{"id":4631,"date":"2025-12-16T10:18:39","date_gmt":"2025-12-16T02:18:39","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=4631"},"modified":"2025-12-16T10:18:39","modified_gmt":"2025-12-16T02:18:39","slug":"202512%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-kap-mllm-assisted-ocr-text-enhancement-for-hybrid-retrieval-in-chinese-non-narrative-documents","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202512%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-kap-mllm-assisted-ocr-text-enhancement-for-hybrid-retrieval-in-chinese-non-narrative-documents\/","title":{"rendered":"202512\u8bba\u6587\u7814\u8bfb-KAP: MLLM-assisted OCR Text Enhancement for Hybrid Retrieval in Chinese Non-Narrative Documents"},"content":{"rendered":"\n<p>\u4f5c\u8005\uff1aHsin-Ling Hsu , Ping-Sheng Lin\u7b49<br>\u5355\u4f4d\uff1a\u53f0\u6e7e\u653f\u6cbb\u5927\u5b66<br>\u6765\u6e90\uff1aArxiv<br>\u65f6\u95f4\uff1a2025.03<\/p>\n\n\n\n<p>\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a<\/p>\n\n\n\n<p>1.1 \u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u4e0e\u6df7\u5408\u68c0\u7d22<br>RAG\uff08Retrieval-Augmented Generation\uff09\u5df2\u6210\u4e3a\u77e5\u8bc6\u5bc6\u96c6\u578b\u4efb\u52a1\uff08\u5982\u95ee\u7b54\u7cfb\u7edf\uff09\u7684\u6838\u5fc3\u6280\u672f\u3002<br>\u6df7\u5408\u68c0\u7d22\uff08Hybrid Retrieval\uff09\u7ed3\u5408\u4e86\uff1a<br>\u7a00\u758f\u68c0\u7d22\uff08Sparse Retrieval\uff0c\u5982 BM25\uff09\uff1a\u57fa\u4e8e\u5173\u952e\u8bcd\u5339\u914d\uff1b<br>\u7a20\u5bc6\u68c0\u7d22\uff08Dense Retrieval\uff09\uff1a\u57fa\u4e8e\u8bed\u4e49\u5d4c\u5165\u76f8\u4f3c\u5ea6\u3002<br>1.2 \u975e\u53d9\u8ff0\u6027\u6587\u6863\u7684\u6311\u6218<br>\u975e\u53d9\u8ff0\u6027\u6587\u6863\uff08Non-narrative documents\uff09\u5982\u8868\u683c\u3001\u8d22\u52a1\u62a5\u8868\u7b49\u5177\u6709\u4ee5\u4e0b\u7279\u70b9\uff1a<br>\u7ed3\u6784\u590d\u6742\uff08\u542b\u5927\u91cf\u8868\u683c\u3001\u5bf9\u9f50\u683c\u5f0f\uff09\uff1b<br>\u8bcd\u6c47\u4e30\u5bcc\u3001\u672f\u8bed\u4e13\u4e1a\uff1b<br>\u7e41\u4f53\u4e2d\u6587\u5b58\u5728\u5927\u91cf\u540c\u4e49\u8bcd\u3001\u8bed\u5e8f\u654f\u611f\u95ee\u9898\u3002<br>OCR\u5904\u7406\u8fd9\u4e9b\u95ee\u9898\u65f6\u6613\u51fa\u9519\uff1a<br>\u5b57\u7b26\u8bc6\u522b\u9519\u8bef\uff08\u5982\u201c0\u201d\u2192\u201cO\u201d\u3001\u201c1\u201d\u2192\u201cl\u201d\uff09\uff1b<br>\u8868\u683c\u7ed3\u6784\u4e22\u5931\uff0c\u5bfc\u81f4\u6570\u636e\u9519\u4f4d\uff1b<br>\u6587\u672c\u788e\u7247\u5316\uff0c\u7834\u574f\u8bed\u4e49\u8fde\u8d2f\u6027\u3002<br>1.3 \u73b0\u6709\u65b9\u6cd5\u7684\u4e0d\u8db3<br>\u73b0\u6709OCR\u540e\u5904\u7406\u65b9\u6cd5\u4e3b\u8981\u5173\u6ce8\u8bed\u8a00\u6d41\u7545\u6027\u6216\u53ef\u8bfb\u6027\uff0c\u800c\u975e\u68c0\u7d22\u6027\u80fd\u4f18\u5316\u3002<br>\u7f3a\u4e4f\u9488\u5bf9\u6df7\u5408\u68c0\u7d22\u53cc\u91cd\u8981\u6c42\uff08\u5173\u952e\u8bcd\u5339\u914d + \u8bed\u4e49\u7406\u89e3\uff09\u7684\u6587\u672c\u9884\u5904\u7406\u7b56\u7565\u3002<\/p>\n\n\n\n<p>\u7814\u7a76\u95ee\u9898<\/p>\n\n\n\n<p>\u8be5\u6587\u65e8\u5728\u89e3\u51b3\u4ee5\u4e0b\u6838\u5fc3\u95ee\u9898\uff1a<br>\u5982\u4f55\u5c06OCR\u8f93\u51fa\u7684\u566a\u58f0\u6587\u672c\u8f6c\u5316\u4e3a\u540c\u65f6\u9002\u914d\u7a00\u758f\u68c0\u7d22\uff08BM25\uff09\u548c\u7a20\u5bc6\u68c0\u7d22\uff08Embedding\uff09\u7684\u9ad8\u8d28\u91cf\u3001\u7ed3\u6784\u4fdd\u7559\u3001\u8bed\u4e49\u5b8c\u6574\u7684\u6587\u672c\uff1f<br>\u5177\u4f53\u5305\u62ec\uff1a<br>\u4fee\u6b63OCR\u5b57\u7b26\u9519\u8bef\uff1b<br>\u91cd\u5efa\u8868\u683c\u7b49\u7ed3\u6784\u4fe1\u606f\uff1b<br>\u91cd\u5199\u6587\u672c\u4ee5\u63d0\u5347\u5173\u952e\u8bcd\u8986\u76d6\uff08BM25\u53cb\u597d\uff09\u548c\u8bed\u4e49\u8868\u8fbe\uff08Dense Retrieval\u53cb\u597d\uff09\u3002<\/p>\n\n\n\n<p>\u89e3\u51b3\u529e\u6cd5-KAP\u6846\u67b6<\/p>\n\n\n\n<p>KAP\u662f\u4e00\u4e2a\u4e24\u9636\u6bb5\u9884\u5904\u7406\u6846\u67b6\uff1a<br>\u9636\u6bb51\uff1aOCR\u5904\u7406<br>\u4f7f\u7528 Tesseract OCR \u63d0\u53d6PDF\u4e2d\u7684\u539f\u59cb\u6587\u672c\u3002<br>\u8f93\u51fa\u901a\u5e38\u5305\u542b\u9519\u5b57\u3001\u683c\u5f0f\u6df7\u4e71\u3001\u8868\u683c\u7ed3\u6784\u4e22\u5931\u7b49\u95ee\u9898\u3002<br>\u9636\u6bb52\uff1aMLLM\u540e\u5904\u7406\uff08\u6838\u5fc3\u521b\u65b0\uff09<br>\u5229\u7528\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\uff08MLLM\uff0c\u5982 Claude-3.7-Sonnet\uff09\u5bf9OCR\u8f93\u51fa\u8fdb\u884c\u589e\u5f3a\u3002\u8be5\u9636\u6bb5\u901a\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1\u7684Prompt\u6a21\u677f\uff08\u539f\u4e3a\u7e41\u4f53\u4e2d\u6587\uff09\uff0c\u5b9e\u73b0\u4e09\u5927\u529f\u80fd\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"879\" height=\"269\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/12\/image-28.png\" alt=\"\" class=\"wp-image-4632\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/12\/image-28.png 879w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/12\/image-28-300x92.png 300w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/12\/image-28-768x235.png 768w\" sizes=\"auto, (max-width: 879px) 100vw, 879px\" \/><\/figure>\n\n\n\n<p>3.1 \u9519\u8bef\u6821\u6b63\uff08Error Correction\uff09<br>\u4fee\u6b63OCR\u4ea7\u751f\u7684\u62fc\u5199\u9519\u8bef\u3001\u6570\u5b57\u9519\u8bc6\u3001\u6807\u70b9\u9519\u4f4d\u3001\u8bed\u5e8f\u98a0\u5012\u7b49\u3002<br>\u4fdd\u6301\u8bed\u4e49\u4e0d\u53d8\u7684\u524d\u63d0\u4e0b\u63d0\u5347\u8bed\u8a00\u6d41\u7545\u6027\u3002<br>\u793a\u4f8b\uff1a\u5c06\u201c\u516c\u56f8\u201d\u7ea0\u6b63\u4e3a\u201c\u516c\u53f8\u201d\u3002<br>3.2 \u5e03\u5c40\u611f\u77e5\u7684\u683c\u5f0f\u91cd\u5efa\uff08Layout-Aware Format Reconstruction\uff09<br>\u5173\u952e\u521b\u65b0\uff1aMLLM\u53ef\u540c\u65f6\u63a5\u6536OCR\u6587\u672c + \u539f\u59cbPDF\u56fe\u50cf\u3002<br>\u5229\u7528\u89c6\u89c9\u4fe1\u606f\u7406\u89e3\u539f\u59cb\u5e03\u5c40\uff08\u5982\u8868\u683c\u884c\u5217\u5173\u7cfb\u3001\u6807\u9898\u4f4d\u7f6e\uff09\uff0c\u91cd\u5efa\u7ed3\u6784\u3002<br>\u7279\u522b\u5f3a\u8c03\u5bf9\u8868\u683c\u5185\u5bb9\u7684\u8fd8\u539f\u3002<br>Prompt\u4e2d\u660e\u786e\u6307\u793a\uff1a\u201c\u53ef\u53c2\u8003\u9644\u56fe\u7406\u89e3\u539f\u6587\u6392\u7248\uff08\u5982\u8868\u683c\u3001\u53d9\u8ff0\u53e5\uff09\uff0c\u5e76\u5224\u65ad\u5404\u6587\u5b57\/\u6570\u5b57\u5728\u539f\u6587\u4e2d\u7684\u4f4d\u7f6e\u4e0e\u542b\u4e49\u3002\u201d<\/p>\n\n\n\n<p>3.3 \u68c0\u7d22\u611f\u77e5\u7684\u91cd\u5199\uff08Retrieval-Aware Rewriting\uff09<br>\u8fd9\u662fKAP\u6700\u6838\u5fc3\u7684\u8bbe\u8ba1\uff0c\u5206\u4e3a\u4e24\u7c7b\u4f18\u5316\uff1a<br>(a) \u7a20\u5bc6\u68c0\u7d22\u4f18\u5316\uff08Dense Retrieval Friendly\uff09<br>\u5c06\u8868\u683c\/\u975e\u7eaf\u53d9\u8ff0\u5185\u5bb9\u8f6c\u5316\u4e3a\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u53e5\uff0c\u4fbf\u4e8e\u8bed\u4e49\u5d4c\u5165\u6a21\u578b\u7406\u89e3\u3002<br>\u793a\u4f8b\uff1a<br>\u539f\u59cb\u8868\u683c\uff1a\u65e5\u671f | \u516c\u53f8 | \u91d1\u989d \u2192 2025\/03\/03 | XX\u516c\u53f8 | $10,000<br>\u91cd\u5199\u540e\uff1a\u201c2025\u5e743\u67083\u65e5\uff0cXX\u516c\u53f8\u8bb0\u5f55\u4e86\u4e00\u7b14$10,000\u7684\u4ea4\u6613\u3002\u201d<br>(b) \u7a00\u758f\u68c0\u7d22\u4f18\u5316\uff08BM25 Friendly\uff09<br>\u5728\u4fdd\u7559\u539f\u59cb\u5173\u952e\u8bcd\u7684\u57fa\u7840\u4e0a\uff0c\u81ea\u7136\u878d\u5165\u540c\u4e49\u8bcd\u6216\u8fd1\u4e49\u8868\u8fbe\uff0c\u63d0\u5347\u5173\u952e\u8bcd\u53ec\u56de\u7387\u3002<br>\u907f\u514d\u8fc7\u5ea6\u66ff\u6362\u5f71\u54cd\u5411\u91cf\u68c0\u7d22\u6548\u679c\u3002<br>\u793a\u4f8b\uff1a<br>\u539f\u6587\uff1a\u201c\u7cfb\u7d71\u53ef\u4ee5\u5206\u6790\u6578\u64da\uff0c\u4ee5\u63d0\u5347\u5546\u696d\u6c7a\u7b56\u80fd\u529b\u3002\u201d<br>\u91cd\u5199\uff1a\u201c\u7cfb\u7d71\u80fd\u5920\u5206\u6790\u6578\u64da\u8207\u76f8\u95dc\u8cc7\u8a0a\uff0c\u5e6b\u52a9\u4f01\u696d\u6216\u516c\u53f8\u505a\u51fa\u66f4\u7cbe\u78ba\u7684\u6c7a\u7b56\uff0c\u4e26\u63d0\u5347\u6574\u9ad4\u7b56\u7565\u3002\u201d<br>\u201c\u6578\u64da\u201d \u2192 \u8865\u5145\u201c\u8cc7\u8a0a\u201d\uff1b<br>\u201c\u6c7a\u7b56\u201d \u2192 \u6269\u5c55\u4e3a\u201c\u6c7a\u7b56\u8207\u7b56\u7565\u201d\u3002<br>\u6ce8\u610f\uff1a\u91cd\u5199\u987a\u5e8f\u662f\u5148\u505aDense\u4f18\u5316\uff0c\u518d\u505aBM25\u4f18\u5316\uff0c\u786e\u4fdd\u4e24\u8005\u517c\u5bb9\u3002<\/p>\n\n\n\n<p>\u5b9e\u9a8c<\/p>\n\n\n\n<p>\u6570\u636e\u96c6\uff1a<br>\u6765\u6e90\uff1aE.SUN Bank \u63d0\u4f9b\u7684\u975e\u516c\u5f00\u6570\u636e\u96c6\uff08AI CUP 2024\u7ade\u8d5b\uff09\u3002<br>\u5305\u542b\u4e09\u7c7b\u6587\u6863\uff1aFAQ\uff08JSON\uff09\u3001\u4fdd\u9669\u6761\u6b3e\uff08PDF\uff09\u3001\u8d22\u52a1\u62a5\u544a\uff08PDF\uff0c\u91cd\u70b9\u7814\u7a76\u5bf9\u8c61\uff09\u3002<br>\u8d22\u52a1\u62a5\u544a\u542b\u5927\u91cf\u8868\u683c\uff0c\u7b26\u5408\u201c\u975e\u53d9\u8ff0\u6027\u201d\u5b9a\u4e49\u3002<br>\u9a8c\u8bc1\u96c6\uff1a\u6bcf\u7c7b50\u4e2a\u95ee\u9898\uff0c\u5171150\u4e2a\uff1b\u901a\u8fc7LLM\u6570\u636e\u589e\u5f3a\u6269\u5c55\u81f3500\u4e2a\u95ee\u9898\uff0c\u91c7\u75289\u79cd\u53d8\u6362\u7b56\u7565\uff08\u5982\u540c\u4e49\u8bcd\u66ff\u6362\u3001\u5173\u952e\u8bcd\u62bd\u53d6\u3001\u53e5\u5f0f\u53d8\u6362\u7b49\uff09\u3002<br>\u6a21\u578b\u9009\u62e9<br>OCR\u5f15\u64ce\uff1aTesseract\uff08\u652f\u6301\u7e41\u4f53\u4e2d\u6587\uff09\uff1b<br>MLLM\uff1aClaude-3.7-Sonnet\uff08\u5f3a\u5927\u591a\u6a21\u6001\u7406\u89e3\u80fd\u529b\uff09\uff1b<br>\u7a20\u5bc6\u68c0\u7d22\u6a21\u578b\uff1atext-embedding-3-large\uff08OpenAI\uff09\uff1b<br>\u5206\u8bcd\u5668\uff1aJieba\uff08\u5e26\u7e41\u4f53\u8bcd\u5178\uff09\u3002<br>Chunking\u7b56\u7565<br>\u5148\u6309\u9875\u9762\u5206\u5272\uff0c\u4fdd\u7559\u4e0a\u4e0b\u6587\uff1b<br>\u518d\u7528\u9012\u5f52\u5206\u5757\uff08Recursive Chunking\uff09\uff0c\u5757\u5927\u5c0f8000 tokens\uff0c\u91cd\u53e0500 tokens\uff0c\u907f\u514d\u8868\u683c\u88ab\u5207\u788e\u3002<\/p>\n\n\n\n<p>Baseline\uff1a\u76f4\u63a5\u4f7f\u7528Tesseract OCR\u8f93\u51fa\uff1b<br>Ablation Study\uff08\u6d88\u878d\u5b9e\u9a8c\uff09\uff1a<br>KAP w\/o Vision\uff1a\u7981\u7528\u56fe\u50cf\u8f93\u5165\uff0c\u4ec5\u7528OCR\u6587\u672c\uff1b<br>KAP w\/o OCR Text\uff1a\u4ec5\u7528\u56fe\u50cf\uff0c\u8ba9MLLM\u4ece\u5934\u63d0\u53d6\uff1b<br>KAP w\/o Rewrite\uff1a\u4e0d\u505a\u68c0\u7d22\u611f\u77e5\u91cd\u5199\uff1b<br>Full KAP\uff1a\u5b8c\u6574\u6d41\u7a0b\u3002<\/p>\n\n\n\n<p>\u5b9e\u9a8c\u7ed3\u679c<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"445\" height=\"448\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/12\/image-29.png\" alt=\"\" class=\"wp-image-4633\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/12\/image-29.png 445w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/12\/image-29-298x300.png 298w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/12\/image-29-150x150.png 150w\" sizes=\"auto, (max-width: 445px) 100vw, 445px\" \/><\/figure>\n\n\n\n<p>\u6838\u5fc3\u5185\u5bb9\u603b\u7ed3\uff1a<br>1.\u5c06MLLM\u7528\u4e8e\u9762\u5411\u6df7\u5408\u68c0\u7d22\u7684OCR\u540e\u5904\u7406\uff0c\u800c\u975e\u901a\u7528\u6587\u672c\u4fee\u590d\uff1b<br>2.\u63d0\u51fa\u201c\u68c0\u7d22\u611f\u77e5\u91cd\u5199\u201d\u53cc\u8def\u5f84\u7b56\u7565\uff1a<br>\u7a20\u5bc6\u8def\u5f84\uff1a\u8868\u683c\u2192\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\uff1b<br>\u7a00\u758f\u8def\u5f84\uff1a\u5173\u952e\u8bcd\u4fdd\u7559+\u540c\u4e49\u6269\u5c55\uff1b<br>3.\u591a\u6a21\u6001Prompt\u5de5\u7a0b\uff1a\u5728\u540c\u4e00Prompt\u4e2d\u6574\u5408\u7ea0\u9519\u3001\u7ed3\u6784\u7406\u89e3\u3001\u91cd\u5199\u6307\u4ee4\uff0c\u9ad8\u6548\u7edf\u4e00\uff1b<br>4.\u6784\u5efa\u9762\u5411\u6df7\u5408\u68c0\u7d22\u7684\u9a8c\u8bc1\u96c6\u589e\u5f3a\u65b9\u6cd5\uff1a\u901a\u8fc79\u79cdLLM\u9a71\u52a8\u7684\u67e5\u8be2\u53d8\u4f53\uff0c\u5168\u9762\u6d4b\u8bd5\u7cfb\u7edf\u9c81\u68d2\u6027\uff1b<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f5c\u8005\uff1aHsin-Ling Hsu , Ping-Sheng Lin\u7b49\u5355\u4f4d\uff1a\u53f0\u6e7e\u653f\u6cbb\u5927\u5b66\u6765\u6e90\uff1aArxiv\u65f6\u95f4\uff1a2 [&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-4631","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4631","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=4631"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4631\/revisions"}],"predecessor-version":[{"id":4634,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4631\/revisions\/4634"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=4631"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=4631"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=4631"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}