{"id":4349,"date":"2025-11-16T13:14:32","date_gmt":"2025-11-16T05:14:32","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=4349"},"modified":"2025-11-16T13:14:32","modified_gmt":"2025-11-16T05:14:32","slug":"202511%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-doclayllm-an-efficient-multi-modal-extension-of-large-language-models-for-text-rich-document-understanding","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202511%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-doclayllm-an-efficient-multi-modal-extension-of-large-language-models-for-text-rich-document-understanding\/","title":{"rendered":"202511\u8bba\u6587\u7814\u8bfb-DocLayLLM: An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding"},"content":{"rendered":"\n<p>\u4f5c\u8005\uff1aWenhui Liao, Jiapeng Wang, Hongliang Li\u7b49<br>\u5355\u4f4d\uff1a\u534e\u5357\u7406\u5de5\u5927\u5b66\uff0c\u963f\u91cc\u4e91\u8ba1\u7b97\uff0cScut-\u73e0\u6d77\u73b0\u4ee3\u4ea7\u4e1a\u521b\u65b0\u7814\u7a76\u9662<br>\u6765\u6e90\uff1aCVPR2025<br>\u65f6\u95f4\uff1a2024.08<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a<\/h2>\n\n\n\n<p>\u5f53\u9762\u5bf9\u6587\u6863\u56fe\u50cf\uff08\u5982\u53d1\u7968\u3001\u8868\u683c\u3001\u5408\u540c\u7b49\uff09\u8fd9\u7c7b\u5bcc\u542b\u5e03\u5c40\u4fe1\u606f\uff08Layout\uff09\u548c\u89c6\u89c9\u7ed3\u6784\uff08Visual Structure\uff09\u7684\u590d\u6742\u6570\u636e\u65f6\uff0c\u6807\u51c6LLM\u663e\u5f97\u529b\u4e0d\u4ece\u5fc3\u3002<br>\u5f53\u524d\u7684\u6587\u6863\u7406\u89e3\uff08Document Understanding, DU\uff09\u65b9\u6cd5\u4e3b\u8981\u5206\u4e3a\u4e24\u7c7b\uff1a<br>OCR\u4f9d\u8d56\u578b\u65b9\u6cd5\uff1a\u9996\u5148\u4f7f\u7528OCR\u5f15\u64ce\u5c06\u56fe\u50cf\u4e2d\u7684\u6587\u5b57\u8bc6\u522b\u51fa\u6765\uff0c\u5e76\u901a\u5e38\u9644\u5e26\u6587\u672c\u5185\u5bb9\u53ca\u5176\u5728\u9875\u9762\u4e0a\u7684\u5750\u6807\uff08bounding boxes\uff09\u3002\u7136\u540e\u5c06\u8fd9\u4e9b\u201c\u5bcc\u6587\u672c\u201d\u4fe1\u606f\u8f93\u5165\u5230\u4e0b\u6e38\u6a21\u578b\u4e2d\u8fdb\u884c\u7406\u89e3\u3002\u8fd9\u79cd\u65b9\u6cd5\u7684\u74f6\u9888\u5728\u4e8eOCR\u7684\u51c6\u786e\u6027\uff0c\u5c24\u5176\u662f\u5728\u5904\u7406\u4f4e\u8d28\u91cf\u3001\u624b\u5199\u6216\u590d\u6742\u7248\u5f0f\u7684\u6587\u6863\u65f6\uff0cOCR\u9519\u8bef\u4f1a\u76f4\u63a5\u4f20\u5bfc\u5e76\u653e\u5927\u5230\u540e\u7eed\u7684\u7406\u89e3\u4efb\u52a1\u4e2d\u3002<br>OCR-free\u65b9\u6cd5\uff1a\u8bd5\u56fe\u76f4\u63a5\u4ece\u539f\u59cb\u50cf\u7d20\u51fa\u53d1\uff0c\u7ed5\u8fc7OCR\u6b65\u9aa4\uff0c\u5229\u7528\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08\u5982\u57fa\u4e8eViT\u7684\u89c6\u89c9\u7f16\u7801\u5668+LLM\uff09\u8fdb\u884c\u7aef\u5230-end\u7684\u7406\u89e3\u3002\u8fd9\u7c7b\u65b9\u6cd5\u867d\u7136\u907f\u514d\u4e86OCR\u9519\u8bef\uff0c\u4f46\u9762\u4e34\u7740\u5de8\u5927\u7684\u6311\u6218\uff1a<br>\u8bad\u7ec3\u6570\u636e\u6c61\u67d3\uff1a\u8bb8\u591a\u7528\u4e8e\u8bad\u7ec3\u8fd9\u4e9b\u591a\u6a21\u6001\u6a21\u578b\u7684\u5927\u89c4\u6a21\u56fe\u6587\u6570\u636e\u96c6\uff0c\u672c\u8eab\u5c31\u5305\u542b\u4e86\u6765\u81ea\u516c\u5f00DU\u57fa\u51c6\uff08\u5982FUNSD, CORD\uff09\u7684\u6570\u636e\u3002\u8fd9\u5bfc\u81f4\u6a21\u578b\u5728\u5b98\u65b9\u6d4b\u8bd5\u96c6\u4e0a\u7684\u8868\u73b0\u53ef\u80fd\u88ab\u9ad8\u4f30\uff0c\u65e0\u6cd5\u771f\u5b9e\u53cd\u6620\u5176\u5728\u672a\u89c1\u8fc7\u7684\u65b0\u6570\u636e\u96c6\u4e0a\u7684\u6cdb\u5316\u80fd\u529b\u3002<br>\u6548\u7387\u4e0e\u6027\u80fd\u74f6\u9888\uff1a\u76f4\u63a5\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u6587\u6863\u56fe\u50cf\u8ba1\u7b97\u6210\u672c\u9ad8\u6602\uff0c\u4e14\u6a21\u578b\u96be\u4ee5\u7cbe\u786e\u6355\u6349\u7ec6\u7c92\u5ea6\u7684\u6587\u672c\u5185\u5bb9\u548c\u590d\u6742\u7684\u5e03\u5c40\u5173\u7cfb\u3002<br>\u56e0\u6b64\uff0c\u9886\u57df\u5185\u4e9f\u9700\u4e00\u79cd\u65e2\u80fd\u6709\u6548\u5229\u7528\u5e03\u5c40\u4fe1\u606f\u3001\u53c8\u4e0d\u8fc7\u5ea6\u4f9d\u8d56OCR\u3001\u540c\u65f6\u5177\u5907\u5f3a\u5927\u96f6\u6837\u672c\uff08zero-shot\uff09\u6cdb\u5316\u80fd\u529b\u7684\u9ad8\u6548\u6587\u6863\u7406\u89e3\u6846\u67b6\u3002<\/p>\n\n\n\n<p>\u8be5\u6587\u65e8\u5728\u89e3\u51b3\u4ee5\u4e0b\u6838\u5fc3\u95ee\u9898\uff1a<br>\u5982\u4f55\u5728\u4e0d\u4f9d\u8d56\u9ad8\u8d28\u91cfOCR\u8f93\u51fa\u7684\u524d\u63d0\u4e0b\uff0c\u6784\u5efa\u4e00\u4e2a\u9ad8\u6548\u7684\u6587\u6863\u7406\u89e3\u6a21\u578b\uff1f<br>\u5982\u4f55\u8bbe\u8ba1\u4e00\u79cd\u65b0\u7684\u8868\u793a\u65b9\u6cd5\uff0c\u65e2\u80fd\u4fdd\u7559\u6587\u6863\u7684\u5173\u952e\u5e03\u5c40\u548c\u7ed3\u6784\u4fe1\u606f\uff0c\u53c8\u80fd\u4ee5LLM\u53cb\u597d\u7684\u65b9\u5f0f\u8f93\u5165\uff0c\u4ece\u800c\u5b9e\u73b0\u5f3a\u5927\u7684\u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b\uff1f<br>\u5982\u4f55\u514b\u670d\u73b0\u6709OCR-free\u65b9\u6cd5\u56e0\u8bad\u7ec3\u6570\u636e\u6c61\u67d3\u800c\u5bfc\u81f4\u7684\u8bc4\u4f30\u504f\u5dee\u95ee\u9898\uff1f<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u89e3\u51b3\u529e\u6cd5-DocLayLLM<\/h2>\n\n\n\n<p>DocLayLLM\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff1a\u4e0d\u76f4\u63a5\u8f93\u5165\u539f\u59cb\u50cf\u7d20\u6216\u5197\u957f\u7684OCR\u5750\u6807\u5e8f\u5217\uff0c\u800c\u662f\u5c06\u6587\u6863\u56fe\u50cf\u201c\u7f16\u8bd1\u201d\u6210\u4e00\u79cd\u7b80\u6d01\u3001\u7ed3\u6784\u5316\u7684\u201c\u6587\u6863\u6807\u8bb0\u8bed\u8a00\u201d\uff08Document Markup Language\uff09\uff0c\u518d\u5c06\u5176\u4f5c\u4e3a\u7eaf\u6587\u672c\u6307\u4ee4\u8f93\u5165\u7ed9LLM\u3002<br>DocLayLLM \u7684\u6574\u4f53\u67b6\u6784\u53ef\u4ee5\u88ab\u770b\u4f5c\u662f\u4e00\u4e2a\u4e24\u9636\u6bb5\u7684\u6d41\u6c34\u7ebf\uff1a\u611f\u77e5\u9636\u6bb5\uff08Perception Stage\uff09\u548c\u63a8\u7406\u9636\u6bb5\uff08Reasoning Stage\uff09\u3002\u5176\u7cbe\u9ad3\u5728\u4e8e\u5982\u4f55\u5728\u8fd9\u4e24\u4e2a\u9636\u6bb5\u4e4b\u95f4\u8bbe\u8ba1\u4e00\u4e2a\u9ad8\u6548\u7684\u3001\u4fe1\u606f\u4e30\u5bcc\u7684\u6865\u6881\u2014\u2014\u5373\u201c\u6587\u6863\u6807\u8bb0\u8bed\u8a00\u201d\u548c\u914d\u5957\u7684\u63a8\u7406\u7b56\u7565\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"870\" height=\"410\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-14.png\" alt=\"\" class=\"wp-image-4350\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-14.png 870w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-14-300x141.png 300w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-14-768x362.png 768w\" sizes=\"auto, (max-width: 870px) 100vw, 870px\" \/><\/figure>\n\n\n\n<p>\u7b2c\u4e00\u90e8\u5206\uff1a\u611f\u77e5\u9636\u6bb5\uff08Perception Stage\uff09<br>\u6587\u6863\u5e03\u5c40\u89e3\u6790\uff08Document Layout Parsing\uff09<br>\u9996\u5148\uff0c\u4f7f\u7528\u4e00\u4e2a\u9884\u8bad\u7ec3\u7684\u3001\u901a\u7528\u7684\u6587\u6863\u5e03\u5c40\u5206\u6790\u6a21\u578b\uff08\u4f8b\u5982\uff0c\u57fa\u4e8eDetectron2\u6216\u7c7b\u4f3c\u67b6\u6784\u7684\u6a21\u578b\uff09\u6765\u5904\u7406\u8f93\u5165\u7684\u6587\u6863\u56fe\u50cf\u3002<br>\u8be5\u6a21\u578b\u4f1a\u68c0\u6d4b\u51fa\u6587\u6863\u4e2d\u7684\u4e0d\u540c\u8bed\u4e49\u533a\u57df\uff08Semantic Regions\uff09\uff0c\u4f8b\u5982\uff1atext, title, list, table, figure, form \u7b49\uff0c\u5e76\u4e3a\u6bcf\u4e2a\u533a\u57df\u751f\u6210\u4e00\u4e2a\u8fb9\u754c\u6846\uff08Bounding Box\uff09\u3002<br>\u533a\u57df\u5185\u6587\u672c\u8bc6\u522b\uff08Intra-region Text Recognition &#8211; \u53ef\u9009\u4f46\u63a8\u8350\uff09<br>\u5bf9\u4e8e\u6bcf\u4e2a\u88ab\u8bc6\u522b\u51fa\u7684\u6587\u672c\u7c7b\u533a\u57df\uff08\u5982text, title\uff09\uff0c\u4f7f\u7528\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684OCR\u5f15\u64ce\uff08\u5982PaddleOCR, Tesseract\uff09\u6765\u8bc6\u522b\u5176\u4e2d\u7684\u6587\u5b57\u5185\u5bb9\u3002<br>\u5173\u952e\u70b9\uff1a\u8fd9\u91cc\u7684OCR\u4ec5\u7528\u4e8e\u83b7\u53d6\u533a\u57df\u5185\u7684\u7eaf\u6587\u672c\u5185\u5bb9\uff0c\u4e0d\u518d\u9700\u8981\u5176\u7cbe\u786e\u7684\u5b57\u7b26\u7ea7\u6216\u5355\u8bcd\u7ea7\u5750\u6807\u3002\u8fd9\u5927\u5927\u964d\u4f4e\u4e86\u5bf9OCR\u7cbe\u5ea6\u7684\u82db\u523b\u8981\u6c42\uff0c\u56e0\u4e3a\u5373\u4f7fOCR\u6709\u5c11\u91cf\u9519\u8bef\uff0c\u53ea\u8981\u4e0d\u5f71\u54cd\u6574\u4f53\u8bed\u4e49\uff0cLLM\u4f9d\u7136\u53ef\u4ee5\u7406\u89e3\u3002<br>\u6784\u5efa\u201c\u6587\u6863\u6807\u8bb0\u8bed\u8a00\u201d\uff08Constructing the Document Markup Language\uff09<br>\u8fd9\u662fDocLayLLM\u6700\u6838\u5fc3\u7684\u521b\u65b0\u3002\u7cfb\u7edf\u5c06\u4e0a\u4e00\u6b65\u5f97\u5230\u7684\u5e03\u5c40\u533a\u57df\u548c\u6587\u672c\u5185\u5bb9\uff0c\u6309\u7167\u5b83\u4eec\u5728\u9875\u9762\u4e0a\u7684\u7a7a\u95f4\u4f4d\u7f6e\u5173\u7cfb\uff08\u4e3b\u8981\u662f\u4ece\u4e0a\u5230\u4e0b\u3001\u4ece\u5de6\u5230\u53f3\u7684\u9605\u8bfb\u987a\u5e8f\uff09\uff0c\u7ec4\u7ec7\u6210\u4e00\u79cd\u7c7b\u4f3cHTML\/XML\u7684\u7ed3\u6784\u5316\u6587\u672c\u3002<\/p>\n\n\n\n<p>\u6784\u5efa\u201c\u6587\u6863\u6807\u8bb0\u8bed\u8a00\u201d\uff08Constructing the Document Markup Language\uff09<br>\u8fd9\u662fDocLayLLM\u6700\u6838\u5fc3\u7684\u521b\u65b0\u3002\u7cfb\u7edf\u5c06\u4e0a\u4e00\u6b65\u5f97\u5230\u7684\u5e03\u5c40\u533a\u57df\u548c\u6587\u672c\u5185\u5bb9\uff0c\u6309\u7167\u5b83\u4eec\u5728\u9875\u9762\u4e0a\u7684\u7a7a\u95f4\u4f4d\u7f6e\u5173\u7cfb\uff08\u4e3b\u8981\u662f\u4ece\u4e0a\u5230\u4e0b\u3001\u4ece\u5de6\u5230\u53f3\u7684\u9605\u8bfb\u987a\u5e8f\uff09\uff0c\u7ec4\u7ec7\u6210\u4e00\u79cd\u7c7b\u4f3cHTML\/XML\u7684\u7ed3\u6784\u5316\u6587\u672c\u3002<br>\u8fd9\u79cd\u8868\u793a\u6cd5\u663e\u5f0f\u5730\u7f16\u7801\u4e86\u6587\u6863\u7684\u5c42\u6b21\u7ed3\u6784\u548c\u8bed\u4e49\u7c7b\u578b\uff0c\u540c\u65f6\u9690\u5f0f\u5730\u901a\u8fc7\u6807\u7b7e\u7684\u987a\u5e8f\u7f16\u7801\u4e86\u7a7a\u95f4\u5e03\u5c40\u4fe1\u606f\u3002\u5b83\u5c06\u590d\u6742\u7684\u591a\u6a21\u6001\u4fe1\u606f\uff08\u56fe\u50cf+\u5e03\u5c40+\u6587\u672c\uff09\u7edf\u4e00\u8f6c\u5316\u4e3a\u4e86LLM\u6700\u64c5\u957f\u5904\u7406\u7684\u7eaf\u6587\u672c\u5e8f\u5217\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"304\" height=\"294\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-15.png\" alt=\"\" class=\"wp-image-4351\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-15.png 304w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-15-300x290.png 300w\" sizes=\"auto, (max-width: 304px) 100vw, 304px\" \/><\/figure>\n\n\n\n<p>\u7b2c\u4e8c\u90e8\u5206\uff1a\u63a8\u7406\u9636\u6bb5\uff08Reasoning Stage\uff09\u4e0e CoT \u7b56\u7565<br>\u8fd9\u662fDocLayLLM\u5c55\u73b0\u5176\u667a\u80fd\u7684\u5730\u65b9\u3002\u5b83\u4e0d\u4ec5\u4ec5\u662f\u5c06 DML \u4e22\u7ed9LLM\uff0c\u800c\u662f\u901a\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1\u7684 Chain-of-Thought (CoT) \u63d0\u793a\u5de5\u7a0b\uff0c\u5f15\u5bfcLLM\u8fdb\u884c\u7b26\u5408\u6587\u6863\u7406\u89e3\u903b\u8f91\u7684\u9010\u6b65\u63a8\u7406\u3002<br>Layout-Type-Related CoT (\u5e03\u5c40\u7c7b\u578b\u76f8\u5173 CoT)<br>\u8fd9\u79cdCoT\u4e3b\u8981\u7528\u4e8e\u952e\u503c\u5bf9\u4fe1\u606f\u62bd\u53d6\uff08KIE\uff09\u7b49\u4efb\u52a1\uff0c\u5176\u76ee\u6807\u662f\u63a8\u65ad\u67d0\u4e2a\u7279\u5b9a\u533a\u57df\u7684\u5e03\u5c40\u7c7b\u578b\u6216\u4ece\u4e2d\u63d0\u53d6\u4fe1\u606f\u3002<br>\u9002\u7528\u573a\u666f: \u5f53\u7528\u6237\u67e5\u8be2\u6307\u5411\u4e00\u4e2a\u5177\u4f53\u7684\u533a\u57df\uff08\u901a\u8fc7bounding box\u7ed9\u51fa\uff09\u65f6\u3002<br>CoT \u6b65\u9aa4:<br>Step 1: \u5b9a\u4f4d\u4e0e\u533a\u57df\u63cf\u8ff0<br>\u201c\u9996\u5148\uff0c\u8bc6\u522b\u67e5\u8be2\u533a\u57df [x1, y1, x2, y2] \u5185\u7684\u6587\u672c\u5185\u5bb9\uff0c\u5e76\u786e\u5b9a\u8be5\u533a\u57df\u5728\u6587\u6863\u4e2d\u7684\u5927\u81f4\u4f4d\u7f6e\uff08\u4f8b\u5982\uff0c\u5de6\u4e0a\u89d2\u3001\u53f3\u4e0b\u89d2\u7b49\uff09\u3002\u201d<br>Step 2: \u5bfb\u627e\u6700\u8fd1\u90bb<br>\u201c\u7136\u540e\uff0c\u627e\u5230\u4e0e\u8be5\u67e5\u8be2\u533a\u57df\u5728\u7a7a\u95f4\u4e0a\u6700\u63a5\u8fd1\u7684\u5176\u4ed6\u5e03\u5c40\u533a\u5757\uff0c\u5e76\u8bc4\u4f30\u5b83\u4eec\u7684\u5e03\u5c40\u7c7b\u578b\uff08\u4f8b\u5982\uff0c\u5b83\u662f\u5426\u7d27\u90bb\u4e00\u4e2a\u2018\u6807\u9898\u2019\u6216\u4f4d\u4e8e\u4e00\u4e2a\u2018\u8868\u683c\u2019\u5185\u90e8\uff1f\uff09\u3002\u201d<br>Step 3: \u57fa\u4e8e\u4e0a\u4e0b\u6587\u63a8\u7406<br>\u201c\u6700\u540e\uff0c\u7efc\u5408\u4ee5\u4e0a\u4fe1\u606f\uff0c\u63a8\u65ad\u51fa\u67e5\u8be2\u533a\u57df\u7684\u5e03\u5c40\u7c7b\u578b\u6216\u6240\u9700\u7684\u4fe1\u606f\u3002\u201d<br>\u4f5c\u7528: \u8fd9\u79cdCoT\u6559\u4f1aLLM\u50cf\u4eba\u7c7b\u4e00\u6837\uff0c\u901a\u8fc7\u5c40\u90e8\u4e0a\u4e0b\u6587\uff08\u9644\u8fd1\u7684\u5143\u7d20\u662f\u4ec0\u4e48\uff09\u6765\u7406\u89e3\u4e00\u4e2a\u672a\u77e5\u533a\u57df\u7684\u542b\u4e49\u3002\u4f8b\u5982\uff0c\u4e00\u4e2a\u5b64\u7acb\u7684\u6570\u5b57\u5982\u679c\u7d27\u6328\u7740\u201cTotal:\u201d\u5b57\u6837\uff0c\u5f88\u53ef\u80fd\u5c31\u662f\u603b\u91d1\u989d\u3002<\/p>\n\n\n\n<p>Geometry-Related CoT (\u51e0\u4f55\u5173\u7cfb\u76f8\u5173 CoT)<br>\u8fd9\u79cdCoT\u4e3b\u8981\u7528\u4e8e\u6587\u6863\u89c6\u89c9\u95ee\u7b54\uff08VQA\uff09\u7b49\u9700\u8981\u7406\u89e3\u5143\u7d20\u95f4\u7a7a\u95f4\u5173\u7cfb\u7684\u4efb\u52a1\u3002<br>\u9002\u7528\u573a\u666f: \u5f53\u95ee\u9898\u6d89\u53ca\u5230\u4e24\u4e2a\u6216\u591a\u4e2a\u6587\u672c\u7247\u6bb5\u4e4b\u95f4\u7684\u76f8\u5bf9\u4f4d\u7f6e\u5173\u7cfb\u65f6\uff08\u4f8b\u5982\uff0c\u201c\u2018Conclusion\u2019\u8fd9\u4e2a\u8bcd\u5728\u2018AIR\u2019\u8fd9\u4e2a\u8bcd\u7684\u54ea\u4e2a\u65b9\u5411\uff1f\u201d\uff09\u3002<br>CoT \u6b65\u9aa4\uff08\u6839\u636e\u60a8\u63d0\u4f9b\u7684\u77e5\u8bc6\u5e93\u5185\u5bb9\uff09:<br>Step 1: \u83b7\u53d6\u5750\u6807<br>\u201c\u2018Concluding\u2019 is located at [35,17,167,49] and \u2018AIR\u2019 at [816,259,858,289].\u201d<br>\uff08\u9996\u5148\uff0c\u4eceDML\u6216\u9644\u52a0\u7684\u5143\u6570\u636e\u4e2d\u63d0\u53d6\u4e24\u4e2a\u5173\u952e\u8bcd\u7684\u7cbe\u786e\u8fb9\u754c\u6846\u5750\u6807\u3002\uff09<br>Step 2: \u6295\u5f71\u5206\u6790<br>\u201cThe vertical projection is 35\u2013167 for \u2018Concluding\u2019 and 816\u2013858 for \u2018AIR\u2019; the horizontal projection is 17\u201349 for \u2018Concluding\u2019 and 259\u2013289 for \u2018AIR,\u2019 with no overlap.\u201d<br>\uff08\u63a5\u7740\uff0c\u5206\u522b\u8ba1\u7b97\u4e24\u4e2a\u6846\u5728X\u8f74\uff08\u5782\u76f4\u6295\u5f71\uff09\u548cY\u8f74\uff08\u6c34\u5e73\u6295\u5f71\uff09\u4e0a\u7684\u533a\u95f4\uff0c\u5e76\u5224\u65ad\u5b83\u4eec\u662f\u5426\u6709\u91cd\u53e0\u3002\u6ca1\u6709\u91cd\u53e0\u610f\u5473\u7740\u5b83\u4eec\u65e2\u4e0d\u5728\u540c\u4e00\u884c\u4e5f\u4e0d\u5728\u540c\u4e00\u5217\u3002\uff09<br>Step 3: \u4e2d\u5fc3\u70b9\u8ba1\u7b97\u4e0e\u65b9\u5411\u5224\u65ad<br>\u201cThe centers are (101,33) for \u2018Concluding\u2019 and (827,274) for \u2018AIR\u2019.\u201d<br>\uff08\u7136\u540e\uff0c\u8ba1\u7b97\u4e24\u4e2a\u8fb9\u754c\u6846\u7684\u4e2d\u5fc3\u70b9\u5750\u6807\u3002\uff09<br>Step 4: \u6700\u7ec8\u7ed3\u8bba<br>\u201cSince the center of \u2018Concluding\u2019 has a smaller X and Y coordinate than \u2018AIR\u2019, it is located to the top-left of \u2018AIR\u2019.\u201d<br>\uff08\u6700\u540e\uff0c\u901a\u8fc7\u6bd4\u8f83\u4e2d\u5fc3\u70b9\u7684\u5750\u6807\uff0c\u5f97\u51fa\u6700\u7ec8\u7684\u7a7a\u95f4\u65b9\u4f4d\u5173\u7cfb\u3002\uff09<br>\u4f5c\u7528: \u8fd9\u79cdCoT\u5c06\u62bd\u8c61\u7684\u7a7a\u95f4\u5173\u7cfb\u95ee\u9898\uff0c\u5206\u89e3\u4e3a\u4e00\u7cfb\u5217\u53ef\u8ba1\u7b97\u7684\u3001\u57fa\u4e8e\u5750\u6807\u7684\u51e0\u4f55\u64cd\u4f5c\u3002\u5b83\u8d4b\u4e88\u4e86\u7eaf\u6587\u672cLLM\u4e00\u79cd\u201c\u7a7a\u95f4\u60f3\u8c61\u529b\u201d\uff0c\u4f7f\u5176\u80fd\u591f\u7cbe\u786e\u56de\u7b54\u5173\u4e8e\u5e03\u5c40\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p>\u7b2c\u4e09\u90e8\u5206\uff1a\u6a21\u578b\u96c6\u6210\u4e0e\u8bad\u7ec3<br>LLM\u9009\u62e9: \u4f7f\u7528\u4e00\u4e2a\u5f3a\u5927\u7684\u5f00\u6e90LLM\uff08\u5982Vicuna-7B\uff09\uff0c\u5e76\u4e14\u51bb\u7ed3\u5176\u6240\u6709\u53c2\u6570\u3002\u8fd9\u4fdd\u8bc1\u4e86\u6a21\u578b\u7684\u901a\u7528\u8bed\u8a00\u80fd\u529b\u4e0d\u4f1a\u88ab\u7834\u574f\uff0c\u540c\u65f6\u4e5f\u6781\u5927\u5730\u8282\u7701\u4e86\u8ba1\u7b97\u8d44\u6e90\u3002<br>\u8f93\u5165\u6784\u9020: \u5c06\u4efb\u52a1\u6307\u4ee4\u3001DML\u5b57\u7b26\u4e32\u4ee5\u53ca\u76f8\u5e94\u7684CoT\u6a21\u677f\uff08\u4f5c\u4e3aFew-shot\u793a\u4f8b\u6216\u7cfb\u7edf\u63d0\u793a\uff09\u62fc\u63a5\u6210\u4e00\u4e2a\u5b8c\u6574\u7684Prompt\u3002<br>\u8bad\u7ec3: \u7531\u4e8eLLM\u662f\u51bb\u7ed3\u7684\uff0c\u8bad\u7ec3\u4ec5\u6d89\u53ca\u4e00\u4e2a\u975e\u5e38\u8f7b\u91cf\u7ea7\u7684\u89c6\u89c9\u9002\u914d\u5668\uff08Visual Adapter\uff09\u6216LoRA\u6a21\u5757\uff0c\u7528\u4e8e\u5fae\u8c03\u6a21\u578b\u4ee5\u66f4\u597d\u5730\u7406\u89e3DML\u7684\u8bed\u6cd5\u548c\u9075\u5faaCoT\u6307\u4ee4\u3002\u8bad\u7ec3\u6570\u636e\u91cf\u9700\u6c42\u5f88\u5c0f\u3002<br>\u63a8\u7406: \u5728\u63a8\u7406\u65f6\uff0c\u6574\u4e2a\u6d41\u7a0b\u662f\u786e\u5b9a\u6027\u7684\uff1a\u56fe\u50cf -> \u5e03\u5c40\u5206\u6790 -> OCR -> DML -> Prompt -> LLM -> \u7b54\u6848\u3002<\/p>\n\n\n\n<p>DocLayLLM \u7684\u5b8c\u6574\u5de5\u4f5c\u6d41\u5982\u4e0b\uff1a<br>\u89c6\u89c9\u611f\u77e5: \u7528\u4f20\u7edfCV\u6a21\u578b\u89e3\u6790\u6587\u6863\u7684\u5e03\u5c40\u7ed3\u6784\u548c\u533a\u5757\u6587\u672c\u3002<br>\u4fe1\u606f\u7f16\u8bd1: \u5c06\u611f\u77e5\u7ed3\u679c\u7f16\u8bd1\u6210LLM\u53cb\u597d\u7684**\u7ed3\u6784\u5316\u6587\u6863\u6807\u8bb0\u8bed\u8a00 **(DML)\u3002<br>\u667a\u80fd\u63a8\u7406: \u901a\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u00a0Layout-Type-Related\u00a0\u548c\u00a0Geometry-Related\u00a0\u4e24\u79cd **Chain-of-Thought **(CoT) \u7b56\u7565\uff0c\u5f15\u5bfc\u51bb\u7ed3\u7684LLM\u5bf9DML\u8fdb\u884c\u6df1\u5ea6\u7406\u89e3\u548c\u63a8\u7406\u3002<br>\u9ad8\u6548\u5b9e\u73b0: \u6574\u4e2a\u7cfb\u7edf\u8bad\u7ec3\u6210\u672c\u4f4e\uff0c\u63a8\u7406\u901f\u5ea6\u5feb\uff0c\u4e14\u5bf9OCR\u9519\u8bef\u9c81\u68d2\u3002<\/p>\n\n\n\n<p>\u8bad\u7ec3\u7b56\u7565\uff08Training Strategy\uff09:<br>\u9ad8\u6548\u6027\u4f53\u73b0\uff1a\u7531\u4e8eLLM\u7684\u53c2\u6570\u662f\u51bb\u7ed3\u7684\uff0c\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u53ea\u9700\u8981\u5fae\u8c03\u4e00\u4e2a\u975e\u5e38\u8f7b\u91cf\u7ea7\u7684\u9002\u914d\u5668\uff08Adapter\uff09\u6216\u8005\u751a\u81f3\u53ea\u8fdb\u884c\u4e0a\u4e0b\u6587\u5b66\u4e60\uff08In-context Learning\uff09\u3002<br>\u8bad\u7ec3\u6570\u636e\u4ec5\u9700\u5c11\u91cf\u7684\uff08\u6587\u6863\u6807\u8bb0\u8bed\u8a00\uff0c\u6307\u4ee4\uff0c\u7b54\u6848\uff09\u4e09\u5143\u7ec4\uff0c\u5373\u53ef\u8ba9\u6a21\u578b\u5b66\u4f1a\u5982\u4f55\u54cd\u5e94\u5404\u79cdDU\u4efb\u52a1\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5b9e\u9a8c<\/h2>\n\n\n\n<p>\u6570\u636e\u96c6\uff1a<br>\u8bad\u7ec3\u96c6\uff1a\u4f7f\u7528\u4e86\u591a\u4e2a\u6807\u51c6DU\u6570\u636e\u96c6\uff08\u5982DocVQA, RVL-CDIP\u7b49\uff09\u6765\u6784\u5efa\u201c\u6587\u6863\u6807\u8bb0\u8bed\u8a00\u201d\u683c\u5f0f\u7684\u8bad\u7ec3\u6837\u672c\u3002<br>\u6d4b\u8bd5\u96c6\uff1a\u5728\u591a\u4e2a\u5177\u6709\u6311\u6218\u6027\u7684\u57fa\u51c6\u4e0a\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u5305\u62ec\uff1a<br>OCR-dependent benchmarks: FUNSD\uff08\u8868\u5355\u7406\u89e3\uff09, CORD\uff08\u6536\u636e\u7406\u89e3\uff09, SROIE\uff08\u7968\u636e\u4fe1\u606f\u62bd\u53d6\uff09\u3002<br>OCR-free \/ VQA benchmarks: DocVQA\uff08\u6587\u6863\u89c6\u89c9\u95ee\u7b54\uff09\u3002<br>\u5173\u952e\u63a7\u5236\uff1a\u4f5c\u8005\u7279\u522b\u5f3a\u8c03\u5e76\u9a8c\u8bc1\u4e86\u4ed6\u4eec\u7684\u8bad\u7ec3\u6570\u636e\u4e0d\u5305\u542bFUNSD\u3001CORD\u3001SROIE\u7b49\u6d4b\u8bd5\u96c6\u7684\u4efb\u4f55\u6837\u672c\uff0c\u4ee5\u786e\u4fdd\u8bc4\u4f30\u7684\u516c\u5e73\u6027\u548c\u5bf9\u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b\u7684\u771f\u5b9e\u68c0\u9a8c\u3002<br>\u5bf9\u6bd4\u65b9\u6cd5\uff1a<br>SOTA OCR-dependent methods: \u5982UDoc, StructEqTable\u7b49\u3002<br>SOTA OCR-free methods: \u5982Donut, Pix2Struct\u7b49\u3002<br>\u8bc4\u4f30\u6307\u6807\uff1a<br>\u6839\u636e\u4e0d\u540c\u4efb\u52a1\u91c7\u7528\u76f8\u5e94\u6307\u6807\uff0c\u5982F1-score\uff08\u4fe1\u606f\u62bd\u53d6\uff09\u3001ANLS\uff08DocVQA\uff09\u7b49\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5b9e\u9a8c\u7ed3\u679c<\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"435\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-16-1024x435.png\" alt=\"\" class=\"wp-image-4352\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-16-1024x435.png 1024w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-16-300x127.png 300w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-16-768x326.png 768w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-16.png 1139w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"377\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-17-1024x377.png\" alt=\"\" class=\"wp-image-4353\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-17-1024x377.png 1024w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-17-300x110.png 300w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-17-768x283.png 768w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/11\/image-17.png 1105w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u6838\u5fc3\u5185\u5bb9\u603b\u7ed3<\/h2>\n\n\n\n<p>\u201c\u6587\u6863\u6807\u8bb0\u8bed\u8a00\u201d\u8868\u793a\u6cd5\uff1a\u6700\u6838\u5fc3\u7684\u521b\u65b0\u3002\u7528\u4e00\u79cd\u7b80\u6d01\u3001\u7ed3\u6784\u5316\u3001LLM\u539f\u751f\u53cb\u597d\u7684\u65b9\u5f0f\uff0c\u7edf\u4e00\u7f16\u7801\u4e86\u6587\u6863\u7684\u8bed\u4e49\u7c7b\u578b\u3001\u5c42\u6b21\u7ed3\u6784\u548c\u7a7a\u95f4\u5e03\u5c40\u4fe1\u606f\u3002<br>\u89e3\u8026\u7684\u4e24\u9636\u6bb5\u67b6\u6784\uff1a\u5c06\u590d\u6742\u7684\u89c6\u89c9\u611f\u77e5\u4efb\u52a1\uff08\u5e03\u5c40\u5206\u6790\u3001\u7c97\u7565OCR\uff09\u4e0e\u9ad8\u7ea7\u8bed\u4e49\u7406\u89e3\u4efb\u52a1\uff08\u7531LLM\u5b8c\u6210\uff09\u89e3\u8026\u3002\u8fd9\u79cd\u6a21\u5757\u5316\u8bbe\u8ba1\u63d0\u9ad8\u4e86\u7cfb\u7edf\u7684\u7075\u6d3b\u6027\u3001\u53ef\u7ef4\u62a4\u6027\u548c\u6548\u7387\u3002<br>\u9ad8\u6548\u7684\u8bad\u7ec3\u8303\u5f0f\uff1a\u901a\u8fc7\u51bb\u7ed3LLM\u5e76\u4ec5\u5fae\u8c03\u8f7b\u91cf\u7ea7\u7ec4\u4ef6\uff0c\u6781\u5927\u5730\u964d\u4f4e\u4e86\u8ba1\u7b97\u8d44\u6e90\u9700\u6c42\uff0c\u4f7f\u5f97\u5728\u6709\u9650\u8d44\u6e90\u4e0b\u4e5f\u80fd\u90e8\u7f72\u5f3a\u5927\u7684DU\u6a21\u578b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f5c\u8005\uff1aWenhui Liao, Jiapeng Wang, Hongliang Li\u7b49\u5355\u4f4d\uff1a\u534e\u5357\u7406\u5de5\u5927\u5b66\uff0c\u963f\u91cc [&hellip;]<\/p>\n","protected":false},"author":66,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-4349","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4349","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/66"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=4349"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4349\/revisions"}],"predecessor-version":[{"id":4354,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4349\/revisions\/4354"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=4349"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=4349"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=4349"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}