MCHDoc：面向多载体中文历史文档阅读的综合基准
MCHDoc: A Comprehensive Benchmark for Reading Multi-Carrier Chinese Historical Documents
为文档图像理解提供了跨载体、跨三千年的高分辨率评测基准，可用于检验 OCR 与多模态大模型在异构版式、材质退化下的字符级识别与 LLM 后校正能力，并可作为文档分析模型领域泛化性的测试床。
原文依据 MCHDoc spans over 3,000 years of history and contains 15,724 high-resolution documents from six major carriers
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论阅读跨多种载体形式的中文历史文档是理解中华文明演进的核心环节，但这项工作仍然劳动密集，并且依赖稀缺的专家知识。尽管近期的大规模模型在孤立的历史文献集合上展现出潜力，但它们并未系统性地考察跨异构载体阅读历史文档的基本能力。为此，我们提出 MCHDoc，一个用于阅读多载体中文历史文档的综合基准。MCHDoc 跨越三千余年历史，包含来自六大载体的 15,724 份高分辨率文档，涵盖材质、版式等方面的丰富变化。该基准模拟专家工作流程，支持页面级与字符级识别，以及在有外部知识和无外部知识两种设定下基于大模型的后校正。我们在 MCHDoc 上系统评测了多种大规模模型。结果表明，即便是顶尖模型在多载体历史文档上也表现吃力。此外，我们的分析凸显了若干使大模型有效适配中文历史文本的关键因素。因此，MCHDoc 为阅读中文历史文档提供了一个标准化、具有挑战性且有历史依据的基准，并为文档分析与数字人文的未来研究奠定了基础。
英文原文摘要 1349 字符
Reading Chinese historical documents across diverse carriers is central to understanding the evolution of Chinese civilization, yet remains labor-intensive and dependent on scarce expert knowledge. Although recent large-scale models show promise on isolated historical collections, they do not systematically probe the fundamental ability to read historical documents across heterogeneous carriers. Therefore, we present MCHDoc, a comprehensive benchmark for reading multi-carrier Chinese historical documents. MCHDoc spans over 3,000 years of history and contains 15,724 high-resolution documents from six major carriers, capturing rich variations in material, layout, etc. Mimicking expert workflows, the benchmark supports page-level and character-level recognition, as well as LLM-based post-correction with and without external knowledge. We systematically evaluate a wide range of large-scale models on MCHDoc. The results show that even top-tier models struggle with multi-carrier historical documents. Furthermore, our analysis highlights several key factors for effectively adapting large models to Chinese historical texts. MCHDoc thus offers a standardized, challenging, and historically grounded benchmark for reading Chinese historical documents and provides a foundation for future research in document analysis and digital humanities.