Mgs Stereotype Detection Eval

mgs-stereotype-detection-eval

Stereotype Detection in LLMs: A Multiclass, Explainable, and Benchmark-Driven Approach — Wu et al. (2024) (arXiv:2404.01768, 2024)

What this evaluates

Evaluates the ability of classifiers to detect and categorize stereotypes across multiple intersecting dimensions (race, gender, profession, religion) in text. It also probes cross-dataset generalization and quantifies generative bias deviation in LLMs.

Datasets

• MGS Dataset (MGSD) — total 51867; splits: test (-1)
• StereoSet — total ?; splits: test (-1)
• CrowsPairs — total ?; splits: test (-1)

Metrics

• Macro F1 Score (primary) — range: [0, 1]
  • The unweighted mean of recall and precision calculated per stereotype dimension (Race, Gender, Profession, Religion) and averaged across classes.
• Precision — range: [0, 1]
  • The ratio of correctly predicted positive stereotype instances to the total predicted positives, macro-averaged across dimensions.
• Recall — range: [0, 1]
  • The ratio of correctly predicted positive stereotype instances to the total actual positives, macro-averaged across dimensions.
• Accuracy — range: [0, 1]
  • The proportion of total instances correctly classified across all stereotype dimensions.

Input / output format

Input: Text snippets or sentences containing potential stereotypical content, labeled with one or multiple stereotype dimensions (Race, Gender, Profession, Religion).

Output: Predicted stereotype dimension(s) and stereotype label per instance.

Scoring recipe

def compute_macro_metrics(y_true, y_pred, classes):
    precisions, recalls, f1s = [], [], []
    for c in classes:
        tp = sum(1 for t, p in zip(y_true, y_pred) if t == c and p == c)
        fp = 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)
        prec = tp / (tp + fp) if (tp + fp) > 0 else 0
        rec = tp / (tp + fn) if (tp + fn) > 0 else 0
        f1 = 2 * prec * rec / (prec + rec) if (prec + rec) > 0 else 0
        precisions.append(prec)
        recalls.append(rec)
        f1s.append(f1)
    return sum(precisions)/len(precisions), sum(recalls)/len(recalls), sum(f1s)/len(f1s)

Common pitfalls

• Evaluating only on a single stereotype dimension ignores the paper's core focus on intersectionality and multi-dimensional detection.
• Cross-dataset generalization scores vary drastically; testing solely on CrowsPairs or StereoSet without MGSD may misrepresent model robustness.
• Explainability tools like SHAP and LIME provide local approximations that may not accurately reflect the model's global decision boundaries.

Evidence (verbatim from paper)

The results in Table [1] show that multiple-dimension stereotype detector consistently outperform single-dimension stereotype counterparts across all stereotype dimensions—Race, Profession, Gender, Religion—as well as in all Macro evaluation metrics: Precision, Recall, and F1 Score.

Citation

@misc{wu2024stereotypedetection,
  title={Stereotype Detection in LLMs: A Multiclass, Explainable, and Benchmark-Driven Approach},
  author={Wu et al. (2024)},
  year={2024},
  note={arXiv:2404.01768}
}

• arXiv: 2404.01768