Extraction of Agent Psychological Behaviors from
Glosses of WordNet Personality Adjectives
Jean-Paul Sansonnet1 and François Bouchet1
1 LIMSI-CNRS,
BP 133, F-91403 Orsay Cedex, France
Abstract. Conversational agents have two main parts: the rational agent performs symbolic reasoning over the model of the system while the psychological
agent is in charge of the interaction with the user. In this context, agent cognitive
modeling focuses on the taxonomy and on the computational implementation of
psychological behaviors, in close relation with the rational reasoning process.
We present here the process of elicitation of a large class of psychological behaviors for their future computational implementation in rational agents. Behaviors
are extracted from a corpus of personality adjectives associated with synsets and
glosses in the WordNet lexical data base. Collected glosses are classified along
the standard FFM - NEO PI-R taxonomy of personality traits in order to constitute clusters, now available as a resource for the community of agent cognitive
modeling.
Keywords: Agent cognitive modeling, Personality traits, WordNet lexical base.
1
1.1
Introduction
Adding a personality to rational agents
According to traditional definitions issuing from Artificial Intelligence and Multi-Agent
Systems, Rational Agents are associated with programs capable of Practical Reasoning
that is building plans and choosing actions to be executed, in order to achieve their
goals. For example, SOAR-based architectures are one of the first attempts at modeling
the cognitive reasoning process of an agent [1] by means of explicit IF-THEN rules.
More recently, the BDI approach of Bratman [2], Rao and Georgeff [3] is a theory
of practical reasoning (deciding what to do next) directed towards situated reasoning
about actions and plans [4]. Recently, authors have proposed to integrate into rational agent architectures psychological notions, for two main reasons: first) to propose
more complete cognitive models of agents; second) to propose agents capable of sustaining more human-like interactions with people, especially ordinary people involved
in conversational activities with assistant agents. For example, Gratch and Marsella [5]
have proposed a model of emotions based on SOAR, with a significant impact upon the
SOAR architecture [5]. Using the agent creation platform JACK [6] that implements
the BDI theory, CoJACK [7, 8] is an extension layer intended to simulate physiological
human constraints like the duration taken for cognition, working memory limitations
(e.g. loosing a belief” if the activation is low or “forgetting the next step” of a procedure), fuzzy retrieval of beliefs, limited focus of attention or the use of moderators
to alter cognition. The eBDI architecture of Jiang et al. [9] implements emotions in a
2
BDI framework. They give a good introduction about the history of the necessity to
implement emotions into rational agents.
1.2
The R&B framework
Although presented here as separate notions, the rational and the psychological reasoning capacities of an agent actually work in quite an intricate manner [10, 11]. Moreover,
most studies mentioned above focus on low-level/transitory psychological notions such
as Emotions and Moods; other notions like Personality Traits [12] that are associated
with high-level/long-lasting features of the personality of a human being, should be
integrated and could be promising for developing agents with consistent characters.
This is the reason why we have proposed a framework dedicated to the study of the
nature of the relationships between the rational and the psychological reasoning capacities: the Rational and Behavioral architecture (R&B) 1 (where ’behavioral‘ has the
particular meaning of ’psychological behavior’). R&B is a generic framework enabling
the computational definition and the experimentation of various rational/psychological
strategies. In recent work based on R&B [13], we have proposed a model where psychological phenomena, such as personality traits, are implemented in terms of influence operators altering the rational process of an agent. Presently the proposed model
has been implemented only with few arbitrary-chosen psychological phenomena (e.g.
a lazy/gloomy/cheerful/. . . agent) [14]. This approach illustrates the principle and the
feasibility of the R&B framework, but it fails to cover significantly the main domains of
the description of the psychology of a person, especially un terms of personality traits.
1.3
The need for eliciting a set of psychological schemes
In the literature on Psychology, a lot of works has been achieved to assess the personality traits of a person, especially by proposing taxonomies of emotions, moods,
traits, affects, stereotypes, social roles etc. (see section 2.1 for a development on traits).
Hence we have to rely upon and to be compliant with state of the art works in Psychology. However from the computational point of view, our main concern is that those
taxonomies are far too generic: e.g. 16 categories in Catell’s model [15] ; 5 categories in
the Five Factor Model [16]. This is the reason why some authors have proposed to refine
traits taxonomies with so-called facets, like for example the NEO PI-R facets [17], thus
resulting in a two-leveled taxonomy (traits-facets). Unfortunately, the proposed facets
are still too generic (see glosses definitions of NEO PI-R facets given in Table 1), so
they cannot lead to a straightforward implementation.
Our proposition can be summarized in the following key-points:
1. We suggest to refine the facets in order to propose a three-leveled taxonomy for describing the personality of a person.
2. The third level of this extended taxonomy will be composed of so-called psychological schemes (or in short schemes) that are intended to provide implementable instantiations of the facets, especially in terms of influences upon actions and plans of the
1
http://www.limsi.fr/~jps/research/rnb/rnb.htm
3
rational agent.
3. We propose to elicit those schemes from a corpus of personality adjectives, which
are in turn associated with their lexical senses and glosses in the WordNet lexical data
base [18]. Indeed, linguistics resources have long been a source for building psychological taxonomies (see section 2.1); moreover WordNet offers three extra advantages: it
is a computerized data base; it has a standard position in natural language processing;
above all, it provides a comprehensive set of lexical semantics senses (called synsets)
that will provide a solid ground for defining the schemes.
4. The scheme elicitation process has been conducted in two main phases, each divided
in two sub phases:
Phase 1
Phase 2
Personality adjectives:
a) selection of a significant subset of personality adjectives and
b) classification into a standard, two-leveled taxonomy (traits-facets)
For each facet:
a) clusterization of involved lexical senses (synsets) into schemes and
b) elicitation of the influence operators associated with the schemes
Works achieved in phase 1 have been recently presented in [19] and the XML resource describing the classification of the selected personality adjectives can be freely
accessed from a Web page 2 of the R&B project.
In this paper, we present the works achieved in phase 2. Also, the XML resource describing the clusterization of the schemes can be freely accessed from the same Web
page 2 of the R&B project.
Outline of the paper: In section 2 we present aspects in Psychology and Computational
Linguistics underlying this study, together with a summary of the process achieved in
phase 1. Section 3 presents the two steps involved in phase 2: first the definition of the
schemes, then the definition of the operators associated with the schemes.
2
Classification of WordNet lexical semantic senses related to
personality adjectives into the FFM/NEO PI-R taxonomy
2.1
Contribution of linguistic resources to works on personality traits
The Five Factor Model (FFM) When one is interested in the taxonomy of the psychological phenomena, especially those related to personality traits, the most successful
paradigm to day is the FFM, which is a convergent research from many authors in
Psychology during the last 20 years. This paradigm has taken upon Cattell’s classification into 16 factors [15], which was still prominent in the 80s and was supported by
Eysenck’s Personality Questionnaires (EPQ), which are questionnaires (generally with
yes/no questions) to assess the personality traits of a person [20, 21]. Another approach
to the taxonomy of traits is based on natural language and more precisely lexical resources [22], such as the glosses found in dictionaries. The lexical hypothesis states that
2
http://www.limsi.fr/~jps/research/rnb/toolkit/taxo-glosses/taxo.htm
4
most of socially relevant and salient personality characteristics have become encoded
in the natural language [23]. The lexical approach has been promoted by Goldberg who
claimed that “personality vocabulary provides an extensive, yet finite, set of attributes
that people speaking a given language have found important and useful in their daily
interactions” [24]. In 1990, Goldberg tried to define a small set of 475 common trait
adjectives grouped into 131 sets of factors that can cover the big five domain [25]; It
issued in 1992 into the 50-item instrument using so-called ‘transparent format’ [16] that
finally led to the FFM. The FFM is based on five large classes of psychological traits,
often named Big Five model or OCEAN, by taking the first letter of the name of each
class (which are listed in the first column of Table 1).
The NEO PI-R facets The FFM being a very generic classification, several authors
have tried to refine this taxonomy by dividing the FFM classes into so-called facets [26,
17, 27]. John et al. [28] have shown that these facet lists have many similarities although
their facet number varies from 16 in [26] to 30 in the so-called NEO PI-R3 proposition
of Costa and McCrae [17]. The 30 facets of the NEO PI-R classification are listed in
the second column of Table 1, together with their glosses. NEO PI-R is a long standing
model that provides a very precise facet list. For these reasons, we will rely on it for the
work presented in section 2.3.
2.2
The lexical semantic senses of the WordNet base
Above, we have mentioned that authors already have classified sets of adjectives in the
FFM model and in the NEO PI-R model. However, when we tried to use these classifications, we encountered two main problems: first, classification tables provided in
the papers are too synthetic to be exploited and it is not easy to access to the original
resource data files. Second, in order to fulfill the requirements of computational definition, we need words but also the precise lexical semantics attached to them (their
senses). For example, the word ‘kind’ has three senses: Tolerant, Genial, Openhearted
(see Table 2) that can be classified into two distinct FFM classes and three distinct NEO
PI-R facets: resp. A-compliance, E-warmth, A-tendermindedness.
This example shows that the word level is insufficient instead we need to work at
the level of the semantic senses.
The WordNet lexical data base [18] comes handy when one has to treat a large
amount of lexical data and one has to access the lexical semantics of words. In WordNet, a word is attached to several so-called synsets that define a unique lexical sense
described by a gloss (also called a short phrase, cf. Table 2 for examples). Moreover,
because the WordNet data base is freely accessible, it makes it easy to build a computer
aided system for the classification process. Note that WordNet has been used to support
research on affective computing, e.g. in the ‘WordNet affect’ project [29, 30]; however
this work is dedicated to the recognition of affects in texts, not to the expression of
psychological behaviors by a CAA, and the chosen classification is mainly ad hoc.
3
NEO PI-R stands for Neuroticism Extraversion Openness Personality Inventory-Revisited
5
Table 1. NEO PI-R facets for the Big five personality trait domain.
OCEAN classes
NEO PI-R 30 facets Facet gloss
Openness
Fantasy
Aesthetics
Feelings
Actions
Ideas
Values
Competence
Orderliness
Dutifulness
Conscientiousness
Achievement-striving
Self-discipline
receptivity to the inner world of imagination
appreciation of art and beauty
openness to inner feelings and emotions
openness to new experiences on a practical level
intellectual curiosity
readiness to re-examine own values and those of authority figures
belief in own self efficacy
personal organization
emphasis placed on importance of fulfilling moral obligations
need for personal achievement and sense of direction
capacity to begin tasks and follow through to completion despite
boredom or distractions
Extraversion
Agreability
Neuroticism
Deliberation
tendency to think things through before acting or speaking
Warmth
Gregariousness
Assertiveness
Activity
Excitement-seeking
Positive-emotions
interest in and friendliness towards others
Trust
Straightforwardness
Altruism
Compliance
Modesty
Tender-mindedness
belief in the sincerity and good intentions of others
Anxiety
Angry-Hostility
level of free floating anxiety
preference for the company of others
social ascendancy and forcefulness of expression
pace of living
need for environmental stimulation
tendency to experience positive emotions
frankness in expression
active concern for the welfare of others
response to interpersonal conflict
tendency to play down own achievements and be humble
attitude of sympathy for others
tendency to experience anger and related states such as frustration and bitterness
Depression
tendency to experience feelings of guilt, sadness, despondency
and loneliness
Self-consciousness
Impulsiveness
shyness or social anxiety
tendency to act on cravings and urges rather than reining them
in and delaying gratification
Vulnerability
2.3
general susceptibility to stress
Summary of the classification process
In work related to phase 1 [19], we have carried out a classification of a set of personality
adjectives in the FFM/NEO PI-R taxonomy. We give here a summary of this process
carried out by two annotators, in three main steps:
Step 1. Personality adjective selection: In order to work on actually used personality adjectives, we have collected a set of personality adjectives Ccoll , from ten different
Internet sources explicitly claiming to provide “lists of adjectives describing person-
6
Table 2. WordNet synsets associated with three commonly used personality adjectives. In column
2, stars mark synsets not related to personality traits.
Adjective Not Synset
friendly
*
*
Pally
Allied
Easy
Favorable
WordNet gloss defining the synset
characteristic of or befitting a friend
of or belonging to your own country’s forces or those of an ally
easy to understand or use
inclined to help or support; not antagonistic or hostile
Tolerant
tolerant and forgiving under provocation
Genial
agreeable, conducive to comfort
Openhearted having or showing a tender and considerate and helpful nature;
used especially of persons and their behavior
kind
lively
*
*
Vital
Eventful
Frothy
Springy
Alert
Racy
full of spirit
filled with events or activity
full of life and energy
elastic; rebounds readily
quick and energetic
full of zest or vigor
ality traits” (sources listed in [19]). Ccoll contains 1,055 distinct adjectives, providing
a first order approximation of the linguistic domain related to personality traits adjectives. Moreover, for each personality adjective a, we have associated a salience rank ra
depending on the number of lists in which a given adjective appears, which represents
its usage frequency in the resources files. We have observed ra values ranging from 1 to
9. The most salient adjectives (ra = 9) are listed in Table 2. Unfortunately, some of the
words appearing in Ccoll are not present in the WordNet data base (N = 29) or present
but not listed as adjectives (N = 21), but their low salience (ra = 1) allowed us to choose
to discard them, leading to a set CWN of 1,005 distinct WordNet-tractable adjectives.
Step 2. Selection of the relevant adjective-synset-gloss triplets: Each adjective
a ∈ CWN has a WordNet entry, which associates with a a set of lexical semantics senses
(its synsets). In WordNet, each synset is a unique identifier and defines a unique sense
by means of a short phrase also called a gloss. It is thus possible to obtain from the
CWN list of adjectives, a base Ball of triplets < ad jective, synset, gloss >. However
as one can see from entries marked with a * in Table 2, not all triplets are related
to the description of the personality of a person. In order to discard these entries, an
annotating process was performed by two annotators (details in [19]), which resulted in
the selection of a base Bsel of 1,356 relevant triplets, containing 904 adjectives.
Step 3. Classification of the triplets into the FFM/NEO PI-R taxonomy: The
NEO PI-R taxonomy defines 6 facets for each of the five OCEAN personality classes,
thus resulting in 30 positions. In NEO PI-R, each facet encompasses both the concept
and its antonym(s), e.g. the facet A-Modesty can stand for adjectives like ‘mild’ as well
as ‘arrogant’. Hence, similarly to Goldberg’s works on 50 bipolar scales [16], each
facet was divided into two poles: noted + for the concept and - for its antonym(s),
thus resulting in 60 target positions noted +/-NEO PI-R. The classification of the 1,356
7
triplets was performed by two annotators (details in [19]), which resulted in the building
of a base, further transformed into an XML resource, freely available on the Web 2 , as
shown in Figure 1.
3
3.1
Elicitation of the psychological schemes
Definition of the psychological schemes
From the process of classification of the triplets into the FFM/NEO PI-R taxonomy we
have obtained a resource base where each facet contains two bipolar lists of triplets. The
glosses contained in a given facet provide several instantiations of the psychological notion associated with the facet. Actually, these instantiations can describe distinct kinds
of actual psychological behaviors, needing to be implemented with distinct operations. This is the reason why, for each facet of the FFM/NEO PI-R taxonomy, we had to
make a new classification of their triplets in terms of so-called psychological schemes.
Schemes are sets of triplets corresponding to an atomic psychological behavior. Indeed,
each scheme is intended to be further associated with a specific computational implementation, for example in terms of influence operators over the rational process of the
agent. Formally, each scheme is a structure of the form:
Scheme ID
FFM/NEO PI-R category
Positive pole ID, Textual definition, {gloss1, .. glossN}
Negative pole ID, Textual definition, {gloss1, .. glossN}
Like facets, schemes are organized into a positive pole and a negative pole. Contrary
to facet-poles that support a two-position scale (concept/antonym opposition), schemepoles are organized on a three-position scale: more than average (+), average (=), less
than average (-). Actually the average position (=) has no triplets associated with, because it implicitly refer to the “average behavior of an average personality”. + and
- poles explicitly denote a personality having an inclination diverging from average.
Technically, the semantics of a scheme is only and uniquely defined by the list of its
associated glosses. However for each scheme pole, we provide also an informal textual
definition synthesizing its glosses.
Example: The facet O-Fantasy, showed in Figure 1, can be divided into three schemes
given in Table 3 (the glosses of the scheme poles are abridged to their count).
3.2
Organization of the facets into psychological schemes
During the process of organization of the facets into psychological schemes 766 glosses
were annotated and clustered into 69 distinct schemes resulting in a total of 2×69 = 138
scheme +/- poles. Figure 2 displays the distribution of the glosses and the schemes in the
FFM/NEO PI-R taxonomy. Globally, one observes a good and homogeneous coverage
in terms of glosses and schemes over OCEAN traits and NEO PI-R facets. There is also
a good and homogeneous coverage of glosses over the schemes: the distribution varies
from min = 2 to max = 40, with a mean of 26 glosses by scheme. However gloss coverage is not complete at the pole level: 11 scheme poles over 138 have no gloss attached
8
O.C.E.A.N. category
NEO PI-R facet
Positive pole
(concept)
Wordnet gloss
[ word, Synset, rank]
Multi synsets
Negative pole
(antonyms)
abridged
Fig. 1. Screenshot of the Web page of the XML file of the personality adjectives classification
9
Table 3. Clusterization of facet Openness-Fantasy (Ofa) into three distinct atomic bipolar
schemes. For poles +/- of scheme CREATIVENESS, we give the set of its associated triplets,
actually abridged here to the gloss part (text italicized)
Scheme’s poles
Identifiers
Textual definition of the scheme
+ abridges ‘more than average’ resp. - ‘less than average’
1. IDEALISTICNESS
+ IDEALISTIC
- PRACTICAL
Ofa
+ inclined to like (even to believe in) imaginary ideas or situations
- inclined to like (or to believe in) imaginary ideas or situations
Gloss
count
1
5
2. SPIRITUALISTNESS Ofa
+ SPIRITUALIST + interested in spirit and soul matters
- MATERIALIST
- interested in spirit and soul matters, even denying the tenets
6
2
3. CREATIVENESS
+ CREATIVE
4
Ofa
+ having power of creativity in thought or action
having the ability to produce or originate
promoting construction or creation
having the ability or power to create
[. . . ] marked by independence and creativity in thought or action
- UNIMAGINATIVE - having power of creativity in thought or action
unimaginatively conventional
deficient in originality or creativity; lacking powers of invention
unimaginative and conformist
lacking spontaneity or originality or individuality
Fig. 2. Distribution in FFM/NEO PI-R taxonomy of glosses (top) and schemes (bottom), where:
N f acet = 30, Nglosse = 766, Nscheme = 69, mean number of glosses per facet = 26 and mean
number of schemes by facet = 2.3
4
10
to them. This is due to the fact that each time a new scheme has been introduced, it was
created with its two +/- poles and some of them remained empty of glosses through the
organization process. An additional list of personality adjectives should be considered
and classified in order to potentially fill these scheme poles. However as a first approximation, the semantics of the empty scheme poles can be derived by taling the antonym
attached to the opposite pole (e.g. in the LOYALNESS scheme, the semantics of - pole
UNFAITHFUL (containing 0 glosses) can be inferred from its + pole LOYAL (containing
8 glosses).
3.3
Association of influence operators to the schemes
Once the WordNet glosses of the set of personality adjectives have been reorganized
into the FFM/NEO PI-R facet taxonomy, they must be associated with an actual computational implementation. For example, in the R&B framework, this part is played by
so-called influence operators. These operators are heuristics operating on/altering the
rational reasoning process of the agent (see [13] for a formal definition). In this context, a final step towards the computational implementation of the schemes is required
because there is often but not always a direct correspondence between a scheme and an
operator. It comes from the fact that schemes are concepts elicited from the linguistic
domain point of view, whereas operators we wish to define are symbolic heuristics (or
meta rules) defined from a computational point of view. This heterogeneity entails three
consequences:
– It can happen that a scheme requires several distinct operators to be implemented;
– Conversely, the same operator can implement distinct schemes, meaning that actually
these schemes are not distinct from a computational point of view;
– Moreover, the +/- poles of the schemes can be inverted with the +/- poles of the operators (especially in the Neuroticism class).
From the set of 69 schemes, it was possible to elicit a set of 56 operators, listed in
Table 4. Because the operators can appear in several positions in the FFM/NEO PI-R
taxonomy, we had to establish another taxonomy for the operators, which is a twoleveled taxonomy. The first reason for that is that, whereas the scheme taxonomy is
psychology-oriented (FFM/NEO PI-R) the operator taxonomy is organized according
to the rational activities of an agent. From the observation of the operator set, a first
salient taxonomic dividing feature separates:
– Mono-agent situations: with operators mainly related to the actions of the agent in the
physical world. In turn, this class of operators is made of four subsets: personal knowledge; personal action; personal satisfaction; control of personal mind states.
– Multi-agent situations: with operators mainly related to interpersonal relationships
(e.g. in the case of a conversational interaction bteween an agent and a human). In turn,
this class of operators is made of three subsets: knowledge exchange; multi-agent action
and control of others; interactional manners and feelings.
Table 4 lists the taxonomy of the No = 56 operators with a short textual definition
and for each category the count of its operators. From this taxonomy, it was possible to
draw three main results:
11
Table 4. 2-level taxonomy of the influence operators (No = 56)
Operator
Textual definition of the operator
Count
1. MONO AGENT
1.1 PERSONAL KNOWLEDGE
memory
capacity to retain facts and tasks to do
fact
degree of factualness required to believe a proposition
commonsense amount of general knowledge in the fact base
learn
interest in learning new facts about the world
26
4
1.2 PERSONAL EXECUTION OF ACTION AND PLANS
adaptation
capacity to adapt actions/plans to events and changes
decision
capacity to make decisions and choices
deliberation capacity to compute the consequences of ones actions
creativity
capacity to create things
danger
attitude towards danger
expectation hopes and fears of possible outcomes
novelty
attitude toward new things
clean
capacity to maintain a clean order of things
method
capacity to achieve a task with method
concern
controls the volatility from task to task, role to role
attention
level of focus in the execution tasks
activity
energy put in involving in and executing tasks
work
attitude towards executing work-based tasks
13
1.3 PERSONAL SATISFACTION
possession
possessivity of resources, advantages
success
desire for success of ones goals
pleasure
attirance to good life, even luxury
art
attirance to art and beauty in things
4
1.4 PERSONAL MOOD/AFFECT CONTROL
moodiness
controls the change speed of moods and affects
emotionality way to react emotionally under stress bad events or failure
fierceness
controls the speed and intensity in reacting to events
assurance
controls the confidence in ones competence, charisma, worth
prone(m)
controls the tendency to experience the mood m
5
2. MULTI AGENT
2.1 INTERACTION: KNOWLEDGE
trust
level of trust upon others told facts, promises, actions
frankness
telling facts to others that are true, complete, not vague
curiosity
interest in learning facts about others
privacy
controls the level of private facts told to others
conversion
capacity to change one’s idea and adopt ideas of others
30
5
2.2 INTERACTION: ACTION AND CONTROL OF OTHERS
15
meeting
level of acceptance to meet with others (form social groups)
cooperation level of acceptance to enter into co action with shared goals
sacrifice
disregard ones advantages to support others
givenness
will to give/share one’ s possessions with others
fairness
how resources, advantages/penalties are distributed over others
forgiveness
indulgence and forgiveness to others misdeeds
commitment will to keep promises, affection, allegiance concerning others
charisma
capacity to drive and entice others
domination
will to exert domination upon others
menace
will to exert all kinds of intimidation upon others
agreement
will to reach agreement in conflicts with others
honesty
honesty in dealings with others
malice
will to obstruct, to hurt, to disadvantage others in hidden manner
conform
conforming to traditions for oneself and others
progress
desire to change traditions with others
2.3 INTERACTION: FEELINGS AND MANNERS
comfort
reacting to others problems by expressing comfort
reassurance asking others for comfort and help when one has problems
perception
capacity of perceptiveness of others feelings
empathy
capacity to experience other feelings
humility
express to others with humble manners
tolerance
reaction to provocations by others
aggression
capacity to pro actively attack, even hurt others
friendliness
manner of warmth and friendliness in expression to others
showiness
manner bright and excessive in expression to others
smalltalk
level of smalltalk in dialogue with others
10
12
Result 1. First level balance: There is a good balance between the two sets of personal
operators (N po = 26) and the interactional operators (Nio = 30). This balance is not artificial: it is the result of the distribution of the lexical semantics attached to the initial
set of personality adjectives. Actually, the linguistic senses reflects the two faces of an
agent: its personal face and its social face.
Result 2. Scheme-operator alignment: The alignment between schemes and operators
is quite good. Actually, most schemes (68%) are in bijection with an operator and 25%
of the others operators occur only in two schemes, as shown in Table 5.
Table 5. Statistics of association between schemes and operators
Type of operators counted
All occurrences in schemes
With 1 occurrence (bijections)
With 2 occurrences
With 3 occurrences
With special operator prone(m)
Frequency % of No
82
38 67.9%
14 25.0%
3 5.4%
7 12.5%
Result 3. Adjective convergence over the operators: Figure 3 plots of the number of
operators hit by the glosses associated with the initial set of 621 personality adjectives.
While the number of adjectives increases from 1 to 621, the number of covered operators increases in a log-like curve from 1 to 56. In the bottom curve, adjectives were
taken by salience order (and in alphabetical order in case of equal rank). The top curve
is the mean of 5 random draws in the 621 adjective set. One can make two observations:
– Most salient adjectives are less good at rapidly covering the operators: this is due to
the fact that most used adjectives are not equally distributed over the facets as shown in
Figure 2. Note from Figure 3 that with a random sample of only 100 adjectives one can
expect to cover 90% of the operators.
– Moreover in the top curve, all operators are covered with 229 adjectives (mean over
5 random draws). With WordNet, one can easily compute lots of synonyms for words;
so it is technically possible to increase the base of adjectives by adding synonyms (with
a loss in semantic relevance), say to double or triple the initial base. However, the top
curve shows that adding new adjectives will probably produce few new schemes, hence
few new operators. We can consider that the resource base built in this work provides a
good coverage of the personality trait domain in terms of schemes and operators.
4
Conclusion and perspectives
We have described here the definition of a taxonomy of psychological behaviors through
a study based 1) on standard personality traits taxonomies in psychology (FFM with
NEO PI-R for the facets) and 2) on a commonly used lexical data base (WordNet). Using the senses of an original selection of 1,005 personality adjectives in that base, we
13
Operator
Count
50
40
30
20
10
0
0
100
200
300
400
500
600
Adjective
Count
Fig. 3. Coverage of operators with adjectives: Bottom curve) ranked set of adjectives; top curve)
mean of 5 random draws in the same set.
have dispatched them over the 60 facets of NEO PI-R, taking into account their polarity. The glosses associated to those senses in WordNet thus provide extra and useful
information about how to computationally implement a given facet. We have finally
shown that although there is often a direct mapping between a psychological facet and
a computational operator, there are several exceptions, which led us to define a second
computationally-oriented taxonomy.
We believe this taxonomy can become a useful resource for future works focusing
on the implementation of psychological behaviors, like we had done in preliminary
works regarding the work or memory operator (cf. examples of a lazy and of a scatterbrained agent developed in [14]). This is confirmed by the fact that several operators
identified in the multi-agent section of our taxonomy correspond to some already wellstudied phenomena, like trust [33] and empathy [34].
Once implemented, a given operator would then have to be evaluated. Considering
the class of multi-agent operators, for example in the context of conversational agents, it
can be done by making human subjects interact with the agent. Considering the class of
mono-agent operators, for example in the context of autonomous agents, it can be done
by observing the agent actual actions. These experiments should confirm that people
would notice differences in behaviors, and that those differences actually correspond to
the personality trait or facet that was intended to be implemented by that operator.
References
1. Laird, J.E., Newell, A., Rosenbloom, P.S.: Soar: an architecture for general intelligence.
Artif. Intell. 33(1) (1987) 1–64
2. Bratman, M.E.: What is intention? In Cohen, P.R., Morgan, J., Pollack, M.E., eds.: Intentions
in Communication. The MIT Press (1990) 15–32
14
3. Rao, A.S., Georgeff, M.P.: BDI agents: From theory to practice. In: Proc. First Int. Conference on Multi-agent Systems (ICMAS-95). (1995) 312–319
4. Allen, J.F., Kautz, H.A., Pelavin, R.N., Tenenberg, J.D.: Reasoning About Plans. Representation And Reasoning. Morgan Kaufmann Publishers, Inc. (1991)
5. Gratch, J., Marsella, S.: A domain-independent framework for modeling emotion. Journal
of Cognitive Systems Research 5(4) (2004) 269–306
6. Howden, N., Rnnquist, R., Hodgson, A., Lucas, A.: Intelligent agents - summary of an
agent infrastructure. In: Proc. of the 5th International Conference on Autonomous Agents,
Montreal (2001)
7. Norling, E., Ritter, F.E.: Towards supporting psychologically plausible variability in AgentBased human modelling. In: Proceedings of the Third International Joint Conference on
Autonomous Agents and Multi-Agent Systems. (2004)
8. Evertsz, R., Ritter, F.E., Busetta, P., Pedrotti, M.: Realistic behaviour variation in a BDIbased cognitive architecture. In: Proc. of SimTecT’08, Melbourne, Australia (2008)
9. Jiang, H., Vidal, J.M., Huhns, M.N.: eBDI: an architecture for emotional agents. In: AAMAS ’07: Proceedings of the 6th international joint conference on Autonomous agents and
multiagent systems, New York, NY, USA, ACM (2007) 1–3
10. Ellsworth, P.C., Scherer, K.R.: Appraisal processes in emotion. In Davidson, R.J., Scherer,
K.R., Goldsmith, H.H., eds.: Handbook of affective sciences. Series in affective science.
Oxford University Press, New York, NY, USA (2003) 572–595
11. Frijda, N.H.: The Laws of Emotion. Psychology Press (2006)
12. John, O.P., Robins, R.W., Pervin, L.A., eds.: Handbook of Personality: Theory and Research.
3rd edn. The Guilford Press (August 2008)
13. Sansonnet, J.P., Bouchet, F.: Expression of behaviors in assistant agents as influences on rational execution of plans. In: Intelligent Virtual Agents (IVA 2010). Volume 6356 of LNAI.,
Philadelphia, PA, Springer-Verlag (2010) 413–419
14. Bouchet, F., Sansonnet, J.P.: A framework for modeling the relationships between the rational and behavioral reactions of assisting conversational agents. In: Proc. of the 7th European
Workshop on Multi-Agent Systems (EUMAS’09), Agia Napa, Cyprus (December 2009)
15. Cattell, R.B., Eber, H.W., Tatsuoka, M.M.: Handbook for the sixteen personality factor
questionnaire (16 PF). Champain, Illinois (1970)
16. Goldberg, L.R.: The development of markers for the big-five factor structure. Psychological
Assessment 4 (1992) 26–42
17. Costa, P.T., McCrae, R.R.: The NEO PI-R professional manual. Odessa, FL: Psychological
Assessment Resources (1992)
18. Fellbaum, C.: WordNet: An Electronic Lexical Database. MIT Press, Cambridge, MA (1998)
19. Bouchet, F., Sansonnet, J.P.: Classification of wordnet personality adjectives in the neo pi-r
taxonomy. In: Fourth Workshop on Animated Conversational Agents, WACA 2010, Lille,
France (2010)
20. Eysenck, H.J., Eysenck, S.B.G.: Manual of the Eysenck Personality Questionnaire. London:
Hodder and Stoughton (1975)
21. Eysenck, S.B.G., Eysenck, H.J., Barret, P.: A revised version of the psychoticism scale.
Journal of Personality and Individual Differences 6(1) (1985) 21–29
22. Baumgarten, F.: Die charaktereigenschaften [The character traits]. Beitrage zue Charakterund Personlichkeitsforschung (1) (1933)
23. Allport, G.W.: Personality: A psychological interpretation (1937)
24. Goldberg, L.R.: Language and individual differences: The search for universal in personality
lexicons. Review of personality and social psychology 2 (1981) 141–165
25. Goldberg, L.R.: An alternative description of personality: The big-five factor structure. Journal of Personality and Social Psychology 59 (1990) 1216–1229
15
26. Saucier, G., Ostendorf, F.: Hierarchical subcomponents of the big five personality factors:
A cross-language replication. Journal of Personality and Social Psychology 76(4) (1999)
613–627
27. Soto, C.J., John, O.P.: Ten facet scales for the big five inventory: Convergence with NEO PIR facets, self-peer agreement, and discriminant validity. Journal of Research in Personality
43(1) (2009) 84–90
28. John, O.P., Naumann, L.P., Soto, C.J.: Paradigm shift to the integrative big five trait taxonomy: History, measurements, and conceptual issues. In: Handbook of Personality: Theory
and Research. 3rd edn. The Guilford Press (2008)
29. Strapparava, C., Valitutti, A.: WordNet-Affect: an affective extension of WordNet. In: Proc.
of the 4th International Conference on Language Resources and Evaluation, Lisbon, Portugal
(2004) 1083–1086
30. Strapparava, C., Mihalcea, R.: Learning to identify emotions in text. In: Proceedings of
the 2008 ACM symposium on Applied computing, Fortaleza, Ceara, Brazil, ACM (2008)
1556–1560
31. Sabouret, N., Sansonnet, J.P.: Learning collective behavior from local interactions. In:
CEEMAS: Second International Workshop of Central and Eastern Europe on Multi-Agent
Systems. Volume 2296 of LNCS., Cracovia, Poland, Springer Verlag (2002) 273–282
32. Sansonnet, J.P., Sabouret, N., Pitel, G.: An agent design and query language dedicated to
natural language interaction. In: AAMAS:2002 First International Joint Conference on Autonomous Agents and Multi-Agent Systems, Bologna, Italy (2002)
33. Ramchurn, S., Huynh, D., Jennings, N.: Trust in multi-agent systems. The Knowledge
Engineering Review 19(1) (2005) 1–25
34. Paiva, A., Dias, J., Sobral, D., Aylett, R., Sobreperez, P., Woods, S., Zoll, C., Hall, L.: Caring
for agents and agents that care: Building empathic relations with synthetic agents. In: Proceedings of the Third International Joint Conference on Autonomous Agents and Multiagent
Systems. Volume 1., IEEE Computer Society (2004) 201