<![CDATA[The increasing number of online news articles published every day has created challenges for readers in obtaining important information efficiently. Automatic text summarization offers a solution by generating concise versions of news articles while preserving essential information. This study aims to compare the performance of extractive and abstractive text summarization models for Indonesian-language news articles, namely IndoBERT, T5-Base, and mT5 XL-Sum. The research employed a quantitative experimental approach using the IndoSum dataset, with 200 news articles from the test split selected as evaluation samples. IndoBERT was implemented as an extractive model through binary sequence classification, while T5-Base and mT5 XL-Sum were evaluated using a sequence-to-sequence generative approach. Model performance was measured using ROUGE-1, ROUGE-2, and ROUGE-L metrics to assess information coverage, syntactic quality, and sequence coherence. The results show that T5-Base achieved the highest performance with ROUGE-1 of 0.5686, ROUGE-2 of 0.4950, and ROUGE-L of 0.5365. IndoBERT obtained competitive results with ROUGE-1 of 0.4377, ROUGE-2 of 0.3291, and ROUGE-L of 0.3784, while mT5 XL-Sum produced the lowest scores among the evaluated models. These findings indicate that a monolingual abstractive model specifically trained for Indonesian language tasks is more effective than extractive and multilingual approaches for Indonesian news summarization. The study contributes empirical evidence regarding the suitability of Transformer-based summarization models and can serve as a reference for future development of automatic summarization systems for Indonesian-language content.]]>
Copyrights © 2026
