Research Of Bidirectional Context-based Collective Entity Linking

The rapid development of the information age has been accompanied by a sharp increase in unstructured data of document,but the ambiguity and diversity of natural language increases the difficulty of text processing.Entity linking technology links the entity in the document to the unambiguous entity in the knowledge base,which can provide important support for the subsequent processing of information.Compared with traditional entity linking technology which ignores the correlation between all entities in the same article,the collective entity linking technology uses such information to improve the accuracy of the algorithm.But it also has following problems in actual use:First,the algorithm should balance accuracy and computational complexity;Second,the model generalization ability must be strong that it can be applied in actual scenarios with data from deferent domains.In response to these problems,this article has conducted an in-depth study of the collective entity linking model,and designed a collective entity linking model based on bidirectional context(BCEL).For an mention in the document,the model collects bidirectional context information composed of the entity set before the position in the document and the entity set after the position,and learns the relationship between the entities and determines which is the object through three combinations.Compared with graph-based collective model,which running time cost increases exponentially with the number of entities,BCEL model has the advantage of linearly running cost increasing with the number of entities.Compared with the model based on unidirectional accumulation of context information between entities,BCEL has the information of all entities in the same article,so the generalization ability of the model is stronger.The model was trained and tested on a public data set named AIDA CoNLL-YAGO.The results show that the model has an accuracy of 95.11%and an increase of 0.47%on the in-domain test set AIDA-B.The model shows that running cost is a linear growth with the increase in the number of linked entities features,which are good for entity linking process in long texts.In addition,although the performance of the model on five cross-domain datasets is lower than that of graph-based models,it is superior to the unidirectional-context collective model.In general,the BCEL model balances accuracy and time cost better,improves generalization ability,and has better application scenarios in actual use.