Research On Deep Learning Based Event Extraction Method

Event extraction is an important subtask of information extraction,a basic technology for intelligent processing of textual information,and has a wide range of applications in the fields of information retrieval,intelligent Q&A and knowledge graph research.At this stage,the research on event extraction methods is mainly focused on judicial field,news field,breaking news field and financial field except for the general field.With the rise of deep learning in the field of artificial intelligence,researchers have found that deep learning-based methods have advantages over machine learning methods in deep-level feature extraction for event extraction.Deep learning-based event extraction is usually converted into classification tasks,annotation tasks,and machine reading comprehension tasks,and the mainstream research approach is to convert event extraction into annotation tasks.The event extraction task itself suffers from overlapping elements,statement ambiguity,inadequate feature fusion,and error propagation error accumulation.The data in the financial field is characterized by information redundancy and discourse order complexity,relatively specialized requirements for event types,and high requirements for event element recognition accuracy.In this paper,we focus on the event extraction method based on deep learning in the financial field,and combine the characteristics of financial field data and the problems faced by the event extraction task to conduct relevant method research.To address the problems of statement ambiguity,element overlap and error propagation in event extraction,this paper proposes a joint event extraction model based on pointer annotation.The coding layer of the model uses the BERT pre-training model to encode the financial text utterances,and the semantic information of the sentences is obtained and incorporated into the event type information for semantic enhancement.The fused sentence information is input to the self-attention layer of the trigger word decoding layer for feature extraction,and finally the trigger word is decoded by the pointer labeling method.Then,the fused sentence information is input to the self-attentive layer of the argument decoding layer for feature extraction,and the event argument is decoded by pointer annotation with the relevant information of the trigger word.Adversarial training is added to enhance robustness and generalization during model training.For the problem of inadequate event feature extraction and the complex discourse order of financial data,this paper designs a joint event extraction model based on PERT pre-training based on the joint event extraction model based on pointer labeling.Financial text sentences are input to the coding layer of the PERT pre-training model to get sentence vectors,fuse event type information,and input the fused sentence information to the self-attention layer of the trigger word decoding layer and the bi-directional long and short term memory network layer for extracting word-to-word features between sentences and features of sentence context,and finally decode the trigger words by the pointer annotation method,after the fused event type and The sentence information fused with event type and trigger word information is input to the self-attentive layer and the bi-directional long-and short-term memory network layer in the argument decoding layer for extracting word-to-word features and sentence context features,and finally decoding the event argument by adding the related information of the trigger word with the pointer annotation method.The two joint event extraction models in this paper are experimentally validated on the Chinese financial dataset Few FC,and the experimental results show that both event extraction models proposed in this paper can effectively improve the results of event extraction,and the test results outperform the results of other comparative experiments,proving the effectiveness and superiority of the models.