Text Similarity Calculation In The Judicial Field Based On Text Representation Learning

The goal of text similarity measurement is to automatically determine a score to indicate the semantic similarity of text.In the judicial filed,the advancement of judicial openness has given the people an opportunity to participate in the judicial.But the problem of "different judgments in similar cases" has subsequently been exposed,which has become one of the main negative factors affecting judicial credibility.Similar cases matching has become one of the important ways to solve this problem.It is of great significance to provide judges with similar cases when they making judgments in order to reduce the deviation of the trial.As a new form of legal service,it satisfies the public’s legal consultation demands and reduces unreasonable litigation expectations.How to measure text similarity in the judicial filed is the core issue of similar cases matching.The texts in the judicial filed have strong characteristics such as particularity,professionalism and rigor.The text is highly stylized,the text is long,the sentence structure is complex and rigorous,and the text contains the legal relationship and facts between the subject and the object in complex scenarios.But the information is scattered in the legal text and redundant.Existing methods are difficult to capture complex legal relationships and essential facts,lack the guidance of domain knowledge and conceptual level knowledge.It’s difficult to represent texts in the judicial field,and there are problems with long text representations.To solve the above problems,this paper proposes text similarity calculation in the judicial field based on text representation learning.Considering that the knowledge graph has a strong ability to express heterogeneous information and flexible modeling capabilities,this thesis constructs the case knowledge graph to represent factual information and legal relations.It’s not only avoids the redundancy of essential facts in the legal field text,but also introduces conceptual level knowledge and domain knowledge,which enriches the presentation of case information.In terms of knowledge representation,the unlabeled and generalization operation of the case knowledge graph to enhance the computing power and feature expression ability of the knowledge graph.In terms of text similarity calculation,this paper constructs a graph convolutional neural network and a siamese convolutional neural network to extract features of the case knowledge graph and uses the feature representation of the case knowledge graph as the text feature vector to calculate the similarity between texts in the judicial filed.And proposed the drop＿node suppression graph convolutional neural network overfitting.This thesis uses the data set of the CAIL.The experimental results on this data set show that the method proposed in this thesis has achieved good performance in judicial filed.