Abstract
This study presents a framework for conducting sentiment analysis using deep learning techniques to evaluate restaurant food reviews collected from FoodPanda, a food delivery platform. We provide BDFoodSent, a unique large-scale dataset that includes over 334,000 reviews from FoodPanda in Bangladesh. These reviews were annotated for both binary (positive/negative) and multi-class (very bad, bad, neutral, good, and very good) sentiment classifications. Bag of Words (BoW) and TF-IDF feature extraction techniques were used to assess the traditional machine learning (ML) models. The models included decision trees, support vector machines, random forests, XGBoost, and k-nearest neighbors. To solve the issue of class imbalance, we used random oversampling, which resulted in a considerable improvement in the performance of the multi-class classification model. In addition, we present a hybrid deep learning architecture called H-Food. The design included embedding, BiLSTM, attention, and convolutional layers. The results of the experiments show that the proposed model achieved an accuracy of 91.42% and an F1-score of 91.35% when applied to binary classification tasks. Additionally, it achieved an accuracy of 78.74% and an F1-score of 78.72% when applied to multi-class classification using oversampled data. Comparative evaluations showed that the proposed architectures outperformed the classical models and baseline CNN design. The findings of this study demonstrate the efficacy of deep hierarchical networks in comprehending the feelings of users and provide insights that may be applied by restaurant services in emerging countries in the future.