Topic modeling from journal article titles is challenging because titles are short and provide limited word co occurrence and contextual information. This study compares Latent Dirichlet Allocation (LDA) and BERTopic for identifying computer science research topics and examining their temporal trends. The corpus comprises 39,532 titles published between 2015 and 2025 across 283 Indonesian journals indexed in the Science and Technology Index (SINTA). Of these, 46.2% were originally written in English and translated into Indonesian before modeling. Both approaches were evaluated using the same corpus, vocabulary, topic coherence (C_v), topic diversity, and three random seeds. LDA was tested across topic counts from 5 to 30, while BERTopic used multilingual sentence embeddings, UMAP, and HDBSCAN and was reduced to 25 topics for comparison. Temporal trends for 2015–2024 were assessed using three statistical tests and robustness checks for journal composition, SINTA ranking, and original title language. At 25 topics, BERTopic achieved higher coherence (0.671, SD = 0.017) and diversity (0.841) than LDA (0.580, SD = 0.039; diversity = 0.693), but left 37.9% of titles unassigned. LDA assigned all titles and produced a more balanced topic-size distribution. Agreement between the partitions was limited (NMI = 0.31), with 11 of 25 LDA topics having one-to-one counterparts. Seven method–topic combinations passed all reported trend checks, including sentiment analysis, deep learning, and IoT monitoring, whereas some steep declines reflected changes in journal composition. The findings reveal a trade-off between topic quality and document coverage and indicate that research trends should be interpreted only when they remain robust across methods and corpus-composition checks.
Copyrights © 2026