Research On The Method Of Law Prediction Based On Feature Fusion Of Global Features And Local Features

In recent years,the interdisciplinary combination of artificial intelligence and the judicial field has gradually become an emerging research hotspot,and the research on the scheme for judicial decision-making assisted by artificial intelligence technology has attracted widespread attention in both academia and industry.Prediction of laws and regulations is one of the important tasks in smart justice.It aims to predict applicable laws and regulations through case description texts,so as to assist decision-making and improve judicial efficiency.Current research mainly focuses on predicting the corresponding legal label by inputting fact description text through machine learning methods or neural network models.The inadequacies of the current research are as follows: 1)The prediction of legal articles focuses on the characteristics of the factual elements of the case,and these elements have commonalities.However,most of the current research work only extracts information on individual cases,and lacks the mining of element information between different data;2)The prediction of legal articles is a problem of the judicial field.The structure of judicial texts is relatively fixed and the length of the text is relatively long.And,different parts of the text have different importance for the prediction of legal articles.The current work lacks the treatment of this problem;3)Article labels There are similarities,and the frequency of label appearance is quite different.The current work lacks consideration of label distribution.Aiming at the above problems,this paper proposes a feature fusion-based law prediction model.The main work are as follows: 1)Aiming at the problem of only extracting information from isolated cases in the task of law prediction,we studied a global information extraction method based on graph structure;through the information sharing of the global graph structure,the feature information of elements can be spread across cases;2)Aiming at the structure and the characteristics of judicial text,we studied the multi-level feature extraction method;by constructing a feature extraction network at the document level and sentence level,the global features and local features of the case description text are integrated;3)Aiming at the distribution of law labels,we studied a multi-task auxiliary training and loss function optimization method based on judicial logic;by constructing auxiliary tasks to promote law prediction,and for the distribution of law labels,we employ a loss function that measures the distribution of labels.Finally,the effectiveness of the method in this paper is proved through the experimental comparison on the real data set,and we explain the effectiveness of the method by conducting ablation experiments and sensitive parameter experiments.