Sarcasm and Irony Detection in Lazada App Reviews Using IndoBERT

Abstract

Digital technology has reshaped consumer behavior, particularly in e-commerce, where Google Play Store reviews provide rich feedback but often include sarcasm and irony that conventional sentiment models misread. This study proposes an Indonesian sarcasm–irony detection model using IndoBERT, a transformer pre-trained on Indonesian corpora. A dataset of 1,998 Lazada app reviews was collected via web scraping and preprocessed through text cleaning, tokenization, and stopword removal with the Sastrawi library. IndoBERT was fine-tuned to classify reviews into three classes: sarcasm, irony, and literal. Performance was assessed using accuracy, precision, recall, F1-score, and a confusion matrix. The model achieved 96.40% accuracy, with F1-scores of 0.9725 (sarcasm), 0.9675 (irony), and 0.9267 (literal). Word cloud visualizations revealed distinct lexical patterns across classes, supporting IndoBERT’s ability to capture contextual cues behind implicit sentiment. The findings indicate IndoBERT is effective for advanced opinion mining in Indonesian e-commerce, with potential applications in customer feedback monitoring, surfacing hidden complaints, and improving recommendation systems beyond surface polarity. Limitations include reliance on a single platform (Google Play) and text-only input, without modeling non-textual signals such as emojis or punctuation intensity. Future work should test cross-platform generalization, incorporate non-textual cues, and apply data augmentation to reduce class imbalance, particularly for the less frequent literal class, to improve robustness for real-world deployment