A BERT-based Topic Modeling Approach for Indonesian Media Discourse Analysis

This research examines the use of BERTopic for topic modeling in Indonesian news articles sourced from seven prominent media outlets. A customized preprocessing pipeline was developed to address noise, multitopic content, and linguistic variation in Indonesian news datasets incorporating stemming, stopword removal, and lemmatization at the sentence level. IndoBERT embeddings were employed to obtain semantic representations, which were then reduced in dimensionality through UMAP and clustering through HDBSCAN. Experiments were performed with various topic configurations and the resulting models were assessed using UMass and NPMI coherence metrics. The findings demonstrated that the integration of IndoBERT embeddings, UMAP dimensionality reduction, and HDBSCAN clustering yielded superior coherence scores relative to other configurations. The 25 topic configuration produced an optimal balance between coherence and topic interpretability. The results confirm BERTopic’s suitability for complex Indonesian news corpora. This study improves the comprehension of topic modeling in low-resource languages and noisy corpora while also presenting avenues for future research that utilize domain-adapted models and supervised topic modeling methodologies.