Abstract: Answer extraction is the key of the Chinese questioning system. Normally answer extraction mainly depends on the pattern of the answer sentence. Since the pattern of the answer sentence is obtained by experts based on the language rules, so it strongly relies on the experts' language knowledge. To overcome this limit, a scheme is proposed to gain the pattern of the answer sentence by pattern learning. The scheme takes the advantage of searching engine to retrieve related documents of question. From these documents the sentences that include the answers are extracted. Then the types of the questions and answers are marked to form questionanswer training corpus to the questions of different types. Then by statistic learning method, the candidates of sentence patterns are abstracted and the weights of the patterns are calculated. Thus, based on the weights the patterns of the answer sentences to the questions of different types are obtained. Answer extraction result of the factoid question shows that the experimental MRAR is up to 0.28, which indicates the effectiveness of the proposed pattern learning scheme. The patterns gained by pattern learning cover the normal answer sentences.

Key words: computer software, question answering system, answer extracting, pattern learning, pattern matching