Research On Some Key Techniques Of Judicial Litigation Case Text Mining

Case text,as the main carrier of case description,has become an important reference resource in the judicial field and provides a data basis for legal artificial intelligence.It is of great significance for the construction of judicial case analysis database to use effective text mining method to quickly and accurately extract useful knowledge automatically from case text and form structured data.Named entity identification,element extraction and judgment prediction of judicial cases,as an effective means to establish case analysis database,can provide reference basis for judicial practice and case prediction,achieve the same judgment of the same case,promote judicial fairness,and have a greater role in promoting the intelligent processing of judicial cases.In recent years,the development of deep learning technology has pushed the research of data-driven natural language processing to a new climax and shown its potential application value.For this reason,this paper mainly adopts the deep learning-based case text mining technology in the judicial field to carry out research on three tasks of judicial text named entity recognition,element extraction and case decision prediction.The main research contents are as follows:(1)Research on the method of ambiguous named entity recognition based on transfer learning:by studying the characteristics of named entity recognition task data in the judicial field,the basic types of entities are summarized and the domain corpus is expanded;According to different entity types,a named entity recognition method based on pre-training is studied.Aiming at the difference problem of entity understanding,a named entity recognition method based on transfer learning is proposed to eliminate entity ambiguity.According to the general knowledge learned from the pre-training transfer learning model,the downstream neural network is fitted,and the character level and word vector level features of the text are represented.The context-dependent information of the text is used to extract the context-bidirectional features and effectively solve the problem of task boundary division of named entities.The effectiveness of the proposed method is verified by experiments.Based on the basic entity,three levels of legal ontology NERP,NERCGP and NERFPP are developed,which expand the corpus of named entity identification in the judicial field,and provide a label standard for the extraction of elements of judicial litigation cases.According to different levels of entities,the result of entity identification is given,which provides a new idea for the judicial field of named entity identification.(2)Research on deep neural network element extraction method based on domain knowledge fusion:based on text classification and according to the relationship between fact elements and feature description,the feature extraction method based on multivariate dichotomy and multi-label classification are constructed respectively.Based on the ontology of named entity identification law,the element label system of element extraction is established,which provides technical support for the supervised element extraction.In view of the characteristics of complex elements of judicial litigation cases and lack of domain knowledge,a deep neural network element extraction method based on domain knowledge fusion is proposed.First,according to the knowledge of unsupervised data,the weight parameters of the data fitting in the judicial field are given.Then the "sentence pair"supervision type data is constructed and the part of speech information is used as the network input to enhance the understanding and generalization of models.Last,the construction of deep circular convolutional neural network provides an effective means to capture the description information of elements scattered in different sentences.Experimental results show that deep cyclic convolutional neural network can effectively mine the deep features of factual elements,improve the accuracy of element extraction,and then improve the network performance.(3)Research on data-oriented method of case decision prediction:based on the in-depth study of the text data of judicial cases,a new method of judicial decision prediction based on joint neural network for unbalanced data is proposed.The method constructs the feature vector of key words,establishes the weight scheme of instance and data balance strategy,which alleviates the problem of data imbalance.On this basis,the joint neural network is initialized according to the co-occurrence patterns among multiple tags in the task,the frequency of co-occurrence patterns is counted,and the dimension reduction is carried out in combination with principal component analysis to realize the judgment prediction of the case,and the cognition ability of the model to the crime and the law is enhanced.The experimental results show that the proposed method provides an effective means for case decision prediction with unbalanced data,which can effectively improve the result of case decision prediction.