The Research On Construction And Application Of Sentiment Analysis Model Of Stock News

Text sentiment orientation analysis is one of the hot issues in the field of NLP.The main goal is to classify the sentimental orientation of the text,and divide the text into "positive","neutral" and "negative" according to different emotional orientations.It also called text sentiment classification.This thesis chooses stock news texts as the research texts and uses deep learning model to express the sentiment orientation of stock news texts and explore the emotional tendency contained in it.The main work includes:First,we studied the stock news text and sentiment classification of stock news text,and analyzes its characteristics.Secondly,we studied the chapter-level stock news text sentiment classification task.Aiming at the shortcomings of the traditional model,we introduced the sentiment word coding,sentence position coding and BiDLSTM.Then we proposed a chapter-level news text sentiment classification model based on Bi-DLSTM and hierarchical attention network,and introduced the modules of the model in detail.Then we conducted experiments to verify the effectiveness of the model.Third,we studied the aspect-level news text sentiment classification model.We introduced sentiment word coding and Bi-DLSTM,and constructed an aspect-level news text sentiment classification model based on hierarchical attention network.We introduced each module of the model in detail,and conducted experiments to verify the effectiveness of the model.Finally,we studied the application of stock news text sentiment classification model on stock quantification,and studied the validity of sentiment classification result data.Then we constructed news sentiment factors,and analyzed the effectiveness of the factors.we constructed a stock selection strategy with the single news factor,through the historical backtesting we verified that the factor has a better stock selection ability.This thesis proposes two kinds of text sentiment classification models,and conducts a large number of qualitative and quantitative experiments under different experimental verification targets.Through experiments,the experimental accuracy rate of chapter-level news text sentiment classification model based on Bi-DLSTM and hierarchical attention network reaches82.94%,and the experiment of aspect-level news text sentiment classification model based on hierarchical attention network correct rate reached 81.15%.The experimental results verify the rationality and effectiveness of the two models proposed in this thesis.This thesis constructed a single-factor stock selection strategy by the news sentiment factor,then did a historical backtesting with historical data from August 31,2018 to August 31,2019.The annualized return of the backtesting reached 23.89%,which significantly outperformed the CSI 500 Index and the CSI 300 Index,verifying the effectiveness of the factor and the factor has a better stock selection ability.