Chinese Discourse Relation Analysis

Information such as entities and events in discourse organized according to acertain structure to express a semantic relation. So if we want to understand thesemantic of the discourse, we should get the information in the discourse andunderstand the structure of the information. And discourse relation recognition aimsto analyze the internal logic structure between sentences hoping to understand thecomposition structure of the discourse based on lexical analysis and phrase structureanalysis. Discourse relation recognition can provide strong technical support aboutunderstanding the internal discourse structure for NLP tasks such as sentimentanalysis, text coherence, etc.Discourse relation analysis contains two subtasks: identifying the elementarydiscourse unit in the sentence and recognizing the logic relation between theelementary discourse units. Now most previous work focused on recognizing thelogic semantic relation between the discourse units manually annotated b y humanbased on the English discourse Treebank. This paper explodes the ChineseDiscourse relation automatic recognition which contains developing the automaticmethod of splitting a sentence into elementary discourse units, recognizingconnective in a sentence and recognizing discourse relation between the discourseunits based on publishing the Chinese discourse Treebank.To split a sentence into elementary discourse units, we firstly analyze theadvantages and disadvantages of using comma, and then we propose a method ofsplitting sentence into elementary discourse unit based on the analysis of phrasestructure which can get a better accuracy.we use the SVM tools to train the Chinese discourse connectives automaticrecognition model based on the building of Chinese discourse connectivesdictionary, which is used to recognize the discourse connectives between thediscourse units.This paper analyze the effect of using Chinese discourse connectives dictionaryto recognize the explicit discourse relation that can be very good and then wedevelop the recognition method based on regulation and SVM classification. Wealso train a SVM model for recognizing the implicit discourse relation, and then wequantify the discourse units in the article based on the word vector and trained anew model which can archive a better effect.At last, we transform the Chinese discourse Treebank in a PDTB-style formatand make it applicable in the network to help researchers to analyze the Chinesediscourse recognition (HIT-CDTB, http://ir.hit.edu.cn/hit-cdtb), and we also develop an online demo to demonstrate the process of Chinese discourse relation analysis.