FATHURRAHMAN, Alif Hafian and Kusumaningrum, Retno and Farikhin, Farikhin (2025) UMPAN BALIK OTOMATIS UNTUK PERTANYAAN URAIAN SINGKAT BERBAHASA INDONESIA MENGGUNAKAN BIDIRECTIONAL AUTO REGRESSIVE FROM TRANSFORMERS (BART) DENGAN TEKNIK AUGMENTASI. Masters thesis, UNIVERSITAS DIPONEGORO.

	Text COVER-1.pdf Download (182kB)
	Text COVER.pdf Restricted to Repository staff only Download (2MB)
	Text BAB 1.pdf Download (356kB)
	Text BAB 2.pdf Download (1MB)
	Text BAB 3.pdf Restricted to Repository staff only Download (879kB)
	Text BAB 4.pdf Restricted to Repository staff only Download (1MB)
	Text BAB 5.pdf Restricted to Repository staff only Download (310kB)
	Text DAFTAR PUSTAKA.pdf Download (493kB)
	Text LAMPIRAN.pdf Restricted to Repository staff only Download (896kB)

Abstract

Penelitian ini berhasil menghasilkan sistem umpan balik otomatis untuk jawaban soal uraian singkat berbahasa Indonesia dengan membandingkan performa dua model pre-trained, BART dan T5. Proses pelatihan menggunakan fine-tuning dan penambahan data melalui data augmentation dengan teknik few-shot prompting berbantuan ChatGPT, sehingga variasi jawaban menjadi lebih beragam. Model BART yang di-fine-tune dengan data hasil augmentasi terbukti mencapai hasil paling optimal dengan skor ROUGE-1 precision 67,74 %, ROUGE-2 precision 50,11 %, ROUGE-L F1 49,72 %, dan BERTScore F1 87,19 %. Validasi human-centered dengan dua penilai menunjukkan ICC sebesar 0,585 (kategori moderat) dan rata-rata skor Likert 5,43, mendukung relevansi umpan balik yang dihasilkan. Sistem ini juga telah diimplementasikan sehingga mampu memberikan umpan balik yang mendekati keluaran evaluator manusia. Temuan ini menegaskan pentingnya strategi transfer learning, fine-tuning, dan augmentasi data berbasis ChatGPT untuk meningkatkan kualitas umpan balik otomatis, sekaligus menjadi dasar pengembangan evaluasi pembelajaran yang lebih efektif di masa depan.
Kata kunci: umpan balik otomatis,BART, T5, ChatGPT, data augmentation, ROUGE, BERTScore,bahasa Indonesia.

This study successfully produced an automatic feedback system for short essay questions in Indonesian by comparing the performance of two pre-trained models, BART and T5. The training process used fine-tuning and data augmentation with ChatGPT-assisted few-shot prompting techniques, resulting in a wider variety of answers. The BART model fine-tuned with augmented data was proven to achieve optimal results with a ROUGE-1 precision score of 67.74%, ROUGE-2 precision of 50.11%, ROUGE-L F1 of 49.72%, and BERTScore F1 of 87.19%. Human- centered validation with two raters showed an ICC of 0.585 (moderate category) and an average Likert score of 5.43, supporting the relevance of the feedback generated. This system has also been implemented so that it is able to provide feedback close to the output of human evaluators. These findings underscore the importance of ChatGPT-based transfer learning, fine-tuning, and data augmentation strategies to improve the quality of automated feedback, while also providing a basis for developing more effective learning evaluations in the future.
Keywords : automatic feedback, BART, T5, ChatGPT, data augmentation, ROUGE,BERTScore, Indonesia Languange

Actions (login required)

View Item