Question Answering Using a Large NLP System

The Microsoft Research question-answering system for TREC-9 was based on a combination of the Okapi retrieval engine, Microsoft’s natural language processing system (NLPWin), and a module for matching logical forms. There is no recent published account of NLPWin, although a description of its predecessor can be found in Jensen et al. (1993). NLPWin accepts sentences and delivers a detailed syntactic analysis, together with a logical form (LF) representing an abstraction of the meaning. The original goal was to construct a framework for complex inferencing between the logical forms for questions and sentences from documents. Many answers can be found with trivial inference schemas. For example, the TREC-8 question What is the brightest star visible from Earth? could be answered from a sentence containing ... Sirius, the brightest star visible from Earth ...by noting that all of the content words from the question are matched, and stand in the same relationships in the question and in the answer, and that the term Sirius is equivalent to the answer's counterpart of the head term in the question, star. The goal of using inferencing over logical forms was to allow for more complex cases, as in Who wrote the play ``Hamlet''? which should not be answered using ... Zeferelli's film of ``Hamlet'' since a film is not a play. The idea of using inferencing for question-answering is not new. It can be found in systems from the 1970s for story understanding (Lehnert, 1978) and database querying (Bolc, 1980), and in more recent work for questions over computer system documentation (Aliod, 1998).