durations (225 ms) and long ISIs (450 ms). This rapid auditory task was used because a similar version has been shown to be related to oral-language problems (Tallal & Piercy, 1973,1974,1975). We considered the rapid auditory task to be primarily a temporal ordering task because it required the participant to report the sequence of different stimuli (Farmer & Klein, 1995). However, this task also had a short-term memory component be-

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

109

cause it required the participants to remember the order of tones presented. Some participants may have also internally rehearsed the articulatory code for the tone labels. However, we argue that if the memory component of this task was responsible for the effects discussed later, there is no reason to predict why rapid auditory ability should be associated only with visual phonological and not with visual orthographic processing.

Standard words. The Word Identification subtest (Woodcock, 1987) contains approximately 80% regular words and 20% irregular words. The test begins with easy words, such asplay, and ends with more difficult words, such as Zeitgeist. Test administration was stopped when the participant pronounced six consecutive words incorrectly. This test was used to measure general reading skill.

Nonwords. The Word Attack subtest (Woodcock, 1987) is a nonword naming measure containing items that range in complexity from easy words at the beginning, such as dat, to more difficult words at the end, such as byrcal. Participants read all nonwords in this task regardless of their accuracy. We considered this task to primarily measure phonological-processing skill.

Exception words. The Exception Words subtest (Adams & Huggins, 1985) requires the participant to read aloud 45 exception words. The items range from easy words, such as ocean, to more difficult words, such as baroque. Participants read all exception words in this task regardless of their accuracy. We considered this task to primarily measure orthographic-processing skill.

Orthographic and pseudohomophone priming. This task involved 15

practice trials followed by a series of 120 test trials. A fixation cross was displayed before each trial, and the participant was asked to press a button to begin each trial. Each trial consisted of a brief presentation of a nonword prime (60 ms), followed immediately by a brief presentation of a real word target (60 ms), which was then immediately followed by a mask of the form XXXXX (500 ms). There was no interstimulus interval between prime and target or between target and mask. Primes were always presented in uppercase, and targets were always presented in lowercase, so that any observed priming effects would have to be attributed to an abstract letter representation and not a lower level case-specific visual representation. The pattern mask was used to disrupt the processing of the target. Without the pattern mask, performance on this task would have been near ceiling for the more advanced readers because of luminance persistence of the target on the computer monitor. All

110

BOOTH ETAL.

stimuli were presented in white letters on a black background in 16-point Courier font. All words were four (1.6 cm) or five (2 cm) letters in length. The participants' task was to write down the target word after each practice and test trial. The dependent variable on this task was the percentage of correct responses. Reaction time was not measured in this task because factors not associated with orthographic and phonological processing were likely to influence writing timefor example, the amount of time to move pencil to the proper location on paper. Participants were encouraged to guess about the identity of the target if they were not sure. No partial credit was given. We are assuming that the participants' written responses reflected access to orthographic and phonological forms, although other factors, such as knowledge of common letter patterns, may have influenced the translation from a mental representation to a written word. There were one within-item and two between-item factors in the priming task. The within-item factor varied the three prime types. The pseudohomophone primes were phonologically identical to the targets (e.g., TUME-tomb). The orthographic primes shared the same overlapping letters with the target words as pseudohomophone primes but different nonoverlapping letters (e.g., TAMS-tomb). The control primes shared no letters in common with target (e.g., USAN- tomb). There were three counterbalancing lists, so that across participants, each prime preceded each target an equal number of times. This meant that the three groups of participants corresponding to the three counterbalancing lists received a different list of prime-target pairs. The two between-item factors were (a) orthographic similarity between prime and target and (b) target frequency. Orthographic similarity was defined on the basis of the formula that takes into account identical letters in the same position, in adjacent positions, and in the first and final positions (Van Orden, Johnston, & Hale, 1988). Our orthographic similarity value (os = .52) was exactly the same for the orthographic and pseudohomophone primes and was similar to, but slightly lower than, those found in other studies (e.g., os = .62-.68; Van Orden et al., 1988). TUME-tomb is an example of a low orthographic similarity pair, whereas HOAP-hope is an example of a high orthographic similarity pair. The low-frequency words had a mean frequency level of 11.5 in 1 million. The high-frequency words had a mean frequency level of 167 in 1 million (Kucera & Francis, 1967).

Results and Discussion

The following analytical procedure was used for Experiment 1 with children with reading impairment and later for Experiment 2 with adults with reading impairment. First, we examined the relation of a number of stimuli, duration, and interstimulus interval to accuracy levels in the rapid temporal ability tasks. Second,

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

111

we examined individual differences in naming accuracy on the exception word and nonword tasks, and then we examined the relation among overall naming accuracy and orthographic and phonological priming. Finally, we examined the relation between rapid temporal ability and orthographic and phonological processing. Rapid Temporal Processing Figure 2 displays accuracy on the rapid visual ability measure as a function of the number of dots in the set and the ISI between each dot in the set. A 6 (dot number: 4, 5, 6,7, 8, 9) x 3 (dot ISI: 200,300,400 ms) analysis of variance (ANOVA) yielded significant main effects for dot number, F(5, 629) = 3.35,p < .05; dot ISI, F(2, 629) = 46.12, p < .001; and a trend for a significant interaction between dot number and dot ISI, F(10, 629) = 2.60, p < .10. The main effects show that sets with a greater number of dots or with shorter ISIs had lower accuracy levels than did sets with a fewer number of dots or with longer ISIs, respectively. However, the interaction suggests that increasing the number of dots at the long ISIs had a minimal effect on accuracy, whereas increasing the number dots at the shorter ISIs reduced accuracy levels substantially. Figure 3 displays accuracy on the rapid auditory ability measure as a function of the number of tones in the set, the duration of the tones in the set, and the ISI between each tone in the set. A 2 (tone number: 2,3) * 2 (tone duration: 75,150 ms) x Children

Number of dots
FIGURE 2 Mean correct on the rapid visual ability measure as a function of the number of dots in the set (4 to 9) and the ISI between each dot in the set (200,300, or 400 ms) for the children in Experiment 1. Error bars indicate one standard error.

112

BOOTH ETAL.

Children
1.00.90-

Mean correct

.80-. 70.60.50.40.
i

Tone.lSI
150,100ms

150,300ms

75,300ms
A

7 5,100ms

Number of tones FIGURE 3 Mean correct on the rapid auditory ability measure as a function of the number of tones in the set (2,3), the duration of the tones in the set (75,150 ms), and the ISI between each tone in the set (100,300 ms) for the children in Experiment 1. Error bars indicate one standard error.

2 (tone ISI: 100, 300 ms) ANOVA yielded a significant main effect for tone number,^, 279) = 8.21,p < .01; tone duration, F(\, 279) = lA9,p< .01; and tone ISI, F(\, 279) = 8.70,p < .01. These main effects indicate that sequences of two tones had higher accuracy levels than sequences of three tones, that sequences of long tones had higher accuracy levels than sequences of short tones, and that sequences with long ISIs had higher accuracy levels than sequences with short ISIs. For the remaining analyses involving rapid temporal or auditory ability, we used an average measure, including all items in each instrument, because increasing the number of items increased the reliability of the measure. This was justified because the correlations between the subscales in the temporal ability measures were high (rave = .69). By subscales, we mean accuracy levels on each combination of number, duration, and ISI factors. Individual Differences in Word Naming Table 1 displays the means for the word naming, rapid temporal ability, and IQ measures for the children with reading impairment. We were interested in whether the children had particular deficits on the exception words or nonwords measures, so we calculated difference scores (d) between exception words and standard words and between nonwords and standard words. These difference scores give an estimate of the children's relative deficits in orthographic and phonological pro-

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING TABLE 1 Means and Standard Deviations for Word Naming, Rapid Temporal, and IQ Measures in Experiment 1 for the Children Who Were Reading Impaired M Naming Standard words Exception words Nonwords Rapid temporal Visual Auditory IQ Verbal Performance SD

113

50.9 38.1 34.7 43.6 66.6 81.9 78.5

18.4 28.1 24.1 21.1 23.2 15.7 13.9

Note. Means for the naming and rapid temporal measures are percentages. Means for the IQ scores are standard scores (M= 100, SD = 15).

cessing. The difference scores for orthographic (d = 12.8) and phonological processing (d= 16.2) were not significantly different, ^(34) = 1. \2,p = .27, suggesting that the children with reading impairment were having difficulty with phonological as well as with orthographic processing. All analyses on the standard words and nonwords measures were calculated on the raw scores, even though normative data are available for these measures (Woodcock, 1987). This was done because the normative data are not sensitive for the adult populations in Experiment 2 for two reasons. First, the normative sample only included adults up to 33 years of age; second,fhere are a limited number of words on the upper end of the range of item difficulties. Nevertheless, we computed percentile ranks, standard scores, and age equivalents on the standard words measure for the reading disabled children in Experiment 1 and the reading disabled adults in Experiment 2 (see Appendix). These calculations show that on average the children and adults scored lower than the 13th percentile in word naming. Clearly, both children and adults in this study were severely reading disabled. These figures also suggest that the adults sampled may be slightly less disabled than the children sampled, but these scores must be interpreted with caution because they are not sensitive to older populations.

Orthographic and Phonological Priming

In the priming analyses, naming accuracy was the one between-participant variable. Naming accuracy was treated as a continuous regressor variable to more accurately represent underlying differences in ability. Because this design requires

114

BOOTH ETAL.

an analysis focusing on a continuous participant variable, only participant analyses could be computed. Naming accuracy was defined as the mean score on the three naming measuresstandard words, exception words, and nonwords. All three naming measures were highly intercorrelated: standard-exception, r(34) = .94, p < .001; standard-non, r(34) = .82,/? < .001; exception-non, r(34) = .77, p < .001. Thus, combining them was a reasonable approach. There were two within-participant independent variablesorthographic similarity and word frequency. On the basis of a previous study of children without disabilities (Booth, Perfetti, & MacWhinney, 1999), we expected the priming effects to depend on the amount of orthographic overlap of the prime with the target and on the frequency of the target. In particular, we expected more orthographic priming for pairs with high orthographic similarity and for high-frequency targets. There were two dependent variables of interest in the priming taskorthographic priming and pseudohomophone priming. Orthographic priming was calculated as the difference (d) between accuracy in the orthographic and control conditions. Because the orthographic primes had overlapping letters and sounds with the target, we must attribute these priming effects to a combination of orthographic and phonemic overlap. Pseudohomophone priming was calculated as the difference (d) between accuracy in the pseudohomophone and orthographic conditions. Because the pseudohomophone primes had the same level of orthographic similarity to the targets as did the orthographic primes, it is possible to view the pseudohomophone priming effects as resulting purely from phonological priming. Planned comparisons were computed separately for the orthographic and pseudohomophone priming dependent variables, and these results are presented in different sections. In summary, orthographic and pseudohomophone priming difference scores were submitted to a 2 (orthographic similarity: high, low) x 2 (frequency: high, low) analysis of covariance (ANCOVA) with naming accuracy as an independent regressor variable. All data are presented as if naming accuracy were a dichoto-mous variable (median split) for clarity of presentation and ease in interpretation, but all statistical analyses used the continuous naming accuracy variable. Figure 4 displays accuracy levels for the three priming conditions for the high and low naming accuracy scores of children.

Pseudohomophone priming. As predicted, the most important finding was that children with high naming accuracy scores benefited more from pseudohomophone priming (d= 5.4%) than did children with low naming accuracy scores (tf = -1.7%), F(l, 139) = 9.12,/? < .01. Presumably, the children with high naming accuracy scores had quicker access to higher quality phonological representations, which allowed them to use this information for target identification more efficiently than the children with low naming accuracy scores. We have ar-

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

115

Children
.70.60' .50-. 40' .30' .20' .10' 0.00. igh Low
i

Correct

Prime
Pseudo

Mean

3Z=
Naming accuracy

Ortho
i

Control

FIGURE 4 Mean correct forthe pseudohomophone (phonological), orthographic, and control priming conditions for the high and low naming accuracy scores of children in Experiment 1. Error bars indicate one standard error.

gued elsewhere that the more precise and redundant (Perfetti, 1992) lexical representations of better readers who do not have impairment allow them to more efficiently activate representations for words and grapheme-phoneme correspondences at these very short presentation durations, and this results in larger priming effects (Booth et al., 1999). This relation held in the this experiment even though our sample was of children who have severe reading impairment.

Orthographic priming. As predicted, children with high naming accuracy scores also exhibited more orthographic priming (d = 21.4%) than did the children withlownaming accuracy scores (d= 8. \%),F{\, 139) = 36.20,/? <. 001. In addition, high orthographic similarity pairs (d= 18.9%) benefited more from priming than did low orthographic similarity pairs (d = 11.0%), F(l, 139) = 8.78,;? < .01. The high similarity pairs showed more priming because the larger amount of letter overlap allowed orthographic information to influence processing to a greater degree. Finally, high-frequency targets (d = 19.0%) benefited more from orthographic priming than did low-frequency targets (d = 10.9%), F(l, 139) = 9.25, p < .01. This can be accounted for by a model of word recognition that assumes that high-frequency targets are driven more strongly by orthographic input because of their increased frequency of exposure (Plaut, McClelland, Seidenberg, & Patterson, 1996). Stronger input for high-frequency targets means that less priming information is needed to produce correct identification. Low-frequency targets are not driven strongly by orthographic input, so more priming information is needed to produce correct identification.

116

BOOTH ET AL.

Comparison of Readers With Impairment to Reading Matched Control Population The data from a group of control children were taken from a separate study that examined the development of orthographic and phonological knowledge (Booth et al., 1999). The control children were not part of this study, so we have a limited number of measures on these children. Our measures included age, phonological priming, orthographic priming, and word naming ability. The control children were matched to the children with reading impairment in terms of their standard word naming ability on Word Identification (Woodcock, 1987). The means on word naming for the control group (M = 49.5, SD = 18.2) and group of individuals with reading impairment (M= 45.6, SD = 19.9) were not significantly different, f(69) = 1.23,p = .22. Because these groups were matched based on reading age, the mean chronological age of the control children (M= 9.2 years, range = 7-11 years) was significantly less than the mean age of the readers with impairment (M= 15.2 years, range= 11-18years), /(69) = 13.81,p< .001. Figure 5 displays the accuracy levels for the priming conditions for the children with reading disabilities and for reading-age-matched children without disabilities. To compare the accuracy levels of these two groups, a 2 (group: nondisabled, disabled) x 3 (prime: pseudohomophone, orthographic, control) ANOVA was computed. This analysis yielded significant main effects for group, F(l, 209) = Children
.70.60' .50' .40' .30-. 20' .100.00.

correct

Prime Pseudo *
Orttio

Mean

Control

Reading group

FIGURE 5 Mean correct for the pseudohomophone (phonological), orthographic, and control priming conditions for the children with reading disabilities in Experiment 1 and for reading-age-matched children without disabilities. Error bars indicate one standard error.

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

117

18.85,/? < .001, and for prime, F(\, 209) = 34.82,/? < .01. Even though the readers with disabilities were matched to the children without disabilities in standard word reading, the readers with disabilities (19.9%) exhibited significantly lower overall accuracy levels than did those without disabilities (30.9%). This suggests that readers with disabilities cannot quickly and automatically activate consistent information between the prime and the target, or that they show more interference from the inconsistent information between the prime and target. The lack of a significant Group * Prime interaction, F(l, 209) = 1.39, p = .25, suggests that both groups exhibited a similar amount of priming. A closer examination of the means presented in Figure 5, however, reveals a trend for the reading-age-matched children without disabilities to show more phonological priming (pseudohomophone - orthographic) than the readers with disabilities. Taken together, these results suggest that the readers who have impairments were not qualitatively different from the controls without disabilities, but rather the readers with impairment were just quantitatively delayed in their development of orthographic to phonological correspondences.

Relation of Rapid Temporal With Orthographic and Phonological Processing Table 2 displays the results of the hierarchical regression equations predicting accuracy on the standard words, exception words, and nonwords measures. We first entered age and then IQ into the equation to partial out these effects so that the predictive power of the rapid perceptual measures could not be attributed to differences in age or IQ. Indeed, these analyses showed that both age and IQ explained unique variance in word naming, Then we determined whether rapid auditory abilTABLE 2 Hierarchical Regression Equations Predicting Accuracy (Multiple R) on the Three Naming Tasks in Experiment 1 for the Children Who Were Reading Impaired Standard Words Age IQ Rapid visual Rapid auditory Rapid auditory Rapid visual .37* .46* .65* .70 .65* .70 Exception Words .42* .53* .73* .73 .62* .73* Nonwords .31* .51* .61* .71* .70* .70

*p < .05 for the unique variance explained by that variable when entered into the equation. Age then IQ were entered into the equation. Next, either rapid visual then rapid auditory or rapid auditory then rapid visual were entered into the equation.

118

BOOTH ETAL.

ity explained unique variance in naming accuracy after partialling for rapid visual ability. Finally, we determined whether rapid visual ability explained unique variance in naming accuracy after partialling for rapid auditory ability. These analyses showed that rapid visual ability explained unique variance in orthographic processing as measured by exception words naming, t(34) = 3.02,p < .01, but not in phonological processing as measured by nonword naming, t(34) = 0.40, p = .67. Conversely, rapid auditory ability explained unique variance in phonological processing, /(34) = 2.69,p < .05, but not in orthographic processing, /(34) = 0.70, p = .48. These results support the differential development model of reading disability presented in Figure 1, which argues orthographic-processing problems are associated primarily with rapid visual ability deficits, whereas phonological-processing problems are associated primarily with rapid auditory ability deficits. The rapid perceptual measures explained unique variance in orthographic or phonological processing despite the fact that rapid visual and rapid auditory ability were significantly correlated, r(34) = .59,p < .001. Furthermore, the rapid perceptual measures explained common variance in orthographic and phonological processing. This may be due to more general abilities shared by the rapid perceptual tasks, such as processing speed or attention. We also examined the unique variance explained by rapid auditory and rapid visual ability in the magnitude of orthographic and phonological priming. Table 3 displays the results of the hierarchical regression equations predicting orthographic and phonological priming. These analyses show that rapid auditory ability explained unique variance in the magnitude of phonological priming, ?(34) = 2.67, p < .05, but not in orthographic priming, /(34) = 0.27, p = .78. These results are consistent with the differential development model presented in Figure 1. However, the differential development model was not supported by the finding that rapid visual ability did not explain unique variance in orthographic priming, ?(34) = 0.70,p = .49. Rapid visual ability may not have explained unique variance in orTABLE 3 Hierarchical Regression Equations Predicting (Multiple R) Pseudohomophone and Orthographic Priming in Experiment 1 for the Children Who Were Reading Impaired iphone Priming Age IQ Rapid visual Rapid auditory Rapid auditory Rapid visual .08 .15 .20 .35* .35* .35 Orthographic Priming .37* .40 .46 .48 .47 .48

*p < .05 for the unique variance explained by that variable when entered into the equation. Age then IQ were entered into the equation. Next, either rapid visual then rapid auditory or rapid auditory then rapid visual were entered into the equation.

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

119

thographic priming because the orthographic primes are both orthographically and phonologically similar to the targets, as compared to the control primes. These null results warrant further research on the relation between rapid visual ability and orthographic priming.

EXPERIMENT 2 As discussed earlier, individuals with surface dyslexia tend to have deficits in orthographic processing, and individuals with phonological dyslexia tend to have deficits in phonological processing. A recent computational model of dyslexia suggests that there is a different pattern of orthographic and phonological deficits for adults than for children with reading disabilities (Harm & Seidenberg, 1999). This model argues that phonological dyslexia results from a pervasive deficit that is characterized by inaccurate, incomplete phonological representations, suggesting that, without intensive auditory and phonemic training, phonological deficits should persist into adulthood because developing accurate phonological representations requires fine-grained discrimination in rapid acoustic transitions during speech input (Tallal &Piercy, 1973,1974,1975). There is some evidence that training can increase auditory and phonemic discrimination in children with language impairment (Tallal et al., 1996) and in Japanese adults who initially could not discriminate between the kl and IV in the English language (McClelland, 1999). As mentioned earlier, deficits in speech perception tend to be weak (Werker & Tees, 1987) and apply only to a subset of individuals with dyslexia (Manis et al., 1997). In contrast to phonological dyslexia, the computational model argues that surface dyslexia results from slower learning or less reading exposure that is characterized by a reading delay (Harm & Seidenberg, 1999). This suggests that orthographic deficits should diminish by adulthood because of increased exposure to written language. The notion that phonological deficits should persist into adulthood, whereas orthographic deficits should diminish by adulthood, is reflected in the differential development model presented in Figure 1 as thicker lines in the auditory-phonological route than in the visual-orthographic route. Indeed, behavioral research suggests that improvements in adults with reading impairment primarily involve gains in orthographic processing and not in phonological processing (Bruck, 1992; Pennington, Lefty, Van Orden, Bookman, & Smith, 1987). Adults with reading impairment may access word meanings based on visual features and spelling patterns but continue to have deficits in phonological processing. Experiment 2 examined the relation of rapid visual and auditory ability with orthographic and phonological processing in a population of adults with reading impairment. On the basis of the differential development model presented in Figure 1, we predicted that the adults with reading impairment would be more

120

BOOTH ETAL.

impaired on the nonword naming measure of phonological processing than on the exception word naming measure of orthographic processing. The differential development model also predicts that initially an inefficient visual system should slow down the process of reading and the acquisition of a sight word vocabulary in children. This slow down should be primarily reflected in limited acquisition of low-frequency exception words. Later, an inefficient visual system should have less of an impact on reading because increasing exposure to written words should allow the acquisition of the low-frequency exception words. In other words, we predicted a relation between rapid visual ability and orthographic processing in children with reading impairment but not in adults with reading impairment. The differential development model also predicts that phonological-processing deficits should persist into adulthood and that there should continue to be a strong relation between rapid auditory ability and phonological-processing skill. In fact, deficits in auditory ability and phonological processing may eventually prevent the attainment of normal levels of orthographic-processing skill because of the interaction between the phonological and orthographic systems in reading development (Plaut et al., 1996). The self-teaching hypothesis discussed earlier also predicts a snowballing effect of early deficits in phonological decoding on later orthographic skill (Share & Stanovich, 1995). Therefore, the differential development model predicts that deficits in rapid auditory ability should be related to orthographic-processing problems in adults with reading impairment.

Method Participants
Participants were 32 adults (M= 34.7 years, range = 19-51 years) from an educational program in the Pittsburgh metropolitan area for adults with specific learning disabilities. Only adults with at least a second-grade proficiency in reading, based on their scores on the Word Identification subtest (Woodcock, 1987) were chosen. All participants had been administered the Wechsler Adult Intelligence Scale (Wechsler, 1985) within the last 5 years. The mean Full IQ was 87 (range = 75-110)10 adults had IQ scores less than 80. All adults also had existing neuropsychological examinations from registered clinical psychologists. The most common DSM-IV diagnoses (American Psychiatric Association, 1994) were Cognitive Disorder (50%), Borderline Intellectual Functioning (33%), Reading Disorder (28%), Mathematics Disorder (28%), Writing Disorder (17%), and Attention Deficit Hyperactivity Disorder (17%). No adults had diagnosed behavior or emotional problems, according to their records. In addition, no adults had motoric problems that interfered with their ability to complete the experimental

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

121

protocol. This adult sample was very similar to the child sample in Experiment 1 except that the adults had a slightly higher mean Full IQ (87 vs. 80, respectively). These samples may have been very similar because both programs are administered by the same nonprofit organization and use similar criteria for admission into the program. Furthermore, many of the adults in this study were enrolled at some time in the children's program.

Materials and Procedure

The materials and procedure for the adults in Experiment 2 were exactly the same as for the children in Experiment 1.

Results and Discussion Rapid Temporal Processing

Figure 6 displays accuracy on the rapid visual ability measure as a function of the number of dots in the set and the ISI between each dot in the set. A 6 (dot number: 4, 5, 6, 7, 8, 9) x 3 (dot ISI: 200, 300, 400 ms) ANOVA yielded significant main effects for dot number, F(5, 575) = 5.43,/? < .001, and dot ISI, F(2, 575) = 8.78, p < .001. The main effects indicate that sets with a greater number of dots or with shorter ISIs had lower accuracy levels than did sets with a fewer number of dots or with longer ISIs, respectively. Figure 7 displays accuracy on the rapid auditory ability measure as a function of the number of tones in the set, the duration of the tones in the set, and the ISI between each tone in the set. A 2 (tone number: 2,3) x 2 (tone duration: 75,150 ms) * 2 (tone ISI: 100, 300 ms) ANOVA yielded a significant main effect for tone duration, F(l, 255) = 4.09,p< .05, and tone ISI, F{\, 255) = 8.33,/- < .01. These main effects show that long tones had higher accuracy levels than short tones and that long ISIs had higher accuracy levels than short ISIs. As with Experiment 1, for the rest of the analyses involving rapid visual or auditory ability in Experiment 2, we used an average measure, including all the items in each instrument. The correlation between the subscales in the temporal-processing measures were high (rave =.71), and increasing the number of the items in the scale increased the reliability of the measure.

Individual Differences
Table 4 displays the means for the word naming, rapid temporal ability, and IQ measures for the adults with reading impairment. To determine relative deficits in orthographic and phonological processing, we calculated the same difference

Adults

300ms 200ms
5 6 8 7

Number of dots
FIGURE 6 Mean correct on the rapid visual ability measure as a function of the number of dots in the set (4 to 9) and the ISI between each dot in the set (200,300,400 ms) for the adults in Experiment 2. Error bars indicate one standard error.

Adults
1.00' .90'

o .70 c
CO

a> S .60 .50 .40

ll II
Two

<
i

ToneJSI
i i

150,300ms 150,100ms 75,300ms A 75,100ms

Three

Number of tones

FIGURE 7 Mean correct on the rapid auditory ability measure as a function of the number of tones in the set (2,3), the duration of the tones in the set (75,150 ms), and the ISI between each tone in the set (100, 300 ms) for the adults in Experiment 2. Error bars indicate one standard error.

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

123

TABLE 4 Means and Standard Deviations for Word Naming, Rapid Temporal, and IQ Measures in Experiment 2 for the Adults Who Were Reading Impaired M Naming Standard words Exception words Nonwords Rapid temporal Visual Auditory IQ Verbal Performance 71.2 69.1 45.2 50.5 75.2 87.6 88.1 12.4 21.1 21.3 21.3 23.7 10.7 10.2 71 55 77 86 89 93 89 SD Child/Adult (%)

Note. Means for the naming and rapid temporal measures are percentages. Means for the IQ scores are standard scores (M = 100, SD = 15). Child/Adult is calculated as the percentage accuracy for children in Experiment 1 divided by the percentage accuracy for adults in Experiment 2. This measure provides an index of magnitude of the difference between children and adults.

scores (d) as those in Experiment 1. The adults in Experiment 2 had larger deficits on nonwords (d = 26.0) than on exception words (d = 2.2), compared to accuracy on the standard words measure, /(31) = 8.52,;? < .001. The findings for adults with reading impairment were consistent with other studies that show a disproportionate improvement in orthographic processing, compared to phonological processing, with increasing reading exposure (Brack, 1992; Pennington et al., 1987). This disproportionate improvement is in contrast to Experiment 1, which found that children with reading impairment performed equally poorly on the exception words and nonwords measures. Indeed, when accuracy levels in children and adults were directly compared, exception word accuracy in children was 55% of adult levels, whereas nonword accuracy in children was 77% of adult levels. These results support the differential development model presented in Figure 1, which argues that adults with reading problems may partially overcome deficits in orthographic processing. Like the children in Experiment 1, the adults in Experiment 2 scored lower on the rapid visual measure than on the rapid auditory measure. Moreover, when the accuracy levels in children and adults were directly compared, it appears that development is marked by the same improvement in visual and auditory processing. Accuracy levels in rapid visual processing in children was 86% of adult levels, whereas accuracy levels in rapid auditory processing in children was 89% of adult levels. Unfortunately, the design of our study cannot determine whether the children or the adults have absolute deficits in temporal processing because our study did not include a control population that was given the rapid perceptual ability measures. However, it is likely that our child and adult samples did have absolute

124

BOOTH ETAL.

deficits because other studies that have used similar tasks have found deficits in populations with reading and oral language impairments (Eden et al., 1995; Tallal & Piercy, 1973, 1974, 1975).

Orthographic and Phonological Priming

As in Experiment 1, orthographic and pseudohomophone priming difference scores (d) were submitted to a 2 (orthographic similarity: high, low) x 2 (frequency: high, low) ANCOVA, with naming accuracy as an independent regressor variable. Naming accuracy was defined as the mean score on the three naming measuresstandard words, exception words, and nonwords. All three naming measures were highly intercorrelated: standard-exception, r(31) = .90, p < .001; standard-non, r(31) =. 87, p < .001; exception-non, r(31) = .72, p < .001. Figure 8 displays accuracy levels for the three priming conditions for the high and low naming accuracy adults.

Pseudohomophone priming. As predicted, adults with high naming accuracy scores {d = 15.1 %) benefited more from pseudohomophone priming than did

Adults
.70.60.50' .40.30-. 20.10' 0.00.

Mean

correct

+
t
<

'rime Pseudo
Ortho

Control

High

N aming accuracy Low

FIGURE 8 Mean correct for the pseudohomophone (phonological), orthographic, and control priming conditions for the high and low naming accuracy adults in Experiment 2. Error bars indicate one standard error.

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

125

adults with low naming accuracy scores (d = 2.8%), F(l, 127) = 9.91,p < .01. The adults with high scores had quicker access to higher quality phonological representations allowed them to use this information for target identification more often than did the adults with low scores. This finding is consistent with the differential development model presented in Figure 1, which argues that there should be pervasive phonological-processing deficits in adults with severe reading problems.

Orthographic priming. In contrast to the ability differences for the children in Experiment 1, adults with high naming accuracy scores (d=29A%) did not exhibit significantly more orthographic priming than did those with low naming accuracy scores (d =21.8%), F{\, 127)= 1.15,;? = .285. This is consistent with the differential development model that argues orthographic deficits may be partially overcome in adult readers with impairments. Like the children in Experiment 1, high orthographic similarity pairs (d = 35.8%) benefited more from orthographic priming than low orthographic similarity pairs (d= 15.4%), F(l, 127) = 35.91,p < .001. In addition, high-frequency targets (d = 30.8%) benefited more from orthographic priming than did low-frequency targets (d= 15.4%), F(l, 127) = 9.32,/? < .01. See Experiment 1 for a discussion of these effects.

Relation of Rapid Temporal With Orthographic and Phonological Processing

As with Table 2 in Experiment 1, Table 5 displays the results of hierarchical regression equations predicting standard words, exception words, and nonwords naming. We found that rapid visual ability did not explain unique variance in
TABLE 5 Hierarchical Regression Equations Predicting Accuracy (Multiple R) on the Three Naming Tasks in Experiment 2 for the Adults Who Were Reading Impaired Standard Words Age IQ Rapid visual Rapid auditory Rapid auditory Rapid visual .16 .57* .58 .71* .71* .71 Exception Words .07 .47* .49 .64* .63* .64 Nonwords .38* .61* .63 .73* .73* .73

*p < .05 for the unique variance explained by that variable when entered into the equation. Age then IQ were entered into the equation. Next, either rapid visual then rapid auditory or rapid auditory then rapid visual were entered into the equation.

126

BOOTH ETAL.

orthographic processing, t(3l) = 0.66,p = .52, or phonological processing, <31) = 0.30, p = .76. This is not due to a ceiling effect or limited variability in performance on the rapid visual ability or naming measures. The adults exhibited levels of variability comparable to those of the children (see Tables 1 and 4). These results support the differential development model presented in Figure 1, which argues for the absence of a relation between rapid visual ability and orthographic processing in adults. In contrast to the findings with rapid visual ability, rapid auditory ability explained unique variance in both orthographic processing, (31) = 2.80, p< .01, and phonological processing, J(31) = 2.84, /? < .01. These results also support the differential development model that argues orthographic- and phonological-processing problems in adults with reading impairment should be associated with deficient rapid auditory ability. Table 6 displays the prediction of pseudohomophone and orthographic priming by the rapid temporal ability measures. Rapid temporal processing did not explain significant variance in priming in the adults with reading impairment after partialling for age and IQ. In contrast for the children with reading impairment in Experiment 1, rapid auditory processing may have explained unique variance in pseudohomophone priming because IQ did not account for much variance in the hierarchical regression equation. In contrast to the children with reading impairment, the correlation between rapid visual and rapid auditory ability in adults was not significant, r(31) = .23,p = .20. This suggests that the temporal-processing deficits in the children with reading impairment may reflect a common underlying problem, whereas the adults' reading problems may result from a more specific deficit in rapid auditory processing.

TABLE 6 Hierarchical Regression Equations Predicting (Multiple R) Pseudohomophone and Orthographic Priming in Experiment 2 for the Adults Who Were Reading Impaired iphone Priming Age IQ Rapid visual Rapid auditory Rapid auditory Rapid visual .23 .35* .37 .38 .35 .38 Orthographic Priming .12 .16 .17 .17 .16 .17

*p < .05 for the unique variance explained by that variable when entered into the equation. Age then IQ were entered into the equation. Next, either rapid visual then rapid auditory or rapid auditory then rapid visual were entered into the equation.

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

127

GENERAL DISCUSSION Support for the Differential Development Model The goal of this study was to test the differential development model of readers with impairment (presented in Figure 1). This required the measurement of rapid visual ability, rapid auditory ability, orthographic processing, and phonological processing in a group of children and adults with reading impairment. The differential development model argues that, in children, rapid visual ability deficits should be associated with orthographic-processing problems and that rapid auditory ability deficits should be associated with phonological-processing problems (Farmer & Klein, 1995). However, the differential development model argues that orthographic-processing deficits are a reading delay condition and should diminish by adulthood. Thus, there should not be a relation between rapid visual ability and orthographic processing in adults. In contrast, the differential development model argues that phonological deficits are associated with incomplete, inaccurate phonological representations and should continue to adulthood (Harm & Seidenberg, 1999). Thus, there should be a strong relation between rapid auditory ability and phonological processing in adults. Indeed, early deficits in rapid auditory ability and phonological-processing skills may even produce deficits with orthographic processing later in development because the orthographic, phonological, and semantic systems are interactively connected (Plaut & Booth, 1999; Plaut et al., 1996). The snowballing effect of early deficits on later reading development in other areas has been referred to as the Matthew Effect (Stanovich, 1986). Thus, there should also be a relation between rapid auditory ability and orthographic-processing skill in adults. The results of this study were largely consistent with the differential development model. The children with reading impairment in Experiment 1 showed large deficits on both the exception words measure of orthographic processing and the nonwords measure of phonological processing compared to standard words processing. Furthermore, individual differences in representing rapid visual information explained unique variance in orthographic processing, and individual differences in representing rapid auditory information explained unique variance in phonological processing. The adults with reading impairment in Experiment 2 showed relatively small deficits in the exception words measure of orthographic processing but relatively large deficits in the nonwords measure of phonological processing. Furthermore, rapid visual ability did not explain variance in orthographic or phonological processing, even though there was substantial variance on all of these measures. Individual differences in representing rapid auditory information, however, did explain unique variance in phonological as well as orthographic processing.

128

BOOTH ETAL.

Extensions of This Study

There are several ways in which the results of this study can be extended. First, our rapid visual and auditory-processing measures may be contaminated by short-term memory demands, so it is difficult to interpret the effects involving perceptual ability to only differences in transient processing. Further studies should examine the relation of individuation tasks in the auditory and visual modalities (McCroskey & Kidder, 1980; Slaghuis & Lovegrove, 1985) with orthographic and phonological processing. Second, this study employed only two measures of orthographic and phonological processing. To test the generality of our findings, further research should examine whether rapid visual processing is related to other orthographic measures (Olson, Kliegl, Davidson, & Foltz, 1985) and whether rapid auditory processing is related to other phonological measures (Wagner et al., 1994). Third, this study only included a limited sample of control children that were given only a subset of the experimental tasks. Although the results of our study produced some provocative findings, further studies should include control populations who are given all the experimental measures so more definite conclusions can be drawn regarding the absolute magnitude of deficits at different ages. Fourth, this study examined children and adults with reading impairment with a wide range of IQ scores. We accounted for this variability by partialling out the effects of IQ from the relevant analyses. In light of the current controversy over absolute- versus discrepant-based definitions of reading impairment (Shaywitz et al., 1992, 1994; Siegel, 1992; Stanovich & Siegel, 1994), future studies should more directly examine the relation of rapid perceptual ability with orthographic and phonological processing in absolute versus discrepant populations.

Implications for Remedial Intervention

The findings of this study have some general, but tentative, implications. We suggest that low word identification skills in children and adults may be associated with a pervasive deficit in phonological and rapid auditory ability. This implies that children and adults with deficits in these areas would benefit from an extensive intervention program involving a combination of auditory discrimination and phonological training (Merzenich et al., 1996; Tallal et al., 1996; Wagner et al., 1994). Because we believe that these processes are inherently interactive, training at one level would not be as successful as training at both the auditory and phonological levels. This study also suggests that lower identification skills in certain children, but not in adults, may be marked by a deficit in orthographic processing that is related to deficits in rapid visual ability. It appears that with increasing reading exposure adults may be able to effectively develop their orthographic-processing skills. This suggests that children with deficits in visual and orthographic processing may

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

129

benefit from a program designed to dramatically increase their exposure to low-frequency exception words through reading.

ACKNOWLEDGMENTS
This research was funded in part by National Institute of Child Health and Human Development Grant 5-F32-HD08255-02, by National Institute of Deafness and Other Communication Disorders Grant 5-R01-HD23998, and by National Institute of Mental Health Grant 5-T32-MH19102. We thank the students, teachers, and administrators at Katherine Dean Tillotson School and the Association for Children and Adults With Learning Disabilities. We thank Joanne Carslisle, Frank Van Santen, and Yasuaki Harasaki for their helpful comments on earlier versions of this article. We also thank Laura Holly Bevan for her assistance in conducting the experiments.

REFERENCES
Adams, M. J., & Huggins, A. W. F. (1985). The growth of children's sight word vocabulary: A quick test with educational and theoretical implications. Reading Research Quarterly, 20, 560- 568. American Psychiatric Association. (1994). Diagnostic and statistical manual of mental disorders (4th ed.). Washington, DC: Author. Blackwell, S. L., Mclntyre, C. W., & Murray, M.E.(1983). Information processed from brief visual displays by learning disabled boys. Child Development, 54, 977-940. Booth, J. R., Perfetti, C. A., & MacWhinney, B. (1999). Quick, automatic, and general activation of orthographic and phonological representations in young readers. Developmental Psychology, 35, 3-19. Brannon, J. R., & Williams, M. C. (1988). Developmental verses sensory deficit effects on perceptual processing in reading disabled. Perception & Psychophysics, 44, 437-444. Bruck, M. (1992). Persistence of dyslexic's phonological awareness deficits. Developmental Psychology, 28, 874-886. Castles, A., & Coltheart, M. (1993). Varieties of developmental dyslexia. Cognition, 47, 149-180. Cohen, J. D., MacWhinney, B., Flatt, M., & Provost, J. (1993). PsyScope: An interactive graphic system for designing and controlling experiments in the psychology laboratory using Macintosh computers. Behavior Research Methods, Instruments, and Computers, 25, 257-271. Demb, J. B., Boynton, G. M., & Heeger, D. J. (1998). Functional magnetic resonance imaging of early visual pathways in dyslexia. Journal ofNeuroscience, 18, 6939-6951. Eden, G. F., Stein, J. F., Wood, H. M., & Wood, F. B. (1995). Temporal and spatial processing in reading disabled and normal children. Cortex, 31, 451-468. Eden, G. F., VanMeter, J. W., Rumsey, J. M., Maisog, J. M Woods, R. P., & Zeffiro, T. A. (1996). Abnormal processing of visual motion in dyslexia revealed by functional brain imaging. Nature, 382, 66-69. Farmer, M. E., & Klein, R. M. (1995). The evidence for a temporal processing deficit linked to dyslexia: A review. Psychonomic Bulletin and Review, 2, 460-493.

130

BOOTH ETAL.

Galaburda, A. M, Menard, M. T., & Rosen, G. D. (1994). Evidence for aberrant auditory anatomy in developmental dyslexia. Proceedings of the National Academy Sciences USA, 91, 8010-8013. Harm, M. W., & Seidenberg, M. S. (1999). Phonology, reading, and dyslexia: Insights from connectionist models. Psychological Review, 106, 491-528. Herman, A. C, Galaburda, A. M, Fitch, R. H., Carter, A. R., & Rosen, G. D. (1997). Cerebral microgyria, thalamic cell size and auditory temporal processing in male and female rats. Cerebral Cortex, 7, 453^64. Jorm, A. F., & Share, D. L. (1983). Phonological recoding and reading acquisition. Applied Psycholinguistics, 4, 103-147. Kucera, H., & Francis, W. N. (1967). Computational analysis of present day American English. Providence, RI: Brown University Press. Liberman, A. M. (1982). On finding that speech is special. American Psychologist, 37, 148-167. Liberman, A. M., & Mattingly, I. G. (1985). The motor theory of speech perception revised. Cognition, 21, 1-36. Livingstone, M., Rosen, G. D., Drislane, F. W., & Galaburda, A. (1991). Physiological and anatomical evidence for a magnocellular deficit in developmental dyslexia. Proceedings of the National Academy of Sciences USA, 88, 7943-7947. Lovegrove, W., Martin, F., & Slaghuis, W. (1986). A theoretical and experimental case for a visual deficit in specific reading disability. Cognitive Neuropsychology, 3, 225-267. Manis, F. R., McBride-Chang, C, Seidenberg, M. S., Keating, P., Doi, L. M., Munson, B., & Petersen, A. (1997). Are speech perception deficits associated with developmental dyslexia? Journal of Experimental Child Psychology, 66, 211-235. Manis, F.R., Seidenberg, M. S., Doi, L.M.,McBride-Chang,C.,& Petersen, A. (1995). On the bases of the two subtypes of developmental dyslexia. Cognition, 58, 157-195. McBride-Chang, C. (1995). Phonological processing, speech perception, and reading disability: An integrative review. Educational Psychologist, 30, 109-121. McClelland, J. L. (1999). Failures to learn and their remediation: A Hebbian account. Manuscript in preparation, Carnegie Mellon University, Pittsburgh, PA. McClelland, J. L., & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88, 375-407. McCroskey, R. L., & Kidder, H. C. (1980). Auditory fusion among learning disabled, reading disabled, and normal children. Journal of Learning Disabilities, 13, 18-25. Merzenich, M. M., Jenkins, W. M., Johnston, P., Schreiner, C, Miller, S. L., & Tallal, P. (1996). Temporal processing deficits of language-learning impaired children ameliorated by training. Science, 217, 77-81. Mody, M., Studdert-Kennedy, M., & Brady, S. (1997). Speech perception deficits in poor readers: Auditory processing or phonological coding? Journal ofExperimental Child Psychology, 64, 199-231. Olson, R. K., Kliegl, R., Davidson, B. J., & Foltz, G. (1985). Individual and developmental differences in reading disability. In G. E. MacKinnon & T. G. Waller (Eds.), Reading research: Advances in theory and practice (pp. 1-64). Orlando, FL: Academic. Pennington, B. C, Lefty, D. L., Van Orden, G. C, Bookman, M. O., & Smith, S. D. (1987). Is phonology bypassed in normal or dyslexic development? Annals of Dyslexia, 37, 62-89. Perfetti, C. A. (1992). The representation problem in reading acquisition. In P. B. Gough, L. C. Ehri, & R. Treiman (Eds.), Reading acquisition (pp. 145-174). Hillsdale, NJ: Lawrence Erlbaum Associates, Inc. Perfetti, C. A., Bell, L., & Delaney, S. M. (1988). Automatic (prelexical) phonetic activation in silent word reading: Evidence from backward masking. Journal of Memory and Language, 27, 59-70. Plaut, D. C, & Booth, J. R. (1999). Individual and developmental differences in semantic priming: Empirical findings and computational support for a single-mechanism account of lexical processing. Manuscript submitted for publication.

ORTHOGRAPHIC AND PHONOLOGICAL PROCESSING

131

Plaut, D. C, McClelland, J. L., Seidenberg, M. S., & Patterson, K. (1996). Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103, 56-115. Reed, M. A. (1989). Speech perception and the discrimination of brief auditory cues in reading disabled children. Journal of Experimental Child Psychology, 48, 270-292. Share, D. L., & Stanovich, K. E. (1995). Cognitive processes in early reading development: Accommodating individual differences into a model of acquisition. Issues in Education, 1, 1- 57. Shaywitz, B. A., Fletcher, J. M, Holahan, J. M., & Shaywitz, S. E. (1992). Discrepancy compared to low achievement definitions of reading disability: Results from the Connecticut Longitudinal Study. Journal of Learning Disabilities, 25, 639-648. Shaywitz, S. E., Fletcher, J. M., & Shaywitz, B. A. (1994). A conceptual framework for learning disabilities and attention deficit-hyperactivity disorder. Canadian Journal ofSpecial Education, 9,1-32. Siegel, L. S. (1992). An evaluation of the discrepancy definition of dyslexia. Journal of Learning Disabilities, 25, 618-629. Slaghuis, W. L., & Lovegrove, W. J. (1985). Spatial-frequency-dependent visible persistence and specific reading disability. Brain and Cognition, 4, 219-240. Stanovich, K. E. (1986). Matthew effects in reading: Some consequences of individual differences in the acquisition of literacy. Reading Research Quarterly, 21, 360-407. Stanovich, K. E., & Siegel, L. S. (1994). Phenotypic performance profile of children with reading disabilities: A regression based approach to the phonological-core variable-difference model. Journal of Educational Psychology, 86, 24-53. Stanovich, K. E., Siegel, L. S., & Gottardo, A. (1997). Converging evidence for phonological and surface subtypes of reading disability. Journal of Educational Psychology, 89, 114-127. Tallal, P. (1980). Auditory temporal perception, phonics, and the reading disabilities in children. Brain and Language, 9, 182-199. Tallal, P., Miller, S. L., Bedi, G., Byma, G., Wang, X., Nagarajan, S. S., Schreiner, C, Jenkins, W. M., & Merzenich, M. M. (1996). Language comprehension in language-learning impaired children improved with acoustically modified speech. Science, 271, & 1-84. Tallal, P., & Piercy, M. (1973). Developmental aphasia: Impaired rate of non-verbal processing as a function of sensory modality. Neuropsychologia, 11, 389-398. Tallal, P., & Piercy, M. (1974). Developmental aphasia: Rate of auditory processing and selective impairment of consonant perception. Neuropsychologia, 12, 83-93. Tallal, P., & Piercy, M. (1975). Developmental aphasia: The perception of brief vowels and extended stop consonants. Neuropsychologia, 13, 69-74. Van Orden, G. C, Johnston, J. C, & Hale, B. L. (1988). Word identification in reading proceeds from spelling to sound to meaning. Journal of Experimental Psychology: Learning Memory and Cognition, 14, 371-386. Victor, J. D., Conte, M. M., Burton, L., & Nass, R. D. (1993). Visual evoked potentials in dyslexics and normals: Failure to find a difference in transient or steady-state responses. Visual Neuroscience, 10, 939-946. Wagner, R. K., Torgesen, J. K., & Rashotte, C. A. (1994). Development of reading-related phonological processing abilities: New evidence of bidirectional causality from a latent variable longitudinal study. Developmental Psychology, 30, 73-87. Walther-Mueller, P. U. (1995). Is there a deficit of early vision in dyslexia? Perception, 24, 919-936. Wechsler, D. (1985). Wechsler Adult Intelligence ScaleRevised. San Antonio, TX: The Psychological Corporation. Wechsler, D. (1991). Wechsler Intelligence Scale for ChildrenThird Edition. San Antonio, TX: Har-court Brace. Werker, J. F., & Tees, R. C. (1987). Speech perception in severely disabled and average reading children. Canadian Journal of Psychology, 4, 48-61.

132

BOOTH ET AL.

Woodcock, R. W. (1987). Woodcock Reading Mastery TestsRevised. Circle Pines, MN: American Guidance Service.

Manuscript received January 19, 1999 Final revision received August 23, 1999 Accepted September 20, 1999

APPENDIX Mean Percentile Ranks, Standard Scores, and Grade Equivalents for the Standard Words Measure for the Children and Adults With Reading Impairment in Experiments 1 and 2, Respectively Children Percentile rank Standard score" Grade equivalentb 6 63 4-0 Adults 12 77 6-10

'M= 100, SD= 15. The first number is the year, and the second number is the month.

LANGUAGE AND LINGUISTICS 7.3:637-657,2006 2006-0-007-003-000171-1

Visual Analysis and Lexical Access of Chinese Characters by Chinese as Second Language Readers
Ying Liu12, Charles A. Perfetti1, and Min Wang3
2

University of Pittsburgh Liaoning Normal University 3 University of Maryland

To assess the learning of word form, pronunciation, and meaning in an unfamiliar writing system, we carried out Event Related Potential (ERP) experiments with learners of Chinese at the end of their first and second terms of Chinese class at an American university. The subjects were required to recognize a target Chinese character or English word with ERP recorded. They named filler targets indicated by a signal 1000ms after the onset of the stimuli. The orthographic processing of characters and words was extracted as a 200ms component by Principle Component Analysis (PCA). The semantic processing was extracted as a 400ms component (N400). The 200ms PCA component was negative at occipital (N200) and positive at frontal electrodes (P200). It was sensitive to visual analysis and lexical access respectively. ERP results showed that the visual analysis of Chinese was more difficult than English at the first term, but not the second term. The lexical access was more difficult and the semantic processing was slower for Chinese than English at both terms. Faster lexical access was obtained for familiar characters at the first term, but not the second term. The separation of visual analysis and lexical access at the second term indicates a threshold style processing of Chinese characters for the learners with moderate reading proficiency. Key words: Chinese as a second language, ERP, sinograms

1. Introduction
One issue that recently raised lots of interest was how the characteristic of writing system influences the reading process. A particular interesting contrast is between Chinese and English. Chinese, a non-alphabetic writing system, provides a case of high contrast for alphabetic systems. Its graphic units, characters, do not represent phonemes, but rather morphemic (meaning-bearing) syllables. This research was supported by National Science Foundation Grant BN0113243 and SBE-0354420, and National Natural Science Foundation of China Grant 30470567.

Ying Liu, Charles A. Perfetti, and Min Wang

Chinese characters are composed of radicals, basic units that can sometimes give a cue to the pronunciation and meaning of characters. For example, 0 ri 'sun' is a simple character consisting of only one radical. There are also lots of compound characters containing two radicals, such as flf qing 'green', which is composed by a top and a bottom radical. There are also even more complex characters composed of three or more radicals. For example, Hff qing 'sunshine' consists of B + ff. This three-radical compound is related in meaning to its left radical and in pronunciation to its right radical (notice that the pronunciation similarity in this case includes phonemes but not tone). But the compound pronunciation and meaning are not always consistent with its radicals (Perfetti, Zhang & Berent 1992, Y. Zhou 1978). In the modern Simplified Chinese writing system there are a total of 7,785 characters with 623 radicals (Li & Liu 1988). Even though a writing system with these properties encourages the hypothesis that reading in Chinese is strictly a visual-form-to-meaning process (e.g., Baron & Strawson 1976, Chen, Yung & Ng 1988, Hoosain & Osgood 1983, Tzeng & Hung 1978, Wang 1973, Zhou & Marslen-Wilson 1996), recent behavioral studies lead to the conclusion that automatic activation of both meaning and pronunciation occurs in reading Chinese characters, as it does in English (Chua 1999, Perfetti & Tan 1998, Perfetti & Zhang 1995, Xu, Pollatsek & Potter 1999, Zhang, Perfetti & Yang 1999). Thus, despite the differences in their input units (characters vs. letters) and mapping functions (syllables vs. phonemes), Chinese and alphabetic systems are similar at this general level, with important processing differences in details (Perfetti, Liu & Tan 2005). More specifically, Chinese is processed at a threshold style and English at a cascade style (Coltheart, Rastle, Perry, Langdon & Ziegler 2001). The threshold style processing of Chinese is illustrated by the interactive constituency model (Perfetti, Liu & Tan 2005). The model is a network of linked units of orthographic, phonological, and semantic constituents across which activation spreads. In the model, a successful lexical access needs the activation of all three constituents. Its input units are radicals and spatial relationship between the radicals. The radical input and the phonological levels of the model can be considered distributed representations, whereas the orthographic and semantic representations can be considered localized representations. Simulation clearly captures the pattern of graphic priming at shorter SOA, whereas it turns into inhibition under longer SOA. The facilitation occurs because visually similar orthographic units are activated by the same radical; so with a graphic prime, the activation level of the target, which shares a radical with the prime, is nearer to threshold than it would be otherwise; hence, initial graphic facilitation without the orthographic unit of prime character itself reaching threshold.

638

Visual Analysis and Lexical Access of Chinese Characters

When an orthographic unit does reach threshold, it sends its activation to the phonological and semantic units, allowing phonological and semantic priming effects to occur. But because it has reached threshold, the orthographic unit of the prime is competing with the target. More importantly, the appearance of the target keeps the prime orthographic unit activated longer than it should be, because of the radical shared by the prime and target orthographic units. This competition can occur with a prime that is not graphically similar to the target, but because it does not share a radical with the target, the target unit can suppress the prime unit very quickly. The net result is a competition that delays the identification of the target longer for the graphically related condition than unrelated control. Thus, two important form priming effects are simulated successfully within the same processing timeline. After a brief period of pre-threshold graphic facilitation, graphic inhibition and phonological facilitation simultaneously emerge, followed by a semantic facilitation. The theory currently applies mainly to native speakers of Chinese. What about those learning Chinese as a second language (CSL)? What relevance is the theory to alphabet-users when they read Chinese? Does a CSL learner read Chinese in a way similar to native Chinese speakers? There are three components related to the issue: orthographic analysis, phonological access, and meaning retrieval. Of these components, orthographic analysis of component radicals and their positional information is important for character recognition, and they feed input to the orthographic units. Shu & Anderson (1999) found that first and second graders can differentiate legal or illegal radical positions in non-characters. Children in higher grades and adults were also found to be sensitive to radical legality effects (Peng, Li & Yang 1997, Taft, Zhu & Peng 1999). Wang, Perfetti & Liu (2003) found CSL learners were also sensitive to the structure composition of Chinese characters, and were able to identify simple characters better than compound characters. This result suggests that learners acquired the orthographic analysis skill rather quickly and process the character form similarly to native Chinese speakers. Liu, Wang & Perfetti (2006) tapped into the issue of constituency processing more specifically by using a primed naming paradigm with a fixed 500ms SGA on the learners. The results showed 61ms of orthographic facilitation at the end of the first term of learning, with no phonological (-5ms) or semantic (-11ms) effect. This pattern is similar to native Chinese speakers when the SOA is as short as 43ms (Perfetti & Tan 1998). However, at the end of the second term, the orthographic facilitation disappeared. Instead, there were 66ms semantic facilitation and a marginal phonological facilitation (27ms) at the end of the second term, which is similar to the longer SOAs of native speakers (Perfetti & Tan 1998). The result showed that the first term learners were not able to reach the orthographic threshold within the 500ms SOA, but the second term learners could. It was concluded that the learners process Chinese in a threshold

639

Ying Liu, Charles A. Perfetti, and Min Wang

style similar to native Chinese speakers. Besides behavioral measures, neuroimaging measures have also been used to compare Chinese and English reading. Evidence from alphabetic research has converged on identification of a network of brain areas that functions in skilled reading, including a left ventral occipito-temporal pathway for orthographic processing, a left superior temporal and inferior parietal region for phonological processing, and a left inferior frontal gyrus for semantic processing (Fiez & Petersen 1998, Mechelli, Gorno-Tempini & Price 2003, Price 2000). Recently, neuroimaging studies of Chinese have developed a picture of the functional neuroanatomy that is partly convergent and partly divergent with the results of alphabetic studies. For example, the left fusiform gyrus is activated in Chinese reading as it is in alphabetic reading (Chee, Tan & Thiel 1999, Chee, Weekes, Lee, Soon, Schreiber, Hoon & Chee 2000, Tan, Liu, Perfetti, Spinks, Fox & Gao 2001, Tan, Spinks, Gao, Liu, Perfetti, Xiong, Stofer, Pu, Liu & Fox 2000). However, Chinese shows additional right hemisphere activation in occipital and fusiform regions (Tan, Laird, Li & Fox 2005, Tan et al. 2001, Tan et al. 2000), a result also seen in an ERP study by Liu & Perfetti (2003). In addition, in frontal areas, Chinese shows less activation in the left inferior frontal gyrus and more activation in the left middle frontal gyrus at BA 9 (Siok, Jin, Fletcher & Tan 2003, Siok, Perfetti, Jin & Tan 2004, Tan et al. 2001, Tan, Spinks, Feng, Siok, Perfetti, Xiong, Fox & Gao 2003, Tan et al. 2000). Even though behavioral experiments can provide valuable information on the reading process, they are indirect measures of brain functions. Functional Magnetic Resonance Imaging (IMRI) and Positron Emission Tomography (PET) are more direct measures, but they cannot separate cognitive processes that happen within tens or hundreds of milliseconds. Event Related Potential (ERP), on the other hand, provides millisecond level neuronal activity data during a cognitive task. ERP recordings produce characteristic voltage shifts (components) that have been associated with reading processes, including the N200, which is sensitive to orthographic and phonological level processing (Kramer & Donchin 1987), and the N400 (N450) which is sensitive to both phonology and meaning (Kutas & Hillyard 1980, Rugg 1984). Bentin et al. (1999) identified a fuller range of time points for reading process associated with inferred brain regions. ERP results also confirmed the ERP components, such as N200 (negative at occipital and positive at frontal electrodes) and N400, which had very similiar temporal features in Chinese reading (Liu, Perfetti & Hart 2003, Valdes-Sosa, Gonzalez, Liu & Zhang 1993). Recent development of high spatial density ERP recording (more than 64 channels) made it possible to more accurately locate the cortex source for the scalp recorded electrical signal. Liu & Perfetti (2003) used 128 channel ERP to compare the reading of

640

Visual Analysis and Lexical Access of Chinese Characters

Chinese and English by Chinese-English bilinguals. The results showed a left to right occipital shift during 100-200ms when Chinese-English bilinguals read Chinese, but not English. It was also found that the temporally similar N400 components of Chinese and English have different sources in the cortex. In order to explore temporal and spatial brain activity during Chinese and English word processing by CSL learners, we adopted the design of the above study and applied it to learners. High density ERP was recorded to compare the reading of Chinese and English at two levels of reading proficiency. The task was delayed naming, a task that (a) allows the examination of a single word reading event and (b) requires a specific reading process, namely the preparation of a spoken word form for overt reading. This task assures that orthographic and phonological processes are engaged and may be detected in the 1000ms time window prior to the actual naming which may cause artifacts in the ERP signal. Furthermore, if ERP records are sensitive to word reading, then we should see a specific ERP indicator of a well-established word processing variable such as word frequency for both Chinese and English. By temporally and spatially comparing the ERP indicators of learners with the results from native Chinese speakers in the literature, we can obtain evidence on whether learners read Chinese in a cascade or threshold style.

2. Methods and materials 2.1 Participants

Twenty four undergraduate students (14 male and 10 female), enrolled in an elementary Chinese class at the University of Pittsburgh, participated in the experiment at the end of their first term (12-15 weeks learning, 12 hours a week). The age of the subjects ranged from 19 to 28. Twenty-two subjects had English as their native language and the other three had alphabetic writing system languages (Tai and Vietnamese). None of the subjects had been formally exposed to any Chinese environment before taking the class. All subjects had normal or corrected to normal vision and were free of medication within one week before the experiment and had no history of neurological diseases. They were paid for their participation. 2.2 Stimuli The Chinese-corpus based character frequency (Li & Liu 1988) does not apply to these learners very well, although they tend to start from more commonly used characters as native Chinese speakers do. Instead, we created a computerized curriculum file based on the textbook (Barnes, unpublished manuscript) and tallied the number of appearances

641

Ying Liu, Charles A. Perfetti, and Min Wang

for each character by a computer program. In total, there were 261 characters in the first term curriculum occurring from 3 to 287 times. Any character that appeared fewer than three times did not enter further material selection. The teaching method used at University of Pittsburgh discouraged any additional study of Chinese beyond the textbook, which enabled our curriculum-based frequency to serve as a corpus-based frequency and provide a very good estimate of the character familiarity level. Furthermore, the subjective familiarity assessed by the same group of subjects was highly correlated with the curriculum-based frequency (Wang, Perfetti & Liu 2003). Thus in the present study, we used this curriculum-based frequency to select Chinese materials. Four experimental conditions were defined by language and frequency: high frequency Chinese characters (43.35/6248 curriculum), low frequency Chinese characters (9.675/6248 curriculum), high (136.1/million) and low frequency (1.2/million) English words (Kucera & Francis 1967). The radical and stroke numbers were matched between high and low frequency Chinese characters. Word length of high and low frequency English words were also matched. Thirty-two experimental stimuli and 8 fillers of the same type appeared in each condition. One Chinese and one English block were presented, each randomly mixed with high and low frequency characters or words.

2.3 Procedure
Participants named 20% of the stimuli in all conditions, with the English and Chinese blocks counter-balanced across subjects. Each trial began with a waiting signal (-***-) that remained in the center of the screen until the subject initiated the trial. Subjects were told not to move their eyes or blink them once a trial had begun, but were encouraged to blink between trials. When the subject pressed the space bar, a fixation "+" appeared on the screen for 500ms, followed by the stimulus, which was exposed for 1500ms. Subjects were told to have the pronunciation of the word in mind and be ready to produce when the naming signal appeared. In 20% of the trials (fillers), a naming signal appeared 1500ms after the onset of the stimulus, followed in two seconds by the waiting signal of the next trial. For the other 80% of the trials (experimental stimuli), a two-second blank interval occurred instead of a naming signal indicating naming was not required.

2.4 EEG recording and averaging

A 15-inch CRT monitor working at 60Hz refresh rate presented the stimuli. A 128 Channel Geodesic Sensor Net (Electrical Geodesies Incorporated, Oregon, USA) recorded the EEG data. All impedances were kept below 40KQ (Ferree, Luu, Russell &

642

Visual Analysis and Lexical Access of Chinese Characters

Tucker 2001). A vertex reference was used in the recording and the data were recomputed off-line against the average reference. Six eye channels allowed rejection of trials with eye movements and blinks. The signals were recorded at 500 Hz. The hardware filter was between 0.1 and 200 Hz. ERPs were averaged off-line over the experimental trials in each condition after the elimination of artifacts followed by a baseline correction and a 30 Hz software low pass filter.

2.5 Data analysis

Twenty subjects provided data for ERP analysis (data from four subjects with a high percentage of trials that contained artifact were rejected). The grand averaged ERP waveforms were calculated for each condition, based on the recordings of experimental trials that contained no artifact, from target onset to 1000ms post-onset. Temporal Principle Component Analysis (PCA) was used to analyze the data instead of traditional mean amplitudes. As a data driven method, PCA extracts a small number of uncorrelated components from a large number of variables, which in temporal PCA are time points. Each PCA component has a factor loading on each time point which is more objective than the human defined 0/1 weighted time window in the mean amplitude method. Simulations have shown that PCA can effectively decompose ERP signal into latent components (Chapman & McCrary 1995, van Boxtel 1998) Our temporal PCA was carried out on subject averages, based on 100 10-ms time samples. Input for the PCA was a data matrix of 10,320 observations (129 electrodes, 20 subjects and 4 stimuli types) by 100 time samples. The computation used correlation matrix with Varimax rotation (Picton, Bentin, Berg, Donchin, Hillyard, Johnson, Miller, Ritter, Ruchkin, Rugg & Taylor 2000). PCA scores were used as dependent measures in ANOVAs to test the effects of experimental manipulations.

3. Results
Grand averaged outputs from 13 electrodes (F3, Fz, F4, C3, Cz, C4, P3, Pz, P4, T7, T8, 01, and 02 in 10-20 system) are shown in Figure 1 in microvolts. By visually inspecting the waveforms, three large shifts at Cz can be seen: a 100ms negative shift, a 200ms positive shift, and a 400ms negative shift. Waveforms at the frontal electrodes have similar polarity as Cz, but the parietal and occipital electrodes are reversed in polarity to Cz at some time points. The most salient reversion is that at 200ms, there are negative shifts at both 01 and 02.

643

Ying Liu, Charles A. Perfetti, and Min Wang

Figure 1: Term 1 Grand Averaged Waveforms

3.1 PCA components

The PCA extracted a small number of uncorrected components from the 100 variables corresponding to 100 10-ms ERP time samples. Five components had eigenvalues larger than 1, explaining 95.5% of the total variance. The eigenvalues of these five components are shown in Figure 2 and the component loadings at each time sample are shown in Figure 3 (note that "component" here does not refer to a voltage shift but to a PCA component). Component 1 (46.9% explained variance), a slow wave component, rises slowly from 200ms to its maximum at 1000ms. This slow-wave component is widely found in PCA on ERP, sometimes a result of the baseline correction and autocorrelated nature of ERP data (Wastell 1981). Component 2 (20.3% explained variance) rises from 150ms with a peak loading at 190ms. The loading lasts until 350ms. The average component scores are positive at all frontal and central electrodes, and negative for all parietal and occipital electrodes. In

644

Visual Analysis and Lexical Access of Chinese Characters

the ERP waveform, a 150-300ms parietal and occipital negative shift and a frontal and central positive shift can be observed (Figure 1). Component 3 (17.2% explained variance) is a very early component. It starts from the onset of stimuli and the component loading turns lower than .5 at 150ms. Component 4 (9.6% explained variance) peaks at 450ms. It lasts from 350ms to 600ms and is negative at most electrodes. Component 5 (1.5% explained variance) peaks at 130ms. Of the five components above, two and four respectively correspond to N200/P200 and N400 based on their latency and shape. These two components fit our temporal window of interest and entered further analysis.

-Eigenvalue

Component Component Component Component Component 12 3 4 5

Figure 2: Term 1 PC A Components Eigenvalues

1 0.8 0.6
rH

a
_ ;

0.4 0.2 0 -0.2 -0.4

L> *
y

'""- -.
-r--"--"r---- -i
_

Component 1 "Component 2 Component 3 - Component 4 800 Component 5

03 O

200

400

600

Figure 3: Term 1 PC A Component Loadings

3.2 ANOVA
We carried out two ANOVAs, one testing the PCA scores for three medial electrodes (Fz, Cz, and Pz) and one for ten lateral electrodes (F3, F4, C3, C4, P3, P4, T7,

645

Ying Liu, Charles A. Perfetti, and Min Wang

T8, 01, and 02). Each was a repeated-measure ANOVA with language (Chinese and English), frequency (high and low), and site (frontal, central, and parietal) as factors. The ANOVA for the lateral electrodes added occipital and temporal to the site factor and hemisphere as the fourth factor. The Greenhouse & Geisser (1959) correction was applied when the sphericity assumption was not satisfied. The Epsilon value is reported here only when adjustment of freedom was performed. Post hoc t-tests between conditions (without adjustment for multiple comparisons) were carried out (t values not presented here) when the overall ANOVA showed reliable condition effects. Both N200/P200 and N400 components produced reliable differences among conditions. The component scores are shown for selected electrodes (Figure 4). N200/P200. Both language and frequency effects were observed in this component. The component scores were positive at frontal and central electrodes and negative at posterior electrodes (medial site effect, F(2,38)=21.60, p<.001, MSE=1.568, Epsilon=.623; lateral site effect, F(4,76)=23.38, p_<001, MSE=2.417, Epsilon=.515). For Chinese stimuli, the P200 component scores were more positive at frontal electrodes, and the N200 scores were more negative at parietal and right occipital electrodes than for English (medial language x site, F(2,38)=6.46, p_<.01, MSE=249, Epsilon=723; lateral language x site, F(4,76)=5.674, p<.01, MSE=390, Epsilon=495). The P200 of Chinese high frequency characters were more positive than low frequency characters at central frontal electrodes (medial frequency x site, F(2,38)=5.78, p<.01, MSE=.096), and the P200 of English high frequency words were more positive than low frequency words at left central electrode (lateral frequency x site, F(4,76)=4.28, p<.05, MSE=.215, Epsilon =.569). N400. English were significantly more negative than Chinese at medial frontal, central and parietal electrodes (medial language, F(l,19)=31.20, p<.001, MSE=.532), right frontal and right central electrodes (lateral language, F( 1,19)= 15.402, p<.01, MSE=.261). Compared with high frequency English words, low frequency English words elicited more negative N400 at a left frontal electrode (lateral frequency, F(l,19) =8.68,p<01,MSE=181).

1.5 -|

o.5;
-0.5 -1

in-

BChinese high Chinese low

o..IsSsiL_Jr

-1.5 N200 Fz

W^
N200 02

I j

English high ^English low

N400 Cz

Figure 4: Term 1 Component Scores of N200 at Fz and 02, N400 at Cz

646

Visual Analysis and Lexical Access of Chinese Characters

4. Experiment 2 4.1 Participants

Seventeen undergraduate students (nine male and eight female), enrolled in the same elementary Chinese class as in Experiment 1, participated in the experiment at the end of their second term (27-30 weeks learning including first term, 12 hours a week). The age of the subjects ranged from 19 to 28. All subjects had normal or corrected to normal vision and were free of medication within one week before experiment and had no history of neurological diseases. They were paid for their participation. Experiment procedures and materials were the same as those used in experiment 1.

4.2 Results
Fourteen subjects provided data for further analysis after the artifact rejection. Grand averaged waveform is shown in Figure 5. The amplitude shifts in the waveforms are similar to experiment 1. PCA extracted a small number of uncorrelated components from the 100 time sample variables. Six components had eigenvalues larger than 1, explaining 95.7% of the total variance. The eigenvalues of these six components are shown in Figure 6. It can be seen that there is a sharp decrease of eigenvalue from the fifth to the sixth component. The first five component loadings are shown in Figure 7 and their features are described below. Component 1 (48.2% explained variance), a slow wave component, rises slowly from 450ms to its maximum at 1000ms. Component 2 (28.5% explained variance) peaks at 430ms, rises from 270 through 600ms. Component 3 (18.9% explained variance) rises from 150ms with a peak loading at 230ms. The loading lasts until 330ms. The average component scores are positive at all frontal and central electrodes, and negative at all parietal and occipital electrodes. There are a 15 0-3 00ms parietal and occipital negative shift and a frontal and central positive shift in the ERP waveform (Figure 5). This component is similar to the N200/P200 in experiment 1. Component 4 (11.6% explained variance) is a very early component. It starts to rise from the onset of stimuli and the component loading turns lower than .5 at 90ms. Component5 (11.1% explained variance) peaks at 120ms. Among the five components above, N200/P200 (component 3) and N400 (component 2) fit our temporal window of interest and entered further analysis.

647

Ying Liu, Charles A. Perfetti, and Min Wang

yx****** [yw jj^r^^s

200

400

600

800

- Chinese__Hgh - &iglish_Hgh

- Chinese_Low Biglish_Low

Figure 5: Term 2 Grand Averaged Waveforms

Eigenvalue

#
OF

Figure 6: Term 2 PC A Components Eigenvalues

Visual Analysis and Lexical Access of Chinese Characters

1
0.8 0.6 0.4 0.2 0 -0.2

Conponent Component -Component -Component Component

1 2 3 4 5

- ',/'

\/

*<""' *'

03

60 C
TH

-O
<ao

200

400

600 800

ms

Figure 7: Term 2 PC A Component Loadings

4.3 ANOVA
The ANOVA procedure was the same as in Experiment 1. The slow wave, N200/P200 and N400 components produced reliable differences among conditions. PCA scores of N200/P200 and N400 are shown in Figure 8 for selected electrodes. N200/P200. The P200 was much more positive for Chinese than English at left and middle frontal (lateral language x lobe, F(4,52)=7.025, p<.01, MSE=.593, Epsilon=.414). N400. There was significant language effect. The N400 was significantly more negative for English than Chinese at Fz, Cz, and Pz, F4, C4, P4 (medial site language effect, F(l,13)=27.619, p<.01, MSE=523; lateral site language effect, F(l,13)=5.309, p<.05, MSE=.681; lateral site language x hemisphere interaction, F(l,13)=4.943, p<.05, MSE=.957). There was also near significant frequency effect in medial site (F(l,13)= 4.401, p=.056, MSE=.192) which was due to the low frequency English words being more negative than high frequency English words at Cz (p<.05 with no adjustment).

S Chinese high Chinese low English high English low N200 Fz N200 02 N400 Cz

Figure 8: Term 2 Component Scores of N200 at Fz and 02, N400 at Cz

649

Ying Liu, Charles A. Perfetti, and Min Wang

5. Discussion
We found that at the first term, both N200/P200 and N400 components showed language difference between Chinese and English stimuli. N200/P200 was larger for Chinese than English at both frontal and occipital electrodes. Conversely, N400 was larger for English than Chinese at frontal and central electrodes. At the second term, P200 was significantly more positive for Chinese than English at left and middle frontal electrodes, and N400 was significantly more negative for English than Chinese at right frontal, central, and parietal electrodes, but N200 did not show language difference at occipital electrodes. Chinese familiarity effect was only observed at the first term: A larger P200 for high familiarity words at frontal electrodes. There was no Chinese familiarity effect at the second term. N200/P200 has been found by Liu & Perfetti (2003) as a component sensitive to orthographical processing. Its amplitude was reduced when the target Chinese character was preceded by an orthographically similar prime character. The reduction has been attributed to the pre-activation of orthographic form of the target. In another study by Liu & Perfetti (2003), Chinese-English bilinguals performed a delayed naming task with ERP collected. The amplitude of N200 was larger for English (the second language) than Chinese (the first language). The N200/P200 in above studies had a similar distribution as the present study: positive at frontal and central electrodes, and negative at occipital electrodes. It has also been found in other paradigms that two components between 180 and 300ms were sensitive to object detection (Potts & Tucker 2001). They were named P2a and N2b respectively. N2b preceded the P200 by nearly 30ms in the object detection task. The P2a and N2b interacted between orbitofrontal cortical areas of salience representation and posterior cortical areas of stimulus feature representations. The N200/P200 in the present study has some functional relations with the P2a/N2b component, because word processing does require feature detection of visual object and meaningful words are salient to the readers. The left occipital/fusiform region is a visual word form analysis region widely found in English reading by fMRI and other methods (Cohen, Lehericy, Chochon, Lemer, Rivaud & Dehaene 2002, Dehaene, Naccache, Cohen, Bihan, Mangin, Poline et al. 2001, Nobre, Allison & McCarthy 1998). In Chinese reading, not only the left but also the right occipital/fusiform regions are involved. Comparing Chinese and English directly, Tan et al. (2001, 2003) found that Chinese materials elicited stronger activation at right occipital/fusiform (visual analysis) and left middle frontal regions (lexical access) than English. In the present study, the N200 component was significantly larger at right occipital and the P200 was significantly larger at left and middle frontal electrodes

650

Visual Analysis and Lexical Access of Chinese Characters

for Chinese than English, which well matched the fMRI findings (Tan, Spinks, Eden, Perfetti & Siok 2005). Because the N200 and P200 were highly correlated temporally, our temporal PCA did not separate them into two components. However, the difference between first and second terms unveiled the functional difference of their underlying processing. For the first term learners, Chinese had larger amplitude at N200/P200 which indicated more visual processing (occipital) and lexical access (frontal and central) effort were needed for processing Chinese because of high demand of visual analysis and lexical retrieval. For the second term learners, the lexical access difference at frontal was retained, which indicated that reading Chinese was still a demanding task at the lexical access level, but visual analysis had been rather quickly accommodated. This result was consistent with the behavioral finding that learners can quickly learn the character structure (Wang, Perfetti & Liu 2003). The spatial and functional separation of the N200/P200 component also provides further evidence that the learners might process Chinese characters in threshold style at the second term. Liu et al. (2006) found significant orthographic priming in the first term learners. In that study, the orthographically similar pairs either shared some strokes that appear at the same spatial position, or one radical at the same position. Both types of similarities facilitated the target identification at visual analysis level. The occipital N200 reflects the fact that the visual processing of Chinese and English had different speed and strength at the first term. However, the visual analysis skill on Chinese characters was significantly improved at the second term so that no significant language difference was found at occipital electrodes, even though very likely there is some kind of fine grain difference not measured by ERP. This finding is consistent with the behavioral result that no orthographic priming was found at second term (Liu, Wang & Perfetti 2006). The observed separation of visual analysis and lexical access at 200ms supports the assumption that accompanying the increase of reading proficiency, the sinogram is lexically accessed after the orthographic processing has been completed at visual analysis level, which indicates a threshold style processing. The first term Chinese familiarity effect suggests that the N200/P200 is an informative indicator of orthographic processing speed. Larger P200 at frontal electrodes reflects the earlier lexical access of high frequency characters in the first term. However, at the second term, visual analysis was faster for both familiar and unfamiliar characters which caused earlier lexical access for both and reduced the speed difference between them. Furthermore, because the vocabulary of the subjects was limited, after the lexical access is started, the difficulty of accessing familiar and unfamiliar characters is not very different. There is a possibility that with more experience in Chinese (a larger vocabulary), the frontal P200 might show more lexical access difficulty for low frequency characters

651

Ying Liu, Charles A. Perfetti, and Min Wang

(Liu & Perfetti 2003). There was also a P200 familiarity effect (familiar > unfamiliar) at left central electrode for English, but only in the first term. So it is possible that this component shows some similarity between Chinese and English processing. However, the underlying source on the cortex might be different because the observed difference is more left lateralized and inferior for English which is consistent with fMRI findings (Tan, Spinks, Eden, Perfetti & Siok 2005). It is hard to explain why the English familiarity is only observed among the first term students. One possibility is that it is a rather localized effect and the p value of ANOVA test is larger than .01, the 14 subjects in the second term might not be able to provide enough statistical power to observe this effect. N400 has been widely used as an indicator of semantic and phonological processing (Kutas & Hillyard 1980, Rugg 1984). The N400 in the present two experiments was only observed in English materials, but not for Chinese characters. It showed that the semantic processing on Chinese of the learners was too slow to be in the measuring range of N400. Even though Liu et al. (2006) found semantic priming with a 500ms SOA, still it is much longer than the 85ms SOA used for native speakers (Perfetti & Tan 1998). Since ERP components require the neuronal activity to reach a certain level to be measured, not being able to observe a strong N400 component indicates that the semantic and phonological activations are still weak even though the orthographical threshold has been reached within 200ms. In summary, we found that for Chinese learners with alphabetic writing system background, processing Chinese orthography can be separated into a visual analysis stage and a lexical access stage. The visual analysis stage focuses on stroke and radical processing which can be learned quite fast and accomplished within 200ms from the character onset. After the completion of visual analysis, lexical access follows immediately. However, the retrieval of semantic and phonological information is still slow and difficult to be observed via ERP. The present study supports the contention that learners read Chinese characters similar to native Chinese speakers after two terms of learning. The learners might start out their character reading in a way similar to English reading. However, because the alphabetic way does not work well on a non-alphabetic system, and to fit the new learned writing system, the brain non-coincidentally develops a processing method that is used by native Chinese speakers: threshold-style processing. Brain regions used by native Chinese speakers are also recruited to accomplish the task. We propose an accommodation hypothesis whereby the brain of alphabetic users accommodates to the Chinese writing system during the learning process.

652

Visual Analysis and Lexical Access of Chinese Characters

References
Baron, J., and C. Strawson. 1976. Use of orthographic and word-specific knowledge in reading words aloud. Journal of Experimental Psychology: Human Perception and Performance 2.3:386-393. Bentin, S., Y. Mouchetant-Rostaing, M. H. Giard, J. F. Echallier, and J. Pernier. 1999. ERP manifestations of processing printed words at different psycholinguistic levels: time course and scalp distribution. Journal of Cognitive Neuroscience 11.3: 235-260. Chapman, R. M., and J. W. McCrary. 1995. EP component identification and measurement by principal components analysis. Brain and Cognition 27.3:288-310. Chee, M. W., B. Weekes, K. M. Lee, C. S. Soon, A. Schreiber, J. J. Hoon et al. 2000. Overlap and dissociation of semantic processing of Chinese characters, English words, and pictures: evidence from-fMRI. Neuroimage 12.4:392-403. Chee, M. W., E. W. Tan, and T. Thiel. 1999. Mandarin and English single word processing studied with functional magnetic resonance imaging. Journal of Neuroscience 19.8:3050-3056. Chen, M. J., Y. F. Yung, and T. W. Ng. 1988. The effect of context on the perception of Chinese characters. Cognitive Aspects of the Chinese Language, ed. by M. Liu, H.-C. Chen, and M. J. Chen, 27-39. Hong Kong: Asian Research Service. Chua, F. K. 1999. Phonological recoding in Chinese logograph recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition 25.4:876-891. Cohen, L., S. Lehericy, F. Chochon, C. Lemer, S. Rivaud, and S. Dehaene. 2002. Language-specific tuning of visual cortex? Functional properties of the Visual Word Form Area. Brain 125.5:1054-69. Coltheart, M., K. Rastle, C. Perry, R. Langdon, and J. Ziegler. 2001. DRC: a dual route cascaded model of visual word recognition and reading aloud. Psychological Review 108.1:204-256. Dehaene, S., L. Naccache, L. Cohen, D. L. Bihan, J. F. Mangin, J. B. Poline et al. 2001. Cerebral mechanisms of word masking and unconscious repetition priming. Nature Neuroscience A.1:152-15%. Ferree, T. C, P. Luu, G. S. Russell, and D. M. Tucker. 2001. Scalp electrode impedance, infection risk, and EEG data quality. Journal of Clinical Neurophysiology 112.3: 536-544. Fiez, J. A., and S. E. Petersen. 1998. Neuroimaging studies of word reading. Proceedings of the National Academy of Sciences of the United States of America 95, 914-921. Greenhouse, S. W., and S. Geisser. 1959. On methods in the analysis of profile data. Psychometrika 24.2:95-112.

653

Ying Liu, Charles A. Perfetti, and Min Wang

Hoosain, R., and C. E. Osgood. 1983. Information processing times for English and Chinese words. Perception andPsychophysics 34:573-577. Kramer, A. F., and E. Donchin. 1987. Brain potentials as indices of orthographic and phonological interaction during word matching. Journal of Experimental Psychology: Learning, Memory, and Cognition 13.1:76-86. Kucera, Henry, and W. N. Francis. 1967. Computational Analysis of Present-Day American English. Providence: Brown University Press. Kutas, M., and S. A. Hillyard. 1980. Reading senseless sentences: brain potentials reflect semantic incongruity. Science 207:203-205. Li, G., and R. Liu. (eds.) 1988. A Dictionary of Chinese Character Information. Beijing: Science Press. Liu, Y., and C. A. Perfetti. 2003. The time course of brain activity in reading English and Chinese: an ERP study of Chinese bilinguals. Human Brain Mapping 18.3: 167-175. Liu, Y., C. A. Perfetti, and L. Hart. 2003. ERP evidence for the time course of graphic, phonological and semantic information in Chinese meaning and pronunciation decisions. Journal of Experimental Psychology: Learning, Memory, and Cognition 29.6:1231-1247. Liu, Y., M. Wang, and C. A. Perfetti. 2006. Threshold-style processing of Chinese characters for adult second language learners. Memory and Cognition (in press). Mechelli, A., M. L. Gorno-Tempini, and C. J. Price. 2003. Neuroimaging studies of word and pseudoword reading: consistencies, inconsistencies, and limitations. Journal of Cognitive Neuroscience 15.2:260-271. Nobre, A. C, T. Allison, and G. McCarthy. 1998. Modulation of human extrastriate visual processing by selective attention to colors and words. Brain 121.7:1357-1368. Peng, D., Y. P. Li, and H. Yang. 1997. Orthographic processing in the identification of Chinese characters. Cognitive Processing of Chinese and Related Asian Languages, ed. by H. C. Chen, 86-108. Hong Kong: The Chinese University Press. Perfetti, C. A., and L. H. Tan. 1998. The time course of graphic, phonological, and semantic activation in Chinese character identification. Journal of Experimental Psychology: Learning, Memory, and Cognition 24.1:101-118. Perfetti, C. A., and S. Zhang. 1995. Very early phonological activation in Chinese reading. Journal of Experimental Psychology: Learning, Memory, and Cognition 21.1:24-33. Perfetti, C. A., S. Zhang, and I. Berent. 1992. Reading in English and Chinese: evidence for a "universal" phonological principle. Orthography, Phonology, Morphology, and Meaning, ed. by R. Frost, 227-248. Amsterdam: North-Holland. Perfetti, C. A., Y. Liu, and L. H. Tan. 2005. The lexical constituency model: some

654

Visual Analysis and Lexical Access of Chinese Characters

implications of research on Chinese for general theories of reading. Psychological Review 112.1:43-59. Picton, T. W., S. Bentin, P. Berg, E. Donchin, S. A. Hillyard, R. Johnson Jr., G. A. Miller, W. Ritter, D. S. Ruchkin, M. D. Ruggs, and M. J. Taylor. 2000. Guidelines for using human event-related potentials to study cognition: recording standards and publication criteria. Psychophysiology 37.2:127-152. Potts, G. F., and D. M. Tucker. 2001. Frontal evaluation and posterior representation in target detection. Cognitive Brain Research 11.1:147-156. Price, C. J. 2000. The anatomy of language: contributions from functional neuroimaging. Journal of Anatomy 197.3:335-360. Rugg, M. D. 1984. Event-related potentials in phonological matching tasks. Brain and Language 23.2:225-240. Shu, H., and R. C. Anderson. 1999. Learning to read Chinese: the development of metalinguistic awareness. Reading Chinese Script: A Cognitive Analysis, ed. by Jian Wang and Albrecht W. Inhoff et al., 1-18. Mahwah: Lawrence Erlbaum Associates. Siok, W. T., C. A.' Perfetti, Z. Jin, and L. H. Tan. 2004. Biological abnormality of impaired reading is constrained by culture. Nature 431:71-76. Siok, W. T., Z. Jin, P. Fletcher, and L. H. Tan. 2003. Distinct brain regions associated with syllable and phoneme. Human Brain Mapping 18.3:201-207. ' Taft, M., X. Zhu, and D. Peng. 1999. Positional specificity of radicals in Chinese character recognition. Journal of Memory and Language 40.4:498-519. Tan, L. H., A. Laird, K. Li, and P. T. Fox. 2005. Neuroanatomical correlates of phonological processing of Chinese characters and alphabetic words: a meta-analysis. Human Brain Mapping 25.1:83-91. Tan, L. H., H.-L. Liu, C. A. Perfetti, J. A. Spinks, P. T. Fox, and J.-H. Gao. 2001. The neural system underlying Chinese logograph reading. Neurolmage 13.5:836-846. Tan, L. PL, J. A. Spinks, C. M. Feng, W. T. Siok, C. A. Perfetti, J. Xiong, P. T. Fox, and J. H. Gao. 2003. Neural systems of second language reading are shaped by native language. Human Brain Mapping 18.3:158-166. Tan, L. FL, J. A. Spinks, G. F. Eden, C. A. Perfetti, and W. T. Siok. 2005. Reading depends on writing, in Chinese. Proceedings of the National Academy of Sciences of the United States of America 102, 8781-8785. Tan, L. H., J. A. Spinks, J. H. Gao, H. L. Liu, C. A. Perfetti, and J. Xiong et al. 2000. Brain activation in the processing of Chinese characters and words: a functional MRI study. Human Brain Mapping 10.1:16-27. Tzeng, O. J. L., and D. L. Hung. 1978. Reading Chinese characters: some basic experimental studies. Acta Psychologica Taiwanica 20:45-49.

655

Ying Liu, Charles A. Perfetti, and Min Wang

Valdes-Sosa, M., A. Gonzalez, X. Liu, X. L. Zhang et al. 1993. Brain potentials in a phonological matching task using Chinese characters. Neuropsychologia 31.8: 853-864. van Boxtel, G. J. M. 1998. Computational and statistical methods for analyzing event-related potential data. Behavior Research Methods, Instruments, and Computers 30.1:87-102. Wang, M., C. A. Perfetti, and Y. Liu. 2003. Alphabetic readers quickly acquire orthographic structure in learning to read Chinese. Scientific Studies of Reading 7.2:183-208. Wang, W. S. Y. 1973. The Chinese language. Scientific American 228:50-60. Wastell, D. G. 1981. On the correlated nature of evoked brain activity: biophysical and statistical considerations. Biological Psychology 13:51-69. Xu, Y., A. Pollatsek, and M. C. Potter. 1999. The activation of phonology during silent Chinese word reading. Journal of Experimental Psychology: Learning, Memory, and Cognition 25.4:838-857. Zhang, S., C. A. Perfetti, and H. Yang. 1999. Whole word, frequency-general phonology in semantic processing of Chinese characters. Journal of Experimental Psychology: Learning, Memory, and Cognition 25.4:858-875. Zhou, X., and W. Marslen-Wilson. 1996. Direct visual access is the only way to access the Chinese mental lexicon. Proceedings of 18th Annual Conference of the Cognitive Science Society, ed. by G. Cottrell, 714-719. Hillsdale: Lawrence Erlbaum Associates. Zhou, Y. 1978. To what degree are the "phonetics" of present-day Chinese characters still phonetic? Zhongguo Yuwen 1978.3:172-177.

[Received 30 May 2005; revised 24 August 2005; accepted 24 November 2005] Ying Liu Learning Research and Development Center University of Pittsburgh Pittsburgh, PA 15260 USA liuying@pitt.edu

656

Visual Analysis and Lexical Access of Chinese Characters

Mn

Charles A. Perfetti1

aiTmfflW*fflil*{u (Event Related Potential) W^THS**^^

1000 S^ftffiJl^^fHiSHtM/iffllJMtlSSyW^^If if ERP f&l$j# frtiT (Principle Component Analysis) JH&HTHIft 200 ^iEfcffS^Wtia u# - SMKE 400 j^fcSKffiaiinx^^ a^Snxjtft$at

@f!Sftff^^i/ffpfi*asi5[^ ERP mmm&m-mim. - m^tm mstfrffi\m$Mmmm u^=mm%mm&ffim5:T mm m *xmmmmimmmw^mm$:miimM &# f^^^j^ mmmftmmmmmmfrMmm&*xmwmm*minT.m*5:mm% mm' mmwm^m mmm ^>>c m=.mm wisum

657

15
BRINGING READING RESEARCH TO LIFE
DECODING, VOCABULARY, AND COMPREHENSION
The Golden Triangle of Reading Skill

Charles Perfetti

Edited by MARGARET G. McKEOWN LINDA KUCAN

The triangle has presented a strong symbol through the ages, representing ideas in religion, astrology, and sexual identity, as well as more specialized denotations. For Example, in mathematics it is the symbol for a small difference, and on the dashboards of modern cars, the symbol for warning and emergency. Naturally, reading has its triangle as well: Triangle models of word identification represent the three constituents of written-word knowledgegraphic form (spelling), phonological form (pronunciation), and semantics (meaning). Triangle models generally denote the class of connectionist reading models (Seidenberg & McClelland, 1989) but, more specifically, versions of these models that exert a semantic influence on word identification (Harm & Seidenberg> 1999; Plaut, McClelland, Seidenberg, & Patterson, 1996).

2010 The Guilford Press A Division of Guilford Publications, Inc. 72 Spring Street, New York, NY 10012 www.guilford.com

THE DVC TRIANGLE

Given the prolific spread of triangle imageryand despite the established status of the triangle as a representation of written word identificationI think proposing a new triangle is easily justified. The DVC triangle is the interconnected set of cognitive-linguistic components that make up general reading skill: decoding, vocabulary, and compre291

THE GUILFORD PRESS New York London

292 BRINGING READING RESEARCH TO LIFE

Decoding,Vocabulary, and Comprehension Vocabulary

293

hension. Each of these is a complex constituent rather than an elementary unit, so each has its own constituents. For example, the decoding constituent consists of orthographic and phonological knowledge; comprehension includes a wide range of basic sentence, extended text, and general-knowledge-based inference procedures; vocabulary includes both a quantitative (number of words) component and a qualitative (specific word knowledge) component. The complexity of the triangle is a matter of grain sizerelatively fine or relatively coarse. For general descriptive purposes, the coarse grain size shown in Figure 15.1 is about right. It expresses the interconnections among decoding, vocabulary knowledge, and comprehension that are central to skilled reading.1 The DVC model in Figure 15.1 is primarily a heuristic, suggesting a way to conceptualize reading skill (it is the triangle itself), while also illustrating causal relations among three critical constituents of reading skill (the sides of the triangle are directional arrows). Decoding leads to a word's meaning, but not to comprehension beyond the word directly. Comprehension affects vocabulary (word meanings are learned from context) but not decoding directly. And both the decoding-vocabulary and the vocabulary-comprehension relations are reciprocally causal. DecodingVocabulary Decoding affects vocabulary directly, because successful decoding events (1) retrieve meanings of familiar words, thus strengthening form-meaning connections, and (2) establish context-dependent links between unfamiliar words and meaning-bearing contexts. Vocabulary (knowledge of the meaning of a word) affects decoding because decoding a word whose meaning is known strengthens the connection between the word's orthographic form (its spelling) and its meaning. This process helps establish a word-specific representation, which is especially helpful for words with exceptional or irregular spelling-pronunciation mappings and theoretically helpful for all words, to some
1

Decoding

Comprehension

FIGURE 15.1. The DVC reading skill triangle. Abilities in decoding, vocabulary, and comprehension combine to produce general reading skill. Because the three constituents are interconnected, limitations in any one will affect at least one other constituent and will accordingly set a limit on overall skill. For some purposes decoding can be interpreted in its broad sense of word identification; for other purposes it is interpreted in its narrow sense of grapheme-phoneme conversion. (See Footnote 1.) These two would be differentiated in a finer grained triangle model.

extent. The importance of this meaning-to-form support is demonstrated by Nation and Snowling (1998), who found that word identification of comprehension-impaired readers was especially slow in identifying words that, theoretically, depend on frequent exposures (exception words). Simply put, as children decode words, they strengthen their vocabulary knowledge; and as children retrieve their knowledge of a word's meaning while decoding it, they strengthen the identifiability of that word. Vocabulary-Comprehension Comprehension is obviously dependent on knowing the meanings of words being read, and the DVC triangle represents this causal direction. At the moment a reader encounters a text, the ability to access the meaning of the word, as it applies in the context of this particular text, is critical. At the same time, achieving some comprehension from a segment of text that contains an unknown word also can cause the reader to learn something about the meaning of that word. Thus the causation runs both ways between word meaning and comprehension. That multiple causes can underlie the general correlation between vocabulary knowledge and comprehension has long been recognized (e.g., Anderson & Freebody, 1981; Beck, McKeown, & Omanson, 1987). Evidence for a reciprocal causation across measurement points, however, has come only recently from a longitudinal study by Wagner (2005).

An important point in interpreting the triangle and in the discussion that follows is that a strict definition of decoding is the conversion of letter strings to phoneme strings. Word identification includes both decoding processes in this narrow sense and the retrieval of word-specific representations that uses knowledge about a word's spelling to identify it. In English, this word-specific process is needed for words whose spellings are exceptions to the dominant grapheme-to-phoneme mappings (Coltheart, Rastle, Perry, Langdon, &c Ziegler, 2001) A finer grained triangle model would be needed to distinguish these two identification processes.

294

BRINGING READING RESEARCH TO LIFE

Decoding,Vocabulary, and Comprehension

295

Decoding-Comprehension? The DVC triangle does not show causal arrows between decoding and comprehension in either direction. This is because the effects of decoding on comprehension are mediated by knowing the meaning of the decoded word. The effects of comprehension on decoding are mediated by achieving enough meaning from the text to verify the identity of a decoded word. Note that this assumption rests on the logic of cognitive event sequences in reading and not on correlations of skill assessments. In practical terms, there is a strong causal relation between decoding and comprehension in that fluent or automatic decoding allows more processing resources to be available for comprehension (Perfetti, 1985). And comprehending a text aids word identification, especially for readers of low word reading skill (Perfetti, 1985; Stanovich, 1980). However, at closer distance, these causal effects depend on word meanings being produced by identification. Thus knowledge of word meanings (or vocabulary knowledge) has a pivotal position between word identification and comprehension (Perfetti, Landi, & Oakhill, 2005).

Decoding skill itself supports self-teaching of written-word representations, which allows children to move from a reading process entirely dependent on phonological coding of printed word forms to a process that accesses words quickly based on their orthography (Share, 1995, 1999).

To clarify the omission of decoding-comprehension effects, note

that assessments of decoding do correlate with assessments of comprehension (Perfetti, 1985), but on the present assumption this correlation reflects a causal connection from decoding to comprehension that is mediated by knowledge of word meanings. The decoding-comprehension correlation may also partly reflect their shared influences from outside the triangle (e.g., phonological knowledge, other linguistic knowledge, and general intelligence). Beyond the more obvious implications for individual differences, there is an interesting, less obvious one. Children with weak decoding skills may have to depend more on the vocabulary -> decoding side of the triangle. Indeed, a semantics-to-decoding connection helps to compensate for weak decoding skills (Snowling, Hulme, & Goulandris, 1994). The DVC triangle representation of individual differences approximates that captured by the lexical quality hypothesis (LQHPerfetti, 2007; Perfetti & Hart, 2001). The LQH claims that knowledge about word forms (phonological, orthographic, and morphemic knowledge) affects reading comprehension in both obvious and less obvious ways. The particular DVC of Figure 15.1, by collapsing distinctions between orthographic and phonological knowledge and between word identification and decoding, misses some important details in lexical quality but captures the broad relations. To convert the triangle into a processing scheme, Figure 15.2 shows a linear flow of knowledge of word form and meaning to the processes of word identification and comprehension, with feedback from comprehension back to the word knowledge level. According to the DVC and the LQH, word meanings are central to comprehension and word identification. However, research on comprehension has often ignored vocabulary to focus on other comprehension issues (e.g., inference making, comprehension strategies). Nevertheless, knowledge of word meanings cannot be ignored in accounts of individual differences. Meanwhile, the search for cognitive mechanism differences, as opposed to knowledge differences, has had the effect

INDIVIDUAL DIFFERENCES IN LEXICAL QUALITY The DVC triangle identifies possible differences in reading skill at each point of the triangle and also at the four causal links (two for decod-ing-vocabulary and two for vocabulary-comprehension). These seven possible sources of reading skill variability obviously are not all independent, and all have causal links to knowledge sources outside the triangle. For example, knowledge of word meanings is affected by pre-literacy exposure to vocabulary, which is stunningly variable across demographic categories (Hart & Risley, 1995). But it is simple enough to summarize key relations in reading skill: 1. Skill in reading comprehension will be affected by skill in decoding and skill in vocabulary (which will not be independent, according to the model). 2. Skill in vocabulary will be affected by skill in comprehension and skill in decoding (which will be independent). 3. Skill in decoding (understood as word identification) will be predicted by vocabulary knowledge.

296

BRINGING READING RESEARCH TO LIFE

The Lexical Quality Hypothesis Word Knowledge to Comprehension Many other factors

Decoding, Vocabulary, and Comprehension

297

nections between the two. These characteristics allow word reading and

meaning retrieval to be rapid and relatively automatic given a familiar

printed word. Just as work on comprehension has sometimes ignored word meaning, research on word reading skill has tended to ignore spelling. We are learning, however, that even among "good readers," differences in knowledge of word spellings lead to differences in word reading processes (Andrews, 2008).

Lexical Knowledge (form, form, meaning) V A

Word Identification Skill

Comprehension

Linguistic and Conceptual Systems

THE DVC TRIANGLE IN THE CONTEXT OF THE READING EDUCATION FIELD

It is useful to ground the DVC triangle in two contexts relevant to reading research. The first places the triangle idea in a personal historical context. My approach to reading has been fairly simple, grounded on the idea that reading is a process built on language. The distinctive part of reading, in this commonsense view, is written-word identification: The singular recurring cognitive activity in reading is the identification of words and the retrieval of their meanings. From this it follows that comprehension depends in part on successful word reading and that skill differences in comprehension can arise from skill differences in word reading. This was not an easy thing to acknowledge. I came to reading from training as a psycholinguist. Naturally, I became interested in reading comprehension as a language problem and, along with my graduate students, considered where "the action" was in reading comprehension skill. Syntactic abilities? Getting meaning from sentences and paragraphs? Higher level language knowledge? Despite the inherent interest of these possibilities to a psycholinguist, I thought the first step was to ask about word reading and the role it played in comprehension. These observations on the role of word identification in reading were the core of a theory of comprehension skill that, over 25 years ago, I referred to as verbal efficiency theory (Perfetti, 1985): Word identification, the rapid retrieval of a word's phonology and meaning, was a limiting factor in comprehension. Verbal efficiency theory is the ancestor of the DVC triangle and the LQH and captures some of the same ideas. I came to conclude, however, that verbal efficiency's emphasis on general processesdecoding, phonological processes, retrieval, memory, automaticitydid not quite capture the importance of knowledge. Verbal efficiency implied that skilled reading was about efficient pro-

FIGURE 15.2. Simple schematic of the flow of information from knowledge about word form and meaning to comprehension through word identification. Lexical quality varies within the word-knowledge component and affects word identification and comprehension and is in turn modified by through both word identification (e.g., self-teaching) and comprehension (e.g., acquiring word meanings).

of suggesting that comprehension skill differences and knowledge differences (especially vocabulary) derive from differences in working memory resources (Daneman, 1988; Gathercole & Baddeley, 1989; Seigneuric, Ehrlich, Oakhill, & Yuill, 2000). Cognitive resource differences cannot, however, explain the massive differences observable in exposure to language that builds the vocabulary children are likely to encounter in written texts (Hart & Risley, 1995). Somehow, and this is central to understanding reading comprehension, children both before and after the beginnings of literacy differ greatly in the numbers of words they know in both the sense of familiarity and specific semantic knowledge. (See Reichle and Perfetti, 2003, for a framework that links this distinction to lexical quality in a memory-based approach to recognizing words.) Such word knowledge is not only instrumental to comprehension but also a signature for literacy. Although here I focus on word meaning as pivotal, I need to emphasize that word-form knowledge is also critical in skilled reading. The development of word reading skill depends on the refinement of word representations by adding spelling knowledge to spoken-word representations. This refinement produces representations that increase in precision, knowledge of all the letters of a word, and redundancy, complete letter knowledge and complete phonemic knowledge with con-

298 BRINGING READING RESEARCH TO LIFE cessing mechanisms and that less-skilled reading was about these same mechanisms operating less efficiently. But where did differences in efficiency arise? Although differences in processing capacity provide a possible answer to this question, it did not seem plausible as the whole story. Skill differences were often specific to reading, and they were subject to practice effects. Instead, I concluded that the major source of reading ability is the knowledge a reader has about words, specific lexical representations. Knowledge plus practice that refines the knowledge and makes it more accessible leads to efficient processing. This perspective on the problem led eventually to the LQH (Perfetti, 2007; Perfetti & Hart, 2001). Although the lexical knowledge emphasis was an important refinement of the verbal efficiency hypothesis, this emphasis was already anticipated in Reading Ability (Perfetti, 1985), which has several observations that emphasized knowledge rather than process. For example, I argued in that volume that the retrieval of a lexical representation is high in quality "to the extent that it contains both semantic and phonetic information sufficient to recover its memory location. ... This quality must be retained long enough for subsequent processes to perform their work. Thus a 'name' without meaning and a meaning without a 'name' are both low quality" (Perfetti, 1985, p. 118). The second context for the DVC is that it aligns well with the research of Isabel Beck. Indeed, it is remarkable that the whole of the triangle can be instantiated by the research career of Beck and her colleagues. Next, I consider each point of the triangle in turn. Decoding First, Beck's early work developing code-based reading instruction (Beck & McCaslin, 1978) sits at the decoding corner of the triangle. Her approach to decoding followed the foundational principles of alphabetic reading by directly teaching the correspondences between letters and phonemes. Its distinctive addition to this basic principle was a procedure to support blending, the integration of phonemes so that the child would learn to map letter sequences to phonological words (m-a-t - /maet/) rather than only to isolated phonemes (m -> /m/; a ->/ae/;t->/t/). The basic good sense of this program, as well as its careful optimizing aspects, flew against the wind of the whole-language movement. Much later, with the return of good sense to the teaching of reading, Beck's Making Sense of Phonics (2006) reestablished the value of decoding instruction for a more receptive practice community. Vocabulary

Decoding,Vocabulary, and Comprehension

299

At the vocabulary corner of the triangle, the story is similar to the decoding story. Direct instruction in decoding was not in vogue at the time that Beck developed a direct instruction program, and neither was direct instruction in vocabulary, although for a different reason. Although no one argued against the value of direct instruction of vocabulary, little instruction actually occurred in schools (Scott, jamieson, & Asselin, 2003). People generally assumed that such instruction was of marginal value, because most word learning occurs incidentally through reading and spoken-language experience. For example, Nagy and Herman (1987) estimated that vocabulary might grow by 2,000 to 3,000 words per year over grades 1-12. Thus, at best, direct instruction might help children learn a tiny percentage of the massive English vocabulary needed for academic success. I believe this observation is correct as far as it goes. But it does not take into account some important additional considerations. First are the massive individual differences that are present in vocabulary knowledge in school-age children, especially differences across different socioeconomic classes, that arise well before children enter school (Hart & Risley, 1995). If one could teach 100 words to a child who knows only 2,000 academically useful words (a possible estimate for a child entering first grade from a very low socioeconomic status [SES] background), the result would be a noticeable 5% gain. Even at 4,000 (2.5%) or 8,000 words (1.25%), it is not clear why one would want to dismiss the gains to vocabulary, except for issues of allotting instructional time. Second is the potential for vocabulary "spread." Words related to the meaning of a new word can be used in connection with the new word, and Beck and colleagues' robust vocabulary instruction (Beck, McKeown & Kucan, 2002) promotes high levels of verbal interaction around a taught word, inevitably strengthening the use of other words and the concepts underlying them. Third is the potential for enhancing a child's lexical awareness, that is, an increased attention to words, their meaning similarities, their differences, and perhaps even their forms. Part of this increased awareness depends on becoming interested in words and engaged in academic language production, as well as comprehension. Although it remains to be seen whether robust vocabulary instruction produces such gains, it is clear that this instruction, as developed by Isabel Beck and her colleagues, includes the kinds of meaningful engagement with language that could promote them.

300

BRINGING READING RESEARCH TO LIFE

Decoding,Vocabulary, and Comprehension

301

Comprehension

lary training. Comprehension of texts allows readers to add new word

issues have tended to localize on a very broad part of it, the higher level
"Comprehension" covers a lot of territory in reading, and the practical comprehension ability to understand extended written tests, as indicated through answering questions, recalling, or summarizing stories. Beck and McKeown (2006) developed Questioning the Author as a way of guiding teachers to support what they saw as the critical component of comprehension: attention to the meaning of the text. As in Beck's other work, this idea is a blend of common sense with research and theory. Getting meaning from a text is about reading the words, encoding their meanings, using sentence structure to form their meanings into semantic content (e.g., propositions), and integrating these meanings with "prior" knowledge and across sentences (Kintsch, 1988; Perfetti, 1985). On this view, the central strategy for reading comprehension is to answer the question of, What does the text say? Other questions (including why questions) engage the reader with the content with the goal of supporting a text-based mental representation of the text. This content-based strategy may be more effective than strategies that aim at a general level (e.g., making inferences, monitoring for confusions) that only indirectly engage semantic content (McKeown, Beck, & Blake, in press).

meanings to their vocabularies, and learning new word meanings

allows readers to comprehend texts that contain those words.

CONCLUSION
The DVC triangle reflects the interdependence of knowledge about word forms (decoding and word identification) and word meanings (vocabulary) and comprehension processes. The LQH formulates these dependencies in terms of the components of word knowledge and its consequences for comprehension. In this framework, once beginning readingdecodinghas been mastered, reading depends on a complex of acquired skills honed by effective reading experiences. Experiences that yield comprehension and also strengthen knowledge of word forms and meanings essentially provide practice for reading skill. Research has contributed substantial knowledge that is of value for reading instruction. We know how to support instruction in decoding so that children can acquire the foundation point of the reading triangle. We are equally sure about the importance of vocabulary but less clear about how to ensure that it keeps up with demands of academic learning. Unlike decoding, which is the great equalizer for unequal opportunity, vocabulary is the reflection of unequal opportunity. Accordingly, it is an even bigger problem to tackle, although we do know how to help children learn word meanings. Comprehension would appear to be the biggest problem, but the research field has provided some useful guidance for comprehension instruction. Once we take into account the vocabulary-comprehension connection, the comprehension issue shrinks a bit. Being able to identify words and use their meanings is a large part of the issue, and with reading practice, especially effective text reading, to support knowledge of word forms and meanings, the comprehension issue becomes one of general language comprehension, certainly a big issue in itself. Because the problems are specialized, researchers typically have pursued one or other corners of the trianglejustifiably so, because each corner represents complexity well beyond what I have implied here. For several reasons, I have found myself working on all the corners and legs of the triangle at one time or another and sometimes at the same time. Truly impressive, however, is what Isabel Beck has done in this triangle. Beyond experimental research, she and her colleagues have taken on the hard problems of instruction, at not one or two but

THE COMPREHENSION-VOCABULARY LEG OF THE TRIANGLE As I noted earlier, comprehension covers a large territory, and educational research in comprehension has attended more to the higher level part of this territory. But comprehension is also about understanding sentences through the meanings of the words they containlocal processes as opposed to global processes. I was able to join Beck, McKeown, and colleagues some years ago in studies that exemplify this level of comprehension and, more important, the link between word meanings and local comprehension. Beck, Perfetti, and McKeown (1982) instructed children in vocabulary and then inserted the newly learned words into sentences and measured the reading (sentence verification) times on the sentences. Children showed gains not only in word meaning measures but also on sentence verification when the sentences contained newly taught words. McKeown, Beck, Omanson, and Perfetti (1983) later found comprehension gains for passages following vocabu-

302

BRINGING READING RESEARCH TO LIFE

Decoding,Vbcabukry, and Comprehension

303

all three points of the triangle. The result is three different projects of systematic, research-based interventions that help students to decode, to learn word meanings, and to comprehend.

REFERENCES Anderson, R. C, & Freebody, P. (1981). Vocabulary knowledge. In J. T. Guthrie (Ed.), Comprehension and teaching: Research reviews (pp. 77-117). Newark, DE: International Reading Association. Andrews, S. (2008). Lexical expertise and reading skill. Psychology of Learning and Motivation, 49, 249-281. Beck, I. L. (2006). Making sense of phonics: The hows and whys. New York: Guilford Press. Beck, I. L., & McCaslin, E. (1978). An analysis of dimensions that affect the development of code-breaking ability in eight beginning reading programs. Pittsburgh, PA: University of Pittsburgh, Learning Research and Development Center. Beck, I. L., & McKeown, M. G. (2006). Improving comprehension with Questioning the Author: A fresh and enhanced view of a proven approach. New York: Scholastic. Beck, I. L., McKeown, M. G., & Kucan, L. (2002). Bringing words to life: Robust vocabulary instruction. New York: Guilford Press. Beck, I., McKeown, M., & Omanson, R. (1987). The effects and uses of diverse vocabulary instruction techniques. In M. McKeown & M. E. Curtis (Eds.), The nature of vocabulary acquisition (pp. 147-163). Hillsdale NJ: Erlbaum. Beck, I. L., Perfetti, C. A., & McKeown, M. G. (1982). The effects of long-term vocabulary instruction on lexical access and reading comprehension. Journal of Educational Psychology, 74, 506-521. Coltheart, M., Rastle, K., Perry, C, Langdon, R., & Ziegler, J. (2001). DRC: A dual route cascaded model of visual word recognition and reading aloud. Psychological Review, 108(1), 204-256. Daneman, M. (1988). Word knowledge and reading skill. In M. Daneman, G. MacKinnon, & T. G. Waller (Eds.), Reading research: Advances in theory and practice (Vol. 6, pp. 145-175). San Diego, CA: Academic Press. Gathercole, S. E., & Baddeley, A. D. (1989). Evaluation of the role of phonological STM in the development of vocabulary in children: A longitudinal study, journal of Memory and Language, 28, 200-213. Harm, M. W., & Seidenberg, M. S. (1999). Phonology, reading acquisition, and dyslexia: Insights from connectionist models. Psychological Review, 106(3), 491-528. Hart, B., & Risley, T. (1995). Meaningful differences in the everyday experience of young American children. Baltimore: Brookes. Kintsch, W. (1988). The role of knowledge in discourse processing: A construction-integration model. Psychological Review, 95,163-182. McKeown, M. G., Beck, I. L., & Blake, R. G. K. (2009). Rethinking reading comprehension instruction: A comparison of instruction for strategies and content approaches. Reading Research Quarterly, 44(3), 218-253.

McKeown, M. G., Beck, I. L., Omanson, R. C, & Perfetti, C. A. (1983). The effects of long-term vocabulary instruction on reading comprehension: A replication. Journal of Reading Behavior, 15(1), 3-18. Nagy, W. E., & Herman, P. A. (1987). Breadth and depth of vocabulary knowledge: Implications for acquisition and instruction. In M. McKeown & M. E. Curtis (Eds.), The nature of vocabulary acquisition (pp. 19-36). Hillsdale, NJ: Erlbaum. Nation, K., & Snowling, M. J. (1998). Semantic processing and the development of word recognition skills: Evidence from children with reading comprehension difficulties. Journal of Memory and Language, 39, 85-101. Perfetti, C. A. (1985). Reading ability. New York: Oxford University Press. Perfetti, C. A. (2007). Reading ability: Lexical quality to comprehension. Scientific Studies of Reading, 11(4), 357-383. Perfetti, C. A., & Hart, L. (2001). The lexical bases of comprehension skill. In D. Gorfien (Ed.), On the consequences of meaning selection (pp. 67-86). Washington, DC: American Psychological Association. Perfetti, C. A., Landi, N., & Oakhill, J. (2005). The acquisition of reading comprehension skill. In M. J. Snowling & C. Hulme (Eds.), The science of reading: A handbook (pp. 227-247). Oxford, UK: Blackwell. Plaut, D. C, McClelland, J. L., Seidenberg, M. S., & Patterson, K. (1996). Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103, 56-115. Reichle, E. D., & Perfetti, C. A. (2003). Morphology in word identification: A word-experience model that accounts for morpheme frequency effects. Scientific Studies of Reading, 7(1), 219-238. Scott, J. A., Jamieson, D., & Asselin, M. (2003). Casting a broad net to catch vocabulary instruction. Elementary School Journal, 103(3), 269-286. Seidenberg, M. S., & McClelland, J. L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96, 523-568. Seigneuric, A. S., Ehrlich, M. F., Oakhill, J. V., & Yuill, N. M. (2000). Working memory resources and children's reading comprehension. Reading and Writing, 13, 81-103. Share, D. L. (1995). Phonological recoding and self-teaching: Sine qua non of reading acquisition. Cognition, 55,151-218. Share, D. L. (1999). Phonological recoding and orthographic learning: A direct test of the self-teaching hypothesis. Journal of Experimental Child Psychology, 72, 95-129. Snowling, M., Hulme, C, & Goulandris, N. (1994). Word recognition and development: A connectionist interpretation. Quarterly Journal of Experimental Psychology, 47A, 895-916. Stanovich, K. E. (1980). Toward an interactive-compensatory model of individual differences in the development of reading fluency. Reading Research Quarterly, 16, 32-71. Wagner, R. (2005, April). Causal relations between vocabulary development and reading comprehension. Toronto, ON, Canada: American Educational Research Association.

In 2005, AA. J. Snowling & C. Hulme (Eds.), The science of reading: A handbook. Oxford: Blackwell.

13
The Acquisition of Reading Comprehension Skill
Charles A. Perfetti, Nicole landi, and Jane Oakhill
How do people acquire skill at comprehending what they read? That is the simple question to which we shall try to make a tentative answer. To begin, we have to acknowledge some complexities about the concept of reading comprehension and what it means to develop it.

Introduction: Simple Ideas about Reading Comprehension

We can expect the comprehension of written language to approximate the comprehension of spoken language. When that happens, then reading comprehension has developed, for practical purposes, to its limiting or asymptotic level. (It is possible for reading comprehension skill to develop so as to exceed listening comprehension skill, but that is another matter.) All other limitations are imposed by linguistic abilities, relevant knowledge, and general intelligence. If we make things more complex than this, we push onto the concept of reading comprehension all these other important aspects of cognition, with the muddle that results from conceptual conflation. This simple idea that the acquisition of reading comprehension is learning to understand writing as well as one understands spoken language has empirical justification. At the beginning of learning to read, the correlations between reading and spoken language comprehension are small (Curtis, 1980; Sticht & James, 1984). This is because at the beginning, children are learning to decode and identify words, so it is these word-reading processes that limit comprehension. However, as children move beyond the beginnings of learning to read, die correlations between reading comprehension and spoken language comprehension increase and then level out by high school (Sticht & James, 1984). As children learn to read words, the limiting factor in reading comprehension shifts from word recognition to spoken language comprehension. For adult college student samples,

228

Charles A. Perfetti, Nicole Landi, and Jane Oakhill

the correlation between scores on reading comprehension and listening comprehension tests reaches r- .90 (Gernsbacher, 1990). If this were the end of the story, then the study of reading comprehension would fold completely into the study of language comprehension. However, there is probably more to the story. First are some methodological considerations. Studies that compare reading comprehension with listening comprehension avoid the confounding of materials, making a clean comparison between the same or equivalent passages with only the "modality" (speech or writing) different. But for most people, what they usually hear is different in content and style from what they read. These differences extend through formal, semantic, and pragmatic dimensions of language. Thus, what is necessary for experimental control is problematic for authenticity. Second, one must make a decision about die speech rate in such comparisons. What is the proper rate for a comparison with reading? The listener's preference? The speaker's preference? A rate equal to the reading rate? Finally, we take note of a more interesting possibility; namely that literacy may alter the way people process spoken language (Olson, 1977). If so, this would boost the correlation of listening and reading comprehension in adulthood. We accept, approximately and in an idealized form, the assumption that reading comprehension is the joint product of printed word identification and listening comprehension, an idea famously asserted by Gough and Tunmer (1986) as a simple view of reading. However, we also must assume that learning to read with comprehension brings enough additional complexities to justify a chapter on how that happens.

A Framework for Comprehension

Comprehension occurs as the reader builds a mental representation of a text message. (For a review of current ideas about reading comprehension in adults, see Kintsch & Rawson, current volume.) This situation model (Van Dijk & Kintsch, 1983) is a representation of what the text is about. The comprehension processes that bring about this representation occur at multiple levels across units of language: word level, (lexical processes), sentence level (syntactic processes), and text level. Across these levels, processes of word identification, parsing, referential mapping, and a variety of inference processes all contribute, interacting with the reader's conceptual knowledge, to produce a mental model of the text. Questions of cognitive architecture emerge in any attempt to arrange these processes into a framework for comprehension. The various knowledge sources can interact freely, or with varying degrees of constraint. For example, computing simple syntactic representations (parsing) probably is more independent of nonlinguistic knowledge than is generating inferences. These issues of cognitive architecture are important, complex, and contentious; we will not discuss them further. Instead, we assume a general framework that exposes the processes of comprehension without making strong assumptions about constraints on their interactions. Figure 13.1 represents this framework schematically.

The Acquisition of Reading Comprehension Skill

229

Comprehension Processes Situation

General Knowledge

Model
Text Representation Parser Meaning and Form Selection Linguistic System Phonology, Syntax, Morphology

1
Lexicon Meaning Morphology Sntax

Word Representaion Word Identification

Orthography Mapping to phonology

Visual Input

Figure 13.1 The components of reading comprehension from identifying words to comprehending texts. Adapted from Perfetti (1999). Within Figure 13.1 are two major classes of processing events: (1) the identification of words, and (2) the engagement of language processing mechanisms that assemble these words into messages. These processes provide contextually appropriate word meanings, parse word strings into constituents, and provide inferential integration of sentence information into more complete representations of extended text. These representations are not the result of exclusively linguistic processes, but are critically enhanced by other knowledge sources. Within this framework, acquiring skill in reading comprehension may include developments in all these components. However, if we focus on reading, as opposed to language comprehension in general, then the unique development concerns printed words. All other processes apply to spoken as well as written language. Children must come to readily identify words and encode their relevant meaning into the mental representation that they are constructing. Although in a chapter on comprehension, we avoid dwelling on word identification, we cannot ignore it completely. Comprehension cannot be successful without the identification of words and the retrieval of their meanings. Both

230

Charles A. Perfetti, Nicole Landi, and Jane Oakhill

children and adults with low levels of comprehension may also have problems with lexical representations, a point to which we shall return later. First we address die sentence and text-level processes that are the defining features of comprehension.

Propositions and mental models The atoms of meaning are extracted from sentences, aggregated through the reading of other sentences of the text and supplemented by inferences necessary to make the text coherent. The bare bones of the text its literal meaning or "text base" consist of propositions (nouns and predicates or modifiers) derived from sentences. They are largely linguistic, based on the meanings of words and the relations between diem (predicates and modifiers), as expressed in a clause. The reader's mental model can be considered an extended set of propositions that includes inferences as well as propositions extracted from actual text sentences. A mental model also may represent text information in an integrated nonpropositional format (Garnham, 1981; Johnson-Laird, 1983), preserving both stated and inferable spatial information in the form of spatial analogues (Glenberg, Kruley, &: Langston, 1994; Haenggi, Kintsch, & Gernsbacher, 1995; Morrow, Greenspan, & Bower, 1987). More typical are texts that are organized, not around space, but about time (Zwaan & Radvansky, 1998). Research has clearly shown that readers are very sensitive to the temporal dimension of narratives (Zwaan, 1996). With this framework of skilled comprehension, we can ask about the acquisition of comprehension skill and differences in comprehension skill. What accounts for comprehension failure? Are the difficulties in comprehension localized in the processes of inference that are needed for the situation model? Or in the processes of meaning extraction that are required to represent the propositions of die text? To address these questions, we examine studies that compare readers who differ in comprehension skill. In most research, the assessment of comprehension is a global one, based on readers' answers to questions following the reading (usually silent, sometimes oral) of very short texts. (For a rare example of an assessment based on the differentiation of comprehension components see Hannon and Daneman, 2001.) We first consider those processes that go beyond understanding the literal meaning of clauses and sentences. We begin with processes commonly viewed as critical to producing higher-level comprehension.

Higher-Level Factors in Comprehension

Among the components of the comprehension framework are three that we highlight in this section: sensitivity to story structure, inference making, and comprehension monitoring. We begin with the last two, which have been proposed as important sources of comprehension development and comprehension problems.

232

Charles A. Perfetti, Nicole Landi, and Jane Oakhill

(Perfetti, Marron, & Foltz, 1996). Yuill and Oakhill (1991) proposed three possibilities to explain inference-making differences between skilled and less-skilled comprehenders: (1) General knowledge deficits restrict less-skilled comprehenders' inference making. (2) Less-skilled comprehenders do not know when, it is appropriate to draw inferences. (3) Less-skilled comprehenders have processing limitations, which hamper their ability to make inferences and integrate text information with prior knowledge. A methodological digression. In sorting through various causal possibilities, there is a pervasive experimental design issue to consider: how to define comparison groups in relation to relative skill and age. One can sample within an age or grade level and compare the more skilled with the less skilled on measures that tap processes hypothesized to produce the differences in comprehension. But any differences in inference making, for example, between a 10-year-old highly-skilled comprehender and a 10-year-old less-skilled comprehender could have arisen because of their differences in comprehension skill or amount of reading. An alternative is to match the children not on chronological age but on "comprehension age"; that is, on their assessed level of comprehension. The comparisons then are between a group of younger children who have attained the same level of comprehension as a group of older children. The older group will be low in comprehension skill relative to their age, whereas the younger group will be average in comprehension relative to their age. These comprehension age matched (CAM) designs allow some of the causal possibilities to be ruled out. If the younger children are better at inferences than the older children, this cannot be attributed to a superior comprehension of the younger group, because the groups have the same absolute level of comprehension skill. Thus, by elimination, a causal link between inference making and comprehension skill becomes more likely. However, all comparisons, whether age- or comprehension-matched, rest on the association of differences, and thus they inherit the limitations of correlational designs for making direct causal conclusions.

Inferences as causal in comprehension skill

In trying to determine the causal status of inference ability in comprehension development, Cain and Oakhill (1999) used the comprehension-match design described above. They compared two groups, one younger and one older, matched on comprehension (CAM) and one group of age-matched skilled comprehenders, as measured by the comprehension score of the Neale Analysis of Reading Ability (Neale, 1997). Thus, less-skilled comprehenders of age 78 were compared with both more skilled comprehenders of the same age and with a younger comprehension matched (CAM) group of age 6. The older two groups were matched on word reading ability according to the Neale accuracy score, whereas the younger CAM group had reading accuracy commensurate with their chronological age, about one year lower than that of the older skilled and less-skilled comprehenders. The three groups read passages and were asked questions that required one of two types of inferences, text connecting or gap-filling. In a text-connecting inference, the reader needed to make a referential link between noun phrases in successive sentences;

The Acquisition of Reading Comprehension Skill

233

for example Michael took the drink out of his bag. The orange juice was very refreshing. Inferring that Michael took orange juice out of his bag is a text-connecting inference. The gap-filling inferences had a more global scope; for example, they required an inference about the setting of a story. One text referred to two children playing in the sand and swimming. Inferring that the children were at the beach would be a gap-filling inference. Cain and Oakhill found that skilled readers and CAM readers were better than less-skilled readers at making text-connecting inferences. On the logic of age-match and comprehension-match comparisons, their conclusion was that comprehension skill is not a cause (it could be a consequence) of text-integration skill (as measured by the ability to make text-connecting inferences). Because skilled comprehenders were better than both the age-matched less-skilled and CAM groups at making such inferences, the causal connection between gap-filling inferences and comprehension was not clarified by the study. If the problems in inference making arise from a poor representation of the text itself, rather than some deficit in the ability to make an inference, then attending to the text could help. When Cain and Oakhill (1999) told children exactly where to look in the text for the relevant information, their performance on the text-connecting inference questions improved, but their performance on the gap-filling inference questions remained poor. The authors concluded that less-skilled readers may have different goals when reading text, perhaps focusing on reading individual words rather than striving for coherence. This suggests that the causal relation between inference making and comprehension could be partly mediated by the reader's standard for coherence. As a working hypothesis, a standard for coherence broadly determines the extent to which a reader will read for understanding, make inferences, and monitor his or her comprehension. A corollary of this hypothesis is that a low standard for text coherence is a general characteristic of low skill comprehenders. Consistent with this possibility, Cain and Oakhill (1996) found that when children were prompted to tell a story, less-skilled comprehenders told stories that had local coherence, but which lacked any overall main point. Cain and Oakhill (1999) proposed that the less-skilled and CAM readers performed more poorly on the gap-filling questions because they failed to know when to use relevant knowledge during reading. They ruled out the availability of the knowledge because a posttest showed equivalent relevant knowledge across the groups. Cain, Oakhill, Barnes, and Bryant (2001) further examined this knowledge question by creating the relevant knowledge. Children were taught an entirely new knowledge base about an imaginary planet ("Gan"), including such facts as "The bears on Gan have blue fur" and "The ponds on Gan are filled with orange juice." Once the knowledge base had been learned to criterion (perfect recall), the children heard a multi-episode story situated on the imaginary planet, and were asked both literal and inferential questions about the story. Correct responses required children to integrate information from the knowledge base with premises from the story. Even when knowledge was controlled in this way, the skilled comprehenders were still able to correctly answer more inference questions than were the less-skilled comprehenders. Not ruled out in either of the above studies are differences in the processing resources (i.e., working memory) that are required to juggle the demands of reading. The retrieval of relevant knowledge, the retention of text information needed for the inference, and

234

Charles A. Perfetti, Nicole Landi, and Jane Oakhill

the building of the inference itself all compete with each other and with other processes (word identification and meaning retrieval). Verbal working memory tasks in fact correlate with both inference tasks and general comprehension measures (Oakhill, Cain, & Bryant, 2003a; Oakhill & Yuill, 1986). However, when we look beyond the correlations, working memory is not the critical factor in comprehension, or at least not the only one. Oakhill et al. (2003a) showed at each of two time points in the study (when the children were age 78 and 89) that inference and text integration skills were predictive of comprehension skill over and above the contribution of working memory, verbal IQ, vocabulary, and word reading accuracy. So, although working memory is likely to contribute to comprehension-related skills like inference making, it is unlikely to be the whole story. Finally, the Cain and Oakhill (1999) study addresses a vexing problem for conclusions about the causal status of inference making. Perfetti et al. (1996) argued that before one can conclude that inference making is a cause of poor comprehension, assurance is needed that the pool comprehender has an effective representation of the basic text meaning (i.e., its literal meaning.) An impoverished representation of the word and clause meanings will make inferences difficult. Cain and Oakhill (1999) addressed this problem by measuring responses to questions about literal content (e.g., asking for the names of the characters which were explicitly given), and found no significant differences (less-skilled readers did show nonsignificantly lower scores). On theoretical grounds, we think the complete separation of inferences from the literal meaning of a text is difficult. In the Construction-Integration processing model of comprehension (Kintsch, 1988), the production of inferences can feed back to literal propositions and strengthen their memory representation. We ought to be surprised to find no differences at all between the literal memory of children who are making inferences and those who are not. Indeed, Cain and Oakhill (1999) showed that literal memory does predict global comprehension; however, they further found that performance on both text-connecting and gap filling inferences predicted comprehension ability even when the ability to answer literal questions (and vocabulary and word reading ability) were controlled. Notice that these results clarify the unique role of inferences in global assessments of comprehension that follow reading. However, they do not verify the assumption that literal text elements are available to the reader when the inference is to be made. As far as we know, although studies have assessed answers to literal questions after reading, the more direct link from a given inference to the text supporting that inference has not been established.

Comprehension monitoring
Readers who strive for coherence in their representation of a text must be able to monitor their comprehension. Monitoring allows the reader to verify his or her understanding and to make repairs where this understanding is not sensible. Skilled readers can use the detection of a comprehension breakdown (e.g., an apparent inconsistency) as a signal for rereading and repair. Less-skilled readers may not engage this monitoring process (Baker, 1984; Garner, 1980). Again the question is why not? This question has not been answered conclusively, but some hints are provided by the many studies on monitoring. For example, a study by Hacker (1997) examined compre-

The Acquisition of Reading Comprehension Skill

235

hension monitoring in seventh-grade, ninth-grade, and eleventh-grade students (mean ages 12 to 16 respectively), with three levels of reading ability within each grade-level. Texts contained three types of detectable problems: contradictory sentences (semantic), various formal writing errors classified as "syntactic" errors (capitalization, verb agreement), and spelling errors. The developmental pattern was increased detection of all categories of text errors with age and, within age, with reading skill. More interesting were the results of an attention manipulation, with students asked to focus on meaning or on form (spelling and grammar). Directing attention to meaning was effective for improved monitoring of meaning errors (with no reduction in detecting form errors) but only for above-average readers. For low-skilled readers, instructional focus appeared not to matter. Thus, for a skilled reader, drawing attention to meaning improves comprehension monitoring. Low reading comprehension appears to be associated with low monitoring performance at all age levels. In the study by Hacker (1997), eleventh-grade low-skill readers were no better than ninth-grade low-skill readers and not as good as seventh-grade skilled readers. The cause of this monitoring problem evades easy explanation. When students were given an additional chance to find the errors with an examiner pointing to the line containing an error, performance improved. However, the least skilled group of readers failed to improve as much as the more skilled groups. This certainly suggests that relevant knowledge is not always used in monitoring and that there are knowledge and basic processing differences that limit monitoring among some low-skilled readers. Thus, not all the problems can be due to a "monitoring deficit." Again, reading with a certain coherence standard is necessary for monitoring to be engaged. It is important to note that observed differences in monitoring comprehension are not independent of the reader's ability to construct an accurate representation of the sentences in the text (Otero & Kintsch, 1992; Vosniadou, Pearson, & Rogers, 1988). Vosniadou et al. (1988) studied first-, third-, and fifth-grade readers' detection of text inconsistencies compared with their detection of false sentences that contradict facts that the child could know from memory. The familiarity of the critical information proved to be important for whether the child could detect an inconsistency, based either on memory or the text. This result, while not surprising, reinforces the important point that retrieving relevant knowledge during reading is essential for monitoring. However, when they controlled the familiarity of the critical information, Vosniadou et al. (1988) found that children were as good at detecting inconsistencies based on two contradictory text sentences as they were at detecting the contradiction of a single sentence with a familiar fact. This finding suggests that at least some problems in monitoring can be characterized as a failure to encode the meaning of a sentence in a way that promotes its comparison with other information, either in the text or in memory. A simple explanation is difficult because comprehension monitoring, like inference making, both contributes to and results from the reader's text representation. This makes it difficult to attribute comprehension problems uniquely to a general failure to monitor comprehension. Any observed problem can result from an incomplete representation of sentence meaning, a failure to activate relevant knowledge at the critical moment, a failure to monitor the coherence of the text with respect either to its internal consistency or the readers' knowledge of the world. Finally, as in the case of inference making, the standard-

236

Charles A. Perfetti, Nicole Landi, and Jane Oakhill

of-coherence hypothesis may be relevant: Comprehension monitoring failures may result from a low standard for coherence.

Sensitivity to story structure The genre of texts (narrative, descriptive, etc.), their linguistic styles, and the various layouts of texts all can present novel problems that are solved only by experience in reading. Among the many text genre possibilities, the simple story of the sort encountered by children in schools has attracted the most attention, and we focus here on this specific text type. The developmental research on this topic has focused on the understanding of story structure (e.g., Smiley, Oakley, Worthen, Campione, & Brown, 1977; Stein & Glenn, 1979). What is interesting about this development is its earliness. Stein and Albro (1997) argue that story understanding depends on knowledge about the intentions that motivate human action, and conclude that this knowledge is typically acquired by age 3. If so, although the application of narrative understanding to written texts can undergo further development with reading experience, we would not expect that story structure "deficits" would limit comprehension skill. Beyond the conceptual bases for narrative, however, is the understanding that the text itself honors the narrative structure through coherence devices. Differences in this sensitivity to text coherence could lead to differences in comprehension. Indeed, a study by Yuill and Oakhill (1991) demonstrated that, when they were required to narrate a story from a picture sequence, the less-skilled comprehenders produced fewer causal connectives and made more ambiguous use of referential ties than did skilled comprehenders. The less-skilled comprehenders also had difficulties in using linguistic elements to make their stories well structured and integrated. Less-skilled comprehenders have been found to have weakness in other aspects of text structure understanding. Cain and Oakhill (1996) required groups of skilled and less-skilled comprehenders, together with a comprehension-age match group, to tell stories prompted by a title, such as "Pirates." The less-skilled comprehenders produced more poorly structured stories than either of the other two groups. Their poorer performance relative to the comprehension-age match group indicates that the ability to produce well-structuted stories is not simply a by-product of having a certain level of comprehension skill. (Again, on the logic of comprehension match, this is because the poor comprehenders and the younger, comprehension-age match group had the same absolute level of comprehension skill.) Rather, an ability to produce a well-structured story is more likely to be associated with the causes of comprehension development. A sensitivity to story structure is one possibility for a cause of this development. A standard for coherence that extends to both production and comprehension is another possibility. Reading comprehension skill is also related to children's knowledge about particular story features: notably titles, beginnings and endings. In one study, more than 80% of skilled comprehenders could give examples of the information contained in a story title, such as "it tells you what it's about and who's in it"; whereas, only about 25% of a same-age group of less-skilled comprehenders were able to do so (Cain, 1996). Some of the

The Acquisition of Reading Comprehension Skill

237

less-skilled readers claimed that the title of a story provides no useful information at all. Less-skilled comprehenders were also less aware that the beginnings of stories might provide useful information about the story setting and characters. Thus, less-skilled comprehenders appear to have less explicit awareness of the features of stories that might help scaffold their mental representation of the text. However, although less-skilled comprehenders are poor at explaining the function of a variety of text features, they must have at least some implicit awareness of the use of such features, because they benefit from integrated and goal-directed titles in both comprehension and production tasks (Cain & Oakhill, 1996; Yuill & Joscelyne, 1988).

The Linguistic-Conceptual Machinery for Comprehension

Below the higher-level aspects of comprehension are the processes that convert sentences into basic semantic content, their propositional meaning. The derivation of propositional meaning requires knowledge about syntactic forms and the meanings of words.

Syntactic processing Since the defining arguments by Chomsky (1965) and early research on the development of language (e.g., McNeill, 1970), the implicit assumption seems to have been that syntax should not be an issue for the development of reading. Competence in the grammar of one's native language is acquired naturally, emerging from biological dispositions through the niters of a local linguistic environment well before entry to school. Reading would naturally use this same grammatical knowledge. However, once differences in syntax between typical spoken forms and typical written forms are acknowledged (O'Donnell, 1974), the simple story is compromised. The question becomes empirical: Are the child's syntactic abilities, cultivated in a natural social environment, enough to meet the challenges of the more formal and more complex syntax that is present in written texts? We should expect that language skill differences lead to individual differences in comprehension, and, in fact, younger less-skilled readers show a wide range of problems with syntax and morphology (Fletcher, Satz, & Scholes, 1981; Stein, Cairns, & Zurif, 1984). The question is whether such problems arise from a syntactic knowledge deficit or from some other source that affects performance on syntactic tasks (such as working memory, lack of practice, or lexical processing limitations). Research with children (Crain & Shankweiler, 1988) and adults (Carpenter, Miyake, & Just, 1994) suggests that syntactic parsing problems can arise from processing limitations rather than a lack of syntactic knowledge. Comprehension difficulties may be localized at points of high processing demands, whether from syntax or other sources. Crain and Shankweiler (1988) concluded that even less-skilled readers have the necessary syntactic abilities to comprehend the relatively complex sentences they used in their studies. For example, children as young as three years can understand restrictive relative clauses such as "A cat is holding hands with a man that is holding hands with a woman."

238

Charles A. Perfetti, Nicole Landi, and Jane Oakhill

Thus, difficulties with syntax, when they are observed, may be in masquerade, with the real problem lying elsewhere. The "elsewhere" has been assigned to verbal working memory ability (Crain & Shankweiler, 1988; Perfetti, 1985) or difficulty processing phonological material (Bar-Shalom, Crain, & Shankweiler, 1993). Nevertheless, there have been few thorough studies of the broader question of the syntactic abilities of less-skilled comprehenders. Accordingly, the conclusion that all syntactic difficulties originate as working memory limitations is too strong. Differences in syntactic processing can be observed in the absence of obvious phonological problems (Stothard & Hulme, 1992). In a study of 7-9-year-olds, Oakhill et al. (2003a) found significant relations between global comprehension skill and a measure of syntactic ability (the TROG, a picture-sentence matching test, also used by Stothard and Hulme, 1992). (Relations were also found for text integration, comprehension monitoring, and working memory.) However, with verbal ability and vocabulary controlled, syntactic ability was significant at only the second of two test points. Although a more precise role for syntactic abilities, free of other factors, remains to be worked out, its role may be genuine, reflecting variability in the development of functional language skills. Finally, gaining experience with syntactic structures that are less common in spoken than written language, e.g., the use of nominalizations, clausal noun phrases, and other more complex structures, is something that benefits from successful reading. Experience with a variety of syntactic structures should increase functional expertise in syntax and reduce the demands of complex structures on working memory.

Working memory systems Understanding a sentence involves remembering words within the sentence, retrieving information from preceding text, parsing the sentence, and other processes that require resources. Working memory one or more systems of limited capacity that both store and manipulate information - is a bottleneck for these processes. The hypothesis that working memory factors are correlated with individual differences in comprehension has received wide support (Baddeley, Logie, & Nimmo-Smith, 1985; Crain & Shankweiler, 1988; Just & Carpenter, 1992; Perfetti & Lesgold, 1977). In addition, the evidence shows it is an active working memory system rather than a passive short-term memory store that is important in reading comprehension skill (Daneman & Carpenter, 1980; Perfetti & Goldman, 1976; Seigneuric, Ehrlich, Oakhill, & Yuiil, 2000). Different subsystems of working memory have been postulated, including one that is specialized for holding and manipulating phonological information (Baddeley, 1979). Phonological working memory has a direct link to reading through the need to keep active the contents of a sentence until the end of a clause or sentence, when integrative processes complete their work and make a verbatim memory less important. A phonological memory system directly affects the comprehension of spoken language. In fact, children who are less skilled in reading comprehension show poorer memory for words they recently heard from spoken discourse (Perfetti & Goldman, 1976). This interdependence of spoken and written language comprehension is important in the analysis of reading

The Acquisition of Reading Comprehension Skill

239

comprehension problems. Whether phonological memory is the critical cause of differences in both spoken and written language comprehension is another matter. As we suggest below, the basic language processing mechanisms, which include more than phonological representations, may affect performance in working memory tasks. Phonological memory processes may affect reading comprehension by an additional pathway through the development of word identification. Dufva, Niemi, and Voeten (2001), in a longitudinal study from preschool through second grade, used assessments of phonological awareness, phonological memory, word identification, and spoken and written comprehension. Structural equation modeling showed an indirect causal link from preschool phonological memory to word recognition development between first and second grade, which was mediated by phonological awareness. Phonological memory showed a similar indirect causal link to reading comprehension, mediated by listening comprehension. The results suggest that the ability to hold and manipulate phonemes in memory may explain the relation between phonemic awareness and reading. Moreover, they suggest that phonological memory supports listening comprehension and thus, indirectly, reading comprehension. Because word identification and listening comprehension are primary determinants of reading comprehension, phonological knowledge prior to literacy could play a role in the development of reading comprehension by either or both of two pathways. A causal path from early phonological knowledge through word identification to later reading comprehension is one possibility. Another possibility is a pathway from phonological processing to listening comprehension to reading comprehension. Of course, both causal pathways could be involved. On either description, working memory capacity is not at the heart of comprehension problems, but rather its correlations with comprehension reflect limitations in phonological processing. Indeed, Crain and Shankweiler (1988) argued that differences in working memory capacity arise from difficulties in phonological processing. In the absence of specifically phonological problems, working memory differences are still observed and can be traced to other language processing weaknesses (Nation, Adams, Bowyer-Crane, & Snowling, 1999; Stothard & Hulme, 1992). The general conclusion appears to be that working memory differences related to reading skill are fairly specific to language processing. Indeed, the even more general conclusion is that language processing weaknesses are at the core of reading comprehension problems. These weaknesses will often be manifest specifically in phonology but they can also be reflected in other aspects of language processing. The assumption of a limited capacity working memory system has been central in theories of cognition generally. An additional implicit assumption is that this system is more or less fixed biologically. However, alternative perspectives on working memory suppose that its limitations are not completely fixed but at least partly influenced by knowledge and experience (Chi, 1978; Ericsson & Delaney, 1999; Ericsson & Kintsch, 1995). If we see working memory as partly fixed and partly "expandable," we move toward a perspective that views the role of effective experience as critical in the development of comprehension skill. Effective experiences in a domain strengthen the functionality of memory resources in that domain. In the case of reading, the effective experience is reading itself (with a high standard for coherence) so as to support the fluent processing that effectively stretches working memory.

240

Charles A. Perfetti, Nicole Landi, and Jane Oakhill

Building conceptual understanding from words Vocabulary has been a slightiy neglected partner in accounts of reading comprehension. This neglect arises not from any assumption that vocabulary is unimportant, but from theoretical interests in other aspects of the comprehension problem. The research strategies have either assumed or verified that relevant vocabulary knowledge is equal between a group of skilled and less-skilled comprehenders, so that experimental designs could focus on inferences, monitoring, working memory, or whatever component of comprehension was the target of interest. Of course, everyone accepts that knowledge of word meanings and comprehension skill are related. The possible causal relations underlying their relationship include several plausible possibilities (Anderson & Freebody, 1981; Beck, McKeown, & Omanson, 1987; Curtis, 1987). Word meanings are instrumental in comprehension on logical as well as theoretical grounds. Nevertheless, the more one reads, the more comprehension brings along increases in the knowledge of word meanings. Sorting out causality is again difficult, and we might expect research designs to follow the lead of the comprehension-match design, making matches based on vocabulary levels. For some purposes, it does not matter whether the causal history is from vocabulary-to-comprehension or comprehension-to-vocabulary. Indeed, the causal relationship is likely to be reciprocal. To the extent that word meanings are inferred from context, then vocabulary growth results from comprehension skill, including inference making. But at the moment a reader encounters a text, his or her ability to access the meaning of the word, as it applies in the context of this particular text, is critical. Not knowing the meanings of words in a text is a bottleneck in comprehension. Because readers do not know the meanings of all words they encounter, they need to infer the meanings of unknown words from texts. This process, of course, requires comprehension and like other aspects of comprehension, it is correlated with working memory (Daneman & Green, 1986). This correlation might reflect working memory's role in learning the meanings of words from context (Daneman, 1988). Note also that inferring the meanings of unknown words from the text is possible only if most words are understood and if some approximation to text meaning is achieved. One estimate is that a reader must know at least 90% of the words in a text in order to comprehend it (Nagy & Scott, 2000). We know very little about the kind of text representation that results when words are not understood. The nature of this representation would depend on all sorts of other factors, from the role of an unknown word in the structure of the text message to the reader's tolerance for gaps in comprehension. Somehow, children's knowledge of word meanings grows dramatically. Nagy and Herman (1987), based on several earlier estimates of vocabulary growth, computed the per-year growth of vocabulary at 3,000 words over grades 112. The gap between the number of words known by the high-knowledge and low-knowledge children is correspondingly large. According to one estimate, a first-grade reader with high vocabulary knowledge knows twice as many words as a first-grade reader with low knowledge, and this difference may actually double by the twelfth grade (Smith, 1941). Differences in word knowledge emerge well before schooling. Large social class differences in the vocabulary heard by children at home produce corresponding differences

The Acquisition of Reading Comprehension Skill

241

in the vocabularies of children as they enter school (Hart & Risley, 1995a). These differences are not about only the conventional meanings of words, but the background knowledge needed to interpret messages that contain these words. Consider this example (from Hart & Risley, 1995a): My wife and I wanted to go to Mexico, but her only vacation time was in July. Interpreting the "but" clause, which needs to be understood as causal for an unstated action (they probably did not go to Mexico), is easier if the reader knows that Mexico is very hot in July and that some people might not want to have a vacation in high heat. Knowledge of this sort is critical in its consequences for understanding even simple texts. Beyond the general importance of word knowledge (and associated conceptual knowledge) are specific demonstrations that children less skilled in comprehension have problems with word knowledge and semantic processing. Nation and Snowling (1998a) compared children with specific comprehension difficulties with a group of skilled comprehenders matched for decoding ability, age and nonverbal ability on semantic and phonological tasks. They found that less-skilled comprehenders scored lower on a synonym judgment task (Do BOAT and SHIP mean the same thing?), although not on a rhyme judgment task (Do ROSE and NOSE rhyme?). Less-skilled comprehenders were also slower to generate semantic category members (but not rhymes) than skilled comprehenders. This suggests that comprehension problems for some children are associated with reduced semantic knowledge (or less effective semantic processing) in the absence of obvious phonological problems. (See also Nation, this volume.) More interesting, however, is that these same less-skilled comprehenders showed a problem in reading low-frequency and exception words. In effect, Nation and Snowling observed a link between skill in specific word identification (not decoding) and comprehension that could be mediated by knowledge of word meanings. Theoretically, such a link can reflect the role of word meanings in the identification of words that cannot be identified by reliable graphemephoneme correspondence rules. Children with weak decoding skills may develop a dependency on more semantically based procedures (Snowling, Hulme, & Goulandris, 1994). Thus, knowledge of word meanings may play a role in both the identification of words (at least in an orthography that is not transparent) and in comprehension. This dual role of word meanings places lexical semantics in a pivotal position between word identification and comprehension. (Notice that figure 13.1 reflects its pivotal position.) This conclusion also accords with an observation on adult comprehenders reported in Perfetti and Hart (2002), who reported a factor analysis based on various reading component assessments. For skilled comprehenders word identification contributed to both a word form factor (phonology and spelling) and a comprehension factor, whereas for less-skilled comprehenders, word identification was associated with a phonological decoding factor but not with spelling or comprehension. This dual role of word meanings in skilled reading also may account for previous observations that less-skilled comprehenders are slower in accessing words in semantic search tasks (Perfetti, 1985). If less-skilled readers have a weak lexical semantic system, then one might expect semantic variables that reflect the functioning of this system to make a difference. For example, concrete meanings are more readily activated than more abstract meanings.

242

Charles A. Perfetti, Nicole Landi, and Jane Oakhill

Nation, et al. (1999) found that an advantage for concrete words was more pronounced for less-skilled readers than skilled readers who were matched for nonword reading (decoding). In a priming study, Nation and Snowling (1999) found that less-skilled com-prehenders are more sensitive to associative strength among related words and less sensitive to abstract semantic relations, compared with skilled comprehenders. Research at this more specific semantic level could help clarify the nature of the semantic obstacles to comprehension.

Word Identification, Decoding, and Phonological Awareness

If word meanings are central to comprehension and important for identification of at least some words, then we have come to an interesting conclusion: Despite trying to ignore word level processing in comprehension, we cannot. In examining the role of working memory, we were forced to conclude that a link to comprehension could go from phonological processing through word identification to comprehension. Even phonological awareness, ordinarily considered only important for decoding, has been found to predict young readers' comprehension independently of working memory (Leather & Henry, 1994). The general association between word identification and reading comprehension skill has been well established for some time (Perfetti & Hogaboam, 1975). This association reflects the fact that word identification skill and comprehension skill develop in mutual support. The child's development of high-quality word representations is one of the main ingredients of fluent reading (Perfetti, 1985, 1991). Such representations must be acquired in large part through reading itself. Instrumental in acquiring these word representations is a process identified by Share (Share, 1995, 1999) as self-teaching. This process allows children to move from a reading process entirely dependent on phonological coding of printed word forms to a process that accesses words quickly based on their orthography. What drives this development of orthographic access is the child's decoding attempts, which provide phonological feedback in the presence of a printed word, establishing the orthography of the word as an accessible representation. Models that simulate learning to read words can be said to implement this kind of mechanism (Plaut, McClelland, Seidenberg, & Patterson, 1996). As children develop word-reading skills, comprehension becomes less limited by word identification and more influenced by other factors. However, even for adult skilled readers, the association between reading comprehension and word identification persists, reflecting either a lingering limitation of word identification on comprehension or a history of reading experience that has strengthened both skills. The word-level skill can be conceived as reflecting lexical quality (Perfetti & Hart, 2001), knowledge of word forms and meanings, which has its consequences in effective and efficient processing. Word level processing is never the whole story in comprehension. However, it is a baseline against which to assess the role of higher-level processes such as comprehension monitoring and inference making (Perfetti et al. 1996).

The Acquisition of Reading Comprehension Skill Which components bring about growth in comprehension skill?

243

To this point, we have examined the acquisition of reading skill largely through studies comparing skilled and less-skilled readers, whether matched on relevant skills or age. Longitudinal studies that track the course of changes in comprehension skill can provide additional information about the causal relations among the components of comprehension, and thus about the course of development. A few such studies have begun to appear. Muter, Hulme, Snowling, and Stevenson (2004) studied young children for two years from their entry into school, assessing a number of abilities, including phonological and grammatical abilities and vocabulary knowledge. Word identification skills, grammatical knowledge, and vocabulary assessed at age 5-6 each predicted unique variance in reading comprehension at the end of the second year of schooling. This pattern confirms the contributions to comprehension of diree factors we have reviewed in previous sections. In a longitudinal study of children in school years 3 to 6, Oakhill, Cain, and Bryant (2003b) extended the study of Oakhill et al. (2003a) by the addition of a third cohort of children and providing a longitudinal analysis of data from ages 78 (Year 3), 89 (Year 4), and 1011 (Year 6). In each age group, there were measures of reading comprehension and reading accuracy, verbal and performance IQ (Time 1 only), working memory (both verbal and numerical span measures), phonemic awareness (phoneme deletion), vocabulary (British Picture Vocabulary Test), syntax (TROG), and measures of three comprehension related skills: inference making, comprehension monitoring and story-structure understanding (story anagram task). The results of multiple regression were applied to a causal path diagram to show the pattern and strength of relations among the various skills across time. The final causal path diagram, with only significant paths included, is shown in figure 13.2. Initial comprehension skill was a strong predictor of later comprehension, and verbal ability (vocabulary and verbal IQ) also made significant contributions to the prediction of comprehension ability across time. Nevertheless, three distinct predictors of comprehension skill emerged, either through direct or indirect links: answering inferential questions, monitoring comprehension (by detecting inconsistencies in text), and understanding story structure (assessed by the ability to reconstruct a story from a set of jumbled sentences). These factors predicted comprehension at a later time even after the auto-regressive effect of comprehension (the prediction of comprehension at later times from comprehension at earlier times) was controlled. With reading accuracy as the dependent variable, the pattern was quite different. The significant predictors were previous measures of reading accuracy and a phoneme deletion measure taken at Time 1. From these analyses a picture of skill development emerges in which certain components of comprehension are predictive of general comprehension skill. Early abilities in inference skill, story structure understanding, and comprehension monitoring all predict a later global assessment of comprehension skill independendy of the contribution of earlier comprehension skill. Finally, to assess growth in skill, Oakhill et al. (2003b) calculated estimates of growth in reading comprehension and reading accuracy, and used these estimates as dependent variables in two further sets of regression analyses. (Verbal and performance IQand vocabulary were entered at the first step, followed by all of the reading-related and language

244

Charles A. Perfetti, Nicole Landi, and Jane Oakhill

Comprehension Tl

Comprehension T3

Verbal IQ TV

BPVS Tl

Story anagram Tl

Monitoring Tl

Monitoring T2

Figure 13.2 Path analysis based on data from longitudinal study by Oakhill, Cain, and Bryant (2003b). Variables measured at Time 1 (age 7-8) predict variables at Time 2 (age 910) and Time 3 (age 1112). Variables shown were significant predictors after the effects of all other variables were removed: a global comprehension measure (COMP), a picture vocabulary test (BPVS), verbal IQ (VIQ), detection of text contradictions (MONITOR), a sensitivity to story structure (Story Anagram), and integrative inferences (INFER). Paths mat linked Time 1 and Time 2 variables but not Time 3 comprehension have been excluded for clarity. Because the original data were standardized, the coefficients shown are directly comparable. variables and working memory measures entered simultaneously.) Although vocabulary and verbal IQ predicted growth in comprehension and reading accuracy, other variables made independent predictions. Story structure understanding was the sole predictor of growth in reading comprehension. Phonemic awareness was the sole predictor of growth in reading accuracy skill. The study confirms that a set of higher-level comprehension components, which, on theoretical grounds, ought to be instrumental in the growth of reading comprehension skill, may indeed be instrumental. Muter et al. (2004) report a slightly different pattern for their younger children. Word identification (Hatcher, Early Word Recognition Test, Hatcher, Hulme, & Ellis, 1994) was important in predicting comprehension, as one might expect for younger children, as were knowledge of word meanings and grammatical knowledge. Because Muter et al. had a comprehension assessment only at the final test point in their study, their study is not directly comparable with the study by Oakhill et al. (2003b). It is possible that all the factors identified in these two studies influence comprehension development, with the strength of their contribution depending upon the level of the child's skill. However, studies that carry out comparable assessments, including tests of comprehension at more than one time point, are needed to test this possibility.

Comprehension Instruction
A failure to develop a high level of comprehensions skill creates a severe obstacle to educational attainment. Accordingly, there is widespread concern about how to improve chil-

The Acquisition of Reading Comprehension Skill

245

dren's reading comprehension. Although we cannot review the research on instruction in comprehension here, we briefly note the wide extent of such research, drawing on a comprehensive review of research on reading (National Institute of Child Health and Human Development [NICHD], 2000). The summary NICHD report refers to 453 studies between 1980 and the time of the review, augmented by a few earlier studies from the 1970s. The 205 studies that met the methodological criteria led the report to identify seven categories of comprehension instruction that appeared to have solid evidence for their effectiveness. These seven include procedures that we characterize as drawing the reader into a deeper engagement with the text - in a phrase, active processing. They include comprehension monitoring, question answering (teacher directed questions) and question generation (student self-questioning), the use of semantic organizers (students making graphic representations of text), and student summarization of texts. Instruction in story structures was also judged to be effective. The NICHD Report concludes that these procedures are effective in isolation in improving their specific target skills (sensitivity to story structures, quality of summarization, etc.) but that improvement of scores on standardized comprehension tests may require training multiple strategies in combination. The procedures that the NICHD Report suggests are effective are consistent with the comprehensions skills we have reviewed in this chapter. Active engagement with the meaning of text helps the reader to represent the text content in a way that fosters both learning (as opposed to superficial and incomplete understanding) and an attraction to reading. However, the NICHD report adds some cautions to its conclusions on behalf of the instruction strategies it recommends. To those, we add our own reservations. Instructional interventions may produce only short-term gains. Two years after the intervention, is the child comprehending better? Answers to this kind of question appear to be lacking. We think the complex interaction among the comprehension components and the role of motivation for reading make real gains difficult to achieve. Internalizing externally delivered procedures so that they become a habit a basic attitude toward texts and learning may be a long-term process. It requires both wanting to read and gaining skill in reading, which go hand in hand.

Conclusion: A More General View of Comprehension Development

We conclude by taking a step back from the details of how skill in comprehension is acquired. With more research, the kind of developmental picture we described in the preceding sections may be confirmed or alternative pictures will emerge, based on different experimental tasks and resulting in a different arrangement of causal relations. Because a detailed model of skill acquisition seems premature, we turn to a more general, speculative account of acquisition. This general model framework, which is illustrated as a highly schematic representation in figure 3.3, can be realized by a number of specific models.

246

Charles A. Perfetti, Nicole Landi, and Jane Oakhill

Fi9lire 13.3 A schematic representation of the major components in the acquisition of reading comprehension skill. The leftright arrows represent increases in skill across with experience and gains in knowledge. Reading comprehension depends on spoken language comprehension throughout development. Early in reading, written word identification (not shown) is a limiting factor for reading comprehension. Reading comprehension has reciprocal relationships with both spoken language comprehension and lexical knowledge. Not represented: general knowledge (which, of course, also increases) and the specific processes of comprehension (e.g., syntactic processing and inference making).

We assume the following: 1. 2. General skill in reading comprehension and its related components increase with reading experience, and, with some component skills, with spoken language experience. Reading comprehension and listening comprehension are related throughout development. Their relation is reciprocal, with experience in each potentially affecting skill acquisition in the other. However, this does not mean that the two are "equal," and substantial asymmetries can develop.

The Acquisition of Reading Comprehension Skill 3.

247

4. 5.

Word identification skill sets a limit on how closely reading comprehension skill can approach listening comprehension skill. It specifically limits comprehension early in reading development. Knowledge of word meanings is central to comprehension. This knowledge derives from multiple sources, including written and spoken comprehension, and grows indefinitely. Higher levels of comprehension require the reader to apply a high standard of coherence to his or her understanding of the text.

The first four assumptions comprise a basic analysis of what is necessary for comprehension. Our review of research on higher-level comprehension processes emphasizes the need for this basic analysis to be taken into account - that is, "controlled for" - in the search for higher-level comprehension factors that are strategic; for example, monitoring comprehension, making inferences. However, we conclude also that the basic analysis provides the necessary, but not sufficient, causal story. For comprehension to develop to higher levels, the reader must adopt a high standard of coherence to care whether the text makes sense. When coherence is a goal, inferences are made to keep things coherent. When coherence is a goal, inconsistencies between text elements or between text elements and the reader's knowledge are resolved rather than ignored or not noticed. All readers find themselves relaxing their standards for coherence occasionally. Unwanted reading and countless nontext distractions can promote this laxity. The goal, however, is adopting the high-standard criterion as the "default." We think skilled readers do this. This brings reciprocal supports into play. Adopting a high coherence standard supports interest in reading, which encourages a high standard of coherence. The result of these influences is more reading and, especially, more effective reading. This surely aids reading comprehension.

Note The authors are grateful to Kate Cain for providing comments on a draft of this chapter, which was prepared while the first author was a visiting research fellow at the University of Sussex. The first author's work on the chapter was supported by a Leverhulme Visiting Professorship and a comprehension research award from Institute for Educational Sciences (US Department of Education).

References Anderson, R. C., & Freebody, P. (1981) Vocabulary knowledge. In J. T. Guthrie (Ed.), Comprehension and teaching: Research reviews (pp. 77-117). Newark DE: International Reading Association. Baddeley, A. D. (1979). Working memory and reading. In P. A. Kolers, M. E. Wrolstad, & H. Bouma (Eds.), Procession of visible language (Vol. 1, pp. 355-370). New York: Plenum. Baddeley, A., Logie, R., & Nimmo-Smith, I. (1985). Components of fluent reading. Journal of Memory and Language, 24(1), 119-131. Baker, L. (1984). Spontaneous versus instructed use of multiple standards for evaluating comprehension: Effects of age, reading proficiency, and type of standard. Journal of Experimental Child Psychology, 38, 289-311. Barnes, M. A., Dennis, M., & Haefele-Kalvaitis. (1996). The effects of knowledge availability and knowledge accessibility on coherence and elaborative inferencing in children from six to fifteen years of age. Journal of Experimental Child Psychology, 61, 216-241. Bar-Shalom, E., Crain, S., & Shankweiler, D. (1993). A comparison of comprehension and production abilities of good and poor readers. Applied Psycholinguistics, 14, 197-227. Beck, I., McKeown, M. & Omanson, R. (1987). The effects and uses of diverse vocabulary instruction techniques. In M. McKeown, & M. E. Curtis (Eds.), The nature of vocabulary acquisition (pp. 147-163). Hillsdale NJ: Erlbaum. Cain, K. (1996). Story knowledge and comprehension skill. In C. Cornoldi & J. Oakhill (Eds), Reading comprehension difficulties: Processes and interventions, pp. 167-192. Mahwah, NJ: Erlbaum. Cain, K., & Oakhill, J. V. (1996). The nature of the relationship between comprehension skill and the ability to tell a story. British Journal of Developmental Psychology, 14, 187-201. Cain, K., & Oakhill, J. V. (1999). Inference ability and its relation to comprehension failure in young children. Reading and Writing, 11, 489-503. Cain, K., Oakhill, J. V., Barnes, M. A., & Bryant, P. E. (2001). Comprehension skill, inference-making ability, and the relation to knowledge. Memory and Cognition, 29(6), 850-859. Carpenter, P. A., Miyake, A., & Just, M. A. (1994). Working memory constraints in comprehension: Evidence from individual differences, aphasia, and aging. In M. Gernsbacher (Ed.), Handbook of Psycholinguistics (pp. 1075-1122). San Diego, CA: Academic Press. Casteel, M. A., & Simpson, G.B. (1991). Textual coherence and the development of inferential generation skills. Journal of Research in Reading, 14, 116-129.

Chi, M. T. H. (1978). Knowledge structures and memory development. In R. Siegler (Ed.), Childrens thinking: What develops? (pp. 73-96). Hillsdale, NJ: Erlbaum. Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge, MA: MIT Press. Crain, K., & Shankweiler, D. (1988). Syntactic complexity and reading acquisition. In A. Davison & G. M. Green (Eds.), Linguistic complexity and text comprehension: Readability issues reconsidered (pp. 167-192). Hillsdale, NJ: Erlbaum. Curtis, M. E. (1980). Development of components of reading skill. Journal of Educational Psychology, 72, 656-669. Curtis, M. E. (1987) Vocabulary testing and vocabulary instruction. In M. McKeown, & M. E. Curtis (Eds.), The nature of vocabulary acquisition (pp. 37-51). Hillsdale NJ: Erlbaum. Daneman, M. (1988). Word knowledge and reading skill. In. M. Daneman,G. MacKinnon and T.G. Waller (Eds.), Reading research: Advances in theory and practice, Vol. 6. San Diego: Academic Press. Daneman, M., & Carpenter, P. A. (1980). Individual differences in working memory and reading. Journal of Verbal Learning and Verbal Behavior, 19, 450-466. Daneman, M., & Green, I. (1986). Individual differences in comprehending and producing words in context. Journal of Memory and Language, 25, 1-18. Dufva, M., Niemi, P., & Voeten, M. J. M. (2001). The role of phonological memory, word recognition, and comprehension skills in reading development from preschool to grade 2. Reading and Writing, 14, 91-117. Ericsson, K. A., & Delaney, P. F. (1999). Long-term working memory as an alternative to capacity models of working memory in everyday skilled performance. In A. Miyake & P. Shah (Eds.), Models of working memory: Mechanisms of active maintenance and executive control (pp. 257-297). Cambridge University Press. Ericsson, K. A., & Kintsch, W. (1995). Long-term working memory. Psychological Review, 102(2), 211-245. Fletcher, J. M., Satz, P., & Scholes, R. J. (1981). Developmental changes in the linguistic performance correlates of reading achievement. Brain and Language, 13, 78-90. Garner, R. (1980). Monitoring of understanding: An investigation of good and poor readers awareness of induced miscomprehension of text. Journal of Reading Behavior, 12, 5-63. Garnham, A. (1981). Mental models as representations of text. Memory and Cognition, 9, 560-565. Gernsbacher, M. A. (1990). Language comprehension as structure building. Hillsdale, NJ: Erlbaum.

Glenberg, A. M., Kruley, P., & Langston, W. E. (1994). Analogical processes in comprehension: Simulation of a mental model. In M. A. Gernsbacher, (Ed.) Handbook of psycholinguistics (pp. 609-640). San Diego, CA: Academic Press. Gough, P. B., & Tunmer, W. E. (1986). Decoding, reading and reading disability. Remedial and Special Education, 7, 6-10. Graesser, A. C., & Kreuz, R. J. (1993). A theory of inference generation during text comprehension. Discourse Processes, 16, 146-160. Graesser, A. C., Singer, M., & Trabasso, T. (1994). Construction inferences during narrative comprehension. Psychological Review, 101, 371-395. Hacker, D. J. (1997). Comprehension monitoring of written discourse across early-to-middle adolescence. Reading and Writing, 9, 207-240. Haenggi, D., Kintsch, W., & Gernsbacher, M. A. (1995). Spatial situation models and text comprehension. Discourse Processes, 19, 173-199. Hannon, B., & Daneman, M. (2001). A new tool for measuring and understanding individual differences in the component processes of reading comprehension. Journal of Educational Psychology, 93, 103-128. Hart, B., & Risley, T. (1995). Learning to talk. Baltimore: Brookes. Hatcher, P.J., Hulme, C. and Ellis, A.W. (1994). Ameliorating early reading failure by integrating the teaching of reading and phonological skills. Child Development, 65, 41-57. Johnson-Laird, P. N. (1983). Mental models. Cambridge, MA: Harvard University Press. Just, M. A. & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review, 99, 122-149. Kintsch, W. (1988). The role of knowledge in discourse processing: A construction-integration model. Psychological Review, 95, 163-182. Leather, C. V., & Henry, L. A. (1994). Working memory span and phonological awareness tasks as predictors of early reading ability. Journal of Experimental Child Psychology, 58, 88-11. Long, D. L., Oppy, B. J., & Seely, M. R. (1997). Individual differences in reader's sentence- and text-level representations. Journal of Memory and Language, 36, 129-145. McKoon, G., & Ratcliff, R. (1992). Inference during reading. Psychological Review, 99, 440-446. McNeill, D. (1970). The acquisition of language: The study of developmental psycholinguistics. New York: Harper and Row. Morrow, D. G., Greenspan, S. L. & Bower, G. H. (1987). Accessibility and situation models in narrative comprehension. Journal of Memory and Language, 26, 165-187. Muter, V. , Hulme, C., Snowling, M. J. & Stevenson, J. (in press). Phonemes, rimes, vocabulary and grammatical skills as foundations of early reading development: Evidence from a longitudinal study. Developmental Psychology.

Nagy, W. E., & Herman, P. A. (1987). Breadth and depth of vocabulary knowledge: implications for acquisition and instruction. In M. McKeown & M. E. Curtis (Eds.), The nature of vocabulary acquisition (pp. 19-36). Hillsdale NJ: Erlbaum. Nagy, W. E., & Scott, J. A. (2000). Vocabulary processes. In M. L. Kamil, P. B. Mosenthal, P. David Pearson, & R. Barr (Eds.), Handbook of reading research (Vol. III, pp. 69-284). Mahwah, NJ: Erlbaum. Nation, K., Adams, J. W., Bowyer-Crane, C., & Snowling, M. J. (1999). Working memory deficits in poor comprehenders reflect underlying impairments. Journal of Experimental Child Psychology, 73, 139-158. Nation, K., & Snowling, M. J. (1998). Semantic processing and the development of word recognition skills: Evidence from children with reading comprehension difficulties. Journal of Memory and Language, 39, 85-101. Nation, K., & Snowling, M. J. (1999). Developmental differences in sensitivity to semantic relations among good and poor comprehenders: evidence from semantic priming. Cognition, 70, B1-B13. National Institute of Child Health and Human Development. (2000). Report of the National Reading Panel. Teaching children to read: An evidence-based assessment of the scientific research literature on reading and its implications for reading instruction (NIH Publication No. 00-4769). Washington, DC: U.S. Government Printing Office. Neale, M. (1997). The Neale Analysis of Reading Ability II. Windsor: NFER-Nelson. Oakhill, J. (1993). Childrens difficulties in reading comprehension. Educational Psychology Review, 5, 223-237. Oakhill, J. V., Cain, K., & Bryant, P. (2003a). The dissociation of word reading and text comprehension: Evidence from component skills. Language and Cognitive Processes, 18, 443-468. Oakhill, J.V. Cain, K.E., and Bryant, P.E. (2003b, May). Prediction of comprehension skill in the primary school years. Paper presented at the biennial meeting of the Society for Research in Child Development, Tampa. Oakhill, J., & Garnham, A. (1988). Becoming a skilled reader. New York: Basil Blackwell. Oakhill, J., & Yuill, N. (1986). Pronoun resolution in skilled and less-skilled comprehenders: Effects of memory load and inferential complexity. Language and Speech, 29, 25-36. O'Donnell, R. C. (1974). Syntactic differences between speech and writing. American Speech, 49, 102-110. Olson, D. R. (1977). From utterance to text: The bias of language in speech and writing. Harvard Educational Review, 47, 257-281. Otero, J., & Kintsch, W. (1992). Failures to detect contradictions in a text: What readers believe versus what they read. Psychological Science, 3, 229-235. Perfetti, C. A. (1985). Reading ability. New York: Oxford Press.

Perfetti, C. A. (1991). Representations and awareness in the acquisition of reading competence. In L. Rieben & C. A. Perfetti (Eds.), Learning to read: Basic research and its implications (pp. 33-44). Hillsdale, NJ: Erlbaum. Perfetti, C. A. (1999). Comprehending written language: A blueprint of the reader. In C. Brown & P. Hagoort (Eds.), The neurocognition of language (pp. 167-208). Oxford University Press. Perfetti, C. A., & Goldman, S. R. (1976). Discourse memory and reading comprehension skill. Journal of Verbal Learning and Verbal Behavior, 14, 33-42. Perfetti, C. A., & Hart, L. (2002). The lexical quality hypothesis. In L. Vehoeven. C. Elbro & P. Reitsma (Eds.), Precursors of functional literacy (pp. 189-213). Amsterdam/Philadelphia: John Benjamins. Perfetti, C. A., & Hogaboam, T. W. (1975). The relationship between single word decoding and reading comprehension skill. Journal of Educational Psychology, 67, 461-469. Perfetti, C. A., & Lesgold, A. M. (1977). Discourse comprehension and sources of individual differences. In P. A. Carpenter & M. A. Just (Eds.), Cognitive processes in comprehension (pp. 141-183). Hillsdale, NJ: Erlbaum. Perfetti, C. A., Marron, M. A., & Foltz, P. W. (1996). Sources of comprehension failure: Theoretical perspectives and case studies. In C. Cornoldi & J. Oakhill (Eds.), Reading comprehension difficulties: Processes and intervention (pp. 137-165). Mahwah, NJ: Erlbaum. Plaut, D. C., McClelland, J. L., Seidenberg, M. S., & Patterson, K. (1996). Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103, 56-115. Seigneuric, A. S., Ehrlich, M. F., Oakhill, J. V., & Yuill, N. M. (2000). Working memory resources and childrens reading comprehension. Reading and Writing, 13, 81-103. Share, D. L. (1995). Phonological recoding and self-teaching: Sine qua non of reading acquisition. Cognition, 55, 151-218. Share, D. L. (1999). Phonological recoding and orthographic learning: A direct test of the self teaching hypothesis. Journal of Experimental Child Psychology, 72, 95-129. Smiley, S. S., Oakley, D. D., Worthen, D., Campione, J. C., & Brown, A. L. (1977). Recall of thematically relevant material by adolescent good and poor readers as a function of written versus oral presentation. Journal of Educational Psychology, 69, 381-387. Smith, M. K. (1941). Measurement of the size of general English vocabulary through the elementary grades and high school. Genetic Psychological Monographs, 24, 311-345. Snowling, M., Hulme, C., & Goulandris, N. (1994). Word recognition and development: A connectionist interpretation. Quarterly Journal of Experimental Psychology 47A, 895-916 . Stein, N. L., & Albro, E. R. (1997) The emergence of narrative understanding: Evidence for rapid learning in personally relevant contexts. Contemporary Issues in Education, 60, 83-98.

Stein, C. L., Cairns, J. S., & Zurif, E. B. (1984). Sentence comprehension limitations related to syntactic deficits in reading-disabled children. Applied Psycholinguistics, 5, 305-322. Stein, N. L., & Glenn, C. G. (1979). An analysis of story comprehension in elementary school children. In R. O. Freedle (Ed.), New directions in discourse processing (pp. 53-120). Hillsdale, NJ: Erlbaum. Sticht, T., & James, J. (1984). Listening and reading. In P. Pearson (Ed.), Handbook of research on reading (pp. 293-317). New York: Longmans. Stothard, S. E., & Hulme, C. (1992). Reading comprehension difficulties in children: The role of language comprehension and working memory skills. Reading and Writing, 4, 245-256. Trabasso, T., & Suh, S. (1993). Understanding text: Achieving explanatory coherence through online inferences and mental operations in working memory. Discourse Processes, 16, 3-34. Van Dijk, T. A., & Kintsch, W. (1983). Strategies of discourse comprehension. New York: Academic Press. Vosniadou, S., Pearson, P. D., & Rogers, T. (1988). What causes childrens failures to detect inconsistencies in texts? Representation versus comparison difficulties. Journal of Educational Psychology, 80, 27-39. Yuill, N., & Joscelyne, T. (1988). Effects of organizational cures and strategies on good and poor comprehenders story understanding. Journal of Educational Psychology, 80, 152-158. Yuill, N., & Oakhill, J. (1988). Effects on inference awareness training of poor reading comprehension. Applied Cognitive Psychology, 2, 33-45. Yuill, N., & Oakhill, J. (1991). Childrens problems in text comprehension. Cambridge University Press. Zwaan, R. A. (1996). Processing narrative time shifts. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 1196-1207. Zwaan, R. A., & Radvansky, G. A. (1998). Situation models in language comprehension and memory. Psychological Bulletin, 123, 162-185.