Judgment Prediction Based On Multi-Class Information Fusion

Traditional method based on machine learning and deep learning can better in crime,law prediction accuracy,but less sample size category accuracy is low,its main reason is that the sample is not balanced,lead to rush the partial model in the learning process,due to less number of training samples is not effectively.Due to the influence of external factors,such as complexity and region,the existing models mostly have the problem that the prediction accuracy of term of imprisonment is not high and cannot reach the applicable level.A single description of a crime cannot provide enough information as input,so additional knowledge must be sought to further improve the effect of sentence prediction.The long description text of crime facts is prone to the problem of long-distance dependence,which directly leads to the decline of the network's ability to extract information and the inability to effectively encode the effective information in the text.Aiming at the above problems,this paper proposes a judgment prediction model based on multi-type information fusion,which can improve the prediction effect by adding structured data based on fact determination on the basis of single crime fact description input.The experimental results show that the decision prediction model based on multi-class information fusion proposed in this paper has high accuracy,good robustness and the ability to learn in a balanced way,and has achieved good results in the task of decision prediction.The main contributions and innovations of the paper are as follows:(1)to solve the problem of unbalanced samples,this paper proposes a method of combining structured data with description of criminal facts.The structured data based on fact determination is extracted from the judgment document,and the structured data information is integrated with the text information.Experimental results show that the method of combining structured data with description of criminal facts can effectively improve the accuracy of small sample data.(2)in view of the difficulty of embedding structured data,this paper divides structured data into continuous data and discrete data,designs continuous data encoder and discrete data encoder respectively,and integrates them with text information.Experimental results show that the continuous data encoder and discrete data encoder designed in this paper can effectively encode the structured data based on fact determination.(3)in view of the long description text of criminal facts,this paper designs a Transformer-HAN structure to improve the text coding ability of the model.Due to the long description text of crime facts,the direct use of RNN encoder will cause the problem of long-distance dependence.Therefore,the paper abandons the traditional RNN encoder and USES HAN encoder.The basic part of HAN adopts Transformer structure to further improve the coding ability of the model.(4)in order to explore whether the multi-task structure is suitable for this task and what kind of task structure can achieve the optimal effect,this paper designs a multi-class comparison model and explores the optimal model on the classifier structure.Experiments show that the multi-task model is more robust than the single-task learning model,and the proposed multi-task structure based on topology is superior to the general multi-task structure,with better results in fl,MP,MR and other indicators.