Speech Act Therory and Epistemic Planning

This paper reviews the use of speech act theory in AI natural language systems, arguing that attempts to embody speech acts as simple STRIPS-like operators are doomed to failure. The diiculties arise partly from the fact that there is no reliable way of associating speech acts and surface forms, and partly from properties of epistemic logic. 1 Background The aim of the current paper is to serve as a warning. When you are trying to think about the ways language gets used, you rapidly encounter situations where what the speaker is trying to achieve by an utterance is clearly at odds with the utterance's standard interpretation. If someone says \Can you pass the salt?" they usually mean \Will you pass the salt?", if they say \It's open" when you knock on the door they usually mean \Come in", and so on. Clearly, in order to understand what is going on in such cases you have to start reasoning about knowledge and belief, about rational action, and about the interactions between the two. But what you don't need to do is to come up with a nite set of \speech acts" lacking in umanbiguous surface identiiers and plug them into the kind of AI planning algorithm that stems from the (extremely interesting and powerful) work of (Fikes and Nilsson, 1971). My sole aim in this paper is to try to divert research eeort from embodied as STRIPS-like operators and plugged into some more or less sophisticated planning system. In order to do this, I will show that a particular strand of research in this area introduces operators which have no eeects and which can only be identiied by doing the kind of reasoning which they were supposed to obviate. I will not, however, propose an alternative. This work has been extremely innuential, and has led to the development of numerous very appealing systems (e.g. (Appelt, 1985; Grosz and Sidner, 1986)). The argument below is simply intended to set some limits on the use of simple minded planning algorithms in this context, and to suggest a set of research problems. Part of the motivation for the analysis presented here is that I do not believe that straightforward application of AI planning theory to the interpretation of speech acts is going to scale up, for quite concrete reasons which are developed below. But I also feel that there is a version of Occam's …