Number Type Identification of Chinese Personal Noun Phrases via Piece-wise Linear Support Vector Machine

Number information plays an important role in many Chinese natural language processing tasks. Different from English, Chinese does not explicitly show grammatical number information by inflection. Mostly existing number type identification methods for Chinese personal noun phrases are based on rules, some machine learning-based methods explore diverse features, including both flat and structured to train the classification models. However, the unbalanced data between different classes in the problem of personal noun phrases number type identification leads to the low classification accuracy in the small sample class. In this paper, we propose a Piece-wise Linear Support Vector Machine(PL-SVM) method to divide the large number class into N parts, and train the local linear SVM to construct classification boundary and fit the train data, to improve the classification accuracy in small sample class. Experiments on ACE2005 show that our method can not only achieves better performance on large number class, but also improves the classification accuracy on small number class, which makes up the flaw of traditional methods in some extent. Furthermore, the experiment results show that our method can improve the resolution accuracy when it is implemented in Chinese coreference resolution.