Leveraging Machine Translation for Toxicity Classification

This research article presents an empirical exploration of multilingual toxicity detection using machine translation (MT) systems. The study compares translation-based classification pipelines against language-specific and multilingual classifiers across 17 languages with varying resource levels. The findings indicate that translation-based methods consistently outperform out-of-distribution classifiers in 81.3% of cases, particularly benefiting low-resource languages. The analysis reveals that traditional classifiers surpass large language model (LLM) judges, especially for languages with limited resources. Additionally, the study examines the impact of MT-specific fine-tuning on LLMs, noting that while it reduces refusal rates, it may adversely affect detection accuracy for low-resource languages. The article concludes with practical recommendations for developing scalable multilingual content moderation systems, emphasizing the significance of translation quality and resource levels in enhancing toxicity detection capabilities.