Studi Komparatif Skenario Fine-Tuning Model Transformer untuk Analisis Sentimen Teks Twitter Berbahasa Indonesia

Akmal, Fadhl and Ma'ruf, Muhammad Rifqi (2025) Studi Komparatif Skenario Fine-Tuning Model Transformer untuk Analisis Sentimen Teks Twitter Berbahasa Indonesia. Project Report. [s.n], [s.l.]. (Unpublished)

Text 5025221028_5025221060-Project_Report.pdf - Accepted Version Download (1MB)

Abstract

Pertumbuhan konten digital berbahasa Indonesia yang pesat menuntut adanya sistem analisis sentimen yang akurat dan efisien. Model transformer berbasis transfer learning telah menunjukkan potensi besar, namun strategi fine-tuning yang paling efektif untuk konteks bahasa Indonesia masih memerlukan eksplorasi lebih lanjut. Penelitian ini melakukan studi komparatif untuk mengevaluasi tiga skenario fine-tuning—Fine-Tuning Standar, Gradual Unfreezing, dan Differential Learning Rates—pada tiga arsitektur model: IndoBERTbase, IndoBERTweet, dan RoBERTa. Pengujian dilakukan pada dua dataset dengan domain berbeda, yaitu ulasan aplikasi (Dataset BBM) dan komentar politik (Dataset Pemilu), dengan metrik evaluasi utama F1-Score. Hasil penelitian menunjukkan bahwa IndoBERTweet secara konsisten menjadi model dengan performa tertinggi di semua skenario, mencapai F1-Score puncak 0.9218 pada Dataset BBM dan 0.7431 pada Dataset Pemilu. Temuan kunci lainnya adalah strategi Fine-Tuning Standar dengan learning rate yang teroptimasi terbukti lebih unggul dibandingkan dua teknik lanjutan yang lebih kompleks. Penelitian ini menyimpulkan bahwa keselarasan domain antara data pre-training model dengan data tugas akhir merupakan factor krusial untuk mencapai performa tinggi, dan sebuah metode yang lebih sederhana namun teroptimasi dengan baik dapat lebih efektif daripada strategi yang kompleks.
====================================================================================================================================
The rapid growth of Indonesian digital content necessitates an accurate and efficient sentiment analysis system. Transfer learning-based transformer models have shown great potential, yet the most effective fine-tuning strategies for the Indonesian context require further exploration. This research conducts a comparative study to evaluate three fine-tuning scenarios—Standard Fine-Tuning, Gradual Unfreezing, and Differential Learning Rates—on three model architectures: IndoBERTbase, IndoBERTweet, and RoBERTa. The models were tested on two datasets with different domains: application reviews (BBM Dataset) and political comments (Pemilu Dataset), using the F1-Score as the primary evaluation metric. The results consistently show that IndoBERTweet achieved the highest performance across all scenarios, reaching a peak F1-Score of 0.9218 on the BBM Dataset and 0.7431 on the Pemilu Dataset. Another key finding is that the Standard Fine-Tuning strategy with an optimized learning rate proved superior to the two more complex advanced techniques. This study concludes that domain alignment between the model's pre-training data and the final task data is a crucial factor for achieving high performance. Furthermore, a simpler yet well-optimized method can be more effective than complex strategies.

Item Type:	Monograph (Project Report)
Uncontrolled Keywords:	Analisis Sentimen, Transfer Learning, Fine-Tuning, BERT, IndoBERT, IndoBERTweet, RoBERTa, Pemrosesan Bahasa Alami, Bahasa Indonesia Sentiment Analysis, Transfer Learning, Fine-Tuning, BERT, IndoBERT, IndoBERTweet, RoBERTa, Natural Language Processing, Indonesian Language.
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Rifqi Ma'ruf
Date Deposited:	23 Jul 2025 01:30
Last Modified:	23 Jul 2025 01:30
URI:	http://repository.its.ac.id/id/eprint/120632

Actions (login required)

View Item