| One of the core tasks of smart justice is to realize credible automatic judicial judgment,the key point of which is to improve automatic judicial judgment.The interpretability of legal judg-ments is allowing machines to understand the facts of the case and give convincing judgments from the judge’s judgment perspective result.The judgment document condenses the judge’s understanding of the case,which is embodied in the sentence description of the case in the judg-ment document.If the causal relationships between the sentences describing the case could be extracted,the judge’s opinion on understanding of the facts of the case is got,which could en-hance the interpretability of automatic judicial decisions.Discourse relation recognition aims to analyze the relationship between sentences.The recognition of implicit relationship between sentences is still a difficult problem,and it is hardly utilized in the judicial field.From the per-spective of judges judging cases,this thesis uses discourse analysis technology combined with knowledge in the legal field to analyze the judgment texts.The causal relations between the case description sentences is identified,and then the causality of the case is deduced,which improves the model’s understanding of the case.The main contents and innovations of this thesis are as follows:(1)An implicit discourse causal relation recognition method PTDRR based on unsupervised constituency parse trees is proposed.For avoid the drawbacks of expensive and low-general syntactic structures,a method using unsupervised constituency parse trees is constructed,which models syntactic differences with vectors to enhance argument interactions.It also employs ON-LSTM to enhance the semantic encoding of arguments.The proposed method has better the generalization and semantic understanding than the origin one.The F1 value of this method on the English dataset PDTB is 52.04%,which improves the absolute F1 value by 6.90%.The F1 value on the Chinese dataset HITCDTB is 73.48%,an absolute F1 value increase of 4.54%,both significantly higher in the baseline model.(2)The causal relation system between the case description sentences in the judgment docu-ments is constructed and the data set CCRTB is annotated.The former general domain causality system did not focus on the core case elements in judicial decisions,thus did not have domain applicability.Based on the case element system from the perspective of criminal investigation,the causal relationship that the judge pays attention to in the judgment of the case is sum-marized.Including the cause of crime,the result of the crime and the causal relationship of the compensatory behavior,and subdivided into the motive of the crime causation,criminal action causation,disability causation,economic compensation causation,and behavioral com-pensation causation,which provide suggestions for the reconstruction of the logic of the case.Further,use the case description of the documents related to the crime of intentional wounding in the public judgment documents of the China Judgements Online Website.As the corpus,the text causality recognition data set CCRTB is manually annotated to provide data support for the research.(3)A case causal identification method EIDRR is proposed,which integrates explicit and im-plicit relations.Multiple judgemental documents may use similar connective pairs,in order to make full use of the information of the connective pairs to improve the causal identification performance of the case,based on the PTDRR method,the explicit connectives extracted by regular matching are used to model the information of connective pairs and causality corre-lation,and improve the performance of relation classification.Experiments on CCRTB show that compared with the PTDRR method,the proposed method improves the absolute F1 value by 4.92%,reaching 77.05%,and it has a better performance on criminal motive causation,disability causation economic compensation causation and behavior compensation causation. |
PDF Full Text Request