Developing a Model of Agreement Negotiation Dialogue

Mare Koit

Institute of Computer Science, University of Tartu, Liivi 2, Tartu, Estonia

Keywords: Dialogue Structure, Dialogue Act, Dialogue Corpus, Knowledge Representation.

Abstract: We are investigating human-human dialogues in the Estonian dialogue corpus with the further aim to

develop a dialogue system which carries out negotiations with a user in a natural language. Two sub-corpora

are analysed and compared: (1) MSN conversations, and (2) everyday dialogues, both phone calls and face-

to-face conversations. In the dialogues, the participants are trying to achieve an agreement about doing an

action. The structure of negotiations is represented as a sequence of dialogue acts. A special case of

negotiation – debate where the participants have contradictory communicative goals – has been

implemented as an experimental dialogue system.

1 INTRODUCTION

Negotiation is a dialogue between two or more

people or parties, intended to reach an understan-

ding, resolve point of difference, or gain advantage

in outcome of dialogue. Collaborative and adversa-

rial negotiations are different. In a discussion, the

parties hold points of view, but are potentially open

to learning from alternative perspectives. Debates,

on the other hand, are marked by an adversarial

approach where each party comes equipped to

promote their position and to undermine that of the

other side (Cummins, 2011).

Rahwan et al (2004) discuss three approaches to

automated negotiation: game-theoretic, heuristic-

based and argumentation-based. The last approach to

negotiation allows agents to argue about their beliefs

and other mental attitudes during the negotiation

process.

Besnard and Hunter (2008) formalize argumenta-

tion by using classical logic. Logical models of

argument support decision making by participants,

guide negotiation and allow reach agreements

(Amgoud and Cayrol, 2002).

Hadjinikolis et al. (2012) provide an

argumentation-based framework for persuasion

dialogues, using a logical conception of arguments

that an agent may undertake in a dialogue game,

based on its model of its opponents.

Amgoud et al. (2015) introduce a target language

for representing arguments mined from natural

language. They propose a formal language (RC

language) for representing reasons and claims of

arguments, and a framework for reasoning about

arguments.

Overviews of the state of art in modelling

agreement negotiation can be found in (Amgoud et

al., 2015), (Besnard and Hunter, 2008), and

(Chesňevar et al., 2000).

We are studying the interactions in natural

language between two participants (A and B) about

doing an action, e.g. fixing an appointment. The

communicative goals of the participants can

coincide or be different. The participants are

presenting arguments and counterarguments during a

dialogue. They can also ask and answer questions in

order to make choices among the arguments for

averting the partner’s counterarguments.

If A and B have contradictory goals when starting

interaction then they are involved into debate. One

participant will achieve his or her communicative

goal (wins debate) and another has to abandon her or

his initial goal (loses debate).

If A and B have a common communicative goal

then they are cooperatively looking for arguments

that support achieving this goal. Still, one of them

can indicate to obstacles which do not allow achieve

the goal. Then the partner has to find arguments for

showing how the obstacles can be exceeded. The

final result of negotiation is whether achieving the

collective goal (win-win model) or its withdrawal if

some of the obstacles cannot be eliminated (lose-

lose model).

We have worked out a dialogue model which

includes a reasoning model as its part and

implemented it in an experimental dialogue system

Koit, M.

Developing a Model of Agreement Negotiation Dialogue.

DOI: 10.5220/0006069201570162

In Proceedings of the 8th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2016) - Volume 2: KEOD, pages 157-162

ISBN: 978-989-758-203-5

157

(DS) (Koit and Õim, 2014; Koit, 2015). In the

current paper, we will further develop the model as

based on the analysis of human-human dialogues.

Our main goal here is to explain how people

negotiate. The further aim is to develop our DS.

The rest of the paper is structured as follows.

Section 2 introduces the used dialogue corpus and

gives the results of the corpus analysis. The structure

of human-human agreement negotiation will be

represented by using dialogue acts. Section 3

discusses some questions related to the implementa-

tion of the structure in a DS which interacts with the

user in a natural language and follows norms and

rules of human-human conversation. Conclusions

will be made in Section 4.

2 CORPUS ANALYSIS

2.1 The Dialogue Corpus

Our study is based on the Estonian dialogue corpus

(EDiC) (Koit and Õim, 2014). It includes three

different kinds of human-human dialogues: (1)

recordings and transcripts of human-human spoken

dialogues, (2) written dialogues collected in simulat-

ions by Wizard-of-Oz method, and (3) (written)

MSN conversations. The corpus also includes log

files of interactions with some DSs. The spoken

dialogues are recorded in authentic situations and

transcribed by using the transcription system of

Conversation Analysis (CA) (Sidnell, 2010). There

are both institutional and everyday phone calls as

well as face-to-face conversations in the corpus.

Dialogue acts (DA) are annotated in the corpus

by using a customized typology (Koit, 2015) which

is based on CA. In the typology, the acts are divided

into two groups – adjacency pair (AP) acts where the

first pair part expects a certain second pair part (like

request – grant), and single (non-AP) acts which do

not expect any response (like acknowledgement ah).

Names of the DAs consist of two parts separated by

a colon: (a) the first two letters present an

abbreviation of the name of an act-group, e.g. DI –

DIrectives, VR – Voluntary Reactions. The third

letter is only used for AP acts – the first (F) or the

second (S) pair part of an AP act; (b) the proper

name of the act. There are acts as DIF: Request,

DIS: Giving information, VR: Acknowledgement,

etc. The total number of the acts is 126.

We are using custom-made web-based software

for annotation of dialogues. An utterance can get

more than one DA tag if it is multifunctional (cf.

Example 1: a phone call of friends A and B; ‘|’

separates the DA tags of a multifunctional

utterance).

(1)

A: .hhhhh ´tulge meile ´pühapäeval

´külla. DIF: Proposal

Please come to visit us on Sunday.

B: ´pühapäeval. QUF: Offering answer |

RPF: Checking

On Sunday?

A: mhmh QUS: Yes | RPS: Repair

Yes.

…

B: okei. DIS: Accept

OK.

Another custom-made software tool enables to

calculate some statistics for the dialogues: the counts

of utterances, words, different DAs, frequency of

words and certain sequences of DAs, etc.

Here we will study two different sub-corpora of

EDiC. The first one consists of 40 MSN

conversations, and the second one of 44 everyday

dialogues where an action is negotiated and argued

(among them 22 phone calls and 22 face-to-face

conversations). We believe that MSN dialogues and

everyday phone calls might be a suitable basis for

the development of a DS which interacts with a user

following the rules of human-human conversation.

Face-to-face dialogues will be analysed for compa-

rison. Our aim is to find out the typical structure of

negotiations in the different types of dialogues as

represented by DAs. Further, we attempt to design a

general structure of argumentation-based negotiation

in order to develop our experimental DS.

2.2 Negotiation in MSN Conversations

The analysed 40 MSN conversations include 3313

utterances in total; the average length of a

conversation is 82.8 utterances (min 24, max 193).

The total number of words is 23,943, i.e. the average

length of a conversation is 599 words. Among the

conversations, there are 17 where agreement

negotiation takes place. (In the remaining 23

dialogues, the participants are discussing about their

everyday experience: visiting lectures, parties,

movies, skiing, etc.). The number of utterances in

the agreement negotiation dialogues is 1427 in total

and the number of words is 9367.

In Example 2, the friends A and B are negotiating

a meeting. A asks a question about the meeting time.

B excludes some days of the current week bringing

out the explanations and proposes the next week for

the meeting. The participants do not appoint a

weekday and time but they agree to continue the

negotiation later.

KEOD 2016 - 8th International Conference on Knowledge Engineering and Ontology Development

158

(2)

A: Kuna me kokku võiksime saada? QUF:

h-question

hen do we meet?

B: hmm.. las ma nüüd mõtlen, kuidas mul

see nödal on. QUS: Giving information

Let me think about this week.

neljap peab kaili sünnat, siis ei soboi

AI: Justification

On Thursday, Kaili has her birthday

party, then it is impossible.

…

aga järgmine nädal? QUF: Offering

answer

How about the next week?

A: Järgmine nädal sobiks paremini vast

küll. QUS: Yes

The next week is better.

B: tore VR: Acknowledgement

Fine.

kas lepime kohe mingi aja kokku?

Do we appoint the time just now?

või räägime järgmine nädal QUF:

Alternative question

Or do we discuss the next week?

A: Ei, räägime järgmine nädal. QUS:

Alternative answer: one

No, let’s discuss the next week.

In another conversation, the participants A and B

are discussing about a surprise for mother’s

birthday. A proposes to call broadcasting where

congratuliations will be delivered. B is doubtful but

A succeeds to convince him by the arguments. They

also determine a hit song for the congratulation in

radio.

The general structure of MSN conversation

where an action is negotiated looks like follows (Fig.

1). The winding brackets ’{ ’and ’}’ connect a part

that can be repeated (0 or more times); round

brackets connect a part that can be missed; ’/’

separates alternatives; ’- -’ starts a comment.

AP acts (directives and questions together with

their second pair parts) form an encompassing

structure of negotiation. Arguments are represented

by non-AP acts (primary single act of giving

information PS: Giving information, and additional

information act of justification AI: Justification).

A: DIF: Proposal/ QUF: Wh-question

{

B: (PS: Giving information/ AI: Justification - -

<argument>)

DIF: Request/ QUF: Wh-question/ QUF:

Offering answer

A: DIS/ QUS: Giving information/ QUS: Yes

(PS: Giving information/ AI: Justification

- -<argument>)

}

- - Decision

B: DIS: Accept/ QUS: Giving information/ DIS/

QUS: Deferral

Figure 1: The structure of MSN negotiation (A makes a

proposal to B to do an action). The winding brackets ’{’

and ’}’ connect a part that can be repeated; round brackets

connect a part that can be missed; ’/’ separates

alternatives; ’- -’ starts a comment.

2.3 Negotiation in Everyday Dialogues

The participants of the everyday dialogues are

acquainted or friends and the initiator A makes a

proposal to the partner B to perform an action. In our

analysed dialogues, positive decision has been

achieved in 25 cases and negative in one case (in a

phone call). The remaining dialogues (out of 44)

finish with the postponement of the decision.

Several DAs are used for giving arguments: non-AP

acts like in MSN conversations as well as AP acts.

2.3.1 Phone Calls

The total number of utterances is 1172 in the

analysed 22 phone calls and the average length of a

dialogue is 53.3 utterances. The number of words is

6,412, i.e. the average length of a call is 291 words.

As compared with the MSN conversations, the

phone calls include 2.8 times less utterances and 3.7

times less words in total. That can be surprising

because typing is more time consuming than

speaking. Still, our phone calls are negotiations but

MSN conversations include more chatting than

negotiation. In addition, the number of MSN

conversations is almost two times bigger than the

number of calls. However, if we only take into

account the 17 agreement negotiation MSN

conversations then we again see that the phone calls

are in average shorter than computer-mediated

written conversations. People are more economical

when speaking by phone as compared with MSN

communication.

Examples (1) and (3) are the typical phone calls

where agreement negotiation takes place.

(3)

A: =kule kas sa ´välja ei ´viitsi

tulla=vä. QUF: Yes/no question

Do you come out with me?

(0.3)

…

B: jah QUS: Yes

Yes.

aga (0.5) prä- (.) noh mul läb natuke

Developing a Model of Agreement Negotiation Dialogue

159

´aega onju. AI: Justification

But I need some time before.

…

A: ´millal me=s ´lähme. QUF: Wh-

question

When do we go?

(0.2)

B: noh ´lähme nimodi:: (.) panen

´riidesse. QUS: Giving information

I need to dress myself.

…

A: no poole viiest ´lähme. DIF:

Proposal

Let’s go half past four.

B: jah. DIS: Accept

yes.

The general structure of a phone call where an action

is negotiated is represented in Fig. 2 (cf. Koit, 2015).

{

A: DIF: Proposal/ QUF: Wh-question/ Yes/no

question

(PS: Giving information/ AI: Justification

- -<argument>)

{

B: DIF: Request/ QUF: Wh-question

(PS: Giving information/ AI: Justification

- - <argument>)

A: DIS: Giving information

(PS: Giving information / AI: Justification

- -<argument>)

}

- - Decision

B: DIS: Accept/ Deferral/ Reject/ QUS: Giving

information/ Yes/ No

}

Figure 2: The structure of everyday negotiation (a phone

call: A makes a proposal to B to do an action). The

winding brackets ’{’ and ’}’ connect a part that can be

repeated; round brackets connect a part that can be missed;

’/’ separates alternatives; ’- -’ starts a comment.

As we see, the structures in Fig. 1 and Fig. 2 are

quite similar with the difference that more than one

proposal can be made by the initiator A of dialogue

in the case of a phone call (in our corpus).

2.3.2 Face-to-Face Conversations

The total number of utterances is 2,362 in the

analysed 22 face-to-face conversations and the

average length of a dialogue is 107.4 utterances. The

number of words is 20,653. Therefore, the face-to-

face conversations are approximately two times

longer than phone calls in total and more or less as

long as MSN conversations (but the number of the

analysed MSN conversations is almost two times

bigger than the number of face-to-face dialogues).

That is not surprising because there are no limits

caused by a medium in face-to-face conversations

(neither phone nor computer). Therefore, we can

expect that the typical structure of a face-to-face

conversation is more complicated as compared with

phone calls and MSN conversations.

In Example 4, the companions A and B are

planning to buy an apartment. A problem is that its

cellar needs reconstruction.

(4)

A: mis=selle ´korteriga teeme. QUF: Wh-

question

What do we do with the apartment?

(6.9)

B: mt=.h mulle jättis see ´väga=hea

´mulje. QUS: Other| OPF: Opinion

I like it very much.

(3.3)

hhh=see:: (0.3) g-=eeee ´korteri

´omanik oli see ´naine, (0.8) mt (0.5)

kes=seal:= eeeeeeee (0.5) ´ootas=see

oli: üsna süm´paatne ´inimene [{--} ]

OPF: Opinion

The owner is a sympathetic woman.

A: [noja=aga=ta=i tead]nud sest

´korterist mitte ´midagi=ju. OPS: Other

| OPF: Assertion

Yes, but she seemed to know nothing

about the apartment.

(2.0)

…

B: a=´uvitav kas nendes ´teistes

majades mida: ´pakk- (0.2) noh milles

seal ´kortereid saada ´on kas ´seal on

ka ´nii madalad ´keldrid vä. QUF:

Yes/no question

Btw, do the other apartments we visited

similarly have that low cellars?

…

A: ma=i ´tea. QUS: Other

I don’t know.

…

selles=mõttes=et meil=poleks ´vahet

et=et=e=.hh (0.9) hh ostad ´korteri

´ära ja tellid ´uurimise tuleb

nagu=´ei: või .hhh (0.2) või: t- ostad

´korteri ära=ja hakkad ´ise

tegema=ja=tuleb=´ei: et=et,=h (4.1)

ma=i ´taha ´uskuda=et seal ´korteri=all

kus need ´gaasijuhtmed jooksevad. OPF:

Opinion

No difference for us: we buy the

apartment and then request an

exploration or we do the reconstruction

ourselves. I don’t believe that the gas

trace is going under the apartment.

B: ei ´seda küll. OPS: Accept

No, indeed.

KEOD 2016 - 8th International Conference on Knowledge Engineering and Ontology Development

160

It is significant that most of the face-to-face

conversations include opinions (the adjacency pair

acts OPF: Opinion/ Assertion and OPS: Accept/

Reject). Thus, the arguments are often represented

by AP acts and not by non-AP acts like in the other

types of the analysed dialogues. The typical

structure of a face-to-face dialogue where an action

is negotiated is given in Fig. 3.

A: DIF: PROPOSAL/ QUF: Wh-question/ Yes/no

question

(PS: Giving information/ AI: Justification

- -<argument>)

{

B: (PS: Giving information/ AI: Justification

- - <argument>)

DIF: Request/ QUF: Wh-question

A: DIS: Giving information

(PS: Giving information/ AI: Justification

- -<argument>)

}

B: OPF: Opinion/ Assertion - - <argument>

A: OPS: Accept/ Reject | OPF: Assertion

- - <argument>

{

B: OPS: Accept/ Reject | OPF: Opinion/

Assertion - - <argument>

A: OPS: Accept/ Reject | OPF: Opinion/

Assertion - - <argument>

}

- - Decision

B: DIS: Accept/ Deferral/ Reject

Figure 3: The structure of everyday face-to-face

negotiation (A makes a proposal to B to do an action). The

winding brackets ’{’ and ’}’ connect a part that can be

repeated; round brackets connect a part that can be missed;

’/’ separates alternatives; ‘|’ separates the DA tags of a

multifunctional utterance; ’- -’ starts a comment.

The initiator A when attempting to convince B to

perform an action presents his arguments by using

non-AP acts (PS: Giving information, AI:

Justification/ Specification/ Explanation) and does

not expect the reaction of the partner. If the partner

B is antagonistic then she takes over the initiative

and presents her arguments as the first pair parts of

the opinion AP (OPF: Opinion/ Assertion) expecting

A’s reaction. Such a reaction is different as

compared with phone calls where the participants

mostly are collaborative.

3 DISCUSSION

The corpus analysis demonstrates that the structure

of everyday face-to-face conversations is more

complicated than in the case of phone calls and

MSN conversations. The reason is that the partici-

pants have more freedom when communicating

face-to-face. On the other hand, both typing (MSN

conversations) and calling (phone calls) set some

limits to the participants as caused by the communi-

cation mode.

As said before, our typology of DAs is based on

Conversation Analysis. The corpus analysis has

demonstrated that single acts (PS: Giving

information, AI: Justification, etc.) are preferred for

representing arguments both in MSN and phone

conversations. Still, the participants have a common

communicative goal in most of the conversations.

On the other hand, when communicating face-to-

face, the participants often have different communi-

cative goals. They express their opinions which

expect reaction of the other side, i.e. they use AP

acts OPF: Opinion/ Assertion and OPS: Accept/

Reject, etc.). For example, there are 723 OPF acts in

total in the analysed face-to-face conversations but

only 82 OPF acts in the everyday phone calls. If the

participants are collaborative and one of them

expresses an opinion then both they usually accept

arguments of each other (using DA ’accept’). If they

are confrontational then at least one of them rejects

the arguments of the partner (using DA ’reject’) and

presents his or her own counterarguments.

Taking into account the results of the analysis,

we can represent the general structure of

argumentation-based negotiation where performing

an action is discussed as a sequence of (generalized)

DAs in Fig.4.

An experimental dialogue system has been

created which carries out (written) debates with a

user about performing an action and presents

counterarguments to the arguments given by the user

(Koit, 2015). Both the computer and the user can

choose their arguments from given sets of utterances

in natural language. The structure of debate

implemented in our DS is presented in Fig. 5.

A: proposal

{

B: question

A: giving-information

}

{

B: assertion/ justification/ giving-information

- - argument

A: accept/ justification/ giving-information/

reject

(assertion)

- - argument

Developing a Model of Agreement Negotiation Dialogue

161

}

B: accept/ deferral/ reject

Figure 4: The structure of negotiation (A makes a proposal

to B to do an action). The winding brackets ’{’ and ’}’

connect a part that can be repeated; round brackets

connect a part that can be missed; ’/’ separates

alternatives;’- -’ starts a comment.

The study demonstrates that the actual human-

human negotiations have more complicated structure

than in our implementation. Therefore, our DS has

to be developed.

A: proposal

assertion - - argument

{

B: reject | assertion - - argument

A: reject | assertion - - argument

assertion - - argument

}

B: accept/ reject

Figure 5: The structure of debate implemented in our DS

(A makes a proposal to B to do an action). The winding

brackets ’{’ and ’}’ connect a part that can be repeated; ‘|’

separates the different functions of a multifunctional

utterance; ’- -’ starts a comment.

4 CONCLUSION AND FUTURE

WORK

We analyse two kinds of human-human negotiation

dialogues with the aim to find out the general

structure of agreement negotiation. We use dialogue

acts in order to represent the structure. This structure

can be taken as a basis of the dialogue manager

when developing DS which interacts with the user in

a natural language following rules of human-human

communication. We believe that such a DS can help

to train the user’s argumentation skills.

We have implemented on the computer a simple

argumentation-based debate where A’s

communicative goal is “B will do an action D” but

B’s goal is “do not do D”. Our further aim is to

develop our DS by implementing the results of the

study.

ACKNOWLEDGEMENTS

This work was supported by the Estonian Ministry

of Education and Research (IUT20-56), and by the

European Union through the European Regional

Development Fund (Centre of Excellence in

Estonian Studies).

REFERENCES

Amgoud, L., Besnard, P. Hunter, A., 2015. Logical

Representation and Analysis for RC-Arguments. In

Proc. of ICTAI, 8 p. https://www.irit.fr/~Leila.Am-

goud/ictai2015-1.pdf

Amgoud, L., Cayrol, C., 2002. A Reasoning Model Based

on the Production of Acceptable Arguments. In Ann.

Math. Artif. Intell. 34(1-3), 197-215.

Besnard, P., Hunter, A., 2008. Elements of Argumentation,

MIT Press, Cambridge, MA.

Chesñevar, C., Maguitman, A., Loui, R., 2000. Logical

models of argument. In ACM Computing Surveys, vol.

32(4), 337–383.

Cummins, T., 2011. Negotiation: Discussion or Debate?

https://commitmentmatters.com/2011/11/29/negotiati-

on-discussion-or-debate/

Hadjinikolis, C., Modgil, S., Black, E., McBurney, P.,

Luck, M., 2012. Investigating Strategic Considerations

in Persuasion Dialogue Games. In STAIRS, 137–148.

Koit, M., 2015. Analysing Human-Human Negotiations

with the Aim to Develop a Dialogue System. In

Proc.of 17th International Conference SPECOM

2015, LNAI, 9319, pp. 73−80, Springer-Verlag.

Koit, M., Õim, H., 2014. A Computational Model of

Argumentation in Agreement Negotiation Processes.

In Argument & Computation, 5 (2-3), pp. 209–236,

Taylor & Francis Online. DOI: 10.1080/19462166.-

2014.915233

Rahwan, I., Ramchurn, S.D., Jennings, N.R., Mcburney,

P., Parsons, S., Sonenberg, L., 2004. Argumentation-

based negotiation. In The Knowledge Engineering

Review, vol. 18(4), 343–375, Cambridge University

Press, DOI: 10.1017/S0269888904000098

Sidnell, J., 2010. Conversation Analysis: An Introduction,

London: Wiley-Blackwell.

KEOD 2016 - 8th International Conference on Knowledge Engineering and Ontology Development

162