Biologically Inspired Cognitive Architectures (2016) 17, 22– 31

Available at www.sciencedirect.com

ScienceDirect

journal homepage: www.elsevier.com/locate/bica

RESEARCH ARTICLE

Analyzing and discussing primary creative

traits of a robotic artist
A. Augello, I. Infantino, A. Manfré, G. Pilato, F. Vella *

National Research Council of Italy (CNR), Via Ugo La Malfa, 153, 90146 Palermo, Italy

Received 11 June 2016; accepted 26 July 2016

KEYWORDS
Cognitive architecture;
Computational creativity;
Humanoid robot
Abstract
We present a robot aimed at producing a collage formed by a mix of photomontage and digital
collage. The artwork is created after a visual and verbal interaction with a human user. The
proposed system, through a cognitive architecture, allows the robot to manage the three dif-
ferent phases of the real-time artwork process: (i) taking inspiration from information captured
during the postural and verbal interaction with the human user and from the analysis of his/her
social web items; (ii) performing a creative process to obtain a model of the artwork; (iii) exe-
cuting the creative collage composition and providing a significant title. The paper explains,
primarily, how the creativity traits of the robot are implemented in the proposed architecture:
how ideas are generated through an elaboration that is modulated by affective influences; how
the personality and the artistic behavior are modeled by learning and guided by external eval-
uations; the motivation and the confidence evolution as a function of successes or failures.
Ó 2016 Elsevier B.V. All rights reserved.

Introduction also comprises the artificial branch named computational

In 1961, Rhodes (1961) discussed in a famous article some
basilar principles of creativity, proposing the relevant
strands that have influenced and inspired researchers on
this topic. Rhodes, analyzing the human creativity, individ-
uated the four strands that are well-known as the big Ps:
people, processes, press, products. Nowadays, the scientific
field dealing with creativity is huge and variegated, and it
* Corresponding author.
E-mail address: filippo.vella@pa.icar.cnr.it (F. Vella).
creativity (Colton, Pease, & Charnley, 2011; Wiggins,
2006). Our approach to artificial creativity, in the present
work, has various points of contact with Rhodes analysis.
The artificial creative agent has to exhibit cognitive capabil-
ities that are typical of a human being, such as personality,
habits, attitudes, and so on. The production of an artwork is
the result of a creative act that depends on the experience
and expertise of the artist and that it is strongly influenced
by motivation, emotions, learning and communicative
capabilities.

http://dx.doi.org/10.1016/j.bica.2016.07.006
2212-683X/Ó 2016 Elsevier B.V. All rights reserved.
Analyzing and discussing primary creative traits of a robotic artist
Products are the tangible outcomes of ideas and they are
characterized by the media of expression, by the emotional
impact on the subject who experiences the artwork, and by
their intrinsic value (e.g. aesthetic, utility, and so on).
In recent years, robotic systems are available at a
reduced cost and with increased performances allowing
cognitive researchers to experiment their developed models
in embodied artificial agents and exploring different artistic
domains such as dance, music, drama, painting. On the
other hand, robotic researchers have the possibility to test
the robots social behaviors in the real environment while
they interact with humans by means of suitable artificial
cognitive architectures (CAs). We think that a CA is manda-
tory to establish a meaningful connection between robots
and humans to express a concept or transmit a message.
The quality of the artistic creation can be bound to the
assessable bunch of meaning that is possible associate to
an artistic piece (Pound, 1934). In this paper we consider
the creation of an artwork as a collage whose meaning
can be obtained with the proper composition of elementary
items.
While some skillful aspects may be undertaken by the
computer, the intent that deals with the creation of the
artefact and its related meaning depends on the developer
who drives the process; this is one of the reason why sys-
tems undertaking creative tasks might be perceived as not
being creative (Krzeczkowska, El-Hage, Colton, & Clark,
2010). A system that creates an output endowed with an
intent has been proposed by Krzeczkowska et al. (2010)
and Cook and Colton (2011). The system searches for news
articles in order to find information that is related with a
given concept, then, a set of images are selected and they
are properly put together in a collage by exploiting services
of image retrieval and image processing.
Similarities and analogies play also a key role in the gen-
eration of new ideas that is at the basis of creative pro-
cesses (Indurkhya & Ogawa, 2012; Stojanov & Indurkhya,
2013). In the specific case of collage creation, they are fun-
damental for the selection of the items to compose. An
agent, interacting with the environment, stores in its mem-
ory the most salient aspects of its perceptions. The stored
concepts can be recalled when the agent experiences some-
thing similar, also if the new experience involves different
sensing quantities.
In this paper, the humanoid starts the collage creation, it
informally takes inspiration, from a verbal interaction with
a human user and from the collection of web items that
she/he has published on his social network profile. The
inspiration phase is strongly influenced by the semantic
associations between the humanoid perceptions and the
elements of the creative language of the artist. In particular
the symbols of the artistic language used by the collagist are
pictorial elements that are stored in its long term memory
and can be used to compose the collage. The inspiration is
characterized by evoking pictorial elements which are con-
sidered similar to the information acquired from the user
through a dialog. The dialog also triggers an emotional
empathy in the robot, that influences its artwork. The artist
composes, in an original manner, the ‘‘basic terms” of its
artistic language, painting them with palettes that are
related to the emotions that have been detected during
the interaction with the user. When the collage is
23
completed, the robot creates a title for the artwork through
an analysis of the concepts arising during the dialog phase.
The paper is structured as follows: the first section
describes how the cognitive architecture supports the cre-
ative process and the inspiration phase; then we describe
the artwork creation process and finally, in the last section
we show some experiments.
Cognitive architecture supporting collage
creation
The creative behavior of the digital artist is modeled
through a cognitive architecture depicted in Fig. 1. The dig-
ital artist is embodied into the humanoid robot NAO Alde-
baran. The NAO is able to draw very simple geometric
shapes and to interact with the user by means of a natural
language dialog implemented by a chatbot. The collage
allows the robot to summarize and express ideas about
the people it interacts with and to compose a set of
selected pictorial items. This technique allows the robot
to expand its limited physical capabilities and to produce
a more elaborated artistic result.
The creation of the artwork is based on two phases, an
inspiration phase, where the robot processes the sensory
inputs and as consequence it selects from its LTM the ele-
ments and the most adequate styles to use for the artwork
creation, and an execution phase where the robot physically
draws the artistic work.
The inspiration phase is based on the interaction with
people, in agreement with the hypothesis that human cre-
ativity only makes sense when placed in the context of rich
social interactions as shown by Csikszentmihalyi (1988) and
Sawyer (2008, 1999).
In this first phase, the robot is endowed with some set of
basilar sketches and artistic styles, which constitutes its
basic repertoire. The repertoire can be extended with other
pictorial elements provided by the user during its interac-
tion with the robot (for example giving some image pub-
lished into its social network profile). The items (sketches,
artistic styles, and images provided by the user) stored in
the robot knowledge base constitute the basic ‘‘terms” of
its artistic language. All the natural language descriptions
of the artistic repertoire are mapped into a semantic space,
built through the well known Latent Semantic Analysis (LSA)
approach, that is well founded for both cognitive (Landauer,
Foltz, & Laham, 1998) and statistical (Pilato & Vassallo,
2015) aspects. The use of a semantic space allows the digi-
tal artist to enrich its creation with associative/intuitive
capabilities. This is a key point, since association and intu-
ition are two peculiar features of human beings in general
and artists in particular. In the proposed architecture what
the user says is mapped in the semantic space. This informa-
tion triggers the activation of some basic items of the artis-
tic repertoire of the robot, according to a semantic
similarity measure. This representation allows the robot to
choose those terms that best fit user profile (or loosely
speaking his/her personality).
The recognition of human emotions through a text anal-
ysis, influences also the creative process of the robot. In
particular, by means of an introspective process, the robot
is able to associate a specific color palette to the perceived
24
Fig. 1 The cognitive architecture of the robotic painter.
emotions. It can use this palette to emphasize the message
represented by the collage. An interaction phase is also per-
formed allowing the robot to draw on the same surface with
the user and producing a collaborative drawing.
During the execution phase, the knowledge selected dur-
ing the interaction phase is analyzed and transformed
according to the creative styles. Then, it is properly incor-
porated and composed into a collage.
A motivational block influences the creation of the col-
lage. The motivation is implemented as a parameter that
is changed through an evaluation process (both internal
and external) through the so called urges, i.e. components
of the PSI architecture, introduced to reflect specific
demands, like physiological (Physiological Well-Being),
social (Affiliation) and cognitive drives (Competence and
Certainty). The motivations are among the most important
characteristics of the model. The Urges blocks manage the
motivations that justify the actions of the agent or modu-
lates the response of the agent to the input signals. The
Urges not only modulate the reactions of the agent but also
push the agent to find new solutions to the problems and
change the importance of the sensory inputs. A detailed
description of motivational process is reported by Augello,
Infantino, Pilato, Rizzo, and Vella (2013), while in this work
we focus on the inspiration and the creative components of
the proposed model.
Artistic inspiration
The painting process is driven by the motivations and by a
set of items the artist can take from in its long term memory
or from the interaction with the user. The digital painter
can either physically stay in front of the people (for exam-
ple during artistic performances), or it can also retrieve
people information by analyzing their social networks posts.
A. Augello et al.
The artworks will be a creative interpretation of the user’s
interests, the conversation topics, his moods and emotions.
This information can be either explicitly expressed by the
user during the conversation or implicitly contained in his
social networks profile. The humanoid exploits text mining
and emotion recognition algorithms to analyze this implicit
information. The interpretation of these perceptions is
strongly influenced by the personality of the artist, by its
emotions and in particular by his previous experiences.
The digital artist has learned during his experience to asso-
ciate colors palettes with emotions. It has also learnt to
associate images with semantic elements, such as words,
sentences, conversation topics and user interests.
The association between images and semantic elements
is obtained by mapping each image into a semantic space
through the vector representation of its corresponding tex-
tual description. As an example, the robot can analyze the
comments inserted by the user when he publishes an image
on the wall of his Facebook page.
An emotion recognition algorithm allows the robot to
identify the user emotions. The perceived emotions will
determine the colors palettes used by the artist.
In what follows, a detailed description of these ‘‘inspira-
tional” processes is given.
Association between sketches and text
The basic items the humanoid can draw during the interac-
tion phase are the sketches that are stored in its Long Term
Memory. These graphical items are also bound to semantic
elements such as words and sentences. This association is
obtained by mapping both images and semantic elements
into a same semantic space through their textual descrip-
tions retrieved by internet or provided by the user. A seman-
tic space is a sub-symbolic representation of natural
language elements, such as words, sentences, documents,
Analyzing and discussing primary creative traits of a robotic artist
in a high dimensional vector space. The natural language
elements represented into this model are considered being
semantically related according to the proximity of their cor-
responding vectors in the space. This concept has been sum-
marized by Sahlgren (2005) in a geometric metaphor of
meaning: meanings are locations in a semantic space, and
semantic similarity is proximity between the locations.
In this work we exploit the latent semantic analysis
methodology; this technique claims a human like general-
ization capability that is similar to induction processes
obtained from associative, semantic, and neural network
models, and it is capable of simulating several human cogni-
tive capabilities such as word-categorization, sentence-
word semantic priming, discourse comprehension
(Landauer et al., 1998). Latent Semantic Analysis (LSA) is
based on a dimensionality optimization of a semantic space
by means of the truncated singular-value decomposition
(SVD) technique (Berry, Dumais, & O’Brien, 1995), which
captures indirect, latent relations between elements of
the space overcoming the issues due to the synonymy prop-
erties of natural language. We assume that this methodol-
ogy can be used to simulate the process of visual
evocation from text that we can model in our humanoid
(Pilato, Vella, Vassallo, & La Cascia, 2010).
We consider a semantic space built by applying the LSA
methodology on a corpus of general culture document, in
particular we have chosen the space General Reading up
to 1st year college available at the http://lsa.colorado.
edu/ page. The visual knowledge of the humanoid, is com-
posed by sketches of its repertoire and the images acquired
by the analysis of the user social profile is sub-symbolically
coded in the semantic space. In this first phase, in order to
start with a first drawing repertoire we have embedded in
its KB the dataset of 20,000 drawings used in Eitz, Hays,
and Alexa (2012). Each piece of the artistic repertoire is
described by a title. The graphical KB of the robot is
extended with the other pictorial elements provided by
the user together with a textual description, for example
a set of images published in the user social profile.
The artistic items are sub-symbolically coded by mapping
their textual descriptions in the semantic space by means of
the folding-in technique. This mapping is exploited by the
robot to evaluate the similarity between the sketches
descriptions and other texts, such as words and sentences
introduced by the user interacting with him or the posts
of the user social profile. The similarity is evaluated by
the cosine distance between the vector associated with a
text and the vector associated with the images. This method
allows the artist to exploit the knowledge of the semantic
space to evoke specific images from a text analysis. Consid-
ering the previous example, it can associate terms such as
flight, fly, pilot and plane with the sketch of an aeroplane
described by the sentence ‘‘flying aeroplane”.
Associations among colors and emotions
A fundamental step in the artwork creation is the choice of
the palette to express and communicate its message. The
choice of the right color set is crucial since it has a relevant
contribution on the emotional impact of the artworks. In our
system, the choice of the most suitable palette of colors
depends on the emotions triggered by the user to the digital
25
artist. We have assumed that the use of color can be com-
pared to the use of a particular tone in a verbal communica-
tion. Therefore the artist must be able: (1) to perceive the
emotions transmitted by the users, (2) to choose a color pal-
ette that can express such an emotional content.
For what concerns the first point, the semantic content
of the conversation (or the content of the user’s social
page) can be related to emotions exploiting a methodology
for emotion recognition from texts. In particular we have
used a Naive Bayesian classifier properly trained on an emo-
tions lexicon to classify a text according to the six Ekmann
fundamental emotion classes. In particular we consider
the lexicon proposed in Strapparava and Mihalcea (2007)
and in Strapparava and Mihalcea (2008). The lexicon is com-
posed of six lists of affective words, each one characterizing
one of the Ekmann emotion classes, extracted from Word-
Net Affect (Strapparava, 2004). As result, we mapped the
semantic content of the user as a point, that we named
emo-knoxel, into a vector space that we named the emo-
tional space. This space is characterized by six dimensions
corresponding to the six fundamental emotions. To each
interaction and to each social profile analysis is then associ-
ated an emo-knoxel and a label representing the emotion
having the strongest component in the correspondent
emo-knoxel.
For what concerns the second point, the robot can
decide which palette to use depending on the emotion that
it desires to communicate through the artefact. The experi-
ence concerning palette-emotion associations is derived
from the analysis of data with textual description available
on the web.
Since color palettes are also characterized by a textual
description, this allows to map also a color palette into
the same emotional space, coding it as an emo-knoxel with
the same technique. The distance between the emo-knoxel
representing what the user says and the emo-knoxel repre-
senting a palette is therefore calculated. This allows to
choose the most appropriate color palette that is that closer
to the user profile emo-knoxel.
A dataset has been built by exploiting data available at
the COLOURlovers web site page,1 a creative community
for graphic designers aimed to the creation and sharing of
colors, palettes and patterns. The palettes shared by the
users of the community are described by a title and a set
of tags summarizing their main features. We hypothesize
that when a user creates a specific palette, she/he often
tries to represent a particular mood, as a consequence sev-
eral descriptive tags are words with an emotional content
that can be used for our mapping.
Interaction module
Our humanoid is able to interact with the user by means of
the ALICE chatbot technology [31]. The user questions are
compared with the information stored in the knowledge
base (KB) of the chabot, formalized according to the AIML
(Artificial Intelligence Mark-up Language) language [31].
The AIML KB is organized into a set of question-answer mod-
ules, called categories, where each category is composed of
a pattern tag, used to define a possible pattern of question
that could be formulated (expresses) by the user, and a
1
http://lsa.colorado.edu/.
26 A. Augello et al.

template tag, representing the appropriate answer to such
question. The tag star is used to match and store the text
introduced by the user in a position detected by the wild-
card symbol *, and other tags are used for dialog manage-
ment purposes, such as keeping track of the last
conversation exchange (that tag) or setting and getting dia-
log variables (set and get tags) or to organize the categories
according to conversation topics (topic tag).
The following categories show how the chatbot can
exploit the that and topic tags to manage the history of a
conversation and to answer according to a discussion topic:
<category>
<pattern>My name is *</pattern>
<template>
<think> <set name =‘‘ topic">hobbies</set>
Hi <starn>! What is you favorite hobby?
</template>
</category>
<topic name = ‘‘hobbies">
<category>
<pattern>photography</pattern>
<that>* what is you favorite hobby</that>
<template>I like it too</template>
</category>
</topic>
to the text. The answers are also analyzed by the emotion
recognition module in order to evaluate the emotional
content in the text. In this phase the robot can tell some
jokes and make queries to the user about his opinion on
specific topics, explicitly provoking him, in order to analyze
his reactions, both verbal and facial. The following is an
example of category defined to introduce the telling of a
joke during the dialog:
<category>
<pattern>JOKE</pattern>
<template>
<random>
<li> Usually a cow hangs his paintings in a
mooooseum </li>
<li> Which painting is never happy? The
Moaning Lisa </li>
<li> Why do artistic wives love football
season? Because their husbands sit on the sofa
long enough for them to be sketched. </li>
</random>
Do you like my sense of humor?
</template>
</category>

In the proposed architecture the AIML KB is extended

with the information represented in a semantic space and
in an emotion lexicon. The template contains specific com-
mands to query these external repositories in order to per-
form a deeper analysis of the user sentences aimed to the
emotion recognition and the evocation of images and
sketches. The template contains also commands for the
execution of specific behaviors by the humanoid, as shown
in the following category, where the template contains a
command to perform a greeting gesture:

<category>
<pattern>HI</pattern>
<template>
//ani = Sit/Gestures/Hey_3//
Glad to see you Can i make you a portrait?
</template>
</category>

The interaction follows a well defined structure. In a first

phase, named ‘‘presentation”, the robot introduces himself
and gives appropriate answers if the user asks details about
its artificial creative behavior. After this phase the robot
investigates about the user interests, addressing the conver-
sation on specific topics. Every answer of the user is mapped
into the semantic space and compared to the vectors of the
space associated to the artistic repertoire of the robot in
order to identify images that are semantically correlated Fig. 2 Timeline of artwork creation.
Analyzing and discussing primary creative traits of a robotic artist 27

In the last phase the robot starts to draw the sketches
evoked by the conversation and invites the user to draw
something with him. It invites the user to explicitly inform
it when he has finished. At the end of this phase all the
mined information, the detected emotions, the evoked
images and the drawn sketches are transmitted to the col-
lage composition module.
Artwork creation
After the inspiration phase, the robot taking into account all
the evoked information starts the execution phase. It draws
a sketch and then combines the other retrieved items in the
final artwork, finally the robot generates a title for his final
artwork. In the following sections we describe the artwork
creation process.
Taking inspiration from the thesis of Wan (2015), we have
developed our drawing system able to reproduce simple
sketches using a marker on a white sheet after verbal inter-
action with a human. During the inspiration phase the robot
chooses, from its LTM, a sketch that best matches with the
topic interaction that has triggered a creative behavior;
the selected sketch is the input of our drawing system.
Our drawing process is composed of two principal phases:
image processing, and figure drawing. In the image-
processing phase, the system takes as input the selected
sketch and performs some geometric transformations,
image enhancement and image segmentation in order to
reduce the noise and increase the accuracy of image
analysis.
The first processing step is the edge detection task that
identifies the line segments of the sketch’s shape. We
employed the Canny edge detection and the Harris Corner
detector to extract the intersection of two edges; these
represent the corner point of the geometric figure and are
used by the robot as guidelines for drawing process.
The drawing process is then executed by the robot
through a marker that it holds in its right hand. Therefore,
the knowledge of the junction points position of two seg-
ments allows the robot to plan the path for drawing the sim-
ple shape with the marker. The corner points detected in
the image are in pixels so a domain transformation, from
image domain to paper domain, from pixel to meters, to
reproduce the target shape is required. The robot must then
calculate all the movement in order to follow the points of
intersection of two edges detected on the image and
accordingly move his arm to draw. For each movement,

Fig. 3 The robot and the user draw in the same sheet during verbal interaction. The robot detects and stores relevant images
(face, final drawing). A quad tree decomposition algorithm determines available spaces in image of the drawing where putting the
collage pieces (images, sketches, and words).
28
Fig. 4 The set of users Images from social media used for the collage, and example of sketches evocated by verbal interaction.
the robot computes the distance from the current position
to the target position and by making use the API of Naoqi
it calculates the configuration needed of all arm joints to
execute the movement and draw the part of the desired fig-
ure (see Fig. 2).
This phase ends when all movements are computed and
executed and the desired figure has been drawn on the
paper.
A. Augello et al.
Collage composition
After collecting all digital materials (see Fig. 4), the cogni-
tive architecture exploits its creative behavior to create a
digital model of the artwork. The artificial artist chooses
four different composition styles. They have been deter-
mined during the learning phase and have an associated
score depending on user feedbacks on previously executed
Analyzing and discussing primary creative traits of a robotic artist
artworks. The style, as previously explained, is coded as a
vector of real values in the range ½0:0; 1:0, and since the
process involves stochastic variables the behavior is not pre-
dictable. The quad-tree decomposition on binary image of
the canvas is produced in cooperation with the human user
and determines a list of squared regions where it is possible
to place an image (see Fig. 3). The regions with area less
than 64 pixels are not considered. We have chosen to resize
the image canvas to dimensions of 1024 by 1024 in order to
reduce the processing time. If N is the number of possible
image regions, N2 ¼ roundðN  NPÞ is the effective number
of images that will be inserted in the collage. All available
images are numbered from 1 to N1 , and a random sequence
of N2 integers in the range ½0; N1  determines the collage
pieces to be included in the collage. The initial image is a
square having a random color taken from the emotional pal-
ette. The parameters forming the style are:
 BC: degree of transparency of the background color (the
color of the emotional palette with high component of
luminosity).
 NP: percentage of utilization of the available space in
the canvas to place images or sketches
 RP: percentage of a max value of r of a Gaussian with
mean 0 to calculate an image rotation in the range
½20; 20 degrees. It determines a list of values with
Gaussian distribution and the rotation of each image.
 NT: percentage of utilization of available space in the
canvas to place texts.
 RT: percentage of a max value of r of a Gaussian with
mean 0 to calculate a text rotation in the range
½20; 20 degrees.
The same procedure used to create collage pieces from
images is applied to texts: in fact, for each word or sen-
tence it is created a corresponding image. The texts have
colors randomly selected from the available emotional pal-
ette. Some examples of the obtained collages are shown in
Figs. 5 and 6.
Title generation
The artwork creation is completed by the choice of a title.
The humanoid generates a title through the analysis of the
most frequent words characterizing the user conversation
(or the user Facebook profile). We have chosen a syntactic
structure in the form of Adjective - Noun. The humanoid
associates to each distinct word its related part of speech
and the frequency counts in the whole interaction (social pro-
file). Then, the adjective of the title structure is filled with
the most frequent adjective and similarly the noun is filled
with the most frequent noun. In this selection a particular pri-
ority is given to the adjectives belonging to the emotion lex-
icon. The most frequent words can be also used in the
previous phase to introduce into the collage also written text.
Some experimental example
In this section we show some results obtained by considering
a first simple experimental setting. In this setting we have
analyzed the text produced by one of the authors of this
29
work, considering both the text produced during the inter-
action with the humanoid and the posts published in her
Facebook social profile. Through the interface One-To-
Many Comparison available at the webpage http://lsa.col-
orado.edu/ we have compared the text extracted by the
interaction with the user and the pictures descriptions, con-
sidering their vector representation in the General Reading
up to 1st year college semantic space. The system returns a
similarity score between 1 and 1 evaluated over a main
text and the other submitted texts.
The interaction between the user and the humanoid has
produced the following dialog:
A: Welcome!
U: Hi! (Greetings Gesture)
A: What’s your name?
U: Agnese.
U: Hi Agnese. I am a digital artist. My artistic
skill is composition of social collages.I call my
artworks social because it is the social
interaction with people that inspires my art. Do
you want other information about me? Please
answer with simple and short sentences.
U: No.
A: Agnese, tell me something about you. Do you have
hobbies?
U: Yes.
A: What is your favorite hobby?
U: Photography.
A: I like too! (Happy Gesture). . .
A: Do you have other hobbies?
U: I play piano and violin.
A: Can i tell you a joke?
U: Sure!
A: Which painting is never happy? The Moaning
Lisa! Do you enjoy my sense of humor?
U: Yes, i think that it is pleasant.
A: Ok, i am so happy. (Happy Gesture). Do you want
to draw with me?
U: Yes.
A: Good! (Happy Gesture). Please draw something
in this paper. I will draw something too thinking
about our conversation. Let me know when you have
finished saying Done!
(Collaborative drawing session)
U: Done.
A: Thanks, now i have to complete my artwork.
(Collage Composition and presentation)
A: This is my artwork. I named it Pleasant
Photography. Do you like it?
U: Yes!
The emotion recognized in the conversation is ‘‘joy”.
The following table shows some of the sketches evoked in
the humanoid mind during the conversation. The results
are ordered by the similarity between each sketch and the
text inserted by the user considering the first 5 results. In
particular the table shows the sketches evoked by the user
answers ‘‘Photography” and ‘‘I play piano and violin” (see
Table 1).
30 A. Augello et al.

Fig. 5 Examples of collages with the generated title.

Fig. 6 Further examples of collages with the generated title.

Table 1 Sketches evoked by the conversation.

Photography I play piano and violin
Camera 0.80 Piano 0.74
TV 0.31 Violin 0.67
Eyeglasses 0.30 Guitar 0.63
Binoculars 0.22 Harp 0.53
Eye 0.22 Tennis-racket 0.49

Table 2 Sketches evoked by the social profile of the user.

Watching a flying aeroplane Snowman and Argimusco Megaliths
Flying saucer 0.93 Snowman 0.94
Aeroplane 0.87 Ice cream cone 0.32
Helicopter 0.70 Hat 0.26
Flying bird 0.67 Backpack 0.23
Parachute 0.22 Foot 0.21
Analyzing and discussing primary creative traits of a robotic artist
A set of posts and images published on the social profile
have been mapped into the semantic space to evaluate their
semantic similarity with the textual descriptions of the
sketches. The sketches that are closer to the user’s social
contents have been selected for the creation of the collage.
The profile posts and the photo descriptions have firstly
been translated from Italian to English.
The following table reports the sketches more similar to
two pictures published by the user. The similarity has been
evaluated by mapping the pictures descriptions in the
semantic space. Table 2 reports the descriptions of the
two pictures.
As explicative results the collage illustrated in Figs. 5 and
6 have been created.
Future work
As future works we are considering to employ a larger set of
sketches that the robot can draw. We are considering also to
enrich the drawing process to effectively draw the sketches
contained in the robot KB. At this moment the humanoid is
able to draw very simple geometric shapes. We will also
enlarge the communication capabilities in dialog in order
to manage a larger set of dialog situations.
References
Augello, A., Infantino, I., Pilato, G., Rizzo, R., & Vella, F. (2013).
Introducing a creative process on a cognitive architecture.
Biologically Inspired Cognitive Architectures, 6, 131–139.
Berry, M. W., Dumais, S. T., & O’Brien, G. W. (1995). Using linear
algebra for intelligent information retrieval. SIAM Review, 37,
573–595.
Colton, S., Pease, A., & Charnley, J. (2011). Computational
creativity theory: The face and idea descriptive models. In 2nd
International conference on computational creativity. .
Cook, M., & Colton, S. (2011). Automated collage generation–With
more intent. In Proceedings of the second international confer-
ence on computational creativity. Citeseer.
Csikszentmihalyi, M. (1988). Society, culture, and person: A
systems view of creativity. Cambridge University Press.
Eitz, M., Hays, J., & Alexa, M. (2012). How do humans sketch
objects? ACM Transactions on Graphics, 31, 44.
Indurkhya, B., & Ogawa, S. (2012). An empirical study on the
mechanisms of creativity in visual arts. In Proceedings of the
31
34th annual conference of the cognitive science society
(pp. 1727–1732). Citeseer.
Krzeczkowska, A., El-Hage, J., Colton, S., & Clark, S. (2010).
Automated collage generation-with intent. In Proceedings of the
1st international conference on computational creativity
(pp. 20).
Landauer, T. K., Foltz, P. W., & Laham, D. (1998). An introduction
to latent semantic analysis. Discourse Processes, 25, 259–284.
Pilato, G., & Vassallo, G. (2015). Tsvd as a statistical estimator in
the latent semantic analysis paradigm. IEEE Transactions on
Emerging Topics in Computing, 3, 185–192. http://dx.doi.org/
10.1109/TETC.2014.2385594.
Pilato, G., Vella, F., Vassallo, G., & La Cascia, M. (2010). A
conceptual probabilistic model for the induction of image
semantics. In 2010 IEEE fourth international conference on
semantic computing (ICSC) (pp. 91–96). IEEE.
Pound, E. (1934). ABC of reading. Yale University Press.
Rhodes, M. (1961). An analysis of creativity. The Phi Delta Kappan,
42, 305–310.
Sahlgren, M. (2005). An introduction to random indexing. Methods
and applications of semantic indexing workshop at the 7th
international conference on terminology and knowledge engi-
neering, TKE (Vol. 5). .
Sawyer, K. (2008). Group genius: The creative power of collabo-
ration. Basic Books.
Sawyer, R. (1999). The emergence of creativity. Philosophical
Psychology, 12, 447–469.
Stojanov, G. K., & Indurkhya, B. (2013). Creativity and cognitive
development: The role of perceptual similarity and analogy. In
2013 AAAI spring symposium series. .
Strapparava, A. V. (2004). Wordnet-affect: An affective extension
of wordnet. In Proceedings of 4th international conference on
language resources and evaluation (LREC 2004)
(pp. 1083–1086).
Strapparava, C., & Mihalcea, R. (2007). Semeval-2007 task 14:
Affective text. In Proceedings of the 4th international workshop
on semantic evaluations, SemEval’07 (pp. 70–74).
Strapparava, C., & Mihalcea, R. (2008). Learning to identify
emotions in text. In Proceedings of the 23rd annual ACM
symposium on applied computing (APPLIED COMPUTING 2008)
(pp. 1556–1560).
Wan, Q. (2015). Developements of drawing capability for Nao
humanoid robot Master’s thesis. Vaasan ammattikorkeakoulu
University of Applied Sciences. Technology and Communication.
Wiggins, G. A. (2006). A preliminary framework for description,
analysis and comparison of creative systems. Knowledge-Based
Systems, 19, 449–458. http://dx.doi.org/10.1016/
j.knosys.2006.04.009.