MSAF-CF: A Multimodal Sentiment Analysis Framework Based on Feature Enhancement and Cross-Fusion

As the rapid dissemination of multimodal information continues, multimodal sentiment analysis (MSA) has increasingly become a pivotal focus in sentiment analysis research, widely applied in fields such as social media analysis and telemedicine. However, existing methods are confronted with significant challenges in practical applications, including modality absence, ineffective cross-modal interaction modeling, and inefficient feature fusion. A novel framework, MSAF-CF, is introduced in this work to address the identified challenges in multimodal sentiment analysis. More specifically, a Transformer-based modality enhancement and generation module (MEGM) is introduced to effectively generate missing modalities and maintain semantic consistency across multimodal inputs. Additionally, a tri-channel selective attention model (TriSelM) is designed to adaptively enhance cross-modal feature interactions through dynamic gating mechanisms, addressing varying modality availability and fusion requirements. Furthermore, a joint optimization strategy is employed to simultaneously optimize modality reconstruction and sentiment prediction tasks, significantly boosting model robustness and generalization. Based on extensive evaluations on two well-established benchmark datasets (CMU-MOSI and CMU-MOSEI), MSAF-CF achieves remarkable performance surpassing current state-of-the-art (SOTA) approaches, particularly in scenarios involving modality loss and misalignment, confirming its efficacy and practical feasibility in real-world multimodal sentiment analysis tasks.