IndoBERT-Based Sentiment Analysis of Indonesian Social Media Discourse on AI-Generated Images

Abstract

The rapid emergence of generative artificial intelligence has disrupted creative ecosystems, prompting widespread discourse across Indonesian social media. However, the exact sentiment structure of this public reaction remains empirically unmapped due to the contextual complexities of informal language. The objective of this research is to evaluate the efficacy of contextual language models by fine-tuning IndoBERT and benchmarking it against classical machine learning classifiers—including Complement Naive Bayes, Logistic Regression, and Support Vector Machine—for classifying social media sentiment. A multi-platform dataset comprising 2,981 Indonesian-language posts from X, Reddit, and YouTube was collected and manually annotated into positive, neutral, and negative classes. To address inherent class imbalance, Synthetic Minority Oversampling Technique was applied to classical models, while class-weighted loss and Masked Language Modeling augmentation were utilized for IndoBERT. Performance was evaluated using macro-averaged F1-score across five repeated stratified random splits. IndoBERT achieved a mean macro-F1 of 0.7131 ± 0.0180, outperforming the best classical baseline by approximately 0.12, demonstrating a pronounced advantage in resolving ambiguous neutral discourse. Negative sentiment heavily dominated the corpus at 61.8%, reflecting a prevailing critical stance toward AI-generated imagery concerning ethical and copyright issues. Furthermore, evaluation variance across random seeds exceeded variance from augmentation strategies, indicating test set composition is a major performance determinant. In conclusion, this study establishes a robust empirical baseline for Indonesian sentiment analysis, proving transformer architectures superior for nuanced public opinion mining.