<![CDATA[The increasing volume of information has led to the phenomenon of information overload, a condition where individuals struggle to filter and comprehend information efficiently within a limited time. To address this issue, automatic text summarization serves as an essential approach. This research aims to assess effectiveness of two transformer-based models, IndoT5 and mBART, by comparing their ability to generate abstractive summaries (highlight) of Indonesian news articles. The abstractive approach allows models to generate new sentences with more natural language structures compared to extractive methods. Fine-tuning for both models was conducted using a dataset comprising 10,410 news articles from Tempo.co, each containing full news content and a corresponding highlight used as a reference. ROUGE and BERT-Score metrics were employed in the evaluation process to assess structural and semantic correspondence between the references and the generated summaries. Results show that IndoT5 outperformed in terms of ROUGE-1 (0.43087), ROUGE-2 (0.29143), ROUGE-L (0.39224), BERT-Score Recall (0.89130), and F1 (0.87708), indicating its capability to generate complete and relevant news highlight. Meanwhile, mBART achieved a higher BERT-Score Precision (0.86717) but tended to generate less informative outputs. The findings of this research are expected to aid in enhancing the coherence and efficiency of abstractive summarization systems.]]>
Copyrights © 2025
