This study investigates sentiment-aware deep learning models for short-term stock price forecasting using NVIDIA (NVDA) as a representative high-volatility technology stock. Four architectures—CNN, LSTM, BiLSTM, and a hybrid CNN–BiLSTM—are evaluated under two configurations: without sentiment and with FinBERT-based financial news sentiment integrated as a continuous contextual feature. Historical OHLV data are combined with sentiment information to enable multimodal learning under a controlled experimental setting. The results demonstrate that recurrent architectures consistently outperform convolution-only models, highlighting the importance of temporal dependency modeling in financial time series. Among all configurations, the hybrid CNN–BiLSTM with FinBERT sentiment achieves the best overall performance, yielding the highest R², the lowest MAE and RMSE, and the smallest overfitting gap. Bootstrap-based confidence intervals indicate stable generalization, while Wilcoxon signed-rank tests confirm that the observed performance improvements are statistically significant. The study also presents a near real-time deployment framework with low inference latency, demonstrating practical applicability for decision-support systems. Overall, the findings show that effective alignment between local feature extraction, bidirectional temporal modeling, and contextual sentiment integration is critical for improving stock price. forecasting accuracy and robustness.