Research On Joint Extraction Of Event Triggers And Arguments

With the amount of information on the Internet increasing dramatically,how to make full use of this massive,diverse,and low-value density information to boost the development of society,economy and technology has become a major challenge for the researchers.In order to solve this problem,Information Extraction technology,which aims to extract structured information that people care about from unstructured text,has become more and more important.Additionally,Event Extraction is one of the key tasks of Information Extraction.Event Extraction aims to extract structured events from unstructured text to use them directly or to serve downstream research and applications.Event Extraction can be further divided into two sub-tasks.The first sub-task is to extract the token in the text that best represents the occurrence of the event.This token is called the trigger.The second subtask is to extract the participants of the event,such as time,people and objects involved in the event.These elements are called arguments.The results of Event Extraction are of great significance for the research on the Text Summary,Question Answering systems and Knowledge Graph.Taking Event Extraction as the research theme,the main research contents of this article are as follows:First of all,this article focuses on the extraction of event triggers,which is also called Event Detection.The existing event detection methods only use the entity type information corresponding to the text as a kind of supplementary information to the text information,which ignore the sequence features that may be contained in the entity type sequence.In this article,an event detection model ETEED is proposed.This model fully learns the sequence features of entity types,and fuses text sequence features with entity type sequence features through a trigger-entity interaction learning module,to realize the extraction of event triggers.Then,this article studies on the second sub-task of the Event Extraction,which is called argument extraction.In most current Event Extraction studies,argument extraction and trigger extraction are divided into two relatively independent extraction stages,which means the trigger extraction model is trained and applied first,then the results of the trigger extraction are used as input of the argument extraction model.This kind of method is called pipeline method.The advantage of this method is that the design and training of the model are relatively simple,but the disadvantage is that the errors of the trigger extraction will be propagated in the argument extraction model.Besides,training in stages will increase the time cost of the research.Therefore,this paper proposes a joint extraction model JPEE,which connects trigger extraction and argument extraction modules by sharing entity type information,and integrates the advantages of pre-trained language model BERT,to achieve the joint extraction of trigger and arguments.Finally,this paper evaluates the Event Detection model ETEED and the joint Event Extraction model PJEE.Extensive experiments conducted on the ACE 2005 public evaluation data set show that the models proposed in this paper can effectively use the entity type information,and outperforms the existing state-of-the-art methods.