PANCASILA-BASED ASPECT CATEGORY SENTIMENT ANALYSIS FOR DETECTING NEGATIVE CONTENT IN INDONESIAN CODE-MIXED SOCIAL MEDIA
TL;DR
A Pancasila-based Aspect Category Sentiment Analysis framework grounded in Indonesia’s five foundational values is proposed, offering a culturally grounded approach to AI-assisted content moderation in Indonesia.
Abstract
Hate speech and value-violating content on Indonesian social media, compounded by code-mixed language, threaten social cohesion. This study proposes a Pancasila-based Aspect Category Sentiment Analysis framework grounded in Indonesia’s five foundational values: Divinity, Humanity, Unity, Democracy, and Social Justice. Four Transformer models were evaluated under Full Fine-Tuning, LoRA, and QLoRA on 41,138 Indonesian code-mixed texts (confidence >= 0.75), annotated via zero-shot LLM inference and validated by two independent experts (k = 0.82; LLM-expert k = 0.81). IndoBERT LoRA achieved the highest in-pipeline F1-Score (0.77), though bootstrap intervals show this isstatistically indistinguishable from several top configurations. Againstan independent expert-validated ground truth (n = 1,000), all models dropped 6.7% on average; IndoBERTweet QLoRA obtained the highest point-estimate generalization (GT F1 = 0.71, 10.27 MB adapter storage), best read as the top of a statistically indistinguishable cluster rather than a confirmed single best model. The framework offers a culturally grounded approach to AI-assisted content moderation in Indonesia.