CONVERSATIONAL AGENT IN ARGUMENTATION

A Model and Evaluation on a Dialogue Corpus

Mare Koit

Institute of Computer Science, University of Tartu, 2 J. Liivi Str, Tartu, Estonia

Keywords: Argumentation, Communicative strategy, Communicative tactics, Dialogue corpus, Dialogue system.

Abstract: Communication between two participants, A and B, is considered, where A has a communicative goal that

his/her partner, B, will make a decision to perform an action D. A computational model of argumentation is

developed which includes reasoning. Communicative strategies and tactics used by participants for

achieving their communicative goals are considered. A simple dialogue system (conversational agent) is

implemented which can optionally play the role of A or B using classified sets of pre-defined Estonian

sentences. For further evaluation of the model and with the aim to develop the dialogue system, the analysis

of the Estonian Dialogue Corpus is carried out. Calls of sales clerks who persuade clients to take training

courses of an educational company are analysed. The calls end mostly with the postponement of the

decision therefore the sales clerks do not achieve their communicative goal.

1 INTRODUCTION

There are many dialogue systems (DS) that interact

with a user in a spoken natural language and help

him/her to solve practical problems, e.g. to book

flights, to get information about bus or train

timetables, to detect computer faults, etc (McTear,

2004). Usually, these tasks do not include

argumentation. Rather practical dialogue is

implemented in such systems. On the other hand,

there are tasks and situations where not only

information search, but also argumentation is

required.

Analysis of human-human dialogues can provide

information about their structure and linguistic

features with the purpose of developing a DS. Some

of the well-known dialogue corpora are the HCRC

Maptask, TRAINS, VERBMOBIL (McTear, 2004).

Our research is based on the Estonian Dialogue

Corpus (EDiC). We investigate the conversations

where the goal of one participant, A, is to get another

participant, B, to carry out a certain action D. This

type of dialogue forms one kind of so-called

agreement negotiation dialogues. In this paper, we

consider the dialogues where sales clerks of an

educational company call another institution (a

manager or another responsible person) and offer

courses of their company. We may expect that a

sales clerk tries to influence the partner in such a

way that s/he decides to book a course for the

employees of his/her institution. We are looking for

ways of argumentation both of the sales clerks and

the clients. Our further goal is to develop a DS

which participates in an agreement negotiation

dialogue with a user in a natural language, optionally

performing the roles of A or B. Because of this, we

have modelled the reasoning processes that people

supposedly go through when working out a decision

whether to perform an action or not.

2 CONVERSATIONAL AGENT

Let us consider a conversational agent as a program

that consists of six modules:

(PL, PS, DM, INT, GEN, LP),

where PL – planner, PS – problem solver, DM –

dialogue manager, INT – interpreter, GEN –

generator, LP – linguistic processor. PL directs the

work of both DM and PS, where DM controls

communication process and PS solves domain-

related tasks. The task of INT is to make semantic

analysis of partner’s utterances and that of GEN is to

generate semantic representations of agent’s own

contributions. LP carries out linguistic analysis and

generation.

Conversational agent uses in its work goal base

GB and knowledge base KB which consists of four

552

Koit M..

CONVERSATIONAL AGENT IN ARGUMENTATION - A Model and Evaluation on a Dialogue Corpus.

DOI: 10.5220/0003141205520555

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 552-555

ISBN: 978-989-8425-40-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

components: (KB

, KB

), where KB

contains world knowledge, KB

– linguistic

knowledge, KB

– knowledge about dialogue and

– knowledge about interacting subjects. KB

contains definitions of dialogue acts and algorithms

that are applied to reach communicative goals –

communicative strategies and tactics. KB

contains

knowledge about evaluative dispositions of

participants towards the action(s) (e.g. what do they

consider as pleasant or unpleasant, useful or

harmful), and, on the other hand, algorithms that are

used to reason about actions.

In this paper, we concentrate on the parts KB

and KB

of the knowledge base – the reasoning

model which uses a model of the motivational

sphere of an agent who is reasoning to do an action

or not, and communicative strategies and tactics

used by agents in order to achieve their

communicative goals.

2.1 Reasoning Model

If a conversation in a natural language takes place

between two agents – A and B – then in the goal

base of one participant (let it be A) a certain goal G

related to B’s activities gets activated and triggers a

reasoning process in A. In constructing his/her first

utterance A must plan the dialogue acts and

determine their verbal form as an utterance r

. This

utterance triggers a reasoning process in B where

two types of procedures should be distinguished: the

interpretation of A’s utterance and the generation of

his/her response r

. B’s response triggers in A the

same kind of reasoning cycle in the course of which

s/he has to evaluate how the realization of his/her

goal G

has proceeded. Depending on this s/he may

activate a (new) sub-goal of G

, and the cycle is

repeated. A dialogue comes to an end, when A has

reached his/her goal or abandoned it.

In general, our reasoning model follows the ideas

realised in the BDI model (Allen, 1994). Our model

consists of two functionally linked parts (Koit and

Õim, 2004): (1) a model of a motivational sphere of

a subject who is reasoning to perform an action D or

not, and (2) reasoning procedures. We represent the

model of motivational sphere by the vector of

“weights” of different aspects of the action (these

are e.g. presence of the resources for doing D,

pleasantness of D, etc.):

w = (w(are-resources),

w(pleasantness), w(unpleasantness), w(usefulness),

w(harmfulness), w(is-obligatory), w(is-prohibited),

w(punishment-for-doing-a-prohibited-action),

w(punishment-for-not-doing-an-obligatory-action)).

In the vector, the components w(pleasantness),

w(usefulness), etc. mean weights of different aspects

of D. For simplicity, it is supposed that the aspects

have numerical values and that in the reasoning

process their values can be summed up. In this way,

the model considers a conversational agent from the

standpoint of an action. Here w(resources) = 1, if the

agent has resources necessary to do D (otherwise 0);

w(obligatory) = 1, if D is obligatory for the

reasoning subject (otherwise 0); w(prohibited) = 1, if

D is prohibited (otherwise 0). The values of other

weights are non-negative natural numbers.

The second part of the reasoning model consists of

reasoning schemes that regulate human action-

oriented reasoning. A reasoning scheme represents

steps that the subject goes through in his/her

reasoning process; these consist in computing and

comparing the weights of different aspects of D; and

the result is the decision – to do D or not.

Three basic factors that regulate reasoning of a

subject concerning D are differentiated: his/her

wishes, needs and obligations (Õim, 1996): (1)

subject may wish to do D, if pleasant aspects of D

for him/her outweigh unpleasant ones; (2) subject

may find useful to do D, if D is needed to reach

some higher goal, and usefulness of D outweighs

harmfulness; and (3) subject can be in a situation

where s/he must (is obliged) to do D – if not doing D

will lead to some kind of punishment.

Respectively,

there are three reasoning procedures (WISH,

NEEDED and MUST) in our model depending on

the factor that triggers the reasoning. Each procedure

represents the steps that a subject goes through in the

reasoning process, computing and comparing

weights of different aspects of D (Koit et al., 2009).

2.2 Communicative Strategies and

Tactics

Communication takes place in so-called

communicative space. The communicative space is

determined as an N-dimensional space where

coordinates characterize the relationships of

participants (in our model, N=5). For example,

communication can be measured on the scales

personal-impersonal, collaborative-confrontational.

A communicative strategy is an algorithm which

is used by a communication participant to achieve

his/her communicative goal. A communicative

strategy for A (who has the goal that the partner B

decides to perform D) can be represented as the

following algorithm.

1. Choose an initial point in the

communicative space.

2. Choose a communicative tactic.

3. Implement the tactic to generate a

utterance: inform the partner of the

communicative goal (decision to do

an action D).

CONVERSATIONAL AGENT IN ARGUMENTATION - A Model and Evaluation on a Dialogue Corpus

553

4. Did the partner agree to do D? If

yes then finish (the communicative

goal has been achieved).

5. Give up? If yes then finish (the

communicative goal has not been

achieved).

6. Change the point in the

communicative space? If yes then

choose a new point.

7. Change the communicative tactic? If

yes then choose a new tactic.

8. Implement the tactic to generate a

utterance (an argument) for doing D.

9. Go to 4.

The participant A can realize his/her communicative

strategy in different ways (using different arguments

for doing D): stress pleasantness of D (i.e. entice B),

stress its usefulness (i.e. persuade B), or stress

punishment for not doing D if it is obligatory (i.e.

threaten B). We call these certain ways of realization

of a communicative strategy communicative tactics.

That can be considered as argumentation: A, trying

to direct B’s reasoning to the positive decision (to do

D), proposes various arguments for doing D while B,

when opposing, proposes counter-arguments.

These three tactics are connected with the

reasoning procedures WISH, NEEDED, and MUST,

respectively. The general idea underlying the tactics

is that A proposes arguments for pleasantness of D

(when enticing), usefulness of D (when persuading)

and punishment of not doing D (when threatening)

trying to keep the weight of pleasantness, usefulness

of doing D or punishment of not doing an obligatory

D high enough and the possible values of other

aspects brought out by B low enough so that it

would bring B to the decision to do D (cf. Koit et al.,

2009).

The tactics for B are collaboration and

antagonism. In the first case, B is interested in doing

D and, in collaboration with A, is looking for

arguments that support his/her positive decision. In

the second case, B only uses arguments against D,

his/her goal is opposite with A’s (like in two player

games). Both A and B can implement a mixed

strategy – change their communicative tactics during

a conversation.

3 CORPUS ANALYSIS

In the following, we carry out corpus analysis, in

order to evaluate the communicative tactics in our

model. For that, 30 phone calls are taken from the

EDiC where sales clerks of an educational company

offer different courses to clients. The action D is ’to

take the offered course’. In the case of institutional

communication, both of enticing and threatening

should be excluded because a clerk is an official

person and s/he is obliged to communicate

cooperatively, impersonally, peacefully, etc (i.e. to

stay in a fixed point of the communicative space).

S/he can only persuade a client.

3.1 Sales Clerks’ Tactics

All the dialogues are recorded in the beginning

phase of negotiations, therefore, B takes a final

decision only in few cases. A typical dialogue starts

with A’s introduction and a question whether B does

know the education company. Then a short overview

of the company is given (e.g. we are an

international company, we are acting six years in

Estonia, we are dealing with sale, service,

management, marketing). All the statements can be

considered as arguments for taking a training course.

Then a proposal is made by A to take some courses.

A points the activities of B’s organisation which

demonstrates that s/he has pre-knowledge about the

institution (e.g. your firm is dealing with retail and

whole sale, therefore you could be interested in our

courses). If B does not make a decision then A asks

B to tell more about B’s institution in order to get

more arguments for usability of the courses for B,

and offers them again. The dialogues end with an

agreement to keep the contact (A promises to send

information materials to B, to call B later). B does

not decide to accept nor reject a course but

postpones the decision.

If A and B have been in contact before, then A

always starts the conversation with pointing to a

previous contact. B has had the time to evaluate the

information about the courses in order to make a

decision. B agrees to take a course only in one

conversation, s/he agrees with reservations in two

dialogues, and does not agree in one dialogue. In the

remaining dialogues, A and B come to the agreement

to keep the contact like in the case of the first

communication. Therefore, B typically postpones the

decision.

3.2 Clients’ Tactics

A’s final goal is that B decides to do D (to take a

course). In the case of collaboration, B actively

looks for arguments for doing D.

In a typical dialogue, A introduces himself, gives

an overview of his company (it offers courses of

management, marketing, sale, customer service,

secretary training), and asks whether B has made

training plans for his employees (i.e. an indirect

proposal to take a course). B argues that his staff is

small, and he has got many offers from other

training companies (i.e. a refusal with two

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

554

arguments). Then A tries to awake B to a certain

course by asking about customers of B’s firm. After

that an offer is made to send a catalogue. Now, B

takes the initiative starting to check the presence of

resources and usability of performing D. At the end

of conversation, A and B agree that A sends a

catalogue and calls B again a week later.

Pure antagonism is expressed in one dialogue. B

has studied the catalogue, and made the negative

decision (but yes, I have studied it and

unfortunately, I’ll say that you are not able to teach

what I want). A is looking for new arguments and

asks questions about activities of B’s company

trying to show that the courses are useful for B.

Anyway, B does not give up, and the dialogue ends

with a resolute refusal. Here, both participants try to

take initiative. A implements the tactic of persuasion

but B does not capitulate.

In most cases, B having studied a catalogue,

starts a conversation with antagonism but goes over

to collaboration, i.e. uses a mixed tactic.

4 CONCLUSIONS

When communicating in a natural language, where

A tries to influence B in order to bring him/her to a

decision to perform an action D, A uses several

arguments in order to increase the weights of the

positive aspects and to decrease the weights of the

negative aspects of the action under consideration.

When B has started the reasoning process, s/he

considers various positive and negative aspects of D.

If the positive aspects weigh more, B will make the

decision to do D, otherwise s/he will make the

decision not to do D. If B indicates a certain aspect

which does not allow him/her to do D then A simply

can choose an argument for attacking this aspect

until there are arguments at his/her disposal. When

reasoning, B can make his/her negative decision on

different steps. For example, if B says that resources

are missing and A indicates that resources can be

obtained then B has to start his/her reasoning again

from the beginning. If B does not indicate a certain

reason of rejection then A can only stress the

usefulness of D (when persuading).

The corpus analysis shows that our argumentation

model is a coarse approximation to real human

argumentation. For example, it does not allow

deviations from the main line of argumentation by

asking questions about possible new arguments.

Nevertheless, we believe that the model can be

useful for training argumentation in such a way that

various arguments are classified and a strong

discipline is set to the order of using different

arguments. For example, when persuading, stress the

usefulness of D until the set of arguments becomes

exhausted. At the moment, a simplified version of

our model is implemented as an experimental DS

which can optionally play the role of A or B using

classified sets of pre-defined Estonian sentences and

can be used as a “communication trainer”.

Similar approaches are described in some other

papers, e.g., presenting arguments in (Elhadad,

1995), car selling agent in (Piwek and van Deemter,

2007). Comparison of the approaches remains for

the further work.

We are continuing our work in the following

directions: (1) analysis of human-human dialogues

in the EDiC in order to verify and to refine the

model, (2) specifying resistance strategies, (3)

developing linguistic knowledge of the DS.

ACKNOWLEDGEMENTS

This work is supported by the European Regional

Development Fund through the Estonian Center of

Excellence in Computer Science (EXCS), and the

Estonian Science Foundation (grant 7503).

REFERENCES

Allen, J., 1994. Natural Language Understanding. Second

Edition, The Benjamins/Cummings Publ. Co.

Elhadad, M., 1995. Generating coherent argumentative

paragraphs. In: Journal of Pragmatics. North Holland,

Elsevier, 24, 189–220.

Koit, M., Õim, H., 2004. Argumentation in the Agreement

Negotiation Process: A Model that Involves Natural

Reasoning. In: Proc. of the Workshop W12 on

Computational Models of Natural Argument. 16

European Conference on Artificial Intelligence.

Valencia, Spain, 53–56.

Koit, M., Roosmaa, T., Õim, H., 2009. Knowledge

Representation for Human-machine Interaction. In:

Proc. of International Conference on Knowledge

Engineering and Ontology Development. Portugal,

INSTICC, 396–399.

McTear, M.F., 2004. Spoken Dialogue Technology.

Toward the Conversational User Interface, Springer.

Õim, H., 1996. Naïve theories and communicative

competence. Reasoning in communication. In:

Estonian in the changing world. Tartu University,

211–231.

Piwek, P., van Deemter, K., 2007. Generating under

Global Constraints: the Case of Scripted Dialogue. In:

Journal of Research on Language and Computation,

5(2), 237–263.

CONVERSATIONAL AGENT IN ARGUMENTATION - A Model and Evaluation on a Dialogue Corpus

555