The Research On Statute Citation Reasoning And Analysis Of Criminal Judgment Documents Based On Factor Information Of Statutes

In the broad framework of judicial transparency,judgment documents,as the carrier of the Chinese People’s Court to exercise judicial power and apply laws to resolve disputes on behalf of the state,whose qualities have a direct influence on the authority and credibility of judicial organs.The traditional quality inspection of judgment documents entirely relies on manual review and proofreading.However,this way greatly increases the workload of judges under the situation of too many cases and too few judges.With the continuous development of informatization of the Chinese People’s Court and the disclosure of judgment documents,the text mining about judgment documents has become a hot research direction.Nevertheless,at present,the work related to quality inspection of judgment documents usually focuses on text correction,like word proofreading,and there is less work on the quality of judgment document reasoning.Besides,lacking of judgment document reasoning is a highlighting problem in the field of Chinese criminal judgment.Based on the current research status,this thesis,taking criminal judgment documents as the data source,utilizing the deep learning model as the basis of algorithms and factor information of statutes as the external knowledge,studies on statute citation reasoning and analysis used in quality inspection of judgment document reasoning.In this thesis,we define statute citation reasoning and analysis at first,and then transform it into the text matching problem between facts and contents of statutes.For the problem,we firstly put forward the concept of two-factor in the content of a statute according to the expression pattern of contents of statutes,and then propose a multi-level deep text matching model based on the factor information of statutes.The key idea of this model is to extract text features on different abstraction levels of texts from an inputted text by convolutional neural networks and compute matching patterns between the inputted text pair at each abstraction level,and utilize the factor information of the statute to assist matching patterns computing,and consider all matching patterns comprehensively at last.For the model,we create the related dataset based on judgment documents of two causes of action by manual annotation,and verify the great performance of the model by a series of comparison experiments.For the problem of extracting factor information of statutes,we firstly divide a sentence of the content of a statute into two categories according to categories of factors the sentence belongs to,and then transform the problem of extracting factor information of statutes into a problem of classifying each sentence of the content of a statute.We build the related dataset by manual annotation,and establish a feature set based on common vocabularies of contents of statutes and expression rules of two types of sentences in order to vector inputted sentences,and then choose the random forest algorithm as the classification algorithm.Comparison experiments show that the random forest algorithm has strong performance in this problem.At last,we put forward some improving methods for shortcomings of the proposed text matching model and the approach of extracting factor information of statutes in this thesis.