Factual Question Answering by Automatic Surface Pattern Learning Using Reformulation Rules

摘要

Many of the high performing factual question answering systems in the recent TREC(Text Retrieval Conference)`s use a fairly extensive list of surface text patterns. In this study, an automaticsurface pattern learning using reformulation rules is proposed. In essence, this is an adaptation of surfacepattern learning first proposed by Deepak Ravichandran and Hovy. In our proposed system, predefined setsof representative question and answer patterns, instead of question answer pairs, are used for answerextraction. The performance of the modified system is measured by using two conventional and standardmetrics-MRR (Mean Reciprocal Rank) and precision. The system`s performance is also contrasted with that ofHovy`s, so as to elicit improvements due to proposed modifications, using the same metrics.