This paper proposes a unified data-to-text pre-training method, UniD2T, which can be applied to various downstream data-to-text generation tasks. The method unifies different types of structured data (tables, key-value data, knowledge graphs) into a graph format and casts data-to-text generation as graph-to-text generation. To effectively encode graph structures, the paper proposes a structure-enhanced pre-training model based on the T5 model. The model introduces two new position and attention matrices to incorporate graph structures into the original Transformer. The position matrix encodes relative positional information of connected nodes, while the attention matrix incorporates explicit connectivity structure. The model is pre-trained on a large corpus of structured data and evaluated on six benchmark datasets, achieving substantially better performance than strong baselines. The model is effective in capturing structural information and generating accurate text for diverse data-to-text tasks. The paper also discusses the impact of graph structure on model performance, the effects of different model sizes, and the zero-shot performance of ChatGPT. The results show that UniD2T outperforms other models in terms of factual consistency, diversity of generated sentences, and overall performance. The paper concludes that the proposed method significantly improves data-to-text generation by effectively modeling graph structures and incorporating structural information into the Transformer.This paper proposes a unified data-to-text pre-training method, UniD2T, which can be applied to various downstream data-to-text generation tasks. The method unifies different types of structured data (tables, key-value data, knowledge graphs) into a graph format and casts data-to-text generation as graph-to-text generation. To effectively encode graph structures, the paper proposes a structure-enhanced pre-training model based on the T5 model. The model introduces two new position and attention matrices to incorporate graph structures into the original Transformer. The position matrix encodes relative positional information of connected nodes, while the attention matrix incorporates explicit connectivity structure. The model is pre-trained on a large corpus of structured data and evaluated on six benchmark datasets, achieving substantially better performance than strong baselines. The model is effective in capturing structural information and generating accurate text for diverse data-to-text tasks. The paper also discusses the impact of graph structure on model performance, the effects of different model sizes, and the zero-shot performance of ChatGPT. The results show that UniD2T outperforms other models in terms of factual consistency, diversity of generated sentences, and overall performance. The paper concludes that the proposed method significantly improves data-to-text generation by effectively modeling graph structures and incorporating structural information into the Transformer.