Research On Financial Event Extraction Technology Based On Deep Learning

Event extraction is a technical means to obtain structured information,which mainly extracts important information required for tasks from the text in the form of events.In specific implementation,the technology usually uses a predefined event framework to filter the information in the text according to the various components of the event in the framework.In the financial field,event extraction can provide basic structured data for a variety of downstream tasks such as risk assessment and knowledge graph.However,affected by many aspects such as language characteristics,text characteristics,and professional domain knowledge,current financial event extraction research faces two main challenges: 1)Up to now,there is no standard definition of financial event extraction task in the financial field,so there is no unified division of event types,and there is no predefined framework used in event extraction task.2)The texts in the financial field represented by financial news generally have the characteristics of long length,rich semantic information,and high syntactic complexity,which brings semantic difficulty to the event extraction model.Aiming at the above problems,based on financial news,this thesis proposes a financial event extraction method based on deep learning,and finally builds a financial event extraction system,which includes:1.The central event extraction method based on the financial event library,its purpose is to locate the central event sentence from the text and perform event extraction.The method includes three components: a financial event library,a central event sentence determination algorithm,and an event argument extraction model based on sequence labeling.Among them,the financial event library is composed of predefined financial event type tables,financial entity relationship tables,and trigger word dictionaries,and is an important basis for subsequent algorithms and models.The central event sentence determination algorithm locates the central event sentence in the text and provides the trigger word information corresponding to the sentence by using three types of judgment criteria.In the end,the central event extraction method uses the idea of sequence labeling,and uses dynamic word representation models,bidirectional long short-term memory networks,and conditional random fields to identify event arguments.2.Event subject extraction method for multi-event text.As a supplement to the central event extraction method,this method is mainly aimed at multi-event type texts that are difficult to process in the central event extraction task,to avoid the degradation of event extraction.This method realizes efficient extraction of event information through event type reconstruction and event subject extraction.Among them,event type reconstruction uses the calculation of mutual information and information entropy between trigger words and event arguments to obtain the degree of co-occurrence between words and find a reasonable combination of phrases to form a combined event type.At the same time,the method uses the form of question and answer triples to encapsulate the components of the event subject extraction task,and converts the extraction task into a machine reading comprehension task.Finally,through the pretrained language model BERT,the event subject extraction of multi-event text was realized in the form of multiple rounds of question and answer.3.Financial event extraction and display system.This system realizes the central event extraction and event subject extraction of financial news text,and realizes interaction with users through a visual interface.