Hate Speech Detection Eval

---
name: hate-speech-detection-eval
description: Evaluates binary and multi-label hate speech detection models on Brazilian Portuguese text. It specifically probes a model's sensitivity to targeted minority groups and its ranking quality under severe class imbalance. Use when the user wants to benchmark on ToxiGen-PT, Portuguese Superset Benchmark, HateBR, OLID-BR, TuPy-E, ToLD-BR, or asks about evaluating this task. Reports Macro-Recall.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2506.10245
  bibtex_key: brito2025toxsyn
  confidence: high
---

# hate-speech-detection-eval

> ToxSyn: Reducing Bias in Hate Speech Detection via Synthetic Minority Data in Brazilian Portuguese — Brito et al. (2025) (arXiv:2506.10245, 2025)

## What this evaluates

Evaluates binary and multi-label hate speech detection models on Brazilian Portuguese text. It specifically probes a model's sensitivity to targeted minority groups and its ranking quality under severe class imbalance.

## Datasets

- **ToxiGen-PT** — total 8960; splits: test (8960)
- **Portuguese Superset Benchmark** — total 6002; splits: test (6002)
- **HateBR** — total ?; splits: test (-1)
- **OLID-BR** — total ?; splits: test (-1)
- **TuPy-E** — total ?; splits: test (-1)
- **ToLD-BR** — total ?; splits: test (-1)

## Metrics

- `Macro-Recall` **(primary)** — range: [0, 1]
  - Unweighted mean of recall scores computed independently for each class (toxic/non-toxic or each minority group).
- `ROC AUC` — range: [0, 1]
  - Area under the Receiver Operating Characteristic curve, measuring the model's ability to rank positive instances higher than negative ones across all classification thresholds.

## Input / output format

**Input**: Raw text strings in Brazilian Portuguese (social media posts/comments).

**Output**: Binary label (toxic vs. non-toxic) or a multi-label vector indicating presence of specific protected minority group targets.

## Scoring recipe

```python
def compute_macro_recall(y_true, y_pred, num_classes):
    recalls = []
    for c in range(num_classes):
        tp = sum(1 for t, p in zip(y_true, y_pred) if t == c and p == c)
        fn = sum(1 for t, p in zip(y_true, y_pred) if t == c and p != c)
        recalls.append(tp / (tp + fn) if (tp + fn) > 0 else 0.0)
    return sum(recalls) / len(recalls)

def compute_roc_auc(y_true, y_scores):
    fpr, tpr, _ = roc_curve(y_true, y_scores)
    return auc(fpr, tpr)
```

## Common pitfalls

- Severe class imbalance in hate speech corpora makes accuracy misleading; macro-recall is required to capture minority group sensitivity.
- Existing benchmarks only label protected groups when content is explicitly toxic, leaving neutral/supportive references unlabeled and skewing evaluation.
- Translated datasets contain unnaturalized phrasing and culturally mismatched slurs that may not reflect authentic Brazilian Portuguese usage.

## Evidence (verbatim from paper)

> Due to class imbalance, we report macro-averaged recall (to measure per-class sensitivity) and ROC AUC (to assess ranking quality).

## Citation

```bibtex
@misc{brito2025toxsyn,
  title={ToxSyn: Reducing Bias in Hate Speech Detection via Synthetic Minority Data in Brazilian Portuguese},
  author={Brito et al. (2025)},
  year={2025},
  note={arXiv:2506.10245}
}
```

- arXiv: 2506.10245