Research Of Coreference Resolution Model Base On Mention Detection Optimization And Capsule Network

In a text,an entity can have many different forms of mentions,and the relationship between these mentions is called coreference relationship.The goal of coreference resolution task is to integrate the objects with coreference in the text,so as to eliminate the ambiguity of coreference and enhance the machine's understanding of natural language.Mention detection and coreference calculation are the key steps of coreference resolution,and the current end-to-end model implementation needs to be further optimized.In terms of mention detection,the current model has a maximum length limitation,and its mention score function can be more precise.Because of the difference between the mention detection task and the coreference resolution task,it is not the best choice to train the mention detection module only by coreference resolution task.In the terms of coreferential relationship calculation,the model does not consider the contextual relationship between mentions.Besides,the higher-order coreference resolution model can calculate the relationship more accurately,and current higher-order algorithms all rely on the coreference score,but the wrong score may affect the performance of algorithms.Based on the above problems,this paper conducts the following research:1)Proposed the SEUL(Start-End mention score and Unlimited Length)optimization scheme of mention detection.In this scheme,a mention detection strategy without length lim-itation is adopted,and the mention score calculation of start and end vectors is added,which improves the recall of mentions and coreference resolution.In addition,the pre-training of the mention detection module has been carried out to strengthen the ability of the module and further improve the overall performance of coreference resolution.2)Constructed a coreference score based on context matching.Considering that the an-tecedent(referred object)can replace the mention which has the co-referential relationship with it in the text,this paper calculates the matching degree between the antecedent and mention's context,which can represent the coreference relationship at this level and provide a new per-spective for the judgment of coreference.3)Proposed a higher-order coreferential resolution algorithm based on capsule network,which is named Capsule Merging.The Capsule Merging algorithm don't need the coreference score,only through the dynamic routing of the capsule network,the local feature aggregation can be realized and the corresponding global feature can be extracted.The global feature is introduced into the calculation of coreference resolution,and then the higher-order coreference resolution model is implemented.Finally,this paper uses the Onto Notes English data set as the experimental data,and con-ducts multiple sets of comparative experiments on the Bert-based end-to-end coreference res-olution model,which proves that the above optimization schemes and the Capsule Merging algorithm can effectively improve the performance of coreference resolution.Through the fu-sion experiment,the work of this paper is comprehensively verified,and its overall performance is better than other comparison models.