Doctoral Dissertation Computation of Semantic Equivalence for Question Answering

There are vast amounts of text data on World Wide Web and in encyclopedias and newspapers. For efficient information access to such text data, Question Answering is an option. Question Answering which this thesis uses is defined that a technique to find an answer from a large amount of text data for a given question written in natural language. Evaluation workshops for Question Answering started in 1999 have motivated many researchers to improve techniques for the task. We take an approach to Question Answering task based on the semantic equivalence, regarding the problem of Question Answering as a computation of semantic equivalence between a question and text. We proposed evaluate methods of semantic equivalence in three levels, (1) approximation of semantic equivalence by structural information of text, (2) evaluation of semantic equivalence using lexico-structural transformation, (3) evaluation of semantic equivalence by extracted semantic relations. At the first level, we conducted structural matching as an approximation of semantic equivalence instead of bag-of-words matching which is utilized for Question Answering mainly in previous works. The proposed method, an extension of Collins’ Tree Kernel, provides three options for similarity measurement and relaxes the constraints of matching with the time complexity O(|T1||T2|). For the second level, we developed a Japanese lexico-structural paraphrasing engine called Kura. We also proposed an answer seeking algorithm for Question ∗Doctoral Dissertation, Department of Information Processing, Graduate School of Information Science, Nara Institute of Science and Technology, NAIST-IS-DD0261015, February 3, 2005.