Entity Event Extraction Based On Representation Enhancement

Information extraction is an essential research direction in natural language processing,and its goal is to extract structured knowledge from unstructured text.Information extraction includes three tasks: entity recognition,relation extraction and event extraction.This paper focuses on entity event extraction tasks.The event extraction includes trigger recognition and argument role classification.At present,most studies only focus on one of the tasks,and a few studies have both.Arguments are usually entities in sentences,therefore,most researches usually extract events in condition of given entities.Based on the actual scene,this paper first recognizes the entities in the sentence,and then extracts events based on the results of entity recognition.The difficulty of entity event extraction task is how to efficiently extract multiple events in the same sentence.The input and output of the single trigger recognition task and the single argument role classification task are relatively regular.The event extraction task is relatively complicated,especially when facing multiple events.At present,most researches usually extract events in condition of given entities,but they don't make full use of the known entity information,especially the span information.In the existing research,when dealing with different trigger-argument pairs,the characteristics between them are often ignored.In addition,some studies have pointed out that the type information of entity and trigger plays an important role in the argument role classification task.However,the current research only deals with type information implicitly,and the disadvantage is that the information which can be introduced is very limited.Based on the above observations,this paper proposes a representation enhancement method.When classifying the argument's role,for each trigger-argument pair,the sentence encoder can focus on the trigger and entity information through explicitly adding token markers to point out the spans of entity and trigger.In addition,the type information recognized in the pre-stage is explicitly added to the token markers,which further enhances the representation of the corresponding trigger and argument.Due to limited resources and in order to introduce entity information in the trigger recognition task,this paper using the pipeline way,which to recognize entities firstly and then to extract events,and we purpose two efficient approximations to our approach.Experimental results on commonly used data sets for information extraction show that the method proposed in this paper has obvious effects.One of the shortcomings of the pipeline method is error propagation,what's more,the way of explicitly adding token markers will cause the input end of each subtask to be different,which will affect the performance of the model.Therefore,this paper proposes a joint method to share sentence encoder among various subtasks.In order to ensure the same input format,the token markers of entity and trigger are no longer explicitly added,and the interaction information between trigger and argument is modeled.On the one hand,the information of the argument list related to the same trigger is used to enhance the trigger representation,on the other hand,the correlation characteristics between arguments can be extracted to enhance argument representation.This paper has verified the effectiveness of the above methods through experiments.