Research And Application Of Entity Linking Based On Semantic Enhancement Embedding

As the underlying task of natural language processing,entity linking is intended to map references in unstructured text to the only entity node in the knowledge base.Many upstream tasks such as knowledge graph construction,question answering system and public opinion analysis need to rely on entity linking to resolve the ambiguity in the corpus.The existing entity linking methods encode the description information in the knowledge base to form the corresponding entity embedding.This method will lead to the loss of information outside the context window and increase the difficulty of distinguishing between similar entities.At the same time,the traditional global model will introduce noise information,which will limit the disambiguation ability of the model.The prior probability of some candidate entities is too high,which leads to the irreversible disadvantage of unpopular entities in the disambiguation process.In order to solve the above problems,this thesis studies the embedding of entity linking and the consistency features.The research contents and innovations are as follows:1)Aiming at the problem that similar entities are difficult to distinguish due to insufficient semantic representation in the embedded expression of entity linking,this thesis proposes a method of entity semantic enhancement embedding to supplement and strengthen the difference semantics of similar entities by using metadata in the knowledge base.The experiment shows that our method can improve the effect of the three baseline models to varying degrees,and the maximum gain of micro-F1 index in a single dataset can reach about 0.7 %.2)In view of the misjudgment of some specific entities due to the document-level topic consistency in the global model disambiguation process and the situation that the unpopular entities are difficult to be hit,this thesis focuses the original document-level consistency disambiguation process on the sentence level by introducing fine-grained topic consistency,thus forming a multi-granularity disambiguation model.A threshold filtering mechanism is introduced to reduce the impact of too high a priori probability on model judgment.The experimental results show that this method can improve the performance of the model to a certain extent.Compared with the mainstream models,the best performance of this model in MSNBC dataset is 0.07% higher than that of other models,and it also has a good performance in other similar datasets.3)In order to show the practicability of the method,this thesis combines the above entity linking work with the upstream task knowledge Q&A,and designs and implements a correctable Q&A system,which can collect user feedback to improve the background entity linking model and Q&A model.