Analyzing Apple, Samsung, and Xiaomi Consumer Reviews (2021–2024): Aspect Classification and Sentiment Analysis using BERT, RoBERTa, and DistilBERT

Abstract

The exponential growth of online consumer reviews has amplified the need for effective analytical techniques to accurately capture user sentiment toward specific product features. This research provides a detailed analysis of Apple, Samsung, and Xiaomi consumer reviews collected from major e-commerce platforms from 2021 to 2024, employing transformer-based Natural Language Processing (NLP) models BERT, RoBERTa, and DistilBERT. Specifically,

two critical tasks were conducted separately: Aspect Classification, identifying eight pre- defined product aspects (Battery, Performance, Camera, Display, Price, Multimedia, Fitness &

Health Tracking, and Customization), and Overall Sentiment Classification, assigning each review a general polarity (positive, neutral, negative). Results indicate that RoBERTa achieved the highest accuracy (F1 score of 0.791) in aspect classification tasks, effectively handling nuanced linguistic variations within reviews. In overall sentiment classification, BERT provided the most balanced performance across sentiment classes, whereas DistilBERT, despite lower accuracy, significantly reduced computational overhead with the fastest training times (approximately 6.3 minutes), highlighting an important trade-off between speed and accuracy. Temporal analysis of over 20,000 reviews from 2021 to 2024 highlights distinct sentiment trends across brands. Apple consistently achieved high user satisfaction, particularly for Performance and Camera aspects, reflecting stable positive sentiment. Samsung demonstrated significant improvements in Fitness & Health Tracking, aligning with evolving consumer preferences for smartwatch features. Xiaomi, while competitive on Price, showed variable sentiment for Battery and Performance, indicating areas for targeted improvement. These findings, derived from a modular approach separating aspect classification and overall sentiment analysis, provide granular insights into consumer preferences over time. This research contributes methodologically by decoupling aspect classification from sentiment analysis, thereby improving transparency and reducing model complexity. While product aspects were extracted using transformer-based models, sentiment analysis was conducted only at the overall review level. This design choice improves scalability and annotation consistency. Practically, the findings offer actionable insights for product improvement, customer satisfaction management, and competitive analysis within the consumer electronics market.