Forgive or forget: Understanding the context of hate in audio retrieval systems

2026-06-04

AI Summary

This work addresses the challenge of retrieving harmful content in text-to-audio retrieval systems caused by contextual dependencies. To mitigate this issue without altering user intent or sacrificing critical semantics, the authors propose a post-hoc causal debiasing framework that integrates causal inference with an emotion-controllable mediation mechanism. The framework introduces two complementary strategies: Forgiveness-based reranking and filtering (Forgive) and counterfactual prompt generation (Forget). By leveraging logit adjustment and counterfactual interventions, it establishes a model-agnostic, context-aware safety mechanism that seamlessly integrates into existing pipelines. Experimental results demonstrate that the proposed approach significantly reduces toxicity in retrieved audio across multiple benchmarks while preserving high retrieval accuracy, effectively balancing system safety and reliability.

Abstract

Handling toxic retrieval in text-to-audio systems is challenging due to contextual dependencies. Existing strategies (e.g., rephrasing, summarization) risk altering intent or omitting details. We propose a post hoc causal debiasing framework with a sentiment-controlled mediator to preserve semantic relevance while suppressing harmful speech. Our approach is model-agnostic and integrates seamlessly with existing retrieval pipelines. We introduce two variants: Forgive, which re-ranks and filters toxic audio via logit adjustment, and Forget, which generates counterfactual toxic prompts to mitigate harmful retrievals. Experiments show consistent toxicity reduction with minimal loss in retrieval accuracy, improving both safety and reliability.

Problem

• hate speech
• audio retrieval
• toxicity mitigation
• contextual dependency
• text-to-audio systems

Innovation

• causal debiasing
• sentiment-controlled mediator
• toxicity mitigation
• audio retrieval
• counterfactual prompting