Sentiment Analysis of Thai Stock Reviews Using Transformer Models

The stock market is typically affected by various factors for a long time, such as politics, economics, and finance. These are expressed through online media that people can easily access today. Moreover, in the digital era, data growth is an exponential trend, and a million records of data are generated through many online platforms over the internet. To utilize that information in time, a stock sentimental analysis system integrated with the transformer base model is proposed. This work applies the transformer base models that can break through NLP limitations from the past. Furthermore, we gather data as fundamental analysis in Thai financial content from a financial institution. However, to compare the result between embedding techniques with baseline, we use multinomial logistic regression in the form of a predictive model and apply the baseline, the term frequency-inverse document frequency (TF-IDF). Our experiment shows that WangchanBERTa and BERT can achieve high accuracy at 92.52% and 89.12%, respectively, and the baseline result is 85.03%. In conclusion, our proposed system can precisely predict stock sentiment in Thai with high accuracy.