Research And Application On Latent Semantic Analysis In Answer System

This paper supported by Jilin Province program, introduces research and application on latent semantic analysis in answer system. Latent Semantic Analysis is a kind of analysis method based on corpus in natural language comprehension field, it thinks that some context relations exist between terms ,between terms and documents, and a semantic structure can be consist of respective relation between many documents and terms. We used the method of matrix decomposition in math to optimize this semantic structure: it computes and deals with the structure to keep the most main relation between documents and terms and eliminates else huge, redundant, minor factor. The structure optimized is not only smarter than the original structure, but also keeps the most main relation, so we can mine the latent semantic relation. In sequent search, we can compute the latent similarity between documents and the distributing status of the key terms in the latent semantic structure. This paper used singular value decomposition and factor decomposition method. Singular value decomposition is one of the most used method in symbolic statistics, frequently used in the solution of the problem of no-limit least cube, computing matrix rank, criterion relation analysis, we used this method to deal with high tank sparseness matrix which expresses the original semantic structure; factor decomposition looks the original semantic structure as consisting of factors, we search a new weighted main factor combination through factor decomposition, which is a predigestion of the original semantic structure and can mostly expresses the inner relation of the original semantic structure and explains the all original semantic structure. Singular value decomposition is used to realize the answer function of answer system: the original semantic structure dealt with singular value decomposition is a lot of knowledge stored in the system knowledge-base. In this paper, it used singular value decomposition to present the latent context relation between terms and documents by the form of weighted vectors. On the basis of simplified semantic structure, we compare the similarity between the question and every knowledge document. By comparison, we gain the answer to the question. k value is a very important parameter in singular value decomposition, which decides the dimension of the optimized semantic structure. In the former research, the range of k value is mainly chosen by the experience gained from a number of experiments, so it is easy to produce windage. If we don't have enough experience, we even choose the wrong range of the k value. In this paper, on the basis of the system requisition we improve on the manual choose to the auto choose executed by the system. The system administrator must not be required a number of the experiment experience, he only executes the auto choose after altering knowledge base to gain the most fit k value. It doesn't require finding the file of k value in a lot of background files, and doesn't require repeat altering background files, so it enhances the maneuverability and safety. Factor decomposition is used to realize the knowledge mining function: the original semantic dealt by factor decomposition is composed of the question document which come from user and the answer documents answered by the answer function. We search new weighted factors by factor decomposition, and the big weights ones are main factors. The combination of main factors could mostly express the inner relation and main semantic structure that is contained in the question text and answer text. These main factors are just the knowledge pointes which system wants to mine. In this paper, we realized latent semantic analysis in the answer system. In the design step, we used UML and Rational Rose; in the realization step, we used popular J2EE technology. This answer system is a part of the province level program Network Educational Resource Management System that is build item by Jilin Province Science Committee and developed by the...