<p>Sentiment analysis, an essential task in natural language processing, has evolved rapidly with the emergence of deep learning architectures. Although numerous studies have applied models such as CNNs, RNNs, and Transformers to sentiment classification, a comprehensive comparative synthesis across these architectures remains limited. Existing reviews often focus on specific model families or narrow application domains, leaving a gap in understanding how different deep learning paradigms collectively influence sentiment analysis performance. This paper addresses this gap by presenting a systematic review and comparative analysis of deep learning-based sentiment analysis techniques published between 2020 and 2025. The study examines major architectures—CNN, RNN, attention-based, transformer, hybrid, and graph-based models—covering their design, datasets, performance metrics, and limitations. Furthermore, it proposes a unified taxonomy of sentiment analysis models and identifies open research challenges related to data imbalance, cross-domain adaptation, multilingual processing, and model interpretability. The review aims to provide researchers with an integrated perspective on current advancements while outlining promising directions for future sentiment analysis research.</p>