Chinese Macro Discourse Structure Parsing

Macro discourse structure parsing aims to lay the foundation for the semantic understanding of discourse by analyzing the structure of discourse of text and providing strong structured information support for downstream tasks such as automatic summary,machine translation generation and evaluation,question and answer systems,and information extraction,Discourse analysis is mainly divided into micro and macro levels.The micro-level discourse analysis focuses on the organization and semantic relationship between two sentences,while the macro-level discourse analysis focuses on the organization and semantic relationship among two paragraphs.Compared with micro-level discourse analysis,macrolevel discourse analysis is more effective in a document.It reveals the main idea and content of the document at a higher level,which is beneficial to a deeper understanding of the text.In this dissertation,we conduct an in-depth study on Chinese macro discourse structure parsing,and the main research contents include the following three parts.(1)To address the problem that the traditional unidirectional parsing model cannot switch the appropriate parsing action according to the current context and thus is prone to error propagation,this dissertation proposes a macro discourse structure parsing method based on the bidirectional fusion.Firstly,this dissertation designs a set of bidirectional parsing algorithms,which enables a discourse structure tree to be constructed from two directions simultaneously,breaking the limitation of the unidirectional parsing model and making it possible to combine and complement the two unidirectional parsing models.Secondly,the base parser and the decision maker are designed in this dissertation.The base parser contains two sub-parsers,a bottom-up parser,and a top-down parser,while the decision maker selects the prediction strategy by considering the state of each sub-parser.Finally,a new error analysis method is proposed to quantify the impact of error propagation.Experimental results on MCDTB 2.0 demonstrate that the method can significantly mitigate the error propagation problem.(2)To address the common problems of flattening encoding in traditional methods,as well as the problems of missing behavior modeling and missing golden bidirectional parsing labels in bidirectional parsing,this dissertation proposes a macro discourse structure parsing method based on oracle selection.First,to address the problem of missing data for bidirectional selection,this dissertation designs a set of training strategies based on oracle selection for obtaining the training data required for bidirectional parsing,which enhances the robustness of the model.Secondly,this dissertation improves the decision-maker by introducing two types of behavioral labels and four types of structured labels,which solve the flat coding and behavioral modeling problems simultaneously and free the decision-maker from the complete dependence on the confidence of the basic parser.Finally,experimental results on MCDTB 2.0 show that the method achieves significant improvement over the baseline system.(3)To address the problems of difficulty in acquiring discourse structure corpus and the lack of a deeper understanding of relational semantics in process-oriented discriminative models,this dissertation proposes a macro discourse structure parsing approach based on distant supervision and generative fusion.First,we propose a discourse structure parser based on generative fusion and design a set of structured target sentence generation templates,using the encoding layer of the generative model T5 as the upstream part,while the downstream part includes two parts,a T5-based decoder,and a pointer network decoder.This joint model is able to deepen the understanding of discourse semantics from both generation and discriminative perspectives.Secondly,this dissertation proposes an oracle annotation method for discourse structure parsing of a large-scale topic structure corpus to obtain a silver-standard macro discourse structure corpus,which is used to solve the small sample problem caused by the difficulty of acquiring a discourse structure corpus.Finally,experimental results on MCDTB 2.0 demonstrate the effectiveness of the method.