Halvino Iqbal Nataprawira, Ida Nurhaida

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

Introduction

Indobert-based sentiment analysis of indonesian social media discourse on ai-generated images. Explore Indonesian social media discourse on AI-generated images. This study uses IndoBERT for sentiment analysis, outperforming classical models and revealing a critical public opinion.

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.

Review

This study addresses a highly pertinent and contemporary issue: the public sentiment surrounding AI-generated images within the Indonesian social media landscape. Recognizing the challenge posed by the contextual complexities of informal Indonesian language in accurately mapping public reaction, the research sets a clear objective to evaluate the performance of contextual language models, specifically IndoBERT, against traditional machine learning classifiers. The timely nature of this topic, coupled with the identified gap in understanding sentiment structures, establishes a strong foundation for the research's significance in contributing to both natural language processing and the broader discourse on AI's societal impact. The methodology employed is robust, involving the collection and manual annotation of a multi-platform dataset of 2,981 Indonesian-language posts from X, Reddit, and YouTube, categorized into positive, neutral, and negative sentiments. Crucially, the authors address the common issue of class imbalance using appropriate techniques: SMOTE for classical models and class-weighted loss with Masked Language Modeling augmentation for IndoBERT. The performance evaluation, utilizing macro-averaged F1-score across repeated stratified random splits, provides a reliable measure of model efficacy. The findings unequivocally demonstrate IndoBERT's superior performance, achieving a mean macro-F1 of 0.7131 ± 0.0180 and outperforming the best classical baseline by a substantial margin of approximately 0.12, particularly excelling in resolving ambiguous neutral discourse. Furthermore, the revelation that negative sentiment heavily dominates the corpus (61.8%) offers a significant insight into prevailing public concerns regarding ethical and copyright issues related to AI-generated imagery. In conclusion, this research successfully establishes a valuable empirical baseline for Indonesian sentiment analysis, unequivocally demonstrating the pronounced advantage of transformer architectures for nuanced public opinion mining in this linguistic context. The study's careful handling of data annotation, class imbalance, and rigorous evaluation methodology strengthens its findings, providing a credible benchmark for future work. The identification of a predominantly critical public stance, driven by ethical and copyright concerns, serves as a crucial insight for policymakers, AI developers, and creators alike. This work not only advances the field of Indonesian NLP but also contributes significantly to understanding the socio-technical implications of generative AI.

Full Text

Comments

You need to be logged in to post a comment.