| The development of information technology has led to the creation of a wide variety and size of data in the network.These data carry a lot of very important and interesting textual information in the field of finance in the form of electronic texts.Most of these textual information is disorganized and unstructured,which is not convenient for people to view and use.Faced with the huge amount of unstructured text information,how to structure it so as to meet people’s needs for quick understanding of the basic situation in the financial field has become an important aspect of online financial analysis.However,the current event extraction suffers from difficulties such as sparse corpus resources,overlapping event information and poor generalizability,which makes event extraction,especially chinese financial event extraction,a great challenge.Compared with general events,chinese financial events have the problem of trigger word overlap in addition to the element overlap problem and element role overlap problem.Therefore,to address the information overlap problem in chinese financial events,this paper proposes an attention mechanism-based chinese financial event extraction method.In order to avoid error propagation to a certain extent and flexibly transform the event extraction task into a sequence labeling task,the whole event extraction model is split into an event type detection model combining the attention mechanism and an event extraction model incorporating event type information.The main research elements are as follows:(1)Proposed an attention mechanism-based event type detection model: A variant of the pre-training model BERT,BERT-wwm,is used to encode word embeddings of text to form a text representation,which avoids the chinese word separation problem and the problem of multiple meanings of a word,and trains the chinese vocabulary so that the semantic information expressed by the chinese words can be fully learned.The attention mechanism is flexibly used to focus on latent key semantic information,weakening unconnected or less important features and strengthening key trigger word features,generating text representations with trigger word features.By calculating the similarity between the text representation with trigger word features and the event type,the event type is detected and event classification is performed,so that the overlapping trigger words can be extracted separately according to the different event types.(2)Proposing an event extraction model that fuses event type information: Because there is a certain connection between event types and event elements,this paper uses conditional layer normalization to fuse the event types obtained from the event type detection model with the contextual representation of the pre-training model BERT-wwm as the next text representation.The event trigger word features and event element features that have the most important impact on event extraction are then focused on through a self-attention mechanism,and finally event extraction is performed by designing a trigger word extractor for specific types and an element extractor for specific roles,so that different types of trigger words and elements for different roles can be extracted separately.(3)Experimental comparisons were conducted with five models on the publicly released FewFC financial event dataset to compare their Precision,Recall and F1 values.In this paper,the reliability of the event extraction model is verified by conducting separate experiments on normal and overlapping sentences,plus a final ablation experiment.In comparison with other models,the extraction results of the model proposed in this paper for event trigger word extraction and event element extraction show that the model achieves good results in dealing with overlapping information,proving the feasibility and effectiveness of the model in this paper. |
