Detoxifying Language Models with a Toxic Corpus

TL;DR

This paper explores combining data-based and decoding-based debiasing methods by using toxic corpus as an additional resource to significantly reduce toxicity in language model outputs.

Contribution

It introduces a novel approach that leverages toxic corpus to enhance existing debiasing techniques, effectively reducing toxicity in language generation.

Findings

Toxic corpus helps decrease generated toxicity.

Combining debiasing methods improves results.

Enhanced reduction of harmful biases.

Abstract

Existing studies have investigated the tendency of autoregressive language models to generate contexts that exhibit undesired biases and toxicity. Various debiasing approaches have been proposed, which are primarily categorized into data-based and decoding-based. In our study, we investigate the ensemble of the two debiasing paradigms, proposing to use toxic corpus as an additional resource to reduce the toxicity. Our result shows that toxic corpus can indeed help to reduce the toxicity of the language generation process substantially, complementing the existing debiasing methods.