Toxicity Classification in Ukrainian
Daryna Dementieva, Valeriia Khylenko, Nikolay Babakov, Georg Groh
TL;DR
This work addresses the lack of Ukrainian toxicity classification resources by introducing a first Ukrainian toxicity corpus and evaluating three cross-lingual transfer approaches (backtranslation, LLM prompting, adapter training) alongside three data acquisition methods (translation from English data, keyword-based semi-synthetic data, and crowdsourcing). It finds that translation-based supervision offers robust cross-domain performance, while crowdsourced data deliver the highest annotation fidelity and model performance on in-domain data, with fine-tuned cross-lingual models excelling on all tests. The study provides practical baselines and releases data and models publicly to support multilingual toxicity detection. It also discusses limitations, notably focusing on obscene lexicon and using English as the sole resource language for translation experiments, suggesting future work on additional languages and broader toxic language categories.
Abstract
The task of toxicity detection is still a relevant task, especially in the context of safe and fair LMs development. Nevertheless, labeled binary toxicity classification corpora are not available for all languages, which is understandable given the resource-intensive nature of the annotation process. Ukrainian, in particular, is among the languages lacking such resources. To our knowledge, there has been no existing toxicity classification corpus in Ukrainian. In this study, we aim to fill this gap by investigating cross-lingual knowledge transfer techniques and creating labeled corpora by: (i)~translating from an English corpus, (ii)~filtering toxic samples using keywords, and (iii)~annotating with crowdsourcing. We compare LLMs prompting and other cross-lingual transfer approaches with and without fine-tuning offering insights into the most robust and efficient baselines.