Dialogue Management in the Agreement Negotiation Process: A Model that Involves Natural Reasoning

Dialogue Management in the Agreement Negotiation Process:
A Model that Involves Natural Reasoning
Mare KOIT
Haldur ÕIM
Institute of Computer Science, Tartu University
Dept. of General Linguistics, Tartu University
Liivi 2
Tiigi 78
50409 Tartu, Estonia
51014 Tartu, Estonia
koit@ut.ee
hoim@psych.ut.ee
the cooperation of agents is modelled, relating it
Abstract
to agent’s mental states and planning processes.
The Shared Plans cooperation model deals with
In the paper we describe an approach to
planning processes in which participate multiple
dialogue management in the agreement
agents, see Lochbaum (1998). The model
negotiation where one of the central roles is
concentrates on group tasks that can be divided
attributed to the model of natural human
into separate, but interacting subtasks, and the
reasoning. The reasoning model consists of
central problem is coordination of intentions and
the model of human motivational sphere,
goals of partners.
and of reasoning algorithms. The reasoning
Di Eugenio et al. (2000) present a model
model is interacting with the model of
BalanceProposeDispose: first, the relevant
communication process. The latter is
information concerning the task is considered
considered as rational activity where central
and discussed, then a proposal is made and,
role play the concepts of communicative
lastly, the decision concerning the proposal is
strategies and tactics.
made – it is accepted or rejected.
In our model we depart from the same type of
Introduction
situation. One agent, A, addresses another agent,
Several researches have modelled the process of
B, with the intention that B will carry out an
argument negotiation in cooperative dialogue
action D. After some negotiation, B agrees or
where one participant makes a proposal to
rejects the proposal.
another participant and as the result of
In this paper we concentrate on the problems
negotiation this is accepted or rejected.
connected with modelling participants as
Chu-Carroll and Carberry (1998) present a
conversation agents who are able to participate
cooperative response-generation model as a
in negotiation in the form of natural dialogue –
recursive cycle Propose-Evaluate-Modify. They
dialogue that is carried out in natural language
concentrate on dialogues of information sharing
and according to the rules of human
and negotiation. An information sharing
communication.
dialogue is started, when the agent recognised a
Such a dialogue can be considered as rational
turn of his/her partner as a proposal, but does not
behaviour which is based on beliefs, wants and
have enough information to decide whether to
intentions of agents, at the same time being
accept it or not. A negotiation dialogue is
restricted by their resources, see Jokinen (1995),
started, when the agent concludes that the
Webber (2000). Conversation agent is a kind of
proposal is in conflict with his/her beliefs and
intelligent agent – a computer program that is
preferences, i.e. tends to reject it.
able to communicate with humans as another
Heeman and Hirst (1995) model cooperation by
human being.
the cycle Present-Judge-Refashion. They use
As it is generally accepted, in a model of
two levels of modelling – planning and
conversation agent it is necessary to represent its
cooperation. On the first level utterances are
cognitive states as well as cognitive processes.
generated and interpreted, on the second level

---

One of the most well-known models of this type
evaluative dispositions of participants towards
is the BDI model, see Allen (1994).
the world (e.g. what do they consider as pleasant
Our main point in this paper is that the general
or unpleasant, useful or harmful), and, on the
concepts of cognitive states and processes used
other hand, algorithms that are used to generate
in BDI-type models should be extended in order
plans for acting on the world.
to include certain factors from human
A necessary precondition of a communicative
motivational sphere and certain social principles
interaction is existence of shared (mutual)
in order to guarantee naturalness of dialogues of
knowledge of interacting agents. This concerns
the type we are concerned with. This is
goal bases as well as all types of knowledge
especially important in connection with the fact
bases; the intersections of the corresponding
that interest in modelling cooperative dialogues
bases of interacting agents A and B cannot be
where partners are pursuing a common goal has
empty: GBA ∩ GBB ≠∅, KBAW ∩ KBBW ≠∅,
considerably increased in recent years. On the
KBAL ∩ KBBL ≠∅, KBAD ∩ KBBD ≠∅, KB ABS
one hand, this is connected with rapid spreading
∩ KBBS ≠∅, KBBAS ∩ KBAS ≠∅.
of Internet-based services. On the other hand,
In this paper we will consider a specific type of
the interest in models of full natural dialogue
dialogue where the communicative goal of agent
derives from the possibility of building speech
A is to get agent B to agree to carry out an
interfaces with different knowledge and
action D – so-called agreement negotiation
databases, see Dybkjær (2000). Both of these
dialogue. We will concentrate here on dialogue
developments broaden the concept of
management in such kind of interaction, i.e. on
naturalness of dialogue considerably and present
the functioning of the module DM.
to it much stronger requirements concerning its
empirical adequacy as it has been generally
2
Dialogue Management
accepted thus far.
2.1
Reasoning Model
1
Model of Conversation Agent
A dialogue participant chooses his/her responses
In our model a conversation agent, A, is a
to the parter’s communicative acts as a result of
program that consists of 6 (interacting) modules:
certain reasoning process. After A has made B a
A = (PL, PS, DM, INT, GEN, LP),
proposal to do D, B can respond with agreement
where PL – planner, PS – problem solver, DM –
or rejection, depending on the result of his/her
dialogue manager, INT – interpreter, GEN –
reasoning.
generator, LP – linguistic processor. PL directs
Because we consider the model of natural
the work of both DM and PS, where DM
human reasoning as one of the important
controls communication process and PS solves
components in attaining naturalness of dialogue
domain-related tasks. The task of INT is to make
as a whole, we will discuss our model of
semantic analysis of partner’s utterances and
reasoning in some detail. From the point of view
that of GEN is to generate semantic
of practical NLP the approach we will present
representations of agent’s own contributions. LP
below may seem too abstract. But without solid
carries out linguistic analysis and generation.
theoretical basis it will appear impossible to
Conversation agent uses in its work goal base
guarantee naturalness of dialogues carried out by
GB and knowledge base KB. In our model, KB
computers with human users. We think that the
consists of 4 components:
model we describe here can be taken as a basis
KB = (KBW, KBL, KBD, KBS),
for the corresponding discussion.
where KBW contains world knowledge, KBL –
Our model is not based on any scientific theory
linguistic knowledge, KBD – knowledge about
of how human reasoning proceeds; our aim is to
dialogue and KBS – knowledge about interacting
model a “naïve theory of reasoning” which
subjects. For instance, KBD contains definitions
humans follow in everyday life when trying to
of communicative acts, turns and transactions
understand, predict and influence other persons’
(declarative knowledge), and algorithms that are
decisions and behavior, see Koit and Õim
applied to reach communicative goals –
(2000). The reasoning model consists of two
communicative strategies and tactics (procedural
functionally linked parts: 1) a model of human
knowledge); KBS contains knowledge about
motivational sphere; 2) reasoning schemes.

---

In the motivational sphere three basic factors
P3. In assessing an action D the values of
that regulate reasoning of a subject concerning D
(internal – wish- and needed-) factors are
are differentiated. First, subject may wish to do
checked before the external (must-) factors.
D, if pleasant aspects of D for him/her
P4. If D is found pleasant enough (i.e. D’s
overweight unpleasant ones; second, subject
pleasant aspects overweight unpleasant ones),
may find reasonable to do D, if D is needed to
then the needed- and must-factors will first be
reach some higher goal, and useful aspects of D
checked from the point of view of their negative
overweight harmful ones; and third, subject can
aspects (“to what harmful consequences or
be in a situation where he/she must (is obliged)
punishments D would lead?”).
to do D – if not doing D will lead to some kind
The rule P4 explains, for example, why in
of punishment. We call these factors wish-,
Figure 1 step 1 is immediately followed by step
needed- and must-factors, respectively.
2.
For instance, in reasoning about some action D
The weights of different aspects of D
(e.g. proposed by another agent), an agent as an
(pleasantness, unpleasantness, usefulness,
individual subject typically starts with checking
harmfulness, punishment for doing a prohibited
his/her wish-factor, i.e. whether D’s pleasant
action or not-doing an obligatory action) must be
aspects overweight unpleasant ones. If this
summed up in some way. Thus, in a
holds, then the subject checks his/her resources,
computational model weights must have
and if these exist, proceeds to other positive and
numerial values. In reality people do not operate
negative aspects of D: its usefulness and
with numbers but, rather, with some fuzzy sets.
harmfulness, and if D is prohibited, then also
On the other hand, existence of certain scales
possible punishment(s). If the positive aspects in
also in human everyday reasoning is apparent.
sum overweight negative ones, the resulting
For instance, for the characterisation of pleasant
decision will be to do D, otherwise – not to do
and unpleasant aspects of some action there are
D.
specific words: enticing, delightful, enjoyable,
There can exist other typical situations. If the
attractive, acceptable, unattractive, displeasing,
agent is an “official” person, or a group of
repulsive etc. Each of these adjectives can be
subjects formed to fulfil certain tasks and/or to
expressed quantitatively. This presupposes
pursue certain pre-established goal(s), then
empirical studies, though.
typically the starting point of reasoning is
We have represented the model of motivational
needed- and/or must-factor.
sphere by the following vector of weights:
This means that there exist certain general
wA = (w(resourcesAD1), w(pleasAD1),
principles that determine how the reasoning
w(unpleasAD1), w(useAD1), w(harmAD1),
process proceeds. These principles depend, in
w(obligatoryAD1), w(prohibitedAD1),
part, on the type of the reasoning agent. Before
w(punishAD1), w(punishAnot-D1),…,
starting to construct a concrete reasoning model
w(resourcesADn), w(pleasADn), w(unpleasADn),
the types of agents involved should be
w(useADn), w(harmADn), w(obligatoryADn),
established. In our implementation the agent is
w(prohibitedADn), w(punishADn),
supposed to be a “simple” human being and the
w(punishAnot-Dn)).
actions under consideration are from everyday
Here D1, …, Dn represent human actions;
life. In this case as examples of such principles
w(resourcesADi)=1, if A has resources necessary
used in our model we can present the following
to do Di (otherwise 0); w(obligatoryADi)=1, if Di
ones. For more details, see Õim (1996).
is obligatory for A (otherwise 0);
P1. People prefer pleasant (more pleasant)
w(prohibitedADi)=1, if Di is prohibited for A
states to unpleasant (less pleasant) ones.
(otherwise 0). The values of other weights are
P2. People don’t take an action of which they
non-negative natural numbers.
don’t assume that its consequence will be a
The second part of the reasoning model consists
pleasant (useful) situation, or avoidance of an
of reasoning schemas, that supposedly regulate
unpleasant (harmful) situation.
human action-oriented reasoning. A reasoning
The following principles illustrate more concrete
scheme represents steps that the agent goes
(operational) rules.
through in his reasoning process; these consist in
computing and comparing the weights of

---

different aspects of D; and the result is the
that the pleasant aspects of D (including its
decision to do or not to do D.
consequences) overweigh its unpleasant aspects.

Figure 1 presents the reasoning scheme that
The same kinds of reasoning schemes are
departs from the wish of a subject to do D.
constructed for the needed- and must-factors.
The scheme also illustrates one of the general
The reasoning model is connected with the
principles referred to above. It explains the order
general model of conversation agent in the
the steps are taken by the reasoning agent: if a
following way. First, the planner PL makes use
subject is in a state where he/she wishes to do D,
of reasoning schemes and second, the KBS
then he/she checks first the harmful/useful
contains the vector wA (A’s subjective
aspects of D, and after this proceeds to aspects
evaluations of all possible actions) as well as
connected with possible punishments.
vectors wAB (A’s beliefs concerning B’s
evaluations, where B denotes agents A may
Presupposition:
communicate with). The vector wAB do not
w(pleas) > w(unpleas).
represent truthful knowledge, it is used as a
partner model.

1) Are there enough resources
When comparing our model with BDI model,
for doing D?
then beliefs are represented by knowledge of the

If not then not to do D.

2) Is w(pleas) > w(unpleas) +
conversation agent with reliability less than 1;
w(harm)?
desires are generated by the vector of weights

If not then go to step 6.
wA; and intentions correspond to goals in GB. In

3) Is D prohibited?
addition to desires, from the weights vector we

If not then to do D.
also can derive some parameters of the

4) Is w(pleas) > w(unpleas) +
motivational sphere that are not explicitly
w(harm) + w(punish )?
D

If yes then to do D.
covered by the basic BDI model: needs,

5) Is w(pleas) + w(use) >
obligations and prohibitions. Some wishes or
w(unpleas) + w(harm) +
needs can be stronger than others: if w(pleasADi)
w(punish )?
- w(unpleasA
D
Di) > w(pleasADj) - w(unpleasADj),

If yes then to do D else
then the wish to do Di is stronger than the wish
not to do D.
to do D

6) Is w(pleas) + w(use) >
j. In the same way, some obligations
(prohibitions) can be stronger than others,
w(unpleas) + w(harm)?

If yes then go to step 9.
depending on the weight of the corresponding

7) Is D obligatory?
punishment. It should be mentioned that adding

If not then not to do D.
obligations to the standard BDI model is not

8) Is w(pleas) + w(use) +
new. Traum and Allen (1994) show how
w(punish
) > w(unpleas) +
not-D
discourse obligations can be used to account in a
w(harm)?
natural manner for the connection between a

If yes then to do D else
question and its answer in dialogue and how
not to do D.

9) Is D prohibited?
obligations can be used along with other parts of

If not then to do D.
the discourse context to extend the coverage of a

10) Is w(pleas) + w(use) >
dialogue system.
w(unpleas) + w(harm) +
w(punish )?
2.2
Communicative Strategies and
D

If yes then to do D else
Tactics
not to do D.

Knowledge about dialogue KBD, which is used

Figure 1. The reasoning procedure that departs
by the Dialogue Manager, consists of two
from the wish of a subject to do D.
functional parts: knowledge of the regularities of

dialogue, and rules of constructing and

combining speech acts.
The prerequisite for triggering this reasoning
The top level concept of dialogue rules in our
procedure is w(pleas) > w(unpleas), which is
model is communicative strategy. This concept
based on the following assumption: if a person
is reserved for such basic communication types
wishes to do something, then he/she assumes
as information exchange, directive dialogue,

---

phatic communication, etc. On the more
dialogue.
concrete level, the conversation agent can realise
a communicative strategy by means of several
communicative tactics; this concept more

1) Choose the communicative
closely corresponds to the concept of
tactic.
communicative strategy as used in some other

2) Implement the tactic to
generate an expression (inform
approaches, see e.g. Jokinen (1996). In the case
the partner of the communicative
of directive communication (which is the
goal).
strategy we are interested in) the agent A can use

3) Did the partner agree to do
tactics of enticing, persuading, threatening. In
D? If yes then finish (the
the case of enticing, A stresses pleasant aspects,
communicative goal has been
in the case of persuading – useful aspects of D
reached).

4) Give up? If yes then finish
for B; in the case of ordering A addresses
(the communicative goal has not
obligations of B, in the case of threatening A
been reached).
explicitly refers to possible punishment for not

5) Change the communicative
doing D.
tactic? If yes then choose the
Which one of these tactics A chooses depends
new tactic.
on several factors. There is one relevant aspect

6) Implement the tactic to
of human-human communication which is
generate an expression. Go to
step 3.
relatively well studied in pragmatics of human

communication and which we have included in

Figure 2. Communicative strategy used by the
our model as the concept of communicative
initiator of communication.
space.

Communicative space is defined by a number of

coordinates that characterise the relationships of

1) If wB(resources)=0 then
participants in a communicative encounter.

present a counterargument in
Communication can be collaborative or
order to point at the presence
confrontational, personal or impersonal; it can
of possible resources or at the
be characterised by the social distance between
possibility to gain them.
participants; by the modality (friendly, ironic,

2) If wB(harm) > wAB(harm) then
hostile, etc.) and by intensity (peaceful,

present a counterargument in
order to downgrade the value of
vehement, etc.). Just as in case of motivations of
harm.
human behaviour, people have an intuitive,

3) If wB(obligatory)=1 &
“naïve theory” of these coordinates. This

wB(punish
) < wAB(punish
) then
not-D
not-D
constitutes a part of the social conceptualisation

present a counterargument in
of communication, and it also should not be
order to decrease the weight of
ignored in serious attempts to model natural
the punishment.

4) If wB(prohibited)=1 &
communication in NLP systems.

wB(punish ) > wAB(punish ) then
In our model the choice of a communicative
D
D

present a counterargument in
tactics depends on the “point” of the
order to downgrade the weight of
communicative space in which the participants
the punishment.
place themselves. The values of the coordinates

5) If wB(unpleas) > wAB(unpleas)
are again given in the form of numerical values.
then

The communicative strategy can be presented as

present a counterargument in
order to downgrade the value of
an algorithm (Figure 2).
the unpleasant aspects of D.

Figure 3 presents a tactic of enticement.

6) Present a counterargument in

In our model there are three different
order to stress the pleasant
communicative tactics that A can use within the
aspects of D.
frames of the directive communicative strategy:

those of enticement, persuasion and threatening.

Figure 3. A's tactics of enticement.
Each communicative tactic constitutes a

procedure for compiling a turn in the ongoing

---

The tactic of enticement consists in increasing
PROPOSAL (author A, recipient B,
B’s wish to do D; the tactic of persuasion
A proposes B to do an action D)
consists in increasing B’s belief of the usefulness
I. Static part
SETTING
of D for him/her, and the tactic of threatening
(1)
A has a goal G
consists in increasing B’s understanding that
(2)
A believes that B in the
he/she must do D.
same way has the goal G

Communicative tactics are directly related to the
(3)
A believes that in order to
reasoning process of the partner. If A is applying
reach G an instrumental goal
the tactics of enticement he/she should be able to
G should be reached
i
(4)
A believes that B in the
imagine the reasoning process in B that is
same way believes that in
triggered by the input parameter wish. If B
order to reach G an
refuses to do D, then A should be able to guess
instrumental goal G should be
i
at which point the reasoning of B went into the
reached
“negative branch”, in order to adequately
(5)
A believes that to attain
construct his/her reactive turn.
the goal G B has to do D
i

Analogously, the tactic of persuasion is related
(6)
A believes that B has
resources for doing D
to the reasoning process triggered by the needed-
(7)
A believes that B will
parameter, and the threatening tactic is related to
decide to do D
the reasoning process triggered by the must-
GOAL: B decides to do D
parameter. For more details see, for example,
CONTENT: A informs B that
Koit (1996), Koit and Õim (1998), Koit and Õim
he/she wishes B to do D
(1999).
CONSEQUENCES

Thus, in order to model various communicative
(1)
B knows the SETTING, GOAL
tactics, one must know how to model the process
and CONTENT
of reasoning.
(2)
A knows that B knows the
SETTING, GOAL and CONTENT
2.3
Speech Acts
II. Dynamic part
Generating procedures (A’s
The minimal communicative unit in our model is
possibilities to build his/her
speech act (SA). In the implementation we make
turn that contains Proposal as
use of a limited number of SAs the
the dominant SA).
representational formalism of which is frames.
A has Goal G; A believes that B
also has Goal G; A believes that
Figure 4 presents the frame of SA Proposal in
in order to reach G, G should be
the context of co-operative interaction. Other
i
reached; A has decided to
SAs are represented in the same form. Each SA
formulate this as Proposal to B
contains a static (declarative) and a dynamic
to do D.
(procedural) part. The static part consists of
Procedures (before formulating
preconditions, goal, content (immediate act) and
the turn) consist in checking
consequences. The dynamic part is made up
whether the preconditions of
proposal hold and in making
from two kinds of procedures: 1) those that the
decisions about information to
author of the SA applies in the generation of a
be added in the turn:
communicative turn that contains the given SA;
- in case of (2) : is G
2) those that the addressee applies in the process
actualised in B? If not, then
of response generation.
actualise it by adding SA
As one can see, such a two-part representation
Inform;
- in case of (4) : does B
contains also rules for combining SAs in a turn,
believe that in order to
and on the other hand, guarantees coherence of
reach B, G should be reached
turn-takings: when we have tagged in KB
i
D
first. If not, then add SA
initiating SAs (such as Question or Proposal),
Explanation (Argument);
then the following chain of SAs follows from
- in case of (6) : if A is not
the interpretation-generation procedures as
sure that B has resources for
applied by participants.
D, then add Question;
- in case of (7) : if A is not
sure that B will agree to do

---

D (for this A should model
3
Process of Dialogue
B’s reasoning), then add
Argument.
Let us describe the case where both A and B are
Procedures of interpretation-
intelligent agents; i.e. computer programs.
generation
1. A constructs
(B’s possibilities to react to
a) the frame exemplar of D, putting in it all
proposal) are started after B
relevant information A has about D;
has recognised SA Proposal:
- in case of (2), (4), (5) : if
b) the model of partner B, putting in it all
B does not have Goal G and/or
relevant information it has about B’s
he/she does not have the
evaluations concerning the contents of the
corresponding beliefs and A
slots in D’s frame.
has not provided the needed
2. A chooses the point in communicative space
additional information, then
from which it intents to start the interaction.
add Question (ask for
3. A starts to apply communicative strategy. A
additional information);
- in case of (6) : if B does
models B’s reasoning process, using B’s
not have Resources for D,
model. First A applies the reasoning scheme
then Reject + Argument;
based on the wish of B. If it results in ‘to do
- in case of (7) : if the
D’, then A actualises the tactic of enticing
decision of B to do D (as the
and generates its first turn which contains a
result of the application of
frame exemplar of Proposal. If the result of
reasoning scheme(s)) is
negative, then Reject +
modelled reasoning results in ‘not to do D’,
Argument.
then A tries reasoning which starts from
needed-factor and then the one triggered by
Figure 4. Speech act Proposal in the context of
must-factor, and according to the result
co-operative interaction.
actualises tactics of persuading or
threatening, and generates the first utterance.
If the application of all reasoning schemes
Such a representation does not guarantee
results in ‘not to do D’, then A abandons its
coherence of dialogic encounters (transactions)
goal.
on a more general level. For instance, it does not
4. B interprets A’s turn and recognises
cover such phenomena as topic change,
Proposal in it. B constructs it’s the exemplar
inadequate responses caused by
representation of D (this may not coincide
misunderstandings; but, more importantly, also
with that of A). B starts reasoning, in the
various kinds of initiative overtakings. For
course of which it may need additional
instance, after rejecting the Proposal made by A,
information from A. On the basis of the
B can, in addition to explaining the rejection by
frame of Proposal B formulates the result of
Argument, initiate various “compensatory”
reasoning as its response turn: yes/no +
communicative activities. Such things are
(maybe) Argument.
normal in human co-operative interaction and
5. A interprets B’s answer and determines
they are regulated by general pragmatic
which point in the dialogue scenario this
principles that require from participants, in
corresponds to. If B’s answer was positive
addition to being co-operative and informative,
(decision to do D), then according to
also being considerate and helpful. In our case
communicative strategy the encounter has
this means that KB
come to its successful end. If B’s answer is
D should also include general
level dialogue scenarios (in the form of a graph)
negative, then according to the dialogue
and formalisations of the mentioned pragmatic
scenario A must formulate a
principles; for an example of the latter, see
(counter)Argument. The communicative
Jokinen (1996).
strategy also allows to choose a new point in
communicative space and/or a new tactic.
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

---

change because of B’s negative answer). A
The user indicated unpleasantness of the action
models anew B’s reasoning, i.e. the process
once more. Thus, the new value of wAB(unpleas)
is repeated cyclically.
will be 5.
C: You can take plane.
4
Dialogue examples
The computer decreased the unpleasantness of D
once more. The new value of wAB(unpleas) is 4.
4.1
Example 1
U: You are right - I shall travel.
The user agreed to do D, the communicative
The example represents a dialogue where the
goal of the computer is achieved.
computer plays A’s role and is implementing the
C: I am glad.
tactic of enticement. The user implements the
simple tactic: point out little pleasantness and
unpleasantness of the action.
4.2
Example 2
Let us suppose that the action D is “travel to
The example represents a dialogue where the
Venice and conclude a contract there”. The
computer is implementing the tactic of
computer has chosen the tactics of enticement
threatening. The user implements the tactic as
and has generated the following user model:
before: point out little pleasantness and
wAB(resources)=1, wAB(pleas)=10,
unpleasantness of the action.
wAB(unpleas)=4, wAB(use)=0, wAB(harm)=10,
The action D is “travel to Venice and conclude a
wAB(obligatory)=1, wAB(prohibited)=0,
contract there” as before. The computer has
wAB(punishD)=0, wAB(punishnot-D)=6.
chosen the tactic of threatening and has
The reasoning procedure WISH on this model
generated the following user model:
yields a positive decision. A following dialogue
wAB(resources)=1, wAB(pleas)=10,
was generated (translated from Estonian):
wAB(unpleas)=4, wAB(use)=4, wAB(harm)=6,
C(omputer): Do you want to travel to Venice?
wAB(obligatory)=1, wAB(prohibited)=0,
Our firm needs to conclude a contract there.
wAB(punish
The computer informed the user about its
D)=0, wAB(punishnot-D)=8.
The reasoning procedure MUST on this model
communicative goal.
yields a positive decision thus threatening is
U(ser): It is not interesting for me.
possible. A following dialogue was generated:
The user pointed out little pleasantness of the
C: You must travel to Venice. Our firm needs to
action. The computer must correct wAB(pleas) in
conclude a contract there.
the user model. The new value of wAB(pleas)
U: It is not interesting for me.
will be 8.
The user pointed out little pleasantness of the
C: You can meet interesting people.
action. The computer must correct wAB(pleas) in
The computer chose its reply from the class of
the user model. The new value of wAB(pleas)
expressions for increasing pleasantness. After
will be 0.
this turn the value of wAB(pleas) will be 9.
C: There are many people interested on getting
U: What’s so pleasant about this trip?
your job.
The user pointed out little pleasantness once
Using the tactic of threatening the computer
more. The corrected value of wAB(pleas) will be
chose its reply from the class of expressions for
8.
increasing punishment for not doing D. After
C: There are many nice men in Italy.
this turn the value of wAB(punish
After this turn wAB(pleas)=9.
not-D) will be 9.
U: What’s so pleasant about this trip?.
U: This trip is too intense for me.
The user pointed out little pleasantness once
The user indicated unpleasantness of D. The
more.
corrected value for wAB(unpleas) will be 5.
C: Refusing will be harmful for your reputation.
C: The company will pay all your expenses.
After this turn wAB(punish
The computer chose its reply from the class of
not-D)=10.
U: This trip is too intense for me.
expressions for decreasing unpleasantness and
The user indicated unpleasantness of D. The
supposes that after its reply the value of
corrected value for wAB(unpleas) will be 14.
wAB(unpleas) will be 4.
C: If you have money you can travel by plane.
U: I don’t like to travel.

---

The computer chose its reply from the class of
aspects of dialogue, on modelling information
expressions for decreasing unpleasantness. The
seeking dialogues in the same lines. This type of
value of wAB(unpleas) will be 13.
dialogue clearly will be the area where in the
U: I don’t like to travel.
next few years already systems will be required
The user indicated unpleasantness of the action
that would be practically reliable, but at the
once more.
same time could follow the rules of natural
The reasoning procedure MUST on the user
human communication.
model will give the negative decision thus
threatening is impossible. The computer can’t
choose new tactics because reasoning
Acknowledgements
procedures WISH and NEEDED will give on the
This research was supported by Estonian
user model the negative decision too. The
Science Foundation (grant No 4467).
computer must give up.
C: I am sorry.
References
James Allen (1994) Natural Language
Conclusion
Understanding. 2nd ed. The Benjamin/Cummings
Publ. Comp., Inc.
At present there exists implemented program
Jennifer Chu-Carroll and Sandra Carberry (1998)
which can play the role of both A and B in a
Collaborative Response Generation in Planning
simple communication situation where the goal
Dialogues. Computational Linguistics, 24/3, pp.
of A is that B would decide to do D. At the
355-400.
moment the computer operates with semantic
Barbara Di Eugenio, Pamela W. Jordan, Richmond
representations of linguistic input/output only,
H. Thomason, Johanna D. Moore (2000) The
the surface linguistic part of interaction is
Acceptance Cycle: An empirical investigation of
provided in the form of a list of possible
human-human collaborative dialogues, to appear
utterances. The work on linguistic processor is
in International Journal of Human Computer
in progress.
Studies.
We have deliberately concentrated on modelling
Laila Dybkjær (2000) Preface. – From Spoken
the processes of reasoning of conversation
Dialogue to Full Natural Interactive Dialogue –
agents, as these processes form the heart of the
Theory, Empirical Analysis and Evaluation. LREC
“cognitive” part of human communication, and
2000 Workshop Proceedings. L. Dybkjaer, ed.
on modelling the use of communicative
Athen, pp. 1-2.
strategies and tactics which constitute the
Peter Heeman and Graeme Hirst (1995)
“social” part of communication.
Collaborating on referring expressions.
Computational Linguistics, 21/ 3, pp. 351-382.
Although the concepts and models we have
reported in the paper may seem too abstract
Kristiina Jokinen (1995) Rational Agency. In
from the point of view of practical NLP, we are
“Rational Agency: Concepts, Theories, Models,
and Applications”, M. Fehling, ed.. Proc. of the
convinced that without serious study and
AAAI Fall Symposium. MIT, Boston, pp. 89-93.
modelling of cognitive and social aspects of
Kristiina Jokinen (1996) Cooperative Response
human communication it will appear impossible
Planning in CDM: Reasoning about
to guarantee naturalness of dialogues carried out
Communicative Strategies. In “TWLT11. Dialogue
by a computer system with a human user.
Management in Natural Language Systems”, S.
As we have so far mostly dealt with agreement
LuperFoy, A. Nijholt & G. Veldhuijzen van
negotiation dialogues, we have planned as one
Zanten, ed. Enschede: Universiteit Twente, pp.
of the practical applications of the system as a
159-168.
participant in communication training sessions.
Mare Koit (1996) Implementing a dialogue model on
Here the system can, for instance, establish
the computer. In “Estonian in the Changing World.
certain restrictions on argument types, on the
Papers in Theoretical and Computational
order in the use of arguments and counter-
Linguictics”, H. Õim, ed. Tartu, pp.. 99-114.
arguments, etc.
Mare Koit and Haldur Õim (2000) Developing a
Second, we have started to work, using our
model of natural dialogue. In “From spoken
experience in modelling cognitive and social
dialogue to full natural interactive dialogue-theory,

---

Empirical analysis and evaluation. LREC2000
Workshop proceedings”, L. Dybkjær, ed. Athen,
pp. 18-21.
Mare Koit and Haldur Õim (1999) Communicative
strategies in human-computer interaction: a model
that involves natural reasoning. In “23. Deutsche
Jahrestag für Künstliche Intelligenz”. Bonn,
http://www.ikp.uni-bonn.de/NDS99/Finals/1_2.ps
Mare Koit and Haldur Õim (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
Õim (1996) Naïve theories and
communicative competence: reasoning in
communication. In “Estonian in the Changing
World. Papers in Theoretical and Computational
Linguictics”, H. Õim, ed. Tartu, pp. 211-231.
David R. Traum and James F. Allen (1994)
Discourse Obligations in Dialogue Processing. In
“Proceedings of the 32nd 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.

---