<![CDATA[Purpose: This study aims to evaluate whether integrating Gibbs Sampling into BERTopic improves topic modeling of short Indonesian social media texts concerning Artificial Intelligence. Baseline BERTopic is used as the primary comparator, while LDA and Top2Vec are included as external baselines representing probabilistic and embedding-based topic-modeling approaches. Methods: A corpus of 9,585 public posts from platform X, collected from January 2024 to May 2025, was modeled using Latent Dirichlet Allocation (LDA), Top2Vec, Baseline BERTopic, and Gibbs-BERTopic. Model performance was assessed using topic coherence, topic diversity, topic uniqueness, number of topics, and document coverage. The number of outliers was compared specifically between Baseline BERTopic and Gibbs-BERTopic because LDA and Top2Vec do not use an equivalent HDBSCAN-based outlier mechanism. Result: Baseline BERTopic generated 58 topics and 4,674 outliers, with coherence, diversity, and uniqueness scores of 0.480, 0.460, and 0.380. Gibbs-BERTopic generated 15 topics without outliers and achieved the highest corresponding scores of 0.510, 0.880, and 0.870. LDA and Top2Vec produced lower scores, with 58 and 84 topics, respectively. These findings indicate that, within the corpus and configuration examined, Gibbs-BERTopic provided a more coherent and distinctive topic representation, broader document coverage, and a more compact topic structure. Novelty: This study extends the evaluation of Gibbs-BERTopic to short Indonesian social media texts and compares it with probabilistic, embedding-based, and BERTopic-based models. The results highlight the importance of evaluating topic models using multiple complementary metrics rather than relying on a single performance measure.]]>
