Technology Category
Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLPMachine Learning: Large Multimodal Models (LMMs)Intelligent Robots: Multimodal Perception & Sensor Fusion
SynthDetoxM: Modern LLMs are Few-Shot Parallel Detoxification Data Annotators
🤖 AI Summary
Multilingual text detoxification is hindered by the scarcity of high-quality parallel annotated data, especially for low-resource languages. To address this, we propose the first synthetic data generation framework for multilingual detoxification: leveraging few-shot prompting, we orchestrate nine open-source large language models to generate detoxified parallel sentence pairs in German, French, Spanish, and Russian; these are rigorously filtered via multi-source toxicity scoring and human verification, yielding SynthDetoxM—a 16k-instance benchmark of high-fidelity parallel sentences, the first systematically validated multilingual synthetic detoxification dataset. Models trained on SynthDetoxM significantly outperform those trained on the real-world multilingual dataset MultiParaDetox under data-constrained settings and surpass all baseline LMs in few-shot evaluation across languages. This demonstrates both the efficacy and scalability of synthetic-data-driven detoxification modeling.
Technology Category
Application Category
📝 Abstract
Existing approaches to multilingual text detoxification are hampered by the scarcity of parallel multilingual datasets. In this work, we introduce a pipeline for the generation of multilingual parallel detoxification data. We also introduce SynthDetoxM, a manually collected and synthetically generated multilingual parallel text detoxification dataset comprising 16,000 high-quality detoxification sentence pairs across German, French, Spanish and Russian. The data was sourced from different toxicity evaluation datasets and then rewritten with nine modern open-source LLMs in few-shot setting. Our experiments demonstrate that models trained on the produced synthetic datasets have superior performance to those trained on the human-annotated MultiParaDetox dataset even in data limited setting. Models trained on SynthDetoxM outperform all evaluated LLMs in few-shot setting. We release our dataset and code to help further research in multilingual text detoxification.
Problem
Research questions and friction points this paper is trying to address.
Address multilingual text detoxification data scarcity
Introduce synthetic multilingual detoxification dataset
Enhance model performance with few-shot LLMs
Innovation
Methods, ideas, or system contributions that make the work stand out.
Few-shot LLMs for detoxification
Multilingual parallel dataset creation
Synthetic data outperforms human-annotated
🔎 Similar Papers
No similar papers found.
