PSICOM: Pragmatics and Semantics In Cognitive Modelling

Informativity biases in models of meaning selection

Informativity biases in models of meaning selection
Hannah Rohde
University of Edinburgh

If we take semantics to encompass the study of meaning in communication, a question arises as to which meanings are favored by interlocutors. While a range of candidate meanings may be possible and even plausible, how do speakers select which meanings to convey and how do listeners make guesses as to the most probable meaning when they try to recover what a speaker has said or when they try to anticipate what a speaker will say next?  

In this talk, I compare two hypotheses for ranking candidate meanings. Under one account, listeners’ guesses simply reflect the probability that different meanings hold true: Speakers are taken to generate sentences that describe the world they see and listeners come to expect sentences about the typical situations speakers find themselves in. A second account combines this component for truth with a component capturing the likelihood that a speaker, knowing some meaning to be true, would select that meaning as one worth conveying to a listener in a spoken utterance. I present a series of psycholinguistic studies measuring listeners’ awareness of speakers’ production likelihoods. For example, although bananas are prototypically yellow, speakers rarely mention this yellowness in their utterances. In an eye-tracking study measuring anticipatory looking, listeners who hear a speaker use a color adjective are found to anticipate subsequent mention of an object for which that color is less typical in the real world. Further studies contrast listeners’ estimates of speakers’ beliefs about the world versus their guesses about the content of the speakers’ utterances, showing how the latter diverge from real-world priors and do so in context-dependent informativity-driven ways.

These findings highlight the importance of establishing not only which meanings are possible and how they are derived, but also which meanings are probable as contributions to coherent discourse, despite — or perhaps as a result of — denoting real-world-improbable situations.