Classifying Emotions in Brazilian Stock Market Tweets

Anonymous

Abstract: Stock markets play an important role in accelerating economic growth of developing countries like Brazil and, typically, leads to strong emotions in people, which may be reflected in their behaviour at social media like X, old Twitter. At the bright side, analysing these emotions could unveil interesting insights about public perception, potentially leading to more accurate and profitable stock market forecasts. Automatic emotion detection in tweets has been explored by many studies in the past years. State-of-the-art pre-trained language models have also been used to this end. We propose to detect emotion on tweets related to the brazilian stock market, which have a few dedicated researches. We trained BERTimbau (Portuguese version) large and base on a free domain training dataset and tested the models on the target domain (Brazilian stock market), a cross-domain approach. Applying Plutchik's wheel in its basic form, in this work we considers only the four main emotion pairs, to wit, joy $\times$ sadness, anger $\times$ fear, trust $\times$ disgust and surprise $\times$ anticipation. Models performance drops to values ranging from F1 nil, for BERTimbau Large and \textit{Joy} to $F1=0.78$, also for BERTimbau Large, but with \textit{Trust}. Results by both BERTimbau Large and Base, after test on the \textit{Free Domain Corpus} test set (same domain used for training) reached almost 100\% accuracy for all emotions.

Paper Type: long

Research Area: NLP Applications

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: Portuguese