Research Of Collective Entity Linking Based On Entity Differences

Massive unstructured text data on the Internet provides a data basis for natural language processing,but the semantic ambiguity of natural language increases the difficulty of text processing.The entity linking technology eliminates the ambiguity of entities in the text by linking the entity mentions in the text to the unambiguous entities in the knowledge base,and provides important support for the subsequent tasks of natural language processing.In the past two years,the entity linking system of chain processing has received extensive attention from scholars at home and abroad because of its simple model structure and high accuracy.After investigation,we found that most of the existing models focus on how to obtain more information in the text and how to make better joint links.Almost no models pay attention to the interactive information between the mention candidate entities.When using linked entities to help follow-up mentions disambiguation,the impact of incorrectly linked entities is also ignored.In response to these two issues,this thesis has improved on the basis of the Dynamic Context Augmentation(DCA)model,and proposed Focus on Entity Differences and Incorrect Entity Effects for Entity Linking(FDIE).The FDIE model believes that the difference between candidate entities is an important feature that distinguishes these candidate entities,and proposes the concept of difference similarity,using the performance of the mention in the difference characteristics between the candidate entities as the difference similarity between the mention and the candidate entity.At the same time,convolutional neural networks and self-attention mechanisms are used to extract more effective features of mentions and candidate entities.Aiming at the influence of the wrongly linked entities,the model deals with the two aspects of reducing the number of wrong entities and reducing the interference of the wrong entities on subsequent mention disambiguation.The model believes that the amount of pre-information required for each mention when linking is different,so each time it is linked,the mention with the highest "confidence" among the unlinked mentions under the current amount of information is selected for linking,thereby reducing the number of linked entities In the case of the wrong entities in the linked entities,the model proposes the concept of contrast consistency,and uses the linked mentions as the contrast of the linked entities to calculate the consistency,thereby reducing the interference of the wrong entities.This article conducted comparative experiments and ablation experiments on six public data sets to evaluate the capabilities of the model.Compared with the benchmark model DCA,the accuracy of the FDIE model on these six public data sets has been improved,and on three of the data sets,it has also achieved higher accuracy than other entity link models in recent years.