A comparison of hard filters and soft evidence for answer typing in Watson

Abstract

Questions often explicitly request a particular type of answer. One popular approach to answering natural language questions involves filtering candidate answers based on precompiled lists of instances of common answer types (e.g., countries, animals, foods, etc.). Such a strategy is poorly suited to an open domain in which there is an extremely broad range of types of answers, and the most frequently occurring types cover only a small fraction of all answers. In this paper we present an alternative approach called TyCor, that employs soft filtering of candidates using multiple strategies and sources. We find that TyCor significantly outperforms a single-source, single-strategy hard filtering approach, demonstrating both that multi-source multi-strategy outperforms a single source, single strategy, and that its fault tolerance yields significantly better performance than a hard filter. © 2012 Springer-Verlag Berlin Heidelberg.