Syntactic parsing is one of the important tasks in the area of Natural Language Processing (NLP). It can provide structural information of sentences in various kinds of application systems such as Machine Translation, Question Answering and Information Extraction. In recent years, dependency syntax received more and more attention because of its simple form and efficient parsing algorithm, and has been widely applied in various NLP tasks. This thesis focuses on Chinese dependency parsing. The main achievements and innovations include: (1) We conduct a comparative study on the representive methods for Chinese de-pendency parsing through using the same data set, data division and evaluation metrics. (2) Deterministic dependency parsing method is a greedy method. This thesis pro-poses probabilistic parse action model to overcome the greediness. This model integrates the advantages of both deterministic method and dynamic programming method. (3) Chinese has some special structures that make parse action decision difficult. This thesis proposes n-phase model for these ambiguous structures in Chinese depend-ency parsing. This model has two classes: n-phase parse action model, short distance and long distance separated model. (4) To analyze and test the performance of parse action model on multiple languages, we attended multilingual dependency parsing shared task in CoNLL-2007. This shared task consists of ten kinds of languages. The evaluation results show that parse action model performs superior to traditional deterministic method on all ten languages. (5) This thesis proposes structure prediction method for chunk parsing. Experimental results show that structure prediction method improves the performance on long chunks, and chunking action model performs superior to traditional deterministic method in all chunk parsing problems.
修改评论