On the Role of Speech Data in Reducing Toxicity Detection Bias

2024-11-12

arXiv.org

AI Summary

This work addresses spurious false-positive bias in text toxicity detection—where references to specific demographic groups trigger erroneous toxicity classifications—and presents the first empirical investigation into whether speech modality mitigates this bias. Leveraging the multilingual MuTox dataset, we introduce fine-grained manual demographic annotations and propose a speech-text joint modeling framework. Using fairness metrics—including demographic parity difference (DPD) and equalized odds (EO)—we systematically compare bias in speech versus text classifiers. Results show that speech input reduces group-associated false-positive rates by 23% on average, with especially pronounced gains on ambiguous and contentious samples; moreover, optimizing classifier architecture yields greater bias reduction than improving ASR transcription quality. Our contributions include: (1) empirically demonstrating speech’s bias-mitigation potential; (2) releasing fully annotated data and a multimodal toxicity construction guideline; and (3) establishing a novel paradigm for fair, robust cross-modal content safety detection.

Technology Category

• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Computer Vision: Bias, Fairness & Privacy

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

Text toxicity detection systems exhibit significant biases, producing disproportionate rates of false positives on samples mentioning demographic groups. But what about toxicity detection in speech? To investigate the extent to which text-based biases are mitigated by speech-based systems, we produce a set of high-quality group annotations for the multilingual MuTox dataset, and then leverage these annotations to systematically compare speech- and text-based toxicity classifiers. Our findings indicate that access to speech data during inference supports reduced bias against group mentions, particularly for ambiguous and disagreement-inducing samples. Our results also suggest that improving classifiers, rather than transcription pipelines, is more helpful for reducing group bias. We publicly release our annotations and provide recommendations for future toxicity dataset construction.

Problem

• Investigating bias reduction in toxicity detection using speech data
• Comparing speech- and text-based toxicity classifiers systematically
• Improving classifiers to reduce group bias in toxicity detection

Innovation

• Leveraging speech data to reduce toxicity detection bias
• Comparing speech- and text-based toxicity classifiers systematically
• Improving classifiers to mitigate group bias effectively