| Question Answering system, as a kind of novel and intelligent information retrieval technique, and as a kind of more efficient and more humanized search engine, has already become one of the important research directions in intelligent information and natural language processing field now. Information retrieval is the important link which connects the question analysis and answer extraction modules in the whole Question Answering system, and its performance quality will directly determine whether the answer extraction can work effectively. Most of the current passage retrieval approaches are based on lexical matching. However, these approaches usually result in the loss of some crucial semantic relations between words, and that is a major reason resulting in false positives problem. Here false positive means that many irrelevant passages share the same question terms with correct ones, but the relations between these terms are different from those in the question.In this thesis, related techniques of shallow semantic parsing in natural language processing are introduced into passage retrieval, and a passage retrieval method based on syntactic dependency relations is proposed, which can better reflect the semantic similarity between question and its candidate passages. For the shorter questions, synonyms and abbreviation expansion and dependency relations based query expansion are added to improve the retrieval recall. Meanwhile, the dependency relations based passage retrieval method also can be combined with synonyms and abbreviation expansion, which makes the method implement fuzzy matching of keywords to some extent.Two parts of experiments about different passage retrieval approaches and their combination of query expansion are carried out to evaluate their performances objectively. The experimental results indicate that dependency relations based passage retrieval method has a great enhancement on both precision and ranking of results compared with others. |
PDF Full Text Request