Trainable question-answering systems
Abraham Poovakunnel Ittycheriah, Richard J. Mammone · Rutgers University eBooks · 2001
Research into automatic question answering systems has become increasingly popular as the internet age ripens. With current search engine technology and the explosion of web content, searching for a particular information need results in large amounts of information retrieved which must then be manually filtered and further searched. Question answering systems help to narrow down this search and in the optimum case presents just the desired answer. One alternative use of question answering systems is in the education field, to help satisfy the queries of students, relieving the work of teaching assistants and professors. This dissertation focuses on investigating statistical approaches to make a trainable question answering system. A question is first analyzed and a prediction is to be made as to what type of answer the user is expecting. For this, a maximum entropy model is derived using features of unigrams and bigrams and word position. Semantic classes of the question focus are shown to improve performance. Second, a fast search of the text database is performed and the top documents relevant to the query are retrieved. Finally, the answer tag prediction and the top documents are input to the answer selection stage. In this thesis, we investigate and report results on a trainable answer selection algorithm. The architecture presented here is similar to those investigated by other participants of the TREC-8 (Text Retrieval and Evaluation Conference) question and answering track. A new formal mathematical formulation of question answering is presented. The new formulation uses a training procedure which is independent of the type of questions. Results of the system are presented on the questions from the TREC conferences in 1999 and 2000.