Answer Extraction for Definition Questions using Information Gain and Machine Learning

Extracting nuggets (pieces of an answer) is a very important process in question answering systems, especially in the case of definition questions. Although there are advances in nugget extraction, the problem is finding some general and flexible patterns that allow producing as many useful definition nuggets as possible. Nowadays, patterns are obtained in manual or automatic way and then these patterns are matched against sentences. In contrast to the traditional form of working with patterns, we propose a method using information gain and machine learning instead of matching patterns. We classify the sentences as likely to contain nuggets or not. Also, we analyzed separately in a sentence the nuggets that are left and right of the target term (the term to define). We performed different experiments with the collections of questions from the TREC 2002, 2003 and 2004 and the F-measures obtained are comparable with the participating systems.