From Study of Human-human Dialogues to Reasoning Model

Conversational Agent in Argumentation Dialogue

Mare Koit

and Haldur Õim

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

Institute of Estonian and General Linguistics, University of Tartu, 2 J. Liivi St., Tartu, Estonia

Keywords: Reasoning, Conversational Agent.

Abstract: We study human-human dialogues where one of the participants tries to influence the reasoning process of

the dialogue partner in order to force the partner to make a decision to perform an action. Our further aim is

to implement a dialogue system which would interact with a user in natural language. A model of the

motivational sphere of a reasoning subject will be presented as a vector which consists of evaluations of

different aspects of the action under consideration. Three reasoning procedures will be introduced, each of

which is triggered by a so-called input factor. We examine the communicative strategies and communicative

tactics that dialogue participants use to achieve their communicative goals. The models are implemented as

a computer program.

1 INTRODUCTION

Several general approaches to the theory of

pragmatics of natural communication have been

used when developing models of dialogue

(D´Andrade, 1987); (Davies and Stone, 1995);

(Lester et al., 2004); (Jokinen, 2009); (Ginzburg and

Fernández, 2010). One of the main problems for

pragmatics is that of modelling the mechanisms

people use to reach their communicative goals. This

is done by manipulating the appropriate reasoning

processes of other participants. Therefore, a

pragmatic model of dialogue should include a

commonsense, naïve model of reasoning and present

the means to influence the reasoning processes that

are regularly used by people in communication.

Most of our naïve theories are and will remain

implicit. It is the task of a real science, such as

psychology or linguistics, to explicate this

knowledge (Õim, 1996).

We study communications where one of the

participants is trying to achieve the partner’s

decision to perform an action and have worked out

our versions of the concepts of communicative

strategies and tactics. Below we focus on these

concepts and the concept of reasoning model,

describing the current state of their implementation

in our dialogue system. In general, our model of

reasoning follows the ideas realized in the well-

known BDI (belief-desire-intention) model (Rao and

Georgeff, 1991).

The paper has the following structure. In section

2 we introduce our reasoning model. Section 3

examines how the reasoning can be influenced in

interaction. We introduce a communicative strategy

and communicative tactics as algorithms used by

participants in order to achieve their communicative

goals. Section 4 discusses how the dialogue model is

and can be implemented. In section 5 we draw

conclusions.

2 REASONING MODEL

2.1 Human Reasoning

Let us start with considering a dialogue example

(Ex. 1) taken from the Estonian dialogue corpus. A

client (participant A) is calling a travel agent (B) and

is looking for a ski trip to Austria (action D). In the

following examples the transcription of conversation

analysis is used (Hutchby and Wooffitt, 1998).

(1)

A : .hh e sooviksin teada: natuke

´Austria suusareisi kohta. | REQUEST |

I’d like to know a bit about the ski

trip to Austria

B : jaa? | CONTINUER |

210

Koit M. and Õim H..

From Study of Human-human Dialogues to Reasoning Model - Conversational Agent in Argumentation Dialogue.

DOI: 10.5220/0004195902100216

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 210-216

ISBN: 978-989-8565-39-6

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

yes

A : .hh et=mm (.) kui ´kaua se ültse

´kestab. | WH-QUESTION |

how long will it last

/---/

A : .hh ja (.) mis: ´hinnaklass see

tuleb kui näiteks apartemendi vari´ant

võtta.= | WH-QUESTION |

and what price category would it be for

the option with accommodation in an

apartment

/---/

A : .hh ja siis selle ´suusatamisega

seal on: (.) ´mägedes suusatamine. |

QUESTION OFFERING ANSWER |

and about skiing will the skiing be in

mountains

B : jah. | YES |

yes

A : ja: kas need´suusad ja ´varustus on

nagu ´lisatasu eest või see on: [(.)

´hinna sees.] | ALTERNATIVE QUESTION|

and the skies and the equipment are

there for extra cost or is it included

in the price

A is collecting information that he needs to make a

decision about the action (here: ski trip). We do not

know exactly how the reasoning process proceeds in

A’s head. Only A’s utterances provide the indirect

signals for the travel agent to help her draw

conclusions about how several aspects of the action

are weighed by A.

2.2 Model of Reasoning Subject

We assume that the reasoning process of a certain

type is triggered by a so-called input factor. We

distinguish between three types of input factors: (1)

the reasoning subject may wish (would like) to

perform action D (wish-factor), (2) the subject may

depart from the assumption that doing D is useful for

him (needed-factor), or (3) that doing D is

obligatory (must-factor). In short, these factors

constitute what can be called the (macro) model of

human motivational system underlying his/her

reasoning whether to take an action or not. When the

reasoning process has started, the subject considers

(weighs) the positive and negative aspects of D: how

pleasant or unpleasant, useful or harmful it is, what

punishment will follow if he does not do D, etc. If

the positive aspects weigh more, the subject will

make the decision to perform D, otherwise the

decision will be not to do D.

When constructing our model of reasoning we

assume that the reasoning subject is somehow able

to evaluate the positive and negative aspects of the

object of reasoning (in our case, action D). Here we

assume that the relevant aspects of D can be

characterized by scales that take certain numerical

values, the so-called weights. Although, in reality

people do not operate with numbers, the existence of

certain scales also in human everyday reasoning is

apparent. What would be a more adequate form for

these scales, is a problem of future research.

We use the following notation: w(pleasant),

w(unpleasant), w(useful), w(harmful) – weight of

the pleasant, unpleasant, useful and harmful aspects

of D, respectively; w(obligatory) – the value shows

whether D is obligatory (=1) or not (=0),

w(prohibited) – whether D is prohibited (=1) or not

(=0), w(punishment-do) – weight of punishment for

performing a prohibited action, w(punishment-do-

not) – weight of punishment for not performing an

obligatory action, w(resources) – the value indicates

whether the subject has resources necessary for

performing D (=1) or not (=0).

According to our present model, the motivational

sphere of a reasoning subject can be represented by

the following vector of weights:

w = (w(resources), w(pleasant),

w(unpleasant), w(useful), w(harmful),

w(obligatory), w(prohibited),

w(punishment-do), w(punishment-do-

not)).

2.3 The Three Reasoning Procedures

The model of the motivational sphere is used by a

reasoning subject when weighing different aspects

of the action under consideration. Reasoning is

triggered by an input factor. As an example, we

present the reasoning procedure WISH that

originates in the wish of a subject to do D (Fig. 1).

The prerequisite for triggering this procedure is

w(pleasant) > w(unpleasant) based on the following

assumption: if a person wishes to do something, then

he assumes that the pleasant aspects of D (including

its consequences) overweigh its unpleasant aspects.

Different aspects of D are subsequently

evaluated and the final decision depends on the

result of the comparison of different values as

outcomes of the evaluation process. For instance, if

the subject does not have enough resources for D,

then, independently of his wish, he will decide not to

do D (step 1 in the procedure). If the subject has the

necessary resources and the weight of the pleasant

aspects exceeds the sum of the unpleasant and the

harmful ones and, in addition, D is not prohibited,

then the subject will decide to perform D (steps 1, 2

and 3), etc. The same kinds of procedures, NEEDED

FromStudyofHuman-humanDialoguestoReasoningModel-ConversationalAgentinArgumentationDialogue

211

and MUST, are constructed for the reasoning

processes triggered by the needed and must factors

(Koit and Õim, 2004).

Prerequisite: w(pleasant) > w(unpleasant)

1) Are there enough resources for doing

D? If not then do not do D.

2) Is w(pleasant) > w(unpleasant) +

w(harmful)? If not then go to step 6.

3) Is D prohibited? If not then do D.

4) Is w(pleasant) > w(unpleasant) +

w(harmful) + w(punishment-do)? If yes

then do D.

5) Is w(pleasant) + w(useful) >

w(unpleasant) + w(harmful) +

w(punishment-do)? If yes then to do D

else do not do D.

6) Is w(pleasant) + w(useful) >

w(unpleasant) + w(harmful)? If yes then

go to step 9.

7) Is D obligatory? If not then do not

do D.

8) Is w(pleasant) + w(useful) +

w(punishment-do-not) > w(unpleasant) +

w(harmful)? If yes then do D else do not

do D.

9) Is D prohibited? If not then do D.

10) Is w(pleasant) + w(useful) >

w(unpleasant) + w(harmful) +

w(punishment-do)? If yes then do D else

do not do D.

Figure 1: Reasoning procedure WISH.

Returning to Ex. 1, we can see how A’s

utterances keep a “trace” of his reasoning process: A

is weighing the duration of the trip, its price, the

skiing location, etc. Every utterance refers to a

specific aspect of D, the action under consideration.

2.4 Discussion

We do not claim that the above three reasoning

procedures – WISH, NEEDED and MUST –

exhaustively cover all the varieties of reasoning on

which human action is based that can be

encountered in the “real life”. There are numerous

kinds of situations not accounted for by our model

so far, although we have dealt with them

theoretically. For instance, there can be situations

where a person has several simultaneously activated

and competing input factors, e.g. two competing

wishes for both of which he holds w(pleasant) >

w(unpleasant). Such situations can be partly

accounted for by certain general principles (e.g.

“From two possible pleasant situations people prefer

the more pleasant one”). There are no major

problems with incorporating such principles into our

model.

On the other hand, these motivational factors are

not independent of each other. Thus, a useful

outcome of an action is in some sense also pleasant

for the subject, punishment is unpleasant (and can be

harmful) for the punished person, etc., but we will

not go into these details in our present model.

3 REASONING IN INTERACTION

As said above, our general goal is to model

reasoning in communication. Our further aim is to

build a conversational agent – a computer program

that can participate in interaction with a human user.

3.1 Human-Human Communication

Both participants, A and B, have their own

communicative goals. In Ex.1, the client’s (A)

communicative goal is to make a decision about a

ski trip. B (a travel agent) is definitely interested in

A’s positive decision. Let us consider how the

communication between A and B continues (Ex. 2).

(2)

B: [neil] on ´väga=ea (.) olemas

näiteks e jaanuarikuuks ´väga=ead

pakkumised Rootsi suusakuurortitesse.

| SPECIFICATION |

there are very good offers to Swedish

holiday resorts in January

A: mhmh | CONTINUER |

hem

B: kus on noh ütleme ´hinnad on innad

on niuksed ´tõeliselt (.) ütleme teevad

Soomele ´ka (.) $ silmad ´ette. $

| ACCOUNT |

the prices are really let’s say much

better than in Finland

A: ah=nii.= | CHANGE OF STATE |

I see

B: =ja ´majutus on väga ´korralik |

SPECIFICATION |

and accommodation is very descent

/---/

et sis kui te saate nagu selle ´tunde

kätte et siis juba ´siis juba minna

´Austriasse. | OPINION |

and after you’ve got the feel of it

then you can go to Austria

A: mhmh | LIMITED ACCEPT |

I see

B has understood that a high price is a problem for

A. Now she takes the initiative and offers another,

cheaper trip to Sweden, trying thus to influence the

client to make a positive decision about another

action. What she can do is to stress the pleasant

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

212

aspects of D (i.e. to entice), or the usefulness of D

(i.e. to persuade), or to threaten B with a

punishment, if he does not do D (threatening is

excluded in the current situation). We will say that B

applies a communicative strategy which can be

realized by using different communicative tactics

(enticement, persuasion and threatening,

accordingly). The dialogue will continue until B

reaches her goal (i.e. the decision of A to perform

the action) or gives up.

3.2 Communicative Goal and

Communicative Strategy

In our approach the communicative strategy is

formalized as an algorithm that a participant applies

to achieve his/her communicative goal.

Two kinds of strategies are important in our case:

attack and defence. In the first case, a participant

tries to press his/her communicative goal onto the

partner. In the second case, s/he averts taking over

the partner’s goal. In the situation under

consideration, the communicative strategy used by B

(attack) can be presented as the algorithm in Fig. 2.

1) Choose the communicative tactics.

2) Implement the tactics to generate a

turn (inform the partner of the

communicative goal – to do D).

3) Did the partner agree to do D? If yes

then finish (the communicative goal has

been reached).

4) Give up? If yes then finish (the

communicative goal has not been reached).

5) Change the communicative tactics? If

yes then choose the new tactics.

6) Implement the tactics to generate a

turn. Go to step 3.

Figure 2: Communicative strategy.

3.3 Communicative Tactics

The conversational agent we are modelling performs

the role of B. In our model there are three different

communicative tactics that B can use as part of its

communicative strategy: enticement, persuasion and

threatening. Each of the communicative tactics

constitutes a procedure for compiling a turn in the

ongoing dialogue: the tactic of enticement consists

in increasing A’s wish to do D; persuasion consists

in increasing A’s belief of the usefulness of D for

him, and threatening consists in increasing A’s

understanding that he must do D.

Communicative tactics are directly related to the

reasoning processes of partner A. For instance, if B

is applying the tactic of enticement it should be able

to imagine the reasoning process of A that is

triggered by the input factor wish. When A at a

certain stage refuses to do D, then B should be able

to guess at which point the reasoning of A went into

the “negative” branch ( “do not do D”), in order to

adequately construct its reactive turn. Similarly, the

tactic of persuasion is related to the reasoning

process triggered by the needed-factor, and the

threatening is related to the reasoning process

triggered by the must-factor. Therefore, in order to

model various communicative tactics, the reasoning

model is used.

3.4 Model of Enticement

When implementing a communicative strategy the

agent B uses a model of the motivational sphere of

partner A – a vector w

– which includes its idea

about weights of the aspects of action D. The more B

knows about A the more similar the vector w

with the actual vector w

of the motivational sphere

of partner A. Here we assume that B has several sets

of statements for increasing/decreasing the weights

of the different aspects of D for partner A. All the

statements have their (numerical) weights as well

(Koit, 2011).

As an illustration, we shortly describe the tactic

of enticement that is based on the reasoning

procedure WISH (Fig. 1). The general idea

underlying this tactic is that B presents A with

statements for pleasantness of D trying to keep the

weight of pleasantness for A high enough and the

values of negative aspects brought out by A low

enough so that weighing positive and negative

aspects would lead A to the decision to perform D.

We suppose that A has a set of statements for

indicating the aspect which weight caused his

rejection. Here we assume that B when enticing uses

each statement only once.

3.5 Discussion

The communicative tactics used by the participants

are not limited to the three that were mentioned

above. Firstly, while influencing the reasoning

process of the partner a participant may repeatedly

use the same argument in order to change a specific

weight in the partner model. Secondly,

communicative tactics need to be specified for A too.

Two different scenarios are possible: (1) A and B

have opposite communicative goals (A does not plan

to do D but B’s goal is to get him do it); (2) both A

and B have the same goal and cooperate with each

other when looking for arguments that support

FromStudyofHuman-humanDialoguestoReasoningModel-ConversationalAgentinArgumentationDialogue

213

achieving it. However, the issue of specifying the

communicative tactics of A will be left for the future

work.

4 CONVERSATIONAL AGENT

We have implemented the described models as a

conversational agent – a DS which interacts with a

user in Estonian and tries to achieve the user’s

decision to perform an action by influencing his

reasoning about the action.

4.1 Architecture

The DS we have been developing consists of the

following functional blocks: natural language

understanding and generation modules, a planner, a

dialogue manager, and a problem solver. The

problem solver enables the system to “tune in” to a

specific problem domain. The other blocks form a

basic interaction system (Jurafsky and Martin,

2008).

The modules use a knowledge base where

different knowledge is kept: linguistic knowledge,

also knowledge about the world (in our case –

frames of actions), communication (communicative

strategy, communicative tactics, dialogue acts), and

users (partners’ models, reasoning procedures).

The natural language understanding module

analyzes the utterances of an input turn, and outputs

their representations as the corresponding

recognized dialogue acts (question, answer, request,

etc.). The task of the planner is to construct a turn of

the DS, either as a response to the user’s turn or as a

turn initiated by the DS itself. In this process, the

planner contacts the problem solver (if a domain

problem has to be solved) and the dialogue manager

to determine the communicative structure of the turn

(dialogue acts). The natural language generation

module will compose the semantic structure

underlying the planned output and transform it into a

linguistic expression.

It is the task of the dialogue manager to

determine how to proceed if the communicative goal

of the DS (to achieve the user’s decision to perform

an action) has not been attained by the preceding

turns. The reasoning model plays a crucial role in

this: the dialogue manager has to decide where the

reasoning of the partner went into a “negative”

branch and try to find new material which is

expected to lead to a positive outcome.

When comparing our reasoning model with BDI

model, then beliefs are represented by knowledge of

the conversational agent with reliability less than 1;

desires are generated by the vector of weights w; and

intentions correspond to goals in goal base. In

addition to desires, from the weights vector we also

can derive some parameters of the motivational

sphere that are not explicitly conveyed by the basic

BDI model: needs, obligations and prohibitions.

4.2 Implementation

Presently, we have implemented a program which

can play the role of B in a simple communication

situation where the goal of B is that A (user) decides

to perform action D. At the moment, the computer

operates only with the semantic representations of

the linguistic input/output. In the current

implementation, ready-made Estonian sentences

(texts) are used both by the computer and the user.

The sentences are classified according to their

function, e.g. for increasing the weight of

pleasantness, for decreasing the weight of

harmfulness, for expressing that pleasantness is too

low, etc. Every sentence has a numerical weight 1.

The work on a linguistic processor is in progress.

At the beginning of the dialogue the computer

expresses the communicative goal (this is its first

turn r

). If the user refuses to do D (after

implementing normal human reasoning which we

are trying to model here), based on the response (r

)

the computer determines the aspect of D the weight

of which does not match the reality and changes this

weight so that the new model will give a negative

result as before but it is an extreme case: if we

increased this weight by one unit (in case of positive

aspects of D) or decreased it (in case of negative

ones) we should get a positive decision. The

computer chooses its response r

from the set of

sentences for increasing/decreasing this weight and

at the same time it increases/decreases this weight in

the partner model by the value of the chosen

sentence. A reasoning procedure based on the new

model will yield a positive decision. Now the user

must choose his response and the process can

continue in a similar way.

4.3 Discussion

A dialogue is generated jointly by the computer and

a user. The computer uses its communicative

strategy and tactics. Let us suppose that after the

computer’s proposal to perform an action D, the user

will create a model of himself, i.e. he will attribute

values to all aspects of D and will reason on the

basis of this model. Of course, creating this model is

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

214

implicit, e.g. the user assesses that doing D would be

more unpleasant than pleasant. By implementing its

communicative strategy and tactics, the computer

has to try to influence the partner model in the way

that would cause the partner to make a positive

decision based on the changed model. The problem

is that the computer does not “know” the real

weights attributed to different aspects of D by the

user. It can only guess these values based on the

user’s negative responses.

At the beginning of a dialogue the computer

randomly generates a user model. At the moment we

have set only one restriction: we require that the

initial model should satisfy the assumption(s) that

underlie the corresponding reasoning procedure.

Thus, for enticing w(pleasant) > w(unpleasant), for

persuading w(useful) > w(harmful) and for

threatening w(obligatory) = 1. When an initial model

is generated the computer uses it as a partner model

and informs the user about its communicative goal.

It chooses a sentence (r

) from a special file of

computer sentences. A user can choose his sentences

(i=1,2,...) from a special file of user sentences, i.e.

he can “play a role” but cannot use unrestricted

texts. If a user has chosen a sentence of refusal, the

computer decides that the user model is inexact and

needs amending. The corresponding class of user

sentences of refusal will be recognized and the

aspect of D determined the weight of which in the

user model was either too small or too great, which

brought about the false decision by the computer.

Based on a valid reasoning procedure (and tactics) a

new value will be computed for this weight, which is

congruent with the negative decision (as explicated

by the user expression).

Our research has a practical aim: to implement a

communication trainer, a computer program that

would allow the user to exercise his abilities to reach

certain communicative goals: (a) getting the partner

to decide to perform an action, or (b) on the

contrary, opposing the partner (Koit, 2012).

5 FUTURE WORK

We have examined here only a very restricted type

of dialogues where the user must play a particular

rigid role. In the future we plan to model such

situations where the computer will take the

participant A’s role. In order to do that, A’s strategies

and tactics need to be modeled.

One of our priorities will be to investigate the

possibilities of adding contextual aspects to the

reasoning model. One option is to include the

personal background of the participants, e.g. by

elaborating the notion of communicative space

(Brown and Levinson, 1999). In our case, the

communicative space is determined by a number of

coordinates, such as social distance between the

partners (far between adversaries, close between

friends), intensity of communication (peaceful,

vehement), etc. Without taking this information into

account, formal reasoning about some action can

easily run into problems such as inconsistency, due

to considering the knowledge in a wrong context,

inefficiency, when irrelevant knowledge is being

considered, or incompleteness, when the relevant

inferences are not made.

ACKNOWLEDGEMENTS

This work is supported by the European Regional

Development Fund through the Estonian Centre of

Excellence in Computer Science (EXCS), the

Estonian Research Council (grant ETF9124), and the

Estonian Ministry of Education and Research (grant

SF0180078s08).

REFERENCES

Brown, P. and Levinson, S. C., 1999. Politeness: Some

universals in language usage. In The discourse reader.

London: Routledge, 321–335.

D´Andrade, R., 1987. A Folk Model of the Mind. In

Cultural Models of Language and Though. London:

Cambridge University Press, 112–148.

Davies, M. and Stone, T., 1995. Folk psychology: the

theory of mind debate. Oxford, Cambridge,

Massachusetts: Blackwell.

Ginzburg, J. and Fernández, R., 2010. Computational

Models of Dialogue. In The Handbook of

Computational Linguistics and Natural Language

Processing, 429–481. Wiley Blackwell Publishers.

Hutchby, I. and Wooffitt, R., 1998. Conversation Analysis.

Principles, Practices and Applications. Cambridge,

UK: Polity Press.

Jokinen, K., 2009. Constructive Dialogue Modelling:

Speech Interaction and Rational Agents. John Wiley

& Sons Ltd.

Jurafsky, D. and Martin, J. H., 2008. Speech and

Language Processing: An Introduction to Natural

Language Processing, Computational Linguistics, and

Speech Recognition. Prentice Hall.

Koit, M., 2012. Developing Software for Training

Argumentation Skills. In Proc. of CMNA 2012,

Workshop of ECAI 2012, 11–15. Montpellier, France.

http://www.cmna.info/

Koit, M., 2011. Conversational Agent in Argumentation:

FromStudyofHuman-humanDialoguestoReasoningModel-ConversationalAgentinArgumentationDialogue

215

Updating of Information States. In Proc. of KEOD

2011, ,375–378. Paris: SciTEC Publications Ltd.

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

Agreement Negotiation Process: A Model that

Involves Natural Reasoning. In Proc. of the Workshop

CMNA. 16th European Conference on Artificial

Intelligence, 53–56. Valencia, Spain.

http://www.cmna.info/

Lester, J. Branting, K. and Mott, B. 2004. Conversational.

Agents. In The Practical Handbook of Internet

Computing. Chapman & Hall.

Õim, H. 1996. Naïve Theories and Communicative

Competence: Reasoning in Communication. In

Estonian in the Changing World, 211–231. University

of Tartu.

Rao, A. S. and Georgeff, M. P. 1991. Modeling Rational

Agents within a BDI-Architecture. In Proc. of the 2nd

International Conference on Principles of Knowledge

Representation and Reasoning, 473–484. Morgan

Kaufmann.

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

216