Preprint Releases 160,000 Article Brazilian Portuguese Financial Dataset

Researchers working on financial natural language processing for Brazilian Portuguese have published a dataset meant to close a gap in the field: most financial text classification work targets English, leaving Portuguese language systems with little domain specific data to train on. The team, led by Leomax Bandeira and six co-authors, built a corpus of roughly 160,000 financial news articles scraped from web sources and then set out to test whether the articles could be sorted into categories without paying for full manual annotation of the whole set.

To test that, the authors annotated subsets of the corpus two ways: by hand, and by asking a large language model to assign labels automatically. They then ran three similarity based classification methods, cosine similarity, K-Nearest Neighbors, and FAISS based nearest-neighbor search, using multilingual semantic embeddings, and compared how each performed depending on which labeling approach fed it.

The results split sharply by label source. On the manually labeled subset, KNN was the strongest method, and FAISS trailed closely behind it. On the LLM-annotated subset, accuracy held up reasonably well, but F1-score, the measure that balances precision against recall and punishes a classifier that leans on the most common category, fell far short. The authors attribute that drop to label imbalance and noise introduced by the automated weak annotations, and they conclude that high quality manual annotation remains necessary for this kind of task, at least with current weak supervision methods.

The numbers

• The corpus runs to approximately 160,000 Brazilian Portuguese financial news articles.
• KNN reached 47.3% accuracy and a 31.9% F1-score on the manually labeled subset, the best result the authors recorded.
• On the LLM-annotated subset, accuracy reached approximately 50% while F1-scores dropped to near 20%, the clearest sign in the paper that an accuracy number alone can hide a weak classifier.

The paper, accepted for presentation at SMC 2026, frames its results as a baseline rather than a finished system, and the authors say the dataset itself is the more lasting contribution: a reference point other groups working on Brazilian Portuguese financial text can build against, rather than a method anyone should deploy as is.