Research On The Construction Method Of Retrieval Question Answering System For The Legal Field

It is a basic national policy of our country to comprehensively promote the rule of law and actively bring into play the active role of the legal system in national governance.Since the founding of the People's Republic of my country,the level of education of our citizens has been greatly improved,but when legal means will not be taken to solve problems,the main reason is that legal awareness is still weak.Building a question-and-answer system oriented to the legal field can provide legal advice to the public,so it has great application value.The question answering system is divided into search question answering system and generative question answering system according to the way of answer generation.In the legal field,due to its professionalism and seriousness,the generative question and answer system is not effective,while the retrieval question and answer system is more effective.However,there are still the following problems in constructing a search question answering system oriented to the legal field:(1)The question and answer corpus in the legal field is scarce,and there is a lack of legal question and answer knowledge base.(2)The keyword-based retrieval method ignores the semantic information between the answers to the questions,and when the answers lack keywords,the correct answer cannot be retrieved.(3)Current text matching methods extract high-level features,and it is difficult to achieve fine-grained semantic alignment.In this paper,the research on the above problems has been carried out.The following results have been achieved:Construct a question-and-answer knowledge base oriented to the legal field.The crawler technology is used to crawl the accepted question answer pairs from professional legal consulting websites to obtain the question and answer knowledge base,and perform noise reduction preprocessing.(2)Research on subject classification methods for the legal field.First,the text vector based on the word vector and the label vector are fused to obtain a single text representation,and then the self-attention map convolutional neural network is used to extract the global features of the text,which is input into the classifier,and finally the problem classification result is obtained.Experimental results on the Chinese legal field data set and multiple English classification data sets show that this method has achieved better results than the baseline model.(3)Multi-step fine-grained question-answer matching method based on BERT.First,use BERT(Bidirectional Encoder Representations from Transformers)to encode the questions and answers separately,and then use the gating mechanism to perform memory distillation processing on the input information.Finally,the semantic matching degree is calculated for the information retained by the distillation,and according to the matching The score is used as the basis for sorting answers,and Top-1 is returned to the user.The experimental results of multiple Chinese question and answer data sets show that the method in this paper is effective.(4)Build a law-oriented search-type question and answer prototype system.Take the first research point and the second research.As the core algorithm,and based on demand analysis,use the relevant flask and VUE framework to build a search-based question and answer system for the legal field.