back to

Seminar: 2/3 - Michael Frank, Stanford University

11:00 to 12:30 PM      at:  5101 Tolman Hall

A probabilistic framework for referential communication

A short, ambiguous message can convey a lot of information to a listener who is willing to make inferences based on the identity of the speaker and the context of the message. Pragmatic inferences of this type are crucial to efficient human communication, and have been characterized informally using tools like Grice's conversational maxims. They may also be extremely useful for word learning. In this talk, I'll propose a probabilistic framework for referential communication based on information theory and present results from a variety of communication tasks with adults and children. In experiments with restricted contexts, the framework allows us to predict judgements about both speakers' word choices and the meanings of novel words with high accuracy. I'll also present some ongoing work that begins to extend this framework to more natural, multi-word utterances and linguistic phenomena like scalar implicature.