Peringkasan Otomatis Risalah Rapat DPR RI Menggunakan IndoNanoT5 dan LongT5 TGlobal

Abstract

Meeting minutes of the Indonesian House of Representatives DPR RI are lengthy, dialogic, and distribute important information across many sections, making it difficult for readers to quickly identify core discussions, decisions, and follow up actions. This study develops and evaluates an abstractive summarization system for 200 DPR RI meeting minutes using a meeting domain based staged training strategy. Domain adaptation is performed through intermediate pre training on the AMI Meeting Corpus translated into Indonesian, followed by fine tuning on the DPR RI minutes dataset. Two approaches are compared. The first approach employs LongT5 with Transient Global Attention TGlobal as an end to end long context model. The second approach uses IndoNanoT5 with chunking and sliding window, a chunk size of 1024 tokens, 127 token overlap, and aggregation of 256 token chunk summaries into a final summary of up to 1024 tokens. Evaluation is conducted using ROUGE 1, ROUGE 2, ROUGE L, and Word F1 metrics. Validation results show that LongT5 TGlobal achieves the best performance after fine tuning, with ROUGE 1 of 0.3031, ROUGE 2 of 0.0924, ROUGE L of 0.1427, and Word F1 of 0.2721. IndoNanoT5 improves after fine tuning, achieving ROUGE 1 of 0.1437, ROUGE 2 of 0.0483, ROUGE L of 0.1098, and Word F1 of 0.1264, but remains affected by context fragmentation and repetition caused by overlap and summary aggregation. This approach supports improved accessibility of legislative meeting information for the public, researchers, journalists, and policymakers in sustainable democratic governance contexts.