A Large-Scale Dataset for Multi-Class Financial News

Overview

Financial Natural Language Processing (NLP) for Brazilian Portuguese remains constrained by a shortage of domain-specific datasets. Addressing this gap, the authors introduce a large-scale corpus of approximately 160,000 Brazilian Portuguese financial news articles gathered from web sources. The work establishes initial benchmarks for a weakly supervised, multi-class classification approach grounded in semantic similarity.

Dataset and Annotation

Subsets of the corpus were annotated through two distinct strategies:

Manual labeling by human annotators
LLM-based annotation using a Large Language Model

The dataset is designed to support weakly supervised multi-class classification based on semantic similarity, with annotations serving as initial benchmarks rather than definitive ground truth.

Methodology

The authors evaluated three similarity-based classification methods using multilingual semantic embeddings:

Cosine similarity
K-Nearest Neighbors (KNN)
FAISS-based nearest-neighbor search

These methods were applied to both the manually labeled and LLM-annotated subsets to assess performance differences under varying label quality.

Key Findings

Experiments reveal that classification performance is highly sensitive to label quality and class distribution:

Manually labeled subset: KNN achieved the strongest results (47.3% accuracy, 31.9% F1-score), with FAISS performing closely behind.
LLM-annotated subset: Accuracy reached approximately 50%, but F1-scores dropped to near 20%, underscoring the adverse effects of label imbalance and noisy weak annotations.

Implications

These results establish foundational baselines for what remains a challenging task. The findings highlight the critical importance of high-quality manual annotations and provide a reference point for future research in Brazilian Portuguese Financial NLP.

Publication Details

Conference: Accepted for presentation at SMC 2026
Subjects: Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv Identifier: arXiv:2610.20630 [cs.CL]
DOI: https://doi.org/10.48550/arXiv.2610.20630
Submitted: 29 September 2026

via ArXiv CL+LG