Steps towards Incremental Semantics for Spoken Dialog Systems

Traditionally, spoken dialog systems have interpreted a user's speech one complete utterance at a time, and operated on one level of processing at a time as well. It is clear, however, that people understand language incrementally: they can backchannel, interrupt, and begin taking actions while the speaker's utterance unfolds. It has also become clear that human language understanding involves the rapid integration of multiple sources of information including syntax, semantics, and pragmatics. On the system side, as a reviewer pointed out, deployment scenarios such as driving or flight simulation would present cases where incremental understanding would nicely enable features such as beginning to act on partial commands, or changing the goal of an action as more information arrives. We have recently shown that computational methods for incremental understanding have various advantages over their nonincremental counterparts, including better parsing (Stoness et al. 2005). In this brief note, we describe work on representations of natural language semantics for incremental understanding. First we describe how rampant use of free variables helps represent not only partial sentences, but utterances as they arrive word-by-word. Second, we show how the commonalities in the meaning of phrases such as those used to indicate fine-grained adjustments (“a bit more to the right”) can be captured in semantic representations. Third, we show how allowing actions (“moving to central park”) to be states can cleanly and concisely represent otherwise messy facts such as interrupted actions. These are three initial steps towards incremental semantics for spoken dialog systems. 2 Semantics for Partial Sentences and for Incremental Processing