Detoxify: A framework for abusive text transformation using LLMs

2026-07-08 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

This study evaluates Large Language Models (LLMs) for transforming abusive text (tweets and reviews) into non-abusive versions while retaining original intent. The study compares Gemini-1.5-flash, GPT-4o, DeepSeek-V3, and Groq using two datasets: one from IIT Guwahati (160,000 entries) and a Twitter (X) dataset (4265 tweets). The methodology involves data acquisition, preprocessing, LLM API configuration, review transformation, and evaluation via sentiment and semantic analysis using BERT-based models like SenWave-BERT and MPNet-base-v2. Results show GPT-4o and DeepSeek-V3 have similar transformation rates and sentiment distributions, maintaining semantic similarity to original text. Gemini achieved higher transformation success rates (e.g., 80% in Batch 1 vs. Groq's 60%) due to configurable relaxed safety settings (BLOCK_NONE), which Groq lacked. Groq often produced longer, overly positive transformations, sometimes altering original semantic meaning, and had the lowest semantic similarity with original input.

Key takeaway

For NLP Engineers developing content moderation systems, your choice of LLM for abusive text transformation is critical. Gemini-1.5-flash offers higher transformation success due to configurable safety settings, but Groq may over-positivize, altering original meaning. GPT-4o and DeepSeek-V3 maintain better semantic similarity. Evaluate models based on your specific balance between aggressive detoxification and preserving original intent, especially considering API limitations and inherent model biases.

Key insights

LLMs can transform abusive text, but model-specific safety settings and inherent behaviors significantly impact transformation success and semantic preservation.

Principles

• LLM safety settings directly influence transformation success rates.
• Semantic preservation varies significantly across different LLMs.
• Overly positive reframing can inadvertently alter original text intent.

Method

A six-stage framework: data acquisition, preprocessing, LLM API configuration, abuse detection, review transformation, and final sentiment/semantic analysis using BERT-based models.

In practice

• Configure Gemini's safety settings to BLOCK_NONE for higher transformation rates.
• Use MPNet-base-v2 for quantifying semantic similarity between texts.
• Employ Hate-BERT for abuse detection, noting its potential limitations with short-form text.

Topics

• Large Language Models
• Abusive Text Transformation
• Sentiment Analysis
• Semantic Analysis
• Content Moderation
• Gemini-1.5-flash

Code references

• pinglainstitute/LLM-reviewtransformation

Best for: AI Engineer, Machine Learning Engineer, Research Scientist, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• As Easy as Rocket Science: Assessing the Ability of Large Language Models to Interpret Negation in Figurative Language — LLMs struggle with combined negation and figurative language, with prompt style significantly impacting performance.
• Detoxification for LLM: From Dataset Itself — Detoxifying LLM pretraining datasets directly at the source fundamentally reduces inherent model toxicity while preserving semantics.
• Surpassing Scale by Efficiency: A Compact 135M Parameter Foundational LLM Natively Adapted for the Bangla Language — A compact 135M parameter LLM for Bangla achieves 1B-tier performance by efficiently merging tokens, overcoming low-resource script challenges.
• Better Literary Translation: A Multi-Aspect Data Generation and LLM Training Approach — Multi-aspect iterative refinement generates superior literary translation data for LLM training.
• Assessing Capabilities of Large Language Models in Social Media Analytics: A Multi-task Quest — Modern LLMs demonstrate varied capabilities across social media analytics tasks, with GPT-4 excelling in authorship verification and Gemini in attribute inference.
• The Tatoxa System for Text Detoxification in Low-Resource Languages: The Case of Tatar — The Tatoxa system achieves state-of-the-art text detoxification for Tatar, emphasizing native data's superiority over cross-lingual transfer.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.