WHO ARE YOU?
On the Acquisition of Information about People for an Agent that Remembers
Nikita Mattar and Ipke Wachsmuth
Artificial Intelligence Group, Bielefeld University, Bielefeld, Germany
Keywords:
Conversation topics, Embodied conversational agents, Human-agent interaction, Long-term interactions,
Person memory, Small talk, Social categories, Stereotypes.
Abstract:
Humans make extensive use of specialized representations to remember people they interacted with. While
current research on embodied conversational agents focuses on the relationship between agent and interlocutor,
the representation of the latter is mostly neglected. But information about others are inevitable for an agent to
adapt to its interlocutors and to establish long-term relationships with them. In this work, we present a model
of Person Memory for virtual agents. We discuss what kinds of information have to be stored about people.
Furthermore, we stress the importance of social categories. In our scenario, we focus on first encounters
between our agent and people. We show how the agent is able to exploit his Person Memory to acquire
information about others during Small Talk and guide the conversation.
1 MOTIVATION
When two people meet for the first time they often
would like to know from each other: Who are you?
How do humans acquire information of others they
encounter for the first time? How and what informa-
tion do they remember from each other?
Research on human-like memory for virtual
agents has gained a lot of attention recently. Agents
equipped with autobiographical memory, e.g. (Kasap
et al., 2009), or episodic memory, cf. (Brom et al.,
2007), are enabled to remember their own experi-
ences. Companion agents greatly benefit from these
memories, in that information from previous interac-
tions can be accessed and used by the agent. How-
ever, in the above approaches the memories are cen-
tered on the agent’s experiences, yet people the agent
interacted with play a minor role.
In this work we present a model of Person Mem-
ory for an embodied conversational agent that focuses
on the representation of people. Our agent MAX has
a cognitive architecture based on BDI. The agent re-
sides on the hallway of the AI group of Bielefeld Uni-
versity. A broad range of Small Talk knowledge al-
lows for short enjoyable interactions with him. But
over the long run, conversations with our agent are
more or less the same. So it is not easy to build
up some kind of relationship over time, since people
may get annoyed by repetitive interactions. To enable
our agent to build up longer lasting relationships, he
should be put in the position to adapt to his interlocu-
tors. In that, he must be able to distinguish between
different kindsof people, in order to know what to talk
about. To remember past interactions would allow
him to avoid repetition of things already said, and to
pick up interesting topics of previous conversations.
Small Talk is considered important to increase
trust and familiarity (Bickmore and Cassell, 2001)
between virtual agents and their interaction partners.
Yet, most of the systems equipped with small-talk
abilities are restricted to common topics, like the cur-
rent weather. Small Talk about topics of interest to
certain individuals has been mostly neglected. In
this paper, we show how information of social cate-
gories can be exploited in order to acquire information
about, and guide a conversation with, a new acquain-
tance.
The paper is organized as follows. In Section 2 re-
lated work is reviewed. In order to describe the Person
Memory component, in Section 3, we review work of
social psychology, dealing with person perception and
social categories. The architecture and Person Mem-
ory is described in Section 4. In Section 5 we present
our scenario, and show how social categories can be
used to determine topics of interest to different kinds
of people. A Summary and future work conclude this
paper.
98
Mattar N. and Wachsmuth I..
WHO ARE YOU? - On the Acquisition of Information about People for an Agent that Remembers.
DOI: 10.5220/0003710900980105
In Proceedings of the 4th International Conference on Agents and Artificial Intelligence (ICAART-2012), pages 98-105
ISBN: 978-989-8425-96-6
Copyright
c
 2012 SCITEPRESS (Science and Technology Publications, Lda.)

2 RELATED WORK
Stereotypes have been found an adequate way in rec-
ommender systems in order to tailor recommenda-
tions to different kinds of people. E.g. GRUNDY
(Rich, 1979) demonstrated how stereotypes can be
used to recommend books to people. GRUNDY
prompted the user for some personality traits at the
beginning of the interaction. It then used this infor-
mation to assign stereotypes to be able to give recom-
mendations on books.
Since GRUNDY, a lot of research in the area of
recommender systems has been done. Goal of these
systems is to provide recommendations on e.g. news,
books, or movies that match the interest of their users
as close as possible. Recommender systems are usu-
ally classified into content-based, collaborative, and
hybrid approaches. For an overview refer to (Ado-
mavicius and Tuzhilin, 2005).
Systems that provide user models for recom-
mender systems developedinto powerful tools. Kobsa
provides an overview of generic user modeling sys-
tems (Kobsa, 2007). But, conversational agents differ
from traditional recommender systems and their user
models. Most conversational agents are designed to
support users in certain tasks, like learning, diet plan-
ning etc., thus the relation between agent and interac-
tion partner plays an important role.
Research on relational agents focuses on the
establishment of long-term relationships with their
users. In (Bickmore and Cassell, 2001) an extended
version of REA, an embodied conversational agent, is
described. Besides of task-related talk, REA is capa-
ble of engaging in Small Talk to deepen the relation-
ship to her interlocutors. While building and main-
taining long-term relationships, and even reidentify-
ing users (Schulman et al., 2008), is a central part,
little information is given on how persons the agents
interact with are represented.
The relationship between agent and interlocutor
has shown to be improved when the agent engages in
Small Talk (Campbell et al., 2009). But, the complex
structure of Small Talk has been mostly neglected. In
a recent work of (Endrass et al., 2011) cultural dif-
ferences in Small Talk are investigated and imple-
mented for multi-agent systems. Following (Schnei-
der, 1988), they categorize Small Talk topics into
three different categories: immediate, external, and
communication situation.
More recently, systems with human-like mem-
ory gained attention. Agents equipped with autobi-
ographic or episodic memories are able to remember
past events they experienced either in a virtual world
on their own (Brom et al., 2007), or with people they
interacted with (Kasap et al., 2009). The agent de-
scribed in the latter, e.g. stores the user and the change
of relationship in its episodic memories. However,
again there is no information on how the user is rep-
resented by the agent. In general, besides the fact
that user modeling is commonly employed in recom-
mender systems, most conversational agents lack a
clear representation of the people they interact with.
3 PERSON PERCEPTION
In order to equip a virtual agent with a representa-
tion of people several questions need to be addressed.
How do humans perceive others? Is there a difference
between the representation of people and other world
objects in humans? How do humans acquire knowl-
edge about others they have not met before? What do
humans remember about other people?
Humans try to simplify the surrounding world by
making use of categories. Categorization helps to fo-
cus on important things by reducing information. Not
only do humans categorize world objects like cups
and cars, but we categorize other people we meet as
well (Stangor and Lange, 1994).
While natural objects are mostly perceived
through their form and function (Rosch as cited in
Dahlgren, 1985), members of social groups are per-
ceived in a more complex way. E.g., it is not al-
ways possible to conceive a social category by ob-
serving someone, but the social role of the individual
needs to be clear. Furthermore, more information is
stored in the representation of a social category, such
as social function, relation between members of cate-
gories, and internal traits (Dahlgren, 1985).
Social categories consist of stereotypical informa-
tion humans have about a certain group of people
(Hamilton, 1979). Humans are able to assign some
social categories even before actually talking to some-
body, e.g. categories like age or gender. The impres-
sion of the other is refined when more information is
acquired, e.g. during conversation. Information like
occupation, place of residence, and individual inter-
ests allow for a fine grained categorization. Social
categories are used as a basis of encoding and the as-
sociated information influence judgment of and be-
havior towards others (Stangor and Lange, 1994).
To sum up, in order to represent people in a cogni-
tive architecture of a virtual agent, it seems not suffi-
cient to only consider individuals the agent has inter-
acted with. Person perception in humans is strongly
influenced by knowledge one has about groups of
people, i.e. social categories. To equip an agent with
such knowledge will enhance its capabilities of react-
WHO ARE YOU? - On the Acquisition of Information about People for an Agent that Remembers
99

ing to different kinds of people the agent has not met
before, and the storage of information about people,
as well.
3.1 What to Remember about People
We identified the following five types of information
to be stored about people which form the basis of our
conception of person memory:
Biographical Facts. During conversation people re-
veal facts about themselves, like their age, occu-
pation, hometown, etc. These kinds of informa-
tion can be compared to autobiographical facts
(Conway, 1987) one remembers about oneself.
Preferences and Interests. Preferences and inter-
ests are valuable when trying to find topics to talk
about or that should be avoided during conversa-
tion. Furthermore, they can be used to determine
similarity between people.
Personality Traits. Information about the personal-
ity of the interlocutor allows to adapt the agent’s
personality. It has been shown that people prefer
to talk to an agent that matches their personality
(Bickmore and Cassell, 2001).
Events and Episodes. When we think about a spe-
cific person previous events one experienced with
the other come to mind. In order to access signif-
icant events the agent experienced with someone,
links to these events should be stored in the rep-
resentation of a person. This enables the agent to
bring up a previous encounter that was significant
for their relationship during conversation.
Relationship Information. Information about the
relationship between the agent and a person en-
ables the agent to determine appropriate topics for
Small Talk (Endrass et al., 2011). Furthermore,
the agent’s emotions are especially influenced by
the relationship to an interlocutor: the agent might
be happier when meeting someone he likes than
meeting someone he has never met before.
4 A PERSON MEMORY FOR A
VIRTUAL AGENT
To address these topics, the cognitive architecture is
extended by a Person Memory to allow virtual agents
to keep track of people they interacted with, and an
Event Memory (not discussed in this paper) for stor-
age and retrieval of past interactions.
Figure 1: The cognitive architecture of the agent extended
with a memory component.
Figure 2: Person and Event Memory of the cognitive archi-
tecture.
4.1 Cognitive Architecture
The cognitive architecture consists of a reactive and
a deliberative perceive-reason-act part (figure 1 top).
While the reactive part directly maps sensor informa-
tion to behaviors, the deliberativepart relies on a BDI-
kernel based on JAM (Huber, 1999). The reactive and
deliberative parts are executed concurrently and cor-
responding behaviors can overrule each other in order
to control the agent (Lessmann et al., 2006). Work-
ing memory (WM) consists of JAM beliefs about the
current situation and the world. In our architecture
the term working memory has to be understood as a
technical term. At present there is no limitation of
information like in human working memory.
Since there was no clear distinction between WM
and long-term memory (LTM) in the agent’s architec-
ture, we introduced a new memory component (figure
1 bottom). It allows access of long-term memories
during reasoning and is composed of two different
modules: Person Memory and Event Memory (fig-
ure 2). The memory modules are connected through
a gatekeeper with the BDI architecture. The gate-
keeper is able to select what information is passed
from and to the working memory of the agent. PMPU
and EMPU are the processing units of both modules
ICAART 2012 - International Conference on Agents and Artificial Intelligence
100

Agent: What is your name?
Visitor: My name is Paul.
<rule>
<match>
<convfunction type="provide.content.name" />
<pattern>My name is $name</pattern>
</match>
<action>
<command function="memorize"
arguments="name:$name"/>
</action>
</rule>
Figure 3: The agent asks his interlocutor for his name (top).
Dialog rule matching the interlocutors response, with en-
coding and transferral of the name into working memory as
corresponding action (bottom).
JAM WM LTM
a)
b)
Figure 4: Types of access to WM and LTM and flow of
information. a) The long-term memory can be queried for
information by JAM rules. Available information is trans-
ferred into WM. b) Activation of related long-term mem-
ories occurs when information enters WM. The additional
information is transferred into WM.
(Person Memory and Event Memory processing unit).
They carry out retrieval, storage, and processing of in-
formation.
The dialog engine of our system follows a pattern
matching approach. Utterances of the interlocutor are
processed by JAM rules. If a rule matches the input
the according JAM plan is executed. Figure 3 shows
an example of the encoding and memory transferral
process. A dialog rule matching the response of the
interlocutor to the agents question triggers the trans-
ferral to WM. The information is encoded as a slot-
value pair. Accordingly, there has to be a dialog rule
for every piece of information our agent should be
able to remember.
4.1.1 Activation
Incoming parts of information can lead to an activa-
tion of already memorized information (figure 4 b).
In the conversation of figure 5 (top), a visitor pro-
vides his name to the agent. The information is trans-
ferred into working memory, passes the gatekeeper,
and leads to an activation of information known about
Agent: Hello, I am MAX. What is your
name?
Visitor: Hi MAX, It is me, Paul.
[Activation of additional LTM
information and transferral to WM]
Agent: Paul from Bielefeld?
[Use of information from WM]
Visitor: Yes, that’s right.
Agent: Nice to meet you again!
Visitor: How many people from Bielefeld
do you know?
Agent: Hmm, I dont know.
[No information in WM]
Agent: Let me think about it!
[Query LTM for information]
Agent: I have met 6 people from
Bielefeld.
Figure 5: Fragment of a conversation between our agent
and a visitor the agent met before. Additional information
is activated and transferred into WM (top). Information is
explicitly querried from LTM (bottom).
a person with that name inside the long-term memory.
The activated information is transferred into working
memory. In the example, the agent is thus able to
use the information and ask an additional question in
order to clarify if the person activated in the Person
Memory refers to the right one.
4.1.2 Recall
As not all the information stored in long-term mem-
ory is activated and transferred into WM at once, the
agent needs a way to recall information form his long-
term memories. In figure 5 (bottom), a short fragment
of a conversation with our agent demonstrates a use-
case of direct access to LTM. The agent does not have
enough information in his WM to answer the question
of the visitor. In this case he queries his LTM for the
number of people he knows from Bielefeld and is able
to answer the question (figure 4 a).
4.2 Representation of Social Categories
In the Person Memory, social categories are defined
in a frame like manner. Table 1 depict excerpts of so-
cial categories of our system. Every slot represents
a stereotypical information that is associated with the
social category. Stereotypical information in social
categories can consist of the first three types of infor-
mation introduced in Section 3.1: Biographical facts,
preferences and interests, and personality traits. In
this work we focus on interests in certain topics.
A confidence value c ∈ [0, 1] and a modifier value
m ∈ (−∞, ∞) are associated with each stereotype. The
confidence of a stereotypical information specifies
WHO ARE YOU? - On the Acquisition of Information about People for an Agent that Remembers
101

Table 1: Excerpts of social category frames: male, female,
computer science student, and sports student.
Slot Value Conf. Mod.
interestedin soccer 0.6 2
interestedin shopping 0.2 -5
...
Slot Value Conf. Mod.
interestedin soccer 0.4 -3
interestedin shopping 0.8 5
...
Slot Value Conf. Mod.
interestedin c. games 0.6 5
interestedin soccer 0.3 -4
interestedin shopping 0.2 -3
...
Slot Value Conf. Mod.
interestedin c. games 0.2 -5
interestedin soccer 0.8 5
interestedin shopping 0.7 3
...
how probable it is that the information applies for a
member of the category. This allows triggering cer-
tain behaviors if a stereotype proves wrong during
conversation. E.g. the agent could be surprised if
he talks to a computer science student who does not
like to play computer games, but loves to play soc-
cer, since his expectancies due to his social categories
would be the opposite. The modifier allows altering
the utility of a stereotypical information for conversa-
tion. How the utilities of information can be calcu-
lated are described in Section 5.1 .
4.3 Representation of Persons
The Person Memory stores information about the
agent himself, people the agent met, and people the
agent heard about. The representation of a person fol-
lows a frame-based approach, as well. Table 2 shows
an empty person frame.
The initial person frame is sparse and only con-
tains slots for a few important information about a
person. The confidence value is used to annotate in-
formation acquired directly from the interlocutor dur-
ing conversation (high confidence), and information
that is e.g. inferred from other information (low confi-
dence). The modifier, again, is used to alter the utility
for an information for conversation.
Following (Rich, 1979), every social category is
annotated with one or more facts that trigger their ac-
tivation. If a fact about a person is revealed during
Table 2: Empty person frame.
Slot Value Conf. Mod.
id
firstname
lastname
gender
age
hometown
social category
...
interest
com. situation
ext. situation
open
reason
origin
im. situation
greeting
Topic Category Topic
ext. situation studies
com. situation
farewell
Figure 6: Structure of the testbed dialog. Topics are catego-
rized in three categories: immediate, external, and commu-
nication situation.
conversation and triggers a social category, the cate-
gory is assigned to the person. From this point, the
stereotypes can be used by the agent, for example, to
select topics that might be interesting to the interlocu-
tor. Depending on the response of the interlocutor the
stereotype might be confirmed or rejected and can be
included in the corresponding person frame.
As one acquires specific knowledge about an indi-
vidual, one is less likely to make use of the social cat-
egories associated with the person (Senay and Keysar,
2009). Therefore, further uses of this information will
refer to the information in the person frame instead of
the one in the social category.
5 SCENARIO
As a testbed for the Person Memory, we designed a di-
alog scenario. Since our agent resides on the hallway
of our group, mostly students approach the agent dur-
ing the day. In our scenario, our agent and a student
meet on our hallway and have a short conversation,
while waiting. To model our dialog, we analyzed the
initial-meeting dialogs of participants of the CUBE-G
corpus (Rehm et al., 2007), and used their structure as
ICAART 2012 - International Conference on Agents and Artificial Intelligence
102

a guideline.
Figure 6 depicts the structure of our dialog. Clas-
sification of topics follows the classification used in
(Endrass et al., 2011). After the initial greeting, the
agent asks the interlocutor why he is waiting, and then
proceeds talking about the student’s subject of study.
Then the interlocutor is asked for his origin.
Up to this point the dialog topics are fixed. This
enables the agent to acquire some initial social cate-
gories. Next, further topics can be discussed during
an open slot in the dialog structure.
We now present how social categories can be ex-
ploited to determine topics that are suitable to talk
about as the conversation continues.
5.1 Exploiting Social Categories
Social categories and the associated stereotypical in-
formation are well suited for Small Talk. Exploit-
ing social categories, an ECA is able to ascribe at-
tributes to people with very little information about
them. This enables the agent to identify topics that
might be interesting to someone, rather than suggest-
ing what to talk about out of the blue.
To select suitable topics for conversation, not only
the interests inferred from the social categories of the
interlocutor have to be taken into account. Since we
aim at building a human-like interaction partner, our
agent is equipped with own interests that might inter-
fere with the ones of his interaction partner. There-
fore, the interests of the agent need to be considered
when selecting a topic for conversation.
5.1.1 Combining Social Categories
For topic selection, stereotypical information of all
social categories of a person are combined to find a
suitable topic.
The utility u(I) of an information I of a social cat-
egory C is calculated according to (1).
u(I) = m(I) ∗ c(I) (1)
In equation (1), m(I) denotes the modifier, c(I) the
confidence of an information I. The total utility value
is acquired by adding all the single utility values of an
information:
for (Information i : informations) {
for (SocialCategory sc : categories) {
utility_i += u( sc.get(i) );
}
}
Figure 7 demonstrates how the utility values for
each interest of a person change during conversation.
Figure 7: Adaptation of normalized utility values of inter-
ests of a person, as more social categories are being ac-
quired.
Figure 8: Utility values of the agent’s, and an interlocutor’s
interests, and calculated final utility values.
5.1.2 Topic Selection
As stated above, not only the interlocutor’s interests
are considered, but the agent’s interests as well. Fur-
thermore, the course of conversation influences the
utility of a topic. In that, a topic that was already
talked about has a lower utility. The final utility u
f
(I)
is calculated according to equation (2).
WHO ARE YOU? - On the Acquisition of Information about People for an Agent that Remembers
103

u
f
(I) =
u
P
(I) ∗ c
A
(I) + u
A
(I) ∗ c
P
(I)
2
∗ r(I) (2)
where u
P
, u
A
denote the utility, and c
P
, c
A
the con-
fidence values for information I of the person and the
agent. r(I) denotes the recency of an information.
Figure 8 shows the final utility values of our
agent’s and an interlocutor’s interests. The informa-
tion with maximum utility, soccer in this case, will be
suggested as topic. After a suggestion of a topic, its
recency value is decreased to prevent that the topic is
suggested again during the ongoing conversation.
5.1.3 Example Dialogs
In this section, we present an example dialog between
our agent and two different interlocutors. We used
four different social categories, with a total of five po-
tential topics to talk about.
Our agent belongs to the social categories “male”,
and “computersciencestudent”. The first interlocu-
tor belongs to the categories “female” and “sportsstu-
dent” (refer to figure 8 for the corresponding utility
values), the second interlocutor belongs to the cate-
gories “male” and “computersciencestudent”.
Agent: Hi, my name is Agent. What is your name?
Paula: Hello Agent, my name is Paula.
Agent: I have never met you here before, Paula.
Are you waiting for someone?
Paula: Yeah, I am waiting for a friend of mine.
Agent: Do you study computer science? Most of
the people I meet here do.
Paula: No. I am a student of sport sciences.
Agent: Oh, nice to finally meet someone from
another discipline.
Agent: So do you live here in Bielefeld?
Paula: Yes, I moved here two years ago.
Agent: Oh, ok.
Agent: Did you see the last match of
Bielefeld’s soccer club?
Paula: Yes, that was a great game.
...
Figure 9: Example dialog 1: Excerpt of a conversation be-
tween our agent and a female sports student.
At first, the conversationsdepicted in figures 9 and
10 do not differ in terms of discussed topics. The ini-
tial questions enable the agent to fill slots of the as-
sociated person frames with information provided by
the interlocutors, and with information inferred from
the interlocutors’ answers. As soon as the open topic
part of the dialog is reached (refer to figure 6), the
agent determines topics by using his knowledge about
the interlocutors inferred from the social categories.
Agent: Hi, my name is Agent. What is your name?
Paul : Hello Agent, my name is Paul.
Agent: I have never met you here before, Paul.
Are you waiting for someone?
Paul : Yeah, my friend told me to be here at 3,
but he didn’t show up, yet.
Agent: Are you a student?
Paul : Yes, I study computer sciences.
Agent: Seems a common place for computer
scientists to hang out.
Agent: Are you from Bielefeld, or did you move
here to study?
Paul : I was born here in Bielefeld.
Agent: As a student of computer science you
probably love to play computer games,
don’t you?
Paul : I used to play a lot. Recently I don’t
have enough time to play.
...
Figure 10: Example dialog 2: Excerpt of a conversation
between our agent and a male computer science student.
Figure 11: Person frame of an interlocutor after initial con-
versation.
E.g., in the first dialog, the agent associates the so-
cial category “sportsstudent” with the interlocutor, af-
ter the interlocutor told she is studying sport sciences.
Later on, the agent infers from this social category
that sport students are interested in soccer. Therefore
he introduces soccer as a topic. Figure 11 shows the
agent’s representation of the second interlocutor after
the conversation depicted in figure 10.
6 SUMMARY AND OUTLOOK
In this work we presented a model of Person Memory
for our embodied conversational agent. It enables him
to remember information about people he interacted
with, and use this information in subsequent interac-
tions. In contrast to traditional user modeling, the aim
of our Person Memory is not to completely adapt the
agent to individuals, but to enable him to respond to
different people appropriately, and in a more human-
like fashion.
Small Talk, while often used in virtual agents to
ICAART 2012 - International Conference on Agents and Artificial Intelligence
104

keep things going, has not been exploited to its full
extent. Research on Small Talk shows that it is not
as random as often perceived. It does follow certain
rules, depends on the situational context, the relation-
ship between the interlocutors, and goes further than
chatting about the weather. We demonstrated how so-
cial categories can be applied to infer topics one could
talk about, during a first encounter. A benefit of social
categories and stereotypical information is the mini-
mal amount of information that is needed about the
other in order to come up with a first impression.
The core of our Person Memory has been imple-
mented and demonstrated in a small scenario. Yet, a
lot of parameters allow for a fine tuning of the system.
Determining these parameters is a tedious task. A
thorough interaction study is needed in order to iden-
tify an optimal set of suitable parameters. Further-
more, we only demonstrated a small subset of the ca-
pabilities of our Person Model. Since we focused on
first encounters of people, no information about the
relationship between agent and interlocutor is avail-
able. But behavior towards others is strongly influ-
enced by how good we can relate to each other. E.g.,
as stated above, the choice of topics during Small Talk
is influenced by the relationship of the interlocutors.
Therefore, as a next step it has to be investigated how
the information of a first encounter, as presented here,
can be used in further interactions. The integration of
relationship information in equation (2) will e.g. fur-
ther enhance topic selection.
To increase believability, emotions of the agent
should be adapted to the situation, as well. As sug-
gested in section 3.1, behavior towards unknown peo-
ple should differ from behavior towards people the
agent already met. Information of previous inter-
actions, like attitude of the interlocutor towards the
agent, will allow altering the agent’s emotions and
mood. Awareness of what was talked about in previ-
ous encounters will help to prevent that the conversa-
tion will get annoying over time: It will allow picking
up topics of interest of both interlocutors and prevent
repetition of already said things.
To conclude, the Person Memory presented in this
work provides a solid foundation for further research
on human-agent long-term relationships. It enhances
the agent’s awareness of persons he interacted with
and allows the agent to react to individuals in a more
human-like fashion.
REFERENCES
Adomavicius, G. and Tuzhilin, A. (2005). Toward the next
generation of recommender systems: A survey of the
state-of-the-art and possible extensions. IEEE trans-
actions on knowledge and data engineering, pages
734–749.
Bickmore, T. and Cassell, J. (2001). Relational Agents: A
Model and Implementation of Building User Trust. In
Proceedings of the SIGCHI conference on Human fac-
tors in computing systems, CHI ’01, pages 396–403,
New York, NY, USA. ACM.
Brom, C., Peˇskov´a, K., and Lukavsky, J. (2007). What Does
Your Actor Remember? Towards Characters with a
Full Episodic Memory. In Proceedings of the 4th
ICVS, ICVS’07, pages 89–101, Berlin, Heidelberg.
Springer-Verlag.
Campbell, R., Grimshaw, M., and Green, G. (2009). Re-
lational agents: A critical review. The Open Virtual
Reality Journal, 11(7).
Conway, M. A. (1987). Verifying autobiographical facts.
Cognition, 26(1):39 – 58.
Dahlgren, K. (1985). The Cognitive Structure of Social Cat-
egories. Cognitive Science, 9(3):379–398.
Endrass, B., Rehm, M., and Andr´e, E. (2011). Plan-
ning Small Talk behavior with cultural influences for
multiagent systems. Computer Speech & Language,
25(2):158–174.
Hamilton, D. (1979). A cognitive-attributional analysis of
stereotyping. Advances in experimental social psy-
chology, 12:53–84.
Huber, M. (1999). JAM: A BDI-theoretic Mobile Agent Ar-
chitecture. In Proceedings of the third annual confer-
ence on Autonomous Agents, pages 236–243. ACM.
Kasap, Z., Ben Moussa, M., Chaudhuri, P., and Magnenat-
Thalmann, N. (2009). Making Them Remember –
Emotional Virtual Characters with Memory. Com-
puter Graphics and Applications, IEEE, 29(2):20–29.
Kobsa, A. (2007). Generic User Modeling Systems. In The
adaptive web, pages 136–154. Springer-Verlag.
Lessmann, N., Kopp, S., and Wachsmuth, I. (2006). Situ-
ated interaction with a virtual human - perception, ac-
tion, and cognition. In Rickheit, G. and Wachsmuth,
I., editors, Situated Communication, pages 287–323.
Mouton de Gruyter, Berlin.
Rehm, M., Andr´e, E., Bee, N., Endrass, B., Wissner, M.,
Nakano, Y., Nishida, T., and Huang, H. (2007). The
CUBE-G Approach - Coaching Culture-Specific Non-
verbal Behavior by Virtual Agents. In Proceedings of
the 38th Conference of ISAGA.
Rich, E. (1979). User Modeling via Stereotypes. Cognitive
science, 3(4):329–354.
Schneider, K. (1988). Small Talk: Analysing Phatic Dis-
course. Hitzeroth, Marburg.
Schulman, D., Sharma, M., and Bickmore, T. (2008). The
Identification of Users by Relational Agents. In Pro-
ceedings of the 7th international joint conference on
Autonomous agents and multiagent systems-Volume 1,
pages 105–111.
Senay, I. and Keysar, B. (2009). Keeping Trackof Speaker’s
Perspective: The Role of Social Identity. Discourse
Processes, 46(5):401–425.
Stangor, C. and Lange, J. (1994). Mental representations
of social groups: Advances in understanding stereo-
types and stereotyping. Advances in experimental so-
cial psychology, 26:357–357.
WHO ARE YOU? - On the Acquisition of Information about People for an Agent that Remembers
105