RTP-LX: Can LLMs Evaluate Toxicity in Multilingual Scenarios?
Adrian de Wynter, Ishaan Watts, Tua Wongsangaroonsri, Minghui Zhang, Noura Farra, Nektar Ege Altıntoprak, Lena Baur, Samantha Claudet
Abstract
Large language models (LLMs) and small language models (SLMs) are being adopted at remarkable speed, although their safety still remains a serious concern. With the advent of multilingual S/LLMs, the question now becomes a matter of scale: can we expand multilingual safety evaluations of these models with the same velocity at which they are deployed? To this end, we introduce RTP-LX, a human-transcreated and human-annotated corpus of toxic prompts and outputs in 28 languages. RTP-LX follows participatory design practices, and a portion of the corpus is especially designed to detect culturally-specific toxic language. We evaluate 10 S/LLMs on their ability to detect toxic content in a culturally-sensitive, multilingual scenario. We find that, although they typically score acceptably in terms of accuracy, they have low agreement with human judges when scoring holistically the toxicity of a prompt; and have difficulty discerning harm in context-dependent scenarios, particularly with subtle-yet-harmful content (e.g. microaggressions, bias). We release this dataset to contribute to further reduce harmful uses of these models and improve their safe deployment.
BibTeX
@article{de Wynter_Watts_Wongsangaroonsri_Zhang_Farra_Altıntoprak_Baur_Claudet_Gajdušek_Gu_Kaminska_Kaminski_Kuo_Kyuba_Lee_Mathur_Merok_Milovanović_Paananen_Paananen_Pavlenko_Vidal_Strika_Tsao_Turcato_Vakhno_Velcsov_Vickers_Visser_Widarmanto_Zaikin_Chen_2025, title={RTP-LX: Can LLMs Evaluate Toxicity in Multilingual Scenarios?}, volume={39}, url={https://ojs.aaai.org/index.php/AAAI/article/view/35011}, DOI={10.1609/aaai.v39i27.35011}, abstractNote={Large language models (LLMs) and small language models (SLMs) are being adopted at remarkable speed, although their safety still remains a serious concern. With the advent of multilingual S/LLMs, the question now becomes a matter of scale: can we expand multilingual safety evaluations of these models with the same velocity at which they are deployed? To this end, we introduce RTP-LX, a human-transcreated and human-annotated corpus of toxic prompts and outputs in 28 languages. RTP-LX follows participatory design practices, and a portion of the corpus is especially designed to detect culturally-specific toxic language. We evaluate 10 S/LLMs on their ability to detect toxic content in a culturally-sensitive, multilingual scenario. We find that, although they typically score acceptably in terms of accuracy, they have low agreement with human judges when scoring holistically the toxicity of a prompt; and have difficulty discerning harm in context-dependent scenarios, particularly with subtle-yet-harmful content (e.g. microaggressions, bias). We release this dataset to contribute to further reduce harmful uses of these models and improve their safe deployment.}, number={27}, journal={Proceedings of the AAAI Conference on Artificial Intelligence}, author={de Wynter, Adrian and Watts, Ishaan and Wongsangaroonsri, Tua and Zhang, Minghui and Farra, Noura and Altıntoprak, Nektar Ege and Baur, Lena and Claudet, Samantha and Gajdušek, Pavel and Gu, Qilong and Kaminska, Anna and Kaminski, Tomasz and Kuo, Ruby and Kyuba, Akiko and Lee, Jongho and Mathur, Kartik and Merok, Petter and Milovanović, Ivana and Paananen, Nani and Paananen, Vesa-Matti and Pavlenko, Anna and Vidal, Bruno Pereira and Strika, Luciano Ivan and Tsao, Yueh and Turcato, Davide and Vakhno, Oleksandr and Velcsov, Judit and Vickers, Anna and Visser, Stéphanie F. and Widarmanto, Herdyan and Zaikin, Andrey and Chen, Si-Qing}, year={2025}, month={Apr.}, pages={27940-27950} }