Towards a scalable theory of alternative utterances
Pragmatic language comprehension often requires reasoning about alternative utterances. For example, we take “Bob ate some of the cookies” to imply that Bob ate some *but not all* of the cookies, as the utterance “Bob ate all of the cookies” is a viable alternative that the speaker chose not to say. However, existing theories of how alternatives are determined fail to explain how humans perform this type of inference in unbounded, naturalistic domains. I propose a series of experiments to implement and test candidate theories at scale, with the goal of improving NLP models as well as our understanding of the cognitive basis of pragmatics.
Elucidating the structure of the comprehender’s noise model
Recent advances in human language comprehension show that people sometimes entertain nonliteral interpretation of sentence by rationally integrating prior semantic expectation and the likelihood that the sentence is corrupted by noise. This inferential process is referred as noisy channel processing. For example, if you hear “The mother sent the gift her daughter”, you might infer that the speaker accidentally omitted a “to”, and she actually meant “The mother sent the gift to her daughter“. This talk explores some other noisy operations people might consider besides the deletion and insertion proposed by previous literatures, and also how the available alternatives to the perceived sentences affect the rate of nonliteral interpretations.