| In recent years,the rapid development of artificial intelligence technology and natural language processing technology has also promoted the process of intelligent development in the judicial field.Artificial intelligence not only provides research theory for intelligent justice,but also provides technical support for the proposal of more intelligent judicial technology.Charges prediction is a core task in the field of intelligent justice,is also a hot research topic in the field of natural language processing,Crime prediction is the core task in the field of intelligent justice,and also a hot research topic in the field of natural language processing.because the purpose of charge prediction is to hope that computers can learn the knowledge of the judicial field,and then have the ability to judge cases with the same ability as human judges and lawyers,the case document text gives the prediction results of charges,Therefore,the purpose of charge prediction is to automatically predict the charges involved in the case according to the given case document text,and to complete the prediction by modeling the criminal facts based on semantic differences.Multi-label charge prediction is usually regarded as a text multi-label classification problem in the judicial field.In this thesis,based on the open electronic documents of China Judicial Documents Network,the multilabel charge prediction method is improved from two aspects of deep learning modeling and introducing semantic attention of crimes.This thesis mainly includes the following aspects:(1)Proposed multi-label Charge Prediction model Based on Case Documents(BCDMCP).After the text content of case documents is input into BCDMCP model,the semantic representation of each text can be obtained by pre-training through BERT model,and the vector representation of words can be obtained.After that,the vector representation of the word is input into BiGRU,and the semantic information is analyzed and extracted by neural network.Then input the case document vector output by BiGRU layer into the attention layer,and give different weight to different semantic vector to highlight the key attention information.Finally,the word vector is classified by Sigmoid layer,and the probability value on each label is calculated by Sigmoid,and the probability value and threshold value are judged to achieve multi-label charge prediction.(2)The Charge Semantic Attention Embedding multi-label Charge Prediction(CSAMCP)is proposed.CSAMCP model will introduce charge text for charge prediction.Firstly,case document text is input into BERT model for pre-training to obtain vector representation of case document text.Then,the pre-trained vector is input into bidirectional gated recurrent unit neural network to learn semantic correlation between vectors.Charges at the same time the text word vector obtained by Glove,then charges a text word vector with case document text semantic vector input to attention,attention mechanism can effectively highlight the case document text characteristic and charges text,highlight the relationship between important information to give greater weight,to optimize the tag classification effect more,Finally,the probability value on each label is calculated by Sigmoid,and the probability value and threshold value are judged to achieve multi-label charge prediction. |
