Computer Science Thesis Proposal

Wednesday, January 20, 2016 - 10:00am


Traffic 21 Classroom 6501 Gates & Hillman Centers



For More Information, Contact:

Historically, shallow semantic parsing has focused on tagging individual predicates expressed by individual lexical units. While fruitful, this paradigm cannot consistently and comprehensively represent semantic relations that take a wide variety of linguistic realizations. It also struggles to represent relations where multiple meanings are entangled in the same expression. One relation type that exhibits both challenges is causation: causation can be expressed by a variety of individual words, multi-word expressions, or patterns of words and syntactic relations that may even span multiple clauses; and it competes for linguistic space with other phenomena such as temporal relations and obligation. To expand shallow semantic parsing to such challenging relations, my thesis will present approaches for annotating and tagging causal language, based on the emerging linguistic paradigm known as Construction Grammar (CxG). CxG places form/function pairings called constructions at the heart of both syntax and semantics, allowing the semantics to be anchored to an enormous variety of forms. In this proposal, I describe a CxG-based scheme and methodology for annotating explicit causal relations in English, and two computational methods for automatically tagging these relations. Both methods combine automatically induced rules for pattern-matching with statistical classifiers. I also describe experiments to date that demonstrate the potential of these methods and the ways in which they can be improved. I propose expanding the methodology to allow each construction instance to be annotated with multiple semantic relations, thus incorporating some of the phenomena intertwined with causality. I also propose improvements to the existing pipelines. Finally, I propose a third tagging method based on more robust structured prediction techniques, in which a syntactic parser based on neural networks can be trained to decorate a parse with construction annotations. Thesis Committee:Jaime G. Carbonell (Co-Chair) Lori Levin (Co-Chair) Ed Hovy Nianwen Xue (Brandeis University) Copy of Proposal Document


Thesis Proposal