Question Answering with Lydia (TREC 2005 QA Track)

The goal of our participation in TREC 2005 was to determine how effectively our entity recognition/text analysis system, Lydia (http://www.textmap.com) [1–3] could be adapted to question answering. Indeed, our entire QA subsystem consists of only about 2000 additional lines of Perl code. Lydia detects every named entity mentioned in the AQUAINT corpus, and keeps a variety of information on named entities and documents in a relational database. We can collect candidate answers by means of information kept in the database. To produce a response for the main task or a ranked list of documents for the document ranking task, we rank the collected candidate answers or documents using syntactic and statistical analyses. A significant distinction from other question answering systems [4–6] presented earlier at TREC is that we do not use web sources such as Wikipedia and Google to generate candidate answers or answers. Rather, we only use syntactic and statistical features of the test set of questions and corpus provided. Our approach is independent of other sources, and finds answers from the text provided. We describe the design of Lydia and associated algorithms in Section 2, and focus on the design and algorithms of the QA system in Section 3. We then analyze the performance of the QA system in Section 4, and conclude this paper with discussion on future directions in Section 5.