Dialogue Management in the Agreement Negotiation Process: 
A Model that Involves Natural Reasoning 
Mare KOIT 
Institute of Computer Science, Tartu University 
Liivi 2 
50409 Tartu, Estonia 
koit@ut.ee 
Haldur ~)IM 
Dept. of General Linguistics, Tartu University 
Tiigi 78 
51014 Tartu, Estonia 
hoim@psych.ut.ee 
Abstract 
In the paper we describe an approach to 
dialogue management in the agreement 
negotiation where one of the central roles is 
attributed to the model of natural human 
reasoning. The reasoning model consists of 
the model of human motivational sphere, 
and of reasoning algorithms. The reasoning 
model is interacting with the model of 
communication process. "/'he latter is 
considered as rational activity where central 
role play the concepts of communicative 
strategies and tactics. 
Introduction 
Several researches have modelled the process of 
argument negotiation in cooperative dialogue 
where one participant makes a proposal to 
another participant and as the result of 
negotiation this is accepted or rejected. 
Chu-Carroll and Carberry (1998) present a 
cooperative response-generation model as a 
recursive cycle Propose-Evaluate-Modify. They 
concentrate on dialogues of information sharing 
and negotiation. An information sharing 
dialogue is started, when the agent recognised a 
turn of his/her partner as a proposal, but does not 
have enough information to decide whether to 
accept it or not. A negotiation dialogue is 
started, when the agent concludes that the 
proposal is in conflict with his/her beliefs and 
preferences, i.e. tends to reject it. 
Heeman and Hirst (1995) model cooperation by 
the cycle Present-Judge-Refashion. They use 
two levels of modelling - planning and 
cooperation. On the first level utterances are 
generated and interpreted, on the second level 
the cooperation of agents is modelled, relating it 
to agent's mental states and planning processes. 
The Shared Plans cooperation model deals with 
planning processes in which participate multiple 
agents, see Lochbaurn (1998). The model 
concentrates on group tasks that can be divided 
into separate, but interacting subtasks, and the 
central problem is coordination of intentions and 
goals of partners. 
Di Eugenio et al. (2000) present a model 
BalanceProposeDispose: first, the relevant 
information concerning the task is considered 
and discussed, then a proposal is made and, 
lastly, the decision concerning the proposal is 
made - it is accepted or rejected. 
In our model we depart from the same type of 
situation. One agent, A, addresses another agent, 
B, with the intention that B will carry out an 
action D. After some negotiation, B agrees or 
rejects the proposal. 
In this paper we concentrate on the problems 
connected with modelling participants as 
conversation agents who are able to participate 
in negotiation in the form of natural dialogue - 
dialogue that is carried out in natural  
and according to the rules of human 
communication. 
Such a dialogue can be considered as rational 
behaviour which is based on beliefs, wants and 
intentions of agents, at the same time being 
restricted by their resources, see Jokinen (1995), 
Webber (2000). Conversation agent is a kind of 
intelligent agent - a computer program that is 
able to communicate with humans as another 
human being. 
As it is generally accepted, in a model of 
conversation agent it is necessary to represent its 
cognitive states as well as cognitive processes. 
102 
One of the most well-known models of this type 
is the BDI model, see Allen (1994). 
Our main point in this paper is that the general 
concepts of cognitive states and processes used 
in BDI-type models should be extended in order 
to include certain factors from human 
motivational sphere and certain social principles 
in order to guarantee naturalness of dialogues of 
the type we are concerned with. This is 
especially important in connection with the fact 
that interest in modelling cooperative dialogues 
where partners are pursuing a common goal has 
considerably increased in recent years. On the 
one hand, this is connected with rapid spreading 
of Internet-based services. On the other hand, 
the interest in models of full natural dialogue 
derives from the possibility of building speech 
interfaces with different knowledge and 
databases, see Dybkjaer (2000). Both of these 
developments broaden the concept of 
naturalness of dialogue considerably and present 
to it much stronger requirements concerning its 
empirical adequacy as it has been generally 
accepted thus far. 
1 Model of Conversation Agent 
In our model a conversation agent, A, is a 
program that consists of 6 (interacting) modules: 
A = (PL, PS, DM, INT, GEN, LP), 
where PL - planner, PS - problem solver, DM - 
dialogue manager, INT - interpreter, GEN - 
generator, LP - linguistic processor. PL directs 
the work of both DM and PS, where DM 
controls communication process and PS solves 
domain-related tasks. The task of INT is to make 
semantic analysis of partner's utterances and 
that of GEN is to generate semantic 
representations of agent's own contributions. LP 
carries out linguistic analysis and generation. 
Conversation agent uses in its work goal base 
GB and knowledge base KB. In our model, KB 
consists of 4 components: 
KB = (KBw, KBL, KBD, KBs), 
where KBw contains world knowledge, KBL - 
linguistic knowledge, KBD - knowledge about 
dialogue and KBs - knowledge about interacting 
subjects. For instance, KBD contains definitions 
of communicative acts, turns and transactions 
(declarative knowledge), and algorithms that are 
applied to reach communicative goals - 
communicative strategies and tactics (procedural 
knowledge); KBs contains knowledge about 
evaluative dispositions of participants towards 
the world (e.g. what do they consider as pleasant 
or unpleasant, useful or harmful), and, on the 
other hand, algorithms that are used to generate 
plans for acting on the world. 
A necessary precondition of a communicative 
interaction is existence of shared (mutual) 
knowledge of interacting agents. This concerns 
goal bases as well as all types of knowledge 
bases; the intersections of the corresponding 
bases of interacting agents A and B cannot be 
empty: GB g n GB a ~:~, KBAw n KBBw ~, 
KBAL n KBBL ~O, KI3AD n KBBD ~, KB ABS 
KBBs :.7~:~, KBBAs ("h KBAs -~:~. 
In this paper we will consider a specific type of 
dialogue where the communicative goal of agent 
A is to get agent B to agree to carry out an 
action D - so-called agreement negotiation 
dialogue. We will concentrate here on dialogue 
management in such kind of interaction, i.e. on 
the functioning of the module DM. 
2 Dialogue Management 
2.1 Reasoning Model 
A dialogue participant chooses his/her responses 
to the parter's communicative acts as a result of 
certain reasoning process. After A has made B a 
proposal to do D, B can respond with agreement 
or rejection, depending on the result of his/her 
reasoning. 
Because we consider the model of natural 
human reasoning as one of the important 
components in attaining naturalness of dialogue 
as a whole, we will discuss our model of 
reasoning in some detail. From the point of view 
of practical NLP the approach we will present 
below may seem too abstract. But without solid 
theoretical basis it will appear impossible to 
guarantee naturalness of dialogues carried out by 
computers with human users. We think that the 
model we describe here can be taken as a basis 
for the corresponding discussion. 
Our model is not based on any scientific theory 
of how human reasoning proceeds; our aim is to 
model a "naive theory of reasoning" which 
humans follow in everyday life when trying to 
understand, predict and influence other persons' 
decisions and behavior, see Koit and C)im 
(2000). The reasoning model consists of two 
functionally linked parts: 1) a model of human 
motivational sphere; 2) reasoning schemes. 
103 
In the motivational sphere three basic factors 
that regulate reasoning of a subje, ct concerning D 
are differentiated. First, subject may wish to do 
D, if pleasant aspects of D for him/her 
overweight unpleasant ones; .second, subject 
may fmd reasonable to do D, if D is needed to 
reach some higher goal, and useful aspects of D 
overweight harmful ones; and tlfird, subject can 
be in a situation where he/she must (is obliged) 
to do D - if not doing D will lead to some kind 
of punishment. We call these; factors wish-, 
needed- and must-factors, respectively. 
For instance, in reasoning about some action D 
(e.g. proposed by another agent), an agent as an 
individual subject typically starts with checking 
his/her wish-factor, i.e. whether D's pleasant 
aspects overweight unpleasant ones. If this 
holds, then the subject checks his/her resources, 
and if these exist, proceeds to other positive and 
negative aspects of D: its usefulness and 
harmfulness, and if D is prohibited, then also 
possible punishment(s). If the positive aspects in 
sum overweight negative ones, the resulting 
decision will be to do D, otherwise - not to do 
D. 
There can exist other typical situations. If the 
agent is an "official" person, or a group of 
subjects formed to fulfil certain tasks and/or to 
pursue certain pre-established goal(s), then 
typically the starting point of reasoning is 
needed- and/or mast-factor. 
This means that there exist certain general 
principles that determine how the reasoning 
process proceeds. These principles depend, in 
part, on the type of the reasoning agent. Before 
starting to construct a concrete reasoning model 
the types of agents involved should be 
established. In our implementation the agent is 
supposed to be a "simple" human being and the 
actions under consideration are from everyday 
life. In this case as examples of such principles 
used in our model we can present the following 
ones. For more details, see Oim (1996). 
P1. People prefer pleasant (more pleasant) 
states to unpleasant (less pleasant) ones. 
P2. People don "t take an action of which they 
don't assume that its consequence will be a 
pleasant (useful) situation, or avoidance of an 
unpleasant (harmful) situation. 
The following principles illustrate more concrete 
(operational) rules. 
P3. In assessing an action D the values of 
(internal - wish- and needed-)factors are 
checked before the external (must-) factors. 
P4. If D is found pleasant enough (i.e. D's 
pleasant aspects overweight unpleasant ones), 
then the needed- and must-factors will first be 
checked from the point of view of their negative 
aspects ("to what harmful consequences or 
punishments D would lead? "~). 
The rule P4 explains, for example, why in 
Figure 1 step 1 is immediately followed by step 
2. 
The weights of different aspects of D 
(pleasantness, unpleasantness, usefulness, 
harmfulness, punishment for doing a prohibited 
action or not-doing an obligatory action) must be 
summed up in some way. Thus, in a 
computational model weights must have 
numerial values. In reality people do not operate 
with numbers but, rather, with some fuzzy sets. 
On the other hand, existence of certain scales 
also in human everyday reasoning is apparent. 
For instance, for the characterisation of pleasant 
and unpleasant aspects of some action there are 
specific words: enticing, delighOCul, enjoyable, 
attractive, acceptable, unattractive, displeasing, 
repulsive etc. Each of these adjectives can be 
expressed quantitatively. This presupposes 
empirical studies, though. 
We have represented the model of motivational 
sphere by the following vector of weights: 
w A = (w(resourcesAol), w(pleasAm), 
w(unpleasAm), w(useAm), w(harmAm), 
w(obligatoryAm), w(prohibitedAm), w(punishgm), w(punishAnot.Di),..., 
w(resourcesADn), w(pleasAon), w(unpleasADn), 
W(useAD~), w(harmAo,), w(obligatoryAD,), 
w(prohibitedADn), w(punishAm), 
W(punishAnot.Dn)). 
Here D~, ..., Dn represent human actions; 
W(resourcesADi)=I, if A has resources necessary 
to do Di (otherwise 0); w(obligatoryAsi)=l, if Di 
is obligatory for A (otherwise 0); 
w(prohibitedADi)=l, if Di is prohibited for A 
(otherwise 0). The values of other weights are 
non-negative natural numbers. 
The second part of the reasoning model consists 
of reasoning schemas, that supposedly regulate 
human action-oriented reasoning. A reasoning 
scheme represents steps that the agent goes 
through in his reasoning process; these consist in 
computing and comparing the weights of 
104 
different aspects of D; and the result is the 
decision to do or not to do D. 
Figure 1 presents the reasoning scheme that 
departs from the wish of a subject to do D. 
The scheme also illustrates one of the general 
principles referred to above. It explains the order 
the steps are taken by the reasoning agent: if a 
subject is in a state where he/she wishes to do D, 
then he/she checks first the harmful/useful 
aspects of D, and after this proceeds to aspects 
connected with possible punishments. 
Presupposition: 
w(pleas) > w(unpleas). 
i) 
for doing D? 
If not then not to do D. 
2) Is w(pleas) > w(unpleas) + 
w(harm)? 
If not then go to step 6. 
3) Is D prohibited? 
If not then to do D. 
4) Is w(pleas) > w(unpleas) + 
w(harm) + w(punishD)? 
If yes then to do D. 
5) Is w(pleas) + w(use) > 
w(unpleas) + w(harm) + 
w(punish~)? 
If yes then to do D else 
not to do D. 
6) Is w(pleas) + w(use) > 
w(unpleas) + w(harm)? 
If yes then go to step 9. 
7) Is D obligatory? 
If not then not to do D. 
8) Is w (pleas) + w (use) + 
w (punishnot. ~) > w (unpleas) + 
w (harm) ? 
If yes then to do D else 
not to do D. 
9) Is D prohibited? 
If not then to do D. 
i0) Is w (pleas) + w(use) > 
w (unpleas) + w (harm) + 
w (punish~) ? 
Are there enough resources 
If yes then to do D else 
not to do D. 
Figure 1. The reasoning procedure that departs 
from the wish of a subject to do D. 
The prerequisite for triggering this reasoning 
procedure is w(pleas) > w(unpleas), which is 
based on the following assumption: if a person 
wishes to do something, then he/she assumes 
that the pleasant aspects of D (including its 
consequences) overweigh its unpleasant aspects. 
The same kinds of reasoning schemes are 
constructed for the needed- and must-factors. 
The reasoning model is connected with the 
general model of conversation agent in the 
following way. First, the planner PL makes use 
of reasoning schemes and second, the KBs 
contains the vector w A (A's subjective 
evaluations of all possible actions) as well as 
vectors w AB (A's beliefs concerning B's 
evaluations, where B denotes agents A may 
communicate with). The vector w As do not 
represent truthful knowledge, it is used as a 
partner model. 
When comparing our model with BDI model, 
then belier are represented by knowledge of the 
conversation agent with reliability less than 1; 
desires are generated by the vector of weights 
WA; and intentions correspond to goals in GB. In 
addition to desires, from the weights vector we 
also can derive some parameters of the 
motivational sphere that are not explicitly 
covered by the basic BDI model: needs, 
obligations and prohibitions. Some wishes or 
needs can be stronger than others: if w(pleasADi) 
- W(unpleasAoi) > w(pleasAoj) - w(unpleasAt~), 
then the wish to do Di is stronger than the wish 
to do Dj. In the same way, some obligations 
(prohibitions) can be stronger than others, 
depending on the weight of the corresponding 
punishment. It should be mentioned that adding 
obligations to the standard BDI model is not 
new. Traum and Allen (1994) show how 
discourse obligations can be used to account in a 
natural manner for the connection between a 
question and its answer in dialogue and how 
obligations can be used along with other parts of 
the discourse context to extend the coverage of a 
dialogue system. 
2.2 Communicative Strategies and 
Tactics 
Knowledge about dialogue KBD, which is used 
by the Dialogue Manager, consists of two 
functional parts: knowledge of the regularities of 
dialogue, and rules of constructing and 
combining speech acts. 
The top level concept of dialogue rules in our 
model is communicative strategy. This concept 
is reserved for such basic communication types 
as information exchange, directive dialogue, 
105 
phatic communication, etc. On the more 
concrete level, the conversation agent can realise 
a communicative strategy by means of several 
communicative tactics; this concept more 
closely corresponds to the: concept of 
communicative strategy as us~l in some other 
approaches, see e.g. Jokinen (1996). In the case 
of directive communication (which is the 
strategy we are interested in) the agent A can use 
tactics of enticing, persuading, threatening. In 
the case of enticing, A stresses pleasant aspects, 
in the case of persuading - usel~ aspects of D 
for B; in the case of ordering A addresses 
obligations of B, in the case of threatening A 
explicitly refers to possible punishment for not 
doing D. 
Which one of these tactics A chooses depends 
on several factors. There is one: relevant aspect 
of human-human communication which is 
relatively well studied in pragmatics of human 
communication and which we have included in 
our model as the concept of communicative 
space. 
Communicative space is defined by a number of 
coordinates that characterise the relationships of 
participants in a communicative encounter. 
Communication can be collaborative or 
confrontational, personal or impersonal; it can 
be characterised by the social distance between 
participants; by the modality (friendly, ironic, 
hostile, etc.) and by intensity (peaceful, 
vehement, etc.). Just as in case of motivations of 
human behaviour, people have an intuitive, 
"naive theory" of these coordinates. This 
constitutes a part of the social conceptualisation 
of communication, and it also should not be 
ignored in serious attempts to model natural 
communication in NLP systems. 
In our model the choice of a communicative 
tactics depends on the "point" of the 
communicative space in which the participants 
place themselves. The values of the coordinates 
are again given in the form of numerical values. 
The communicative strategy can be presented as 
an algofithra (Figure 2). 
Figure 3 presents a tactic of enticement. 
In our model there are three different 
communicative tactics that A can use within the 
frames of the directive communicative strategy: 
those of enticement, persuasion and threatening. 
Each communicative tactic constitutes a 
procedure for compiling a turn in the ongoing 
dialogue. 
i) Choose the communicative 
tactic. 
2) Implement the tactic to 
generate an expression (inform 
the partner of the communicative 
goal). 
3) Did the partner agree to do 
D? If yes then finish (the 
communicative goal has been 
reached). 
4) Give up? If yes then finish 
(the communicative goal has not 
been reached). 
5) Change the communicative 
tactic? If yes then choose the 
new tactic. 
6) Implement the tactic to 
generate an expression. Go to 
step 3. 
Figure 2. Communicative strategy used by the 
initiator of communication. 
i) If wB(resources)=0 then 
present a counterargument in 
order to point at the presence 
of possible resources or at the 
possibility to gain them. 
2) If w s(harm) > w as(harm) then 
present a counterargument in 
order to downgrade the value of 
harm. 
3) If wB(obligatory)=l & 
w B (punish .... o) < w~ (punishno~-~) then 
present a counterargument in 
order to decrease the weight of 
the punishment. 
4) If wB (prohibited) =l & 
w ~(punis~) > w ~(punis~) then 
present a counterargument in 
order to downgrade the weight of 
the punishment. 
5) If wB(unpleas) > w~(unpleas) 
then 
present a counterargument in 
order to downgrade the value of 
the unpleasant aspects of D. 
6) Present a counterargument in 
order to stress the pleasant 
aspects of D. 
Figure 3. A's tactics of enticement 
106 
The tactic of enticement consists in increasing 
B's wish to do D; the tactic of persuasion 
consists in increasing B's belief of the usefulness 
of D for him/her, and the tactic of threatening 
consists in increasing B's understanding that 
he/she must do D. 
Communicative tactics are directly related to the 
reasoning process of the partner. IrA is applying 
the tactics of enticement he/she should be able to 
imagine the reasoning process in B that is 
triggered by the input parameter wish. If B 
refuses to do D, then A should be able to guess 
at which point the reasoning of B went into the 
"negative branch", in order to adequately 
construct his/her reactive turn. 
Analogously, the tactic of persuasion is related 
to the reasoning process triggered by the needed- 
parameter, and the threatening tactic is related to 
the reasoning process triggered by the must- 
parameter. For more details see, for example, 
Koit (1996), Koit and 0im (1998), Koit and Oim 
(1999). 
Thus, in order to model various communicative 
tactics, one must know how to model the process 
of reasoning. 
2.3 Speech Acts 
The minimal communicative unit in our model is 
speech act (SA). In the implementation we make 
use of a limited number of SAs the 
representational formalism of which is flames. 
Figure 4 presents the frame of SA Proposal in 
the context of co-operative interaction. Other 
SAs are represented in the same form. Each SA 
contains a static (declarative) and a dynamic 
(procedural) part. The static part consists of 
preconditions, goal, content (immediate act) and 
consequences. The dynamic part is made up 
from two kinds of procedures: 1) those that the 
author of the SA applies in the generation of a 
communicative turn that contains the given SA; 
2) those that the addressee applies in the process 
of response generation. 
As one can see, such a two-part representation 
contains also rules for combining SAs in a turn, 
and on the other hand, guarantees coherence of 
turn-takings: when we have tagged in KBD 
initiating SAs (such as Question or Proposal), 
then the following chain of SAs follows from 
the interpretation-generation procedures as 
applied by participants. 
PROPOSAL (author A, recipient B, 
A proposes B to do an action D) 
I. Static part 
SETTING 
(i) A has a goal G 
(2) A believes that B in the 
same way has the goal G 
(3) A believes that in order to 
reach G an instrumental goal 
G i should be reached 
(4) A believes that B in the 
same way believes that in 
order to reach G an 
instrumental goal G i should be 
reached 
(5) A believes that to attain 
the goal Gi B has to do D 
(6) A believes that B has 
resources for doing D 
(7) A believes that B will 
decide to do D 
GOAL: B decides to do D 
CONTENT: A informs B that 
he/she wishes B to do D 
CONSEQUENCES 
(i) B knows the SETTING, GOAL 
and CONTENT 
(2) A knows that B knows the 
SETTING, GOAL and CONTENT 
II. Dynamic part 
Generating procedures (A's 
possibilities to build his/her 
turn that contains Proposal as 
the dominant SA). 
A has Goal G; A believes that B 
also has Goal G; A believes that 
in order to reach G, Gi should be 
reached; A has decided to 
formulate this as Proposal to B 
to do D. 
Procedures (before formulating 
the turn) consist in checking 
whether the preconditions of 
proposal hold and in making 
decisions about information to 
be added in the turn: 
- in case of (2) : is G 
actualised in B? If not, then 
actualise it by adding SA 
Inform; 
- in case of (4) : does B 
believe that in order to 
reach B, G~ should be reached 
first. If not, then add SA 
Explanation (Argument); 
- in case of (6) : if A is not 
sure that B has resources for 
D, then add Question; 
- in case of (7) : if A is not 
sure that B will agree to do 
107 
D (for this A should model 
B's reasoning) , then add 
Argument. 
Procedures of interpretation- 
generation 
(B's possibilities to react to 
proposal) are started after B 
has recognised SA Proposal: 
- in case of (2), (4), (5) : if 
B does not have Goal G and/or 
he/she does not haw~ the 
corresponding beliefs and A 
has not provided the needed 
additional information, then 
add Question (ask for 
additional information) ; 
- in case of (6) : if B does 
not have Resources for D, 
then Reject + Argument; 
- in case of (7) : if the 
decision of B to do D (as the 
result of the application of 
reasoning scheme (s)) is 
negative, then Reject + 
Argument. 
Figure 4. Speech act Proposal in the context of 
co-operative interaction. 
Such a representation does not guarantee 
coherence of dialogic encounters (transactions) 
on a more general level. For instance, it does not 
cover such phenomena as topic change, 
inadequate responses caused by 
misunderstandings; but, more importantly, also 
various kinds of initiative overtakings. For 
instance, after rejecting the Proposal made by A, 
B can, in addition to explaining the rejection by 
Argument, initiate various "compensatory" 
communicative activities. Such things are 
normal in human co-operative interaction and 
they are regulated by general pragmatic 
principles that require from participants, in 
addition to being co-operative and informative, 
also being considerate and helpful. In our case 
this means that KBo should also include general 
level dialogue scenarios (in the form of a graph) 
and formalisations of the mentioned pragmatic 
principles; for an example of the latter, see 
Jokinen (1996). 
3 Process of Dialogue 
Let us describe the case where both A and B are 
intelligent agents; i.e. computer programs. 
1. A constructs 
a) the frame exemplar of D, putting in it all 
relevant information A has about D; 
b) the model of partner B, putting in it all 
relevant information it has about B's 
evaluations concerning the contents of the 
slots in D's frame. 
2. A chooses the point in communicative space 
from which it intents to start the interaction. 
3. A starts to apply communicative strategy. A 
models B's reasoning process, using B's 
model. First A applies the reasoning scheme 
based on the wish of B. If it results in 'to do 
D', then A actualises the tactic of enticing 
and generates its first turn which contains a 
frame exemplar of Proposal. If the result of 
modelled reasoning results in 'not to do D', 
then A tries reasoning which starts from 
needed-factor and then the one triggered by 
must-factor, and according to the result 
actualises tactics of persuading or 
threatening, and generates the first utterance. 
If the application of all reasoning schemes 
results in 'not to do D', then A abandons its 
goal. 
4. B interprets A's turn and recognises 
Proposal in it. B constructs it's the exemplar 
representation of D (this may not coincide 
with that of A). B starts reasoning, in the 
course of which it may need additional 
information from A. On the basis of the 
frame of Proposal B formulates the result of 
reasoning as its response turn: yes/no + 
(maybe) Argument. 
5. A interprets B's answer and determines 
which point in the dialogue scenario this 
corresponds to. If B's answer was positive 
(decision to do D), then according to 
communicative strategy the encounter has 
come to its successful end. If B's answer is 
negative, then according to the dialogue 
scenario A must formulate a 
(counter)Argument. The communicative 
strategy also allows to choose a new point in 
communicative space and/or a new merle. 
To formulate the counter-argument, A uses 
information from the exemplar of D (it may 
be updated on the basis of B's negative 
answer) and its model of B (which it had to 
108 
change because of B's negative answer). A 
models anew B's reasoning, i.e. the process 
is repeated cyclically. 
4 Dialogue examples 
4.1 Example 1 
The example represents a dialogue where the 
computer plays A's role and is implementing the 
tactic of enticement. The user implements the 
simple tactic: point out little pleasantness and 
unpleasantness of the action. 
Let us suppose that the action D is "travel to 
Venice and conclude a contract there". The 
computer has chosen the tactics of enticement 
and has generated the following user model: 
wAa(resources) = 1, wAB(pleas) =10, 
wAB(unpleas)=4, w~a3(use)=0, wAB(harm)=l 0, 
wAB(obligatory)= 1, wAB(prohibited)=0, 
wAB(puniShD)=0 , WAB(puniShnot.D)=6. 
The reasoning procedure WISH on this model 
yields a positive decision. A following dialogue 
was generated (translated from Estonian): 
C(omputer): Do you want to travel to Venice? 
Our firm needs to conclude a contract there. 
The computer informed the user about its 
communicative goal. 
U(ser) : It.is not interesting for me. 
The user pointed out little pleasantness of the 
action. The computer must correct win(pleas) in 
the user model. The new value of wAS(pleas) 
will be 8. 
C: You can meet interesting people. 
The computer chose its reply from the class of 
expressions for increasing pleasantness. After 
this turn the value ofwaB(pleas) will be 9. 
U: What's so pleasant about this trip? 
The user pointed out little pleasantness once 
more. The corrected value of wAS(pleas) will be 
8. 
C: There are many nice men in Italy. 
After this turn wAB(pleas)=9. 
U: This trip is too intense for me. 
The user indicated unpleasantness of D. The 
corrected value for wm(unpleas) will be 5. 
C: The company will pay all your expenses. 
The computer chose its reply from the class of 
expressions for decreasing unpleasantness and 
supposes that after its reply the value of 
wAa(unpleas) will be 4. 
U: I don "t like to travel. 
The user indicated unpleasantness of the action 
once more. Thus, the new value of wAB(unpleas) 
will be 5. 
C: You can take plane. 
The computer decreased the unpleasantness of D 
once more. The new value of wAB(unpleas) is 4. 
U: You are right - I shall travel. 
The user agreed to do D, the communicative 
goal of the computer is achieved. 
C: I am glad. 
4.2 Example 2 
The example represents a dialogue where the 
computer is implementing the tactic of 
threatening. The user implements the tactic as 
before: point out little pleasantness and 
unpleasantness of the action. 
The action D is "travel to Venice and conclude a 
contract there" as before. The computer has 
chosen the tactic of threatening and has 
generated the following user model: 
wAB(resources) = 1, wAB(pleas)= 10, 
wAB(unpleas)=4, wAB(use)=4, wAB(harm)=6, 
WAB(obligatory)=l, wAa(prohibited)=0, 
wAB(puniShD)=0, w~(punishnot.D)=8. 
The reasoning procedure MUST on this model 
yields a positive decision thus threatening is 
possible. A following dialogue was generated: 
C: You must travel to Venice. Our firm needs to 
conclude a contract there. 
U: It is not interesting for me. 
The user pointed out little pleasantness of the 
action. The computer must correct wAB(pleas) in 
the user model. The new value of wAB(pleas) 
will be 0. 
C: There are many people interested on getting 
your job. 
Using the tactic of threatening the computer 
chose its reply from the class of expressions for 
increasing punishment for not doing D. After 
this turn the value of wAB(punish~ot.D) will be 9. 
U: What ~ so pleasant about this trip?. 
The user pointed out little pleasantness once 
more. 
C: Refusing will be harmful for your reputation. 
After this turn WAB(punishnot.D)=10. 
U: This trip is too intense for me. 
The user indicated unpleasantness of D. The 
corrected value for waB(unpleas) will be 14. 
C: lf you have money you can travel by plane. 
109 
The computer chose its reply fi:om the class of 
expressions for decreasing unpleasantness. The 
value ofwAS(unpleas) will be 13. 
U: 1don "t like to travel 
The user indicated unpleasantness of the action 
once more. 
The reasoning procedure MUST on the user 
model will give the negatiw~ decision thus 
threatening is impossible. The computer can't 
choose new tactics because reasoning 
procedures WISH and NEEDED will give on the 
user model the negative decision too. The 
computer must give up. 
C: 1am sorry. 
Conclusion 
At present there exists implemented program 
which can play the role of both A and B in a 
simple communication situation where the goal 
of A is that B would decide to do D. At the 
moment the computer operates with semantic 
representations of linguistic input/output only, 
the surface linguistic part of interaction is 
provided in the form of a list of possible 
utterances. The work on linguistic processor is 
in progress. 
We have deliberately concentrated on modelling 
the processes of reasoning of conversation 
agents, as these processes form the heart of the 
"cognitive" part of human communication, and 
on modelling the use of communicative 
strategies and tactics which constitute the 
"social" part of communication. 
Although the concepts and models we have 
reported in the paper may seem too abstract 
from the point of view of practical NLP, we are 
convinced that without serious study and 
modelling of cognitive and social aspects of 
human communication it will appear impossible 
to guarantee naturalness of dialogues carried out 
by a computer system with a human user. 
As we have so far mostly dealt with agre ment 
negotiation dialogues, we have planned as one 
of the practical applications of the system as a 
participant in communication training sessions. 
Here the system can, for instance, establish 
certain restrictions on argument types, on the 
order in the use of arguments and counter- 
arguments, etc. 
Second, we have started to work, using our 
experience in modelling cognitive and social 
aspects of dialogue, on modelling information 
seeking dialogues in the same lines. This type of 
dialogue clearly will be the area where in the 
next few years already systems will be required 
that would be practically reliable, but at the 
same time could follow the rules of natural 
human communication. 
Acknowledgements 
This research was supported by 
Science Foundation (grant No 4467). 
Estonian 

References 
James Allen (1994) Natural Language 
Understanding. 2nd ed. The Benjamin/Cummings 
Publ. Comp., Inc. 
Jennifer Chu-Carroll and Sandra Carberry (1998) 
Collaborative Response Generation in Planning 
Dialogues. Computational Linguistics, 24/3, pp. 
355-400. 
Barbara Di Eugenio, Pamela W. Jordan, Richmond 
H. Tlaomason, Johanna D. Moore (2000) The 
Acceptance Cycle: An empirical investigation of 
human-human collaborative dialogues, to appear 
in International Journal of Human Computer 
Studies. 
Laila Dybkj~er (2000) Preface. - From Spoken 
Dialogue to Full Natural Interactive Dialogue- 
Theory, Empirical Analysis and Evaluation. LREC 
2000 Workshop Proceedings. L. Dybkjaer, ed. 
Athen, pp. 1-2. 
Peter Heeman and Graeme Hirst (1995) 
Collaborating on referring expressions. 
Computational Linguistics, 21/3, pp. 351-382. 
Kristiina Jokinen (1995) Rational Agency. In 
"Rational Agency: Concepts, Theories, Models, 
and Applications", M. Fehling, ed.. Proe. of the 
AAAI Fall Symposium. MIT, Boston, pp. 89-93. 
Kristiina Jokinen (1996) Cooperative Response 
Planning in CDM." Reasoning about 
Communicative Strategies. In "TWLT11. Dialogue 
Management in Natural Language Systems", S. 
LuperFoy, A. Nijholt & G. Veldhuijzen van 
Zauten, ed. Enschede: Universiteit Twente, pp. 
159-168. 
Mare Koit (1996) lmplementing a dialogue model on 
the computer. In "Estonian in the Changing World. 
Papers in Theoretical and Computational 
Linguieties", H. Oim, ed. Tartu, pp.. 99-114. 
Mare Koit and Haldur Oim (2000) Developing a 
model of natural dialogue. In "From spoken 
dialogue to full natural interactive dialogue-theory, 
Empirical analysis and evaluation. LREC2000 
Workshop proceedings", L. Dybkj~r, ed. Athen, 
pp. 18-21. 
Mare Koit and Haldur 0im (1999) Communicative 
strategies in human-computer interaction: a model 
that involves natural reasoning. In "23. Deutsche 
Jahrestag fiir Kfmstliche Intelligenz". Bonn, 
http://www.ikp.uni-bonn.de/NDS 99/Finals/1 2.ps 
Mare Koit and Haldur Oim (1998) Developing a 
model of dialog strategy. In "Text, Speech, 
Dialogue - TSD'98 Proceedings". Brno, pp. 387- 
390. 
Karen Lochbaum (1998) A Collaborative Planning 
Model of Intentional Structure. Computational 
Linguistics, 24/4, pp. 525-572. 
Haldur Oim (1996) Na~'ve theories and 
communicative competence: reasoning in 
communication. In "Estonian in the Changing 
World. Papers in Theoretical and Computational 
Linguictics", H. C)im, ed. Tartu, pp. 211-231. 
David R. Traum and James F. Allen (1994) 
Discourse Obligations in Dialogue Processing. In 
"Proceedings of the 32rid Annual Meeting of the 
Association for Computational Linguistics (ACL- 
94)", pp 1-8. 
Bonnie Webber (2000) Computational Perspectives 
on Discourse and Dialogue. In "The Handbook of 
Discourse Analysis". D. Schiffrin, D. Tannen, H. 
Hamilton, ed. Blackwell Publishers Ltd. \
