Comparative Sentiment Analysis of GrabFood Reviews Using BiLSTM and BiGRU

Abstract

The exponential growth of user-generated reviews on digital platforms has made manual sentiment interpretation of Online Food Delivery (OFD) services increasingly impractical. GrabFood, operating within the Grab ecosystem, has accumulated over 16.1 million reviews on the Google Play Store, necessitating an automated and scalable approach to sentiment monitoring. Conventional labeling approaches, including star-rating proxies and lexicon-based annotation, are inadequate for capturing contextual nuance, negation, and informal linguistic patterns prevalent in Indonesian-language OFD reviews. Furthermore, limited research has systematically compared BiLSTM and BiGRU architectures within a transformer-assisted labeling framework for Indonesian OFD sentiment analysis. This study aims to implement RoBERTa-based automatic sentiment labeling and to comparatively evaluate BiLSTM and BiGRU models for three-class sentiment classification of GrabFood reviews. A corpus of 265,500 raw reviews was collected via web scraping, filtered to 17,709 reviews through rigorous preprocessing, and annotated using the w11wo/indonesian-roberta-base-sentiment-classifier. Random Oversampling was applied to address class imbalance. BiLSTM and BiGRU models were trained and benchmarked against Support Vector Machine (SVM) and Naïve Bayes baselines. BiLSTM achieved 86% accuracy while BiGRU attained 85%, both substantially outperforming SVM (82%) and Naïve Bayes (77%). However, BiGRU demonstrated superior convergence speed and more stable per-class performance, particularly on the neutral category (F1: 51% vs. 50%). Transformer-assisted automatic labeling combined with bidirectional recurrent architectures constitutes an effective and scalable pipeline for Indonesian OFD sentiment classification, with neutral sentiment remaining the primary classification challenge.

GS Cited Analysis