Modelling Attitudes of a Conversational Agent
Mare Koit
a
Institute of Computer Science, University of Tartu, J. Liivi 2, Tartu, Estonia
Keywords: Conversational Agent, Negotiation, Reasoning Model, Attitude, Knowledge Representation.
Abstract: The paper introduces a work in progress on modelling of attitudes of a conversational agent in negotiation.
Two kinds of attitudes are under consideration: (1) related to different aspects of a negotiation object (in our
case, doing an action) which direct reasoning about the action, and (2) related to a communication partner
(dominance, collaboration, communicative distance, etc.) which are modelled by using the concept of
multidimensional social space. Attitudes of participants have been annotated in a small sub-corpus of the
Estonian dialogue corpus. An example from the sub-corpus is presented in order to illustrate how the models
describe the change of attitudes of human participants. A limited version of the model of a conversational
agent is implemented on the computer. Our further aim is to develop a dialogue system to train the user’s
negotiation skills by interacting with him in a natural language.
1 INTRODUCTION
Modelling of conversational agents and development
of dialogue systems (DS) is aimed to make interaction
of human users with the computer more convenient.
Conversational agents communicate with users in
natural language making travel arrangements,
answering questions about weather or sports, routing
telephone calls, acting as a general telephone
assistant, or performing even more sophisticated tasks
(Jurafsky and Martin, 2013).
We are studying the dialogues where one
participant (A) requests her partner (B) to do an action
D and proposes several arguments for doing D, trying
to influence B. The paper introduces a model of
conversational agent, concentrating on the attitudes of
the agent when it is involved into negotiation with a
user or with another conversational agent. Two kinds
of attitudes are under consideration: (1) related to
different aspects of a negotiation object (in our case,
it is doing an action), and (2) related to a
communication partner, or social attitudes.
Our aim is here to justify some aspects of the
model on actual human-human dialogues in order to
include these aspects into the DS that interacts with
the user in the Estonian language. We currently limit
us with verbal interaction and do not consider
nonverbal means.
a
https://orcid.org/0000-0002-7318-087X
The paper is structured as follows. Section 2
describes the related work. Section 3 introduces a
model of conversational agent, including
representation of information states that the agent
passes during negotiation as well as the participants’
attitudes that are changing in negotiation. An
authentic human-human negotiation is analyzed in
Section 4 in order to demonstrate how well work the
models of attitudes. Section 5 presents an
implementation and Section 6 discusses how the
model of conversational agent can be used when
developing a DS. Section 7 draws conclusions.
2 RELATED WORK
Negotiation is simultaneously a linguistic and a
reasoning problem, in which intent must be formulated
and then verbally realized. Such dialogues require
agents to understand, plan, and generate utterances to
achieve their goals (Traum et al., 2008; Lewis et al.,
2017). Automated negotiation agents capable of
negotiating efficiently with people must rely on a good
opponent modeling component to model their
counterpart, adapt their behavior to their partner,
influencing the partner’s opinions and attitudes (Oshrat
et al., 2009). There are several approaches to model
change of a person’s attitude, incl. the Elaboration
Koit, M.
Modelling Attitudes of a Conversational Agent.
DOI: 10.5220/0008067102250232
In Proceedings of the 11th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2019), pages 225-232
ISBN: 978-989-758-382-7
Copyright
c
 2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
225

Likelihood Model (ELM) that is a theory of thinking
process, Social Judgment Theory that emphasizes the
distance in opinions, and Social Impact Theory which
emphasizes the number, strength and immediacy of the
people trying to influence a person to change its mind.
A naïve, intuitive model of functioning of the human
mind reflects knowledge that human beings have and
use about their partners in everyday communication
(Õim, 1996).
For virtual agents, the expressions of attitudes in
groups is a key element to improve the social
believability of the virtual worlds that they populate as
well as the user’s experience, for example in
entertainment or training applications (Ravenet et al.,
2015; DeVaultand et al., 2015; Gratch et al., 2015;
Callejas et al., 2014).
Concept of social attitude or interpersonal stance in
interaction (being polite, distant, cold, warm,
supportive, contemptuous, etc.) is considered in
(Carofiglio, 2009). Ravenet et al. (2012) show the
influence of dominance and liking in the nonverbal
behavior depending on the gender of the speaker.
These attitudes can be conveyed by words and voice
features but also by nonverbal means – facial
expression, body movement, and gestures (Knapp and
Hall, 2009).
Computational approaches to dialogue fall into two
categories of computational task: dialogue modelling
and dialogue management. A dialogue system will
have both dialogue modeling and dialogue
management components (Traum, 2017). The
functions of the dialogue manager can be formalized in
terms of information state update. An information state
includes beliefs, assumptions, expectations, goals,
preferences and other attitudes of a dialogue
participant’s that may influence the participant’s
interpretation and generation of communicative
behavior (Bunt, 2014).
An interesting and useful kind of dialogue systems
rapidly developing in the last years are embodied
conversational agents (Harthold et al., 2013; Ravenet
et al., 2015; Dermouche, 2016; Jokinen, 2018). Such
agents are interacting with human users in a natural
language, and they are designed to play a certain social
role in interaction.
3 CONVERSATIONAL AGENT
AND INFORMATION STATE
UPDATES
Let us consider interaction between a conversational
agent A and its partner B (which can be whether
another conversational agent or a human user). A
initializes the interaction by requesting B to do an
action D. The process is determined if the following
is given (cf. Koit, 2018):
1) set G of communicative goals where both
participants choose their own initial goals (G
A
and G
B
,
respectively). In our case, G
A
= “B makes a decision
to do D“
2) set S of communicative strategies of the
participants. A communicative strategy is an
algorithm used by a participant for achieving his/her
communicative goal. This algorithm determines the
activity of a participant at each communicative step
3) set T of communicative tactics, i.e. methods of
influencing the partner when applying a
communicative strategy. For example, A can entice,
persuade, or threaten B in order to achieve its goal G
A
4) set R of reasoning models which is used by
participants when reasoning about doing the action D.
A reasoning model is an algorithm the result of which
is a positive or negative decision about the reasoning
object (in our case, the action D)
5) set P of participant models, i.e. a participant’s
depiction of the attitudes himself/herself and his/her
partner in relation to the reasoning object:
P = {P
A
(A), P
A
(B), P
B
(A), P
B
(B)}
6) set of world knowledge
7) set of linguistic knowledge.
A conversational agent passes several information
states during interaction starting from initial state and
going to following states by applying update rules.
Information state represents cumulative additions
from previous actions in the dialogue, motivating
future actions. There are two parts of an information
state of a conversational agent (Traum and Larsson,
2003) – private (information accessible only for the
agent) and shared (accessible for both participants).
The private part of an information state of the
conversational agent A consists of the following
information: (a) current partner model and social
attitudes in relation to the partner, (b) communicative
tactics t
i
A
which A has chosen for influencing B, (c)
the reasoning model r
j
which A is trying to trigger in
B and bring it to the positive decision (it is determined
by the chosen tactics, e.g. when enticing, A tries to
increase B’s wish to do D), (d) set of dialogue acts
DA={d
1
A
, d
2
A
, …, d
n
A
} which A can use, (e) set of
utterances for increasing or decreasing the values of
B’s attitudes in relation to D (arguments for/against
of doing D) U={u
i1
A
, u
i2
A
, …, u
iki
A
}.
The shared part of an information state contains
(a) set of reasoning models R={r
1
,…,r
k
}, (b) set of
communicative tactics T={t
1
, t
2
, …, t
p
}, and (c)
KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development
226

dialogue history p
1
:u
1
[d
1
], p
2
:u
2
[d
2
],…, p
i
:u
i
[d
i
]
where p
1
=A, p
2
, etc. are A or B.
There are two categories of update rules that will
be used by a conversational agent for moving from
the current information state into the next one: (1) the
rules used by the agent in order to interpret the
partner’s turns and (2) the rules used in order to
generate its own turns.
3.1 Reasoning Model and Attitudes
related to Conversation Object
The reasoning process of a subject about doing an
action D consists of steps where the resources,
positive and negative aspects of D will be weighed. A
communication partner can take part in this process
only implicitly by presenting arguments to stress the
positive aspects of D and downgrade the negative
ones.
Our used reasoning model includes two parts: (1)
a model of (human) motivational sphere that
represents the attitudes of a reasoning subject in
relation to the aspects of the action under
consideration, and (2) reasoning procedures. It is a
kind of BDI (Belief-Desire-Intention) model
(Bratman, 1999).
We represent the model of motivational sphere of
a communication participant as a vector with
numerical coordinates that express the attitudes of the
participant in relation to different aspects of the action
D (Koit and Õim, 2014):
w
D
= (w(resources
D
), w(pleasant
D
), w(un¬ple¬a-
sant
D
), w(useful
D
), w(harm¬ful
D
), w(ob¬li-
ga¬tory
D
), w(punishme¬nt-not
D
), w(pro¬hi¬bi-
ted
D
), w(pu¬nish¬ment-do
D
)).
Here w(pleasant
D
), w(unpleasant
D
), etc, indicate
the pleasantness, unpleasantness, etc. of D or its
consequences; w(punishment-do
D
) is the punishment
for doing a prohibited action and w(punishment-not
D
)
– the punishment for not doing an obligatory action.
The value of w(resources
D
) is 1 if the reasoning
subject has all the resources needed for doing D (or 0
if some of the resources are missing), w(obligatory
D
)
is 1 if the action is obligatory for the subject
(otherwise 0), w(prohibited
D
) is 1 if the action is
prohibited (otherwise 0). The values of the other
coordinates can be numbers on the scale from 0 to 10.
The model of motivational sphere is used by the
subject when reasoning about doing D.
The reasoning itself depends on the determinant
which triggers it. With respect to a naïve theory, there
are three kinds of determinants that can cause humans
to reason about an action D (Õim, 1996): his/her wish,
need and obligation. Therefore, three different
prototypical reasoning procedures can be described.
Every procedure consists of steps passed by a
reasoning subject and it finishes with a decision: do
D or not. When reasoning in order to make a decision,
B considers his resources as well as different positive
and negative aspects of doing D. If the positive
aspects (pleasantness, etc.) weigh more than negative
(unpleasantness, etc.) then the decision will be “do D”
otherwise “do not do D”.
Presumption: w(pleasant)  w(unpleasant).
1) Is w(resources) = 1? If not then go to 11.
2) Is w(pleasant) > w(unpleasant) + w(harmful)? If not then go
to 6.
3) Is w(prohibited) = 1? If not then go to 10.
4) Is w(pleasant) > w(unpleasant) + w(harmful) +
w(punishment-do)? If yes then go to 10.
5) Is w(pleasant) + w(useful) > w(unpleasant)+w(harmful) +
w(punishment-do)? If yes then go to 10 else go to 11.
6) Is w(pleasant) + w(useful)  w(unpleasant) + w(harmful)? If
not then go to 9.
7) Is w(obligatory) = 1? If not then go to 11.
8) Is w(pleasant) + w(useful) + w(punishment-not) >
w(unpleasant) + w(harmful)? If yes then go to 10 else go to 11.
9) Is w(prohibited) = 1? If yes then go to 5. else go to 10.
10) Decide: do D. End.
11) Decide: do not do D.
Figure 1: Reasoning procedure WISH.
Let us present two reasoning procedures (WISH
and NEEDED) which will be used in the example in
the following section. One procedure is triggered by
the wish and the other by the need of the reasoning
subject to do the action D. Both procedures are
presented as step-form algorithms in Fig.1 and 2,
respectively. (We do not indicate here the action D.)
Presumption: w(useful)  w(harmful).
1) Is w(resources) = 1? If not then go to 8.
2) Is w(pleasant) > w(unpleasant)? If not then go to 5.
3) Is w(prohibited) = 1? If not then go to 7.
4) Is w(pleasant) + w(useful) > w(unpleasant) + w(harmful) +
w(punishment-do)? If yes then go to 7 otherwise go to 8.
5) Is w(obligatory) = 1? If not then go to 8.
6) Is w(pleasant) + w(useful) + w(punishment-not) >
w(unpleasant) + w(harmful)? if not then go to 8.
7) Decide: do D. End.
8) Decide: do not do D.
Figure 2: Reasoning procedure NEEDED.
We use two vectors (w
B
D
and w
AB
D
) which capture
the attitudes of communication participants in
relation to the action D under consideration. Here w
B
D
is the model of motivational sphere of B who has to
make a decision about doing D; the vector includes
B’s (actual) evaluations (attitudes) of D’s aspects, it
is used by B when reasoning about doing D. The other
vector w
AB
D
is the partner model that includes A’s
Modelling Attitudes of a Conversational Agent
227

beliefs concerning B’s attitudes, it is used by A when
planning the next turn in dialogue. We suppose that A
has some preliminary information about B in order to
compose the initial partner model before making the
proposal to do D. Both the models will change as
influenced by the arguments presented by the
participants in negotiation. For example, every
argument presented by A targeting the pleasantness of
D should increase the corresponding values of
w
B
D
(pleasant) as well as w
AB
D
(pleasant
D
).
3.2 Social Attitudes related to
Communication Partner
In order to model the attitudes of a participant in
relation to a communication partner we use the
dimensions that characterize the relationships of
participants in a communicative encounter.
Communication can be collaborative or
confrontational, personal or impersonal; it can also be
characterized by the social distance of participants
(near, far), etc. People have an intuitive, naïve theory
of these dimensions; the values can be expressed by
specific words. Still, we use numerical values as
approximations to the words in our model (like in the
model of human motivational sphere in the previous
subsection). The following dimensions of such a
‘social space’ can be specified:
1. Dominance (on the scale from dominant to
submissive)
2. Communicative distance to the partner (from
near to far)
3. Cooperation (from collaborative to
confrontational)
4. Politeness (from polite to impolite)
5. Personality (from personal to impersonal)
6. Modality (from friendly to hostile)
7. Intensity (from indolent to vehement).
We use the numbers 1, 0 and -1 for the values of the
coordinates of social space. For example, the value of
communicative distance is -1 if the person feels
closeness in relation to his/her communication
partner and 1 if he/she is far from the partner. The
value 1 on the scale of modality means friendly and
the value -1 means hostile interaction. On any scale,
0 is the neutral value. Still, a bigger number of values
than three can be considered on every scale. It is also
possible to use continuous scales instead of discrete
values.
The attitudes of participants in relation to the
partner can be represented by the (7-dimensional)
vectors s
AB
and s
BA
, respectively which determine the
‘points’ in social space. The participants can be
located in different points. For example, customers
sometimes angrily communicate with a service man
who, as an official person, has to remain neutral or
even friendly. Moreover, the participants can also
‘move’ from one point to another during
communication. For instance, the participants who
were on confrontational positions at the outset can
reach the collaborative one at the end.
4 ANALYSIS OF
HUMAN-HUMAN
NEGOTIATION
In order to evaluate the model of conversational agent
and especially, the models of attitudes, we are
studying human-human dialogues.
Our current analysis is based on the Estonian
dialogue corpus (Hennoste et al., 2008). The biggest
part of the corpus – about 1000 spoken dialogues – is
recorded in authentic situations and transcribed by
using the transcription system of Conversation
Analysis (Sidnell and Sivers, 2012). Each
transcription is provided with a header that lists
situational factors (meta-knowledge about the
dialogue session), which affect language use –
participants’ names, social characteristics, relations
between participants in the situation, specification of
situation (private/public place, private/institutional
conversation), etc.
For the current study, we have chosen five phone
calls from the corpus where the acquaintances are
negotiating about doing an action by one of them (cf.
Koit, 2019). We are studying how the attitudes of the
negotiation participants are changing and how well
the models describe the changes. To do so, we
annotated both the attitudes in relation to the
negotiation object and in relation to the
communication partner in every dialogue move using
corresponding vectors.
In the next example (s. the following Table 1, left
column) taken from our analyzed sub-corpus, mother
A entices her son B to bake gingersnaps (action D).
She presents several arguments in order to
increase the son’s wish to do the proposed action by
stressing the pleasantness of the action until he finally
agrees. In the current study, the initial attitudes (the
coordinates of the vectors w
AB
and w
B
, s. Table 1)
have been determined by an informal analysis of the
whole dialogue text. Further, we suppose that every
argument presented by A will change the targeted
attitude (the pleasantness of D) by the value 1.
(Therefore, we suppose that all the arguments have
the weights equal to 1; still, this is a simplification.)
KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development
228

Table 1: Dialogue example (A – mother, B – son).
Utterances
A’s partner model w
A
B
(A’s beliefs about B’s
attitudes in relation to D)
B’s model w
B
(B’s actual attitudes
in relation to D)
A’s attitudes s
AB
in relation to B
B’s attitudes s
BA
in relation to A
/---/ (1) A: ´küsimus.
A question.
(0.6) .hhhhh kas sulle pakuks ´pinget
´piparkookide ´küpsetamine.
Would you like to bake gingersnaps?
(2) B: ´praegu.
Just now?
(3) A: jah.
Yes.
(1,6,2,1,1,1,0,0,0)
WISH (Fig.1) gives the
decision “do” (steps 1, 2,
3, 10)
(1,1,5,2,1,1,1,0,0)
WISH is not
applicable
(presumption is not
fulfilled: 1<5);
NEEDED (Fig.2)
gives the decision
„do not do“ (steps 1,
2, 5, 6, 8)
(1,1,1,0,1,1,0)
(4) B: .hhhhhhh ma=i=´tea vist ´mitte.
I don’t know, perhaps no.
(1,6,2,2,1,1,1,0,0,0)
WISH gives „do not do“
(-1,1,-1,0,1,0,0)
(5) A: ja=sis gla´suurimine=ja=´nii.
And then glazing and so on.
(0.6)
(1,2 3,2,1,1,1,0,0,0)
WISH gives „do“
(1,2,5,2,1,1,1,0,0)
NEEDED gives „do
not do“
“
(6) B: ´ei, ´ei, ´ei ei=´ei.
No, no, no, no, no.
(0.9)
(1,3,2,2,1,1,1,0,0,0)
WISH gives „do not do“
(-1,1,-1,0,1,0,1)
intensity –
vehement
(7) A: me saaksime nad
´vanaema=jurde ´kaasa võtta.
We can take them with us when
going to grandmother.
(1,2,3,2,1,1,1,0,0,0)
WISH gives „do“
(1,3,5,2,1,1,1,0,0)
NEEDED gives „do
not do“
“
(8) B: ´präägu ei=´taha.
I just don’t want.
(1.3)
(1,2,2,2,1,1,1,0,0,0)
WISH gives „do not do“
(-1,1,-1,0,1,0,0)
intensity –
neutral
(9) B: aga (.) noh, kas sa mõtled nagu
.hhh kui sa tuled ´koju=vä.
But what do you think – after you
come home?
(10) A: .hhh ei
No.
ma mõtlen: kui mind kodus ei=´ole.
I think when I’m not home.
(11) B: aa.
Aha.
(0.5) .hhh et ´lähen ostan ´tainast=vä.
Then I’ll go to buy paste, right?
(12) A: ja=niimodi=jah,
Yes, right.
(1,3,2,2,1,1,1,0,0,0)
WISH gives „do“
(1,4,5,2,1,1,1,0,0)
NEEDED gives „do“
“
(-1,1,0,0,1,0,0)
cooperation –
neutral
(-1,1,1,0,1,0,0)
cooperation –
collaborative
(13) .hhh sinna:: ´Pereleiva
´kohvikusse võiksid minna @ ´võiksid seal
endale ühe ´kohvi lubada=ja @ (2.7) teha
ostmise ´mõnusaks=ja (0.8) ja=siis tulla
´koju=ja? (1.7) ´piparkooke teha=ja
And you could go to Pereleiva cafe
and take a coffee in order to make buying
pleasant for you, and then go home to
bake gingersnaps.
(1,4,2,1,1,1,0,0,0)
WISH gives „do“
“
(1.2)
(14) B: okei?
OK.
/---/
(1,5,5,2,1,1,1,0,0)
Both NEEDED and
WISH are applicable,
WISH gives „do“
“
(15) A: .hhhhhhhhhh (0.2) ja ´siis ma
tahtsin sulle öelda=et ´külmkapis on: ´sulatatud
või tähendab=ned ´külmutatud ja ´ülessulanud
´maasikad ja ´vaarika´mömm.=hh
And I’d like to tell you that there are
frozen strawberries and raspberries in
the refrigerator.
(1,5,2,1,1,1,0,0,0)
WISH gives “do” (
“
(16) B: jah
Yes.
(17) A: palun ´paku endale sealt.
Please help yourself. /---/
(1,6,5,2,1,1,1,0,0)
WISH gives „do“
“
Modelling Attitudes of a Conversational Agent
229

A initiates the dialogue. Her communicative goal
is to convince B to do the proposed action D (baking
gingersnaps). A is using the partner model w
AB
(corresponding to her image of B) by supposing that
B has all the resources to do the action (the value of
the first coordinate is 1), the pleasantness (6) is much
greater than the unpleasantness (2), the usefulness
and the harmfulness are equal (both 1), the action is
obligatory (value 1) for B (because son is obliged to
fulfil mother’s requests). Still, mother will not punish
(0) son if he does not fulfill the request. Further, the
action being obligatory is not prohibited (0) and no
punishment (0) will follow if it will be done. A applies
the reasoning procedure WISH (Fig. 1) in the partner
model and achieves the decision “do D” (s. the Table,
2
nd
column).
At the same time, B‘s actual attitudes are different
(model w
B
, s. the Table, 3rd column). He cannot apply
the reasoning procedure WISH which presumption is
not fulfilled (Fig.1). Instead, he applies the procedure
NEEDED (Fig.2) which unfortunately gives the
decion „do not do D“ as indicated by B’s answer
(utterance 4).
Now A has to introduce the changes into the
partner model in order to get the same decision like B
got and to present an argument for the pleasantness
(utterance 5). Influenced by A’s argument, B
increases his attitude in relation to the pleasantness
(by 1 as we suppose) and applies the reasoning
procedure NEEDED in his changed model.
The dialogue continues in the similar way until
A’s argument presented by the utterance (10) makes
it possible for B by applying the procedure NEEDED
to get the decision “do” (11). It turns out that A’s next
argument (13) will increase the pleasantness for B in
such a way that he can apply the reasoning procedure
WISH (prerequisite is fulfilled) therefore, at this
moment he started to want to do the action. A does not
stop but presents one more argument (15) which
increases B’s wish once more. Therefore, both A and
B finally achieved their communicative goal.
The initial social attitudes have been determined
by using meta-knowledge about the dialogue session
that are given in the header of the transcript and by
the analysis of the first turns. Mother dominates over
son (the value on the dominance scale 1 for mother
and -1 for son), communicative distance is ‘near’ for
the both participants (value 1), mother expresses
cooperativity (value 1) but son – antagonism (value
-1), the politeness is neutral (0), communication is
personal (value 1), modality is 1 for mother and 0 for
son, intensity is neutral for both. As seen in the Table
(two last columns), mother keeps her social attitudes
during the whole negotiation but son’s utterance (6) –
strong rejection – expresses the increased modality
(value 1) which decreases to 0 in his next utterance
(8). Son’s utterance (9) demonstrates that antagonism
has decreased – he started to doubt in his previous
rejection and asks an adjusting question (value 0 on
the cooperativity scale). By the utterance (11) son
expresses cooperativity (value 1). Therefore, the final
attitudes can be represented by the vectors s
AB
=
(1,1,1,0,1,1,0) and s
BA
= (-1,1,1,0,1,1,0),
respectively. Thus, the participants have approached
one to another during the negotiation.
Here we do not consider the problem of how to
determine automatically the initial attitudes of both
kinds. This needs semantic analysis of dialogues that
is currently not available for Estonian. In addition,
recognition of social attitudes is hard without taking
into account nonverbal means.
5 IMPLEMENTATION
A limited version of the conversational agent is
implemented as a simple DS that interacts with a user
in written Estonian. Information-state dialogue
manager is used in the implementation. The
programming language is Java.
The agent plays A’s role and the user B’s role. A’s
communicative goal is “B will do D”. The computer
has ready-made sentences (assertions) for expressing
of arguments, i.e., for stressing or downgrading the
values of different aspects of the proposed action,
which depend on its user model. The user (B) can
choose one of two actions – traveling to a certain
place or becoming a vegetarian. The user can
optionally use another set of ready-made sentences or
put in free texts. In the last case, keywords are used
in order to analyze the texts. In the current
implementation, the social attitudes of participants
are not included.
Starting a dialogue, A determines a partner model
w
AB
D
, fixes its communicative strategy and chooses
the communicative tactics that it will follow, that is,
the computer respectively determines a reasoning
procedure that it will try to trigger in B’s mind. A
applies the reasoning procedure in its partner model,
in order to ‘put itself’ into B’s role and use suitable
arguments when convincing B to decide to do D. The
user is not obliged (but can) to follow neither certain
communicative tactics nor reasoning procedures.
He/she is also not obliged to fix his/her attitudes in
relation to D by composing the vector w
B
D
. However,
A does not ‘know’ B’s attitudes but it only can choose
its arguments on the basis of B’s rejection and/or
counterarguments. Respectively, A is making
KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development
230

changes in its partner model during a dialogue. The
process runs in the similar way as described in the
Table above.
6 DISCUSSION
Our model of conversational agent considers two
kinds of attitudes of a dialogue participant: (1) the
attitudes in relation of doing an action (which is the
negotiation object), and (2) the attitudes related to a
communication partner. Both kinds of attitudes are
changing in dialogue as influenced by the arguments
of the communication partner.
To model the attitudes in relation to a reasoning
object (in our case, doing an action), we use the vector
which coordinates correspond to the different aspects
of the action (its pleasantness, usefulness, etc.). We
evaluate these aspects by giving them discrete
numerical values on a numerical scale. Still, people
do not operate with numbers and what is more, with
the exact values of the aspects of an action. They
rather make ‘fuzzy calculations’, for example, they
believe that doing an action is more useful than
harmful and therefore it is needed to do it. In addition,
if a reasoning object is different (not doing an action
like in our case) then the attitudes of a reasoning
subject can be characterized by a different set of
aspects.
Social attitudes are modelled by using the concept
of social space that dimensions correspond to the
different relations of communication participants:
dominance-subordination, social distance, politeness,
etc. Currently, we use the values 1, 0 and -1 for the
coordinates. Still, it is possible to operate with
continuous scales instead of discrete values (cf.
Mesiarová-Zemánková, 2016). It is also possible to
use words of a natural language for the values. For
example, the intensity of communication can be
indolent, restrained, vehement, etc. However,
annotation of the points of social space in written
dialogues (transcripts of spoken dialogues) is difficult
and subjective already with three different values (1,
0, -1). So far, we have not done that automatically.
Further empirical research is needed in order to
determine the list of dimensions of communicative
space, their relations and values on different scales
(which can be different). Linguistic cues can be used
for recognizing of values of some coordinates. For
example, if a participant uses the 2nd person singular
form of pronouns in Estonian then he/she is indicating
a short communicative distance (-1) and a big value
on the personality scale (+1). Feeling words can be
used for recognizing the values of some coordinates,
for example, please and thank indicate politeness.
The comments of transcribers in transcripts of spoken
dialogues help to determine the modality of
communication (for example, the comment
((violently)) indicates the value 1 on the intensity
scale). The dialogue act tags also contribute to
recognizing of some coordinates. Opinion mining
(Liu, 2015) can be used to automatically annotate
communication points. Still, the small size of the
Estonian dialogue corpus does not yet allow
implement statistical or machine learning methods.
How to use the proposed models in human-
computer systems? The conversational agents and
especially, the conversational characters, which have
recently become popular, take into account only the
features of a limited field (e.g., a virtual guide of an
art exhibition). At the same time, the agents can be
created which could be ‘tuned’ to behave according
to certain locations in social space depending on the
user. For example, a travel agent gives information
about a trip but it can also add various advices being
neutral, advertising or even intrusive.
7 CONCLUSION
The paper describes a model of conversational agent
that we are implementing in an experimental dialogue
system. We are considering argument-based
negotiations where the goal of the initiator (A) is to
get the partner (B) to carry out a certain action D.
We consider two kinds of attitudes expressed by
participants in negotiation about doing an action: (1)
related to the action, and (2) related to a
communication partner. We represent the first kind of
attitudes as coordinates of a vector of motivational
sphere of a participant who is reasoning whether to do
an action or not. The second kind of attitudes is
represented by using the concept of social space – a
mental space where communication takes place.
We introduce information states of the
conversational agent. Update rules allow move from
one information state into another. An information
state of the agent A includes a partner model that
consists of evaluations of different aspects of the
action under consideration. The attitudes included
into the partner model as well as the actual attitudes
of the partner are changing during the interaction,
based on the arguments and counter-arguments
presented. The conversational agent acts in social
space that represents its attitudes related to the partner
(social distance, politeness, etc.). The partner
similarly behaves according his/her attitudes that are
expressed in utterances. We consider here only verbal
Modelling Attitudes of a Conversational Agent
231

communication, non-verbal means are not taken into
account.
In order to evaluate our attitude models we
analyze a small sub-corpus of the Estonian dialogue
corpus formed by human-human phone calls. The
analysis demonstrates how the attitudes of the
negotiation participants are changing and how the
different points in social space are visited during
conversation.
We have implemented an experimental
conversational agent, which argues for doing an
action by interacting with the user in written Estonian.
We believe that including the model of social space
into the system will make the interaction more
human-like. This needs more advanced processing of
Estonian and remains for the further work.
ACKNOWLEDGEMENTS
This work was supported by institutional research
funding IUT (20-56) of the Estonian Ministry of
Education and Research, and by the European Union
through the European Regional Development Fund
(Centre of Excellence in Estonian Studies).
REFERENCES
Bratman, M.E. 1999. Intention, Plans, and Practical
Reason, CSLI Publications.
Bunt, H. 2014. A context-change semantics for dialogue
acts, Computing Meaning, 4, 177–201.
Callejas, Z., Ravenet, B., Ochs, M., Pelachaud, C. 2014. A
Computational model of Social Attitudes for a Virtual
Recruiter, AAMAS, 8 p.
Carofiglio, V., De Carolis, B., Mazzotta, I., Novielli, N.,
Pizzutilo, S. 2009. Towards a Socially Intelligent ECA,
IxD&A 5-6, 99–106.
Dermouche, S. 2016. Computational Model for
Interpersonal Attitude Expression, ICMI, 554–558.
DeVaultand, D., Mell, J., Gratch, J. 2015. Toward Natural
Turn-Taking in a Virtual Human Negotiation Agent,
AAAI Spring Symposium on Turn-taking and
Coordination in Human-Machine Interaction, 9 p.
Gratch, J., Hill, S., Morency, L.-P., Pynadath, D., Traum,
D. 2015. Exploring the Implications of Virtual Human
Research for Human-Robot Teams, VAMR, 186–196.
Hartholt, A., Traum, D., Marsella, S.C., Shapiro, A.,
Stratou, G., Leuski, A., Morency, L.P., Gratch, J. 2013.
All together now: Introducing the Virtual Human
toolkit, IVA, 368–381.
Hennoste, T., Gerassimenko, O., Kasterpalu, R., Koit, M.,
Rääbis, A., and Strandson, K. 2008. From Human
Communication to Intelligent User Interfaces: Corpora
of Spoken Estonian, LREC, 2025–2032.
Jokinen, K. 2018. AI-based Dialogue Modelling for Social
Robots, FAIM/ISCA Workshop on Artificial
Intelligence for Multimodal Human Robot Interaction,
57-60.
Jurafsky, D., Martin, J. 2013. Speech and Language
Processing. An Introduction to Natural Language
Processing, Computational Linguistics, and Speech
Recognition. Prentice Hall; 2nd edition.
Knapp, M.L., Hall, J.A., 2009. Nonverbal Communication
in Human Interaction. Wadsworth Publishing.
Koit, M. 2019. Changing of Participants’ Attitudes in
Argument-Based Negotiation, ICAART, 2, 770−777.
Koit, M. 2018. Reasoning and communicative strategies in
a model of argument-based negotiation, Journal of
Information and Telecommunication TJIT, 2, 14 p.
Taylor & Francis Online
Koit, M., Õim, H. 2014. A computational model of
argumentation in agreement negotiation processes,
Argument & Computation, 5 (2-3), 209–236. Taylor &
Francis Online.
Lewis, M., Yarats, D., Dauphin, Y.N., Parikh, D., Batra, D.
2017. Deal or No Deal? End-to-End Learning for
Negotiation Dialogues, EMNLP, 2443–2453.
Liu, B. 2015. Sentiment Analysis. Mining opinions,
sentiments, and emotions. Cambridge University Press.
Mensio, M., Rizzo, G., Morisio, M. 2018. The Rise of
Emotion-aware Conversational Agents: Threats in
Digital Emotions. The 2018 Web Conference
Companion, Lyon, France, 4 p.
Mesiarová-Zemánková, A. 2016. Sensitivity analysis of
fuzzy rule-based classification systems by means of the
Lipschitz condition, Soft Computing, 103–113.
Oshrat, Y., Lin, R., Kraus, S. 2009. Facing the Challenge of
Human-Agent Negotiations via Effective General
Opponent Modeling, AAMAS, 377–384.
Õim, H. 1996. Naïve Theories and Communicative
Competence: Reasoning in Communication, Estonian
in the Changing World, 211–231.
Sidnell, J., Stivers, T. (Eds.) 2012. Handbook of
Conversation Analysis. Boston: Wiley-Blackwell.
Ravenet, B., Cafaro, A., Biancardi, B., Ochs, M.,
Pelachaud, C. 2015. Conversational Behavior
Reflecting Interpersonal Attitudes in Small Group
Interactions, IVA, 375–388.
Ravenet, B., Ochs, M., Pelachaud, C. 2012. A
computational model of social attitude effects on the
nonverbal behavior for a relational agent, WACAI, 9 p.
Rosenfeld, A., Zuckerman, I., Segal-Halevi, E., Drein, O.,
Kraus, S. 2014. NegoChat: A Chat-based Negotiation
Agent, AAMAS, 525–532.
Traum, D. 2017. Computational Approaches to Dialogue,
The Routledge Handbook of Language and Dialogue.
Traum, D., Larsson, S., 2003. The Information State
Approach to Dialogue Management, Current and New
Directions in Discourse and Dialogue, 325–353.
Traum, D., Marsella, S., Gratch, J., Lee, J., Hartholt, A.
2008. Multi-party, Multi-issue, Multi-strategy
Negotiation for Multi-modal Virtual Agents, IVA, 117–
130.
KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development
232