Question Paraphrase Generation for Question Answering System

Abstract

The queries to a practical Question Answering (QA) system range from keywords, phrases, badly written questions, and occasionally grammatically perfect questions. Among different kinds of question analysis approaches, the pattern matching works well in analyzing such queries. It is costly to build this pattern matching module because tremendous manual labor is needed to expand its coverage to so many variations in natural language questions. This thesis proposes that the costly manual labor should be saved by the technique of paraphrase generation which can automatically generate semantically similar paraphrases of a natural language question. Previous approaches of paraphrase generation either require large scale of corpus and the dependency parser, or only deal with the relation-entity type of simple question queries. By introducing a method of inferring transformation operations between paraphrases, and a description of sentence structure, this thesis develops a paraphrase generation method and its implementation in Chinese with very limited amount of corpus. The evaluation results of this implementation show its ability to aid humans to efficiently create a pattern matching module for QA systems as it greatly outperforms the human editors in the coverage of natural language questions, with an acceptable precision in generated paraphrases.