<![CDATA[Indonesian Supreme Court decisions on Human Trafficking (TPPO) are lengthy and structurally complex, rendering manual review inefficient for legal practitioners. Existing abstractive summarization research for Indonesian text concentrates on news and social media, while no publicly benchmarked XSum-style dataset exists for the Indonesian legal domain. This study has two explicit objectives: (i) an XSum-structured legal summarization dataset is constructed from 404 TPPO decisions, and (ii) four fine-tuned Transformer models (T5 Base Indonesia, mT5 Small, DistilBART CNN, BART Large XSum) are benchmarked against extractive and classical abstractive baselines. The method couples an n8n-based PDF extraction pipeline with CSV-sourced verdict statements as reference summaries, followed by fine-tuning and evaluation using ROUGE-1/2/L and BERTScore F1, complemented by paired bootstrap significance testing (n=10,000). Results show T5 Base Indonesia attains the highest ROUGE-L of 39.49 and BERTScore F1 of 74.82, while mT5 Small achieves the highest ROUGE-1 of 44.97, all significantly outperforming Seq2Seq+Attention (ROUGE-1 27.31) and First-2-Sentences (ROUGE-1 10.86) with p<0.001. The contributions are: an XSum-formatted TPPO dataset, an automated extraction pipeline, and a comprehensive benchmark spanning extractive, classical abstractive, and Transformer-based methods. These findings offer practical benefits for legal document analysis and judicial information retrieval in Indonesia.]]>
Copyrights © 2026
