Mining Auxiliary Sentencing Rules For Prosecuting Agency Based On Deep Learning

Auxiliary sentencing rule mining technology is one of the core contents of"Intelligent Court",and it is also a research hotspot of the combination of law and artificial intelligence.Auxiliary sentencing can provide reference for the daily work of judges and prosecutors,and help to maintain the fairness and correctness of our legal system.At present,there are sentencing prediction modules in the judicial system of many local courts which build "intelligent courts*' in our country.However,the accuracy of Named Entity Recognition(NER)of legal documents is not high,and the automaticity of knowledge map construction is poor.This paper discusses the NER task,entity relationship extraction and sentencing rule mining of legal documents,establishes LSTM-CRF model for the named entity recognition of legal documents,the named entity relationship extraction model based on Graph-LSTM model and ATTENTION mechanism,and finally builds a sentencing rule mining model based on Graph-LSTM combined with text features.The experiments prove the model has a high accuracy rate and meets the requirements of the judicial field.Firstly,by introducing CRF model into the named entity recognition model based on LSTM,the problem that LSTM model can not capture the constraints between tags is avoided.Especially combining the domain characteristics of legal documents,the entity recognition model is constructed.By changing the size of training corpus,the performance of centralized machine learning under different data sets is compared and analyzed.Secondly,a named entity relationship extraction model of legal documents based on Graph-LSTM model and ATTENTION mechanism is established.By analyzing the syntactic dependence of legal documents,the semantic and grammatical characteristics of entities are obtained.Graph-LSTM model is constructed based on document structure diagram.Weak label data sets are obtained by remote monitoring.ATTENTION mechanism is introduced to eliminate noise at the end of the model.Comparing with other machine learning model experiments,the performance of the model is analyzed.Finally,a mining model of auxiliary sentencing rules is designed,The accuracy of experimental analysis and cross validation with five folds change with the number of iterations,and the model is compared with other machine learning experiments,which proves the high accuracy of the model.