Researches On Learning Approach And Application Of Structured Support Vector Machine

Support vector machine (SVM) has solid theoretical basis of Statistical learning theory (SLT), perfect form of mathematics, intuitive geometric explanation and good generalization ability. SVM has been a powerful tool for solving many problems in the domain of data mining. However, for most of practical applications, there are many kinds of complex structural data, such as tree structure, mesh structure and queue structure, etc. It is very difficult for the traditional support vector machine to solve the problem including this kind of data. Structured support vector machine (SVM-Struct) may be a suitable approach for solving these complex, mutual interdependence and structural data processing problems. Therefore, researching learning approach and applications of structured support vector machine will possess important theoretic significance and practical application value.To research the principle of SVM-Struct, this thesis takes Chinese parsing as an example through constructing structural feature function (?)(x, y), and then Chinese parsing model is built. Because complex and structural data embed in Chinese parsing, the proposed approach can be evaluated effectively. The main research works are concluded in the following:(1) Analyzing the basis principle of SVM-Struct learning systematically, and discussing the construction of approaches some structural feature function (?)(x, y).(2) Proposing a structural support vector machine (SVM-Struct) approach for Chinese parsing through constructing structural feature function (?)(x,y), and weight context-free grammar Chinese parsing model is built. Combining with CYK (Cocke, Kasami, Younger)algorithm, Chinese parsing is then accomplished.(3) Testifying the proposed approach on the public mini-corpus samples from the institute of computational linguistics, Peking University. Comparing with the traditional probabilistic context free grammar (PCFG), the experiment results demonstrate that the proposed SVM-Struct approach is feasible and effective for Chinese parsing.By deep research on SVM-Struct, structured support vector machine approach for Chinese parsing is presented, which can extend the application area of SVM-Struct. The research results will not only richen the SVM learning theory, but provide a novel research approaches in Chinese parsing as well.