Beruflich Dokumente
Kultur Dokumente
Abstra t
Analyses of dialogue that in
orporate the insights of spee
h a
t theory presuppose that an ut-
teran
e gets asso
iated with a
onversational move type (CMT). Due to diÆ
ulties that beset
attempts to integrate CMTs into grammar in early generative work, as well as the per
eived
problems
on
erning multifun
tionality, CMT information is typi
ally not in
luded in most for-
mal grammati
al analyses. We provide arguments as to why CMT does need to be integrated in
grammati
al analysis of
onversation. We oer a proposal for su
h an integration
ou
hed in Head
Driven Phrase Stru
ture Grammar (HPSG). We sket
h explanations as to why our proposal does
not run into the foundational and empiri
al pitfalls that have beset previous proposals.
1 Introdu tion
Categorizing utteran
es in terms of a notion of illo
utionary for
e or
onversational move type
(CMT) is
ommon in
orpus-based work (for some re
ently proposed CMT taxonomies, see (Car-
letta et al., 1996), (Core & Allen, 1997)). Indeed any analysis of dialogue that in
orporates the
insights of spee
h a
t theory presupposes that an utteran
e ultimately gets asso
iated with a
CMT. Nonetheless, there exist few attempts to integrate su
h notions into
ontemporary formal
grammati
al work. In part, this is due to the fa
t that most grammati
al formalisms to date
have been designed with monologue or text in mind, where this issue is easier to put aside than
in
onversational settings. A more prin
ipled reason for this la
una is perhaps the phenomenon
of multifun
tionality (see e.g. (Allwood, 1995)): it is often the
ase that a given utteran
e serves
more than one purpose|an assertion
an fun
tion also as an oer, a query as a suggestion et
.
This has often led to the feeling that issues pertaining to CMT belong entirely to the realm of
pragmati
s. Although no worked out pragmati
theory as to how CMTs get assigned to utteran
es
has emerged to date, the one in
uential series of attempts to subsume CMT into the grammar,
based on the Performative Hypothesis (PH) is generally viewed to have been a resounding failure
(see (Levinson, 1983), pp. 247-263).
In this paper we argue that CMT
an and should be integrated in the semanti
analyses
provided by the grammar. That is, CMT is a parameter of meaning
onventionally asso
iated
with
ertain words and
lasses of phrases. For instan
e, in hearing an utteran
e by A of a senten
e
su
h as (1a), we
laim that a
ompetent interlo
uter B knows that its meaning is the template
2
BI-DIALOG 2001 (P. K uhnlein, H. Rieser, & H. Zeevat (eds)).
Paper 1, Copyright
2001, Jonathan Ginzburg and Ivan A. Sag and Matthew Purver
fginzburg,purvergd
s.k
l.a
.uk
sag
sli.stanford.edu
http://www.d
s.k
l.a
.uk/ fstaff/ginzburg, pg/purver g
http://www-
sli.stanford.edu/~sag.
s
hemati
ally given as (1b), not simply the proposition (1
). That is, B knows that in order to
ground A's utteran
e she must try to instantiate the parameters A; t; l; P within the template
given in (1b) in su
h a way as to satisfy the
onstraints provided by the grammar (e.g. A must
be the speaker, t must be a time the day after utteran
e time, P ranges over a set that in
ludes
fassert, threaten, promise,: : :g, but not over, for instan
e, fask, ex
laim, apologize,: : :g)1
Although there are a variety of versions of the PH, they essentially boil down to positing that all
(English) matrix senten
es have the form I illo
-verb S, where I is the rst person singular pronoun
and illo
-verb is a verb from the
lass of performative verbs (e.g. assert, ask, order, bet, : : :). For
all matrix senten
es whi
h do not have this form overtly, the PH involves the assumption that
the `illo
utionary prex' I illo
-verb is not realized at the surfa
e but is represented at some other
synta
ti
level. In its formulations in the 1970s, at least, the PH ran into a variety of problems,
the most serious of whi
h revolved around the diÆ
ulty of maintaining a
oherent denition of
truth for de
laratives. The diÆ
ulty arises from the parallelism that the PH enfor
es between
senten
es that la
k an overt illo
utionary prex (e.g. (2a)) and expli
it performatives (e.g. (2b)):
(2) a. Snow is bla
k.
b. I
laim that snow is bla
k.
Su
h a parallelism is untenable be
ause it either
on
ates the truth
onditions of quite
on-
tingent senten
es su
h as (2a) with those of (2b), whi
h, essentially, be
ome true on
e they are
uttered. Alternatively, the parallelism requires a mysterious ltering away of the semanti
ee
t
of the illo
utionary prex. Despite the diÆ
ulties for the PH, we argue that in fa
t there are
good reasons for assuming that the
ontents spe
ied by the grammar do
ontain CMTs as a
onstituent. Our rst argument
on
erns the existen
e of words that a
tually
arry their CMT on
their sleeve. Examples of su
h words are given in (3):
(3) a. [Context: A sees B as she enters a building℄ A: Hi.
b. [Context: A enters train
arriage, sees B leave℄ A: Bye.
. [Context: in a bus queue A slips and unintentionally pushes B℄ A: Sorry.
d. [Context: B is a bus
ondu
tor who gives A a ti
ket.℄ A: Thanks.
A
ompetent speaker of English might paraphrase ea
h of these utteran
es as in (4):
1 How any of these values get instantiated, if indeed B manages to do so,
an involve highly
omplex reasoning
(involving e.g. domain-spe
i
knowledge, reasoning about intentions et
) with whi
h of
ourse the grammar as
su
h provides no assistan
e. However, the use of su
h reasoning to resolve the value of a
onstituent of
ontent also
ae
ts
onstituents of
ontent (e.g. tense and anaphora) that lie un
ontroversially within the realm of semanti
s.
Hen
e, this
annot be used as an argument against integrating CMT within grammati
al analysis.
3
(4) a. A greeted B.
b. A bid farewell to B.
. A apologized to B (for having pushed her).
d. A thanked B (for giving her a ti
ket).
This
an be used as eviden
e that these words are asso
iated with meanings s
hematized as
in (5). In these representations, the main predi
ate
onstitutes the CMT asso
iated with the
utteran
e, whereas m(es)s(a)g(e)-arg indi
ates the semanti
type of the propositional/des
riptive
ontent sele
ted by the CMT. Note a
ontrast illustrated in (4): whereas both [the relations
denoted by℄ apologize and thank sele
t
lausal
omplements (whose denotations)
onstitute the
des
riptive
ontent, there is no su
h sele
tion by greet and bid-farewell. This provides some of the
motivation for assuming that these latter should not spe
ied for a msg-arg, in other words that
su
h spee
h a
ts have no des
riptive
ontent.
(5) a. Hi: greet(speaker,addressee,msg-arg:none)
b. Bye: bid-farewell(speaker,addressee,msg-arg:none)
4
. A: Did Belula resign? B: Did WHO resign? (unambiguously: Whoi are you asking
whether i resigned?)
If su
h paraphrases are the
orre
t basis for an analysis of su
h utteran
es, this indi
ates
that in reprise utteran
es at least CMT (the CMT of the pre
eding utteran
e, to be pre
ise)
an
be
ome a
onstituent of the des
riptive
ontent of an utteran
e.3 In other words, CMT be
omes
a
onstituent of the
ontent the grammar in
ontrovertibly needs to build up.
In fa
t, following (Ginzburg & Sag, 2000), we suggest that reprise utteran
es provide a probe
that allows one to lter away the indire
t for
e of an utteran
e and establish a single dire
t CMT
with a given utteran
e.4 Consider (7), uttered outside a West End theater
urrently showing a
best selling musi
al:
(7) (1) Stina: I have a ti
ket for tonight's performan
e.
(2) Padraig: You have a ti
ket for tonight's performan
e?
(3) Stina: Yes.
is a re
ipe for disaster, as the robot will hassle any approa
hing theatre-goer, rather than solely
loudly de
laiming touts.
3 This
laim was originally made, independently, by (Ginzburg, 1992) and (Ja
obs, 1991).
4 Using reprises as su
h a probe was rst suggested to us by Ri
hmond Thomason in an oral dis
ussion that
followed presentation of (Ginzburg & Sag, 1999).
5 Our dis
ussion of these data is of ne
essity all too brief. As dis
ussed in (Ginzburg & Cooper, 2001), reprises
exemplify an additional reading dubbed the
onstituent-reading, whi
h involves a request for reformulation of the
import of the reprised (sub)-utteran
e. Thus, for an referential NP utteran
e, as in (i), this will be understood as
a request for referen
e resolution:
(i) Andie: Did Jo leave?
Bo: Jo?
Andie: Your
ousin.
Given this, reprises su
h as (7[2℄) will also yield readings paraphrasable as (ii), where the inferred
omponent of
ontent is not ne
essarily ltered away:
(ii) Shi: What do you mean by saying you have a ti
ket for tonight's performan
e?
yes, however, is an inappropriate response to this reading.
5
3 Integrating CMT into a
onstraint-based grammar
We adopt a version of HPSG developed in (Sag, 1997; Ginzburg & Sag, 2000). The
ontent asso-
iated with signs, phrasal or lexi
al, is drawn from a situation theoreti
ontology. The ontology
distinguishes inter alia questions, propositions, fa
ts, situations/events, and out
omes. Informa-
tion about phrases is en
oded by
ross-
lassifying them in a multi-dimensional type hierar
hy.
Phrases are
lassied not only in terms of their phrase stru
ture s
hema or X-bar type, but also
with respe
t to a further informational dimension of
lausality. Clauses are divided into inter
alia de
larative
lauses whi
h denote propositions, interrogative
lauses denoting questions, ex
la-
mative
lauses denoting fa
ts, and imperative
lauses denoting out
omes. Ea
h maximal phrasal
type
an inherit from both these dimensions. This
lassi
ation allows a spe
i
ation of systemati
orrelations between
lausal
onstru
tion types and types of semanti
ontent.
We note two
onsiderations that an a
ount integrating CMT information into the grammar
needs to heed:
In order to avoid the problems asso
iated with the PH, one has to ensure that the way in
whi
h CMT information enters into the
ontent of a sign does not ae
t the assignment of
(non-CMT)
ontent. One must also ensure that a sign that has CMT information (of the
urrent utteran
e)
annot be embedded as a daughter of another sign.
In order to des
ribe reprise utteran
es, one must have the means to let signs with CMT
information be inputs to grammati
al
onstraints, e.g. to build questions whose queried
proposition
ontains CMT information.
We will satisfy these requirements by making a ner grained distin
tion than usually made
with respe
t to \matrix" (non-embedded) signs. Whereas all signs that
annot be
omplements
of an embedding predi
ate bear the spe
i
ation i(ndependent)
(lause):+, we will introdu
e
a further partition among su
h signs, depending as to whether or not they
an play a role in
re
ursive operations of the grammar. Those that
annot will be designated as root:+. Before we
an illustrate how this a
tually works, we need to bring CMTs into the pi
ture.
Our approa
h is
onsistent with various ontologies of CMTs. The minimal su
h ontology one
ould posit involves a 1{1 relationship between what is often
alled the
ontent of a sign, i.e.
entities of type message (proposition, question, out
ome, fa
t, : : :) and CMTs: propositions are
asso
iated with the CMT of asserting, questions with asking, out
omes with ordering, and fa
ts
with ex
laiming. This involves positing a type illo
(utionary)-rel as the immediate supertype of
these four CMTs:
(9) illo
-rel
Arguably, su
h a relationship between message types and CMTs
onstitutes something like
a default. But ea
h of the afore-mentioned subtypes of message
learly does have other uses:
questions
an be used `rhetori
ally' (also known as a reassertion of a resolved question), out
omes
an be suggested, propositions
an feature in threats and so on. Thus, an adequate view of
6
utteran
e
ontent needs to allow for a ri
her ontology of CMTs and for the CMT asso
iated with
a given message-type to be underspe
ied. This renement is easy to implement by (a) positing
more maximal subtypes of illo
-rel (e.g. threat-rel, promise-rel, reassert-rel et
) and (b) positing
types intermediate between illo
-rel and the leaves of the hierar
hy in (9) (e.g. a type prop-illo
-
rel whi
h would subsume all propositional CMTs|assert-rel, threat-rel, promise-rel et
.). In this
abstra
t, as in our implementation at present, we maintain the more simplisti
view, enshrined in
(9).
The nal ingredient we need as far as phrases go is a
onstraint that determines the appropriate
ontent value for utteran
es, i.e. for root
lauses. We propose that the
ontent of every root
lause be a proposition whose soa value is of type illo
-rel. This proposition represents the belief
an agent forms about the (full, dire
t illo
utionary)
ontent of an utteran
e. More spe
i
ally, this
is the
ontent a speaker will assign to her utteran
e, as will an addressee in
ase
ommuni
ation is
su
essful. Given (9), this will mean that a root
lause will be resolved so as to have as its
ontent
a proposition whose soa value is of one of the subtypes of illo
-rel.
In order to ensure that root
lauses have
ontents in whi
h CMT information is represented,
we posit a type root-
l and propose a
onstru
tional treatment of root utteran
es in terms of a
non-bran
hing phrasal type (hd-only-ph) that embeds message-denoting senten
es as arguments
of an illo
-rel. The
onstraints idiosyn
rati
to this
onstru
tion, akin to a `start' symbol in a
ontext free grammar, are illustrated in (11):6
(11) root-
l:
2 3
root +
6 v 7
6head 7
6 vform n 7
6 2 37
6 proposition 7
6 7
6 6 2 377
6 6 illo
-rel 77
6
ont 6 1 777 2 3
6 6
6soa 4utterer 77
6 3 557 root
6
6
4 addressee
msg-arg 2 7
7 ! H4i
+5
6 2 3 7
ont 2
6 7
6
-spkr 1 7
6
6
ntxtj
-indi
es 4
-addr 35 7 7
6 4 7
6
-time 7
6 7
6slash f g 7
4 5
store fg
wh fg
Note that the arguments of the illo
-rel are identied with the appropriate individuals in the
ontext of utteran
e. As mentioned above, we now distinguish root
lauses from other independent
lauses in terms of positive versus negative spe
i
ations for the feature root.7
Let us illustrate the ee
ts of the
onstraint in (11). (12a) has an analysis as a polar question
in whi
h it expresses the question in (12b). Therefore, given (10) and (11), the
ontent su
h a
lause gets as a root utteran
e (ignoring tense) is (12
):
6 The
onstraint here relates the mother to its (sole) daughter, denoted with a large bold fa
ed H.
7 On this view, signs are [root ℄ by default. Sin
e this is the
ase, we will suppress [root ℄ spe
i
ations
on all phrases other than instan
es of the type root-
l.
7
(12) a. A: Did Jo leave?
b. 2question 3
6params fg
2
7
37
6
6 proposition 7
6 7
6 6sit 2 77
6 6 2 377
6prop 6 leave-rel 77
6 6 77
4 4soaj nu
l 4leaver 3 555
time 4
b
kgrd named( 3 ,Jo), t-pre
ede( 5 , 4 )
2 3
-spkr 1
j
ntxt
-indi
es 4
-addr 65
-time 5
2 3
. proposition
6sit 0 7
6 2 37
6
6
quants hi 7
6 6 2 377
7
6 6 ask-rel 77
6
6
6
6 6utterer 1 7
777
6 6 6 6 777
6 6addressee 7 7
6 6 2 3777
6 6 6 question 777
6 6 777
6params f g
6 6
6 6 7777
6soa 6 6 6 2 7
377 7
6 6nu
l 6 6 proposition 7777
6 6 6 6 7777
6 6 6msg-arg 6 6sit 2 77777
6 6 6 6 6 2 377777
6 6 6 6prop 6 leave-rel 77777
6 6 6 6 6 77777
6 6 4 4 4soaj nu
l 4leaver 3 555577
4 4 55
time 4
b
kgrd named( 3 ,Jo), t-pre
ede( 5 , 4 )
2 3
-spkr 1
j
ntxt
-indi
es 4
-addr 65
-time 5
So far we have fo
ussed on phrases. However, one
an within this approa
h equally des
ribe
words su
h as those dis
ussed in (3){(5). For instan
e, the word hi
an be des
ribed as follows:8
2 3
(13) phon
2
h hi i 3
" #
6 7
6 interj 7
6 6
at head 77
6 6 i
+ 77
6 6 7
6 6 2 377
7
6 6 proposition 77
6 6 6 0 77
6 6 sit 77
6ssjlo
6 6 2 3777
6 6 6 quants h i 777
7
6 6 6 2 3777
6
ont 6 6 greet-rel 7 7
6
6 6 6 6 777
7
7
6 6 6soa 6 6 i 77777
7 7
utterer
6 6 6 7
4addressee j 5557
6 nu
l 7
6 4 4 4 57
6 time k 7
6 7
6 7
6arg-st h i 7
6 2 3 7
6
-spkr i 7
6 7
4
txtj
-indi
es 4
-addr j 5 5
-time k
8 A lexi
al entry for e.g. good morning would involve adding the assumption in b
kgrd that the time k is lo
ated
before noon.
8
Note that hi is spe
ied as i
:+, whi
h means that it
annot be embedded. However, it
is underspe
ied for root. This means, as we will soon see, that this entry
an be the head-
daughter of a reprise
onstru
tion (re
e
ting
ases su
h as [Context: A is a
rusty brigadier, B a
raw re
ruit℄ B: Hi. A (growls): Hi? (= Are you greeting me)).9; 10
Finally, we explain brie
y how the CMT of the previous utteran
e enters as a
onstituent
of the
ontent of
ertain reprise utteran
es. We assume the a
ount developed in (Ginzburg
& Cooper, 2001) of how
lari
ations arise during attempted integration of an utteran
e in a
onversationalist's information state (IS). Simplifying somewhat, on this view a ne
essary
ondition
for B to ground an utteran
e by A is that B manage to nd values for the
ontextual parameters
of the meaning of the utteran
e. What happens when B
annot or is at least un
ertain as to how
he should instantiate in his IS a
ontextual parameter i? In su
h a
ase B needs to do at least
the following: (1) perform a partial update of the existing
ontext with the su
essfully pro
essed
omponents of the utteran
e (2) pose a
lari
ation question that involves referen
e to the sub-
utteran
e ui from whi
h i emanates. Sin
e the original speaker, A,
an
oherently integrate a
lari
ation question on
e she hears it, it follows that, for a given utteran
e, there is a predi
table
range of < partial updates +
onsequent
lari
ation questions>. These we take to be spe
ied by
a set of
oer
ion operations on utteran
e representations.11 Indeed we assume that a
omponent
of dialogue
ompeten
e is knowledge of these
oer
ion operations.
One su
h operation is dubbed parameter fo
ussing by (Ginzburg & Cooper, 2001). This involves
a (partially updated)
ontext in whi
h the issue under dis
ussion is a question that arises by
instantiating all
ontextual parameters ex
ept for i and abstra
ting over i. In su
h a
ontext,
one
an
onrm that i gets the value B suspe
ts it has by uttering with rising intonation any
apparently
o-referential phrase whose synta
ti
ategory is identi
al to u1's (see (6a,b) above).
One
onstru
tion type appropriate for this
ontext are reprise interrogative
lauses (repr-int-
l).
In the framework of (Ginzburg & Sag, 2000) they are des
ribed by means of the following s
hema:
(14)2repr-int-
l 3
2 " #3
2 3 +
6 question 7 i
6 7 6
at head
n 7
6
ont 4params 3 57 6 vform 7
6
6 prop
1 soajnu
ljmsg-arg
7
2 7 ! H 6
6
ont 2
7
7
6 7 6 7
6
4store 1 7
5
4store ( 1 ℄ 3 ) 5
2
b
kgrnd prev-utt( 0 ),
ont( 0 , 1 ) ℄ 2 b
kgrd
To illustrate this: a reprise of (12a)
an be performed using (15a). This
an be assigned the
ontent in (15b) on the basis of the s
hema in (14):12
9 As with all CE utteran
es, this one
an be understood in a number of ways. In this
ase, the
onstituent
reading alluded to in footnote 5 is possibly even more prominent. It would yield a reading paraphrasable as what
do10you mean by saying hi to me.
Underspe
ifying hi for root might suggest that it
ould fun
tion as the head daughter in (11), thereby yielding
an unwanted reading I assert that I greet you. However, in the framework of (Ginzburg & Sag, 2000) all headed
phrases are subje
t to the Generalized Head Feature Prin
iple (GHFP), whi
h involves the synsem value of the
mother of a headed phrase and that of its head daughter being identi
al by default. This means that the head
daughter of a root-
l is spe
ied to be
atjhead:v[n℄; hi (as its relatives bye, sorry, thanks et
) is spe
ied as
atjhead: interj, and hen
e
annot serve as the head daughter of a root-
l.
11 The term
oer
ion operation is inspired by work on utteran
e representation within a type theoreti
framework
reported in (Cooper, 1998).
12 Note that the previous utteran
e identied the utterer of the ask-rel with the speaker of that utteran
e (this is
ensured by the
onstraint in (11) on the type root). Hen
e, the utterer of the ask-rel in the
ontent of the reprise
must also be that individual, indi
ated as spkr 0 in (15).
9
(15)
Did WHO leave?
S
2 3
repr-int-
l
6store f g 7
6 2 37
6 question 7
6 ( 7
6 6 ) 77
6 6 ind k 77
6 6params 4 77
6 6 restr fperson(k )g 77
6 6 77
6 6 2 2 3377
6
6
6 ask-rel 77
6 77
6 6 6 6utterer spkr 0 7777
6 6 6 6 2 77
37777
6 6 6 6 question 77
6
ont 6 6 6 7777
6 6 6 6 6 77
77777
6 6 6 6 6
params
7 77
6 6 6 6 6 2 377777
6prop 36 6soajnu
l 6
6 6 6 proposition 77777
6 6 6 77777
6 6 6 6msg-arg 26 6sit s 777777
6 6 6 6 6 2 3777777
quants h i
6 6 6
6 6 6prop 6 7 7777
6 6 6 6 6 6 777777
6 6 6 6 4 4 soa 4 leave-rel 557
57777
6 4 4 4 nu
l 5557
6 leaver k 7
6 7
4 5
b
kgrnd prev-utt( 0 ),
ont( 0 , 3 )
S
2 3
pol-int-
l
4store f 4 g5
ont 2
V NP V
In this paper, we have presented a number of arguments that indi
ate the need to integrate
CMT information in grammars intended to analyze
onversational intera
tion. One su
h argu-
ment
on
erns the proper analysis of words su
h as hi, thanks, sorry whi
h
an stand alone as
omplete utteran
es. Another arguent derives from the
onsideration of reprise utteran
es. We
have sket
hed brie
y the basi
s of an HPSG in whi
h CMT information is integrated. This gram-
mar has been implemented as part of the SHARDS system (Ginzburg, Gregory, & Lappin, 2001).
In future work we hope to show how grammars of this type
an, when integrated with domain
knowledge, oer insightful solutions to the many puzzles posed by multifun
tionality.
A knowledgements
We would like to thank three anonymous BIDIALOG reviewers for very useful
omments. The
resear
h des
ribed here is funded by grant number R00022269 from the E
onomi
and So
ial Re-
sear
h Coun
il of the United Kingdom and by grant number GR/R04942/01 from the Engineering
and Physi
al S
ien
es Resear
h Coun
il of the United Kingdom.
10
Bibliography
11