Research On The Key Issues Of English Event Extraction

Event extraction is the starting point for information extraction which based on events,with high theoretical and application value.At present,related research mainly has the following problems.Firstly,it pays too much attention to the sentence,but ignores the document-level information.Secondly,it is difficult to detect all events by facing multiple events existing in the same sentence.Finally,it pays attention to the pipeline model of event recognition and argument identification,which leads to error propagation.In view of the above problems,our study includes the following three aspects:Firstly,towards for better event detection,this paper proposes a sequence-to-sequence approach,which uses attention mechanism to construct a unified model combing local characters,words,entities,and global events co-occurrence in one document.Experimental results on LDC2017E02 corpus show that this method can effectively improve the performance of event recognition.Secondly,this paper proposes a method of event detection based on ELMo and graph convolutional networks,by using dependency analysis to establish the connection of many kinds of events,and using graph convolutional network to learn the dependency analysis.Experimental results show that this method can further improve the performance of event recognition.Finally,due to the size of the corpus,this paper proposes a joint model which combines local features.The method learns the connection between events through a graph convolutional network,and learns the semantic relationship between event arguments and event types through memory units.The experimental results on the TAC KBP 2016 testing show that the joint events extraction can effectively improve the performance of event recognition and argument identification.