Japanese Dependency Analysis Based On Improved Svm And Knn

This paper presents a method of Japanese dependency structure analysis based on improved Support Vector Machine (SVM). Japanese dependency analyzer based on SVM has been proposed and has achieved high accuracy. The efficient way to improve dependency accuracy farther is to increase the training data. However, the increase of training data will bring a great amount of training cost and decrease the parsing efficiency. We delete those samples that are unused or not good to improve the classifier's performance, and then train the reduced training set with SVM to obtain the final classifier. Furthermore, we combine improved SVM with K nearest neighbors(KNN) to improve the performance of dependency analyzer. Experiments using the Kyoto University Corpus show that the method outperforms previous systems as well as the dependency accuracy and the parsing efficiency.