Text sentiment analysis is a fundamental task in natural language processing, with extensive applications in social media monitoring, market research, and public opinion analysis. This study systematically investigates the performance of classical and deep learning models for this task. The methodology begins by employing Word2Vec technology to convert textual data into dense word vectors, effectively capturing semantic relationships within the data. These generated embeddings are then used as input for four distinct classification algorithms: Support Vector Machine (SVM), Recurrent Neural Network (RNN), Long Short-Term Memory network (LSTM), and Gated Recurrent Unit (GRU). Comparative experiments were conducted on a unified dataset to ensure a fair evaluation of each model's efficacy in text classification. The results demonstrate a clear trade-off: the SVM model delivers robust and computationally efficient performance on smaller-scale datasets, serving as a strong baseline. In contrast, the more complex LSTM and GRU models excel on larger datasets. Their gated architectures are specifically designed to overcome the limitations of simple RNNs by effectively capturing long-range contextual dependencies and temporal features within sequences, which significantly enhances classification accuracy. The analysis concludes that model selection is highly dependent on data scale and computational resources. Furthermore, it proposes promising future optimization directions, suggesting that hybrid approaches which integrate the feature extraction strengths of SVMs with the sequential modeling power of advanced recurrent networks like LSTM could potentially yield superior performance.