Research Of Chinese Discourse Structure Parsing Based On CDTB Corpus

Discourse analysis aims to study basic characteristics of discourses and inner seman-tic relations of discourse units,which then leads to understanding of discourses as whole.Discourse parsing is the key sub-task of discourse analysis.Its purpose is to process co-herent discourse text into proper structured data,which will be beneficial to performances of downstream discourse applications.Compared with English,the development of Chi-nese discourse analysis technologies drop behind due to the lag of computational Chinese discourse theories and Chinese discourse language resources construction.In addition,the highly sparseness and ambiguity of discourses semantics make Chinese discourse parsing a challenge.In this thesis,we focus on technologies about constructing Chinese discourse parser based on Chinese Connective-driven Dependency Treebank(CDTB)Corpus.Main contents are as follows:(1)We propose a discourse elementary unit segmentation method based on Long Short Term Memory Networks and Graph Convolutional Networks as a substitute of classic one which relies on heavy hand-crafted features.(2)The building procedures of a transition-based Chinese discourse parser is finely pro-posed in this thesis.The experiment results we gave show the effectiveness of different neural network approaches in encoding sub-structure of discourse during parsing in a bottom-up manner.(3)To alleviate the impact of ambiguity when parsing from bottom-up,we propose a top-down fashion Chinese discourse parser.It build discourse hierarchical structure by rank-ing all split points among elementary discourse units and binary splitting them into different levels in discourse tree recursively.Experiment results show notable improve-ments when using this approach.(4)In the end,we proposed an end-to-end Chinese discourse parser based on work of this thesis and we gave our conclusion on how error propagate in parsing pipeline.