Klasifikasi Multi-Label Dan Ekstraksi Entitas Pada Ulasan Aplikasi Blu by BCA Digital Menggunakan IndoBERT

Abstract

The growth of digital banking services in Indonesia has heightened the need to understand factors influencing users' application continuance intention. However, prior studies remain limited to single-label classification and general sentiment analysis, lacking the ability to capture the complexity of information in Indonesian-language user reviews in a structured manner. This study aims to perform multi-label classification based on four Expectation-Confirmation Model (ECM) factors—Confirmation, Perceived Usefulness, E-satisfaction, and Perceived Security—and to extract six Named Entity Recognition (NER) entities from Blu by BCA Digital application reviews using IndoBERT. The dataset was collected from Google Play Store and Apple App Store covering January to December 2025, yielding 3,389 Indonesian-language reviews after filtering. The study employs a single-task approach, applying oversampling and Focal Loss for multi-label classification, and token augmentation with Conditional Random Field (CRF) for NER. Annotation validation using Krippendorff's Alpha yielded average values of 0.856 for intent labels and 0.919 for NER entities. Results show that the best classification model achieved an F1-Score of 0.798 with a Hamming Loss of 0.131, while the best NER model achieved an F1-Score of 0.812. This study demonstrates that IndoBERT is effective for analyzing digital banking application reviews in identifying ECM factors and extracting domain-specific entities, thereby offering potential support for developers in automatically understanding user needs.