Research On The Method Of Judicial Public Opinion Summarization Enhanced By Domain Knowledge

With the continuous expansion of the scale of the internet,public opinion events,especially high-risk legal-related events,occur frequently,which has severely affected the trial work of the courts.Therefore,generating brief summaries for legal-related public opinion news is of great significance for users to quickly understand and handle it on time.However,generating summaries for legal-related public opinion news can be regarded as a domain-specific task,which is difficult to generate concise,fluent,readable,and factual summaries only rely on a data-driven approach.Therefore,this thesis focuses on the research of domain knowledge enhanced legal-related public opinion news summarization tasks,including legal-related news and topic,boosting the summarization quality by deep mining and effective use of different types of domain knowledge.The main contributions and innovations are summarized as follows:(1)A summarization model of legal public opinion news by exploiting comments information is proposed.Due to the data-driven summary model lacks key information guidance,the ability to identify key information in the source document is insufficient,which usually leads to the problem of content deviation between the generated summarization and the source document.The legal public opinion news on the Internet will generate a large number of user comments,which usually pay close attention to the focus of the event and are effective auxiliary resources for summarization generation.Therefore,this thesis proposes a legal public opinion summarization model mining important case elements from user comments as domain knowledge to promote the model's performance.A dual-channel de-noising module and a two-way attention module are proposed to solve the modeling and integration problems of key case elements in noisy comments and realize the summarization generation under user comments constraints.The experimental results show that this method can better model the key information of the source document compared with the traditional keyword-based summarization methods,and greatly improve the informativeness of the generated summaries.(2)A summary method of legal public opinion news by incorporating topic information is proposed.There are many differences in terms and topic expressions of public opinion events on different topics,and there may be a problem of topic inconsistency between the generated summaries by the existing model and the source document.To tackle this problem,this thesis proposes to model the summary task from the perspective of the topic,treat the source document as a mixture of many different topics,where the word-level topic words are used to capture the core topic of the source document,and the document-level document topic probability distribution is used to further generate the summary consistent with the original topic from the global constraint,to generate a highquality summary.The experimental results show that integrating multi-granularity topic information can effectively solve the key information modeling problem in the scenario with no comments or other external resources in the legal-related news summarization task,and improve the core information identification ability,and further boost fluency and topic consistency summaries generation.(3)A method of legal public opinion summary enhanced by elements graph is proposed.The existing legal-related public opinion news summarization model may generate summaries that conflict with the fact of the source documents,that is,there is a factual deviation between the generated summary and the source document.However,the keywords and topic words,which are discrete knowledge,are difficult to describe the factual framework of events.To address this issue,this thesis proposes to extract the key case elements from the source document,and construct the elements graph that represents the core ideas and facts of the case as domain knowledge,and use the graph transformer network and pre-training language model to effectively model the element relational graph to help the model better understand the global information of the case,and generate a brief,factually consistent summary.The experimental results show that in the legal public opinion summary task,the introduction of explicit structured case representation can improve the factual consistency of the summary.(4)A case topic summary method enhanced by topic interaction graph is proposed.Case topic summary is the task of generating a summary from user comments on case-related topics,which is also an essential task for legal-related news.However,the length of topic-related comments varies greatly,there are many noisy data,and it is difficult to select important information from the noisy comments.Therefore,this thesis proposes a graph-based two-stage topic summary generation method.which constructs a topic interaction graph that including case elements and key sentences.The relationship between important case elements in the comments is effectively modeled,and the graph structure is used to realize noise data filtering and important information selection.Experimental results show that the introduction of the topic interaction graph in the case topic summary task can effectively reduce the difficulty of processing complex documents based on retaining important information,and improve the quality of case topic summary generation.