TOWARDS A FORMAL MODEL OF KNOWLEDGE

ACQUISITION VIA COOEPERATIVE DIALOGUE

Asma Moubaiddin

Department of Linguistics and Phonetics, Faculty of Arts, The university of Jordan, Amman, Jordan

Nadim Obeid

Department of Computer Information Systems, King Abdullah II School for Information Technology

The university of Jordan, Amman, Jordan

Keywords: Dialogue, Argumentation, Nonmonotonic Logic, Knowledge Acquisition/Learning, Formalization.

Abstract: We aim, in this paper, to make a a first step towards developing a model of knowledge acquisition/learning

via cooperative dialogue. A key idea in the model is the concept of integrating exchanged information, via

dialogue, within an agent's theory. The process is nonmonotonic. Dialogue is a structured process and the

structure is relative to what an agent knows about the world or a domain of discourse. We employ a

nonmonotonic logic system, NML3, which formalizes some aspects of revisable reasoning, to capture an

agent's knowledge and reasoning. We will present a formalization of some basic dialogue moves and the

protocols of various types of dialogue. We will show how arguments, proofs, some dialogue moves and

reasoning may be carried out within NML3.

1 INTRODUCTION

We aim, in this paper, to make a a first step towards

developing a model of Knowledge Acquisition

(KA)/learning via cooperative dialogue. A key idea

in the model is the concept of integration; an agent

learns a collection of propositions concerning some

situation by integrating it within its knowledge about

that situation. Agents may switch roles.

We assume that each of the participants in a

dialogue has a certain well-define role, determined

by the type of dialogue, the goal of that type of

dialogue, and the rules for making moves in it. We

shall, following (Frans, Van Emeren and

Grootendorst, 1992; Walton, 1992; Walton and

Krabbe, 1995), adopt a model of dialogue that is

based on a commitment. Agents are computational

entities that have knowledge and possess the ability

to acquire and manipulate (modify, derive) through

reasoning their knowledge.

We shall assume that the agents are cooperative,

abide by the rationality rules, e.g. rules of relevance

(cf. Grice 1975) and rational in the sense that they

fulfil their commitments and obligations in a way

that truthfully reflects their beliefs and intentions.

The types of dialogue we will be considering in

this paper are: information-seeking, inquiry and

persuasion. A dialogue is initiated through

questioning. An answer to a question, about a

particular situation, may confirm what the agent

accepts/knows or it may somehow require a process

of belief revision. This suggests that the process of

incorporation of new information into an agent

theory be modelled nonmonotonically. We employ

for capturing an agent's knowledge and reasoning a

three-valued based nonmonotonic logic, NML3,

which formalizes some aspects of revisable

reasoning and is amenable to implementation.

Within NML3, we present a formalization of some

basic dialogue moves and the rules of protocols of

some types of dialogue. The rules of a protocol are

nonmonotonic in the sense that the set of

propositions to which an agent is committed and the

validity of moves vary from one move to another.

We will show how proofs, some dialogue moves

and reasoning may be carried out within NML3.

We shall begin, in section 2, with a presentation

of NML3 employed to capture an agent's

knowledge and reasoning. In section 3 we present

the types of dialogue and in section 4 we present a

formalization of some dialogue moves, rules of

protocols of some types of dialogue and the process

182

Moubaiddin A. and Obeid N. (2007).

TOWARDS A FORMAL MODEL OF KNOWLEDGE ACQUISITION VIA COOEPERATIVE DIALOGUE.

In Proceedings of the Ninth International Conference on Enterprise Information Systems - HCI, pages 182-189

DOI: 10.5220/0002350801820189

 SciTePress

of integration. We show in section 5 how proofs and

reasoning are carried out in NML3. Section 6 is

concerned with learning and dialogue.

2 REASONING WITH

INCOMPETE INFORMATION

The agent’s partial knowledge and reasoning

capability are expressed in a non-monotonic Logic,

NML3. The language L

NML3

is that of Kleene’s

three-valued logic extended with the modal operator

M (Epistemic Possibility). Staring with T (true), F

(false) and a set of atoms: p, q, r, …, more

complicated Well-Formed Formulae are formed via

closure under ~ (negation), & (conjunction), V

(disjunction) and → (implication). That is, if A and

B are WFF, then so are ~A, A&B, AVB, A→B and

MA.. In NML3, L is the dual of M, LA ≡ ~M~A.

(Obeid, 1996) defines a truth-functional implication

⊃ that behaves exactly like the material implication

of classical logic, as follows:

A ⊃ B = M(~A&B)V~AVB.

Non-monotonic reasoning is represented via the

epistemic possibility operator M. Using M, we may

define the operators U (undefined), D (defined) and

¬ (classical negation) as follows:

UA ≡ MA&M~A

DA ≡ ~UA

¬A ≡ DA & ~A

Formal Semantics

Definition 2.1 A model structure for L

NML3

<W, R, g> where W is a non-empty set of

information states, R is a binary relations on W and

g is a truth assignment function for atomic WFF. R

can be interpreted as epistemic possible extension

between states. Given w, w

are members of W, we

shall write w R w

to mean that the information state

is an epistemic possible extension of the

information state w.

We employ the notation

,w ⎟=g A (resp.

=⎜g A) to mean that A is accepted as true (resp.

false) at w in

with respect to g and

⎟=g A (resp.

=⎜g A) to mean that A is accepted as true (resp.

false) at every w in

with respect to g. For

convenience, reference to g will be omitted except

when confusion may arise.

Definition 2.2 Let A, B be wffs then, the truth "⎟="

and the falsity "=⎜" notions are recursively defined

as follows:

(i)

,w ⎟= T

(ii)

,w ⎟= p iff g(w,p) = true for atomic p

(iii)

,w ⎟= A&B iff

,w ⎟= A and

,w ⎟= B

(iv)

,w ⎟= ~A iff

,w =⎜ A

(v)

,w ⎟= MA iff (∃w

∈W)(wRw

and

⎟≠ ~A)

(i')

,w =⎜ F

(ii’)

,w =⎜ p iff g(w,p) = false for atomic p

(iii')

,w =⎜ A&B iff

,w =⎜ A or

,w =⎜ B

(iv')

,w =⎜ ~A iff

,w ⎟= A

(v')

,w =⎜ MA iff (∀w

∈W)(if wRw

then

⎟= ~A)

An Axiomatic System

NML3 is the smallest set of sentences of L

NML3

which is closed under the following axiom schema

and inference rules. We shall write ├

NML3

to mean

that A is a theorem of NML3.

Axiom Schema

(a1) A ⊃ (B ⊃ A&B)

(a2) A ⊃ (B → A)

(a3) A&B → A (a3') A&B → B

(a4) (A → B) ⊃ [(B → C) ⊃ (A → C)]

(a5) ~~A ≡ A (i.e., ~~A ⊃ A and ~~A ⊃ A)

(a6) ~(A&B) ≡ (~A V ~B)

(a7) A → MA

Inference Rules

Modus Ponens (MP) for ⊃ together with:

(R1) From ~AVB infer ~MAVB

(R2) From A ⊃ B infer MA ⊃ MB

(R3) From the ability to infer ~A infer MA

NML3 is sound and complete. One of the

advantages of NML is that defaults of Reiter's

defaults logic (Reiter 1980) can be represented as

sentences in the object language in the system. It can

be shown that there is a one-to-one correspondence

between extensions of a default theory and

appropriate minimal information states which

provide the semantic account (models) of the system

NML3. For more details (cf Obeid, 2005).

3 DIALOGUE

Dialogue is an exchange of messages between

two(or more) participants. Every dialogue has a goal

and requires cooperation between the participants to

fulfil its goal. This means that each participant has

an commitment to work towards fulfilling its own

goal and a commitment to cooperate with the other

participant’s attempt to realize their own goals.

(Walton and Krabbe, 1995) provides a typology

of dialogue types between two agents. For each type

of dialogue, they formulate an initial situation, a

primary goal, and a set of rules. These constitute a

model, representing the ideal way reasonable,

TOWARDS A FORMAL MODEL OF KNOWLEDGE ACQUISITION VIA COOEPERATIVE DIALOGUE

183

cooperative agents participate in the type of dialogue

in question. It is important to note that in the course

of communication, there often occurs a shift from

one type of dialogue to another. Dialogue

embedding takes place when the embedded dialogue

is functionally related to the first one. For instance, a

persuasion dialogue may require an information-

seeking sub-dialogue.

Information seeking (IS): In IS dialogue, one

participant extracts information from another

provided that it can be provided. An IS dialogue is

initiated when a participant lacks some information.

There may not be a need for proof in IS dialogue and

it is not necessary to establish a collective belief.

Inquiry:

The basic goal of inquiry is information

growth so that an agreement could be reached about

a conclusive answer of some question. The goal is

reached by a incremental process of argumentation

that employs established facts in order to prove

conclusions beyond a reasonable doubt. In short, the

aim is to acquire more reliable knowledge to the

satisfaction of all involved. Inquiry is a cooperative

type of dialogue and correct logic proofs are

essential. It is the most relevant type of dialogue in

collective decision making and learning. Inquiry

may require persuasion and vice-versa.

Persuasion:

The goal of persuasion dialogue is for

one agent to persuade the other participant(s) of its

point of view and the method employed is to prove

the adopted thesis. The initial reason for starting a

persuasion dialogue is a conflict of opinion between

two or more agents and the collective goal is to

resolve the issue. Argument here is based on the

concessions of the other participant. Proofs can be of

two kinds: (1) to infer a proposition from the other

participant’s concessions; and (2) by introducing

new premises probably supported by evidence.

Clearly, a process of knowledge update/belief

revision takes place here.

4 DIALOGUE SYSTEM

A dialogue system is a formal model that aims to

represent how a formal dialogue should proceed. It

defines the rules of the dialogue.

The topic language, L

Topic

, is a logical language

which consists of propositions that are the topics of

the dialogue. L

Topic

is associated with a logic Σ (in

our case NML3) which determines the inference

rules, the defeat relations between the arguments and

defines the construction of proper arguments and

dialogue moves. The choice of Σ (whether

monotonic or nonmonotonic) has an impact on the

entire dialogue system.

The communication language, L

COM

, specifies

the locutions which the participants are able to make

in the dialogue. One of the most influential agent

communication languages is KQML (Finin et al.

1994). Our proposed system uses a KQML-type of

language. We will assume that every agent

understands this language and that all agents have

access to common argument ontology, so that the

semantics of a message is the same for all agents.

4.1 Some Basic Dialogue Moves

A dialogue, D, is a sequence M

, . . . ,M

. A move is

a quadruple as follows:

= <ID(M

),PL(M

),LOC(M

),TARGET(M

where

(1) ID(M

), the identifier of M

, is i (i.e., indicating

that M

is the i

element of the sequence in the

dialogue).

(2) PL(M

) is the player of the move.

(3) LOC(M

) is the locution of the move from L

Topic

(4) TARGET(M

) is the target of the move.

If M

is a reply to a message in M

where

j <i then TARGET(M

) = M

Every dialogue system specifies its own set of

locutions. There are, however, several basic types of

communication primitives. Among these are:

Assert A: an agent g states A.

Retract A: this move is a countermove to Assert A.

In NML3, Retract A by g does not commit g to

Assert ¬A.

Accept A: An agent g accepts/concedes a

proposition A given by another agent.

Reject A: a countermove to Accept A.

Reject A by g does not commit g to Accept ¬A.

Question A: An agent g questions/asks from

another, g1, for information about A (e.g.,

whether A is derivable from its

theory, i.e, whether Σ( g

) ⏐− A.

Chanllenge A: This move is made by one agent g

for another g1, to explicitly state a proof (an

argument supporting) for A.

4.2 Knowledge Update/Process of

Integration

The way a dialogue affects an agent's knowledge or

Knowledge Base (KB) depends on how the agent

reacts to exchanged information.

Accepting a proposition A by an agent g1 entails

that A is not inconsistent with its KB, KB(g

). We

may distinguish the following cases:

(I) There is only one extension of KB(g

)and either

ICEIS 2007 - International Conference on Enterprise Information Systems

184

(a) A is derivable from KB(g

) using g

's logic

(b) ¬A is not derivable from KB(g

(II) There are many extensions of KB(g

) and A is

derivable in one whereas ¬A is derivable in another.

Accepting A would be a commitment to the

extension(s) of KB(g

) where A holds.

Rejecting by g1 a proposition A asserted by g

may entail:

(a) ¬A is derivable from KB(g

)

(b2) neither A nor ¬A is derivable from KB(g1).

A case of rejection without justification.

It is important to add that the issue of rejection

and/or acceptance, by an agent, say g1, of

propositions asserted by another agent, g2, is quite

complicated by various rules such as (temporal)

persistence, denying accepting previous assertions

and citing contradictory support for a proposition

and its negation.

4.3 Update Rules of Dialogue Moves

Let COMIT

(g) represent the commitment set of

agent g at a move that has an identifier i. COMIT

(g)

is a set of propositions from L

Topic

which the agent g

is committed to (e.g., prepared to hold on to) at that

point in the dialogue. During the dialogue,

propositions are added to and/or deleted from the

commitment set.

Given such background, we give the update rules

that specify how commitment stores are modified by

the move (cf. Maudet and Evrard 1998).

Let j < i, M

a move played by g1, and M is a

move by g as a reply to M

, then

(1) M = <i, g, Assert A, M

COMIT

(g)=COMIT

i-1

(g)∪{A} and

COMIT

) = COMIT

i-1

This step adds A to COMIT

i-1

(g) to result in

COMIT

(g) and

g can offer a proof of A.

(2) M = <i, g, Retract A, M

COMIT

(g)=COMIT

i-1

(g) - {A} and

COMIT

) = COMIT

i-1

This step deletes A from COMIT

i-1

(g) to result in

COMIT

(g), i.e., A is deleted from g's theory.

(3) M = <i, g, Accept A, M

)

COMIT

(g) = COMIT

i-1

(g) ∪ {A} and

COMIT

) = COMIT

i-1

Agent g accepts A from g

The impact of this step is

that A will be added to COMIT

i-1

(g) to yield

COMIT

(g) This can be possible if the locution of

the message M

is Assert A. The impact of this

message will be an update of g’s theory with A.

(4) M = <i, g, Reject A, M

)

COMIT

(g) = COMIT

i-1

(g) – {A} and

COMIT

) = COMIT

i-1

Agent g rejects A from g

This can only be possible

if the locution of the message M

is Assert A. The

impact of this message could either be no change to

g’s theory as it is in contradiction with A, in which

case, COMIT

(g)=COMIT

i-1

(g)–{A} = COMIT

i-1

(g))

or an update of g’s theory by retracting A.

(5) M = <i, g, Question A, M

)

This move does not alter either of COMIT

(g) or

COMIT

). In this case g is asking from g1 for

information about A (e.g., whether A is derivable

from its theory, i.e, whether Σ( g

) ⏐− A.

(6) M = <i, g, Chanllenge A, M

)

This move does not alter either of COMIT

(g) or

COMIT

). In this move g is forcing g1, to

explicitly state a proof (an argument supporting) A.

It is important to note that a participant in a

dialogue must keep track of the conversational

record between them and record what has been

accepted, challenged or rejected.

4.4 Rules of Protocols of Different

Types of Dialogue

Information-Seeking: If the information seeker is

g and the other agent is g

(1) g makes a Question move such as M

= <i, g,

Question A, M

) where M

is a move made

earlier by g

and l > i.

(2) g

replies with the move M

where the identifier

is k and its target M

, where k > I, as follows:

(i) M

= <k, g1, Assert A, M

> or

(ii) M

= <k, g1, Assert ¬A, M

> or

(iii) M

= <k, g1, Assert UA, M

UA means that for g

the truth value of A is

undefined.

(3) g either accepts g

response using an Assert

move or challenges it with a Challenge move.

UA initiates an inquiry sub-dialogue between the

agents or the information-seeking dialogue is

terminated.

(4) g

replies to a Challenge move with a proof

using a move M

= <r, g1, Assert S, M

> where

S is a proof of A in Σ( g

(5) Go to step (3) for each sentence in S.

Inquiry. The following is an inquiry-protocol about

a proposition A involving g and g

(1) g seeks a support/proof for A. It begins with an

Assert move that asserts B →A or asserts B ⇒

TOWARDS A FORMAL MODEL OF KNOWLEDGE ACQUISITION VIA COOEPERATIVE DIALOGUE

185

A, for some sentence B or a move that asserts

UA.

(2) g

either accepts B →A or accepts B ⇒ A using

an Accept move or challenges either of B→A

and B ⇒ A as appropriate with a Challenge

move.

(3) g replies to a challenge with Assert move that

provide a proof P in Σ(g) of the last proposition

challenged by g

(4) Go to step (2) for every proposition C ∈ P. That

is, substitute C for B →A or B ⇒ A.

(5) g

seeks a support/proof for B, i.e., it replies

with an Assert move that asserts E → B or

asserts E ⇒ B, for some sentence E or a move

that asserts UB.

(6) If COMIT(g) ∪ COMIT(g

) |- A then

the dialogue terminates successfully.

(7) The agents reverse roles and the appropriate

agent seeks a support/proof for E (step 5).

Persuasion. The agent g is trying to persuade g

accept A.

(1) g begins with a move that asserts A.

(2) g

replies with a move that

(i) accepts A or

(ii) asserts ¬A or

(iii) challenges A.

(3) two possibilities:

(a) If the answer of g

in the previous step is

(ii), then goto to step (2) with the roles of

the agents reversed and ¬A in place of A..

(b) If the answer of g

in the previous step is

(iii) (challenge), then

(α) g should reply with a move that

provide/asserts a proof P of A in Σ(g)

(β) go to step (2) for every for every

proposition C ∈ P.

5 ARGUMENTATION AND

PROOF IN NML3

It is clear from Section 4 that arguments have an

essential role to play in situations of conflict. They

can be used by an agent to increase the degree of

compatibility between its knowledge/beliefs and

those of other agents; one agent can persuade

another to adopt one or more propositions that it

accepts by presenting proofs/support for those

propositions (cf. Reed et al. 1997). In Artificial

Intelligence (AI), It is used in different ways: (1) to

structure knowledge where the aim is to determine

how utterances form arguments and how arguments

can be decomposed (cf. Toulmin, 1958); (2) to

model dialectical reasoning and deal with argument

construction (cf. Dung, 1995). It is important

present an argument in such a way so that it appeals to the

other participant knowledge. It allows an agent to

critically

question the validity of information

presented by another participant, explore multiple

perspectives and/or get involved in belief revision

processes.

5.1 Argumentation Framework

An Argumentation Framework (AF) system should

capture and represent the constituents of arguments

(e.g., the propositions which are taken into

consideration). These may include facts, definition,

rules, regulations, theories, assumptions and

defaults. They can be represented as formulae or sets

of formulae. It should also capture the interactions

and reactions between arguments and constituents of

arguments such as undercutting. Furthermore, some

notion of preference over arguments may be needed

in order to decide between conflicting arguments.

Definition 5.1.. Let Σ be a logical system. An

argument in Σ is a triple P = <S, A> where

S is a set of Well-Formed Formulae (WFF) and

A is a WFF of the language of Σ such that

(1) S is consistent

(2) S |-

A (A follows from S in Σ)

(3) S is minimal. No proper subset of S satisfies

(1) and (2) exists.

An argument in a logical system Σ is simply a proof

in Σ. S may need to be ordered. Thus, minimality in

condition (3) may not necessarily be set-theoretic.

S is called the support of P and A is its

conclusion. We shall use Support(P) (resp. Conc(P))

to denote that S is a support of P (resp. A is a

conclusion of P).

If the logical system Σ contains defeasible

implications/rules, then it would be worthwhile

distinguishing between a defeasible argument and a

non-defeasible/classical argument.

Definition 5.2 A defeasible argument is a proof P =

<S, A> where S contains some defeasible

implications. A non-defeasible/classical argument is

a proof that does not contains any defeasible

implication(s) or rely on any un-discharged

assumptions.

It is important to note that in a

defeasible/nonmonotonic theory, an agent could

provide a argument for both a proposition and its

negation, i.e., the theory of the agent may have

multiple extension (cf. Reiter (1980)). Thus, the

need for a notion of undercutting.

Definition 5.3 Let P1 and P2 be two argument in Σ.

Then Undercut(P1, P2) iff (∃ B ∈ Support(P2) such

ICEIS 2007 - International Conference on Enterprise Information Systems

186

that B ≡ ¬Conc(P1) where ” ≡” is the equivalence

of classical logic.

P2 undercuts P1 if, and only if, P2 has a formula

that is the negation of the conclusion of P1.

Propositions in agents theories may need to be

ordered to reflect some preference between

propositions needed to choose between conflicting

arguments. Such order could reflect the degree of

belief or truth in the proposition or some other

measure of preference.

5.2 Proof Method For NML3

One of the essential features of the proof system is

that it allows free and complete access to all stages

of the proof process.

The proof method proceeds by the construction

of a tableau (Beth, 1987). This is a tree-structure in

which all the possible models allowed by the

premises and negated conclusion are set out and

examined for consistency. The construction of the

tree is governed by rules for each logical connective

in the language. These rules are closely related to the

semantics of the language. A complete set of such

rules for all truth-functional connectives is given in

(Jeffrey, 1967).

The concept of refutation is considerably more

straightforward in classical logic than it is in NML3.

In the former case, to prove that a set of premises

implies a conclusion A, it is sufficient to show that

~A cannot be true if the premises are. We have seen

that in NML3, “A V ~A” is not a theorem. If we

find no consistent models taking ~A as the negation

of our conclusion, we will not have proven that A

follows from our premises. We will have shown

only that it might.

In NML3 (cf. Obeid 2000), we need to consider

the following cases:

(1) A is true (resp. false) if we can find no

consistent models for M~A (resp. MA)

(2) A is true (resp. false) or unknown if we can find

no consistent models for ~MA (~M~A).

The tableau rules for the connectives & and V

are the same as those for classical logic. The tableau

rule is defined as follows: the negation of an atomic

non-modalized formula A is ~MA and the negation

of MA is simply ~MA.

The tableau rule for → can easily be shown to be

of the form:

A → B

___|____

| |

~MA ~M~B

The rule

(R3) If we cannot infer ~A, infer MA

require special attention If there exists an open

branch of a proof tree which includes a formula

~MA, or more than one, all including ~MA, then we

might fire rule (R3) to try to infer nonmonotonically

MA and thereby derive a contradiction and finish the

proof. This is achieved by setting the target formula

for proving that ~A is true against the original

premise set, and running the proof process. If we fail

to prove ~A is true, we may infer that it is

consistent, and pass MA back to the parent proof.

It is our strategy that we only attempt to derive a

proof nonmonotonically if we fail to close all the

paths in a tree with the tableau rules. We are able to

decide whenever we wish whether or not we can

infer ~A, but it only makes sense to try when we

know we need to. This means at present that we wait

for the monotonic proof process to stop before

looking for a way to apply the rule (R3).

If we fire (R3) thereby inferring a formula MA,

we may close off any model in the proof including

the formula ~MA. We may allow several

applications of (R3) in one proof, thereby closing

different branches of the proof in different ways.

6 DIALOGUE AND REASONING

WITHIN NML3

In this section we give two examples that show

how dialogue, together with the reasoning within the

system NML3, is carried out. We shall not present

formally the proofs in NML3 due to lack of space.

Example 6.1. Consider a case where we have two

agents, g1 and g2, cooperate in a diagnostic task of a

series of batteries. g1 is in charge of testing the

voltage of the batteries and g2’s task is to find out

which battery is faulty.

Consider a battery which when operating

normally has a voltage between 1.2 volts and 1.6

volts. We use Batt(B) to mean that B is a battery,

Volt(B,V) to mean that the voltage of B is V and

ok(V) to mean that 1.2 < V < 1.6.

Suppose that we have Batt(B

) and Batt(B

) and

g1 observed that “OBS=Volt(Series(B

),1.45)”.

Then it cannot be the case that both B

and B

are

working normally. To appreciate how subtle and

intuitive the results are, we shall consider what g2

can infer in such a situation:

(i) should g2 infer that if one of the batteries is not

working normally then the other is?

(ii) should g2 infer that if one of the batteries is

working normally then the other is not?

It is a straightforward exercise to show that the

answer to (i) is negative and the answer to (ii) is

positive.

TOWARDS A FORMAL MODEL OF KNOWLEDGE ACQUISITION VIA COOEPERATIVE DIALOGUE

187

Example 6.2 Assume that we have two agents g1

and g2. g2 needs a lift in a car. It notices a car

parked in front of John's house and decides to ask g1

the following query: can John drive? g1 knows that

john is skilled to drive and has a car. Using classical

logic, g1 should fail to drive that John can drive

because it does not know whether John has a driving

licence. Using NML3, g1 can give g2 the answer:

Yes. To be more helpful, g1 may give the answer:

Yes, if John has a driving licence. Such an answer

can be reached because NML3 allows g1 to make

the assumption that John has a driving licence(if

there is no information to the contrary).

However, if g1 has learnt from another agent g3

that John does not have a driving licence, then g1

may give an answer: No or more informatively: No,

because John does not have a driving licence.

7 RELATED WORK

Acquiring knowledge from a domain expert is

considered to be one of the most important and

difficult stage in developing a successful Knowledge

Based System (KBS) (Smith, 1996). The process of

KA, together with representation, is defined as

being a means whereby information is extracted,

structured and organized (Jeng, 1996). (Chan, 1995)

proposes that there exists a need to avoid looking at

KA process as an entire process but perceives it

more like a series of identifiable phases: (1)

knowledge elicitation (to obtain information from

the expert), (2) knowledge analysis (to make sense

of the data acquired in the former stage) and (3)

knowledge representation.

In multi-agent communication languages, such

as KQML (Finin et al. 1993) and COSY (Haddadi,

1996), the emphasis is at the level of individual

messages, along with a relative neglect of overall

task, knowledge modeling. The framework of the

COMMONKADS methodology (Schreiber et al.,

1994) provides a comprehensive conceptual

modeling approach, ranging from organizational

analysis to system design and implementation.

However, it does not provide us with the formalism

and the reasoning mechanism that allow us to learn

from the message exchanged.

Most existing spoken dialogue systems focus on

simple and constrained tasks. Some examples are

found in (Pellom et al. 2001; Xu and Rudnicky

2000; Chu-Carroll 2000).

There has been other work on modelling

dialogue for complex task domains such as the

TRAINS system (Allen et al. 2001) and its

successor, TRIPS (Blaylock et al. 2002). TRIPS is a

distributed, agent-based cooperative dialogue

system. Its components act asynchronously and

communicate with each other by message passing.

Issues in supporting multi-modal interfaces have

been addressed in (McGlashan 1996) which

provides a combination of graphical and speech

modalities. Work in (Traum et al. 2003) follows the

framework of the TRINDI project (Larsson 2000)

which aims to model multi-modal dialogue for

multiple participant interaction.

An attempt in (Paek and Horvitz 2000) is made

to build a probabilistic model (using Bayesian

networks) of possible uncertainties at different levels

of human-computer conversation. Thus the system

would be able to identify actions that maximize the

expected utility of achieving mutual understanding.

8 CONCLUDING REMARKS

In this paper, we have made a first step towards

developing a model of KA/learning via cooperative

dialogue. A key idea in the model is the concept of

integrating exchanged information within an agent

theory. Dialogue is a structured process and the

structure is relative to what an agent knows about

the world or a domain of discourse. We have

employed a logic system NML3 which formalizes

some aspects of revisable reasoning. We have

presented a formalization of some basic dialogue

moves and the protocols of various types of

dialogue. We have also given some indication as to

how arguments, proofs, appropriate dialogue moves

and reasoning may be carried out within NML3.

On the linguistic side, the question of how a

collection of propositions is assigned as the semantic

interpretation to a linguistic message is not trivial.

Lexical and (to a lesser extent) structural ambiguity

are sensitive to what an agent knows about the

world, but "unfortunate" interpretations may still be

consistent with respect to an agent's theory.

There is a general tendency to consider

inconsistency, in agent's, say g, theory, to be a

problem that concerns only g. However, in

cooperative activities that involve more than one

agent, it may be of interest to the other agents to

know about, or minimally to be aware, of the way

inconsistency or exchanged information is dealt with

by g. This is because in such cases, one agent may

regard another agent's knowledge as in some weak

sense an extension of its own. Thus, there may be a

need to define a notion of compatibility which is

weaker and more permissive than localized logical

consistency.

ICEIS 2007 - International Conference on Enterprise Information Systems

188

REFERENCES

Allen J., Byron D., Dzilovska M., 2001, Towards

Conversational Human-Computer Interaction, AI

Magazine, 22(4), 27-37.

Blaylock N., Allen J., and G. Ferguson G., 2002,

Synchronization in an Asynchronous Agent-Based

Architecture for Dialogue Systems, In Proc. of 3rd

SIGdial Workshop on Discourse and Dialog, 1-10.

Chan, C.W., 1996, Knowledge Modeling for Constructing

an Expert System to support reforestation decisions,

Knowledge Based Systems, 9, 41-59.

Chu-Carroll j., 2000, MIMIC: An Adaptive Mixed

Initiative Spoken Dialogue System for Information

Queries, In Proc. 6

ACL Conf. on Applied Natural

Language Processing, 97-104.

Dung, P.M., 1995, On the acceptability of arguments and

its fundamental role in non-monotonic reasoning, logic

programming and n-person games, Artificial

Intelligence, 77, 321-357.

Finin, T., Fritzson R. Mackey D., McEntire R., 1994,

KQML as an agent Communication Language, In:

Proc. of 13

International Conf. on Information and

Knowledge Management, ACM Press, New York.

Frans H. van Emeren F. H. and Rob Grootendorst R.,

1992, Argumentation, Communication and Fallacies,

Hillsdale, N. J. Erlbaum.

Grice, H P, 1975, Logic and Conversation, The Logic of

Grammar, ed. Donald Davidson and Gilbert Harman,

Encino, California, Dickenson, 64-75.

Haddadi, A., 1996, Communication and Cooperation in

Agent Systems, Lecture Notes in Computer Science

Series No. 1056, Springer-Verlag.

Jeng, B., 1996, Interactive Induction of Expert

Knowledge, Expert Systems with Applications 10,

393-401.

Larsson S., Traum D., 2000, Information State and

Dialogue Management in the TRINDI Dialogue Move

Engine Toolkit. Natural Language Engineering, 6,

323-340.

Maudet, N. and Evrard F., 1998, A generic framework for

dialogue game implementation. In: Proceedings of the

Second Workshop on Formal Semantics and

Pragmatics of Dialog, Universite Twente, The

Netherlands.

McGlashan S., 1996, Towards Multimodal Dialogue

Management. In Proc. of 11

Twente Workshop on

Lanugage Technology, 13-22.

Obeid N., 1996, Three Valued Logic and Non-monotonic

Reasoning, Computers and Artificial Intelligence, Vol.

15, No. 6, 509-530.

Obeid N., 2000, Towards a Model of Learning Through

Communication, International Journal of Knowledge

and Information Systems. Vol 2, 498-508, Springer-

Verlag, USA.

Obeid N., 2005, A Model-Theoretic Semantics for Default

Logic, WSEAS Transactions on Computers, Vol. 4,

No. 6, 581-590.

Reed, C.A., Long, D.P., Fox, M. and Garagnani, M., 1997,

Persuasion as a Form of Inter-Agent Negotiation, In:

Lukose, D., Zhang, C., (eds), Proc. of 2nd Australian

Workshop on DAI, Springer Verlag, Berlin

Reiter R., 1980, A Logic for Default Reasoning, Artificial

Intelligence 13, 81-132.

Schreiber A., Wielinga B. J., Akkermans J., Van De

Velde W. and R. de Hoog R., (1994: CommonKADS -

A Comprehensive Methodology for KBS

Development, IEEE Expert 9(6), 28-37.

Smith, P. (1996) An Introduction to Knowledge

Engineering, International Thompson Computer Press.

Toulmin, S. 1958, The uses of argument, Cambridge

University Press, England.

Paek T., Horvitz E., 2000, Conversation as Action Under

Uncertainty, In Proc. of 16

Conf. on Uncertainty in

Artificial Intelligence, 455-464.

Pellom B., Ward W., Hansen J., Hacioglu K., Zhang J.,

Yu X., Pradhan S., 2001, University of Colorado

Dialog Systems for Travel and Navigation, In Proc. of

2001 Human Language Technology Conference.

Traum D., Rickel J., Gratch J., Marsella S., 2003,

Negotiation Over Tasks in Hybrid Human-Agent

Teams for Simulation-Based Training, In Proc. of 2

International Joint Conference on Autonomous Agents

and Multiagent Systems, 441-448.

Walton, D. N., 1992, Types of dialogue, dialectical shifts

and fallacies. In: Emeren, F. H. v., Grootendorst, R.,

Blair, J. A., and Willard, C. A., editors, Argumentation

illuminated, 133–147, Amsterdam.

Walton, D. and Krabbe, E., 1995, Commitment in

Dialogue: Basic Concepts of Interpersonal Reasoning.

State University of New York Press, USA.

Xu W., Rudnicky A., 2000, Task-Based Dialog

Management Using an Agenda, In Proc. of

ANLP/NAACL Workshop on Conversational Systems,

42-47.

TOWARDS A FORMAL MODEL OF KNOWLEDGE ACQUISITION VIA COOEPERATIVE DIALOGUE

189