This paper addresses the challenge of cross-dataset zero-shot transfer learning in graph data, a critical issue in graph learning due to the continuous emergence of new graphs and the difficulty of human labeling. The authors introduce ZeroG, a novel framework designed to enable cross-dataset generalization. ZeroG leverages a pre-trained language model to encode both node attributes and class semantics, ensuring consistent feature dimensions across datasets. It also employs a prompt-based subgraph sampling module to enrich the semantic and structural information of extracted subgraphs. Additionally, a lightweight fine-tuning strategy is adopted to reduce overfitting and maintain the zero-shot learning capabilities of the language model. The effectiveness of ZeroG is demonstrated through comprehensive experiments on seven benchmark datasets, showing significant improvements in cross-dataset zero-shot transferability. The paper provides a comprehensive analysis of existing approaches and highlights the unique challenges and contributions of ZeroG in the field of graph learning.This paper addresses the challenge of cross-dataset zero-shot transfer learning in graph data, a critical issue in graph learning due to the continuous emergence of new graphs and the difficulty of human labeling. The authors introduce ZeroG, a novel framework designed to enable cross-dataset generalization. ZeroG leverages a pre-trained language model to encode both node attributes and class semantics, ensuring consistent feature dimensions across datasets. It also employs a prompt-based subgraph sampling module to enrich the semantic and structural information of extracted subgraphs. Additionally, a lightweight fine-tuning strategy is adopted to reduce overfitting and maintain the zero-shot learning capabilities of the language model. The effectiveness of ZeroG is demonstrated through comprehensive experiments on seven benchmark datasets, showing significant improvements in cross-dataset zero-shot transferability. The paper provides a comprehensive analysis of existing approaches and highlights the unique challenges and contributions of ZeroG in the field of graph learning.