To investigate the effects of linguistic features at different levels of granularity on fund news sentiment classification, a classification method integrating cross-attention with multi-level semantic features is proposed. First, to address the high cost of manual annotation and the limited size of labeled datasets, fund news texts are initially annotated using a sentiment lexicon, latent Dirichlet allocation (LDA)-based topic modeling, and semantic representation methods. Manual review of boundary samples and random inspection of non-boundary samples are subsequently conducted to improve label quality and data reliability. Second, RoBERTa is employed as the backbone semantic representation model, while WoBERT is introduced to encode fine-grained sentiment semantics. Part-of-speech features, lexical semantics, dependency-syntactic structures, and aspect-level sentiment features are then integrated through a cross-attention mechanism. Finally, model training is optimized using a layer-wise learning-rate strategy and fast gradient method (FGM) adversarial training. Experimental results show that the proposed method achieves a Macro-F1 score of 88.08% on the fund news sentiment classification task and outperforms the baseline models overall. The results demonstrate that the proposed method can effectively improve sentiment classification for fund news and provide technical support for intelligent financial text analysis.