Friday, January, 25, 2008, 12-2 p.m.

Human communication as a predictive process

What is going on in the mind of someone who is speaking, or listening to someone else speak? In this talk, I’ll explore the idea that when humans communicate, they engage in a process of joint prediction. When talking, speakers use a rich set of cultural and experiential priors to produce behavior that they predict will change the beliefs or behavior of others. Speakers use semantic cues to activate appropriate linguistic units. These words and chunks, along with other developing contextual cues, then serve to activate subsequent linguistic units as speakers generate the utterances they believe are most likely to bring about changes in listeners' beliefs or behavior. At the same time, listeners, far from being passive decoders of tokens of meaning, are using broadly the same process to predictively build up their understanding of what is being said. Listeners use both learned semantic cues to words, and words themselves as cues to other words, in order to predict the behavior and intentions of speakers. Successful communication thus relies both on the collaboration between speaker and listener, and the degree to which shared prior knowledge enables mutual predictability.

An attractive property of this approach is that it allows human communication to be couched in terms compatible with theories of learning. I’ll describe a series of computational and empirical studies in which learning models are used to predict and explain surprising human performance on tasks related both to learning word meanings and to the processing of multi-words utterances. Conceiving of communication as prediction offers a solution to a puzzle first noted by Darwin: why do English speaking children struggle to learn the meanings of color words? I’ll show how these struggles can be remedied by changes to the ordinary configuration of semantic and linguistic cues employed in communicating in English. Finally, I’ll outline some of the developmental peculiarities of human brains that underlie our remarkable capacity for learning communicative priors, and by relating these to formal learning models, suggest why these abilities seem to be subject to a sensitive period.