Added Toxicity Mitigation at Inference Time for Multimodal and Massively Multilingual Translation
This addresses a critical safety issue in large-scale translation systems for users and developers, though it appears incremental as it builds on existing systems like SEAMLESSM4T.
The paper tackles the problem of added toxicity in multimodal and massively multilingual machine translation, where translations become more toxic than the input, by presenting MinTox, a novel inference-time pipeline that mitigates this issue. It achieves significant reduction, filtering out approximately 25% to 95% of added toxicity across domains and modalities while maintaining translation quality.
Added toxicity in the context of translation refers to the fact of producing a translation output with more toxicity than there exists in the input. In this paper, we present MinTox which is a novel pipeline to identify added toxicity and mitigate this issue which works at inference time. MinTox uses a toxicity detection classifier which is multimodal (speech and text) and works in languages at scale. The mitigation method is applied to languages at scale and directly in text outputs. MinTox is applied to SEAMLESSM4T, which is the latest multimodal and massively multilingual machine translation system. For this system, MinTox achieves significant added toxicity mitigation across domains, modalities and language directions. MinTox manages to approximately filter out from 25% to 95% of added toxicity (depending on the modality and domain) while keeping translation quality.