LoRA-Enhanced Sentiment-Aware Topic Modeling for Indonesian Generative AI Perception

Abstract

Public understanding of generative AI in low-resource language contexts remains underexplored, particularly in relation to how sentiment aligns with thematic discussions on social media. In Indonesia, empirical studies examining this interaction at scale are still limited. This study introduces a sentiment-aware topic modeling framework that integrates parameter-efficient fine-tuning of IndoBERT using Low-Rank Adaptation with topic discovery via BERTopic. The approach enables large-scale analysis of Indonesian social media data under constrained computational settings. Analysis of Indonesian Twitter discourse shows that general discussions of Generative AI are largely neutral and cautious, contrasting with more optimistic trends reported in Western contexts. In comparison, enthusiast communities exhibit predominantly positive sentiment, while ethics-related discussions display balanced polarization. These results highlight the contextual nature of public perception across different discussion domains. The findings demonstrate the applicability of parameter-efficient NLP methods for sentiment and topic analysis in under-resourced languages and provide insights relevant to technology development and policy formulation.