Research On Event Coreference Resolution From Raw Texts

Event coreference resolution is an important task in the field of information extraction.This task focuses on identifying whether different event mentions refer to the same event in the real world and linking coreferential event mentions to an event chain.Researches on event coreference resolution can be divided into within-document and cross-document event coreference resolution.This dissertation focuses on within-document event coreference resolution from raw texts.Unlike early event coreference resolution methods which are often based on labeled event mentions,event coreference resolution from raw texts(without labeled event information)is more challenging but practical.The main contents of this dissertation are as follows:(1)Multitask Learning for Event Coreference Resolution from Raw TextsTo address the issues of the error propagation and the difficulty of using event type information under pipeline method,this dissertation proposes a multitask learning method for event coreference resolution,which combines event subtype re-detection task and event coreference resolution task under multitask framework.The method can update and use the event subtype information when identifying event coreference and suppress error propagation by a correction mechanism.Experimental results on both KBP 2016 and KBP 2017 show that our model outperforms the state-of-the-art baselines.(2)Global Context for Event Coreference Resolution from Raw TextsTo address the issues of missing context information and high time complexity under joint method,this dissertation proposes a global context based event coreference resolution model which combines the advantages of pipeline methods and joint methods.This method uses Longformer to encode the whole document and model the semantic features directly.Meanwhile,this dissertation introduces the idea of contrastive learning into event coreference resolution and uses loss-function constraints instead of manual rules to generate semantic representations that are conducive to clustering.Experimental results on both KBP 2016 and KBP 2017 show that our model outperforms the state-of-the-art baselines.(3)Incorporating Paragraph Information and Generation Model to Event Coreference Resolution from Raw TextsTo address the issues that existing researches often ignore the paragraph structure information,this dissertation proposes an event coreference resolution method which incorporates paragraph information and generation model.This dissertation introduces macro discourse structure to help event coreference resolution,which can take advantage of structure and relations between paragraphs.Moreover,on top of the joint method,this dissertation introduces an encoder-decoder style generation model to further boost the performance.Experimental results on both KBP 2016 and KBP 2017 show that our model outperforms the state-of-theart baselines.Through the above three methods,this dissertation addresses the problems under the previous pipeline method and joint method,and introduces the generation method into event coreference resolution,which improves the performance of this task and provides reference for future research.