Research On Classification Of Judicial Documents Based On Bert-LSTM Model

With the continuous improvement of judicial construction in China,a large number of judicial documents are open on the Internet,and there are a lot of valuable information in the judicial documents.It will be of great value to dig these information.In the context of deep learning,text classification is one of the important basic tasks in the field of natural language processing.This study conducts text classification experiments based on the data of referee documents.By classifying the cases in the judicial documents,we can provide reference for judicial decisions to improve the efficiency of the judges’ handling cases and provide valuable reference for the "same case and verdict" in the judicial field.A judgment document contains a large amount of information.It is of great significance to classify the judicial documents and facilitate judicial personnel to consult.This paper is the classification of the referee documents is discussed by using the Bert model.This method is mainly based on the multi-layer feature fusion of Bert pre training language model to encode,and vector representation of text information in referee documents is carried out by Bert as service tool,and deeper text features are extracted to classify the referee documents.The accuracy of the model in the classification of multi label of the cases of adjudication documents is 87.42%.In addition,this paper also uses a method based on Bert LSTM to analyze and classify the referee documents in a deeper level.This method uses the pre training model Bert to provide word vector,and fine tune the word vector by the Bert model and the context of the referee document.The vector representation which combines the full text semantic information is input into the LSTM model.The self attention mechanism is used to give higher weight to the important information in the text,and LSTM is used to encode and fuse the sequence to obtain the final category information.The experiment shows that the performance of Bert LSTM model has been improved obviously,and the accuracy of the classification is 1.12% compared with the classification of referee documents based on Bert pre training model.