Research On Event Extraction Based On Deep Learning

Event extraction refers to detecting and extracting participants,time,location and other information which users are interested in from unstructured texts containing event information automatically.It is the basis of other high-level Natural Language Processing tasks,and has been widely used in the fields of knowledge graphs,text retrieval,recommendation systems,and so on.In recent years,deep learning models performed well in Natural Language Processing,and the theoretical research of feature extraction technology based on deep learning became more mature.In this thesis,the research on use of deep learning model to improve the performance of the event extraction task has been carried out.The main research contents are listed as follows:(1)To solve the problem that the polysemy of a word in trigger extraction task is not well represented and wordfeatures are not extracted sufficiently in traditional methods,the event extraction method based on pre-training model and multi-feature fusion is proposed in this thesis.This process is divided into two steps: event trigger extraction and event argument extraction.Firstly,the BERT pre-training model is used for text representation.Secondly,the lexical features extracted by convolutional network and the sentence level features extracted by graph convolutional network are fused to obtain event features.Finally,the event features is fed into the classifier to extract the event triggers,and conditional random field model is adopted to perform constraint of trigger label sequences to realize extraction of trigger accurately.In the event argument extraction task,the extraction process is roughly the same as the trigger extraction.The only difference is thatthe event type features represented by the trigger word are added when fusing the lexical features and the sentence features.The experiments based on the ACE 2005 corpus demonstrate that,F1 scores can reach71.1% in event trigger classification task and 56.4% in event argument classification task under test datasets.(2)From the perspective of optimizing the event extraction process,the idea of event extraction without trigger words is proposed,where detection of the position of event trigger words is discarded.The event type label and event argument role label are recombined into new labels using the combination label strategy,and Bi LSTM model with Attention mechanism is used to perform sequence labeling of new labels on event sentences to achieve the purpose of event type classification and event element extraction.Firstly,the BERT pre-training model is used for text representation.Secondly,the Bi LSTM model is used to encode the event sequence in both directions,and the attention mechanism is introduced to calculate the contribution of the sequence words to the event features.Finally,the task of event extraction is realized by a single model,and the effectiveness of the proposed method is demonstrated using the Du EE Chinese datasets.