BREAKTHROUGH IN QUESTION-ANSWERING AT THE NINTH TEXT RETRIEVAL CONFERENCE (TREC-9).

At the NIST-hosted TREC-9, researchers experienced a breakthrough in text retrieval technology that could result in improved products on the Web. In the second running of the Question-Answering track, the goal was to retrieve text snippets rather than documents in response to a question. Over 500 questions drawn from donated query logs were used for testing. These real questions made the task noticeably more difficult than the TREC-8 task, but despite the increased difficulty, the best-performing system was able to extract a correct answer to the question more than 65 % of the time. This system, from Southern Methodist University in Dallas, Texas, demonstrated the power of integrating multiple natural language processing techniques with abductive reasoning to produce a final score almost twice that of the other 27 groups that participated in the testing. Based on past TREC experience, this type of breakthrough will lead to major changes in the techniques used by systems in next year's TREC and then to improve d question-answering products on the Web.