<![CDATA[Background: Indonesian social media toxicity detection remains difficult because harmful meaning is often produced through profanity, indirectness, ridicule, code-mixing, identity reference, and culturally situated politeness rather than through abusive vocabulary alone. Objective: This study examines how toxicity, politeness, and cultural pragmatics interact in Indonesian social media NLP annotation and asks what contextual evidence is required to distinguish abuse, hate speech, satire, criticism, and banter. Method: Using 60 anonymised Indonesian Twitter/X items from a public multi-label hate speech and abusive-language corpus, this study applies a mixed computational-pragmatic design combining source labels with manual interpretation of target, severity, facework, contextual validation, and risk. Results: Findings show that toxicity classes are evenly distributed, but target, domain, and severity differentiate surface abuse from identity-targeted harm. Facework cues such as direct address, laughter, ridicule, and abusive labels appear unevenly across toxicity classes, demonstrating that politeness and impoliteness cannot be inferred from lexical toxicity. Cultural-pragmatic coding shows that religion, ethnicity, gender, platform traces, political indexing, and local-language ambiguity require contextual validation before model outputs can be treated as reliable. Implication: Indonesian NLP moderation should combine classification with pragmatic annotation. Novelty: This study reframes toxicity detection as culturally grounded, target-aware, and pragmatically validated harm interpretation.]]>
Copyrights © 2026
