Goals and Actions in Natural Language Instructions

Human agents are extremely flexible in dealing with Natural Language instructions: they are able both to adapt the plan they are developing to the input instructions, and vice versa, to adapt the input instructions to the plan they are developing. Borrowing the term from [Lewis 1979], I call this two-way adaptation process accommodation. In this proposal, I first define accommodation in the context of processing instructions. I then provide evidence for the particular inferences I advocate, and for the further claim that such inferences are directed by the goal to achieve which certain action is performed. The evidence I provide comes from my analysis of naturally occurring instructions, and in particular of purpose clauses and of negative imperatives. Finally, I propose a computational model of instructions able to support accommodation inferences. Such model is composed of: a speaker / hearer model of imperatives, based on the one presented in [Cohen and Levesque 90]; an action representation formalism based on a hybrid system, á la KRYPTON [Brachman et al. 1983a], whose primitives are those proposed in [Jackendoff 1990]; and inference mechanisms that contribute to building the structure of the intentions that the agent develops while interpreting instructions.