Diffusion Language-Shapelets (DiffShape) is a semi-supervised time-series classification model that improves model interpretability by generating discriminative shapelets. The model combines self-supervised diffusion learning with contrastive language-shapelets learning. In self-supervised diffusion learning, real subsequences are used as conditions to generate shapelets that are more similar to real subsequences. In contrastive language-shapelets learning, natural language descriptions of time series are used to enhance the discriminability of generated shapelets. The model is evaluated on the UCR time series archive and achieves state-of-the-art performance, outperforming existing methods in both classification accuracy and interpretability. DiffShape automatically generates shapelets for each time series, improving model interpretability and classification performance. The model uses a diffusion process to generate shapelets and incorporates natural language descriptions to enhance their discriminability. The model is trained using a combination of classification loss and contrastive learning. The results show that DiffShape achieves superior performance compared to existing methods, demonstrating its effectiveness in semi-supervised time-series classification.Diffusion Language-Shapelets (DiffShape) is a semi-supervised time-series classification model that improves model interpretability by generating discriminative shapelets. The model combines self-supervised diffusion learning with contrastive language-shapelets learning. In self-supervised diffusion learning, real subsequences are used as conditions to generate shapelets that are more similar to real subsequences. In contrastive language-shapelets learning, natural language descriptions of time series are used to enhance the discriminability of generated shapelets. The model is evaluated on the UCR time series archive and achieves state-of-the-art performance, outperforming existing methods in both classification accuracy and interpretability. DiffShape automatically generates shapelets for each time series, improving model interpretability and classification performance. The model uses a diffusion process to generate shapelets and incorporates natural language descriptions to enhance their discriminability. The model is trained using a combination of classification loss and contrastive learning. The results show that DiffShape achieves superior performance compared to existing methods, demonstrating its effectiveness in semi-supervised time-series classification.