Research On Similar Case Matching Method For Private Lendin

Since the Supreme People’s Court issued the "Guiding Opinions on Similar Cases Retrieval",similar case retrieval has become an important part of intelligent justice construction in China.Similar case matching,as an important technology for similar case retrieval,plays an important role in improving the accuracy of similar case retrieval.This thesis takes private lending cases as the object,and it has good application and theoretical value to carry out the research on the matching method of similar cases.Private lending cases are relatively common in the judicial field,but they contain many elements and detailed information.Conventional text matching methods are difficult to distinguish the differences between the judgment documents of the two cases.This thesis uses deep learning technology to carry out research on similar case matching methods according to the characteristics of private lending cases,aiming to identify the case elements of the query cases,use the element information to find candidate cases,and obtain the most similar cases through further semantic reasoning matching.The main research contents of this thesis are as follows:(1)Case element recognition method incorporating label informationMany case elements in the judgment documents of private lending can be classified according to their attributes,and each category corresponds to a different number of element labels,but the scarcity of data and the imbalance of labels will make it difficult for the classifier to accurately and completely identify the elements.Considering that attributes and feature tags have semantic information,that can promote classifiers to recognize effectively,this thesis uses the pre-training language model BERT to build encoders,fuses tag attributes and feature tags with the coding representation of adjudicative documents through different fusion strategies,and finally identifies the elements of each type of attribute with the help of Softmax classifier.The experimental results show that the F1 on the validation set and the test set is improved by 0.036 and 0.0469,respectively,compared with the baseline model BERT.This method not only can accurately identify multiple feature tags,but also is more effective than other methods in identifying several types of features with less data.(2)Similar case matching method fused with segmentation coding and affine mechanismPrivate lending judgment documents are typical long texts.The method of coding long text as a fixed representation for similarity calculation that ignores the detailed information,and it is difficult to find the subtle differences between judgment documents that are very similar in themselves.Because reducing observation granularity can increase semantic information and increasing parameter learning in the process of judgment interaction can better explore the differences between each other,this thesis designs a coding method for coding non-governmental lending judgment documents in segments,and introduces affine transformation to build an interactive scorer when matching interaction,so that the interactive process can learn more subtle differences between texts.Experiments on CAIL2019-SCM dataset demonstrate the validity of this encoding method and affine interactive scorer in matching non-governmental lending judgment documents.The accuracy of the method designed in this thesis is 1.89% higher than that of the best model.(3)Similar case matching system for private lendingCombined with the user’s needs for matching similar cases in the private lending field,based on the elements recognized from the judgment documents and the case semantic reasoning matching model,different front-end and back-end frameworks are used to construct a similar case matching prototype system suitable for the private lending field.The system provides users with the function of automatically analyzing referee documents and providing reference for similar cases.