Sentiment analysis of stock-related tweets is a challenging task, not only due to the specificity of the domain but also because of the short nature of the texts. This work proposes SA-MAIS, a two-step lightweight methodology, specially adapted to perform sentiment analysis in domain-constrained short-text messages. To tackle the issue of domain specificity, based on word frequency, the most relevant words are automatically extracted from the new domain and then manually tagged to update an existing domain-specific sentiment lexicon. The sentiment classification is then performed by combining the updated domain-specific lexicon with VADER sentiment analysis, a well-known and widely used sentiment analysis tool. The proposed method is compared with other well-known and widely used sentiment analysis tools, including transformer-based models, such as BERTweet, Twitter-roBERTa and FinBERT, on a domain-specific corpus of stock market-related tweets comprising 1 million messages. The experimental results show that the proposed approach largely surpasses the performance of the other sentiment analysis tools, reaching an overall accuracy of 72.0%. The achieved results highlight the advantage of using a hybrid method that combines domain-specific lexicons with existing generalist tools for the inference of textual sentiment in domain-specific short-text messages.

This work was partially supported by national funds through Fundação para a Ciência e a Tecnologia (FCT) with references UIDB/50021/2020 and UIDB/00315/2020

Sentiment analysis,Sentiment classification,Sentiment lexicon,Stock market,Tweets

• Computer and Information Sciences - Natural Sciences
• Media and Communications - Social Sciences
• Other Social Sciences - Social Sciences