Hybrid Lexicon-Machine Learning Approach for Aspect-Based Sentiment Analysis on Large-Scale Indonesian App Reviews: A Case Study of Telegram

Abstract

Mobile application reviews provide critical unstructured feedback for developers, yet processing large-scale Indonesian app reviews remains hindered by informal language noise and severe class imbalance. While deep learning offers strong context modeling, its computational burden is impractical for industrial-scale deployment exceeding hundreds of thousands of entries. This study proposes a scalable, computationally efficient Hybrid Lexicon-Machine Learning framework for Aspect-Based Sentiment Analysis (ABSA) on large-scale Indonesian user reviews, using Telegram as a case study. We processed 399,527 reviews from the Google Play Store through a pipeline integrating TF-IDF n-grams, explicit domain-adapted lexicon scoring (incorporating negation and intensifier heuristics), and review length metadata into a LightGBM classifier. To overcome the scarcity and linguistic ambiguity of neutral ratings, a binary classification strategy (Positive vs. Negative/Critical) was implemented. The model achieved Accuracy of 80.4%, Macro-F1 of 76.7%, and AUC-ROC of 0.855 on a strictly held-out test set, peaking at 0.933 for Privacy. Ablation experiments confirm that incorporating explicit lexicon scores consistently improves Macro-F1 by up to 5.1% over standalone TF-IDF. Aspect distribution revealed critical bottlenecks in Privacy (73.0% negative) and Notification (66.1% negative), providing actionable roadmaps for software maintenance.