Teacher gesture recognition aims to identify and interpret teacher gestures within academic settings. It has been applied in domains such as teaching performance evaluation, the optimization of online education, and special needs education. However, the background similarity of teacher gestures, the inter-class similarity, and the intra-class variability limit the recognition capabilities of visual neural networks. In this paper, a Natural Language-Driven Teacher Gesture Recognition (NLD-TGR) framework is proposed. To mitigate the effects of background similarity, textual descriptions for each frame are generated using GPT-4o, guided by prompts specifically designed to describe the teacher's hand posture in the frames. Then, we combine video features with text features mapped to a high-dimensional space to create semantically-enhanced fused features. To overcome the limitations of one-hot labels in capturing inter-class and intra-class relationships, we embed semantically interpreted category names into a textual feature space. Gesture classification is then performed by computing the similarity between these textual embeddings and the fused feature representations. The experimental results validate the effectiveness of the proposed method, which achieves state-of-the-art performance with an accuracy of 93.7% on the TBU-G teacher gesture benchmark. [For the complete proceedings, see ED675583.]