Repository Institusi

Perbandingan kinerja model Distilbert dan Albert dalam deteksi emosi teks

Mursyid, Shafwan (2026) Perbandingan kinerja model Distilbert dan Albert dalam deteksi emosi teks. Sarjana thesis, UIN Sunan Gunung Djati Bandung.

[img]
Preview
Text
1_cover.pdf

Download (2MB) | Preview
[img]
Preview
Text
2_abstrak.pdf

Download (2MB) | Preview
[img]
Preview
Text
3_skbebasplagiarism.pdf

Download (513kB) | Preview
[img]
Preview
Text
4_daftarisi.pdf

Download (2MB) | Preview
[img]
Preview
Text
5_bab1.pdf

Download (2MB) | Preview
[img] Text
6_bab2.pdf
Restricted to Registered users only

Download (2MB) | Request a copy
[img] Text
7_bab3.pdf
Restricted to Registered users only

Download (2MB) | Request a copy
[img] Text
8_bab4.pdf
Restricted to Registered users only

Download (2MB) | Request a copy
[img] Text
9_bab5.pdf
Restricted to Registered users only

Download (2MB) | Request a copy
[img] Text
10_daftarpustaka.pdf
Restricted to Registered users only

Download (2MB) | Request a copy
[img] Text
11_lampiran.pdf
Restricted to Repository staff only

Download (25MB) | Request a copy

Abstract

NDONESIA: Emosi menjadi bagian yang selalu ada dalam komunikasi manusia, dan saat ini emosi tersebut diungkapkan lewat teks di media sosial, ulasan produk, maupun berbagai layanan digital lainnya. Kemampuan mendeteksi emosi pada teks menjadi penting untuk mendukung berbagai kebutuhan, mulai dari pemantauan sentimen publik sampai layanan pelanggan berbasis kecerdasan buatan. Model Transformer besar seperti BERT memang unggul dalam memahami konteks bahasa, tetapi ukurannya yang besar membuatnya kurang praktis diterapkan pada perangkat dengan sumber daya terbatas. Penelitian ini bertujuan mengimplementasikan dan membandingkan performa dua model Transformer ringan, yaitu DistilBERT dan ALBERT, dalam mendeteksi emosi teks berdasarkan metrik akurasi, presisi, recall, F1 Score, dan efisiensi komputasi. Penelitian dilakukan menggunakan dataset GoEmotions yang telah dipetakan menjadi lima kelas emosi, yaitu joy, sadness, anger, fear, dan surprise, dengan kedua model dilatih menggunakan pendekatan full fine-tuning mengikuti kerangka Machine Learning Life Cycle. Hasil pengujian pada 5.883 data uji menunjukkan DistilBERT unggul pada metrik klasifikasi agregat dengan akurasi 76,78% dan F1 Score 77,58%, sementara ALBERT memperoleh akurasi 76,20% dan F1 Score 76,91%. DistilBERT juga lebih cepat, baik dari sisi waktu pelatihan (662 detik) maupun waktu inferensi (3,02 milidetik per sampel), sedangkan ALBERT unggul dari sisi ukuran model yang jauh lebih kecil yaitu 46,75 MB dengan 11,69 juta parameter. Kedua model yang telah dilatih kemudian diterapkan pada aplikasi web demo berbasis FastAPI untuk menunjukkan hasil deteksi emosi secara langsung dan berdampingan. ENGLISH: Emotion has always been a part of human communication, and today it is increasingly expressed through text on social media, product reviews, and various other digital services. The ability to detect emotion in text has become important for supporting many needs, ranging from monitoring public sentiment to AI based customer service systems. Large Transformer models such as BERT are capable of understanding language context very well, but their large size makes them less practical to deploy on devices with limited computational resources. This research aims to implement and compare the performance of two lightweight Transformer models, DistilBERT and ALBERT, in detecting text emotion based on accuracy, precision, recall, F1 Score, and computational efficiency. The research was conducted using the GoEmotions dataset, which was mapped into five emotion classes, namely joy, sadness, anger, fear, and surprise, with both models trained using a full fine-tuning approach following the Machine Learning Life Cycle framework. Testing on 5,883 test samples shows that DistilBERT outperforms ALBERT on aggregate classification metrics, achieving an accuracy of 76.78% and an F1 Score of 77.58%, while ALBERT obtained an accuracy of 76.20% and an F1 Score of 76.91%. DistilBERT is also faster, both in training time (662.45 seconds) and inference time (3.02 milliseconds per sample), while ALBERT excels in model size, which is far smaller at 46.75 MB with only 11.69 million parameters. Both trained models were then deployed in a FastAPI based web demo application to display the emotion detection results side by side in real-time.

Item Type: Thesis (Sarjana)
Uncontrolled Keywords: DistilBERT; ALBERT; deteksi emosi teks; GoEmotions; Transformer
Subjects: Special Computer Methods > Artificial Intelligence
Divisions: Fakultas Sains dan Teknologi > Program Studi Teknik Informatika
Depositing User: SHAFWAN MURSYID MURSYID
Date Deposited: 19 Nov 2026 08:01
Last Modified: 19 Nov 2026 08:01
URI: https://digilib.uinsgd.ac.id/id/eprint/145682

Actions (login required)

View Item View Item