Abstract: This study introduces a frequency-based window-enhanced feature fusion method for text classification (F-WCNN) to address the inadequate semantic representation of low-frequency words and the interference caused by common terms. Words are first stratified into high- and low-frequency groups using the percentile method. A hierarchical modeling strategy is then employed: high-frequency words are integrated with pre-trained embeddings and TF-IDF weighting to alleviate the influence of common terms, while low-frequency words are enriched with contextual information via a sliding-window mechanism to strengthen semantic expressiveness. The fused representations are subsequently processed by a multi-channel convolutional neural network incorporating an attention mechanism and focal loss, which enhances the recognition of hard-to-classify samples and improves robustness under class imbalance. Experimental results on news text classification tasks demonstrate that F-WCNN achieves superior performance compared with baseline methods, exhibiting strong generalization ability and practical value.