Research On The Method Of Sentence Prediction For Theft Case

With the continuous promotion of "Smart Courtt" construction,artificial intelligence technology is gradually applied to judicial intelligent trial assistance,one of the tasks is sentence prediction.Sentence prediction means predicting the length of sentence based on the description of case facts,crime amount and other key information of the case.As one of the key tasks in case prediction,sentence prediction plays an important role in the"Smart Court".It can not only provide a benchmark and reference help for the court in sentencing,but also serve the general public at the grassroots level.Therefore,the study of sentence prediction is of great significance and value.The current sentence prediction methods suffer from strong specialization and insufficient related research.Most traditional methods use deep neural networks to first encode the textual information of a case,then identify the key case elements for trial-such as "recidivism","surrender ",etc.,and then regress or classify them to predict the final sentence.This approach requires training using a large amount of data and is very dependent on the size of the dataset,yet such datasets involve specialized areas of justice and are costly to obtain.The neural network approach does not take into account the trial process and trial logic in the prediction process,which makes the predicted sentence weakly interpreted.Meanwhile,the lack of deeper use of external domain knowledge such as relevant laws and regulations,as well as the problem of low utilization of dependencies between subtasks,makes the accuracy of prediction results also low.The main research works and results of this thesis to address the above issues are:(1)A sentence prediction model for theft cases based on inverse deduction learning fused with trial logic is proposed.The model is based on abductive learning approach that takes full account of trial logic.First,a legal knowledge base is established for the characteristics of theft cases,and it is made more fully applicable to the identification of key circumstance elements and sentence calculation.The knowledge base provides weakly supervised information for the neural network in the identification of key circumstance elements,and can provide a basis and constraint for the sentence results trained using unlabeled data in the sentence calculation after comparing the knowledge base by the legal basis extracted by the canonical.Second,a sentence calculation method that is more consistent with the sentencing rules is designed with reference to the Sentencing Guideline.Therefore,under the condition of the same training data,the MAE error of this model in the experiment of responding to the legal documents of theft cases is reduced by 1.63 months compared with the domain-free knowledge,and the error is only 2.62 months,which makes the results more accurate to assist in the sentencing in the judicial trial process.(2)A multi-task learning-based sentence prediction model for theft cases is proposed.In the previous method,although the sentence prediction method is designed to solve the problem of insufficient consideration of trial logic by referring to the flow of judge’s trial cases,and to improve the interpretation of prediction by utilizing external law-related knowledge,the dependency relationship between the two subtasks of sentence prediction,key circumstance element identification and sentence calculation,is underutilized as well as the efficiency of underlying information acquisition is also low.In this section,the identification of key circumstance elements and sentence calculation are jointly learned through a multi-task learning framework to share the underlying parameters and optimize the sentence calculation method to fit the framework,so that they can play a supervisory role on each other and improve the accuracy of prediction.Experiments show that this method achieves superior predictive accuracy in terms of sentencing compared to traditional methods on the same dataset,with a reduction in Mean Absolute Error(MAE)to 2.37 months and an improvement in R2Score to 0.88.