Sepia : semantic parsing for named entities

Abstract

People's names, dates, locations, organizations, and various numeric expressions, collectively called Named Entities, are used to convey specific meanings to humans in the same way that identifiers and constants convey meaning to a computer language interpreter. Natural Language Question Answering can benefit from understanding the meaning of these expressions because answers in a text are often phrased differently from questions and from each other. For example, "9/11" might mean the same as "September 11th" and "Mayor Rudy Giuliani" might be the same person as "Rudolph Giuliani". Sepia, the system presented here, uses a lexicon of lambda expressions and a mildly context-sensitive parser to create a data structure for each named entity. The parser and grammar design are inspired by Combinatory Categorial Grammar. The data structures are designed to capture semantic dependencies using common syntactic forms. Sepia differs from other natural language parsers in that it does not use a pipeline architecture. As yet there is no statistical component in the architecture. To evaluate Sepia, I use examples tp illustrate its qualitative differences from other named entity systems, I measure component performance on Automatic Content Extraction (ACE) competition held-out training data. and I assess end-to-end performance in the Infolab's TREC-12 Question Answering competition entry. Sepia will compete in the ACE Entity Detection and Tracking track at the end of September.