A Study Of Intelligent Question Answering For Legal Texts

Legal intelligence,which aims to apply artificial intelligence techniques to empower machines to understand legal data to help solve various challenges in the legal field,is one of the critical application scenarios for natural language processing.Text-based question answering,which aims to build systems that can answer questions in natural language form,is one of the most challenging challenges in natural language processing.Legal text-based question answering,also known as legal question answering（Legal QA）,is a typical application of legal intelligence and an essential part of real-world legal advice.An excellent legal question and answer system can be of great convenience to many people and is therefore of great research value.In recent years,with the continuous disclosure of large-scale legal text data represented by adjudication documents and the rapid development of natural language processing technology,research on the application of artificial intelligence technology to empower the judicial field has made remarkable progress.The results of these studies will help judicial practitioners improve the efficiency of handling cases.Furthermore,legal question answering has also received significant attention from academia and industry as an actual application of legal intelligence.As a result,it is gradually becoming one of the hot spots in legal intelligence research.However,at present,there are still many challenges in Legal QA research: 1)the construction of open judicial big data is costly,and there is still a lack of scenario-oriented open legal text data related to newer events,such as the data of adjudication documents related to the COVID-19 pandemic; 2)current semantic matching models oriented towards retrieval-based Q＆A suffer from the problem of inconsistency between semantic representation learning and retrieval-based goals; 3)current legal model structure of the open-domain question-and-answer approach cannot effectively exploit legal knowledge,and the current knowledge exploitation approach suffers from the problem of constraining model performance.Therefore,to address the above challenges,this paper focuses on Legal QA and conducts research and experimental analysis on the key technologies related to the collection and construction of legal text data for specific scenarios,semantic matching models and knowledge exploitation methods in legal open domain Q＆A.The main contributions of this paper are summarised as follows:Firstly,to address the lack of relevant public judicial datasets for the COVID-19 pandemic,this paper collects and manually annotates a series of datasets from relevant official websites to support the implementation of retrieval-based question answering systems.Besides,to address the inconsistency between the supervised training objectives of current semantic matching models and the design of retrieval-based schemes,we propose a semantic relation learning network assisted by a contrastive learning task.The network introduces supervised contrast learning at the sentence level jointly trained with a semantic matching task to solve the above-proposed problem.Experimental results on the datasets show that the proposed approach achieves a significant performance improvement over previous models.Based on the above contributions,this paper develops a legal intelligent question answering system based on the We Chat public platform,which provides a useful reference for the design and implementation of related question answering systems.Secondly,to address the problem that the current open-domain question answering pipeline method cannot be directly applied to the legal open-domain question answering task JEC-QA,we propose a complement-retrieve-answer network for JEC-QA.The network consists of subject information completion,relevant basis extraction,and a question and answer model.In addition,to address the problem that the current state-of-the-art method cannot effectively utilise knowledge information,a question-evidence relation graph-based question answering model is proposed.Specifically,in this model we first construct a question-reliance graph and then use graph attention to select and aggregate the evidence representations in order to make effective use of the evidence information on the relation graph and to improve the performance of the model.Experiments on the benchmark dataset show that the proposed method enables the model to improve the accuracy of answer prediction significantly.