LET’S TALK TOPICALLY WITH ARTIFICIAL AGENTS!

Providing Agents with Humanlike Topic Awareness

in Everyday Dialog Situations

Alexa Breuing and Ipke Wachsmuth

Artiﬁcial Intelligence Group, Bielefeld University, Bielefeld, Germany

Keywords:

Embodied conversational agents, Human-agent interaction, Topic detection, Information retrieval, Wikipedia.

Abstract:

Spoken interactions between humans are characterized by coherent sequences of utterances assigning a the-

matical structure to the whole conversation. Such coherence and the success of a meaningful and ﬂexible

dialog are based on the cognitive ability to be aware of the ongoing conversational topic. This paper presents

how to enable such topically coherent conversations between humans and interactive systems by emulating

humanlike topic awareness in artiﬁcial agents. Therefore, we ﬁrstly automated human topic awareness on

the basis of preprocessed Wikipedia knowledge and secondly transferred such computer-based awareness to

a virtual agent. As a result, we contribute to improve human-agent dialogs by enabling topical talk between

human and artiﬁcial conversation partners.

1 MOTIVATION

Topic awareness plays an important role in human

conversations. Besides resolving linguistic references

and ambiguities which often arise in natural lan-

guage talks, it enables the interlocutors to interaction-

ally produce coherent sequences of spoken utterances.

More precisely, every spoken contribution may raise

new potential topics whose actual realization depends

on the co-participant’s acceptance by picking up one

of these topics within his or her reply (Svennevig,

1999). Hence, a topic can be described as a joint

project (Clark, 1996) as it is jointly established dur-

ing ongoing conversations. Furthermore, being aware

of topics helps us to touch the right subject according

to the social circumstances enclosing the interactional

situation. Assuming an everyday small talk conver-

sation, for example, so-called unsafe topics such as

religion and death should be avoided (Endrass et al.,

2011). Altogether, the competence to talk topically

constitutes a basic requirement to carry on meaning-

ful, ﬂexible, and appropriate conversations with other

persons.

Embodied conversational agents (ECAs) are vir-

tual characters possessing humanlike conversational

behaviors to establish an intuitive human-machine in-

terface (Cassell et al., 2000). That is, they are ca-

pable of holding face-to-face conversations with hu-

mans by understanding and producing speech, ges-

tures, and facial expressions. Nevertheless, they often

fail to converse in great depth and hence to mutually

establish a topical talk with their human opponent. In

addition, many ECAs lack in simulating a sense for

the adequacy of certain topics during dialog. To rem-

edy these weaknesses, the artiﬁcial interlocutor needs

to be aware of ongoing and potential conversational

topics like humans.

To provide conversational agents with artiﬁcial,

humanlike topic awareness in everyday interactions

two main tasks need to be automatized: First, the de-

tection of topics raised in ongoing natural language

dialogs and second, the adequate integration of the

resulting topic information into the agent’s underly-

ing system architecture. This paper introduces an ap-

proach tackling both tasks: We show how to connect

well-established linguistic information retrieval meth-

ods with beneﬁts originated from collaborative work

provided by Wikipedia to automatically detect dialog

topics. Additionally, we present how to utilize the ob-

tained information to improve the conversational abil-

ities of virtual computer characters regarding topic

handling.

The rest of the paper is organized as follows. In

the next section we introduce our notion of dialog

topics establishing the basis for the present work.

Subsequently, the several processes of our automatic

topic detection approach are described in Section 3.

Thereby we especially emphasize the application of

Breuing A. and Wachsmuth I..

LET’S TALK TOPICALLY WITH ARTIFICIAL AGENTS! - Providing Agents with Humanlike Topic Awareness in Everyday Dialog Situations.

DOI: 10.5220/0003745900620071

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 62-71

ISBN: 978-989-8425-96-6

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

collaborative knowledge provided by Wikipedia. Sec-

tion 4 highlights the embedding of the resulting topic

information into the existing architecture of the con-

versational agent Max. As a result, we contribute

in emulating humanlike topic awareness in artiﬁcial

agents as described by means of our dialog scenario

in Section 5. Moreover, we present how to evaluate

our model in the near future. In Section 6 we give

an overview of related work before closing the paper

with a short conclusion and discussion.

2 INTRODUCING DIALOG

TOPICS

Assuming dialogs to be face-to-face conversations be-

tween two partners, a dialog topic emerges from a

joint activity performed by both interlocutors (Juraf-

sky and Martin, 2009). That is, considering single

utterances to specify a dialog topic is insufﬁcient as

they do not have topics in isolation. They rather pro-

vide topic suggestions (Schank, 1977). However, the

topic formulation of the particular topic is done at dif-

ferent levels of abstraction and from different subjec-

tive positions (Svennevig, 1999). Speaker A, for ex-

ample, might categorize a dialog about Whiskey and

Brandy using the term “Alcohol”, whereas speaker B

might choose the term “Drinks” or “Spirits” referenc-

ing the same topic. According to this, we deﬁne a

dialog topic to be an independent, self-selected cat-

egory superordinate to a co-constructed sequence of

dialog contributions (Breuing et al., 2011).

2.1 Topic Shifts

A dialog topic subordinates a sequence of coherent

dialog contributions as wholes (Bublitz, 1989; Sven-

nevig, 1999). Hence, they generalize the concepts

mentioned in these contributions to a certain degree.

A potential topic shift in dialogs occurs, once previ-

ous concepts and concepts coming up subsequently

cannot be generalized to one topic anymore. If at-

tending to the new concepts opens a completely dif-

ferent dialog topic and comes along with a drop of the

present one, we refer to this kind of shift as topic leap

(Svennevig, 1999).

On the other hand, a topic shift might happen

gradually. Imagine the following dialog sequence:

A: “In which city do you live?”

B: “Munich.”

A: “Ah, then you are a fan of Bayern Munich?”

B: “Actually no. I like Arsenal.”

By mentioning the concept “city”, speaker A sug-

gests to talk about places. Speaker B agrees to this

topic by replying with an utterance containing the

concept “Munich” specifying a German city. “Mu-

nich” in turn is unrelated to the upcoming topic

“Sports”, however, it is conceptually closely con-

nected to Munich’s local soccer club “Bayern Mu-

nich”. Thus, the dialog merges seamlessly from the

topic “Places” to the topic “Sports”. Hobbs calls this

phenomenon topic drift (Hobbs, 1990).

2.2 Selection of Dialog Topics

Raising an issue requires choosing a dialog topic ﬁrst.

Thereby, the amount of possible topics is constrained

due to the given dialog scenario, the personal relation

between the dialog partners, and their cultural back-

ground. Accordingly, not every dialog topic is appro-

priate for everyday small talk conversations.

Referring to Schneider (Schneider, 1988), there

are three groups of basic options for topic selection:

(1) The immediate situation involves all topics ad-

dressing the frame of the dialog situation.

(2) The external situation represents the larger con-

text of the immediate situation and hence of its

topics.

(3) The communication situation refers to the con-

versation partners and holds private topics such as

hobbies or family.

A typical small talk starts with a topic related to the

immediate situation and continues with topics from

the external or communication situation. Due to these

social conventions, most small talk structures are very

similar and ease striking up a conversation with other,

especially unknown persons.

3 AUTOMATIC DETECTION OF

DIALOG TOPICS

Constituting a matter of course for humans, the au-

tomatic detection of dialog topics poses a great chal-

lenge. Given a dialog situation as deﬁned before, it

has to meet several requirements. First of all, the

underlying processes have to work online. As di-

alogs are continuous and demand adaptive moment-

by-moment decisions, it is necessary to incessantly

provide the system with information about the current

topic situation directly inﬂuencing the agent’s conver-

sational behavior. Additionally, this information has

to be processed within a short time frame to guar-

antee humanlike reaction time. Moreover, the wide

LET'S TALK TOPICALLY WITH ARTIFICIAL AGENTS! - Providing Agents with Humanlike Topic Awareness in

Everyday Dialog Situations

range of possible topics, for example being discussed

in everyday conversations, calls for a dynamic han-

dling of previously unknown contributions. This in

turn presumes an access to huge amounts of previ-

ously unlearned topics and how they are correlated.

According to the dynamic factor and for further rea-

sons assigned subsequently, the online encyclopedia

Wikipedia proved to be the ideal knowledge source.

3.1 Topics Provided by Wikipedia

According to our deﬁnition, dialog topics are con-

sidered to be categories subordinating a sequence of

dialog contributions. The Wikipedia category sys-

tem is composed of categories subordinating articles

presented by natural language texts. Utilizing the

similarity between dialog-based utterance-topic rela-

tions and Wikipedia-based article-category links con-

stitutes the basis for our dynamic topic detection ap-

proach. Generally speaking, we identify a dialog

topic by mapping the several utterances to Wikipedia

articles and specifying their shared Wikipedia cate-

gories as potential topics. Thus, the detection process

does not simply consider the terms contained in the

utterances, but is capable of identifying a topic t with-

out having a priori knowledge of the domain underly-

ing t.

A big advantage of accessing Wikipedia for our

purpose is the fact that its encyclopedic knowledge is

constructed collaboratively by numerous volunteers.

Hence, Wikipedia provides huge amounts of infor-

mation whose maintenance is done by others. Fur-

thermore, the resulting description and categorization

of concepts reﬂect the participants’ perception of con-

ceptual structures and delivers insights into the human

understanding of topics and their relations.

3.2 Online Detection

Within our approach, realizing an automatic topic de-

tection mainly involves the implementation of auto-

matic processes that identify potential topics, track

ongoing topics, detect topical shifts, and label the co-

herent dialog sequences. To ensure an online work-

ing topic detection the ﬁrst two tasks need to be per-

formed continuously, that is on every incoming utter-

ance. Their outcomes simultaneously affect the re-

maining processes. In the following, the several tasks

are described in more detail. Additionally, Figure 1

gives an overview of the presented topic detection ap-

proach and illustrates the relations between its associ-

ated processes.

3.2.1 Identiﬁcation of Potential Topics

Referring to Schank (1977), an utterance said in re-

sponse to an input provides both a conceptual inter-

section to the present dialog topic and a new concep-

tualization introducing potential new topics. Accord-

ingly, to automatically identify potential topic direc-

tions, at ﬁrst every single dialog contribution has to

be conceptualized by identifying its contained con-

cept terms. Therefore, the system ﬁrst preprocesses

the present utterance by means of the Stanford Part-

Of-Speech Tagger (Toutanova and Manning, 2000).

Afterwards, all nouns and proper nouns are identiﬁed

and speciﬁed as concept terms. Moreover, the system

extracts the verbs contained in the present utterance

and transforms them to their substantive as provid-

ing potential conceptual information as well. There-

fore we make use of the online dictionary Wiktionary.

Then, the system searches for a Wikipedia article giv-

ing a concept description for the substantive. If a cor-

responding article can be found, as for example given

for the term “swimming”, the substantive is consid-

ered as a concept term furthermore. In case no article

is found, the substantive is not considered as a con-

cept term as probably not providing conceptual infor-

mation (like the term “doing”). In addition, auxiliary

verbs such as “having” are excluded in the ﬁrst place.

In order to detect named entities consisting of

more than one word, adjectives and/or nouns, and

proper nouns appearing successively are tested for

their lexical “togetherness”. Therefore we make use

of the concept information provided by Wikipedia in

terms of single articles (Gabrilovich and Markovitch,

2007). More precisely, each of these potential named

entities are mapped onto the set of all Wikipedia arti-

cles A

wiki

twice: once as a whole and once noun-wise.

This mapping process is accomplished via a mapping

function

f : cterm 7→ A

wiki

(1)

where cterm is either the potential named entity or

a single noun. To realize f , we built up an Apache

Lucene (McCandless et al., 2010) search index con-

taining documents for every Wikipedia article includ-

ing information about their titles, textual descriptions,

textual anchors of their incoming links, and redirects.

This allows us to estimate both mappings by means of

the Lucene similarity score

score(q, d) = Σ

t∈q

(t f (t ∈ d)·id f (t)·b

·n(q, d)) (2)

where t f (t ∈ d) speciﬁes the term frequency of each

term t ∈ cterm in d, id f (t) indicates the general im-

portance of t within all documents, b

refers to the

ICAART 2012 - International Conference on Agents and Artificial Intelligence

Figure 1: Overview of the processes involved in our automatic topic detection approach.

ﬁeld boost in case of an exact match of cterm in the ar-

ticle title, and n(q, d) combines Lucene-internal nor-

malization factors. The outcome providing the bet-

ter result determines the ﬁnal composition of the con-

cept term. By this, Wikipedia is acting as a concept

identiﬁer. As a result of the conceptualization step,

a set of concept terms providing the basis for the au-

tomatic detection of potential dialog topics is deter-

mined. Thus, for the utterance “Ah, then you are a

fan of Bayern Munich?” the concept terms “fan” and

“Bayern Munich” are speciﬁed.

One concept term can be related to more than

one topic although in various extents. Within our ap-

proach, the automatic assignment of concepts to top-

ics is implemented by mapping all concept terms to a

set of predeﬁned Wikipedia categories. Therefore, a

number of categories from Wikipedia best presenting

a set of topics possibly addressed in the given dia-

log scenario has to be speciﬁed previously. Basically,

every category contained in the Wikipedia category

system can be considered to present a potential dialog

topic. But it is advisable to choose those categories

having a high degree of abstraction as best reﬂecting

more general topic areas such as “Sports” or “Poli-

tics”.

Subsequently, for every chosen category all sub-

ordinated Wikipedia articles are extracted, that is, all

articles assigned to the considered category or to at

least one of its subcategories. Afterwards, the rele-

vant information parts are stored in a second Lucene

index. More precisely, documents for every prede-

ﬁned Wikipedia category including ﬁeld speciﬁca-

tions about its title as well as information about the

titles and textual contents of their subordinated arti-

cles are set up. Thereby, articles that are related to

one predeﬁned category several times are contained

accordingly often in the category document to boot

its importance within the presented topic area.

To retrieve a list of categories representing pos-

sible topics sorted in descending order according to

their relatedness to the concept term cterm we search

the index for each category document d matching

cterm in a query q on the basis of the scoring formula

presented in equation (2). As a result, each concept

term of the present utterance is represented as a vec-

tor within a space of predeﬁned Wikipedia categories

constituting potential dialog topics. For the rest of

the paper, we refer to these vectors capturing the rela-

tive importance of the dialog topics for the considered

concept term as concept topic vectors.

3.2.2 Identiﬁcation of Dialog Topics

As stated before, a dialog topic is established con-

sensually from both conversation participants. That

is, a single utterance does not have topics in isolation

but rather provide topic suggestions (Schank, 1977).

Based on this idea we have to consider at least two

successive utterances to deﬁne a topical intersection.

Accordingly, the topic tracking process begins with

the second dialog contribution.

To detect topical overlaps between two successive

dialog contributions, we compare each of the concept

topic vectors speciﬁed for one utterance with each

of the concept topic vectors of the subsequent utter-

ance separately using the cosine similarity. That is,

we quantify the similarity between two concept terms

cterm

and cterm

of successive utterances utt

and

utt

on the basis of their concept topic vector repre-

sentations

−→

V (cterm

) and

−→

V (cterm

) via

LET'S TALK TOPICALLY WITH ARTIFICIAL AGENTS! - Providing Agents with Humanlike Topic Awareness in

Everyday Dialog Situations

sim(cterm

, cterm

) =

−→

V (cterm

)·

−→

V (cterm

)

−→

V (cterm

)||

−→

V (cterm

(3)

where cterm

∈ utt

and cterm

∈ utt

If the comparing process detects a signiﬁcantly

high similarity between two concept topic vectors,

that is, their similarity is higher than a given similar-

ity threshold, a topical overlap between utt

and utt

is identiﬁed. For every topical overlap, the involved

concept topic vectors are summed up resulting in a

new vector, called dialog topic vector. The several

components in this vector provide probabilities for

each predeﬁned Wikipedia category possessing a rela-

tion to the considered concept terms. If a probability

again exceeds a given probability threshold, its corre-

sponding category constitutes the current topic of the

ongoing dialog. In case the described conditions are

fulﬁlled several times within one topic tracking pro-

cess, the system is not able to determine one single

Wikipedia category to be the current dialog topic but

rather keeps all topic options open. Figure 2 graph-

ically presents possible results of the topic tracking

process for our example dialog introduced in 2.1 by

means of a bar diagram. As reaching the threshold

represented by the horizontal line in black, the cate-

gories “Regions” and “Sports” constitute the dialog

topics within this illustration.

Utterances which do not provide any concept in-

formation, like the utterance “Okay”, have no impact

on the probabilities for the several dialog topics.

3.2.3 Topic Shift Detection

As mentioned before, we distinguish between a topic

leap as described by Svennevig (1999) and a topic

drift as introduced by Hobbs (1990). Based on this,

systems are capable of detecting radical topic shifts

enabling the particular conversational agents to gen-

erate an appropriate conversation behavior. Accord-

ing to this, the agent might refer to this topic leap via

a suitable utterance such as “What made you think of

this topic?”.

To distinguish between the two types of topic shift

automatically, the transition from one dialog topic to

the next is evaluated based on the outcomes of the

topic tracking process. That is, if no topical over-

lap between the utterances utt

and utt

can be de-

termined, the system detects a topic leap. In contrast,

a topic drift is characterized in that topical overlaps to

both the old and the new dialog topic exist during the

topic transition as shown in Figure 2.

(a) Results for utterances 1 and 2.

(b) Results for utterances 2 and 3.

Figure 2: Bar diagrams presenting results of the topic track-

ing process for our example dialog (see 2.1). For lack of

space only a subset of the bars can be presented. Actually,

all active topics in terms of bars are displayed.

3.2.4 Topic Labeling

To be able to refer to a dialog topic later on, for exam-

ple in another dialog, a descriptive topic label has to

be deﬁned. Wikipedia provides topic labels in terms

of category titles. Thus, a topic can be labeled with

the title of the Wikipedia category that constitutes the

current dialog topic. Thereby, the labels do not have

to be mentioned during dialog before as they are al-

ready existent. However, some category titles might

need to be changed to more intuitive labels allowing

a more humanlike term for a dialog topic. The cate-

gory title “Leisure”, for instance, can be replaced by

“Hobbies” as the latter provides a more intuitive label

for a dialog topic raised in smalltalk conversations.

ICAART 2012 - International Conference on Agents and Artificial Intelligence

4 MAKING ARTIFICIAL AGENTS

MORE TOPIC AWARE

So far, we described how to detect topics in ongo-

ing dialog automatically by means of collaborative

knowledge provided by Wikipedia. However, to emu-

late humanlike topic awareness in artiﬁcial agents our

presented topic detection model needs to be embed-

ded into the agent’s underlying system architecture.

More precisely, the agent’s conversational behavior

has to be adapted by means of the gained topic in-

formation to enable coherent dialogs between human

and artiﬁcial interlocutors. In this section we propose

our approach for improving the conversational abili-

ties of the ECA Max by integrating topic information

into the agent’s existing dialog system.

4.1 The Conversational Agent Max

Max (Kopp et al., 2005) is a virtual character act-

ing as a conversational computer interface that allows

for face-to-face dialogs with humans in German lan-

guage. By means of keyboard-based, textual inputs

human users are able to communicate with the agent.

Max is capable of responding to these inputs with spo-

ken language realized by a synthesized voice. Figure

3 shows Max in his current state acting as a museum

guide where he provides information about the exhi-

bition and involves human visitors in everyday small

talks.

Figure 3: Max at the Heinz Nixdorf MuseumsForum in

Paderborn, Germany.

4.2 Max’ Existing Dialog System

The agent’s verbal communication is realized by a di-

alog system consisting of three modules successively

processing the input of the human dialog partner. In

a ﬁrst step the interpreter of the dialog system deter-

mines the meaning of the user’s input text. The result

of this analysis is then delivered to the dialog man-

ager. By accessing the dialog knowledge, the dia-

log manager chooses an according answer which is

sent to behavior planning afterwards. The behavior

planning component translates the answer into a mul-

timodal utterance for the virtual character.

Both the interpretation of natural language inputs

and the generation of an adequate response to the

user’s utterance are based on a set of rules. Thereby

the interpretation is composed of two steps: First, the

identiﬁcation of modiﬁers specifying the expression

type such as negation, agreement, or greeting. Sec-

ond, the identiﬁcation of the conversational function

reﬂecting the pragmatic and semantic meaning of the

considered utterance. These processes currently em-

ploy about 1.200 rule plans which are selected and

executed via pattern matching processes. These rules

in turn direct the choice of an adequate response.

Due to the rule-based input interpretation cover-

ing a broad spectrum of possible utterances and an

additional, Wikipedia-based question answering com-

ponent (Waltinger et al., 2011), the agent’s system

never fails in computing an appropriate reply. Hence,

Max never stays speechless even if an input cannot

be decoded in detail. Nevertheless, the system has

not yet been able to establish coherent sequences of

dialog contributions as humanlike topic awareness is

not accessible for the agent. The integration of our

online topic detection model into the ECA’s system

architecture is twofold: First, we contribute to im-

prove human-agent conversations by enabling topi-

cal dialogs between human and artiﬁcial conversation

partners. Second, the existing human-machine inter-

face provides an optimal platform for the evaluation

of our approach.

4.3 Integrating Topic Information

The complete system underlying the ECA Max is

based on a multi-agent system composed of several

interacting agents. The conversational behavior, for

example, is realized via a dialog system in terms of

an intelligent dialog agent. According to this, we

built up a topic agent implementing the presented pro-

cesses and integrated this agent into the existing sys-

tem as shown in Figure 4.

The topic agent obtains every dialog contribution,

that is the user’s inputs as well as the agent’s out-

puts, and constantly provides up-to-date information

about the current topic situation of the ongoing dia-

log. It is directly connected to the dialog agent due

to interdependencies. More precisely, the interpreter

of the dialog agent sends its interpretation results to

the topic agent which decides on the topical relevance

LET'S TALK TOPICALLY WITH ARTIFICIAL AGENTS! - Providing Agents with Humanlike Topic Awareness in

Everyday Dialog Situations

Figure 4: Integration of our topic detection model into the existing architecture of the ECA Max.

of the considered utterance on the basis of the iden-

tiﬁed modiﬁer. That is, if an utterance is speciﬁed as

a greeting or farewell, the topic agent does not con-

sider it as being topically relevant. Additionally, if

one interlocutor proposes a dialog topic directly and

the interpreter speciﬁes a rejection in response to this

suggestion, the topic agent again assigns the proposed

topic to irrelevant topics. To give an example, if Max

says “Let’s talk about music!” and his human dia-

log partners answers with “I don’t want to talk about

music!”, the topic agent does not identify “Music” to

be the dialog topic even if it was mentioned in two

successive utterances.

The topic agent in turn sends the results of its

topic detection process to the dialog manager which

has an impact on the conversational behavior of Max.

For this purpose, the rules contained in the knowl-

edge base of the dialog agent are topically arranged

to distinguish between their adequacies according to

the given dialog setting. In the following, an example

extract of the resulting rule library based on the agent

architecture JAM (Huber, 1999) is given.

/** TOPIC: REGIONS **/

Plan {

NAME: ’askFor-homeTown’

GOAL: get user’s home town

BODY: <act> Where are you from? </act>

...}

Plan {

NAME: ’tell-favoriteCountry’

GOAL: tell system’s favorite country

BODY: <act> I love Portugal. </act>

...}

/** TOPIC: SPORTS **/

Plan {

NAME: ’reply-likingSports’

GOAL: tell system’s interest in sports

BODY: <act> Yes, I like soccer. </act>

...}

/** TOPIC: POLITICS **/

Plan {

NAME: ’askFor-politicalAttitude’

GOAL: get user’s party affiliation

BODY: <act> What’s your preferred

political party? </act>

...}

The topical classiﬁcation of the rules allows their

execution based on the dialog situation. Given a

ﬁrst encounter. the dialog participants would not talk

about their political afﬁliation, for instance. Accord-

ingly, Max avoids making use of the rules dealing

with so-called unsafe topics. That is, he neither uses

such a rule pro-actively raising a topic nor reactively

to answer a user question. According to the latter, he

rather gives an evasive answer (as shown in Figure 4).

5 THE DIALOG SCENARIO

In our scenario, a human participant has a face-to-

face small talk encounter with the virtual agent Max.

Thereby, the human dialog partner expresses him or

herself via keyboard-based text inputs whereas the

artiﬁcial interlocutor answers with spoken language

based on speech synthesis. Thus, the contributions of

either side exist as textual information redundantiz-

ing additional speech recognition processes. More-

over, preprocessing steps to handle incomplete and

non-standard sentences are not required as typed in-

puts mostly consist of complete sentences containing

little abbreviations and slang expressions. However,

ICAART 2012 - International Conference on Agents and Artificial Intelligence

textual inputs preclude the perception of topic ending

indicators (such as repetitions, pauses, laughter, etc.

(Howe, 1991)). Thus, they can not be considered in

the process of topic detection although often used in

human conversation.

Enabling a coherent dialog between Max and a hu-

man user necessitates the presetting of a topical struc-

ture allowing to introduce the emulated topic aware-

ness into the dialog and to cause a corresponding con-

versational behavior on the agent’s part. Schneider

(1988) assigns a structure to a typical small talk se-

quence as follows:

1. Question.

2. Answer.

3. Reply.

4. Further turns.

Furthermore, a typical small talk topically covers

the immediated, external, and communication situa-

tion (Schneider, 1988). In their study, Endrass et al.

(2011) identiﬁed a typical distribution of these top-

ics within a dialog between Germans. Thus, Ger-

mans address less of the immediated and approxi-

mately equivalent of the external and communication

situation during small talk. According to these ﬁnd-

ings, and considering the conditions arising from the

fact that Max is situated in an university environment,

the beginning of topical small talks with the conver-

sational agent Max is structured as follows: In his

ﬁrst turn, Max asks the interlocutor for his or her sub-

ject of study as most potential dialog partners are stu-

dents. Subsequently, the agent tries to ﬁnd out the in-

terlocutor’s origin. If successful, Max is able to deter-

mine the interlocutor’s favorite football club from this

knowledge and to continue with the football topic. In

case the human does not want to talk about football

or sports in general, he or she has the opportunity to

suggest another topic. Thus, the ﬁrst dialog topics

are solely initiated by Max. This is important insofar

as this scenario also establishes a basis for the devel-

opment of a personal memory for the agent. This re-

quires the gathering of a lot of information concerning

social categories about the human interlocutor.

5.1 Planned Evaluation

Upon successfully completing a preliminary eval-

uation identifying the topics of newspaper arti-

cles, which has shown an average accuracy of 61.0

(Breuing et al., 2011), we plan to accomplish a more

adequate evaluation considering and addressing the

introduced dialog scenario. Accordingly, we searched

for a corpus comprising dialog information of Ger-

man small talks occurring during ﬁrst encounters be-

tween two persons. The CUBE-G corpora (Endrass

et al., 2011) provides analyzed records of 21 ﬁrst

interaction scenarios each between a student and a

professional actor and each lasting around ﬁve min-

utes. Amongst others, the dialogs were tested for the

amount of topics and topic shifts which is why the

corpus contains topical annotations for each recorded

small talk. Thus, the CUBE-G corpus presents the

optimal basis for our following evaluation.

In preparation for the planned evaluation, we al-

ready determined a list of predeﬁned main categories

that represent typical dialog topics for everyday small

talks. Thereby, we omitted so-called unsafe top-

ics (see Section 1) and especially focused on topics

raised in the given university scenario. Table 1 shows

the resulting list of main categories. Moreover, we

downloaded the German database dump from May

14, 2011 and built up a Lucene index containing all

information parts relevant for our purpose.

Table 1: List of predeﬁned main categories adequate for our

dialog scenario.

Main Category

Science Economics

Family Education

Studies Literature

Mass media Music

Arts Health

Ecology Digital media

Sports Occupations

Fashion Food and drink

Leisure Transport

Intimate relationships Regions

The next step is the preprocessing of the corpus in

that incomplete sentences and expressions are com-

pleted to adapt the recorded utterances to the condi-

tions given by the fact that human-sided utterances

are based on keyboard inputs. Then, we will accom-

plish the evaluation by automatically identifying the

dialog topics and topic shifts within the CUBE-G in-

teractions by means of our proposed method to sub-

sequently compare the results with the manual anno-

tations included in the corpus. If showing promising

performance, a user study evaluating the application

of emulated human topic awareness in the agent Max’

conversational behavior will be scheduled next.

6 RELATED WORK

A lot of work has been carried out on ofﬂine topic

identiﬁcation. A prevalent model was developed

LET'S TALK TOPICALLY WITH ARTIFICIAL AGENTS! - Providing Agents with Humanlike Topic Awareness in

Everyday Dialog Situations

in the context of the Topic Detection and Tracking

(TDT) research program (Allan, 2002). Within the

TDT research, Allan determined ﬁve tasks (i.e., Story

Segmentation, First Story Detection, Cluster Detec-

tion, Tracking, and Story Link Detection) for detect-

ing the several topics outlined in a text-based news-

cast. Further ofﬂine approaches compute the co-

herence between documents via similarity measures

(e.g., (Makkonen et al., 2004; Zhang and Wang,

2010)). Others rank Wikipedia articles according

to their relevance to a given text fragment, for ex-

ample via text classiﬁcation algorithms (Gabrilovich

and Markovitch, 2007) or by simply exploiting the

Wikipedia article titles and categories (Sch

onhofen,

2006). One recent approach uses the Wikipedia cate-

gory network as a conceptual taxonomy and derives a

directed acyclic graph for each document by mapping

terms to a concept in the category network (Chahine

et al., 2011).

Approaches for the online identiﬁcation of topics

in natural language dialogs are rare. One work re-

alizing a “Dynamic Topic Tracking” of natural lan-

guage conversations between a human and a robot

roughly adapted the ﬁve tasks from the TDT project

(see above) to make the robot more situation aware in

human-robot interaction (Maas et al., 2006). Thereby

the amount of topics and the according topic names

are created dynamically by gathering the topic names

from content words most occurring in the dialog utter-

ances. On the contrary, existing taxonomies can serve

as a source for topic labels, for example derived from

the online encyclopedia Wikipedia (Breuing et al.,

2011; Waltinger et al., 2011). Furthermore, con-

versation clusters visually highlight topics discussed

in conversations using Explicit Semantic Analysis

based on Wikipedia articles (Bergstrom and Kara-

halios, 2009).

7 CONCLUSIONS AND FUTURE

WORK

We presented an approach for the automatic emula-

tion of humanlike topic awareness in ongoing small

talk dialogs to extend the conversational abilities of a

virtual agent in human-agent interactions. More pre-

cisely, we proposed solutions for both tasks the auto-

matic identiﬁcation of dialog topics and the integra-

tion of the resulting topic information into the agent’s

existing system architecture. The several associated

processes fulﬁll the requirements given by a face-to-

face encounter between a human and a conversational

agent and enable both a coherent and socially ade-

quate dialog between the human and the artiﬁcial in-

terlocutors. Thereby, we exploit Wikipedia knowl-

edge and hence the beneﬁts originated from collab-

orative work (namely the existence of information

whose maintenance and expansion is carried out by

numerous volunteers and the reﬂection of the partici-

pants’ common perception of conceptual structures).

In future, we will extend our approach by detect-

ing and linking topical afﬁliations to previous dialog

topics to handle short side trips to past topics. More-

over, we will resolve ambiguities by taking into ac-

count the current dialog topic to inﬂuence the concept

detection process.

ACKNOWLEDGEMENTS

This work is kindly supported by the Deutsche

Forschungsgemeinschaft (DFG) in the context of the

KnowCIT research project in the Center of Excel-

lence Cognitive Interaction Technology (CITEC) at

Bielefeld University. We thank Birgit Endrass and

Elisabeth Andr

e from the University of Augsburg for

providing parts of their CUBE-G corpus.

REFERENCES

Allan, J. (2002). Topic Detection and Tracking: Event-

based Information Organization. Kluwer Academic

Publishers.

Bergstrom, T. and Karahalios, K. (2009). Conversa-

tion Clusters: Grouping Conversation Topics through

Human-Computer Dialog. In Proceedings of the Inter-

national Conference on Human Factors in Computing

Systems (CHI09).

Breuing, A., Waltinger, U., and Wachsmuth, I. (2011).

Harvesting Wikipedia Knowledge to Identify Topics

in Ongoing Natural Language Dialogs. In Proceed-

ings of the 2011 IEEE/WIC/ACM International Joint

Conference on Web Intelligence and Intelligent Agent

Technology, pages 445–450.

Bublitz, W. (1989). Topical Coherence in Spoken Dis-

course. Studia Anglica Posnaniensia, 22:31–51.

Cassell, J., Bickmore, T., Campbell, L., Vilhj

almsson, H.,

and Yan, H. (2000). Human Conversation as a Sys-

tem Framework: Designing Embodied Conversational

Agents. In Cassell, J., Sullivan, J., and Churchill, E.,

editors, Embodied Conversational Agents, pages 29–

63. MIT Press.

Chahine, C. A., Chaignaud, N., Kotowicz, J.-P., and

ecuchet, J.-P. (2011). Conceptual Indexing of Doc-

uments Using Wikipedia. In Proceedings of the 2011

IEEE/WIC/ACM International Joint Conference on

Web Intelligence and Intelligent Agent Technology,

pages 195–202.

Clark, H. H. (1996). Using Language. Cambridge Univ.

Press.

ICAART 2012 - International Conference on Agents and Artificial Intelligence

Endrass, B., Rehm, M., and Andr

e, E. (2011). Planning

Small Talk Behavior with Cultural Inﬂuences for Mul-

tiagent Systems. Computer Speech and Language,

25(2):158–174.

Gabrilovich, E. and Markovitch, S. (2007). Computing Se-

mantic Relatedness Using Wikipedia-based Explicit

Semantic Analysis. In Proceedings of the Inter-

national Joint Conference on Artiﬁcial Intelligence

(IJCAI-07).

Hobbs, J. (1990). Topic Drift. In Dorval, B., editor, Con-

versational Organization and Its Development, pages

3–22. Ablex Publishing.

Howe, M. (1991). Collaboration on Topic Change in Con-

versation. Kansas Working Papers in Linguistics,

16:1–14.

Huber, M. J. (1999). JAM: A BDI-Theoretic Mobile Agent

Architecture. In Proceedings of the Third Interna-

tional Conference on Autonomous Agents (Agents99),

pages 236–243.

Jurafsky, D. and Martin, J. H. (2009). Speech and Language

Processing. Pearson Prentice Hall.

Kopp, S., Gesellensetter, L., Kr

amer, N., and Wachsmuth,

I. (2005). A Conversational Agent as Museum Guide

– Design and Evaluation of a Real-World Applica-

tion. In Proceedings of Intelligent Virtual Agents (IVA

2005).

Maas, J. F., Spexard, T., Fritsch, J., Wrede, B., and Sagerer,

G. (2006). BIRON, what’s the Topic? A Multi-Modal

Topic Tracker for Improved Human-Robot Interac-

tion. In Proceedings of the IEEE International Work-

shop on Robot and Human Interactive Communica-

tion (ROMAN), pages 26–32.

Makkonen, J., Ahonen-Myka, H., and Salmenkivi, M.

(2004). Simple Semantics in Topic Detection and

Tracking. Information Retrieval, 7(3-4):347–368.

McCandless, M., Hatcher, E., and Gospodneti

c, O. (2010).

Lucene in Action. Manning, 2 edition.

Schank, R. C. (1977). Rules and Topics in Conversation.

Cognitive Science, 1(4):421–441.

Schneider, K. P. (1988). Small Talk: Analysing Phatic Dis-

ourse. Hitzeroth.

Sch

onhofen, P. (2006). Identifying Document Topics Us-

ing the Wikipedia Category Network. In Proceedings

of the 2006 IEEE/WIC/ACM International Conference

on Web Intelligence (WI’06).

Svennevig, J. (1999). Getting Acquainted in Conversation.

John Benjamins Publishing.

Toutanova, K. and Manning, C. D. (2000). Enriching

the Knowledge Sources Used in a Maximum Entropy

Part-of-Speech Tagger. In Proceedings of the Joint

SIGDAT Conference on Empirical Methods in Natural

Language Processing and Very Large Corpora, pages

63–70.

Waltinger, U., Breuing, A., and Wachsmuth, I. (2011). In-

terfacing Virtual Agents With Collaborative Knowl-

edge: Open Domain Question Answering Using

Wikipedia-based Topic Models. In Proceedings of

the International Joint Conference on Artiﬁcial Intel-

ligence (IJCAI 2011), pages 1896–1902.

Zhang, X. and Wang, T. (2010). Topic Tracking with

Dynamic Topic Model and Topic-based Weighting

Method. Journal of Software, 5(5):482–489.

LET'S TALK TOPICALLY WITH ARTIFICIAL AGENTS! - Providing Agents with Humanlike Topic Awareness in

Everyday Dialog Situations