Sentiment analysis is a fundamental task in Natural Language Processing that focuses on identifying the emotional polarity of textual data. It plays an important role in extracting insights from domains such as e-commerce, customer feedback, and social media. Recent advances in pre-trained language models have significantly improved sentiment classification by enabling effective knowledge transfer through fine-tuning. In this study, we propose a novel approach to Vietnamese sentiment analysis that combines PhoBERT-V2, a state-of-the-art RoBERTa-based model for Vietnamese, with an extended version of Vietnamese SentiWordNet, a sentiment-specific lexical resource. Our proposed model, named CombViSA, integrates contextual embeddings from PhoBERT-V2 with sentiment-aware lexical features to overcome the limitations of existing sentiment lexicons. Experimental results on two benchmark datasets, VLSP 2016 and AIVIVN 2019, demonstrate that CombViSA achieves state-of-the-art performance, obtaining an F1-score of 0.95 on VLSP 2016 and significantly outperforming strong baselines. This work contributes to advancing sentiment analysis for low-resource languages and supports real-world applications such as opinion mining, brand monitoring, and customer experience analysis.