Research On Knowledge Base Question Answering Method For Complex Semantics In The Judicial Fiel

With the construction and advancement of the Smart Court Project,the judicial intelligent question answering system provides an efficient and convenient professional platform to meet the legal consultation needs of users,which not only reduces the burden of judicial practitioners,provides legal assistance to the public,but also broaden the channels of law popularization.With the continuous development of knowledge graphs,knowledge base question answering technology shows great advantages over traditional text retrieval methods,and brings huge breakthroughs to judicial intelligent question answering tasks.However,due to the particularity of its domain attributes and application scenarios,the knowledge base question answering system for the judicial field still faces two difficult problems.One is the lack of knowledge graphs and question answering datasets in the judicial field,the penetration is not deep enough;the second is that most of the user’s questions contain complex semantics,this thesis divides them into two types of complex questions:combination type and multi-hop type,the knowledge base question answering technology has insufficient processing ability for these complex questions.Based on this,this thesis conducts research on knowledge base question answering methods for complex semantics in the judicial field.The thesis mainly completed the following research work:(1)Build knowledge graphs and complex question answering datasets in the judicial fieldDue to the lack of knowledge graphs in the judicial field,this thesis selects judicial judgment documents as the data base to construct a judicial knowledge graph.Firstly,the data of judicial judgment documents are obtained and analyzed.Secondly,the judicial knowledge is modeled,and the entity categories,the entity attributes,the relationship between entity categories and attributes is defined.Then,knowledge is extracted by artificially designed rules,and knowledge fusion is carried out through knowledge fusion.The information is represented as triples in the knowledge graph,and finally the constructed judicial knowledge graph is stored in the Neo4j graph database.In addition,on the basis of the knowledge graph,this thesis constructs a complex question answering dataset through artificially designed templates,which provides strong data support for the research work on question answering methods in the knowledge base in the subsequent chapters.(2)Semantic parsing knowledge base question answering method based on combination question decompositionIn the current judicial intelligent question answering platform,a large number of legal questions are usually complex questions that are combination type,which are composed of several simple questions.Therefore,decomposition and synthesis strategy is an effective way to solve such complex questions.However,during decomposition,the knowledge base question answering model how to fully understand the complex semantics in the question,how to determine the decomposition method,and how to ensure the quality of the sub-questions after decomposition,these are some key issues faced by this task.Therefore,aiming at the above difficulties,this thesis introduces a questiondecomposition semantic parsing model that integrates factual text.The processing of complex questions is divided into three stages:decomposition-extraction-analysis.First,the complex questions are decomposed into simple sub-questions,and then extract the key information in the question,and finally generate a structured query.At the same time,a factual text base is constructed,the triples are converted into sentences described in natural language,and the attention mechanism is used to obtain richer knowledge representation.Experiments show that the model achieves better performance in solving combination complex questions,which proves the effectiveness of the decomposition strategy.(3)Retrieval-ranked knowledge base question answering method based on multi-hop relational reasoningAnother large number of legal questions are multi-hop complex questions involving multiple knowledge information.Such questions usually require models to make inferences and judgments based on existing knowledge.However,the multi-hop process will generate long paths and bring huge computational costs,and the model does not have long-term memory and will forget the path,and in the case of an incomplete knowledge base,how to improve the reasoning ability of the model has become the key issues that restrict the effect of question answering.Therefore,aiming at the above difficulties,this thesis introduces a multi-hop reasoning retrieval ranking model based on sample retrieval,which mainly adopts the method of retrieving similar samples and the key-value memory network to perform long-path reasoning on an incomplete knowledge base.First,input the semantic representation vector of the question,and similar samples are retrieved from the training set according to the subject entity.Secondly,in the inference process of each hop,the model uses the attention mechanism to update the semantic representation vector of the question and the set of similar samples.Finally,the storage and reasoning capabilities of the key-value memory network are used to determine the path of the next hop,and iterate this process.Experiments show that the model achieves good results in answering complex questions of multi-hop type,which verifies the effectiveness of retrieval and reasoning techniques.(4)Knowledge base question answering prototype system for judicial fieldThis thesis firstly analyzes the needs of the intelligent question answering in the judicial field,takes the constructed judicial knowledge graph as the data base,encapsulates the above knowledge base question answering algorithms into interfaces,and builds a knowledge base question answering prototype system for the judicial field,it realizes functions such as retrieve historical files,key information question answering,retrieve similar cases and laws and regulations question answering,which provide efficient and convenient question answering services for legal practitioners and the public.