Question answering system returns precise and concise answers for user questions in natural language, and its core technology is answer extraction. Based on weight importance of different keywords in user’ questions and returned documents, a method for computing keyword weight is proposed. In the meantime, the weighted Latent Semantic Analysis technique is also introduced in this process. Experimental results show that the MRR of the proposed method is better than that of Vector Space Model, and gets a more satisfactory performance.