Research And Application Of Text Event Extraction Methods

The goal of text event extraction is to detect the event instance in the text and,if it exists,to identify the event type and all its arguments and attributes.In recent years,although the research of text event extraction methods has received a lot of attention,compared with entity extraction and relational extraction,the existing methods still have the problems of complex model and low extraction accuracy.In this paper,starting from the feature coding and model structure,the pipelined method is adopted,and the deep pre-training language model,graph convolution network and attention mechanism are introduced to improve the existing Improve the existing text event extraction model.The main research work is as follows:1.In order to solve the problem that the traditional context-free word vector does not take into account the overall meaning of the sentence and can not directly obtain the sentence-level features,an event detection method based on BERT and graph convolution network is proposed.In this method,BERT word vector is introduced to enrich the feature representation of sentences and words,and syntactic structure is introduced to capture longdistance dependencies and identify event triggers.Experimental verification is carried out on the benchmark data set,and compared with other representative trigger word extraction methods,the experimental results show that the performance of the proposed method is better than other methods on the experimental data set.The F1 scores on the two major tasks of event detection are 77.6% and 74.2% respectively,which verifies the effectiveness of the proposed method.2.In view of the fact that the existing event extraction methods are easy to ignore the influence of trigger words and other event arguments on candidate event arguments in the argument extraction stage,an event argument extraction method based on Bi-GRU and Multi-Attention mechanism is proposed.This method combines BERT word vector and other basic featuresin feature selection.In the construction of the model,the Bi-GRU network is used to encode the features,and the multi-attention mechanism can divide the sentence into three parts to calculate the attention weight.The experimental results show that the performance of the method proposed in this paper is better than the existing methods in performing event argument extraction tasks on the experimental data set,and the F1 scores of argument recognition and argument role classification tasks are 69.2% and 61.6%,respectively.The combination of deep language model,multi-attention mechanism and BiGRU helps to improve the quality of feature embedding.3.Based on the above research results,the two proposed methods are applied to the news event extraction field,and a news event extraction prototype system is designed and implemented.In this system,the event extraction module completes the extraction in the way of pipeline.Trigger word extraction and argument extraction respectively adopt the event trigger word extraction method based on BERT and graph convolution network and the event argument extraction method based on Bi-GRU and improved attention mechanism.The system provides a visual Web interface,which helps to improve the work efficiency of enterprises.