Law Knowledge Graph Construction Based On BERT Models And Its Applications In Case Retrieval

With the comprehensive advancement of China’s judicial process,the Supreme People’s Court has issued relevant regulations to formally institutionalize the retrieval of similar cases,making it one of the necessary links in the trial process.At present,the application of case-like retrieval is in its infancy.The Supreme Court and the local higher people’s courts have not yet formed a unified standard for the discussion of relevant systems and the construction of the platform.There are still many problems in its practice,such as the fact that the relevant cases obtained by the retrieval can not completely cover the keywords,the matching rate of the special plot fields related to the case is low,and the similarity degree has no specific quantitative indicators,resulting in a low overall retrieval quality.Therefore,this paper applies knowledge graph to the legal field,and designs a scheme based on similarity index to improve the accuracy of case retrieval.Specifically,the main work of this paper includes the following three aspects:(1)Definition of legal knowledge graph and data preprocessing.Based on the relevant characteristics of theft cases,this paper analyzes the key points,selects suspect,victims,case details,etc.as nine types of entities,and selects collusion,profit,understanding,etc.as seven types of relationships,thus defining the basic content covered by the knowledge graph.Complete the manual annotation of the data set with the theft legal judgment document as the basic data set.In particular,for the named entity recognition task,first use the LTP of Harbin Institute of Technology to pre-label the location,time and person type entities,and then further expand the entity type to complete the overall labeling work.(2)The construction of legal knowledge graph.In this paper,the named entity recognition task is completed based on Bidirectional Encoder Representations from Transformer(BERT),and the relationship between labels is constrained by conditional random field(CRF),which further improves the model effect.At the same time,the relationship extraction task is completed according to the predefined relationship type.Finally,the result triplet is visualized.(3)A kind of case retrieval scheme based on knowledge graph.This paper puts forward corresponding solutions to the problems existing in the current keyword-based query method of case-like retrieval,designs a case-like retrieval scheme based on similarity calculation,focuses on the keywords that have a direct impact on the sentencing results in the judgment documents,and takes the factors such as the case method,stolen goods,and the total amount involved into the consideration of case-like cases,and creates a case-like comparison index-similarity index from multiple aspects,It also combs the overall retrieval process.Finally,this paper applies this scheme to the actual retrieval scenario,and the experimental results show that this scheme can improve the accuracy of case retrieval and provide judicial practitioners with highly relevant case reference.