<p>Annotated data often includes ambiguous labels, posing a partial label learning (PLL) challenge. PLL aims to classify instances that have multiple potential labels but only one correct label. While recent studies have shown that contrastive learning methods are effective for learning image features, significant obstacles remain when applying these methods to PLL. In this work, we introduce contrastive learning for PLL and propose a convolutional attention contrastive learning for PLL (CACL-PLL) framework, designed to learn category-specific prototypes that represent category labels and assist in label disambiguation. Specifically, we incorporate a combination of strong and weak augmentation, and in the encoder component, we propose a convolutional attention encoder alongside a convolutional attention momentum encoder. The encoded features are then passed to the proposed convolutional attention projection head and a fully convolutional network classifier. Additionally, we introduce a new label disambiguation strategy, which involves updating a label confidence matrix and a label correction matrix, where the update of the label correction matrix is guided by the category prototype. Finally, we design a novel loss function that assigns weights to the classification loss of two strongly enhanced samples, working alongside the contrastive loss function to supervise model training. Extensive experiments demonstrate the effectiveness and superior performance of CACL-PLL in PLL.</p>