Law Recommendation Algorithm Based On Hierarchical Structure

The Chinese social economy’s swift expansion and the successive promotion of the construction of democracy and the rule of law have led to a gradual rise in the public’s legal awareness.Therefore,seeking for judicial assistance has become their first choice in dealing with disputes.At the same time,the types and aggregate of criminal cases accepted by people’s courts at all levels are also increasing year by year.To meet the increasing demand of the people in judicial efficiency,using AI technology to assist the judiciary has become an unavoidable trend in the development of intelligent justice in the future.Although there are many existing studies on the prediction of legal judgments,which also verify the effectiveness of external knowledge of legal provisions for judgment prediction,the use of legal provisions is insufficient,and the relationship between legal provisions and the facts of the case and the hierarchical structure of legal provisions are ignored.At this time,there is a “category imbalance” in the existing judgment documents,which is easy to cause the problem of “imbalance” in model training,which in turn leads to the fact that the laws that appear less often are not easy to predict accurately.Therefore,this paper will use the natural hierarchical structure of laws to learn the semantic expression of large categories of crimes and legal articles,and use the attention mechanism to change the classification task into a matching task to increase the interpretability of the model.At the same time,the text data enhancement technology and the pre-trained language model introduce more prior information to relief the problem of model training " out-of-balance" caused by the "category imbalance" of the law.Specifically,the work of this article is as follows:(1)Based on the natural hierarchy of legal articles,design a decomposition strategy,that is,using the hierarchy of legal articles,extract the joint part and the remaining part from the initial semantics of the law as the semantic expression of the general crime and the legal article,respectively,and then predict the parent label and sub-label based on these two parts.(2)The collaborative attention mechanism is used to transform the legal recommendation task from the traditional classification problem to the matching problem,and the semantic matching between candidate labels(major types of crimes and legal articles)and the facts of the case is analyzed,so as to increase the interpretability of the model.(3)To balance the corpus with unbalanced categories,text enhancement strategy EDA is employed to enhance some of the data,and Lawformer,a pre-trained language model Lawformer,is employed to iteratively train the augmented corpus to introduce more prior information.In an effort to test the availability of the model mentioned in this article,experiments were carried out on CAIL2018 and compared with six benchmark models and two ablation models.Experimental results show that the hierarchical dependency of legal articles and the introduction of prior information by text enhancement technology can improve the performance of legal recommendations as a whole.At the same time,while mixing EDA and Lawformer performs slightly better than using EDA alone,models using Lawformer require more computing resources and time.