Research On Entity Linking Algorithm Based On Knowledge Graph

With the rapid popularization and development of mobile Internet tech-nology in the world,the text data on the Internet has seen explosive growth.How to use natural language processing technology more efficiently and intel-ligently to analyse and process these text data has become the focus concern of academia and industry.As the most basic technique in the field of natural lan-guage processing,Entity linking can link the entity mentions in free texts to the corresponding entities in the target knowledge graph,effectively solving the problem of ambiguity caused by the phenomenon of synonym and polysemy which commonly exists in natural language.Entity linking can help the ma-chine to truly understand the semantic information of named entities and can be of great significance to the development of relevant fields such as information retrieval,automatic question answering and knowledge graph expansion.In this paper,we focuse on the core issue of entity linking with a structured knowledge graph,and mainly research on the problem of ranking the candidate entities.What's more,we designed and implemented a knowledge base ques-tion answering(KB-QA)system by a combination of the entity linking module and relation extraction module.The main work of the paper and research inno-vation are summarized as follows:1.We proposed a Deep Semantic Match Model(DSMM)for entity linking by using the structural information in knowledge graph.The traditional studies on entity linking mainly focus on designing handcrafted features for the entities and entity mentions using their contextual texts to perform entity linking.This kind of features can not represent the internal mean-ings of words or entities,lack adaptability to different scenarios and re-quire tedious feature engineering and expensive computation.In response to these problems,the DSMM algorithm uses character granularity,word granularity recurrent neural network and knowledge representation learn-ing method to extract the features of surface form and contextual semantic.The experimental results show that the DSMM algorithm achieves 94.3%accuracy of entity linking on the CoNLL benchmark dataset,which has been improved to a certain extent compared with 92.6%of the mainstream baseline model.2.In order to further verify the effect of DSMM algorithm in practice,we designed and implemented a KB-QA system.The system decomposes the task of KB-QA into two subtasks:getting the subject entity of the ques-tion through the DSMM entity linking method,and then obtaining the re-lation using a relation extractor,by means of which the system effectively reduces the number of candidate(subject entity,relation)pairs without affecting the recall.