Existing multi-class text classification techniques prevailingly adhere to a uniform training paradigm that consists of an encoder for text encoding, a linear classifier for label prediction, and one-hot vectors for label representation. However, this conventional paradigm considers labels as independent categorical variables, overlooking the implicit semantic information contained within labels. This may result in models exhibiting overconfidence, particularly in scenarios with semantically similar labels or inaccurately labeled datasets. To address this, we introduce the Label-Guided Self-Knowledge Distillation (LG-KD) framework. LG-KD captures the semantic similarity between texts and labels, then generates a revised label distribution (i.e., instructive knowledge) using semantic correlations between texts and labels to refine model predictions, similar to a student model learning from its own teacher. A comprehensive evaluation of five benchmark datasets shows the effectiveness of LG-KD over other baselines.