Reserch On Event Coreference Resolution Based On Deep Learning

As an important sub-task of information extraction,Event coreference resolution,is to correctly identify the coreference relationship between even mentions in a document,and has certain guiding significance for topic detection,text summarization,machine translation and other natural language processing(NLP)tasks.Aiming at event coreference resolution task,this paper constructs a Uyghur event coreference resolution framework based on deep learning models,and uses the deep semantic information implicit in the text and the event characteristics to identify the event coreference in a document.Which facilitates machine understanding and supports NLP tasks such as machine translation and information extraction.At present,the research on event coreference resolution mainly focuses on English and Chinese,and there is no research on the Uyghur language.At the same time,in the process of studying event coreference resolution,we find that reasonable use of deep semantic information and good event representation can effectively promote the research of event coreference resolution.Therefore,for event coreference resolution task,the following two researches have been done:(1)Without feature engineering and sophisticated linguistic constraints,bidirectional long short term memory(Bi-LSTM)is used to capture hidden semantic features of event sentences,thus,semantic modeling of event sentences and mining deep semantic information.The neural tensor network(NTN)is used to find the semantic interactions between event pairs.At the same time,according to the characteristics of event expression,a topic embedding based on event trigger words is introduced,thus,the final multi-level and powerful event representation is obtained by the fusion of multi-level semantic representation,and the F value reaches 74.20%(2)With the effective application of gating mechanism in Bi-LSTM and the powerful feature extraction ability of convolutional neural network(CNN),a framework of event coreference resolution based on gated convolutional neural network(GCNN)is constructed.The word embedding is used as the event semantic carrier,and the event semantic information extracted from event sentences by the GCNN is taken as the semantic feature.In addition,according to the characteristics of event coreference,a set of hand-crafted features is constructed,including event attribute features,event trigger lexical features and event distance features.Finally,the semantic features and extracted manual features are fused as input of the classifier to effectively complete event coreference resolution task,with F value reaching77.97%.