A computational model of power in collaborative

negotiation dialogues

Lydia Ould Ouali1 , Nicolas Sabouret1 and Charles Rich2

1
LIMSI-CNRS, UPR 3251, Orsay, France
Université Paris-Sud, Orsay, France
{ouldouali, nicolas.sabouret}@limsi.fr
2
Worcester Polytechnic Institute
Worcester, Massachusetts, USA
[email protected]

Abstract. This paper presents a conversational agent that can deploy

different strategies of negotiation based on its social power. The under-
lying computational model is based on three principles of collaborative
negotiation from the literature in social psychology. The social behavior
of the agent is made visible through its dialogue strategy. We evaluated
our model by showing that these principles are correctly perceived by
human observers on synthetic dialogues.

1 Introduction

As they rise in popularity, artificial conversational agents become more present in

daily applications in which they play different roles such as pedagogical robots,
companion robots for children or for the elderly, etc. In these situations, agent
and user have to collaborate through their interaction in order to achieve shared
goals. Such example of collaboration can be found in intelligent tutoring agents
[14], where agent and learner collaborate to achieve exercises. They compare
their respective knowledge on the problem to be solved and discuss possible
solutions. Such confrontation offers a personalized teaching to the learner.
As illustrated in the above example, user and the agent typically negotiate
in a collaborative manner about the way to achieve the shared goal, depend-
ing on individual expertise and preferences. This specific type of discussion is
called collaborative negotiation. Unlike adversarial negotiation [3], collaborative
negotiation assumes that each participant is motivated by the goal of finding a
trade-off that best satisfies the interests of both participants, instead of one that
maximizes his own interest [22, 5].
Our goal is to build conversational agents capable of credible collaborative
negotiation with a human user. To this end, we need to understand how human
beings behave in such situation. Indeed, as stated by [3], negotiation is a multi-
faceted process which involves social interaction and affects as well as personal
preferences and opinions. Therefore, it is crucial to understand the impact of
social aspects of the negotiation process. Researchers in social psychology and

© Springer International Publishing AG 2017 259

J. Beskow et al. (Eds.): IVA 2017, LNAI 10498, pp. 259-272, 2017.
DOI 10.1007/978-3-319-67401-8_35
260 L. Ouldouali et al.

communication [8, 7] showed that social power directly affects the strategies of
negotiators. Therefore, we claim that conversational agent which is able to nego-
tiate with a user must be able to adapt their negotiation strategies to different
levels of power.
In this paper, we present a conversational agent that aims to contribute
to essentials aspects of conversation agents in terms of negotiation and social
behaviors. We propose an agent capable of negotiating with the user. In addition,
it can deploy different strategies of negotiation based on the social power it wants
to express.
In the next section, we discuss existing works on interpersonal power in the
domain of social psychology and affective computing. Section 3 presents the
computational model of negotiation, based on a set of utterance types and a
model of preferences. It implements a general model of negotiation based on three
principles of collaborative negotiation from the literature in social psychology.
In section 4, we present an experiment conducted with two virtual agents on a
negotiation about restaurants showing that the principles are correctly perceived
by human observers.

2 Related work

The notion of social power has been widely studied in the fields of interpersonal
communication and psychology [13]. It can be defined as the capacity to produce
intended effects and to influence the behavior of the other person in the conver-
sation [8]. In the context of communication and negotiation, power is a dyadic
variable that takes place during the dialogue. Behaviors related to power can
contribute either positively or negatively to the dialogue. Positive contributions
include keeping the conversation going and making quick decisions. Negative
contributions include not considering the partner (e.g. not giving the occasion
to express his opinion) and appearing offensive. In our work, we focus on nego-
tiation dialogues, in which several researchers have already shown the impact of
social power[6, 4].

2.1 Behaviors of power in dialogue

During a dialogue, power can manifest through verbal and nonverbal behaviors.
At the nonverbal level, a wide range of behaviors have been associated with the
relation of power in kinesthetic behaviors (facial expression, body movements
and gestures) and voice (speaking duration, speaking intensity, voice control and
pitch) [4]. Based on this work, several conversational agents have been developed
with the ability to exhibit social power through nonverbal behavior, such as gaze
[16], body movements [19] or head tilt [11] in relation to high-power and low-
power perception.
However, power is also expressed through verbal behaviors. A considerable
body of research in social science and communication has documented the effects
of power on negotiation behaviors and outcomes. De Dreu [7] demonstrated that
A computational model of power in collaborative negotiation dialogues 261

high-power negotiators have higher aspirations, demand more and concede less.
Galinsky [10] affirms that power increases the action orientation: high-power
negotiators control the flow of the negotiation. In addition, high power increase
task orientation and goal-directed behavior. Giebels [12] shows that this leads
powerful negotiators to end up with the larger share of the pie.
Furthermore, power affects the way that negotiators gather information about
their partners [6]. Low-power negotiators have a stronger desire to develop an
accurate understanding of their negotiation partner, which leads them to ask
more diagnostic rather than leading questions.
It was also shown that high-power negotiators are self-centered and tend not
to pay attention to the preferences of the less powerful negotiators [9, 7].
Our goal is to develop a model of dialogue for Virtual Agents which considers
these properties related to social power. We want to make visible the strategies
deployed during the negotiation depending on the power. In order to implement
these different behaviors, we identified three principles related to the relation of
power and their impact on the strategy of negotiation
1. Level of demand and concession: High-power negotiators show a higher
level of demand than the low-power ones. In addition, low-power negotia-
tor’s demand decrease over time and the negotiator tends to make larger
concessions than high-power negotiators. [7]
2. Self vs other: Low-power negotiators consider the preferences of the other
in the negotiation, whereas high-power negotiators are self-centered and only
interested in satisfying their own preferences. [9, 7]
3. Controlling the flow of the negotiation: High-power negotiators tend
to make the first move [17] and take the lead in the negotiation. Low-power
negotiators aim to construct an accurate model of other preferences, which
leads them to ask more questions about other preferences rather than keeping
the negotiation going (e.g by making proposals)[6].
In the context of collaborative negotiation, we must combine these principles
with the goal of finding a trade-off that satisfies both negotiators.

2.2 Similar work in the literature

Only a few researchers have considered the expression of power in the verbal
behavior of a conversational agent. [1] developed an agent that expresses social
power through gaze and linguistic features. They demonstrated that linguistic
personality traits influence the perception of power. However, this work does
not consider how power affects the strategies of negotiation in dialogue. More
recently, [3, 18] consider trust, expression of emotions as anger and happiness as
dimensions of the negotiation strategy of a virtual agent. However, this research
focuses more on the negotiation aspect than on the expression of social power.
In our work, we want to investigate the expression of power through the dialogue
strategy, which has not been considered by previous work.
In a different context, [20] studied how the communication of preferences in
a negotiation with a human can impact the negotiation outcomes. They consider
262 L. Ouldouali et al.

situations in which negotiators can use deception to gain an unfair advantage and
can hide their intentions. We do not consider this case in our work. Our agents
are not only collaborative, but they also give correct information about their
preferences and they do not try to hide any information. Previous work such as
[23] showed that, even in collaborative negotiation, it is not always possible or
advisable to give information about all preferences.

3 Model of negotiation based on the relation of power

In this section, we present our model of dialogue for a Virtual Agent in the
context of collaborative negotiation based on power. First, we present the data
structure for the agent’s preferences and the topics of the negotiation. Second, we
present the implementation of the principles of behaviors of power in negotiation
discussed in this section 2.1.
3.1 Domain model
The overall goal of a negotiation is to choose an option in a set of possible
options O. The evaluation of each option by participants is based on a set of
criteria that reflect the option’s characteristics. Let us consider a set C of n
criteria and let C1 , . . . , Cn be their respective domains of values. O can be simply
defined as the cross-product C1 × . . . × Cn and each option o ∈ O is a tuple
(v1 , . . . , vn ), making the simplifying assumption that all options are available.
For instance, in a dialogue about restaurants, the criteria might be the type of
cuisine and the price, we could have the option: (French, expensive).

3.2 Preference model

The conversational agent is defined with a set of preferences, formalized as a set
of partial orders ≺i on each Ci . For instance, if the agent prefers French food to
Italian, Italian ≺cuisine French.
For a given criterion i ∈ C, for a given value v ∈ Ci , the agent computes its
satisfaction satself (v ≺i ) for this value as the number of values it prefers less
in the partial order ≺i , normalized in [0,1]:
   
|{v : v = v ∧ (v ≺i v  )}|
satself (v, ≺i ) = 1 − (1)
(|Ci | − 1)
This notion of satisfaction is generalized to any option o = (v1 , . . . , vn ) ∈ O
as a simple average3 : n
satself (vi , ≺i )
satself (o, ≺) = i=1 (2)
n
The satisfaction represents the tendency for an agent to accept or to propose
a possible value or option in the negotiation dialogue: the closer to 1, the more
it will be selected early in the negotiation.
3
There exists a great amount of literature in theoretic decision making on how to
combine multiple criteria using Order Weight Averages or Choquet’s integrals, for
instance. We are not concerned with this question of criteria aggregation in this
paper.
A computational model of power in collaborative negotiation dialogues 263

3.3 Dialogue model

Negotiators communicate during the negotiation via utterances. Each utterance

type has a specific set of arguments and is associated with a specific expression in
natural language (NL). We use five utterance types, based on the work of Sidner
[22] and two additional utterances to close the negotiation. Table 1 summarizes
these generic utterance types. Only the NL generation of these utterances has
to be specialized in the application domain. In this paper, we will illustrate this
model on a collaborative negotiation about restaurants. The value /v/ in Table
1 refers to this NL format to express a value.
Each utterance type takes as parameter either a criterion value v ∈ Ci ,
an option o ∈ O or a criterion type i ∈ C. They can be separated into three
groups. Information moves (AskValue/AskCriterion and StateValue) are used
to exchange information about the participant’s likings. In human-human nego-
tiation dialogue, we observed that partners tend to express what they like or do
not like (e.g I (don’t)like Chinese restaurants) rather than binary comparison
(e.g I like Chinese more than French). This is the reason we use unary operators
for the information utterances.
Negotiation moves (Propose, Accept and Reject) allow the agent to make or
to answer to proposals. The agent can propose, accept and reject both values
(“Let’s go to a Chinese restaurant”) or options (“Let’s go to Chez Francis”).
Closure moves (NegotiationSuccess or NegotiationFailure) are used to end
the dialogue. Examples of dialogues for restaurants are given in section 4.
The decision process for utterance selection is based on our three principles
and is described in section 3.4. To perform this utterance selection, the agent
keeps track of all statements and proposals during the dialogue. For each criterion
i ∈ C, we build the set Si ⊆ Ci of statements that the agent has made about this
criterion. This avoids restatements of previous information. We also build the
sets Ai ⊆ Ci and Ui ⊆ Ci of values which have been stated by the interlocutor
as liked or disliked through StateValue utterances. We assume that Ai ∩ Ui = ∅.
We also maintain the sets Pi ⊆ Ci , Ti ⊆ Ci and Ri ⊆ Ci of all proposed,
accepted and rejected values for each criteria. These will be used to make relevant
proposals. Similarly, we consider P ⊆ O, T ⊆ O and R ⊆ O the sets of all
proposed, accepted and rejected options in the dialogue.

Satisfiability Using the sets Ai and Ui that represent the interlocutor’s likings,
the agent can compute the satisfiability of a value v ∈ Ci for the interlocutor
(i.e. the other person) using the following formula:
⎧
⎨ 1 if c ∈ Ai
satother (v) = 0 if c ∈ Ui (3)
⎩
0.5 otherwise

Note that some values might remain unknown: in a collaborative negotiation,

they can be considered as being potentially satisfiable. Therefore, we give them
an arbitrary value fixed to 0.5.
264 L. Ouldouali et al.

Utterance type NL generation Postcondition

StateValue(v) I (don’t) like /v/. Speaker : v ∈ Si
Hearer:
v ∈ Ai is likable, v ∈ Ui
otherwise
AskValue(v) Do you like /v/ ?
AskCriterion(i) What kind of /i/ do you like ?
ProposeOption(o) Let’s go to /o/. o∈P
ProposeValue(v) Let’s go to a /v/. v ∈ Pi
AcceptOption(o) Okay, let’s go to /o/. o∈T
AcceptValue(v) Okay, let’s go to a /v/. v ∈ Ti
RejectOption(o) I’d rather choose something else. o∈R
RejectValue(v) I’d rather choose something else. v ∈ Ri
NegotiationSuccess We reached an agreement.
NegotiationFailure Sorry, but I no longer want to
discuss this.

Table 1: List of utterance types in the model of dialogue.

This function is generalized to any option o = (v1 , . . . , vn ) ∈ O:

n
satother (vi , Ai , Ui )
satother (o, A, U) = i (4)
n
The satisfiability represents the expectation an agent has about the other
accepting or rejecting a proposal. It will be used to select the most tolerable
offer to make in the negotiation. The concept of tolerable is presented in the
next section.

3.4 Decision based on power in negotiation

In section 2.1, we identified three principles related to the relation of power

which affects negotiators strategies and behaviors. In this section, we present
the computational theory implementing each principle.
We denote the agent’s belief of its power pow ∈ [0, 1]. It is a constant for a
given agent in a given interaction.

Level of demand and concession In collaborative negotiation, both partic-

ipants reduce their level of demand over time because they want to reach an
agreement. However, according to our first principle, the level of demand should
be higher for high-power agent and concessions should be greater for low-power
ones. To model this behavior, we use a concession curve illustrated in Figure 1.
Let self(pow, t) be a time varying value, following the concession curve:

pow if (t  τ)
self(pow, t) = δ (5)
max(0, pow − ( pow · (t − τ))) otherwise

where is t  0 is the number of open or rejected proposals, τ > 0 is the

minimum number of proposals before the concession begins and δ > 0 is a
computational parameter of the concession slope.
A computational model of power in collaborative negotiation dialogues 265

The value self(pow, t) represents the weight an agent gives to its self-satisfaction
relative to the satisfaction of its partner. The higher the agent’s power gets, the
higher its demands get. In addition, the slope decreases faster for low values of
power.
These behaviors of demand and concession
are implemented in the computation of an ac-
ceptability value. Based on the value of satisfia-
bility, the agent is able to tell if it likes a value
or not.
The acceptability of a value v ∈ Ci is defined
as a boolean function: Fig. 1: Concession curve
acc(pow, v, t) = satself (v, ≺i )  (β·self(pow, t))
(6)
where β > 0 is a parameter of the theory that defines the weight given to
the level of demand.
This boolean function is generalized to any option o ∈ O: acc(pow, o, t) =
satself (o, ≺)  (β · self(pow, t)). This acceptability is used by the agent to
decide whether he accepts a proposal or not.

Self vs other According to our second principle, high-power negotiators give

more weight to their own satisfaction. To implement this principle in the context
of collaborative negotiation, we compute how much a given proposal is tolerable
considering the satisfiability for both the agent and its partner.
For a given criteria i ∈ C, let us consider the subset Vi ⊆ Ci of values that
are acceptable for the agent:
Vi (pow, t) = {v ∈ Ci : acc(pow, v, t)} (7)
This set corresponds to all the proposals an agent could make at a time of the
negotiation.
We compute the tolerability of a given value v ∈ Vi (pow, t) by balancing
between the agent’s preferences and the likings of its partner. We assume that
the agent gives a weight to its partner satisfaction which is complementary to
its self-satisfaction:

tol(v, t, ≺i , Ai , Ui , pow) = self(pow, t) · satself (v, ≺i )

(8)
+ (1 − self(pow, t)) · satother (v, Ai , Ui )

And we generalize this function to any option o = (v1 , . . . , vn ) ∈ O:

n
tol(vi , t, ≺i , Ai , Ui , pow)
tol(o, t, ≺, A, U, pow) = i (9)
n
The agent will always propose the most tolerable element in Vi :

propose(Vi , ≺i , pow) = arg max(tol(v)) (10)

v∈Vi
266 L. Ouldouali et al.

Summary of general parameters

– π ∈[0,1] : boundary between submissive and dominant used in choosing an
utterance type
– β: a value that represents the minimum score that a value has to get to be
positively satisfiable to the agent preferences in the negotiation. Note that
β = const × self(dom, t).
– τ > 0 : the minimum number of open or rejected proposals before concession
begins
– δ > 0 : parameter in slope of concession curve.
– α > 0: the maximum number of successive statement moves.

Controlling the flow of the negotiation According to our third principle,

high-power negotiators tend to lead the negotiation. We implemented this prin-
ciple through the choice of utterance types presented in Table 2. We defined a
threshold π to split the spectrum of power in two.
Depending on the power, the previous utterance u−1 type and the current
dialogue state, the agent selects the first utterance type in Table 2 for which the
condition is satisfied. For instance, a high-power agent will stop the negotiation
as soon as all the remaining options are unacceptable (line 2). A low-power
agent will reject and state a preference, so as to explain why the proposal is not
acceptable (line 14). If there is no open proposal, the low-power agent will ask
for new information (line 18 -19).
In our model, an agent can express more than one utterance during his speak-
ing turn. This is modeled with the sign ”+” in Table2.
Note that a high-power agent will focus on keeping the negotiation going by
choosing negotiation moves (ProposeValue /ProposeOption, RejectValue /Rejec-
tOption, AcceptValue/ AcceptOption) as presented in lines (4 to 10). The agent
prioritizes the negotiations moves rather than exchanging information about the
preferences. Indeed, as presented in line 3, after α speaking turns dedicated to
sharing information, the agent will rather make proposals than stating his likes.
An example is presented in the dialogue 2.
On the contrary, a lower power negotiator will focus on building an accurate
model of its partner’s likings in order to take the fairest decision. It will focus
more on information moves (StateValue or AskValue/ AskCriterion) as seen in
line(18-20). Moreover, the negotiation moves are restricted by conditions which
ensure that the agent gathered enough information about its partner likes before
to express a proposal (line 16-17).

4 Evaluation
In order to evaluate our model, we conducted a controlled study in which par-
ticipants judge the behaviors of two agents implemented using our model. Our
system is written in Java using the DISCO platform [21]. This allowed us to
produce synthetic dialogues between two artificial agents with different values
of power and varying preferences.
A computational model of power in collaborative negotiation dialogues 267

Line Utterance type Condition

pow > π

nb
1 NegotiationSuccess ∃o ∈ T ∪ P, acc(pow, o, t)
2 NegotiationFailure ∀o ∈ O, ¬acc(pow, o, t)
3 StateValue(v) type(u−1 ) = AskPreference ∧ n < α
where n is the number of successive statement moves
4 AcceptValue(v)+ ∃v ∈ Pi / acc(pow, v, t) ∧ ∃i ∈ C, acc(pow, c, t)
ProposeValue(c)
5 AcceptValue(v)+ ∃v ∈ Pi / acc(pow, v, t) ∧ ∃o ∈ O/ v ∈ o ∧ acc(pow, o, t)
ProposeOption(o)
6 RejectValue(v)+ ∃v ∈ Pi / ¬acc(pow, v, t) ∧ ∃i ∈ C, acc(pow, c, t)
ProposeValue(c)
7 RejectValue(v)+ ∃v ∈ Pi / ¬acc(pow, v, t) ∧ ∃o ∈ O/ acc(pow, o, t)
ProposeOption(o)
8 RejectOption(o1 )+ ∃o1 ∈ P / ¬acc(pow, o1 , t) ∧ ∃o2 ∈ O, acc(pow, o2 , t)
ProposeOption(o2 )
9 ProposeValue(v) ∃v ∈ Ci / tol(v, t, ≺i , Ai , Ui , pow)
10 ProposeOption(o) ∃o ∈ O / tol(o, t, ≺i , Ai , Ui , pow)
11 Negotiation success ∃o ∈ T
pow  π

12 AcceptValue(v) ∃i ∈ C, ∃v ∈ Pi , acc(pow, v, t)
13 AcceptOption(o) ∃o ∈ P, acc(pow, o, t)
14 RejectValue(v)+ t < τ ∧ (∃i ∈ C, ∃v ∈ Pi , ¬acc(pow, v, t)).
StateValue(v)
15 RejectOption(o)+ t < τ∧(∃o ∈ P, ¬acc(pow, o, t)∧∃v ∈ o, ¬acc(pow, v, t)).
StateValue(v)
16 ProposeValue(v) ∃i ∈ C, ∃v ∈ Ci , v ∈ Ai ∧ acc(pow, v, t)
17 ProposeOption(o) ∀i ∈ C, ∃v ∈ Ci , v ∈ Ti ∧ v ∈ o
18 AskValue(v) t > τ ∧ ∃i ∈ C, ∃c ∈ Pi , ¬acc(c, t)
19 AskCriterion(i) ∃i ∈ C, Ai ∪ Ui = ∅
20 StateValue(v) ∃i ∈ C, Ci ∩ Si = ∅
21 ProposeValue(v) ∃v ∈ Ci / tol(v, t, ≺i , Ai , Ui , pow)
22 ProposeOption(o) ∃o ∈ O / tol(o, t, ≺i , Ai , Ui , pow)

Table 2: Selection of utterance types

4.1 Study design

We simulate a collaborative negotiation for choosing a restaurant using four

criteria (cuisine, ambiance, price and location) for a total of 420 options. An
example of dialogue is given in the figure 2. The following parameter values were
used in our simulation: τ = 2, π = 0.5, α = 2, β = 1 and δ = 0.1. We generated
three preferences sets and we measured the difference between the preference
sets using Kendall’s distance [2]. We manipulated two simulation parameters:
the power of both agents (named pow-a and pow-b) and the preference sets.
This later affects the generation of dialogues in term of values and length. Table
3 summarizes the 4 experimental conditions that results from this combination.
Note that we only consider one configuration of social power for the similar
preference sets condition, because the produced dialogues with other values of
power are all very similar in this case. The first speaker (Speaker A) is always
the high-power agent, as stated by our third principle of leading the dialogue.
Our goal is to show these dialogues to human observers so as to evaluate how
the relation of power is perceived in the different dialogues.
268 L. Ouldouali et al.

Preferences A B Label
0.9 0.4 Dialogue 1
Different preferences (Kendall’s tau = 0.96) 0.7 0.4 Dialogue 2
0.7 0.2 Dialogue 3
Similar preferences (Kendall’s tau = 0.46) 0.7 0.4 Dialogue 4

Table 3: Initial condition’s setting for generating dialogues

A: "Let’s go to a Chinese restaurant."

B: "I don’t like Chinese restaurants, let’s choose something else."
A: "Let’s go to a cheap restaurant."
B: "Do you like expensive restaurants?"
A: "I don’t like expensive restaurants."
...
B: "What kind of atmosphere do you like?"
A: "Let’s go to a cheap restaurant."
B: "Okay, let’s go to a cheap restaurant."
A: "Let’s go to Sap. It’s a quiet, cheap Japanese restaurant on the south side."
B: "Okay, let’s go to Sap.

Fig. 2: Excerpt of Dialogue 2.

4.2 Hypotheses

Based on our three principles and the literature on the perception of social power
in negotiation, we investigated four hypotheses:

– H1: The higher-power speaker will more strongly be perceived as self-centered

than the lower-power speaker.
– H2: The lower-power speaker will be more strongly perceived as making
larger concessions than the higher-power speaker.
– H3: The higher-power speaker will more strongly be perceived as demanding
than the lower-power speaker.
– H4: The higher-power speaker will more strongly be perceived as taking the
lead in the negotiation than the lower-power speaker.

4.3 Experimental Procedure

We conducted a between-subject study using the online crowdsoursing website

CrowdFlower 4 . Each participant was shown only one dialogue. Speaker A and
B were described as two friends trying to negotiate a restaurant to have dinner.
Participants were asked to read the assigned dialogue and answer a questionnaire.
We defined two questions for each hypothesis. Two test questions were in-
cluded to check the sanity of the answers. We eliminated participants providing
wrong answers to those questions. Each one of these questions was to be an-
swered on a 5 points Likert scale ranging from “I totally disagree” to “I totally
agree”.
4
https://www.crowdflower.com/
 A computational model of power in collaborative negotiation dialogues 269

Hypothesis question 1 question 2

H1 Speaker (A/B) is self-centered. Speaker (A/B) takes his friend’s preferences
into account in the choice of the restaurant.
H2 Speaker (A/B) makes concessions in the Speaker (A/B) gives up his position in the
negotiation. negotiation
H3 Speaker (A/B) is demanding Speaker (A/B) presses his position in the ne-
gotiation.
H4 Speaker (A/B) takes the lead in the ne- Speaker (A/B) takes the initiative in the ne-
gotiation. gotiation

Table 4: Questions asked on the questionnaire.

A total of 120 native English subjects participated in the experiment (30 for
each condition). Each subject received 25 cents and we excluded 15 participants
because they failed the sanity check.

4.4 Results and discussion

Table 5 summarizes the results of our study, which strongly support all of our
four hypotheses. The higher-power speaker was correctly perceived as more self-
centered (H1), making less concessions (H2), more demanding (H3) and leading
the negotiation (H4).
We first computed the correlation for each pair of questions corresponding
to a given hypothesis (the average correlation is .5). Since there is a strong
correlation, we can use the data to compare the speakers’ behaviors on each
dialogue. Note that our data are not normally distributed. This is the reason
why we used a non-parametric Wilcoxon signed-rank test to verify that our
paired sets of data significantly different from each other.
The higher-power speaker was largely perceived as more self-centered, as
assumed by hypothesis H1, with a large effect size. For instance, if you consider
dialogue 1 on Table 5, the statistical test indicates that the higher-power agent
was significantly perceived as more self-centered with (Z=-5.28 and p<0.001 ).
In addition, a significant difference in the level of concessions expressed in all
the dialogues was revealed (H2). Indeed, the high-power agent was perceived as
making less concessions. The effect size showed a medium effect for dialogues 2
to 4, and a large effect for Dialogue 1 with (Z=-5.34 and p<0.001 ).
Hypothesis H3 was also confirmed by the Wilcoxon signed-rank test, where
the higher-power agent was defined as the most demanding negotiator, with a
large effect size observed for all the dialogues.
The (H4) hypothesis was confirmed. The Wilcoxon ranked test revealed that
the high-power agent was perceived as significantly more leading the dialogue
than the low-power agent, with a large effect size for the dialogues 1, 2 and 4,
and a medium effect size for Dialogue3 as depicted in Table 5.
Finally, we made a post-study analysis by comparing the participants’ judg-
ments on the behaviors of Speaker A across different dialogues. We computed
the differences between the evaluations of Speakers A and B in Dialogue 1 and
Dialogue 2 (power of Speaker B remains unchanged at 0.4 whereas the power of
Speaker A changes from 0.7 to 0.9).
270 L. Ouldouali et al.

Dialogue1 Dialogue2 Dialogue3 Dialogue4

SpeakerA SpeakerB SpeakerA SpeakerB SpeakerA SpeakerB SpeakerA SpeakerB
Mean ± SD 3.9 ± 1.1 2.2± 0.9 3.6 ±0.9 2.2 ±0.8 2.8 ±1.1 2.13± 0.7 3.4 ± 1 2 ±0.9
p-value 9.75E−08 5.14E−08 0.002 6.23E−08
H1
Z-Wilcoxon test −5.28 −5.34 −3 −4.93
Effect size 0.51 0.52 0.3 0.47
Mean ± SD 2.2 ± 1.1 4.3± 0.8 2.5 ±1.2 3.8 ±1.04 2.7 ±1.2 3.6± 0.8 2.3 ± 1 3.2 ±1.2
p-value 7.07E−08 3.71E−05 = 0.01 1.73E−04
H2
Z-Wilcoxon test −5.34 −4.05 −3.13 −3.69
Effect size 0.52 0.39 0.32 0.35
Mean ± SD 4.1 ± 0.8 2.6± 1.1 4.03 ± 0.8 2.7 ±0.9 3.5 ±1.1 2.3± 1 3.8 ± 1.8 1.8 ±0.8
p-value 2.93E−08 4.77E−07 1.19E−04 2.56E−09
H3
Z-Wilcoxon test −4.62 −4.96 −3.80 −5.86
Effect size 0.45 0.49 0.39 0.56
Mean ± SD 4.2 ± 0.9 2.3± 1.1 3.8 ±0.9 2.6 ±1.07 3.8 ±0.9 2.8± 1.1 4.5 ±0.5 1.9 ± 0.9
p-value 2.44E−07 3.28E−05 0.03 7.04E−10
H4
Z-Wilcoxon test −5.11 −4.08 −2.86 −6.09
Effect size 0.5 0.4 0.29 0.57

Table 5: Summary of the obtained results for each hypothesis

Our hypothesis was that a greater difference in power would lead to a better
perception of behaviors related to power. However, power in the dialogue is
interpersonal by nature, which means that participants rate the power of Speaker
A as opposed to the behavior of Speaker B. For this reason, aggregating the
evaluations from different dialogues does not make sense. This explains why we
obtained mixed results on this aspect. Indeed, a tendency was observed (p 
0.1) for self-centeredness, concessions and the level of demand. Only the lead of
dialogue was clearly better perceived (p = 0.043) when the power increases.

Also note that in this experiment, we studied the perception of all the princi-
ples related to power simultaneously. This could be seen as a limit of this study:
we did not investigate the perception of each principle individually. However,
during previous experiments, we detected that the principles are interdepen-
dent, which makes a separate evaluation of each of them in a dialogue difficult
to perform.

We focus in this study on situations where agents have a complementary

relation of power (A high-power side and low-power side). We did not study
situations in which both agents are high-powered or low-powerful. The reason is
that we want to assess whether the social power is perceived by human observers.
This supposes that we aim to analyze the relation of dominance, as defined by
[4] in the context of collaborative negotiation. Indeed, [4] defines the relation
of dominance as the ability to exert behaviors of power. It refers to context
and relationship-dependent interactional patterns in which one actors assertion
of control is met by acquiescence from another. Therefore, in the relation of
dominance, one actor plays the dominant role and expresses high-powered’s be-
haviors whereas, the other actor plays the submissive role and exerts behaviors
of low-power individual.
A computational model of power in collaborative negotiation dialogues 271

4.5 Conclusion

Our research aims to model a conversational agent which is able to deploy dif-
ferent strategies of negotiation depending on its representation of social power.
Based on research in social psychology, we defined three behavioral principles
related to power in collaborative negotiation. We proposed a model of utter-
ance selection based on modeling of preferences and the implementation of these
principles. We showed that the behaviors related to social power are correctly
perceived in the resulting dialogues. Our findings validate our model of dialogue
in general and specifically confirmed the coherence of the generated behaviors
of power.
Future works will focus on using this model in a human-agent interaction.
It was shown by [15] that having a model of the other impacts the negotiation
strategy and improves the outcomes. Therefore, we aim to use our dialogue model
to build a representation of the interlocutor’s social power, following a theory of
mind approach. We hope to show that an agent that adapts its own strategy to
the perceived power of its interlocutor will be a better collaborative negotiator.

References

1. Bee, N., Pollock, C., André, E., Walker, M.: Bossy or wimpy: expressing social
dominance by combining gaze and linguistic behaviors. In: Proc. IVA 2010. pp.
265–271 (2010)
2. Brandenburg, F.J., Gleißner, A., Hofmeier, A.: Comparing and aggregating partial
orders with kendall tau distances. Discrete Mathematics, Algorithms and Applica-
tions 5(02), 1360003 (2013)
3. Broekens, J., Jonker, C.M., Meyer, J.J.C.: Affective negotiation support systems.
Journal of Ambient Intelligence and Smart Environments 2(2), 121–144 (2010)
4. Burgoon, J.K., Dunbar, N.E.: Nonverbal expressions of dominance and power in
human relationships. The Sage handbook of nonverbal communication. Sage 2
(2006)
5. Chu-Carroll, J., Carberry, S.: Response generation in collaborative negotiation.
In: Proceedings of the 33rd annual meeting on Association for Computational
Linguistics. pp. 136–143. Association for Computational Linguistics (1995)
6. De Dreu, C.K., Van Kleef, G.A.: The influence of power on the information search,
impression formation, and demands in negotiation. Journal of Experimental Social
Psychology 40(3), 303–319 (2004)
7. De Dreu, C.K., Van Lange, P.A.: The impact of social value orientations on nego-
tiator cognition and behavior. Personality and Social Psychology Bulletin 21(11),
1178–1188 (1995)
8. Dunbar, N.E., Burgoon, J.K.: Perceptions of power and interactional dominance
in interpersonal relationships. Journal of Social and Personal Relationships 22(2),
207–233 (2005)
9. Fiske, S.T.: Controlling other people: The impact of power on stereotyping. Amer-
ican psychologist 48(6), 621 (1993)
10. Galinsky, A.D., Gruenfeld, D.H., Magee, J.C.: From power to action. Journal of
personality and social psychology 85(3), 453 (2003)
272 L. Ouldouali et al.

11. Gebhard, P., Baur, T., Damian, I., Mehlmann, G., Wagner, J., André, E.: Explor-
ing interaction strategies for virtual characters to induce stress in simulated job
interviews. In: Proc. AAMAS 2014. pp. 661–668 (2014)
12. Giebels, E., De Dreu, C.K., Van De Vliert, E.: Interdependence in negotiation:
Effects of exit options and social motive on distributive and integrative negotiation.
European Journal of Social Psychology 30(2), 255–272 (2000)
13. Kecskés, I., Romero-Trillo, J.: Research trends in intercultural pragmatics, vol. 16.
Walter de Gruyter (2013)
14. Kerly, A., Ellis, R., Bull, S.: Calmsystem: a conversational agent for learner mod-
elling. Knowledge-Based Systems 21(3), 238–246 (2008)
15. Klatt, J., Marsella, S., Krämer, N.C.: Negotiations in the context of aids prevention:
an agent-based model using theory of mind. In: Proc. IVA 2011. pp. 209–215 (2011)
16. Lance, B., Marsella, S.C.: The relation between gaze behavior and the attribution
of emotion: An empirical study. In: Proc. IVA 2008. pp. 1–14 (2008)
17. Magee, J.C., Galinsky, A.D., Gruenfeld, D.H.: Power, propensity to negotiate, and
moving first in competitive interactions. Personality and Social Psychology Bulletin
33(2), 200–212 (2007)
18. de Melo, C.M., Gratch, J., Carnevale, P.J.: Humans versus computers: Impact of
emotion expressions on people’s decision making. IEEE Transactions on Affective
Computing 6(2), 127–136 (2015)
19. Mignault, A., Chaudhuri, A.: The many faces of a neutral face: Head tilt and
perception of dominance and emotion. Journal of nonverbal behavior 27(2), 111–
132 (2003)
20. Nazari, Z., Lucas, G.M., Gratch, J.: Opponent modeling for virtual human ne-
gotiators. In: International Conference on Intelligent Virtual Agents. pp. 39–49.
Springer (2015)
21. Rich, C., Sidner, C.L.: Robots and avatars as hosts, advisors, companions, and
jesters. AI Magazine 30(1), 29 (2009)
22. Sidner, C.L.: An artificial discourse language for collaborative negotiation. In:
AAAI. vol. 94, pp. 814–819 (1994)
23. Swartout, W.R., Gratch, J., Hill Jr, R.W., Hovy, E., Marsella, S., Rickel, J., Traum,
D., et al.: Toward virtual humans. AI Magazine 27(2), 96 (2006)