A CONVERSATIONAL AGENT FOR INFORMATION RETRIEVAL

BASED ON A STUDY OF HUMAN DIALOGUES

A. Loisel, G. Dubuisson Duplessis, N. Chaignaud, J-Ph. Kotowicz and A. Pauchet

LITIS - EA 4108, INSA Rouen, Avenue de l’Universit´e - BP 8, 76801, Saint-

Etienne-du-Rouvray Cedex, France

Keywords:

Conversational agent, Dialogue System, Mixed-initiative Interaction.

Abstract:

This study strives to improve medical information search in the CISMEF system by including a conversational

agent to interact with the user in natural language. Experimentation has been set up to obtain human dialogues

between a user dealing with medical information search and a CISMEF expert reﬁning the request. We extend

the GODIS dialogue system with dialogue strategies in order to support system digressions. A model of an

artiﬁcial agent has been implemented.

1 INTRODUCTION

The CISMEF system (www.cismef.org (Darmoni

et al., 2000)) aims to provide access to resources in the

medical domain for healthcare professionals, medi-

cal students but also patients. CISMEF proposes a

graphical user interface and a query language to build

queries using a controlled vocabulary called MeSH

(Medical Subject Headings).

However, users face problems with its use: (i) The

query language is complex and users haveto know the

CISMEF terminology; (ii) Users are confused by the

presentation of query results.

Natural language dialogue interfaces provide a

good solution to these problems. It has been argued

that a mixed-initiative conversational agent provides

an easier, interactive and natural access to informa-

tion (Androutsopoulos et al., 1995).

Our assumption is that the analysis of H-H inter-

actions brings essential hints to design a H-C dialogue

system even though the nature of these interactions is

not strictly identical. We have set up an experiment

to obtain a corpus of human dialogues between a user

playing the role of a patient who searches medical in-

formation and a CISMEF expert trying to help him

to build the query and to obtain answers. The hand-

analysis of these interactions highlights their discur-

sive structure and their linguistic features. Our dia-

logue model is based on this analysis of the corpus.

We have shown that the concepts of GODIS (Lars-

son, 2002) partly meet our needs. We present in this

article the essential elements of the corpus analysis

and the arguments that led us to adopt this system.

2 APPROACHES TO DIALOGUE

SYSTEM DESIGN

2.1 Approaches to Dialogue

Management

Several approaches to dialogue management exist and

no approach clearly dominates the others. The sim-

plest one is the ﬁnite-state (FS) approach (McTear,

2004) that represents the structure of the dialogue as

a ﬁnite-state transition network. In practice, this ap-

proach turns out to be rigid and limited to system-

directed dialogue. The frame-based approach con-

siders the dialogue as a process of ﬁlling in a frame

which contains a series of slots (Aust et al., 1995).

It is less rigid than the FS approach. However, the

possible contributions of the system are ﬁxed in ad-

vance. The plan-based approach (Allen and Per-

rault, 1980) comes from classical AI. It combines

planning techniques such as plan recognition with

ideas from speech act theory. This approach is

rather complex from a computational perspective and

requires advanced Natural Language Understanding

(NLU) components in order to infer the intentions

of the speaker. The logic-based approach represents

the dialogue and its context in some logical formal-

ism and takes advantage of mechanisms such as in-

ference (Hulstijn, 2000). Most works are still on a

theoretical level with this approach. More recent ap-

proaches aim to automatically learn dialogue policies

with machine learning techniques such as reinforce-

ment learning (Frampton and Lemon, 2009). These

312

Loisel A., Dubuisson Duplessis G., Chaignaud N., Kotowicz J. and Pauchet A..

A CONVERSATIONAL AGENT FOR INFORMATION RETRIEVAL BASED ON A STUDY OF HUMAN DIALOGUES.

DOI: 10.5220/0003736703120317

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 312-317

ISBN: 978-989-8425-95-9

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

approaches require an extensive effort of annotation

since a large amount of annotated data is necessary.

2.2 GODIS/IBIS Systems

The Question Under Discussion (QUD)

model (Ginzburg, 1996) is a theory of dialogue

which aims to precisely analyze the properties of

issues. Ginzburg comprehends dialogues as moves

(e.g. ask) that modify the conversational board.

Based on the QUD theory, GODIS and the IBIS

family of systems (Larsson, 2002) are computational

dialogue models which deal with inquiry-oriented di-

alogue. They use a simpliﬁed version of the semantics

of the questions. Three types of question are possible:

yes-no questions (?P), wh-questions (?x.P(x)) and al-

ternative questions ({?P

, ?P

, . .., ?P

}). IBIS also

considers taskAction which is task-dependent. Ques-

tion and taskAction are collectively designated as is-

sue. Three dialogue moves are available, namely ask

(to ask a question), answer (to answer a question) and

request (to request an action from the system).

Issues are integrated in dialogue plans which are

sequences of abstract actions. GODIS is Information-

State-Update-based (Larsson and Traum, 2000; Bos

et al., 2003) and then uses an information state (IS),

close to the conversationalboard. This complex struc-

ture comprises a private part and a shared part. The

private part corresponds to the inner state of the agent

(one interlocutor) and the shared part deﬁnes the con-

versational board, which memorizes shared informa-

tion between the two interlocutors. The update rules

and the selection rules can change the IS. Accord-

ing to the initial dialogue move, the reactive dialogue

moves are generated.

The main beneﬁt of GODIS is the notion of ac-

commodation (Larsson, 2002) which brings ﬂexibil-

ity to the dialogue. This mechanism enables the user

not to answer the current issue, to correct a previous

answer, to answer an unasked question or to move to

another dialogue plan. However, GODIS and IBIS

only represent a simple task with a limited number of

dialogue moves. We discuss in more details the ad-

vantages and drawbacks of GODIS in section 3.3.

3 CORPUS COLLECTION AND

ANALYSIS

3.1 Corpus Collection

An experiment has been set up to obtain human dia-

logues between experts and users, dealing with med-

ical information search. The users were voluntary

members of the LITIS laboratory (secretary, PhD stu-

dents, researchers and teachers), who wanted to ob-

tain answers about medical inquiries. These users

represent the targeted audience for the system since

they are not medical specialists. Experts were two

members of our project, trained in the CISMEF sys-

tem and its terminology. The experimentation took

place as follows: one expert and one user were facing

a computer using the advanced search interface and

recording all the queries along with their answers in

a log. The expert was in charge of conducting the

search by conversing with the user and verbalizing

each action, inquiry and answer. The experimenta-

tion ended when relevant documents were found or

when it seemed that no answer existed in the system.

Having two distinct experimenters increased the vari-

ability of the recorded dialogues. A textual corpus has

been constructed from the transcription of the 21 dia-

logues between the two experts and the 21 volunteers.

It contains 37 000 words.

3.2 Corpus Analysis

We have hand-analyzed the corpus at 3 levels (Loisel

et al., 2008): (i) An analysis of dialogue acts (as con-

textualized speech acts (Bunt, 1996)): a list is pro-

posed according to the illocutionary goals of the utter-

ances of the corpus. (ii) An analysis of sub-dialogues

which considers the dialogue as a sequence of stages.

(iii) An analysis of issues: we listed issues appearing

in each sub-dialogue.

Dialogue acts (Bunt, 1996) ﬁrstly aim to modify

the context and are interpreted according to the cur-

rent situation. This notion is adequate to our study.

Each utterance of our corpus can be broken down

into segments associated with a dialogue act. A list

of dialogue acts has been built according to linguis-

tic features found in the corpus. This taxonomy is

adapted from (Weisser, 2003) who synthesized exist-

ing annotating schemes. Our taxonomy is close to ex-

isting ones such as DAMSL (Core and Allen, 1997).

We set up a list of 36 dialogue acts containing classic

ones (e.g., Answer, Inform, RequestInfo) and ground-

ing acts (Traum, 1994) (e.g., Accept, Acknowledge,

Conﬁrm, Refuse).

We discovered and extracted a global structure for

the corpus dialogues. This structure consists of sub-

dialogues combined with sequence relations. We es-

tablished a list of 14 sub-dialogues like opening, re-

quest formulation and evaluation of the results sub-

dialogues. Their main characteristic is that they are

opportunistic and some of them are optional. For in-

stance, the opening sub-dialogue which includes the

greetings between the user and the system may be

A CONVERSATIONAL AGENT FOR INFORMATION RETRIEVAL BASED ON A STUDY OF HUMAN

DIALOGUES

313

skipped. The study of the dialogues also shows di-

gressions from the expert or the user with inciden-

tal dialogues. For example, a deﬁnition sub-dialogue

may intervene opportunely to bring a deﬁnition of a

CISMEF term to the user.

We established a list of issues for each sub-

dialogue. We simplify the problem by considering

that issues are only a means to obtain CISMEF terms.

That is why we restrict our description semantics to

the ﬁrst-order logic. A total of 44 issues have been

listed in the analysis.

3.3 Discussion on the Corpus Analysis

The analyses of the corpus showed that we were

facing the same kind of dialogues than IBIS, i.e.,

inquiry-oriented dialogue. Then, our idea is to reuse

this already existing system with some adaptations.

The IBIS approach is well adapted to the dia-

logues of our corpus. Firstly, it is based on an explicit

task which does not require much reasoning on the

users’ intentions. The sub-dialogues extracted from

the corpus combined with the corresponding issues

match dialogue plan in IBIS. Secondly, we reckon

that the sub-dialogues could be left explicitly or im-

plicitly. This feature is caught by the accommodation

mechanism. Finally, IBIS dialogue act list is close to

our taxonomy. In particular, our set of grounding acts

corresponds to the Interactive Communication Man-

agement (ICM) (Larsson, 2002) in IBIS.

However, these acts are not sufﬁcient. The di-

alogue acts Inform, Offer and Suggest are missing.

They allow the expert to propose relevant information

opportunistically according to the current results.

Strategies of dialogue manage turns between the

user and the system to efﬁciently conduct the di-

alogue (Caelen, 2003). Caelen distinguishes ﬁve

strategies of dialogue : (i) The directive strategy is

used to lead entirely the user. (ii) The reactive strat-

egy is used when the system executes the orders of the

user. (iii) The constructive strategy introduces digres-

sions to temporarily leave the current goal for a new

one. This strategy is adopted to present examples or

a previous experience useful to the current situation.

(iv) The cooperative strategy adjusts the current goal

to the user’s one. The system tries to ﬁt the goal of the

user by suggesting new information opportunistically

or offering choices, while staying in the same topic.

(v) Finally, the argumentative strategy appears when

the user disagrees with the goal proposed by the sys-

tem. This strategy is mostly found in argumentative

dialogues but our application is not concerned with it.

If we take a closer look at IBIS, we realize that the

directive and reactive strategies of dialogue are avail-

able with the set of dialogue moves provided. As a

matter of fact, the directive strategy consists in the

classic progress in the dialogue plan. Besides, re-

active strategies are managed. Indeed, user-initiated

digressions are allowed with ask-move and request-

move. Moreover, the accommodation mechanism en-

ables this kind of digressions with answer-move.

Cooperative and constructive strategies can be

collectively viewed as system-initiated digressions.

This kind of digressions is not possible with the cur-

rent IBIS system since dialogue plans are rigid and

predetermined. However, it can be interesting to use

system digressions in an information search dialogue.

For instance, we can imagine a system that proac-

tively presents information related to the user request.

That is why one of our goals is to integrate coop-

erative and constructive strategies into IBIS.

4 THE COGNI-CISMEF AGENT

4.1 Integration of Cooperative and

Constructive Strategies in IBIS

The dialogue manager models the sub-dialogues ob-

served in the corpus with a plan library, represents the

common ground and controls the information state

(IS). While it is based on IBIS, it additionally in-

cludes a model of questions with several satisfactory

answers, question accommodation and action accom-

modation, intentional relations and ﬁnally dialogical

strategies. In this section, only parts of the plan li-

brary and dialogical strategies are described.

4.1.1 The Plan Library

Our model uses two kinds of plans described in the

ﬁrst-order logic using “?” to represent the questions:

(i) question plans (PlanQ) which aim at answering

inquiries by returning data; (ii) action plans (PlanA)

which perform sequences of actions. These plans use

a list of actions coming from IBIS: Findout(q) allows

the system to ask a question q by generating the ask-

move. The system asks this question iteratively until

it is answered or cancelled. AssumeAction(a) adds an

action into the IS. Forget and ForgetAll delete knowl-

edge from the IS. This list is not exhaustive.

For instance, ﬁgure 1 corresponds to the action

plan DocumentSearch which begins with the three

ﬁrst steps of a search: the query formulation, the

query building and the display of the current query.

Then, a yes-no question is asked by a Findout so that

the user can validate the query. If the response is neg-

ative, the system forgets the previous results and re-

ICAART 2012 - International Conference on Agents and Artificial Intelligence

314

PlanA(DocumentSearch,

(IfThen(not AddKeyword(m))

(AssumeAction(QueryFormulation),

AssumeAction(QueryBuilding),

AssumeAction(QueryDisplay),

Findout(QuerySatisfaction),

IfThen(not QuerySatisfaction)

(Forget(Done(QueryBuilding)),

Forget(Done(AddQuery)),

Forget(Resolved(QuerySatisfaction)),

Forget(Resolved(?x2.AddKeyword(x2))),

Forget(Resolved(?x3.AddSubheading(x3))),

Forget(not QuerySatisfaction),

AssumeAction(QueryBuilding)),

AssumeIssue(?x.Documents(x)),

AssumeAction(EvaluationListeDocuments),

Findout(?NewSearch),

IfThen(NewSearch)

(ForgetAll, AssumeAction(DocumentSearch))

Confirm(DocumentSearch)))

Figure 1: Action plan DocumentSearch.

initiates the plan with the action QueryBuilding to re-

ﬁne the query. Otherwise, the query is performed and

the results are appraised. Unless there is an explicit

exit dialogue act uttered by the user or by the system,

or an exit planned by the current plan, the system re-

mains in this plan. All these actions are linked by

satisfaction-precedence relations.

4.1.2 The Dialogical Strategies

The Cooperative strategies modify plans in order to

propose suggestions or help. The dialogue has to

remain in the same context. The current plan is al-

ways active and the system proposes to the user new

information acquired dynamically during the search.

Practically speaking, it dynamically builds new issues

(questions, actions or propositions) and adds them to

the current plan (in shared/plan, while keeping the

other ﬁelds of the IS unchanged.

A new action called CooperativeAction is created

in three ways: (i) CooperativeAction (?x) to ask a new

question to the user; (ii) CooperativeAction(action)to

add a new taskAction; (iii) CooperativeAction(p(x))to

propose an answer or a suggestion to a pending ques-

tion. The system can also answer its own questions by

consulting the database. This predicate can be added

at any place into a plan, where it is relevant to propose

suggestions to the user.

When this action is found in the current plan, an

update rule called ExecCooperativeAction is applied

(Figure 2). It runs the procedure FindCooperative-

PlanAction which ﬁnds new information to be pre-

sented to the user. Finally, the rule adds a new plan

action to the top of the stack private/plan.

For example, one major issue for a search en-

RULE : ExecCooperativeAction

CLASS : PlanExecution

PRE :

(

empty(

/private/plan

top(

/private/action

, CooperativeAction(Q)),

EFF :











P1=content(

/shared/com

)

P2=content(

/private/bel

)

Concatenate(P1, P2, Ps)

FindCooperativePlanAction(Q,Ps,PlanAction)

Add(

/private/plan

, PlanAction)

Figure 2: Update rule ExecCooperativeAction. PRE stands

for preconditions on the IS. EFF stands for effects on the IS.

PlanA(ResultEvaluation,

(InformIntent(ResultEvaluation),

AssumeIssue(?n.NbDocuments(n)),

InformIntent(ImproveQuery),

IfThen(n isTooHigh)

(Inform(TooManyDocuments),

CooperativeAction(?x.Issue(x) TooManyDocuments)),

IfThen(n isTooLow)

(Inform(NotEnoughDocuments),

CooperativeAction(?x.Issue(x) NotEnoughDocuments)),

ForgetIssue(),

ForgetAction(),

AssumeAction(DocumentSearch),

Assume(CorrectNbDocuments),

Confirm(ResultEvaluation)))

Figure 3: Action plan ResultEvaluation.

gine is to ﬁnd the correct number of documents to

be presented to the user. This issue concerns the Re-

sultEvaluation (ﬁgure 3) plan which labels the num-

ber of documents found with TooManyDocuments or

NotEnoughDocuments. This label is added to the IS

through the action plan Inform. If this number is too

high, the plan calls the CooperativeAction plan, to re-

ﬁne the query by proposing suggestions dynamically

pushed in private/bel and private/plan. A similar plan

is executed when no or very few documents are found.

Then, the search has to be launched again, taking into

account the user’s answers to these suggestions. By

plan accommodation, the user can add new terms or

address new issues to the query. The plans Forge-

tAction and ForgetIssue are used to clean the IS. The

plan AssumeAction is performed to relaunch the new

query. Finally, if the number of documents is correct,

the result can be reﬁned by sorting documents to pro-

pose the most relevant ﬁrst.

The Constructive strategies aim to bring explana-

tion sequences during the dialogue. It diverts the sys-

tem from its current plan: new information has been

acquired that makes the system propose digressions,

hints or examples. This new information come from

different sources: opportunely during the dialogue

(e.g., the user speciﬁes that he is a medical expert) or

A CONVERSATIONAL AGENT FOR INFORMATION RETRIEVAL BASED ON A STUDY OF HUMAN

DIALOGUES

315

from the task model which can, e.g., recommend doc-

ument resource types. The system tries to launch such

digression plans at each update of the IS. It is done

through a rule database (domain-dependent) which

proactively scans the IS to run new additional plans.

If a rule matches the current situation, a new action or

question (goal of the corresponding plan) is added in

shared/issues or shared/action to be performed. The

manager ﬁnds the corresponding plans and then pro-

duces the suitable dialogue moves. When the con-

structive sequence is ended, the digression issue is

removed from shared/issues. The previous context

is recovered since the previous issue is now ﬁrst in

shared/issues. Thus, the constructive rules are some

kind of accommodation rules where the system ac-

commodates to its own chosen plans rather than those

of the user. This mechanism allows the system to

leave (temporarily) the current plan.

In our application, the system can propose infor-

mation focusing on a certain type of documents, even

if the user did not ask for it. A rule is then added

to propose a certain type of documents according to

the user identity, before running a query. This rule

is then initiated by the CooperativeAction plan and

added into the DocumentSearch plan.

IF (exists User(patient) in /private/bel)

AND (exists DocumentSearch in /private/action)

AND (neg exists RessourceType(x) in /private/bel)

THEN Launch(Suggest(Plan(Question(RessourceType(x)))))

4.2 The Task Model

The task model gives access to advanced function-

nalities of CISMEF to the other modules composing

COGNI-CISMEF. It includes: (i) The CISMEF ter-

minology which enables the system to determine sub-

headings, deﬁnitions, hypernyms and hyponyms as-

sociated with a term. It also contains a list of terms

used by patients in order to detect medical keywords

in user utterances. (ii) A query builder which uses

terms detected by the NLU component. It builds a

query as a human expert would do in the graphical

interface. (iii) A result interpreter which reﬁnes the

query if necessary. It includes various search strate-

gies (e.g., if the query returns too many documents, it

is necessary to specify it by adding new terms). This

interpreter suggests to the dialogue manager to add

relevant issues in order to restrict the query. (iv) The

CISMEF engine which runs the query so as to access

to medical documents.

4.3 Natural Language Understanding

and Generation

COGNI-CISMEF receivesthe user’s utterance in text-

written natural language in French and carries out two

main analyses: a semantic one and a pragmatic one.

The semantic analysis starts with a recognition of

textual units. Afterwards, recognition and labeling of

CISMEF terms is carried out. These terms are iden-

tiﬁed by the database via the task model.

The pragmatic analysis is driven by a rule-based

dialogue act interpreter. The rules used are based on

linguistic markers, such as the tense and the mode of

the verb, interrogation markers, acceptance markers,

etc., associated with the immediate dialogue co-text

(the previous statement of the dialogue). Finally, the

results of these analyses are concatenated to form a

proposition usable by the dialogue manager.

The sentence generator is rather simple since it

translates a logical proposition coming from the di-

alogue manager into an instantiated incomplete sen-

tence. However, the system uses mechanisms that

produce more natural utterances. For example, a ran-

dom generator diversiﬁes the most frequent utterances

as greetings, agreements, etc.

5 DIALOGUE EXAMPLE

It is possible to perform a full dialogue by select-

ing appropriate terms. Listing 1 presents a real com-

mented dialogue between the COGNI-CISMEF sys-

tem (CC) and a user (U).

Listing 1: Example of dialogue with COGNI-CISMEF.

: Welcome to Cognicismef. What can I do for you?

% Opening sub -dialogue + choice sub-dialogue

: Give me a definition of paludism

% Definition sub-dialogue

: The term means: "Also called malaria (...)"

Now, what can I do for you?

% Answer + come back to the choice sub -dialogue

: Search documents about it

% Query formulation sub-dialogue

: OK. Do you want to add the keyword "paludism"?

% Issue establishment + query building sub-

dialogue (dependance relation)

: Yes

: OK. Do you want to add something else?

: Are there therapeutics?

: OK. Do you want to add the meta -term "

therapeutic" or the subheading "therapeutic"?

% Query building sub-dialogue

: What is a subheading?

% Definition sub-dialogue

: Subheadings are general concepts associated

with a keyword to precise its meaning

ICAART 2012 - International Conference on Agents and Artificial Intelligence

316

: Add the subheading therapeutic

% Come back to the query building sub-dialogue

: OK. Do you want to add something else?

: No

: OK. The query has been modified. Here is the

current query:

keyword(paludism), subheading(therapeutic)

Now, does it suit you?

% Query execution sub-dialogue

: Yes

: OK. I have performed the query. I have found

results. The list includes 11 documents.

That is too much. We can choose more

specialized terms.

% Result evaluation sub-dialogue + proposition

of query expansion

: Are you patient or health professional?

% Issue accommodation

: Patient

: We can add patient specialized document.

% Query building sub-dialogue

: Yes

: OK. The query has been modified. Here is the

current query:

keyword(paludism), subheading(therapeutic),

resourcetype(patient)

Now, does it suit you?

: Yes

: OK. I have performed the query. I have found

results. The list includes 1 document.

Here are the titles:

"Modele de chapitre pour les manuels PCIME"

Which number of document do you want?

% Document selection sub -dialogue

: the first

: The selected document is:

"Modele de chapitre pour les manuels PCIME"

Now, do you want something else?

: No

6 CONCLUSIONS

This article presents the design of a human-computer

dialogue system, based on QUD theory and IBIS,

from corpus collection to implementation. Our con-

versational agent requires enhancements: at least, the

semantic analysis needs a bigger lexicon. Then, the

system will be able to be evaluated. It will consist in

a comparison of the users’ requests using CISMEF

with those proposed by the librarian of the medical

library of the hospital of Rouen and those built by

users using COGNI-CISMEF. We will measure the

improvements (in terms of precision and recall) re-

spectively of the COGNI-CISMEF system and the li-

brarian compared to the request using only CISMEF.

REFERENCES

Allen, J. and Perrault, C. (1980). Analyzing intention in

utterances. Artiﬁcial Intelligence, 15(3):143–178.

Androutsopoulos, I., Ritchie, G., and Thanisch, P. (1995).

Natural language interfaces to databases–an introduc-

tion. Natural Language Engineering, 1(01):29–81.

Aust, H., Oerder, M., Seide, F., and Steinbiss, V. (1995).

The philips automatic train timetable information sys-

tem. Speech Communication, 17(3-4):249–262.

Bos, J., Klein, E., Lemon, O., and Oka, T. (2003). DIPPER:

Description and formalisation of an information-state

update dialogue system architecture. In SIGdial Work-

shop on Discourse and Dialogue, pages 115–124.

Bunt, H. (1996). Dynamic interpretation and dialogue the-

ory. Dialogue and instruction.

Caelen, J. (2003). Strategies of dialogue. In Speech Tech-

nology and Human-Computer Dialogue Conference,

pages 27–42.

Core, M. and Allen, J. (1997). Coding dialogs with the

damsl annotation scheme. In AAAI Fall Symposium

on Communicative Action in Humans and Machines,

pages 28–35.

Darmoni, S., Leroy, J., Baudic, F., Douyre, M., Piot, J., and

Thirion, B. (2000). Cismef : a structured health re-

source guide. Methods of Information in Medicine,

39:30–35.

Frampton, M. and Lemon, O. (2009). Recent research

advances in reinforcement learning in spoken dia-

logue systems. The Knowledge Engineering Review,

24(04):375–408.

Ginzburg, J. (1996). Interrogatives: Questions, facts, and

dialogue. The Handbook of Contemporary Semantic

Theory, pages 385–422.

Hulstijn, J. (2000). Dialogue models for inquiry and trans-

action. PhD thesis, University of Twente.

Larsson, S. (2002). Issue-based dialogue management. PhD

thesis, Department of Linguistics, G¨oteborg Univer-

sity.

Larsson, S. and Traum, D. (2000). Information state

and dialogue management in the TRINDI dialogue

move engine toolkit. Natural language engineering,

6(3&4):323–340.

Loisel, A., Kotowicz, J., and Chaignaud, N. (2008). An

issue-based approach to information search mod-

elling: Analysis of a human dialogue corpus. In Text,

Speech and Dialogue, pages 609–616.

McTear, M. (2004). Spoken dialogue technology: toward

the conversational user interface. Springer-Verlag

New York Inc.

Traum, D. (1994). A Computational Theory of Grounding

in Natural Language Conversation. PhD thesis, Uni-

versity of Rochester.

Weisser, M. (2003). Spaacy: A tool for annotating dialogue.

International Journal of Corpus Linguistics, 8(1).

A CONVERSATIONAL AGENT FOR INFORMATION RETRIEVAL BASED ON A STUDY OF HUMAN

DIALOGUES

317