<p>Multimodal Sentiment Analysis (MSA) utilizes heterogeneous multimodal data to better understand human affective states. Current methods focus on extracting shared/specific features or developing fusion strategies. However, these methods often ignore the varying contributions of different modalities in sentiment analysis and struggle to eliminate non-affective interference. Aiming at the above situation, this paper proposes a UMRFormer-Enhanced ALMT with Modality-Specific Attentions for Multimodal Sentiment Analysis. We design a novel UFO-MultiScale-Rotary Transformer (UMRFormer) to replace the conventional Transformer architecture in the ALMT framework, which improves feature learning and suppresses noise interference. Additionally, the improved SE-1D Attention and CBEAM Dual Attention modules enhance feature representation of auxiliary modalities, and incorporates a gating mechanism within the Adaptive Hyper-Modality Learning (AHL) module to mitigate non-affective interference, thereby optimizing the generation of hyper-modality features. Experiments on three popular MSA datasets (CMU-MOSI, CMU-MOSEI, and CH-SIMS) demonstrate that our proposed model achieves competitive performance.</p>