Research On Multi-label Charge Prediction Algorithm Based On Hierarchical Attention Mechanism Sequence Generation Network

With the development of artificial intelligence technology represented by deep neural network,the judicial field is moving forward in the direction of intelligence and automation.As an important part of the judicial trial,the task of charge prediction is directly related to the verdict and nature of the whole case.At present,the research of single-label charge prediction based on the description of crime facts has been relatively mature.However,in the practical application scenario,the situation of "multiple crimes for one person" and "combined punishment for several crimes" also exists at the same time,which requires the research on the prediction task of multi-label chargeThe traditional multi-label charge prediction method of criminal facts mostly uses the threshold neural network,that is,manually sets the prior threshold value on the basis of the crime probability vector,and completes the prediction task according to the single-label and multi-class method.Some other researches use the problem transformation method to transform the multi-label charge prediction into the parallel single-label charge prediction task.This method constructs and trains the classification model on each label,then integrates the classification model and finally realizes the multi-label charge prediction.Although the calculation logic of this method is simple,it fails to take into account the logical connection between crimes,such as the crime of drug trafficking and the crime of allowing others to take drugs,leading to a decrease in the recall rate index when predicting cases with intrinsic criminal relation.Moreover,when the sample label space is too large,the complexity of the model will increase,which is not easy to use in the real scene.On the basis of analyzing the previous research work,on the one hand,this study transforms the multi-label charge prediction task into the charge sequence generation task,and integrates the logical relation between the charges into the model from the perspective of machine translation,that is,completes the mapping from the description of the criminal facts to the charge sequence.The recurrent neural network unit is used for chain calculation and serial output of charge sequence,which improves the effect of multi-label prediction.At the same time,because it is not necessary to build parallel training data sets,this method also reduces the labor cost of data processing in large sample label space.On the other hand,in view of document level characteristic of criminal facts,the traditional recurrent neural network can no longer meet the modeling needs of long-sequence texts,because the chain structure of the recurrent neural network is prone to gradient disappearance or gradient explosion during model training.In this study,the model uses the network structure based on hierarchical attention mechanism,breaks down the criminal facts according to the sentence structure,and conducts the attention modeling at the word and statement levels respectively to obtain the text representation vector.Through this hierarchical mechanism,the model can excavate the key words and key statements in the criminal facts,thus alleviating the problem of information loss caused by the excessively long text of the criminal facts and improving the accuracy of the model prediction.Through the comparison experiment on the multi-label data set,the sequence generation model based on the hierarchical attention mechanism proposed in this study significantly improves the accuracy and recall rate of the multi-label charge prediction task and the model complexity was reduced.Experimental results fully verify the effectiveness of the hierarchical attentional mechanism and increasing the logical association of charges on the prediction task of multi-label charges.Through the experimental comparison with multiple baseline models on two multi-label data sets,the sequence generation model based on hierarchical attention mechanism proposed in this study has significantly improved the accuracy and recall rate in the task of multi-label charge prediction.The experimental results fully verify the validity of the hierarchical attention mechanism and increasing the logical association of charges to predict the multi-label charges.