Research On Event Extraction Method Based On Attention Mechanism

With the arrival of the artificial intelligence era,information sources are becoming more abundant,information transmission rates are more efficient,and new information is generated all the time.How to accurately and quickly mine target information from the massive information of the Internet is getting more and more attention.For information mining of text,the field of Natural Language Processing defines aspects such as Information Extraction.In Information Extraction aspect,Event Extraction tasks extract structured event information from unstructured text,which is the basis for building Event Graph and other advanced applications.This dissertation is mainly for the two subtasks of Event Extraction,namely event trigger detection and recognition(VDR)and event argument role extraction(EAE)tasks.The VDR task detects whether an event trigger is included in a sentence and identifies the type of event triggered by the detected event trigger.The EAE task extracts event arguments from the event sentences and determines their role in the event.This dissertation explores the problem in VDR task and EAE tasks.On the basis of learning existing models in the aspect of Event Extraction,we focus on the intrinsic relationship between event information and the relationship between different types of events and various argument entities.The main contents are as follows:By analyzing the existing VDR methods,it is found that the existing methods mainly extract the lexical level features of the candidate trigger words and their neighbors,ignoring the relationship between the events,which leads to event recognition errors,especially when the trigger words are polysemous.In this dissertation,the bidirectional recurrent neural network is used to integrate the sentence global information into the candidate trigger encoding,and the attention mechanism is used to mine the relationship between the events,so that the event information in the context is applied to the VDR task.Experiments show that this method has a certain effect on solving the polysemy phenomenon.For the EAE task,the current method combines the trigger features with the candidate argument features to classify,ignoring the relationship between the event type and the argument entity type.In a sentence containing multiple events,an entity may play different roles in multiple events,and multiple entities may also play the same role in the same event.Aiming at the argument role extraction of multi-event sentences,this dissertation proposes an event argument extraction structure based on dependency analysis and multi-head supervised attention mechanism.The bidirectional recurrent neural network is used to extract the global information of the sentence,and the relationship between the event type and the entity type is extracted through the multihead supervised attention mechanism.Finally,the classification structure is constructed in combination with the above information and the dependency path between the trigger and the candidate argument.Experiments show that this method improves the performance of event arguments extraction.