Context-Aware Sentiment Analysis of E-Commerce Reviews Using a BERT-CNN-BiLSTM Hybrid Model

Abstract

User-generated product reviews are an essential source of information in e-commerce; nevertheless, the huge volume and varying quality of review texts make extracting insights difficult. The conventional approach to sentiment classification is limited in terms of recognizing contextual and aspect-oriented sentiment clues in the text. This study proposes a hybrid architecture that uses contextual sentiment clues for e-commerce reviews sentiment classification. The experiments are conducted using FABSA dataset which contains reviews annotated in terms of multiple aspects–sentiments pairs. Every review is expanded in terms of aspect-oriented sentiment classification samples. This makes it possible to learn fine-grained polarities of sentiments associated with specific aspects of product reviews. To tackle the problem of class imbalance in the data, the experiments employ the method of stratified oversampling in combination with the use of class-weighted cross-entropy loss function. The empirical results show that the improved hybrid BERT–CNN–BiLSTM model achieves 92.20% validation accuracy, 92.25% weighted F1-score, and 86.36% macro F1-score. The most noticeable progress has been made in the case of neutral class, where the F1-score has increased by 17.1 percentage points showing the improvements in minority-class recognition and decrease of class imbalance. The architecture-level ablation study shows the superior performance of the BERT–CNN–BiLSTM model compared to its simplified versions based on BERT, BERT–CNN, and BERT–BiLSTM architectures. A contextual comparison with reported FABSA baselines suggests competitive performance, although this comparison should not be interpreted as a strict leaderboard result because the baselines were not reproduced under identical experimental settings. Overall, the findings demonstrate that combining contextual transformer representations with local convolutional features and bidirectional sequential modeling can improve class-balanced sentiment classification in aspect-expanded e-commerce review data.