A Brazilian Dataset for Financial Sentiment Analysis: An Active Learning-Based Annotation Framework
Resumo
This paper introduces a new expert-annotated dataset for financial sentiment analysis in Brazilian Portuguese, constructed from earnings call transcripts of the banking sector. The dataset was developed using a two-stage active learning pipeline designed to reduce annotation cost while preserving data quality. In the first stage, clustering and active learning were combined to filter relevant sentences from noisy transcripts. In the second stage, a Query by Committee strategy combining the transformer models BERTimbau and DeB3RTa was employed to select informative samples for expert annotation. The resulting dataset contains 719 annotated sentences with agreement analysis across multiple annotators. We further evaluated BERTimbau, DeB3RTa, and the external transformer model BERTuguês under three experimental settings, including fine-grained eight-class classification and complementary three-class scenarios. In the eight-class setting, BERTuguês and BERTimbau achieved comparable F1-macro scores (0.43±0.03 and 0.43 ± 0.01, respectively). In the sentiment-only three-class setup, BERTimbau achieved the best performance (0.68± 0.06). These results suggest that the dataset supports both fine-grained and coarse-grained sentiment analysis while remaining useful for models not involved in the active learning process.
Referências
Abilio, R., Coelho, G. P., and da Silva, A. E. A. Replication Data for: A Brazilian Dataset for Financial Sentiment Analysis: An Active Learning-Based Annotation Framework. DOI: 10.25824/redu/HUQPYK, 2026. Repositório de Dados de Pesquisa da Unicamp.
Alawaji, R. and Aloraini, A. Sentiment Analysis of Digital Banking Reviews Using Machine Learning and Large Language Models. Electronics 14 (11): 2125, 2025.
Ashrafi Asli, S. A., Sabeti, B., Majdabadi, Z., Golazizian, P., Fahmi, R., and Momenzadeh, O. Optimizing Annotation Effort Using Active Learning Strategies: A Sentiment Analysis Case Study in Persian. In Proceedings of the Twelfth Language Resources and Evaluation Conference. Marseille, France, pp. 2855–2861, 2020.
Cochran, W. G. Sampling Techniques. Wiley, New York, 1977.
da Silva, F., Roman, N., and Carvalho, A. Stock market tweets annotated with emotions. Corpora 15 (3): 343–354, 2020.
Fleiss, J. L. Measuring nominal scale agreement among many raters. Psychological Bulletin 76 (5): 378–382, 1971.
Januário, B. A., Carosia, A. E. d. O., da Silva, A. E. A., and Coelho, G. P. Sentiment Analysis Applied to News from the Brazilian Stock Market. IEEE Latin America Transactions 20 (3): 512–518, 2022.
Kakimoto, G., Haddadi, S., Araújo, P., Silva, F., dos Reis, J., and Reis, M. B2T: A dataset of tweets in portuguese language about brazilian banks. In Proceedings of the VI Dataset Showcase Workshop (DSW). Florianópolis, SC, Brazil, pp. 1–11, 2024.
Kaseb, A. and Farouk, M. Active learning for Arabic sentiment analysis. Alexandria Engineering Journal vol. 77, pp. 177–187, 2023.
Krippendorff, K. Computing Krippendorff’s alpha-reliability. [link], 2011. Accessed in: 2026-05-04.
Lan, Y., Wu, Y., Xu, W., Feng, W., and Zhang, Y. Chinese fine-grained financial sentiment analysis with large language models. Neural Computing and Applications 37 (30): 24883–24892, 2025.
Li, D., Wang, Z., Chen, Y., Jiang, R., Ding, W., and Okumura, M. A Survey on Deep Active Learning: Recent Advances and New Frontiers. IEEE Transactions on Neural Networks and Learning Systems 36 (4): 5879–5899, 2025.
Liebenlito, M., Inayah, N., Choerunnisa, E., Sutanto, T. E., and Inna, S. Active learning on Indonesian Twitter sentiment analysis using uncertainty sampling. Journal of Applied Data Sciences 5 (1): 114–121, 2024.
Lin, J. Divergence measures based on the Shannon entropy. IEEE Transactions on Information Theory 37 (1): 145–151, 1991.
Malo, P., Sinha, A., Takala, P., Korhonen, P., and Wallenius, J. Good Debt or Bad Debt: Detecting Semantic Orientations in Economic Texts. Journal of the American Society for Information Science and Technology 65 (4): 782–796, 2014.
Mostafavi, S., Yahyavi, Y., and Ravanmehr, R. Systematic literature review on sentiment analysis using transformers. International Journal of Data Science and Analytics 22 (1): 49, 2026.
Pires, H., Paucar, L., and Carvalho, J. P. DeB3RTa: A Transformer-Based Model for the Portuguese Financial Domain. Big Data and Cognitive Computing 9 (3), 2025.
Pustejovsky, J. and Stubbs, A. Natural Language Annotation for Machine Learning: A Guide to Corpus-Building for Applications. O’Reilly, Sebastopol, CA, 2013.
Raicu, I. Financial Banking Dataset for Supervised Machine Learning Classification. Informatica Economica 23 (1): 37–49, 2019.
Romero, V., Assis, G., Carvalho, J., Mann, P., and Paes, A. Opportunities vs. Risks: Exploring Automatic Annotation of Financial Polarity Biases via Large Language Models. In Symposium on Knowledge Discovery, Mining and Learning (KDMiLe). Fortaleza, CE, Brazil, pp. 1–8, 2025.
Santos, L., Bianchi, R., and Costa, A. FinBERT-PT-BR: Análise de Sentimentos de Textos em Português do Mercado Financeiro. In Anais do II Brazilian Workshop on Artificial Intelligence in Finance. João Pessoa, PB, Brazil, pp. 144–155, 2023.
Souza, F., Nogueira, R., and Lotufo, R. BERTimbau: Pretrained BERT Models for Brazilian Portuguese. In Intelligent Systems. Number 12319 in Lecture Notes in Computer Science. Cham, pp. 403–417, 2020.
Zago, R. and Pedotti, L. BERTugues: A Novel BERT Transformer Model Pre-trained for Brazilian Portuguese. Semina: Ciências Exatas e Tecnológicas vol. 45, pp. e50630–e50630, 2024.
