Research On Legal Judgement Prediction Based On Multi-Task Joint Modeling

In recent years,methods of automatic legal judgement prediction which can effectively help judges improve work efficiency and simplify judicial process have become an important research direction of judicial artificial intelligence.The purpose of automatic legal judgment prediction is to apply natural language processing technology to judicial data,automatically predict the defendant’s accusation,sentence,relevant articles of law and similar cases according to the description of the input cases.It can not only provide professional legal advices for judicial staff,but also provide efficient but cheap legal advisory services for the non-professional who are in need.In the past,most researches did not consider the internal logic between tasks,and examined multiple tasks separately.Due to the neglect of the relevance between tasks,which was quite different from the process of artificial judgment,the judgment results predicted in the same case were split or even contradictory,which led to the lack of sufficient interpretability and many obstacles in practical application.Considering this defect,these models are not included in our research.In this paper,we mainly study the multi-task joint models of legal judgment prediction.Based on the correlation between tasks,the multi-task joint models examine the multiple tasks of judgement prediction in the same model.They can simulate the process of artificial judgment.There are three tasks of judgment prediction in this paper: prediction of relevant articles,accusation and sentence.That is,according to the given description of the case,we need predict the relevant articles and accusation and sentence of the defendant one after another.The data of the experiment is the criminal cases of single crime with single defendant.In this paper,some data from CAIL2018 is used as test data to compare the experimental results of a variety of multi-task joint models of legal judgement prediction.And these models are classified according to the impact direction between the tasks.In addition,combined with the advantages of two models,we propose a new model,MPBFN-WCA model integrating the keywords of the articles of law.The experimental results show that the multi-task models of bi-feedback are better than the multi-task models of one-way feedback in most indicators of three judgement tasks.And the new model is slightly better than MPBFN-WCA model in terms of law recommendation and accusation prediction,but has no obvious advantages in terms of sentence prediction.