Research On Sentiment Classification Of Financial News Base On Deep Learning

Nowadays,people,s production and lifestyle are deeply affected by countless Internet news.Especially in the financial field,financial news plays a vital role.Financial news generally focuses on reporting events related to listed companies,and also discloses information on listed company announcements,financial statements,and other financial related information.It is of great value to listed companies as well as investors and financial industry practitioners.By classifying the sentiment of financial news,we can grasp the trend of financial markets better,and also provide a basis of investors to make investment decisions.Based on this,this paper carries out the following research.Firstly,in order to improve the efficiency of data collection,this paper innovatively designed a method for extracting the text of the news webpage based on XGBoost.The accuracy of content extraction on 10 well-known financial news portals reached 97.63%,greatly reducing the cost of web crawler's Development.On this basis,this paper builds a financial news corpus.Aiming at the problem of insufficient financial news corpus for research,this paper proposes to collect the "positive and negative news of listed companies"published by a well-known financial information provider on the snowball website as a corpus seed set through the Internet crawling method.In view of the limited number of collected seed sets,the method of searching online is used to expand the news,and finally a sentiment classification dataset with 17149 financial news is constructed and publicy released.Secondly,the deep learning and attention mechanism is introduced into the sentiment classification of financial news.After fully considering the advantages and disadvantages of CNN model,RNN model and attention mechanisms,and combined with the characteristics of financial news texts,a two-channel LSTM-CNN model based on attention mechanism is proposed.The model is used to train and predict on the financial news sentiment classification dataset constructed above.Compared with the traditional CNN model,RNN model and bidirectional LSTM model based on attention mechanism,the experimental results show that the proposed method can gain the accuracy score of 96.4%,which fully verified the validity of the model.In addition,this paper also extracts the weight vector from attention layer,and gives the different shades of color according to the number of the weight,so as to visualize the news text.