Performance and Energy Efficiency Benchmarking of Sentiment Analysis Models for Jakarta Restaurant Reviews

Abstract

The food and beverage industry in Jakarta has been influenced by online reviews via digital platforms such as Google Maps, necessitating sentiment analysis in the automated way. Manual analysis of sentiment in the text is unfeasible, whereas models find it difficult to identify localized linguistic phenomena (slang, colloquial, code-switching) amid highly imbalanced data. This work explores the task accuracy and energy efficiency of models in classical machine learning (SVM, Naive Bayes, Logistic Regression), recurrent deep learning (LSTM, BiLSTM), and transformers (BERT, DistilBERT) approaches. The Google Place API was used to collect a dataset containing real Jakarta restaurant reviews. Energy use during inference and corresponding carbon emissions in kgCO2e were monitored precisely using the CodeCarbon library. The experimental results indicate that due to the extreme imbalance of the dataset, the total mode collapse was observed for BiLSTM and DistilBERT, confining their task accuracy in the range of the majority class at 0.7571. On the contrary, Logistic Regression managed to overcome the imbalance problem and achieved the best F1-Score of 0.7722 and 0.8143 accuracy score. Additionally, classical machine learning models consumed approximately 67% less energy than the full BERT model.