Automatic Semantic Role Labeling Of Chinese FrameNet Based On Support Vector Machine Model

In recent years, natural language processing is focusing on the semantic role labeling. Based on the Chinese FrameNet and combining with the characteristics of Chinese language itself, the problem of Chinese semantic roles auto-labeling is researched in this paper, which will play an important role in building the large scale semantic corpus and developing the natural language processing technology, such as Chinese information retrieval, Chinese question answering, and Chinese information extraction.As the Chinese parser has not been yet quite perfect, in this paper the Chinese semantic role labeling can be seemed as the sequence labeling problem of using words as the marked units. Here, we use support vector machine as classifiers to train the models which combine the word and base-chunk level features respectively. In this thesis, the task of Chinese FrameNet semantic role labeling is described as:Given a target word and its frame, then identify the boundaries of frame elements and label the correct frame element names for them in a Chinese sentence.25 Frames are selected from Chinese FrameNet Corpus as our training and testing set, then 3×2 fold cross validation is carried out, of which the mean can be used as the evaluate index. In the experiments, we provide sevral window sizes for every kind of candidate features, and then choose the features with their window size by orthogonal array.In addition, two types of semantic role labeling model are constructed here:one is based on word-level features, we achieve 59.65% F-score with the identification and classification separately in the semantic role labeling, and 58.72% F-score can be obtained with combination of the identification and classification; the other one is based on base-chuck features, where we get 59.67% and 58.92% F-score respectively in the above two strategies.Experimental results show:(1) the two steps strategy is better than only one step in semantic role labeling; (2) after putting the base-chuck information in semantic role labeling, the performances are enhanced a little; (3) the phenomenon of higher percise and lower recall are obviously reflected in the two models.