| In the context of the "smart court" strategy,our country has put forward new requirements for text understanding in the legal field.How to make machines automatically understand natural language texts in the legal field and deal with them accordingly has become an urgent problem that needs to be solved.This paper researches the extraction of case elements and the correlation analysis between news and cases.The purpose is to extract case elements from legal documents and news texts,and analyze the correlation between news and cases on this basis,so as to provide a deeper understanding of texts in the legal field.The key problem is how to extract case elements from legal documents and news texts according to the characteristics of the case elements,and how to use the case elements to improve the accuracy of the correlation predicting between news and cases.In view of the fact that the existing methods do not extract the case elements based on the characteristics of the case elements and the legal field text,and do not analyze the correlation between the differences between the news and the case description,this paper expands the research of the case element extraction and the correlation analysis between the news and the case.The research has mainly completed the following points:(1)A method of corpus construction is proposed,and the corpus required for the experiment is constructed.Xpath-based web crawler technology is used to collect news from the internet,rules are used to generate a knowledge base of case elements from legal documents,and tagging corpora are obtained through tagging technology and distance supervising methods to provide data preparation for subsequent experiments.Through the analysis and cleaning of 17191 documents in the referee document network by rules,4311 groups of case elements were obtained and constructed as a knowledge base of case elements.Through distance supervising,3449 news documents were crawled,and data sets were constructed from them by sentence.By analyzing the hot news in recent years,13 special cases were selected,and 4513 articles related to the case were crawled.By establishing the correlation between news and cases,4607 pairs of news-case correspondence data were obtained.(2)The case element is a brief description of the case-related events.Extracting the case elements in the news text has great significance for downstream case field natural language processing tasks.In view of the case field relevance and intrinsic relevance of the case elements,this paper proposes a joint case element extraction method based on case domain correlation and graph convolutional network:modeling sentence contextual information by bi-directional long short-term memory networks,then using it to predict the case field correlation for guarantying the elements' relevance of cases by joint learning,and modeling the dependency relationship of candidate elements by graph convolutional network to capture its intrinsic relevance.The experiments show that the method proposed in this paper improves accuracy rate by 6.6% in extracting case elements.(3)The correlation analysis of news and cases is a key point on news comments analysis,which is to predict whether the news and the case are correlative,and is similar to the text similarity calculation task.Siamese network is one of the most effective method to text similarity calculation task,and it can model two texts with similar structures and semantics.But siamese network may loss accuracy due to the unbalance of texts and the redundancy of news text.To solve these problems,we proposed unbalanced siamese network.Because the news headline contains main information,we compressed the news text with title removing redundant information.Because the case element contains the main semantic information of the case,we encoded the new text and the case by unbalanced siamese network using case element as supervisory information.Finally we predict the correlation between news and case texts.Experiment results show that the proposed model improved accuracy by 2.5%compared to baseline,and the method we proposed solved the problems what traditional siamese network faced.(4)In order to meet the country's strategic needs for a "smart court public opinion monitoring system",this article designed and built a prototype system through software engineering.The system collects news data from the internet,analysis the correlation between the case and the news through the element extraction model and the news case correlation analysis model,and displays it to users. |
PDF Full Text Request