Cognitive Science Dinner Talk

Aaron White

Department of Linguistics, University of Rochester

366 Meliora Hall

Frequency, acceptability, and natural language ontology

Human experience burgeons with complex relationships among objects, both abstract and concrete. Human languages allow us to label these relationships in extremely specific ways: belief is different from doubt is different from hope. In contrast to the expressiveness we see among words with relational meanings, the space of primitive semantic types that enter into such relational meanings ? e.g. individual, event, proposition ? is often assumed to be relatively coarse-grained. This assumption rests, in large part, on data from predicates that take nominal and adpositional arguments, often ignoring distributional complexity in the domain of verbal and clausal arguments.

In this talk, I focus on two questions: (i) what is the granularity of the primitive semantic type space; and (ii) what is the distribution of complex semantic types constructed from those primitive types. To investigate these questions, I leverage two datasets that cover the syntactic distribution of the entire English verbal clause-embedding lexicon: one based on frequencies extracted from a large corpus and another based on a large acceptability judgment task. Using multi-task matrix factorization models that encode minimal assumptions about the structure of the semantic type space, I demonstrate that the semantic type information carried by both types of distributions is substantially more complex than the standard assumptions suggest. I discuss the implications of these findings for theories of the lexicon.