Financial Text Reading Comprehension Based On LDA And BERT

In the financial field,the financial reports publicly released by enterprises often contain important investment information such as the basic information of the enterprise,business scope,operation and financial status.Investors often take this as the basis for investment decision-making when they conduct investment activities.Therefore,financial texts such as financial reports are an important reference for investors to make investment decisions during investment activities.It is an important research direction that how to make machines intelligently read and understand financial texts through machine learning methods,so as to provide convenience for investors.The task of machine reading comprehension in the field of finance is to extract answers from relevant texts by machine learning to answer a given question.The previously proposed machine reading comprehension methods mostly use sequential structure model to obtain the interactive information between texts,which can not fully obtain the interactive information between texts,and the processing methods of multi document long text data in the financial field are not common.The financial text machine reading comprehension model proposed in this paper can better obtain the interactive information between questions and texts,answer the questions raised by users,reduce the time for users to screen information,and assist investors in investment decision-making by reading relevant financial texts.The main contributions of this paper are as follows:·By introducing the recall stage and adopting the parallelizable self attention mechanism,this paper greatly improves the training speed of the model.By introducing the recall phase,the proposed model dynamically segments the long text data of multiple documents,and carries out rough recall to reduce the data entering the subsequent reading and understanding module.Moreover,we abandon the traditional network structures such as RNN and CNN and adopt a parallelizable attention mechanism,which greatly reduces the training and reasoning time of the model.·This paper obtains sufficient text interaction information by superimposing multilayer self attention layers,and improves the reading comprehension ability of the model by designing auxiliary tasks.Through the superposition of multi-layer attention mechanism,the deep semantic interaction information of text and problem is obtained,and appropriate auxiliary tasks are introduced,so that the model can learn the information related to reading comprehension task.·This paper constructs a machine reading and understanding data set of 200000 financial texts.This paper constructs a machine reading comprehension dataset containing 200000 records of data.Each data contains three fields: question,answer and relevant text.The dataset can be used for subsequent research.We have conducted relevant research on machine reading comprehension of financial texts.A multi-stage,multi task learning model based on self attention mechanism is constructed,and experiments show the effectiveness of our model.Our model provides a new idea for the machine reading and understanding task of multi document long text in the financial field.