Research On Event Extraction Based On Representation Learning

Event extraction task aims to extract event information of specified type from a given natural language text and form a structured expression,which is helpful for tasks such as automatic summarization and natural language generation.Event extraction is a challenging task because it is difficult to obtain event semantic information in complex natural texts.Event extraction remains a challenging task because it is difficult to obtain event semantic information in complex natural texts.This dissertation conducts an in-depth study on the event extraction task from the aspect of representation learning,which includes the follow-ing three aspects:(1)At present,most event detection methods focus on capturing long-distance and local information in sequence,ignoring the effect of argument on event detection.To solve this problem,an event detection method based on argument sequence representation is proposed.Firstly,candidate arguments are screened according to the distribution probability of event types.Then,the attention-based recurrent neural network can capture the important clues in the sequence and extract the candidate arguments that contribute most to trigger word ex-traction.Secondly,combining the candidate arguments with the context representation of the trigger words using the Bidirectional Long Short-Term Memory model to extract the trigger words.The experimental results on ACE 2005 show that our method reduces the interference of noise entities on event trigger word extraction and achieves the comparable performance as the current optimal event detection system.(2)Aiming at the lack of practicality of event detection based on annotation entity in-formation,this dissertation proposes an event detection method based on hybrid neural net-work and raw corpus.First,the bidirectional recurrent neural network layer is used to encode the input.Secondly,the obtained entity context representation is further passed to the neural network combined with the self-attention and gated convolution to extract the event trigger word by sharing the information.The experimental results on ACE 2005 show that the method can capture the dependence between entities and trigger words,and effectively im-prove the performance of event detection.(3)Aiming at the problem of dependency between entity,trigger and argument in the current event extraction method,this dissertation proposes a joint extraction method based on graph convolution network for entity,trigger and argument.This dissertation introduces the syntactic short arc to enhance the flow of information between events.Secondly,the graph convolutional network is used to learn the context representation of graph nodes,thus capturing the dependencies between events.Finally,the sequence and syntactic context rep-resentations are fused with sentence-level gated attention mechanism to capture the internal interaction between triggers,arguments and arguments.The experimental results on ACE 2005 show that the proposed method can improve the performance of event extraction.In this dissertation,we propose three effective solutions to the problem of event extraction,and achieve good performance,which will provide a reference for the further research on event extraction.