Collection Mémoires et thèses électroniques
Accueil À propos Nous joindre

Chapter 5[*] A Pragmatic Approach based on Social Commitments and Arguments

Table des matières

In this chapter, we propose a formal approach for modeling the pragmatics of agent communication. This pragmatics captures the evolution and the dynamics of agent conversations. This approach is based on the combination of the social approach and the argumentative approach. The link between commitments and arguments that we establish in this chapter enables us to capture both the public and the reasoning aspects of agent communication pragmatics. On the basis of this approach we also propose a layered communication model and a conversational agent architecture.

Agent communication pragmatics deals with the way that agents use communicative acts when conversing. Pragmatics is related to the dynamics of agent interactions and to the way of connecting individual acts while building complete conversations. In the domain of agent communication, many researchers addressed pragmatics. For example Dastani and his colleagues (2000), Fornara and Colombetti (2003, 2004) and Pitt and Mamdani (2000) proposed the notion of protocols as a pragmatic mechanism. Pasquier and his colleagues (Pasquier and Chaib-draa, 2003), (Pasquier et al., 2003) proposed a cognitive coherence theory for this pragmatics. However, these approaches do not specify the evolution of conversations and they are specified informally or semi-formally. In addition, protocol-based approaches do not indicate how agents select their communicative acts. In the cognitive coherence approach, this aspect is addressed using the cognitive dissonance theory that enables agents to cognitively react to a statement. However, this approach does not allow agents to argue, for example, in order to persuade another agent or to negotiate with it.

In this chapter, we propose theoretical foundations for an approach to agent communication pragmatics. This approach uses three fundamental elements: social commitments, actions, and arguments. As illustrated in Figure 5.1, these elements are separated in three levels. The first level includes social commitments that agents use in their conversations. The second level includes actions that agents apply to the commitments. The speech acts that agents perform when conversing are defined in terms of these actions. The third level is composed of arguments that agents use to support their actions applied to the commitments. The evolution of agent conversations is represented by the notion of commitment state. Agents use their argumentation systems in order to be able to select the appropriate communicative acts to be performed considering the current state of the conversation.

The purpose of this chapter is to introduce our pragmatic approach that we will use in Chapter 6 in which we develop our framework called commitment and argument network. This framework models the connection between the communicative acts in a conversation and the evolution of this conversation. It is also a means that helps agents to communicate. We also use this pragmatic approach to propose a new persuasion protocol that we develop in Chapter 9.

This chapter is organized as follows. In Section 5.2, we present our social commitment-based framework. In Section 5.3, we introduce the notion of commitment state. A taxonomy of social commitments is given in Section 5.4. In Section 5.5, we establish the link between commitments and arguments. In Section 5.6, we present our communication model. In Section 5.7, we conclude the chapter by a discussion.

A social commitment SC is a public commitment made by an agent (called the debtor ), and directed towards a set of agents (called creditors ) (Castelfranchi, 1995), indicating that some fact is true or that some action will be performed. A commitment is an obligation in the sense that the debtor must respect and behave in accordance with this commitment. A representation of this notion as directed obligations using a deontic logic is proposed in (Herrestad, 1995). Commitments are social in the sense that they are expressed publicly and governed by some rules. This means that they are observable by all the participants. The main idea is that a speaker is committed to a statement when he made this statement or when he agreed upon this statement made by another participant and acts accordingly. What is important here is not that an agent agrees or disagrees upon a statement, but rather the fact that the agent expresses agreement or disagreement. Consequently, social commitments are different from the agent’s private mental states like beliefs, desires and intentions. This notion allows us to represent agent conversations as observed by the participants and by an external observant, and not on the basis of the internal agents’ states.

In our framework, we distinguish between the social commitment which can be modeled as an object, and the social commitment content. This distinction will be discussed latter in this section. The commitment content is characterized by a time tφ , which is generally different from the utterance time denoted tu, and from the time associated with the commitment and denoted tsc . tφ is the time described by the utterance, and thus by the content φ . Time tsc that can be used as an identifier of the commitment refers to the time during which the commitment holds. It can correspond to a fixed value or an interval. When it is an interval, this time is denoted [ t , t ]. When a temporal bound is instantiated, it takes a numerical value that respects the time unit used by agents. For example, let us consider the following utterance U sent by agent Ag1 to agent Ag2 : U : I will give you 5$ at 5PM. We can describe the content by the following predicate: φ = Give ( Ag1 , Ag2 , 5 $ ). We have: tφ = 5 PM . The commitment time tsc is an interval: tsc = [ tu , 5 PM ] with tu is the utterance time (Figure 5.2).

If the commitment is satisfied or violated we have tsc = [ tu , tφ ]. However, if the commitment is withdrawn, we have: tsc = [ tu , tw ], with tw the withdrawal time (Figure 5.3). Time tsc indicates the time during which the commitment holds, i.e. the time during which the commitment is active . Time tφ indicates the moment at which the commitment must be satisfied.

We denote a social commitment as follows:

SC ( Ag1 , A* , tsc , ( φ , tφ ))

where Ag1 is the debtor, A* is the set of the creditors ( A* = A / { Ag1 }, where A is the set of participants), tsc is the time associated with the commitment, φ its content and tφ the time associated with the content φ . A social commitment can be identified by tsc . Logically speaking, a commitment is a public propositional attitude. The logical semantics of this notion is defined in Chapter 7. The content of a commitment can be a proposition or an action. A detailed taxonomy of the social commitments that we use in our approach will be discussed latter. To simplify the notation, we suppose throughout this chapter that A = { Ag1 , Ag2 }.

In order to model the dynamics of conversations, we interpret speech acts as actions performed on commitments. A speech act is an abstract act that an agent, the speaker , performs when producing an utterance U and addressing it to another agent, the addressee . According to the Speech Act Theory (Searle, 1969), (Searle and Vanderveken, 1985), the primary units of meaning in the use of language are not isolated propositions but rather speech acts of the type called illocutionary acts . Assertions, questions, orders and declarations are examples of these illocutionary acts. For the moment, our interpretation of a speech act can be denoted by:

SA ( ik , Ag1 , Ag2 , tu , U ) = def Act ( Ag1 , tu , SC ( Ag1 , Ag2 , tsc , ( φ , tφ )))

where = def means "is interpreted by definition as".

The definiendum ( SA ( ik , Ag1 , Ag2 , tu , U )) is defined by the definiens ( Act ( Ag1 , tu , SC ( Ag1 , Ag2 , tsc , ( φ , tφ )))) as an action performed on a social commitment. SA is the abbreviation of "Speech Act", ik is the identifier of the speech act, Ag1 is the speaker, Ag2 is the addressee, tu is the utterance time, U is the utterance and Act indicates the action performed by the debtor on the commitment: Act ∈ { Create , Withdraw , Reactivate , Violate , Satisfy }.

These five actions are the actions that the debtor can apply to a commitment and reflect only the debtor’s point of view. However, we must also take into account the creditor when modeling a conversation which is, by definition, a joint activity. The following example illustrates this aspect:

U1 : Quebec is the capital of Canada.

U2 : No, the capital of Canada is Ottawa.

The utterance U1 leads the debtor to create a commitment whose content is "Quebec is the capital of Canada". On the other hand, the utterance U2 highlights a creditor’s action on this content that is in this case a refusal. We thus propose to model the creditors’ actions, which are applied to the commitment contents and not to the commitments themselves (Figure 5.4). This separation between the commitment and its content enables us to remain compatible with the semantics of commitments, i.e. the fact that only the debtor can handle its commitments. The creditor can only handle the content of the debtor’s commitment. Hence, we must differentiate between the actions applied to a commitment Act and the actions performed on the content of a commitment Act-content : Act-content ∈ { Submit-content , Accept-content , Refuse-content , Challenge-content , Change-content , Suspend-content, Justify-content , Defend-content , Attack-content }. We denote an action applied to the content of an Agi ’s commitment as follows:

Act-content ( Agk , tu , SC ( Agi , Agj , tsc , ( φ , tφ )))

where i , j , k ∈ {1, 2} and i j .

Agent Agk can thus act on the content of its own commitment (in this case we get k = i ) or on the content of another agent’s commitment (in this case we get k = j ).

Thus, a speech act leads either to an action on a commitment when the speaker is the debtor, or to an action on a commitment content when the speaker is the debtor or the creditor. When an agent acts on the content of a commitment created by another agent, we refer to this as " taking a position on a commitment content ". [3] However, it should be noted that the same utterance can lead both to take a position on the content of an existing commitment and to create a new commitment. Generally, a speech act leads to an action on a commitment and/or an action on a commitment content. Formally, in our framework a speech act can be defined using BNF notation as follows:

This definition will be enriched when we establish the link between social commitments and arguments (Section 5.5)

Let us take the previous example:

U1 : Quebec is the capital of Canada.

The utterance U1 leads to the creation of a new commitment:

U2 : No, the capital of Canada is Ottawa.

The utterance U2 leads at the same time to a positioning on the content of the commitment created following the utterance U1 and to the creation of another commitment. Formally:

A commitment can evolve and be transformed as a result of the actions that the debtor performs on it (creation, withdrawal, reactivation, violation and satisfaction). Its content may also be transformed as a consequence of the actions that the debtor and the creditors apply to it (change, acceptance, justification, etc.). Therefore, agents act on their own commitments and on the contents of both their commitments and other agents’ commitments. These actions lead to the transformation of these commitments and commitment contents. Hence, the notion of state makes it possible to capture the evolution of commitments and their contents. However, we must distinguish between the notion of the commitment state (Verdicchio and Colombetti, 2002) and the notion of the content state relative to this commitment as we propose here. Indeed, whenever an agent acts on its commitment, the commitment state is affected; whereas when an agent acts on the content of a commitment, the content state is transformed. Consequently, the notion of commitment state alone does not reflect the conversation dynamics since it only captures the debtor’s actions on its commitment. The two states (the commitment state and the content state of the commitment) reflect this dynamics. This notion is of great importance since it allows us to keep a trace of the dialogue evolution in so far as each speech act leads to an action performed on a commitment or on its content. Contrary to the notion of the commitment store (Hamblin, 1970) which allows us only to track "who said what", the notion of state makes it possible to illustrate how participants change the dialogue state by performing actions on existing commitments or on their contents.

Here are the states that we propose to use in our model. Once created, a commitment will take the active state and its content takes the submitted state . This expresses the fact that the content is presented for possible negotiation. A commitment can be in one of four states: active , satisfied , withdrawn , and violated (Figure 5.5). A commitment content can be in one of nine states: submitted , changed , refused , accepted , challenged , justified , contradicted , suspended , attacked and defended . These states and the operations which trigger them depend on the commitment type. We notice that justification, contradiction, attack and defense are argumentation-related actions. This means that their semantics is defined using the argumentation notions (this aspect will be detailed in Chapter 7).

The set of different states of a commitment whose identifier is tsc is denoted and the set of different states of a commitment content whose identifier is tsc is denoted and are finite and ordered sets. The ordering relation between the elements of these sets is defined as follows:

Definition 5.2 iff the commitment (the commitment content) whose identifier is tsc was in state before to be in state

The current state of a commitment (commitment content) whose identifier is tsc is the biggest element of the set ( ) according to the ordering relation .

The following example illustrates this notion of state and its evolution:

U1 : The book is not allowed during the test.

U2 : Why?

U3 : Because the answers are given in this book.

U4 : Ok, Thank you.

By utterance U1 , agent Ag1 creates a commitment, whose state is "active". The state of the content is "submitted". Formally:

By utterance U2 , agent Ag2 challenges the content of the commitment identified by tsc 1. This commitment always remains in the "active" state, but its content takes the state "challenged". Formally:

By utterance U3 , agent Ag1 creates a new commitment. The state of this commitment is "active", and the state of its content is "submitted". By the same utterance, this agent justifies the content of its commitment identified by tsc 1. The state of this commitment is always "active" and "justified" becomes the current state of its content. Formally:

By utterance U4 , agent Ag2 accepts the content of the commitment identified by tsc 2. Thus, "satisfied" becomes the current state of this commitment and "accepted" becomes the state of its content. Consequently, this agent also accepts the content of the commitment identified by tsc 1. Thus, "satisfied" and "accepted" are the current states respectively of this commitment and its content. Formally:

In the literature (Walton and Krabbe, 1995), (Singh, 1999), (Fornara and Colombetti, 2002), several commitment types have been proposed. In our approach we distinguish absolute commitments , conditional commitments and commitment attempts .

Absolute commitments are commitments whose fulfillment does not depend on any particular condition. An absolute commitment is denoted:

ABC ( Ag1 , Ag2 , tabc , ( λ , tλ )))

Two types can be distinguished: propositional commitments and action commitments .

A. Propositional Commitments

Propositional commitments are related to the state of the world. They are expressed by assertives or by speech acts of declaratory and expressive types. They can be directed towards the past, the present, or the future. We denote a propositional commitment as follows:

PC ( Ag1 , Ag2 , tpc , ( p , tp ))

where p is the proposition on which Ag1 commits.

Example:

U : The door is open

SA ( I0 , Ag1 , Ag2 , tu , U ) = def Create ( Ag1 , tu , PC ( Ag1 , Ag2 , tpc , ( open ( door ), tp )))

such that tpc = tp .

Because propositional commitments are particular cases of social commitments, the relationship between tu and tpc is similar to the one existing between tu and tsc .

Action commitments (also called commitments to a course of action ) are directed towards the present or the future and are related to actions that the debtor is committed to perform. The fulfillment and the violation of such commitments depend on the performance of the underlying action and the specified delay. This type of commitment is typically conveyed by promises. We denote an action commitment as follows:

AC ( Ag1 , Ag2 , tac , (( α , p ), tα ))

where α is the action to be performed, and by performing α , the proposition p becomes true. The relationship between the symbol action α and the proposition p is similar to the relationship existing between α and p in the operator < α > p of dynamic logic. Adding a proposition to the notation of an action commitment enables us to define the semantics of this commitment using this operator. This aspect will be detailed in Chapter 7. We notice here that p does not need a temporal argument because if α is performed at tα , then p becomes true at this moment.

Example:

U : I will give you 10 dollars in one hour

The state diagram of an absolute commitment is similar to that of Figure 5.5. Figure 5.6 presents the state diagram associated with the content of such a commitment. It contains the possible states for the commitment contents and the transitions corresponding to the operations, which can be applied to these contents. The dotted transitions in the figure correspond to the creditor’s actions and the non-dotted transitions correspond to the debtor’s actions. These operations are reflected by the participants’ utterances. Thus, the debtor can submit a commitment content, contradict it, justify it, defend it and change it. The creditor can accept this content, refuse it, challenge it and attack it.

A commitment towards the present or the future can be interpreted either as a propositional commitment or as an action commitment. For example, the utterance "tomorrow the door will be open" may be interpreted as a propositional commitment made by the speaker on a future state of the world. It can also be interpreted as an action commitment if the speaker is responsible for opening the door in question. Therefore, the commitment made by the speaker depends on the conversation context. It is in this sense that social context is a fundamental issue in communication (Moulin, 1998). In particular, this allows us to handle properly indirect speech acts (Bouzouba and Moulin, 1999).

In our framework, there is no explicit relation between propositional commitments and action commitments. When the current state of the world does not satisfy a propositional commitment, we speak about a violation of this commitment. There is no rule indicating that an agent develops an action commitment to make the content of its propositional commitment true when this commitment becomes violated. A propositional commitment is a commitment about a state of the world that the debtor agent can or cannot realize. In contrast, an action commitment is a commitment about an action that the debtor commits to perform in the present or in the future.

Absolute commitments do not consider the conditions that may restrain their fulfillment. However, in several cases, agents need to make commitments not in absolute terms but under given conditions. Another commitment type is therefore required in order to be able to capture situations defined by certain conditions. These commitments are said to be conditional . The structure of a conditional commitment , which must reflect the underlying condition, is different from the structure of a social commitment. We denote a conditional commitment as follows:

CC ( Ag1 , Ag2 , tcc , (( p , tp ), ( λ , tλ )))

This commitment expresses the fact that if p is true at time tp , then Ag1 will be committed towards Ag2 to perform λ or so that λ is true at time tλ . The future for a conditional commitment depends not only on time but also on the satisfaction of the underlying condition. Like for absolute commitments, we can distinguish between conditional commitments about propositions denoted:

PCC ( Ag1 , Ag2 , tpcc , (( p , tp ), ( p’ , tp’ )))

and conditional commitments about actions denoted:

ACC ( Ag1 , Ag2 , tacc , (( p , tp ), (( α , p’ ), tα )))

The relationship between the action symbol α and the proposition p’ is similar to the one existing between this action and this proposition in action commitments.

This distinction is implicit since according to the axiom ( A 1) a conditional commitment becomes an absolute commitment when the condition is satisfied.

where tabc = tcc .

The state diagram associated with a conditional commitment is similar to that of Figure 5.5. The state diagram associated with the content of such a commitment is identical to that of the content of an absolute commitment (Figure 5.6). Indeed, we can consider any social commitment as a conditional commitment whose underlying condition is always true. Thus, we have the following (syntactical) equivalence:

Example:

U : If industrial countries ratify the Kyoto Protocol, it can take effect

where tp’ = tp .

The commitments described so far directly concern the debtor who commits either that a certain fact is true or to perform certain action. For example, these commitments do not allow us to explain the fact that an agent asks another one to be committed to perform an action (by a speech act of a directive type). To solve this problem, we propose the concept of commitment attempt inspired by the notion of pre-commitment proposed in (Colombetti, 2000). We consider a commitment attempt as a request made by a debtor to push a creditor to be committed. Thus, when an agent Ag1 requests another agent Ag2 to do something, we say that the first agent is trying to induce the other agent to make a commitment. In this chapter, we denote a commitment attempt as follows:

CT ( Ag1 , Ag2 , tct , ( φ , tφ ))

where φ is the content of the commitment attempt. This formulation seems more intuitive than Colombetti’s one according to which the agent Ag2 is the debtor and the agent Ag1 is the creditor. In Chapter 7, we will improve this notation in order to be able to express the semantics of this type of commitments using an existential qualifier.

A commitment attempt about a proposition p is denoted:

PCT ( Ag1 , Ag2 , tpct , ( p , tp ))

A commitment attempt about an action α whose performance makes true a proposition p is denoted:

ACT ( Ag1 , Ag2 , tact , (( α , p ), tα ))

The relationship between the action symbol α and the proposition p is similar to the one existing between α and p in action commitments.

Example:

U : Could you call me at 4PM?

A commitment attempt is thought of as a type of social commitment because it conveys content which is made public once the attempt is performed. However, in our approach, there is a true commitment only after the creditor agent reacts in response to the commitment attempt by accepting it or by refusing it. We speak here about the " co-construction " of social commitments by the two interlocutors. This idea is similar to the one proposed by Rousseau, Moulin and Lapalme (1996) in which agents co-construct speech acts using their private mental states. The debtor and the creditor of a commitment attempt can act both on the attempt and on its content. On the one hand, the creditor agent reserves the right to accept a commitment attempt, to refuse it or to suspend it (for example by asking for a period of time for thought). It can also challenge the content of a commitment attempt. On the other hand, the debtor agent can withdraw a commitment attempt. It can also change the content of a commitment attempt and justify it. The states of a commitment attempt and those of its content can also be described by a state diagram. Figure 5.7 illustrates the state diagram associated to the content of a commitment attempt. Like a social commitment, a commitment attempt can be absolute ( ABCT ) or conditional ( CCT ). An absolute commitment attempt is denoted:

ABCT ( Ag1 , Ag2 , tabct , ( γ , tγ ))

A conditional commitment attempt is used for example when an agent asks another one to do some thing if a certain condition is true. A conditional commitment attempt about a proposition is denoted:

CCTP ( Ag1 , Ag2 , tcctp , (( p , tp ), ( p’ , tp’ )))

A conditional commitment attempt about an action is denoted:

CCTA ( Ag1 , Ag2 , tccta , (( p , tp ), (( α , p’ ), tα )))

where p is the underlying condition.

The refusal and the acceptance of a commitment attempt automatically lead to the creation of a new commitment that is in the active state. The two following rules illustrate this characteristic when the commitment attempt relates to a proposition or an action γ :

Syntactically, if λ is an action, ¬ λ indicates that this action will not be performed.

When the commitment attempt relates to a condition, the rules R 1 and R 2 become:

According to rule R 1’, refusing a commitment attempt which relates to a condition consists of refusing its content without committing towards its condition. However, according to rule R 2’, accepting a commitment attempt consists of accepting it under its condition, which leads to a conditional commitment.

Argumentation is based on the construction of arguments and counter-arguments (arguments attacking other arguments), the comparison of these various arguments and finally the selection of the arguments that are considered to be acceptable. A defeasible argumentation system essentially includes a logical language L , a definition of the argument concept, a definition of the attack relation between arguments and finally a definition of acceptability. In our model the formal definitions of these notions are inspired by (Elvang-Goransson et al., 1993). Here Γ indicates a knowledge base, stands for classical inference and ≡ for logical equivalence.

Definition 5.3 An argument is a pair ( H , h ) where h is a formula of L and H a sub-set of Γ such that: i ) H is consistent, ii ) H h and iii ) H is minimal, so that no subset of H satisfying both i and ii exists. H is called the support of the argument and h its conclusion.

Definition 5.4 Let ( H1 , h1 ) , ( H2 , h2 ) be two arguments.

( H1 , h1 ) attacks ( H2 , h2 ) iff H2 ¬ h1. In other words, an argument is attacked if and only if there exists an argument for the negation of its conclusion.

The concept of acceptability is defined as follows (Dung, 1995):

Definition 5.5 An argument ( H , h ) is acceptable for a set S of arguments iff for any argument ( H’ , h’ ): if ( H’ , h’ ) attacks ( H , h ) then ( H’ , h’ ) is attacked by S.

Intuitively, an argument is acceptable if it is not attacked, if it defends itself against all its attackers, or if it is defended by an acceptable argument.

According to (Dung, 1995), any argumentation system includes two essential elements: one element is used to build and generate arguments, the other is used to analyze these arguments by determining their acceptability. This view is important for our communication model. Indeed, agents must reason about their own mental states in order to build arguments in favor of their future commitments, as well as about other agents’ commitments in order to be able to take position with regard to the contents of these commitments. Surely, an argumentation system is essential to help agents to act on commitments and on their contents. However, reasoning about other social attitudes should be taken into account in order to explain agents’ decisions. This aspect will be discussed in Chapter 9, in which we highlight the importance of agents’ trustworthiness to decide, in some cases, about the acceptance of arguments.

The systems proposed in the literature, for example in (Dung, 1995), (Vreeswijk, 1997), (Amgoud, 1999) do not discuss how arguments can support communicative actions. We will specify this here. In fact, before committing to some proposition h being true (i.e. before creating a commitment whose content is h , the speaker agent must use its argumentation system to build an argument ( H , h ). On the other side, the addressee agent must use its own argumentation system to select the answer it will give (i.e. to select the appropriate manipulation of the content of an existing commitment). For example, an agent Ag1 accepts the commitment content h proposed by another agent Ag2 if Ag1 is able to build an argument which supports this content from its knowledge base which is assumed to be consistent. If Ag1 has an argument ( H’ , ¬ h ), then it refuses or attacks this commitment content. If Ag1 does not have any argument for h , or for ¬ h , then it must ask for an explanation. In this case, Ag2 must justify the content h .

Thus, an agent should always use its argumentation system before creating a new commitment or positioning itself on an existing commitment and on its content. Consequently, an argument of an agent Ag1 must support an action of this agent on a given commitment and/or on its content. Formally, an agent Agk ’s argument supporting its action at time tu on a given commitment is denoted:

Arg ( Agk , H , Act ( Agk , tu , SC ( Agi , Agj , tsc , ( φ , tφ ))))

An Agk ’s argument supporting its action at time tu on a given commitment content is denoted:

Arg ( Agk , H , Act-content ( Agk , tu , SC ( Agi , Agj , tsc , ( φ , tφ ))))

with H being the support of the argument and the agent identifiers i , j and k verify:

i , j , k ∈ {1, 2} and i j .

In the first formula, H is the support of the action Act performed by agent Agk on the commitment identified by tsc . In the second formula, H is the support of the action Act-content performed by agent Agk on the content of this commitment. We notice that this support holds at the moment of the action. Thus, according to the nonmonotonicity of arguments, it is possible that this support becomes invalid if new information becomes available for Agk . In this case, Agk must update its knowledge base by removing the invalid argument and adding the new valid argument.

We notice that there is a logical relation between arguments supporting actions and arguments supporting propositions. The argument supporting an action is the argument supporting the proposition that becomes true when the action is performed. This relation is similar to the relation existing between actions and propositions in a dynamic logic (Harel, 1984). In this logic, the semantics of an action is defined as follows:

This means that in a Kripke structure (the model) the action is satisfied in a world iff there is an accessible world in which the proposition becomes true ( is called accessibility relation ). The idea is that by doing the action the proposition becomes true in an accessible world. In our approach an argument H supporting an action Act (respectively an action Act-content ) performed on a commitment whose content is (respectively on the content ) is satisfied in a world iff there is an accessible world in which H supports or is the accessibility relation associated with the action Act (respectively Act-content ). For example, the argument supporting an acceptance action of a social commitment content is the argument supporting this content.

In fact, the relation between H and the commitment content depends on the values of Act and Act-content . Thus, for an absolute or a conditional commitment we have the following axiom:

For example, the first rule indicates that if Act takes the value " Create " or " Satisfy ", then H supports .

To illustrate this idea, let us take the following example between agents Ag1 and Ag2 that we dealt with in Section 5.3:

U1 : The book is not allowed during the test.

U2 : Why?

U3 : Because the answers are given in this book.

U4 : Ok, Thank you.

We suppose that the Ag1 ’s knowledge base contains the arguments ( H , φ ) and ( φ , φ ), and the Ag2 ’s knowledge base contains the argument ( φ , φ ) where H = Give ( Answers , Book ) and = ¬ Allow ( Book , Test ). By utterance U3 , agent Ag1 presents the support H in order to justify the content φ of the commitment identified by tsc 1. Formally we have:

For a commitment attempt we have the following axiom:

An agent can create a commitment attempt related to a proposition p , if it does not have any argument for p or for ¬ p . This reasoning is also valid for a commitment attempt related to a condition or an action. In the case of an action α the agent does not have any argument for or against the proposition p that becomes true by performing α . We notice here that this aspect cannot be verified because the fact that an agent does not have an argument for or against a proposition is related to its internal state. The idea is that an agent can create a commitment attempt related to a proposition p even if it has an argument for p . In addition, the creation of a commitment attempt related to an action can also depend on the context. For example, to create a commitment attempt in the form of an order, the debtor must have the social capacity to give orders to the other agent.

When considering the creation of a new commitment, the agent must also have a reason supporting it (a kind of goal to be achieved). In our approach, this reason is considered as an argument for the action, which is different from the argument that supports the content. Let us take the following example:

To create this commitment, Ag1 must have a reason to do it, as for example in order to ask Ag2 to buy the book. This reason can be considered as an argument supporting the creation action that is different from the argument supporting the content, corresponding for example to "this book is interesting because its editors are well known authors". It is thus significant in this case to distinguish the argument supporting the creation action itself and the argument supporting the content. Generally, the arguments supporting the creation actions are not expressed in speech acts, but correspond to agents’ private goals. We define an argument supporting a creation action as follows:

Definition 5.6 An argument supporting a creation action of a commitment is a pair ( , Create ( Ag1 , tu , SC ( Ag1 , Ag2 , tsc , ( , tφ )))) such that i) is a proposition representing an agent Ag1’s goal and ii ) .

The proposition can be any Ag1 ’s goal. For example, the goal can be just to inform another agent that some thing is true.

After the introduction of the argumentation in our approach, we note that a speech act can lead to an action not only on a commitment as explained in Section 5.2, but also on an argument. An agent can thus accept, refuse, defend or attack an argument. Thus, using BNF notation, we have the following definition improving Definition 5.1:

In Chapter 6 (Section 6.3), we will give a detail example illustrating this definition.

In the previous sections we proposed a formulation of the pragmatics of agent communication using the notion of actions that agents perform on commitments and the arguments enabling agents to select the communicative act to be performed. In this section, we propose an architecture of a communication model in which this approach takes place. In fact, this model combines the three approaches discussed in our taxonomy of the prior approaches (see Chapter 4). It is based on a hybrid approach that we call MSA (Mental-Social-Argumentative). Indeed, if they are taken individually, these three approaches do not allow us to model all the aspects of agent communication. For this reason, we suggest to combine them in a unified approach. In addition, conversation is a cognitive and social activity, which requires a mechanism making it possible to reason about mental states, about what other agents say (public aspects), and about the social aspects (conventions, standards, obligations, etc). These three approaches are thus not exclusive but rather complementary.

The MSA approach has the advantage of capturing simultaneously the mental aspect that characterizes agents participating in a conversation, the social aspect that reflects the context in which these agents communicate, and the reasoning aspect which is essential to be able to take part in conversations. The combination of commitments and arguments seems essential to us because agents must be able to justify the claims to which they are committed and to justify their actions on commitments. This justification cannot be made if agents do not have the necessary argumentation mechanisms. In addition, the combination of commitments and private mental states is necessary because public commitments reflect these mental states that contain additional information motivating the agent’s communicative acts. Finally, the combination of argumentation and mental states is significant because agents have to reason about their mental states before committing in a conversation.

The communication model is composed of three layers: the conversation layer, the commitment/argument layer and the cognitive layer (see Figure 5.8). The abstraction levels justify this stratification in layers. The conversation layer is directly observable because it is composed of the speech acts that agents perform. These acts are not performed in an isolated way, but within a particular conversation. The commitment/argument layer is used to correctly manage the social commitments and the arguments that are related to the conversation. Finally, the cognitive layer is used to take into account the private mental states of agents, the social relations and other elements that agents use in order to communicate.

In order to allow conversational agents to suitably use the communication model, this model must be compatible with the agent architecture. Thus, we propose an architecture of conversational agents, which is composed of three models: the mental model, the social model and the reasoning model (Figure 5.8). The mental model includes beliefs, desires, goals, etc. The social model captures the social concepts such as conventions, roles, etc. Social commitments constitute a significant component of this model. The commitments that an agent makes public when performing speech acts are different from the private mental states, but these two elements are not independent. Indeed, social commitments reflect mental states. Thus, agents must use their reasoning capabilities to reason about their mental states before producing or manipulating social commitments. The agent’s reasoning capabilities are represented by the reasoning model using an argumentation system. The conversational agent model also involves by general knowledge, such as the knowledge of the conversation subject. An agent will use this knowledge in order to build the common ground that it must share with its partners. The notion of common ground introduced by the philosophers of language Clark and Haviland (1974) indicates the set of knowledge, beliefs, and presuppositions, which agents believe that they share during their conversations.

In this chapter, we argued that the three approaches discussed in Chapter 4 can be successfully combined in one pragmatic approach. This unified approach has the advantage of capturing the external public aspect of agent communication and the internal private aspect of agents. The main idea of this approach is that agent communication is considered as actions that agents perform on social commitments and arguments. The dynamics of agent conversation is represented by this notion of actions. In addition, the notion of commitment state enables us to reflect the evolution of agent conversations. The current state of a conversation is clearly described by the state of the different commitments. The use of argumentation allows agents to participate in complete conversations because at each moment they can select the next action to be performed. This approach can be used to specify protocols that are more flexible than classical protocols in the sense that participating agents can make decisions by reasoning about the current state of their conversations. Because it captures the private and the public aspects of agent communication, this approach can also be used to specify the different dialogue types according to the classification proposed by Walton and Krabbe (1995) (see Chapter 2). Thus, in Chapter 9, we will show how it can be used to specify dialogue games in the case of a persuasion dialogue game protocol. We also used this approach to develop a computational model for the dialogization and the implicit information in a communicational model (Bouzouba et al., 2004). A comparison between our approach and other proposals will be made in the next chapter.



[3] The term " taking position " is inspired by the work done by Rousseau, Moulin and Lapalme (1996) and extended by Bouzouba and Moulin (1998) and Bouzouba, Moulin and Kabbaj (2001). In these proposals, agents communicate by taking positions on the agents’ private mental states which are exchanged by agents while conversing.

© Jamal Bentahar, 2005