Integrating Graph-Based And Transition-Based Chinese Dependency Parser

Syntax parsing is one of the most important tasks of natural language processing. Dependency analysis is widely used in machine translation, information retrieval and automatic abstract. Dependency relations present relations between words, and are easy to be converted into semantic dependency. Word is the smallest element of sentence, and the dependency based on words analysis can represent deep syntax relation, This thesis researches Chinese dependency between words.At present, Nivre's algorithm has been used for English and Japanese dependency analysis, and has achieved good research results. The grammatical structure of Chinese is far different from other languages and dependency analysis is very complex. Currently the methods of Chinese dependency parsing can be divided into transition-based approach and graph-based approach, the main representative of transition-based approach is Nivre's algorithm and the main representative of transition-based approach is Maximum spanning tree (MST) algorithm.Nivre's algorithm belongs to deterministic dependency parsing and parses dependency-relation with the features around the words. It performs parsing by greedily taking the highest-scoring transition out of every parser state and the parsing history can be used in next parsing until we have derived a complete dependency graph. MST algorithm get a model for scoring possible dependency graphs for a given sentence, perform parsing by searching for the highest-scoring graph. We can't get the part parsing result until the sentence is parsed over. In this paper, we propose two kinds of integrating methods according to the complementary relationship of Nivre's algorithm and maximum spanning tree algorithm. One method is based on the impact factors of existence that Maximum spanning tree algorithm act as the basic algorithm modified by the existence of the results of Nivre algorithm .The other method is based on the impact factors of dependence degree that Maximum spanning tree algorithm act as the basic algorithm modified by the dependency degree of the results of Nivre algorithm.Experiments using Penn Chinese Treebank5.0 show that the proposed two combination methods are better than the original single algorithm and the model based on the impact factors of dependence degree reaches the highest accuracy which is 86.87%.