Research And Implementation Of A Complex Question-Answering System Based On Knowledge Graphs

With the advancement of knowledge graph technology,question answering technology based on knowledge graph emerged as the times require,and has become a research hotspot in academia and industry.Knowledge graph Q&A can accept users’ natural language questions,use knowledge graphs as data support,use the rich semantic correlation information of graphs to understand user intentions,and directly return accurate answers,which significantly improves the efficiency of users in obtaining information.Constructing a highly available question answering system for vertical fields is also an important direction for many industries to achieve intelligent transformation.Existing studies perform well on answering simple questions,but poorly on answering complex questions such as multi-hop questions.In addition,there are too few training data for complex problems,and high-quality data sets are urgently needed to train models.This paper conducts research on the above issues,mainly including:(1)A complex KGQA model named CT-KGQA is designed and implemented.The model introduces the knowledge graph embedding model ComplEx to model the graph,and uses the pre-trained language model mT5 to understand the semantics of questions.Embed question entities,question sentences,and candidate answer entities into the same complex number space,and use the ComplEx scoring function to sort candidate entities to obtain answers.The model is free from the constraints of adjacent edges,and it also performs well in the case of extremely sparse knowledge graphs.Compared with the baseline model,it has achieved the best performance.(2)A method for generating multi-hop questions is proposed.This paper proposes the Multi-KGQG task as a new downstream task of the pre-trained language model mT5,and realizes the generation of multi-hop questions based on knowledge graphs.(3)Starting from data acquisition,a domain knowledge graph of people in the Taiwan Strait was constructed through technical means such as named entity recognition,relationship extraction,entity disambiguation,and attribute fusion,and a multi-hop question containing more than 3,700 questions was constructed based on this graph data set.(4)Apply the model proposed in this paper to KGQA,combined with existing engineering technology,design and implement an automatic question answering system in the field of characters in the Taiwan Strait.