Research And Implementation Of Automatic Question Answering Technology Based On Knowledge Graph

Knowledge graph is currently a popular research field with a variety of downstream applications,and its practicality has gained widespread attention from both academia and industry.This thesis focuses on the automatic question-answering technology based on knowledge graph.This task presents challenges due to the large scale of the knowledge graph and the heterogeneity between question-answer triplets.Many current research methods use complex rules to narrow down the search space and only reason on a subset of the knowledge graph due to memory overhead issues,which is not convenient for generalizing to larger graphs.In addition,for more complex multi-hop questions with syntactic structures,many methods solve them as semantic matching tasks,but only consider the word-level semantics of the relationship path,or simply randomly initialize the relation name,and the generalization ability relies on the naming rules of the relationship.To address these issues,this thesis conducts related research work,and the main research contents and contributions are summarized as follows:(1)A single-hop KBQA method based on pointer network and retrieval and reranking :This thesis presents a single-hop KBQA approach that can be adapted to larger scale knowledge graphs.The method accesses the knowledge graph by retrieving the reordering framework in the entity link and relation detection subtasks,overcoming the memory overhead of computing on large-scale graphs.The method also considers both character and semantic feature scores in the ranking phase as well as correlation between subtasks using multi-task learning in the training phase.Finally,several related experiments on two datasets,Simple Questions and NLPCCICPOL-2016-KBQA,demonstrate the effectiveness of the method for both the overall task and the subtasks.(2)A multi-hop KBQA method based on knowledge representation learning:This thesis presents a new relational path extraction method for multi-hop KBQA tasks that can alleviate the generalisation dependence of previous methods on relational naming rules.The method introduces knowledge representation learning,pre-trains the knowledge graph using Rotat E,a rotation-based knowledge embedding model,and views the multi-hop relational path embedding as a transformation from the first to the last relation.Finally,several related experiments on two datasets,Meta QA and Web Questions SP,demonstrate that the method can effectively improve Hits@1scores for multi-hop KBQA tasks.(3)Knowledge graph-based encyclopedic question and answering system: Combining the current situation of digital transformation of education and the analysis of the needs of Sequoia Encyclopedia,this paper collects knowledge related to computer hot words from two sources and organizes and fuses them to build a computer encyclopedia knowledge map.Then,this thesis extended the KBQA method proposed in the previous chapters by adding a similarity matching-based FAQ question and answer module,and finally implemented a computer encyclopedia question and answer system based on Vue and Flask frameworks to interact with users with intelligent Sequoia and put it into use,giving the platform a certain degree of intelligence and providing some reference value for subsequent research in the field of computing pedagogy.This thesis first proposes a strategy to complete the knowledge graph question and answer subtask in two stages using a retrieval rearrangement framework to overcome the memory overhead problem of scaling the model to larger-scale graphs,and also designs a matching method combining multiple feature information and a multi-task learning training approach to further improve the model effectiveness.The thesis then focuses on the multi-hop chaining problem in complex problems,and proposes a new multi-hop relational path extraction method based on knowledge representation learning,which alleviates the problem that the effect of previous methods is limited by relational naming rules and achieves better multi-hop quiz performance.Finally,based on practical requirements,this paper extends and applies the proposed method to demonstrate its practicality and provide reference value to related research in the field of computational pedagogy.