Research On Entity Linking Based On Graph Model And Semantic Representation

In recent years,Entity Linking is one of the most important research areas which aroused tremendous interests from both industrial and academic communities.This task has a huge role in promoting potential applications like knowledge base population,question answering,information extraction and information retrieval.Generally speaking,a typical entity linking system consists of the following three modules,candidate entity generation,candidate entity disambiguation and unlinkable mention detection.The candidate entity disambiguation module is a key component for the entity linking system.Candidate entity disambiguation module can be broadly divided into probability generation model based approaches,topic model based approaches,graph based approaches and so on.Graph based approaches which models document-level topical coherence of candidate entities have greater advantage than approaches which only models the context of single entity.However,graph based approaches do not make full use of unambiguous entities.With the increasing number of unambiguous entities,semantic information of entity referent graph is not enriched.To solve the above problems,we propose an entity linking algorithm based on graph model and semantic representation which can play the core role of unambiguous entities by constructing entity referent graph.Specifically,the main contributions of this thesis are as follows:1)Propose a referent graph construction method.The entity referent graph contains all mentions and candidate entities in the same document which constitutes a basis for the entity disambiguation algorithm based on dynamic PageRank in referent graph.The number of vertices,the number of edges,and the weight of edges are dynamically changed with the increasing number of unambiguous entities.2)Propose a dynamic entity disambiguation algorithm based on PageRank in referent graph.The disambiguation algorithm takes into account the semantic consistency between candidate entities and the core role of unambiguous candidate entities.The algorithm picks out the ambiguous candidate entity which gains the highest score as the target entity each round and gradually completes the selection process for candidate entities disambiguation.3)Conduct multiple experiments in three publicly available datasets to evaluate three algorithms in terms of overall accuracy,linkable entity disambiguation and unlinkable mention detection.The experimental results show that the proposed algorithm has the advantages of capturing rich semantic representations from knowledge base and achieve the best performance over the comparative algorithms.