Research On The Method Of Chinese Micro-Discourse Analysis

In recent years,the focus of research has shifted from the word and sentence to the paragraph and discourse in the field of Natural Language Processing(NLP).Discourse anal-ysis is one of the most important foundations of NLP,which attracts more and more research-ers' attention.There are three subtasks in discourse analysis:1)Tree construction,aims to build the discourse structure framework;2)Nuclearity recognition,aims to identify the nu-clearity between arguments;3)Relation recognition,aims to identify the semantic relations between arguments.Through discourse analysis,a complete discourse tree can be con-structed to help understand the overall semantic of discourse.At present,Chinese discourse analysis mainly focuses on the micro-level research(that is,Chinese micro discourse analysis),which only refers to the identification of structural and semantic relations within or between sentences in a paragraph.Compared with the success of English discourse analysis,there are still in the infancy stage in Chinese micro discourse analysis.Therefore,this dissertation focuses on the Chinese micro discourse analysis,which includes the following three aspects:(1)Research on the method of micro discourse tree constructionThere are two issues in previous work on discourse tree construction,i.e.,error accu-mulation and connectives influence in Shift-Reduce algorithms.To address the above issues,this dissertation proposes a tensor-based neural network with the multi-stage strategy and connective deletion mechanism to construct a discourse tree.Firstly,encode arguments by Bi-LSTM and Attention,and use tensor-based transformation to capture the semantic rela-tionship between arguments.Then,aiming at the error accumulation in the Shift-Reduce algorithm proposes a multi-stage strategy.Finally,aiming at the connectives has a great im-pact on Shift-Reduce action prediction,proposes an intra sentence connective deletion mech-anism.Experimental results show that the proposed model outperforms various baselines.(2)Research on the method of micro discourse nuclearity recognitionThere are two issues in the existing popular models,i.e.,it is easy to identify multinu-clear with high semantic similarity as mononuclear and tend to identify longer discourse units as Nucleus.This dissertation proposes a method based on Gated Memory Network(GMN)to recognize nuclearity in Chinese discourses.Firstly,used Bi-LSTM and CNN to capture the remote and local information of each argument.Then,merged two arguments information as a gated unit.Finally,used the gated unit to capture relatively important feature representation from the basic discourse unit to identify the Nucleus.Experimental results show that the proposed model outperforms various baselines.(3)Research on the method of micro implicit discourse relation recognitionChinese micro discourse relation recognition is to identify the semantic relation be-tween arguments.The explicit micro discourse relation recognition has achieved high accu-racy.However,the implicit relations recognition is always a difficult task.Data sparsity is the main problem restricting the development of Chinese discourse relation recognition.This dissertation proposes a method for Chinese implicit discourse relation recognition,which expands the training data by combining active learning and multi-task learning method.This method aims to reduce the noise as much as possible when expands the training data set.Firstly,used active learning to select some explicit data through the classification uncertainty based on BERT,and then the connectives in the explicit data are removed and regarded as pseudo-implicit training data.Finally,used multi-task learning to boost implicit discourse relation recognition by using the pseudo-implicit training data.Experimental results show that the proposed model outperforms various baselines.This dissertation proposes effective solutions on the three subtasks of discourse analysis,i.e.,tree construction,nuclearity recognition and implicit relation recognition,which will provide a reference for further research of Chinese micro discourse analysis.