A Study Of Chinese Named Entity Recognition For Judicial Field

With the continuous disclosure of judicial big data,how to apply artificial intelligence technology to the judicial field to assist judicial workers to improve the efficiency and fairness of case handling has gradually become a hot spot in the research of judicial artificial intelligence.Named entity recognition technology is the basic and key work in this field.It would play an important role in the downstream judicial questions and answers,the prediction of prison terms and charges,and the construction of judicial knowledge map.However,the research on judicial named entity recognition is still in its infancy,and there are still many outstanding problems.Based on the pre-training language model and the idea of multi task joint learning,this paper makes a more in-depth study on some prominent problems.The main work is as follows.Firstly,aiming at the problems of judicial entity classification errors and boundary errors in judicial named entity recognition,this paper proposes a judicial named entity recognition method based on pre-training language model BERT and multi-task joint learning.In this method,the BERT model is used as the multi-task sharing layer,and then the named entity recognition task is trained upon the sharing layer.At the same time,the judicial text classification and judicial Chinese word segmentation tasks are used as aids to learn the text category features and word segmentation features respectively,and the global information and vocabulary information are added to the named entity recognition sequence.Finally,the conditional random field is used to obtain the optimal sequence annotation,and the judicial entity recognition results are obtained.Secondly,aiming at the problem of nested named entity recognition in the judicial field,this paper proposes a nested named entity recognition method based on fragment extraction.Based on the above multi-task framework,this method builds a named entity recognition network based on fragment extraction.The network inputs the text features extracted from the multi-task sharing layer into the beginning and end recognition network for entity boundary detection.Finally,the recognized candidate judicial entity segments are sent to the entity classifier to get the final entity category.Finally,the experimental comparisons show that the proposed methods have achieved good results on the judicial smoothing entity dataset CAIL2021-Flat and Drug and the judicial nested entity dataset CAIL2021-Nested.At the same time,ablation experiments verify the effectiveness of the modules added in the methods.