Research And Implementation Of Document-level Event Extraction Technology In Financial Field

The amount of text data in the current financial field is increasing,and quickly extracting event-related information in the texts will help people quickly grasp financial information.At present,more financial text data is presented in the form of long document,and the traditional sentence-level event extraction method is poor in the performance of financial documentlevel texts.Based on this situation,in recent years,there has been a lot of research work on document-level event extraction applied to financial texts.The purpose of document-level event extraction is to extract structured and complete document-level event information from the document.It mainly provides the abilitiy of solving scattered event arguments and multiple events problems that cannot be provided by sentence-level event extraction.For the above issues,the following research work has been done on the document-level event extraction task applied to financial texts:1)This thesis proposes an event element extraction algorithm integrating long text information.In the document-level text,the effect of text encoding is very important to improve the performance of event argument extraction.The main problem is how to encode the information of document-level text as a whole and how to fuse the information between sentences.Existing methods do not consider the difference in information content of financial texts with different granularities.On the basis of LSTM and conditional random field,the traditional named entity recognition methods,this thesis fuses single sentence encoding information and multiple sentences encoding information,conducts information fusion under the guidance of the gating mechanism,and guides the model to learn more extensive knowledge.The experimental results show that the proposed algorithm can effectively improve the effect of event argument extraction at the financial document-level text.2)This thesis proposes a document-level text encoding algorithm that integrates entity information.Different from the information encoding of sentence-level text,the length of document-level text is longer,and it cannot be encoded into the model as a whole.However,when we extract events from the whole text,we need to get the information of the whole text,such as the relationship between the front and back of the sentence and the position relationship between the event elements.Existing methods do not consider the dependency information among entity mentions.This thesis designs positional relationship and reference relationship for entity mentions,and the relationship information between entity mentions is fused to guide text encoding.Experimental results show that the proposed algorithm can effectively improve the encoding effect of document-level financial text.3)This thesis implements an experimental system for financial document-level event extraction.Based on the research of the aforementioned algorithms,this thesis designs and implements a document-level financial text event extraction system based on the B/S architecture.The system has the functions of uploading text,extracting event information from the text,and displaying event information in a structured manner.