Few-shot learning remains challenging due to the high cost of constructing large labeled datasets. To address this, Distribution Calibration has shown promise by leveraging similar base classes to enhance performance. However, traditional Euclidean-based metrics can be unstable in high-dimensional spaces. Inspired by research on semantic similarity, we propose a semantic similarity-based distribution calibration method that selects support classes based on intrinsic inter-class semantics rather than raw feature distances. Our approach computes semantic similarity using high-dimensional features and introduces a regularization term to ensure the covariance matrix's stability. Experiments on mini-ImageNet, CUB, and tiered-ImageNet demonstrate consistent performance improvements and robustness across varied hyperparameter settings.