Sentiment Analysis of News on Corporation Using KoBERT

This study explores the accuracy level of the sentiment analysis of news article sentences from Korean newspaper, using KoBERT which is a modified version of BERT developed by Google. For comparison, we use MBERT which is the multilingual version of BERT, Google Sentiment Analysis provided through Google API, and dictionary based approach. This paper finds that the accuracy level of the sentiment classification based on KoBERT is the highest at 85.7%, achieving a significantly higher level of accuracy compared to the other three models. MBERT shows the next highest accuracy level at 77.5% and the other two models provide even lower accuracy rate. We further investigate whether the sentiment classification results obtained from these four models could predict future stock return. Using cumulative future stock returns for 3 or 5 days after the news on corporation publishes, we find that the sentiment score based on the sentiment classification from the KoBERT model predicts future return better than the other three models. Overall, these findings would serve as a reference for conducting further studies related to sentiment analysis on accounting and financial text.