Reflect, Reason, Rephrase ($\textup{R}^3$-Detox):\\An In-Context Learning Approach to Text Detoxification

ACL ARR 2025 May Submission142 Authors

Abstract: Traditional content moderation censors harmful content, which can often limit user participation. Text detoxification offers a better alternative, promoting civility without silencing voices. However, prior approaches oversimplify the task by treating detoxification as a one-step process, neglecting the deep contextual analysis needed to remove toxicity while preserving meaning. In this paper, we introduce $\textup{R}^3$-Detox—a Reflect, Reason, and Rephrase framework that enhances detoxification through a structured three-step process, all executed within a single prompt. First, we instruct the LLM to analyze potential toxic words or phrases, guided by Shapley values from toxicity detectors, to counteract potential hallucinations. Next, the model assesses the overall toxicity of the sentence based on these identified elements. Finally, leveraging this prior analysis, the model reasons about necessary modifications to eliminate toxicity while maintaining meaning. We apply this framework and Self-Reflection models to enrich offensive content paraphrasing datasets—ParaDetox, Parallel Detoxification, and APPDIA—by adding explicit detoxification reasoning to each instance, which originally contained only input sentences and their paraphrases. We evaluate our methodology using In-Context Learning, comparing $\textup{R}^3$-Detox against state-of-the-art methods on the same datasets. Experimental results show that our approach outperforms existing methodologies, even in instruction-following models.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: Sentiment Analysis Stylistic Analysis and Argument Mining, Computational Social Science and Cultural Analytics, Generation, Information Extraction

Contribution Types: NLP engineering experiment

Languages Studied: English

Keywords: Controlled Text Generation, Large Language Models, Prompt-based Learning, Step-by-Step Reasoning, Model-based Evaluation

Submission Number: 142