Bias Correlation Mitigation Eval

---
name: bias-correlation-mitigation-eval
description: This evaluation protocol assesses the effectiveness of individual and joint bias mitigation strategies across toxicity detection and word embeddings. It probes whether debiasing for one social identity correlates with or affects bias levels in others, and measures the trade-off between bias reduction and model utility. Use when the user wants to benchmark on Jigsaw Toxicity Dataset, CoNLL 2003, or asks about evaluating this task. Reports AUC.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2205.12391
  bibtex_key: cheng2022biascorrelations
  confidence: high
---

# bias-correlation-mitigation-eval

> Toward Understanding Bias Correlations for Mitigation in NLP — Cheng et al. (2022) (arXiv:2205.12391, 2022)

## What this evaluates

This evaluation protocol assesses the effectiveness of individual and joint bias mitigation strategies across toxicity detection and word embeddings. It probes whether debiasing for one social identity correlates with or affects bias levels in others, and measures the trade-off between bias reduction and model utility.

## Datasets

- **Jigsaw Toxicity Dataset** — total 403957; splits: train (-1), val (-1), test (-1)
- **CoNLL 2003** — total ?; splits: test (-1)

## Metrics

- `AUC` **(primary)** — range: [0, 1]
  - Area Under the Receiver Operating Characteristic curve. Measures the model's ability to distinguish between classes across all classification thresholds.
- `micro-F1` — range: [0, 1]
  - Micro-averaged F1 score calculated globally by counting total true positives, false negatives, and false positives.
- `Accuracy` — range: [0, 1]
  - Ratio of correctly predicted instances to the total number of instances.
- `MACs` — range: [0, 1]
  - Mean Absolute Cosine distance used to quantify bias removal across social identities. Higher values indicate greater bias reduction.
- `F1/Precision/Recall` — range: [0, 1]
  - Standard classification metrics evaluated on downstream NER, POS tagging, and POS chunking tasks to measure embedding utility.

## Input / output format

**Input**: Text snippets with toxicity and identity annotations (Jigsaw); pre-trained word vectors (L2-reddit corpus).

**Output**: Predicted toxicity scores/classes; debiased word embeddings; downstream task predictions (NER, POS, Chunking).

## Scoring recipe

```python
def evaluate(y_true, y_pred, embeddings, lexicons):
    auc = roc_auc_score(y_true, y_pred)
    f1 = f1_score(y_true, y_pred, average='micro')
    acc = accuracy_score(y_true, y_pred)
    macs = [mean_absolute_cosine_dist(embeddings, lex) for lex in lexicons]
    return {'AUC': auc, 'micro-F1': f1, 'Acc': acc, 'MACs': macs}
```

## Common pitfalls

- Confusing individual bias metrics with joint bias metrics, which are only applicable when debiasing for multiple identities simultaneously.
- Assuming debiasing always improves prediction performance; the protocol explicitly notes an inherent debiasing-accuracy trade-off during training.
- Sequential hard-debiasing can amplify bias for subsequent identities, making it less effective than joint debiasing.

## Evidence (verbatim from paper)

> We use AUC, micro-F1, and accuracy (Acc.) scores as the evaluation metrics. For bias mitigation, following [19, 22], we use the standard individual bias metric introduced in Section 3.1 for independent debiasing models (i.e., Gender, Race, and Religion). As the individual bias metric is not suitable when multiple forms of bias are present, we measure bias using the joint bias metric described in Section 4.1 when debiasing for multiple forms of biases (i.e., Ge+Ra, Ge+Re, Ra+Re, and Joint).

## Citation

```bibtex
@misc{cheng2022biascorrelations,
  title={Toward Understanding Bias Correlations for Mitigation in NLP},
  author={Cheng et al. (2022)},
  year={2022},
  note={arXiv:2205.12391}
}
```

- arXiv: 2205.12391