Research And Implementation Of Event Extraction For Financial Announcement

The goal of event extraction research for financial announcements is to extract structured events of interest to investors from announcements.The financial announcement refers to the current operating status of a listed company announced by itself to the market and investors.The announcement will exert a great influence on investors if it contains events like equity repurchase,pledge,freeze,increase or decrease of holdings.At present,the event extraction research for financial announcements is still in the preliminary stage,and the field-related training data is scarce.However,the deep learning model depends on the scale and quality of the training data and the cost of manually data-tagging too high.Therefore,the use of knowledge base and distant supervision to make up for the lack of training data has become a focus of the research.The existing work is to construct a financial announcement event knowledge base through domain experts,and then mark the structure events in the knowledge base back into the announcement data to obtain the training corpus required by the model.The event extraction task for financial announcements typically faces the following two problems:(1)Event elements are distributed in multiple sentences;(2)An announcement contains multiple events.None of the existing research methods of event extraction in the financial field can effectively solve the two problems.This thesis proposes a TDJEE model,a document-level joint event extraction model based on Transformer,to solve the abovementioned problems and propose an approach which can build a knowledge base automatically so as to realize the automatic labeling required by the model.The main contributions of this thesis are as follows:(1)A Fonduer-based approach to build the event knowledge base: the limitation of traditional semi-automatic approaches to build a knowledge base is analyzed,and the method of constructing an event knowledge base based on the Fonduer system combining machine learning,weakly supervised learning,rules and dictionaries is proposed.The results show that the event knowledge base can be constructed automatically with the Fonduer system as a basis.(2)A labeling method of training data based on the distant supervision: The assumptions made during data-labeling with traditional the distant supervision method is too positive,and a lot of noise will be introduced and semantic drift will occur.Therefore,this thesis improves the way of aligning the knowledge base with unstructured text in the distant supervision method,and matches the event in the knowledge base with the text through directional links,so as to obtain the training corpus required by the model.(3)A document-level joint event extraction model based on Transformer: The TDJEE model is implemented based on the Joint Model approach and consists of two sub-models,i.e.event element extraction model and event type detection model.The models share the same input layer and BERT layer and are optimized by an objective function.The experimental results show that the TDJEE model can effectively solve the problems that event elements are distributed across sentences and an announcement contains multiple events.Based on the above work,in order to apply the event extraction model in the financial field to real-life scenarios,this thesis builds a financial announcement event extraction service based on the Django framework.The results show that:(1)The F1 value on the test data of the event knowledge base built based on Fonduer system reached a minimum of 77.4% and a maximum of 89.9%,which reveals that there is a big difference among different types of event knowledge bases in construction performance;(2)The TDJEE model solves the problem that event elements are distributed across sentences and one announcement contains multiple events,and the F1 value of the overall model reaches 76.7%.