Research On Key Issues Of Temporal Relation Between Events

In texts,there are various relations between events,such as coreferential relation,temporal relation,causal relation,and so on.Event temporal relation identification aims to detect the existence of temporal relation between events and classify them into different categories.This task plays an important role in natural language understanding and impacts much on many NLP applications,such as text summarization,story timeline construction,automatic question and answering.Previous studies on event temporal relation identification mainly focus on traditional feature engineering.Considering automation and domain applicability,in this paper,we propose a neural model combining various information to improve the performance of event temporal relation identification.In particular,our work includes:(1)Event temporal relation identification based on GCNAt present,identifying event temporal relation via event shortest dependency path has become the mainstream.This method can remove redundant information and extract key information.However,it also causes the problem of information loss.Moreover,current study focuses on semantic processing,but ignores the syntactic struct-ures.To issue these problems,this paper proposes a model framework combining Long and Short Time Memory Network(LSTM)and Graph Convolutional Network(GCN).Dependent path only considers semantic information in local contexts of two events.The method uses the original sentence as input,combination of the semantic and syntactic information can automatically extract features from word sequences and dependency syntax,which can learn more event information and entity information and get richer representations of events.Experimental results show that the proposed model can effectively improve the identification performance.(2)Event temporal relation identification based on dependency and discourse relationExisting studies only consider the relevant information of two events at the sentence level and encoding the local information of their context,while the relations between the events from the perspective of discourse has been ignored.In view of this problem,this paper proposes a method of event temporal relation identification which combines the sentence-level dependency relation with the text-level discourse relation.In this method,the relation between events is represented in three parts:the event sentence;the dependent path information of the event sentences;the discourse relation information of the basic text unit between the elementary discourse units(EDUs).The neural network model based on this representation system can capture event information from two different levels,which significantly improves the performance of temporal relation identification(3)Event temporal relation identification based on multi-task learningWe propose a multi-task learing approach to address the data imbalance problem caused by a large number of VAGUE type relations.In particular,a binary classification auxiliary task is introduced to determine whether the temporal relation between events is VAGUE.At the same time,focal loss are used as the objective function to better focus on the instances difficult to classify.Experiment results show that the proposed approach can effectively alleviate the data imbalance problem.