Research On Event Coreference Resolution Via Neural Network

Event coreference resolution is one of the most important tasks in information extrac-tion.It is helpful to understand text information and discover the relationship between events.It has important application value in the fields of information extraction,intelligent question and answer,machine reading comprehension and so on.At present,the research of corefer-ence resolution mainly focuses on entity coreference resolution,and the research on event homonym resolution is less.Because of the flexible expression of events and the complex relationship between events in the text,the task of event coreference resolution is challeng-ing.Event coreference resolution can be divided into document-level event coreference res-olution and cross-document-level event coreference resolution.This paper mainly focuses on document-level event coreference resolution.The main research contents include the fol-lowing three aspects(1)Aiming at the problem that the existing methods of event coreference resolution cannot well mine the deep semantics of events,this paper proposes an event coreference resolution method based on multi-decomposable attention mechanism.Firstly,according to the linguistic characteristics of coreference events,a variety of semantic features are adopted;secondly,decomposable attention is used to weigh the features,and a large number of irrel-evant features and interference information are filtered while extracting the similarity infor-mation,and the neural network similarity model is used to judge whether the events are coreferential or not.Finally,the global optimization method is used to solve the conflict results in the model.Experiments on ACE 2005 and KBP corpus show that this method is superior to multiple baseline systems in many indicators(2)To solve the problem that most event coreference resolution studies rely too much on annotated information,an event coreference resolution framework is proposed,which consists of three parts:event extraction,realis recognition and event coreference resolution.Secondly,event extraction and related event attributes are extracted from corpus using con-volution neural network and recurrent neural network.Attention mechanism encodes events,extracts deep features in event structure,and obtains the event coreference results.Experi-ments on KBP 2015 and KBP 2016 corpus show that this method is superior to multiple baseline systems in many indicators.(3)In view of the small scale of the existing event coreference corpus,which makes the deep neural network easy to over-fit,this paper proposes an event coreference resolution method based on distant supervision and reinforcement learning.Firstly,the similarity model based on multi-head attention mechanism is established as the basic model of event corefer-ence resolution;secondly,the distant supervision method is used to annotate data in large-scale unsupervised corpus;thirdly,aiming at the high error rate of automatic annotation data,a sample selector is designed by reinforcement learning mechanism,through screening re-mote monitoring.Samples of data sets improve the data quality of remote monitoring data sets.The data expansion method proposed in this paper does not need manual intervention.Experiments on KBP 2015 and KBP 2016 corpus show that this method is superior to mul-tiple baseline systems in many indicators.Aiming at the problems existing in event coreference resolution task,this paper pro-poses three effective methods of event coreference resolution,which greatly improves the performance of event coreference resolution.These methods have good research value in academic research and practical application.