Accurately forecasting stock prices remains a challenging task due to the complex interaction between market dynamics and external information such as financial news. Existing multimodal forecasting approaches often process news and market data independently and therefore fail to identify which news events are truly relevant to past price movements. To address this limitation, we propose a cross-attention-based approach that allows historical price information to guide the relevance weighting of financial news, explicitly modeling the dependency between past price dynamics and textual information. The approach dynamically focuses on news articles according to their relevance to recent market behavior, allowing the model to highlight price-informative signals while reducing the influence of noisy or unrelated content. Experiments on real-world Vietnamese stock datasets show that the proposed cross-attentive integration can improve forecasting performance for selected stocks and forecasting horizons. In particular, reductions in MAE, RMSE, and MAPE are observed in several long-horizon prediction settings. These results suggest that relevance-aware news integration, guided through cross-attention, can enhance stock price prediction under specific conditions, although such improvements are not uniform across all datasets or horizons.