Experiments with Web QA System and TREC2004 Questions
Dmitri G. Roussinov, José Antonio Robles-Flores, Yin Ding · 2004
We describe our first participation in TREC. We only competed in the Question Answering category and limited our runs to factoids. Our approach was to use an open domain QA system that finds the answer among Web pages indexed by a commercial search engine, then project the answer to the TREC test collection. Our Web QA takes advantage of the redundancy of the Web, obtains the candidate answers by pattern matching, then performs probabilistic triangulation of them to assign a final score. Our novel contributions are 1) the probabilistic triangulation algorithm, 2) more powerful pattern language than presented in prior research, and 3) use of semantic features of expected answers instead of relying on an elaborate hierarchy of question types. Although, we were able to run only first 91 out of 230 factoid questions before the submission deadline, we find our result encouraging, and if interpolated to the entire questions set, would be above the mean performance on factoid questions.