Stock trend prediction remains a challenging task due to the complex interplay of multi-source information and inherent non-stationary characteristics of financial markets. To address the limitations of existing methods in handling noisy textual data and capturing multi-scale temporal patterns, this paper proposes a novel news-driven framework that synergizes structured news representation with adaptive time-frequency analysis. Specifically, we classify financial news into local and global news, and design a sparse attention module to capture both fine-grained local news correlations and macro-level global interactions. Then a cross-attention mechanism is introduced to dynamically modulate local news features under global contextual guidance. To handle non-stationary price sequences, a wavelet transform attention module is proposed to decompose stock technical indicators into high-frequency fluctuations and low-frequency trends, enabling the extraction of transient signals and long-term dependencies through a dual-path attention mechanism. The framework further employs cross-modal attention to fuse news and price features, and a temporal aggregation module to model temporal dependencies. Experimental results on our self-constructed A-share datasets demonstrate that our model outperforms state-of-the-art baselines, highlighting its effectiveness in resolving the ambiguity of traditional text representations and improving prediction performance in low signal-to-noise ratio price sequences.