Answers are grounded in this paper and cite their sources.
Toxic chat in online games exposes young players to harassment, while moderation of Indonesian chat is hampered by slang and non-standard spelling. This study compares monolingual IndoBERT and multilingual mBERT for four-class toxicity classification (neutral, violence, racist, harassment) of Indonesian Roblox and Minecraft chat to identify the more suitable model for moderation. Both models were fine-tuned with identical settings on the Indonesian Chat Dataset (10,702 chats) using three seeds, and evaluated with accuracy, macro-averaged precision, recall, F1-score, and McNemar’s test. IndoBERT reached a mean macro F1-score of 0.8541 ± 0.0019 versus 0.8263 ± 0.0090 for mBERT, significant in two of three seeds. Neutral was the hardest class, and IndoBERT let 5.8% of toxic chats pass as neutral versus 8.3% for mBERT. Excluding 106 duplicated test chats did not change the ranking. Under these settings, IndoBERT performed better, although labels lack inter-annotator agreement and data cover two games only.
No takes yet. Share an insight, caveat, or question.
Sonny et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: