Research And Implementation Of Improved KBQA System For Complex Questions

In recent years,with the emergence of artificial intelligence technology,the in-telligent question answering systems obtain a rapid development and become a con-venient and natural way to interact with the prevailing structured data in a variety of domains.Nowadays,knowledge graph based question answering(KG-QA)has be-come a research hot spot which has attracted a large number of researchers.Knowledge graph is one of the most important data resources in the era of artificial intelligence and aims to describe the various entities and relations that exist in the real world.The KG-QA systems accept the questions in natural language form as input and return the corresponding entity from knowledge graph as the answer.Nowadays,KG-QA systems are widely used in various fields and provides a con-venient way for people to interact with structured data.With the proceeding of more decision-making usage scenarios,the user's information needs,i.e.,input questions be-come more complex and ad hoc.Recently,we find that the comparison,relation,and opinion questions are witnessed significant growth,especially in some product domains.However,most of the current KG-QA methods cannot appropriately handle the inherent complex relation and coverage characteristics within the questions.In order to solve these problems,we propose an improved neural network based KG-QA systems.Complex questions usually imply complex relations which cover more hops that leads to massive growth of the candidate set.Therefore,the key of complex questions answering is relation detection for complex questions.In this paper,we divide the framework into three components,which are topic entity recognition,re-lation detection and constraint detection.In topic entity recognition stage,based on the sequence labeling model which is prevailing used in name entity recognition,we step further to propose a heuristic method for matching and a re-rank model,in order to im-prove the performance.In relation detection stage,we propose to utilize the relation information with the questions and knowledge graph in a mutual way,improving the fi-nal question answering performance.Specifically,we design local and global attention models for relation detection.Finally,we combine the topic entity recognition model and relation detection model as the final question answering system.Besides,we also design a text matching based constraint detection strategy in order to solve the questions with constraintFor each components and the whole system in our paper,we conduct the exper-iments respectively.Experiments on our new dataset and common datasets reveal its advantages both in accuracy and efficiency.