Towards Computational Models for a Long-term Interaction with

an Artiﬁcial Conversational Companion

Sviatlana Danilava

, Stephan Busemann

, Christoph Schommer

and Gudrun Ziegler

University of Luxembourg, 6 Rue Coudenhove-Kalergi, Luxembourg, Luxembourg

Deutsches Forschungszentrum f¨ur K¨unstliche Intelligenz (DFKI) GmbH, Stuhlsatzenhausweg 3, Saarbr¨ucken, Germany

Keywords:

Artiﬁcial Companions, Models of Interaction, Human-Computer-Interaction, Instant Messenger.

Abstract:

In this paper we describe a design approach for an Artiﬁcial Conversational Companion according to ear-

lier identiﬁed requirements of utility, adaptivity, conversational capabilities and long-term interaction. The

Companion is aimed to help advanced learners of a foreign language to practice conversation via instant mes-

senger dialogues. In order to model a meaningful long-term interaction with an Artiﬁcial Conversational

Companion for this application case, it is necessary to understand how natural long-term interaction via chat

between human language experts and language learners works. For this purpose, we created a corpus from

instant messenger-based interactions between native speakers of German and advanced learners of German as

a foreign language. We used methods from conversation analysis to identify rules of interaction. Examples

from our data set are used to illustrate how particular requirements for the agent can be fulﬁlled. Finally, we

outline how the identiﬁed patterns of interaction can be used for the design of an Artiﬁcial Conversational

Companion.

1 INTRODUCTION

The vision of artiﬁcial agents that “listen to spoken

sentences” and are “nice personalities” and our “lit-

tle electronic friends” is not new, see for example

(Winograd and Flores, 1987). In 2005, Y. Wilks in-

troduced the notion of an Artiﬁcial Companion as “an

intelligent and helpful cognitive agent which appears

to know its owner [...], chats to them [...], assists them

with simple tasks” (Wilks, 2005). The agent must

be able to maintain a sustained discourse over a long

time period, serve interests of the user and have a lot

of personal knowledge about the user (Wilks, 2010).

The idea to use conversational agents (chatbots) in

second language language (L

) acquisition is long of

interest (Kerly et al., 2007; Zakos and Capper, 2008).

Language acquisition “requires meaningful interac-

tion in the target language [...] in which speakers

are concerned not with the form of their utterances

but with the messages they are conveying and un-

derstanding” (Krashen, 1981). Predictability of re-

sponses, lack of personality and inability to remember

the interaction history are reported as shortcomings

of current agents for conversation training (Shawar

and Atwell, 2007). Agent designers battle against

these issues by creating more sophisticated patterns

for domain-unrestrictedlanguage understanding, stor-

ing information about the user and already used re-

sponses (Jia, 2009). However, the design of the agents

is still focused on the content of responses, but not on

language as a co-constructed meaningful action.

We consider the application scenario where ad-

vanced learners of a foreign language practice conver-

sation in dialogues with an Artiﬁcial Conversational

Companion (ACC). In earlier work, we identiﬁed the

minimum requirements that an artiﬁcial agent must

satisfy in order to be mentioned as an ACC (Danilava

et al., 2012). We reﬁned the requirements for the

application scenario of conversation training in L

acquisition. We focused on interaction via instant

messenger (IM) because it combines the advantages

of spoken and written communication being concep-

tually oral and medially written (Koch and Oesterre-

icher, 1985).

In this paper, we describe our ACC design ap-

proach based on empirical data from IM dialogues

according to the earlier identiﬁed requirements of

conversation (natural language understanding and

generation (NLU/NLG), cognitive abilities, emo-

tional competence, socio-cultural competence), util-

ity, adaptivity and long-term interaction. We see

a long-term interaction as a process of construction

241

Danilava S., Busemann S., Schommer C. and Ziegler G..

Towards Computational Models for a Long-term Interaction with an Artiﬁcial Conversational Companion.

DOI: 10.5220/0004255402410247

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 241-247

ISBN: 978-989-8565-38-9

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

of longer dialogues over a stretched period of time,

where the length of each dialogue and the number of

dialogues do not have any predeﬁned minimum value.

It remains open, when the interaction ends.

We extract patterns from dialogues between hu-

mans that will help make an interaction with an ACC

close to a natural interaction that is co-constructed by

all the participants as a meaningful activity according

to rules of social interaction and by means of the se-

lected communication medium.

We explain the methodology and brieﬂy describe

the data set in Section 2. We illustrate our design ap-

proach on examples from the dataset in Section 3 fol-

lowed by conclusions in Section 4.

2 METHOD

In order to model a long-terminteraction with an ACC

via IM dialogue, it is necessary to understand how

natural long-term IM-based interaction between hu-

man language experts and language learners works.

We created and used data from natural interactions for

this type of analysis. Language experts provided in-

teraction patterns for the future ACC, and language

learners offered information for user modelling.

2.1 Data

IM chat is subject to intensive research in con-

versation analysis (Orthmann, 2004; Nardi et al.,

2000), computer-mediated collaborative work (Jiang

and Singley, 2009) and natural language process-

ing (Forsythand and Martell, 2007). The data sets

used come from natural workspace interaction (Avra-

hami and Hudson, 2006) or interaction experiments

(Solomon et al., 2010). These data sets are not avail-

able as a resource for the research community. Cor-

pora from multi-user open chatrooms are available

for research, see for example (L¨udeling, 2009; Lin,

2012). They do, however,not satisfy the requirements

of dialogic interaction between language experts and

language learners over a long period of time. For

this reason, we created our own corpus of IM based

expert-learner dialogues for German as a focus lan-

guage. Space limitations preventus to offer more than

a concise description of the data set.

Voluntary participants – 4 German native speakers

and 9 advanced learners of German as L

– commu-

nicated within 4-8 weeks via IM. Each chat session

Visit http://wiki.uni.lu/mine/Sviatlana+Danilava.html

for a detailed description of the data collection.

took between 20 and 90 minutes. The parties com-

municated with the same partner for the complete du-

ration of the experiment. The participants produced a

total of 72 dialogues, which correspond to ca. 2.500

minutes of IM interaction, ca. 4.800 messages of 10

tokens average length, ca. 52.000 tokens in total, and

ca. 6.100 unique tokens.

Since the data set is quite small, quantitativemeth-

ods are not applicable to achieve statistically signiﬁ-

cant results. However, qualitative methods from eth-

nomethodology and conversation analysis can deliver

reliable results in understanding rules of interaction

based on small-scale data sets. Examples from the

data set taken from dialogues with different partici-

pants pairs illustrate the most importanttypes of rules.

We use the notation N for experts and L for learners.

2.2 Modeling Approach

Top-down ACC design approaches according to ab-

stract requirements suffer in general from the impos-

sibility to foresee all potential ways for the interac-

tion ﬂow. In contrast, a bottom-up data driven design

could lead to a huge number of abstract classes with

unclear relations between them and make the system

design unmanageable. We combine the top-down re-

quirements for ACCs with the bottom-up approach

commonly used in conversational analysis.

Conversational analysis usually is performed in

three steps (cf. e.g. (Orthmann, 2004)):

1. Looking through the data without preconceptions

about what may be found.

2. Pattern synthesis: more abstract, generalised de-

scription of structures found.

3. General abstract description of interaction rules.

In particular, we focused on rules in interaction,

where the participants make the meaningful activity

and the social interaction (social closeness or dis-

tance) explicit. There are dialogue patterns speciﬁc

for learners (e.g. different types of errors) and ex-

perts (e.g. error correction), and in patterns that dis-

close disruptive factors in interaction as, for example,

too long response time. Finally, we outline how the

detected patterns serve as a basis for computational

models of long-term interaction in general and for de-

sign of an ACC for conversation training in particular.

3 DATA ANALYSIS

3.1 Patterns for Utility

As reported by the volunteers, the motivation to par-

ticipate in the experiment was their willingness to

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

242

help the organiser, willingness to improve their lan-

guage skills and conversational competence, practice

conversation in German, willingness to do something

new, to get in contact with people from other coun-

tries and not to be passive. Furthermore, we observed

in the data two actions focused on language learning:

error correction and explanation of new lexical ma-

terial. These activities provide valuable patterns for

both, utility and NLU/NLG.

3.1.1 Awareness of the Meaningfulness

Helping the organiser is documented by participants

especially in closing sequences, where they “did their

job”. Example 1 shows participants’ awareness in a

particular interaction sequence: N says “ok, we have

produced a lot of text” expressing his personal inter-

pretation of the meaning of this interaction. Similar

sequences also occur in dialogues produced by other

pairs.

Example 1: Talking about the meaningful activity.

Time Sndr Message Body

18:44:44 N ok, wir haben jetzt eine große Menge Text

produziert. Es war sehr sch¨on mit Dir zu

plaudern. F¨allt Dir noch was ein?

Example 2 illustrates how participants demonstrate

that they “did their job” for the whole experiment (We

explain the use of parentheses in Sec.3.4).

Example 2: Talking about activity completion.

Time Sndr Message Body

21:56:18 N Ich habe ¨ubrigens vorgestern mit unserer Or-

ganisatorin gesprochen. Wir haben unserer

Gespr¨achsanzahl heute erf¨ullt :-)

21:57:28 L Ich hoffe, dass unser Chat f¨ur sie nutzbar

ist)))

After each pair of participants completed 8 dialogue

sessions, they could choose to keep communicating

or to abandon the chat. A task like conversational

training can also be ﬁnished after a certain number

of sessions. A possible scenario for an ACC is to in-

quire with the user from time to time if the interaction

should continue or come to an end.

3.1.2 Error Recognition and Error Correction

Learners often produce ungrammatical sentences. For

an ACC, even pattern-based NLU is challenging due

to incorrect use of lexical items. Therefore, error

models are important for utility and are an essential

part of NLU. There are error-tagged corpora for learn-

ers’ language. However, they are created from con-

ceptually written data (usually essays) and do not con-

tain corrections (L¨udeling et al., 2005; Boyd, 2010).

Typical errors depend among others on the level of

language proﬁciency, and on ones’s native language.

Statistical error models offer a basis for grammar and

style error recognition. However, they are usually

based on native speakers’ data and cannot deal with

wrong use of lexical items (Crysmann et al., 2008).

Automatic error recognition is problematic because of

contextual factors rendering an otherwise grammati-

cal expression invalid. In Example 3, N corrects “bin

ich frei” even though this is a grammatical sentence.

Example 3: Error correction.

Time Sndr Message Body

19:41:15 L Zur Zeit bin ich frei um Diplomarbeit zu

schreiben

19:41:58 N Falls ich das korrigieren darf: Du ”hast” frei,

nicht ”bist”. :-)

19:42:48 L Danke

19:43:16 L Ich habe frei :)

Besides explicit embedded repair sequences as in Ex-

ample 3, the data set also exhibits error corrections,

for example, in the form of indirect repairs (Exam-

ple 7: ”2 Teste” - ”2 Tests”) or a direct correction

response to a erroneous sentence. The choice of the

correction form depends on the level of social inter-

action and the error type. There is a conﬂict between

the learners’ expectation that the native speakers help

them to improve their skills by correcting errors and

the native speakers’ wish not to be boring and there-

fore not to correct too much.

3.1.3 Models for Conversational Training Task

German participants introduced idiomatic expres-

sions, which they thought the partners did not know,

and explained the meaning (Example 4).

Example 4: Introducing new lexical material.

Time Sndr Message Body

20:01:24 N Was macht die Kunst?

20:03:29 L Was bedeutet dieser Ausdr¨uck? Ich verstehe

nicht (((

20:04:09 L Meinst du Tanzen?

20:04:44 N Das habe ich mir schon gedacht :-)

[

explaining

]

German participants explained the meaning of some

words in form of jokes (Example 5).

Lexical error correction or explanation of unknown

words can be implemented with external resources for

idioms and meaning explanations, including but not

limited to Wikipedia or online dictionaries.

TowardsComputationalModelsforaLong-termInteractionwithanArtificialConversationalCompanion

243

Example 5: Introducing and explaining multiple meanings

of treffen.

Time Sndr Message Body

20:32:40 N Ich ﬁnde Witze mit doppelter Wortbedeutung

ganz lustig.

20:33:02 L Ich probiers mal, vielleicht verstehst du ihn.

20:33:32 N Im Wald treffen sich zwei J¨ager, beide tot.

20:35:33 L Ich brauch deine Hilfe. Gib mir bitte einen

Fingerzeig!

20:37:33 N Okay :-)

[

explaining

]

3.2 Patterns for NLU/NLG

Conversation is co-constructed by all participants. It

contains at least a content part (e.g. topic) and a man-

agement part where the participants display that they

are talking and intend to keep doing so. Responsive-

ness is an observable phenomenon for the interaction

management. A huge amount of research is focused

on the the content part. To our knowledge, respon-

siveness has never been taken into account for chat-

bot design. IM responsiveness is inﬂuenced by un-

controlled elements like parallel activities of the par-

ticipants, typing speed, network delays, experience in

chat interaction. Time stamp and the length of the re-

ceived message and the produced response are, how-

ever, analysable for both, researchers and participants.

Interaction parties share the understanding of what

is acceptable. In Example 6, L replied to N’s question.

More than 5 minutes later L posts a request whether N

still has time for chat, displaying that the time interval

is too long. N replies with an excuse offer, displaying

the awareness of the acceptable time interval.

Example 6: The Length of the allowed time interval.

Time Sndr Message Body

17:46:42 N wie bist Du zu diesem Chat-Projekt gekom-

men?

17:47:14 L meine Lektorerin hat mir gesagt.

[

explaining

]

17:52:33 L [

N firstname

], wenn du schon keine Zeit

hast, dann schreibe dann bitte) sonst kann ich

sehr lange schreiben)))

17:53:08 N oh, Verzeihung! bin wieder da, sorry!

17:54:40 L ok)

Models for presence requests and back-channeling

can be described according to responsiveness crite-

ria. The initial value of acceptable responsiveness can

be conﬁgured according to average values from in-

teractions between native speakers, and then adapted

according to the learner’s behavior. Responsiveness-

based rules can be deﬁned for repairs, self-repairs,

repetitions and interaction management at unit (topic

and action) boundaries.

3.3 Required Cognitive Abilities

The participants did not know anything about their

partners prior to the experiment. Initially, they asked

their partners about their names, age, locations and

occupation. This knowledge might be required for

initialising appropriate social closeness/distance in in-

teraction (see also Sec. 3.5). Some facts remain im-

portant for a longer period of time (e.g. names, or lo-

cations), some can be put aside after a particular time

(e.g. examinations), or maybe forgotten.

Once some facts are learned, it is appropriate to

ask about the state of the fact later. For example, when

some of the participants told that they have exams,

their partners asked about the results later (Example

7), which also would be an appropriate reaction for

the ACC.

Example 7: Simple inferences: talking about signiﬁcant

events (23. May and 31. May).

Time Sndr Message Body

18:29:10 L leider auch nicht((( morgen schreibe ich 2

Teste in Deutsch und Englisch. und wie du

verstehst, habe ich noch nicht sie gelernt=))

18:30:02 N 2 Tests? ok, klar, dann erstmal viel Erfolg

dabei!

18:29:23 N hey, wie waren Deine Pr¨ufungen?

18:35:21 L [...] alle Tests wurden schon SEHR gut

geschrieben. [...]

Similar patterns exist for football games (the exper-

iment took place during a European soccer cham-

pionship) and can be generalised for all signiﬁcant

events. This can be implemented as an inference rule

that takes an important event, its date and information

about expected results into account. The ACC may

ask about results later.

3.4 Patterns for Emotional Competence

Human participants use paralinguistic and prosodic

cues to show emotions in IM messages. They are ex-

pressed through capitalisation, spelling, punctuation

or timing (Orthmann, 2004). The most analysed re-

source for displaying emotions in chat are emoticons

(e.g. “:-)” ) and emotive language (e.g. “haha”, “hi-

hihi”). There are also reactive tokens (e.g. news mark-

ers “oh, wirklich?”, “oh, nein!”) expressing emotions.

Emotionalcompetence of the agentconcerns emo-

tion recognition, interpretation and response genera-

tion. There are the same problems in correct emotion

recognition and interpretation as for lexical items. In

a multi-cultural dialogue, culture-speciﬁc items are

used in addition. In Example 7, “(((” and “))” rep-

resent “sadness” and “joy”, respectively. These sym-

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

244

bols are ambiguous to opening and closing parenthe-

ses. This use of parentheses to display joy or sadness

is typical for the learners, but was never used by the

experts.

In Example 6, L makes a presence request con-

taining “)” and “)))”, which we interpret as an indica-

tor of politeness. After the response of her partner N,

L accepts his excuse by “ok)”. In both cases, smiles

cannot be interpreted as an indicator of joy.

Patterns for the use of reactive tokens, emoticons

and emotive language in different actions – for exam-

ple, corrections, presence requests, making appoint-

ments, initiating ending sequences, and marking unit

boundaries – can be extracted from the data set.

3.5 Understanding Social Interaction

The participants use the form of address, the greeting

forms, emotions, politeness, variations in the lexicon

size or syntax to display social and emotional close-

ness or distance (Koch, 1994). The meaning of these

resources can be interpreted differently by represen-

tatives of different cultures. However, there must be

a set of universal rules for social interaction, as oth-

erwise people would not be able to manage any inter-

cultural communication without prior training.

We illustrate how the participants use the form of

address: “Sie” (formal, third person plural) vs. “du”

(familiar, second person singular) to ﬁnd out their de-

gree of social closeness in dialogues. We found pat-

terns for explicit or implicit negotiation, whereby the

implicit negotiation can take place within one interac-

tion or stretch over multiple dialogues.

Example 8 shows an explicit negotiation: L asks,

if it is good to write using “du”, N answers that “it

would be strange to chat using “Sie””.

Example 8: Explicit Negotiating Social Closeness.

Time Sndr Message Body

17:18:08 L Und ist es gut, wenn ich ”auf Du” schreibe?

17:19:13 N oh, hab gar nicht gefragt... nat¨urlich, auf

”Sie” chatten w¨are irgendwie seltsam :-)

17:19:48 L ok=)

Example 9 describes an implicit negotiation: L starts

with “Sie” and changes the form of address to “du” in

her second turn. Her second turn repeats the question

from her ﬁrst turn (“I don’t know your [=Sie] name.”)

but reformulates it using “du” (“What’s your [=du]

name?”).

To determine the appropriate degree of social close-

ness, the participants searched for similarities be-

tween them and their partners (age, location, occupa-

tion). At the current stage of research, it is not clear,

which level of social interaction is appropriate for

Example 9: Implicit Negotiating Social Closeness.

Time Sndr Message Body

19:57:31 L Hallo! Entschuldigung, Ich weiß nicht, wie

heißen Sie. [...]

mit freundlichem Gruß, L

firstname,

lastname, [...]!

19:59:57 N Hallo

L firstname

, das ist ¨uberhaupt kein

Problem! Ich hoffe, alle Probleme sind

gel¨ost und wir k¨onnen ein bisschen chatten.

20:01:58 L Ja, nat¨urlich! wie heißt du?

human-machine dialogues. It cannot be taken for

granted that learners would use “Sie” in the interac-

tion with a machine, but the machine could start with

the polite form. In addition to the forms of address,

use of reactive tokens, humour and sharing private

information are examples of interaction phenomena,

where social interaction is analysable.

3.6 Adaptivity Mechanisms

Similar to the book suggestion system described in

(Rich, 1979), the ACC must be able to build indi-

vidual user models from a very low amount of per-

sonal knowledge in the ﬁrst conversation. As we il-

lustrated in Section 3.3, information about age, gen-

der, location and occupation of the participants was

sufﬁcient to initialise the level of social interaction for

humans. The amount of knowledge about the partners

increased over time. This knowledge can be used by

the system in highly adaptive user models.

Adaptivity and anticipation may affect the single

sessions (adaptivity within one interaction, for exam-

ple topic sequences, responsiveness and user’s mood)

and the complete interaction history (for example pre-

ferred topics, changes in user’s lexicon, error track-

ing, topic selection according to user’s interests). The

adaptation of a Companion’s language (lexicon, style)

to the user’s language is not a general goal, because

we cannot take it for granted that the learners use the

foreign language correctly.

There are mutual dependencies between emo-

tional, social and cultural aspects of the ACC design,

which need to be explicitly modelled to trigger the

adaptivity mechanisms correctly. For example, re-

sources used for social interaction need to be selected

according to social closeness. Emotions are embed-

ded into the context of activity and are also adapted to

the social interaction according to social closeness or

distance. Telling jokes (Example 4), using colloquial

expressions or showing emotions when something is

not clear (Example 3) may be misplaced in an inter-

action with a large social distance.

TowardsComputationalModelsforaLong-termInteractionwithanArtificialConversationalCompanion

245

3.7 Chances for Long-term Interaction

A long-term interaction with an ACC cannot be en-

forced. Potential users of an ACC – the learners – re-

ported that they are curious to chat with the system for

the ﬁrst time, but they will not keep interacting with

the system if it does not make sense for them. The

goal of the ACC designers is to create necessary con-

ditions to make long-term interaction possible. How-

ever, it cannot be a system requirement to achieve a

dialogue of a particular duration in multiple sessions.

As Examples 1 and 2 show, the interaction parties

are aware of them engaging in a joint activity mean-

ingful for both of them, that this activity will span

multiple sessions, and that after a particular number

of sessions of hopefully pleasant interaction, the ac-

tivity is completed, or its meaning has changed and

they keep interacting. This was not avoidable for the

data collection with volunteers. The interaction with

an ACC does not necessarily have to end after a par-

ticular number of conversations.

We cannot take it for granted that the learners will

accept the ACC as a language expert. The users will

try to ﬁnd out what the ACC does not understand. The

ACC designers need to anticipate as much as possi-

ble of such scenarios in order to let the machine look

smart or funny, but still polite according to the level

of social closeness.

4 CONCLUSIONS

We collected data from human-human IM dialogues

that revealed valuable patterns for social interaction

and activity in the context of conversational training.

In contrast to most existing Companion prototypes,

we use empirical data for the design of an ACC. We

described how we use these empirical data in order to

satisfy the requirements of conversation, utility, adap-

tivity and long-term interaction.

To effectively use the patterns for the ACC design,

a consistent, holistic rule framework will account for

the interdependencies in recognising patterns in real-

time during the interaction and in producing appropri-

ate system (re-)actions in terms of utterance content

and interaction management.

ACKNOWLEDGEMENTS

We would like to thank all the participants of the data

collection experiment for their voluntary work.

REFERENCES

Avrahami, D. and Hudson, S. (2006). Responsiveness in in-

stant messaging: Predictive models supporting inter-

personal communication. In HCI, pages 731–740.

Boyd, A. (2010). EAGLE: an error-annotated corpus of be-

ginning learner german. In Proc. of LREC. ELRA.

Crysmann, B., Bertomeu, N., Adolphs, P., Flickinger, D.,

and Kl¨uwer, T. (2008). Hybrid processing for gram-

mar and style checking. In Proc. of the 22nd Int. Conf.

on Comp. Linguistics, pages 153–160. Coling.

Danilava, S., Busemann, S., and Schommer, C. (2012).

Artiﬁcial Conversational Companion: a requirement

analysis. In Proc. of ICAART, pages 282–289.

Forsythand, E. N. and Martell, C. H. (2007). Lexical and

discourse analysis of online chat dialog. In Proc. of

the International Conference on Semantic Computing,

ICSC ’07, pages 19–26. IEEE Computer Society.

Jia, J. (2009). CSIEC: A computer assisted english learn-

ing chatbot based on textual knowledge and reasoning.

Know.-Based Syst., 22(4):249–255.

Jiang, H. and Singley, K. (2009). Exploring bilingual, task-

oriented, document-centric chat. In Proceedings of

GROUP ’09, pages 229–232.

Kerly, A., Hall, P., and Bull, S. (2007). Bringing chat-

bots into education: Towards natural language nego-

tiation of open learner models. Know.-Based Syst.,

20(2):177–185.

Koch, P. (1994). Schriftlichkeit und sprache. In Schrift

und Schriftlichkeit. Ein interdisziplin¨ares Handbuch

internationaler Forschung, pages 587–604. Walter de

Gruyter.

Koch, P. and Oesterreicher, W. (1985). Sprache der

N¨ahe – Sprache der Distanz. M¨undlichkeit und

Schriftlichkeit im Spannungsfeld von Sprachtheorie

und Sprachgeschichte. In Romanistisches Jahrbuch,

volume 36, pages 15–43. Walter de Gruyter.

Krashen, S. (1981). Second Language Acquisition and Sec-

ond Language Learning. Oxford: Pergamon.

Lin, J. (2012). Automatic Author Proﬁling of Online Chat

Logs. Kindle Edition.

L¨udeling, A. (2009). Corpus Linguistics: An International

Handbook. Mouton de Gruyter.

L¨udeling, A., Walter, M., Kroymann, E., and Adolphs, P.

(2005). Multi-level error annotation in learner cor-

pora. In Corpus Linguistics.

Nardi, B. A., Whittaker, S., and Bradner, E. (2000). Interac-

tion and outeraction: instant messaging in action. In

Proc. of the ACM conf. on Computer supported coop-

erative work, pages 79–88.

Orthmann, C. (2004). Strukturen der Chat-Kommunikation:

konversationsanalytische Untersuchung eines Kinder-

und Jugendchats. PhD thesis, Freie Universit¨at Berlin.

Rich, E. (1979). User modeling via stereotypes. Cognitive

Science, 3:329–354.

Shawar, B. A. and Atwell, E. (2007). Chatbots: Are they

really useful? LDV-Forum, 22(1):29 – 49.

Solomon, J., Newman, M., and Teasley, S. (2010). Speak-

ing through text: the inﬂuence of real-time text on dis-

course and usability in im. In Proc. of the 16th ACM

int.l conf. on Supporting group work, pages 197–200.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

246

Wilks, Y. (2005). Artiﬁcial companions. Interdisciplinary

Science Reviews, 30(2):145 – 152.

Wilks, Y. (2010). Is a companion a distinctive kind of rela-

tionship with a machine? In Proc. of the 2010 Work-

shop on CDS, pages 13–18. ACL.

Winograd, T. and Flores, F. (1987). Understanding Com-

puters and Cognition: A New Foundation for Design.

Addison-Wesley Longman Publishing Co., Inc.

Zakos, J. and Capper, L. (2008). CLIVE - an artiﬁcially in-

telligent chat robot for conversational language prac-

tice. In Proceedings of the 5th Hellenic conference on

Artiﬁcial Intelligence: Theories, Models and Applica-

tions, pages 437 – 442. Springer-Verlag.

TowardsComputationalModelsforaLong-termInteractionwithanArtificialConversationalCompanion

247