Aspect-Based Sentiment Analysis (ABSA) has gained increasing attention due to its wide applicability in domains such as marketing, e-commerce, and social media analytics. Nevertheless, research on ABSA in Vietnamese remains limited, especially in the context of e-commerce product reviews, which restricts the development of practical sentiment analysis applications for this language. To address this gap, this study aims to develop and evaluate hybrid models that integrate PhoBERT, a pre-trained language model optimized for Vietnamese, with classical machine learning techniques to enhance aspect-level sentiment classification. Two approaches are proposed: (i) combining PhoBERT embeddings with a Naive Bayes classifier and (ii) integrating PhoBERT with a Hidden Markov Model (HMM) to capture sequential sentiment dynamics. Experiments conducted on a Vietnamese product review dataset collected from Tiki, a leading e-commerce platform, show that the PhoBERT + Naive Bayes model achieves strong performance with an accuracy of 94.54% and an F1-score of 0.95, making it well suited for real-time and interpretable sentiment analysis. In contrast, the PhoBERT + HMM model underperforms with only 32.96% accuracy, exhibiting high inference time and prediction instability. These findings demonstrate the superiority of the PhoBERT + Naive Bayes model while highlighting the limitations of probabilistic sequence models like HMM when combined with transformer-based embeddings. The study contributes to advancing ABSA in low-resource languages and underscores the potential of hybrid modeling approaches to balance accuracy, scalability, and interpretability.