Legal Judgment Elements Extraction For Factual Description Of Legal Documents

In recent years,the prediction of legal judgments for intelligent judicial services has become a research hotspot in the field of natural language processing.The legal judgment elements extraction is one of the important subtasks of legal judgment prediction research,which aims to identify different judgment features from the fact description of legal documents automatically.The existing research mainly extracts judgment elements words or sentences from the fact description of legal documents.Based on the CAIL2019 "Legal Judgment Elements Extraction" task,we focus on the sentence-level legal judgment element extraction method for fact description.At present,the research on the task of legal judgment element extraction mostly adopts the idea of text classification.When the existing methods deal with sentences with large difference in length,they often add many irrelevant filling vectors to the short sentences,which has a certain impact on the model results.Therefore,we propose a legal judgment elements extraction approach based on mask mechanism.At the same time,the analysis results of a large number of real legal cases show that the judgment elements are usually closely related to the law article involved in the case,and most of the existing methods ignore the information of the law article.Based on this,we further propose the legal judgment elements extraction approach with law article-aware mechanism.Finally,we integrate the legal judgment element extraction method into the judgment prediction task and propose a legal judgment prediction approach based on legal judgment elements extraction.The main work includes:Firstly,we propose a legal judgment elements extraction approach based on mask mechanism.We formalize the legal judgment element extraction task as a multi-label classification model of fact description sentences and propose a legal judgment element extraction approach combining BERT and CNN.At the same time,in order to weaken the negative impact of different sentence lengths on the effect of model,we further integrate the multi-head self-attention mechanism based on the mask method into the BERT-CNN model.The method is validated on the dataset of the CAIL2019 legal judgment element extraction task.Secondly,we propose the legal judgment elements extraction approach with law article-aware mechanism.According to the process of judge’s judgment,combining with the Marriage Law of the People’s Republic of China,we propose an attention mechanism integrating the semantic information of the law article.Then we integrate the semantic information of the law article into the fact description sentence through the attention mechanism of the law article.Finally,we use the multi-label classification model to identify the judgment elements in the fact description.Experimental results show that our method achieves significant improvements than other state-of-the-art baselines on the dataset of legal judgment element extraction task.Thirdly,we propose the legal judgment prediction approach based on legal judgment elements extraction.We use the judgment element extraction approach to extract high-quality judgment element sentences,then we transform the legal judgment prediction task into three subtasks:law article prediction,charge prediction,and penalty prediction.We construct the prediction models based on the legal judgment elements.Experimental results show that the F1-values of the three judgment prediction tasks are improved,which further verifies the effectiveness of legal judgment element.