CCC: Chinese Commercial Contracts Dataset for Documents Layout Understanding

In recent years, Visual document understanding tasks have become increasingly popular due to the growing demand for commercial applications, especially for processing complex image documents such as contracts, and patents. However, there is no high-quality domain-specific dataset available except for English. And for other languages like Chinese, it is hard to utilize current English datasets due to the significant differences in writing norms and layout formats. To mitigate this issue, we introduce the Chinese Commercial Contracts (CCC) dataset to explore better visual document layout understanding modeling for Chinese commercial contract in the paper. This dataset contains 10,000 images, each containing various elements such as text, tables, seals, and handwriting. Moreover, we propose the Chinese Layout Understanding Pre-train Transformer (CLUPT) Model, which is pre-trained on the proposed CCC dataset by incorporating textual and layout information into the pre-train task. Based on the VisionEncoder-LanguageDecoder model structure, our model can perform end-to-end Chinese document layout understanding tasks. The data and code are available at https://github.com/yysirs/CLUPT.