A Large-Scale Dataset for Multi-Class Financial News Classification in Brazilian Portuguese
A large dataset of 160,000 Brazilian Portuguese financial news articles was created to help build systems for financial NLP where data is scarce. Researchers tested three similarity-based classification methods—Cosine Similarity, KNN, and FAISS—using semantic embeddings. Findings show that manually labeled data outperforms LLM-generated labels significantly in terms of accuracy and F1-score, proving that high-quality human annotation is crucial for accurate financial text classification.