Experiments in Questions and Relationships at the University of Iowa

Our general approach to question answering is one that attempts to front-load much of the evaluation of corpus semantics, allowing for a more search engine-like user experience. By shifting natural language parsing forward in the process, we can amortize this very expensive step against a number of downstream extraction processes that mine the text for named entities, relationships, etc. Redefinition of extraction specifications hence does not require reparsing of the source text. As for last year, we use a tgrep-like extraction grammar designed for predicate-based extensibility using it in mapping sentence parse trees to relational structure. This overall approach handles not only factoid answers, but definitional answers and those requiring inference across multiple extracted relationships.