Multimodal sentiment analysis (MSA) is crucial for applications like human-computer interaction. While pre-trained models have achieved remarkable performance across general domains, fine-tuning them solely on label-supervised sentiment tasks often leads to an over-reliance on global representations, thereby overlooking fine-grained sentiment indicators. However, it is precisely these fine-grained features that drive changes in sentiment. To address this issue, we propose a Sentiment-Aware Multitask Learning framework that enables the model to simultaneously understand both coarse-grained and fine-grained sentiment information. In addition to label-supervised learning, we aim to identify fine-grained tokens within the data that influence sentiment, which are then masked to prompt the model to reconstruct the missing components. Besides, to better integrate multimodal features, some late fusion methods introduce learnable modules, such as a linear layer. However, linear-layer late fusion methods necessitate dataset-specific retraining, while non-parametric late fusion methods like voting or averaging inadequately integrate multimodal features. Inspired by information theory, we propose an Entropy-Driven Fusion method that preserves the plug-and-play nature of non-parametric techniques while enabling more effective multimodal fusion. Experiments demonstrate that our approach achieves state-of-the-art results on the MSA task.