Research On Evaluation Method Of Legal Case Simulation System Based On Recognizing Textual Entailment

In recent years,artificial intelligence technology has been widely used in the judicial field,providing it with strong technical support and promoting information modernization and intelligence in the judicial field.This thesis studies the evaluation method of legal case simulation system based on recognizing textual entailment.The claim evaluation needs to judge whether the claim is established according to the judicial decision,which is the most difficult one in the evaluation task.The method realizes the claim evaluation by converting it into recognizing textual entailment.The main work of the thesis includes the following aspects:(1)Developed a claim and judicial decision entailment data set.The main work of this part is to extract the claim and the judicial decision from the existing large-scale judgment of the referee document net,combine the characteristics of the judgment itself and the actual needs of the legal case simulation system evaluation,design the appropriate labeling specification,and complete the data set development.(2)Studied the evaluation method of claim based on Decomposable Attention.This method is the baseline system.It mainly uses Bidirectional Long Short-Term Memory to learn the context information,and then uses the Decomposable Attention mechanism to obtain the interaction information of the sentence pairs.After comparing the interactive information,the classifier is used to determine the entailment relation.(3)A evaluation method of claim based on combined attention mechanism is proposed.Since Long Short-Term Memory tends to lose some important long-distance dependencies in long texts,this method adds Self-Attention mechanism to capture the dependence information inside sentences of arbitrary length.At the same time,through the Highway Network to solve the training difficulties caused by the increase of network depth,and achieved good results.(4)A evaluation method of claim based on Memory Network is proposed.Aiming at the problem of small historical memory generated by cell state and it is impossible to accurately record the entire contents of the text in Long Short-Term Memory,the model uses the readable and writable external memory storage module in the Memory Network,takes advantage of the inference group to train,and finally obtains the flexible memory module to increase the memory.At the same time,in the sentence coding,the Multi-Headed Self-Attention mechanism is combined to capture the internal structure information of sentences at different aspects to enhance the structural semantics of the sentence.The experimental results show that compared with the baseline model,the evaluation method of claim based on the combined attention mechanism has better performance,and the model based on Memory Network has better results than the former two.