Speech act theory

Speech act theory [6] [65] treats language as action, so that it can be planned just as any physical action can. In common with physical acts, speech acts have certain preconditions, and can be said to have effects in the environment. In contrast to physical acts, the effects are more in the mental state and subsequent actions of the hearer rather than a direct physical effect. A speech act is a composition of different acts. The locutionary act is the act of saying something, whereby the hearer becomes aware of an utterance. The success of the speech act depends on the success of the locutionary act. Associated with the locutionary act is an illocutionary act. There are different types, of differing "illocutionary force", often identified by different explicit performative verbs, such as "inform" or "request". To define illocutionary acts within the STRIPS planning system, Bruce [8], Perrault and Allen [51] and Allen and Perrault [1] developed a set of schemas. They defined a request act, whose precondition was that the speaker intends the act being requested, and whose effect was that the hearer intended the act. They also defined a family of inform acts, whose preconditions were that the speaker knew the proposition being informed and whose effects were that the hearer knew the proposition. These schemas are repeated here.

name:           request
parameter:      P
precondition:   intend(speaker,P)
effects:        { intend(hearer,P),
                  know(hearer,intend(speaker,P))
                }

name:           inform
parameter:      P
precondition:   know(speaker,P)
effect:         { know(hearer,P),
                  know(hearer,know(speaker,P))
                }

name:           informref
parameter:      TERM,P
precondition:   knowref(speaker,TERM,P)
effect:         { knowref(hearer,TERM,P),
                  know(hearer,knowref(speaker,TERM,P))
                }

name:           informif
parameter:      P
precondition:   knowif(speaker,P)
effect:         { knowif(hearer,P),
                  know(hearer,knowif(speaker,P))
                }

While this was a good first cut, there are some problems, the most important of which is that they ignore the autonomy ( free will ) [78] of the hearer in deciding whether to believe what is informed, and deciding whether to intend what is requested. This is due to the use of effects lists. Appelt, in his logical formulation of the planning of speech acts, also uses hard-coded effects lists to realise effects [5]. A weaker and more acceptable set of effects is that the hearer only believes that the speaker believes P or that the hearer believes that the speaker intends P. Even these are too strong, since if an agent is allowed to insincerely request and inform, the hearer might not recognise sincere acts as being sincere, and so their effects would not occur. It seems that effects lists need to be done away with altogether, and this approach is successfully taken in this thesis. A second problem is the use of the term "know" instead of "believe" in the preconditions. For example, suppose agent A believes $\lnot(P)$, while agent B believes $P$. While it is reasonable that agent B could inform agent C of P, the plan rules would not allow agent A to generate this expectation, since agent A could not say that agent B knows P. The preconditions should instead only refer to the speaking agent's belief. This principle will be adopted in designing the planner's negotiation acts. A third problem is that the effects ignore the nested beliefs of the agents. For example, an effect of the inform act should be that the speaker believes that the hearer believes P, and so on.

In speech act theory, the perlocutionary effect of an act is the state that the system reaches as a result of the act. While Allen and Perrault ignore the hearer's autonomy and define this using the STRIPS effects list, the perlocutionary effect in fact depends on the following choices of the hearer. For instance, the perlocutionary effect of an inform act depends on the hearer's choice of how to revise his beliefs, and what subsequent action he will take as a result. Grice [27] accommodates the hearer's response when he describes meaning as the "effect the speaker intends by the hearer's recognition of that intent". This description motivates the speaker's use of plan recognition and response from the hearer's perspective when planning an illocutionary act, rather than computing the effect using STRIPS effects lists. Without considering the hearer's response, the speaker could not know the meaning of his utterance.