Research On Document Level Financial Event Extraction Method With Integrating Multi-level Sematic Representation

Event extraction means that via deep processing of unstructured text,the computer extracts its corresponding constituent elements according to predefined event types,and finally obtains structured event information.By the work of event extraction,complex and lengthy text can be deeply understood.Therefore,in recent years,event extraction has attracted more and more attention in the field of natural language processing,and has been widely used in many fields including finance and medical treatment.For vertical fields such as finance,financial text is generally presented in the form of a document,and its event arguments are scattered in multiple sentences of the document,and a document may contain multiple events.To this end,this paper proposes two event extraction methods that capture the deep semantic expression of context,and verifies its effect on document-level financial event extraction.This paper proposes an event extraction method that integrates external multi-level semantic representation,and stores sentence context information related to the reasoning entity through an external memory mechanism,enriching its semantic expression,so as to carry out event reasoning better.This method first stores the modeled sentence representation of the document in memory at one time through addressing operations,then calculates the similarity between the inference entity and the sentence information in the memory,integrates the relevant sentence information into the representation of the inference entity.Finally,a global reasoning algorithm is used for the entity to obtain a structured representation of the corresponding financial event type.In order to cope with the challenge of arguments-scattering in document-level event extraction,this paper proposes an event extraction model that integrates multi-level semantic representation of attention.First,introduce a pre-trained language model based on the multi-head attention mechanism,capture the semantic features of different levels of the text through the attention mechanism,thereby more effectively extract the argument entities in the financial text,and then use the attention to merge the contextual information of the entities to extract the reasoning argument entity performs global event reasoning,and fill the predefined event table to obtain a structured representation of the event.Experiments on the two proposed models on the public document-level financial event extraction data set,and compared with the current mainstream financial event extraction methods,the experimental results show that the external memory mechanism can capture entity-related context information to a certain extent.But the overall performance improvement is small.The F1 value of the model incorporating the multi-layer semantic representation of attention reached 83.2% and 68.5% respectively on single event and multiple event types,which exceeded the best model Doc2 EDAG in the comparison method(its F1 was 82.3%,67.3%).At the same time,without distinguishing between single events and multiple events,F1 reached 77.6%,which was 1.3 percentage points higher than the optimal model for comparison.