Legal Judgement Prediction Based On Causal Inference And Multi-expert FTOPJUDGE

In recent years,my country’s judicial field has been committed to promoting the digital intelligence of the judiciary to improve the work efficiency of judicial personnel,simplify work processes and make trials more fair and just.As a key part of smart justice,legal judgment prediction uses deep learning related technologies to deeply understand the description of the facts of the case and introduce additional knowledge as auxiliary key information to predict the charges,laws and sentences of the defendant.This greatly improves the work efficiency of judges,and also provides high-quality reference information for people who have no legal background but still want to understand the content of the case.Since most of the current legal judgment prediction models start from the data dimension,whether it is structured data or unstructured data,it is a simple superposition of data,ignoring the causal relationship between data and results,so the model may learn and The information irrelevant to the results even falls into the trap of probabilistic association and affects the fairness of the model decision,which is obviously contrary to the original intention of legal artificial intelligence.Secondly,for the three sub-tasks of legal judgment prediction,most of the current work considers their relationship to be complementary based on the actual process of judge judgment.But what happens in reality doesn’t necessarily match up with what happens in a neural network,and sometimes it’s quite different,and little work takes this into account.Thirdly,adjudication documents usually contain a lot of legal terminology,and it is difficult for the model to understand its specific meaning only through the traditional word embedding mechanism.Finally,the small sample problem caused by the small number of certain cases is still common in legal judgment prediction,and these problems need to be solved urgently.Aiming at the above problems,this thesis proposes a legal judgment prediction model based on causal inference and multi-expert FTOPJUDGE mechanism,which provides new perspectives and ideas for legal judgment prediction,and the experimental results prove that the model has superior performance.The main innovations and work of the thesis are as follows:(1)Aiming at the problem that the existing model cannot find the real causal relationship between the description of the facts of the case and the law,this paper proposes a framework for processing unstructured text using a causal inference algorithm,which does not require too much manual intervention.Good mining of the information in the text,while bringing better interpretability.And by integrating causal information into the neural network,the neural network has the ability of causal inference to achieve better results.From the experimental results,the F1 values of the law prediction task and the crime prediction task have been improved by 3.71%and 2.44%respectively compared with the current best performing model,which also proves that it can solve the small sample problem to a certain extent.(2)For the problem of how to determine the relationship between the three sub-tasks of decision prediction and design a suitable model structure according to the relationship between the tasks,this thesis first verified through ablation experiments that the three tasks of decision prediction are not purely complementary in the neural network.There is a certain degree of competition,so this thesis proposes a multi-expert mechanism to set up its own neural network for each task,and also set up a neural network that serves multiple tasks at the same time,which meets the needs of each task.While meeting the needs of each task and alleviating the conflict between multi-tasks,it also realizes the common progress of related exchanges between multi-tasks.At the same time,in order to better adapt to the output of the multi-expert mechanism,TOPJUDGE is transformed by using the fully connected network as the basic structural unit,and FTOPJUDGE is proposed.From the experimental results,compared with the current best-performing legal judgment prediction model,the overall improvement of various indicators,especially the 10.82%improvement in the Acc.of legal clause prediction,proves that the multi-expert FTOPJUDGE proposed in this thesis is effective It achieves the balance between the three tasks in decision prediction and improves the overall performance of the model.(3)In view of the problem that it is difficult for the model to understand the legal terminology in the judgment documents,this thesis uses the pre-trained language model Lawformer to provide the model with prior knowledge,which has trained tens of millions of Chinese legal documents so that the model has a stronger generalization ability.The experimental results show that the pre-trained language model does help the downstream model to understand legal documents,and has made progress in multiple evaluation indicators.