M. Noer Fadli Hidayat, Didik Dwi Prasetya, Triyanna Widiyaningtyas
Sentiment analysis of Indonesian user reviews is challenged by informal language and frequent nonverbal cues such as Unicode emojis and ASCII emoticons. This study aimed to quantify the benefit of explicitly modeling ASCII emoticons together with Unicode emojis and text for three-class sentiment classification (negative/neutral/positive). SIEmo-LSTM is a tri-modal pipeline that (i) maps Unicode emojis using an emoji sentiment resource, (ii) detects, normalizes, and converts ASCII emoticons into descriptive tokens using Emot and LEED (93.3% successful conversion), and (iii) encodes the unified sequence using IndoBERT as a contextual feature extractor and refines it with a Bi-LSTM layer before multiclass prediction. Experiments used 304,570 Ruangguru app reviews (2016–2023), a tri-modal subset of 2,527 reviews, and a 70/20/10 train/validation/test split. Class imbalance was addressed using Random OverSampling (ROS). The full Text+SE+IE configuration with ROS achieved up to 0.9935 Accuracy and 0.9967 Macro-F1, outperforming text-only and text+Unicode baselines, while Random UnderSampling (RUS) consistently degraded performance. These findings imply that treating ASCII emoticons as a first-class affective modality—alongside Unicode emojis and text—improves robustness and class-balanced sentiment recognition for Indonesian user-generated reviews. © (2026), (Dr D. Pylarinos). All rights reserved.
Department of Electrical Engineering and Informatics, Universitas Negeri, Malang, Indonesia; Informatics Engineering Department, Faculty of Engineering, Universitas Nurul Jadid, Indonesia
