<p>Multimedia content understanding requires effectively integrating information from various modalities such as image, text, and audio. Traditional multimodal learning models often struggle with semantic misalignment and inadequate cross-modal interactions. To address these challenges, we propose TE-MFM (Transformer-Enhanced Multimodal Fusion Model), a novel framework that leverages a multi-level transformer architecture with cross-modal attention and semantic-guided alignment. Unlike existing methods, TE-MFM innovatively incorporates bidirectional cross-modal attention to capture mutual interactions more effectively and introduces a semantic-guided alignment mechanism that explicitly enforces consistent feature representation across modalities. TE-MFM dynamically captures fine-grained dependencies across modalities and aligns them within a unified semantic space, significantly improving performance in multimedia tasks such as emotion recognition, cross-modal retrieval, and event classification. Extensive experiments on benchmark datasets demonstrate that TE-MFM achieves state-of-the-art results, showing both superior accuracy and robustness.</p>