Guessing What Sentences Mean
Speaker:Tommy Herbert
Institution: University of Edinburgh
Abstract:
I'm trying to give machines the ability to turn an ordinary sentence into a logical form - a semantic representation. A deep representation would give the computer access to information like semantic scope, predicate argument structure, and temporal and spatial relations. If the system was also robust enough to handle most sentences, it would be useful for tasks like automating e-mail responses, building dialogue systems and acquiring ontologies.
The Redwoods corpus contains about 20,000 sentences, each annotated with a semantic representation. My method is to apply machine learning techniques, supplemented with some symbolic modelling, to this corpus and acquire a mapping from natural language strings to logical forms.
The first step is estimating the lexical predicates in the sentence, and I'll talk about methods for doing that and present some experimental results.