Skripsi

FINE TUNING INDOBERT UNTUK ANALISIS SENTIMEN PADA ULASAN PENGGUNA APLIKASI ZALORA DI GOOGLE PLAY STORE

---

This study investigates sentiment in user reviews of the Zalora Android application on Google Play using a hybrid approach that fine-tunes IndoBERT and integrates a TF-IDF Logistic Regression ensemble with inference-time adjustments. A corpus of 29,066 reviews was preprocessed and split with a stratified hold-out; class imbalance was handled on the training partition. The pipeline applies probability calibration and neutral-class thresholding to stabilize decisions where polarity is ambiguous. On the test set (20%), the best model achieved 92.15% accuracy with 0.92 macro-F1, showing balanced performance across positive, neutral, and negative classes. Ablation indicates that neutral-thresholding and calibration reduce over-confidence and improve minority-class consistency, while the lightweight TF-IDF ensemble adds robustness on out-of-domain phrasing. Error analysis highlights confusions concentrated near rating-3 reviews, aligning with expected ambiguity. These findings demonstrate that a calibrated, threshold-aware IndoBERT pipeline yields reliable sentiment classification for Indonesian e-commerce reviews and offers actionable signals for product and service improvement.

---

Availability

Detail Information

Series Title

-

Call Number

T1894642025

Publisher

Indralaya : Prodi Sistem Informasi, Fakultas Ilmu Komputer Universitas Sriwijaya., 2025

Collation

xiv, 87 hlm.; ilus.; tab.; 29 cm.

Language

Indonesia

ISBN/ISSN

-

Classification

006.350 7

Content Type

Text

Media Type

-

Carrier Type

-

Edition

-

Subject(s)

Prodi Sistem Informasi
Analsis Sintemen

Specific Detail Info

-

Statement of Responsibility

MI

Other version/related

File Attachment

Comments

---

You must be logged in to post a comment