Detoxifying Text with MaRCo: Controllable Revision with Experts and Anti-Experts
This addresses the challenge of mitigating online hate, especially subtle toxicity and microaggressions, with a novel approach that shows strong performance gains.
The paper tackled the problem of subtle toxicity in text detoxification by introducing MaRCo, a method that combines controllable generation and rewriting using a Product of Experts with autoencoder language models, resulting in rewrites preferred 2.1 times more in human evaluation.
Text detoxification has the potential to mitigate the harms of toxicity by rephrasing text to remove offensive meaning, but subtle toxicity remains challenging to tackle. We introduce MaRCo, a detoxification algorithm that combines controllable generation and text rewriting methods using a Product of Experts with autoencoder language models (LMs). MaRCo uses likelihoods under a non-toxic LM (expert) and a toxic LM (anti-expert) to find candidate words to mask and potentially replace. We evaluate our method on several subtle toxicity and microaggressions datasets, and show that it not only outperforms baselines on automatic metrics, but MaRCo's rewrites are preferred 2.1 $\times$ more in human evaluation. Its applicability to instances of subtle toxicity is especially promising, demonstrating a path forward for addressing increasingly elusive online hate.