{"id":2506,"date":"2025-03-14T16:24:04","date_gmt":"2025-03-14T08:24:04","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=2506"},"modified":"2025-03-14T16:25:22","modified_gmt":"2025-03-14T08:25:22","slug":"202503-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-table-meets-llm-can-large-language-models-understand-structured-table-data-a-benchmark-and-empirical-study","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202503-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-table-meets-llm-can-large-language-models-understand-structured-table-data-a-benchmark-and-empirical-study\/","title":{"rendered":"202503 \u8bba\u6587\u7814\u8bfb-Table Meets LLM: Can Large Language Models Understand Structured Table Data? A Benchmark and Empirical Study"},"content":{"rendered":"\n

\u4f5c\u8005\uff1a<\/strong> Yuan Sui, Shi Han, Mengyu Zhou, Mingjie Zhou, Dongmei Zhang<\/p>\n\n\n\n

\u6765\u6e90\uff1a<\/strong> International Conference on Web Search and Data Mining\uff08ACM\uff09<\/p>\n\n\n\n

\u65f6\u95f4\uff1a<\/strong> 2024.04<\/p>\n\n\n\n

\u4e00\u3001\u80cc\u666f<\/strong><\/h2>\n\n\n\n

    1.\u7ed3\u6784\u5316\u6570\u636e\u7684\u91cd\u8981\u6027\uff1a\u8868\u683c\u5e7f\u6cdb\u5e94\u7528\u4e8e\u95ee\u7b54\u3001\u4e8b\u5b9e\u9a8c\u8bc1\u7b49\u4efb\u52a1\uff0c\u4f46\u5176\u590d\u6742\u7ed3\u6784\uff08\u5982\u5408\u5e76\u5355\u5143\u683c\u3001\u884c\u5217\u5173\u7cfb\uff09\u5bf9\u6a21\u578b\u7406\u89e3\u80fd\u529b\u63d0\u51fa\u6311\u6218\u3002<\/p>\n\n\n\n

2.\u73b0\u6709\u5de5\u4f5c\u7f3a\u4e4f\u7cfb\u7edf\u6027\u8bc4\u4f30\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5bf9\u8868\u683c\u7ed3\u6784\u7406\u89e3\u7684\u6df1\u5ea6\uff0c\u4e14\u8f93\u5165\u8bbe\u8ba1\uff08\u5982\u5e8f\u5217\u5316\u65b9\u6cd5\uff09\u7f3a\u4e4f\u7edf\u4e00\u6807\u51c6\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u6838\u5fc3\u521b\u65b0<\/strong><\/h2>\n\n\n\n

1.SUC\u57fa\u51c6\u6d4b\u8bd5\uff1a\u9996\u4e2a\u7cfb\u7edf\u6027\u8bc4\u4f30LLMs\u8868\u683c\u7ed3\u6784\u7406\u89e3\u80fd\u529b\u7684\u57fa\u51c6\uff0c\u6db5\u76d67\u9879\u4efb\u52a1\uff08\u8868\u5206\u533a\u3001\u8868\u5927\u5c0f\u68c0\u6d4b\u3001\u5408\u5e76\u5355\u5143\u683c\u68c0\u6d4b\u3001\u5355\u5143\u683c\u67e5\u627e\u3001\u884c\u5217\u68c0\u7d22\u7b49\uff09\uff0c\u8986\u76d6\u7ed3\u6784\u89e3\u6790\u4e0e\u4fe1\u606f\u68c0\u7d22\u80fd\u529b\u3002<\/p>\n\n\n\n

    2.\u81ea\u6211\u589e\u5f3a\u63d0\u793a\u65b9\u6cd5\uff1a\u901a\u8fc7\u4e24\u9636\u6bb5\u63d0\u793a\uff08\u9996\u5148\u751f\u6210\u4e2d\u95f4\u7ed3\u6784\u77e5\u8bc6\uff0c\u518d\u7ed3\u5408\u4e0b\u6e38\u4efb\u52a1\uff09\u63d0\u5347\u4e86\u6a21\u578b\u6027\u80fd\uff0c\u65e0\u9700\u989d\u5916\u8bad\u7ec3\uff0c\u9002\u914d\u591a\u79cd\u8868\u683c\u683c\u5f0f\uff08HTML\u3001JSON\u7b49\uff09<\/p>\n\n\n\n

\u4e09\u3001\u65b9\u6cd5<\/strong><\/h2>\n\n\n\n

3.1 SUC\u57fa\u51c6\u8bbe\u8ba1<\/strong><\/h3>\n\n\n\n

\u8861\u91cf\u8868\u683c\u7ed3\u6784\u7406\u89e3\u80fd\u529b\u7684\u4e24\u4e2a\u65b9\u9762\uff1a<\/strong><\/p>\n\n\n\n

1.\u7ed3\u6784\u89e3\u6790\u80fd\u529b\uff08Partition & Parsing\uff09<\/strong><\/p>\n\n\n\n

2.\u4fe1\u606f\u68c0\u7d22\u80fd\u529b\uff08Search & Retrieval\uff09<\/strong><\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

    1.\u8861\u91cf<\/strong>\u7ed3\u6784\u89e3\u6790\u80fd\u529b\u7684\u4efb\u52a1\u8bbe\u8ba1\uff1a<\/strong><\/p>\n\n\n\n

    1\uff09.\u8868\u5206\u533a\uff08Table Partition\uff09\uff1a\u8981\u6c42\u6a21\u578b\u8bc6\u522b\u8f93\u5165\u6587\u672c\u4e2d\u8868\u683c\u7684\u8fb9\u754c\uff08\u8d77\u59cb\u548c\u7ed3\u675f\u6807\u8bb0\uff09\uff0c\u533a\u5206\u8868\u683c\u4e0e\u5176\u4ed6\u8865\u5145\u4fe1\u606f\uff08\u5982\u63cf\u8ff0\u3001\u4e0a\u4e0b\u6587\uff09\u3002\u8f93\u5165\u8bbe\u8ba1\u5305\u542b\u6df7\u5408\u6587\u672c\uff0c\u6a21\u578b\u9700\u8f93\u51fa\u8868\u683c\u7684\u8d77\u59cb\u548c\u7ed3\u675f\u6807\u8bb0\uff08\u5982HTML\u4e2d\u7684<table>\u548c<\/table>\uff09\u3002<\/p>\n\n\n\n

    2\uff09.\u8868\u5927\u5c0f\u68c0\u6d4b\uff08Size Detection\uff09\uff1a\u7edf\u8ba1\u8868\u683c\u7684\u884c\u6570\u548c\u5217\u6570\uff08\u4f8b\u5982\u8f93\u5165\u201c\u8868\u683c\u6709\u591a\u5c11\u884c\u548c\u5217\uff1f\u201d\uff09\uff0c\u9a8c\u8bc1\u6a21\u578b\u5bf9\u57fa\u672c\u7ed3\u6784\u7279\u5f81\u7684\u63d0\u53d6\u80fd\u529b\u3002<\/p>\n\n\n\n

    3\uff09.\u5408\u5e76\u5355\u5143\u683c\u68c0\u6d4b\uff08Merged Cell Detection\uff09\uff1a\u5b9a\u4f4d\u8868\u683c\u4e2d\u5408\u5e76\u5355\u5143\u683c\u7684\u4f4d\u7f6e\uff08\u5982\u201c\u68c0\u6d4b\u8de8\u8d8a\u591a\u884c\u7684\u5355\u5143\u683c\u201d\uff09\uff0c\u6d4b\u8bd5\u6a21\u578b\u5bf9\u590d\u6742\u7ed3\u6784\u7684\u7406\u89e3\u3002<\/strong><\/p>\n\n\n\n

    2.\u8861\u91cf<\/strong>\u4fe1\u606f\u68c0\u7d22\u80fd\u529b\u7684\u4efb\u52a1\u8bbe\u8ba1<\/strong>\uff1a<\/p>\n\n\n\n

    1\uff09.\u5355\u5143\u683c\u67e5\u627e\uff08Cell Lookup\uff09\uff1a\u6839\u636e\u5185\u5bb9\u5b9a\u4f4d\u5355\u5143\u683c\u7684\u884c\u5217\u7d22\u5f15\uff08\u5982\u201c\u67e5\u627e\u2018Antoine\u2019\u7684\u4f4d\u7f6e\u201d\uff09\u3002<\/p>\n\n\n\n

    2\uff09.\u53cd\u5411\u67e5\u627e\uff08Reverse Lookup\uff09\uff1a\u6839\u636e\u884c\u5217\u7d22\u5f15\u8fd4\u56de\u5355\u5143\u683c\u5185\u5bb9\uff08\u5982\u201c\u7b2c3\u884c\u7b2c2\u5217\u7684\u503c\u662f\u4ec0\u4e48\uff1f\u201d\uff09\u3002<\/p>\n\n\n\n

    3\uff09.\u884c\u5217\u68c0\u7d22\uff08Column\/Row Retrieval\uff09\uff1a\u6309\u5217\u540d\u6216\u884c\u53f7\u63d0\u53d6\u6574\u5217\u6216\u6574\u884c\u6570\u636e\uff08\u5982\u201c\u5217\u51fa\u2018Year\u2019\u5217\u7684\u6240\u6709\u503c\u201d\uff09<\/p>\n\n\n\n

3.2 \u81ea\u6211\u589e\u5f3a\u63d0\u793a\uff08Self-Augmented Prompting\uff09<\/strong><\/h3>\n\n\n\n
\"\"<\/figure>\n\n\n\n

    <\/strong>\u7b2c\u4e00\u9636\u6bb5\uff1a\u4e2d\u95f4\u77e5\u8bc6\u751f\u6210<\/strong><\/p>\n\n\n\n

    <\/strong>\u5173\u952e\u503c\u8303\u56f4\u8bc6\u522b\uff1a\u8981\u6c42\u6a21\u578b\u751f\u6210\u8868\u683c\u4e2d\u7684\u5173\u952e\u6570\u503c\u8303\u56f4\uff08\u5982\u201c\u8bc6\u522b\u2018Races\u2019\u5217\u7684\u6700\u5c0f\u503c\u548c\u6700\u5927\u503c\u201d\uff09\u3002\u683c\u5f0f\u81ea\u89e3\u91ca\uff1a\u6a21\u578b\u81ea\u52a8\u751f\u6210\u8868\u683c\u683c\u5f0f\u63cf\u8ff0\uff08\u5982\u201c\u8868\u683c\u4f7f\u7528HTML\u6807\u7b7e\uff0c\u6bcf\u884c\u7531<tr>\u5305\u88f9\u201d\uff09\u3002\u7ed3\u6784\u7279\u5f81\u63d0\u53d6\uff1a\u8f93\u51fa\u5408\u5e76\u5355\u5143\u683c\u4f4d\u7f6e\u3001\u8868\u5934\u7c7b\u578b\u7b49\uff08\u5982\u201c\u7b2c2\u884c\u7b2c1\u5217\u8de82\u884c\u201d\uff09\u3002<\/p>\n\n\n\n

    \u7b2c\u4e8c\u9636\u6bb5\uff1a\u4e0b\u6e38\u4efb\u52a1\u4f18\u5316<\/strong><\/p>\n\n\n\n

    <\/strong>\u7ed3\u5408\u4e2d\u95f4\u77e5\u8bc6\uff1a\u5c06\u751f\u6210\u7684\u4e2d\u95f4\u77e5\u8bc6\u4f5c\u4e3a\u9644\u52a0\u63d0\u793a\u8f93\u5165\uff0c\u4f18\u5316\u95ee\u7b54\u6216\u4e8b\u5b9e\u9a8c\u8bc1\u4efb\u52a1\u3002<\/p>\n\n\n\n

\u4f18\u52bf\uff1a<\/strong><\/p>\n\n\n\n

    <\/strong>\u6a21\u578b\u65e0\u5173\u6027\uff1a\u65e0\u9700\u5fae\u8c03\uff0c\u9002\u914dGPT-3.5\u3001GPT-4\u7b49\u4e0d\u540c\u6a21\u578b\u3002<\/p>\n\n\n\n

    \u52a8\u6001\u77e5\u8bc6\u6ce8\u5165\uff1a\u901a\u8fc7\u4e2d\u95f4\u6b65\u9aa4\u5f15\u5bfc\u6a21\u578b\u5173\u6ce8\u7ed3\u6784\u7279\u5f81\uff0c\u51cf\u5c11\u5bf9\u663e\u5f0f\u6807\u6ce8\u7684\u4f9d\u8d56\u3002<\/p>\n\n\n\n

    \u7075\u6d3b\u9002\u914d\u591a\u4efb\u52a1\uff1a\u9002\u7528\u4e8e\u8868\u683c\u95ee\u7b54\uff08HybridQA\uff09\u3001\u4e8b\u5b9e\u9a8c\u8bc1\uff08TabFact\uff09\u3001\u6587\u672c\u751f\u6210\uff08ToTTo\uff09\u7b49\u573a\u666f\u3002<\/p>\n\n\n\n

\u56db\u3001\u5b9e\u9a8c<\/strong><\/h2>\n\n\n\n

4.1\u6570\u636e\u51c6\u5907<\/strong><\/h3>\n\n\n\n

1.\u91c7\u7528\u4e94\u4e2a\u516c\u5f00\u6570\u636e\u96c6\uff1aTabFact\u3001HybridQA\u3001SQA\u3001Feverous\u3001ToTTo\uff0c\u6765\u8bc4\u4f30LLM\u5728\u8868\u683c\u63a8\u7406\u548c\u4e8b\u5b9e\u9a8c\u8bc1\u7b49\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n

2.\u91c7\u7528\u4e0d\u540c\u683c\u5f0f\u7684\u6570\u636e\u96c6\uff0c\u5305\u62ecCSV\u3001JSON\u3001XML\u3001markdown\u3001HTML\u548cXLSX\u3002\u786e\u5b9aLLM\u662f\u5426\u6709\u80fd\u529b\u6b63\u786e\u89e3\u6790\u4e0d\u540c\u7684\u683c\u5f0f\u6e90\uff0c\u5e76\u786e\u5b9a\u54ea\u79cd\u7c7b\u578b\u7684\u8f93\u5165\u8bbe\u8ba1\u6700\u9002\u5408LLM\u3002<\/p>\n\n\n\n

4.2  SUC\u57fa\u51c6\u8bc4\u4f30\u5b9e\u9a8c\u7ed3\u679c<\/strong><\/h3>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n

1.HTML\u683c\u5f0f\u6700\u4f18\uff0c\u5728\u8868\u683c\u89e3\u6790\u4efb\u52a1\u4e0a\u6bd4NL+Sep\u683c\u5f0f\u9ad8\u51fa6.76%\u3002<\/p>\n\n\n\n

2.\u96f6\u6837\u672c\u5b66\u4e60\u6548\u679c\u8f83\u5dee\uff0c\u7cfb\u7edf\u5728\u96f6\u6837\u672c\u6761\u4ef6\u4e0b\u7684\u51c6\u786e\u7387\u4e0b\u964d\u7ea630%\u3002<\/p>\n\n\n\n

3.\u8f93\u5165\u987a\u5e8f\u5f71\u54cd\u6027\u80fd\uff0c\u5c06\u5916\u90e8\u4fe1\u606f\uff08\u5982\u95ee\u9898\u63cf\u8ff0\uff09\u7f6e\u4e8e\u8868\u683c\u524d\uff0c\u6709\u52a9\u4e8eLLM\u66f4\u597d\u5730\u7406\u89e3\u8868\u683c\u7ed3\u6784\u3002<\/p>\n\n\n\n

4.\u683c\u5f0f\u8bf4\u660e\u548c\u5206\u533a\u6807\u8bb0\u7684\u4f5c\u7528\uff1a\u683c\u5f0f\u8bf4\u660e\u63d0\u9ad8\u4e86\u7ed3\u6784\u89e3\u6790\u4efb\u52a1\u7684\u51c6\u786e\u6027\uff0c\u4f46\u53ef\u80fd\u964d\u4f4e\u68c0\u7d22\u4efb\u52a1\u7684\u8868\u73b0\u3002\u5206\u533a\u6807\u8bb0\u5728\u5408\u5e76\u5355\u5143\u683c\u68c0\u6d4b\u4efb\u52a1\u4e2d\u63d0\u5347\u4e86\u51c6\u786e\u7387\uff0c\u4f46\u5bf9\u5176\u4ed6\u4efb\u52a1\u5f71\u54cd\u8f83\u5c0f\u3002<\/p>\n\n\n\n

4.3 \u81ea\u6211\u589e\u5f3a\u63d0\u793a\u5b9e\u9a8c<\/strong><\/h3>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n

1.\u81ea\u6211\u589e\u5f3a\u63d0\u793a\uff08Self-augmented Prompting\uff09\u663e\u8457\u63d0\u5347LLM\u6027\u80fd\uff1a\u5173\u952e\u6570\u503c\u8303\u56f4\u8bc6\u522b\u65b9\u6cd5\u5728HybridQA\u4efb\u52a1\u4e0a\u63d0\u9ad8\u4e862.13%\uff0c\u5728SQA\u4efb\u52a1\u4e0a\u63d0\u9ad8\u4e862.72%\u3002\u7ed3\u6784\u5316\u4fe1\u606f\u63cf\u8ff0\u65b9\u6cd5\u5728Feverous\u4efb\u52a1\u4e0a\u63d0\u5347\u4e861.72%\u3002<\/p>\n\n\n\n

2.GPT-4\u5728\u6240\u6709\u4efb\u52a1\u4e0a\u4f18\u4e8eGPT-3.5\uff0c\u5176\u5728TabFact\u4efb\u52a1\u4e0a\u7684\u51c6\u786e\u7387\u63d0\u9ad8\u4e866.36%\u3002<\/p>\n\n\n\n

\u4e94\u3001\u603b\u7ed3\u4e0e\u601d\u8003<\/strong><\/h2>\n\n\n\n

    1. \u8bba\u6587\u63d0\u51fa\u4e86SUC\u57fa\u51c6\u6d4b\u8bd5\u6765\u8bc4\u4f30LLMs\u7684\u7ed3\u6784\u7406\u89e3\u80fd\u529b\uff0c\u540c\u65f6\u901a\u8fc7\u5b9e\u9a8c\u9a8c\u8bc1\u4e86\u9009\u62e9\u5408\u9002\u7684\u8f93\u5165\u8bbe\u8ba1\u7ec4\u5408\u53ef\u4ee5\u663e\u8457\u63d0\u9ad8LLMs\u5bf9\u7ed3\u6784\u5316\u6570\u636e\u7684\u7406\u89e3\u3002<\/p>\n\n\n\n

    2. \u8bba\u6587\u540c\u65f6\u63d0\u51fa\u4e86\u81ea\u589e\u5f3a\u63d0\u793a\u65b9\u6cd5\uff0c\u5e76\u9a8c\u8bc1\u4e86\u5176\u5728\u591a\u4e2a\u8868\u683c\u63a8\u7406\u6570\u636e\u96c6\u4e0a\u7684\u6709\u6548\u6027\u3002      <\/p>\n\n\n\n

    3. LLMs\u5177\u6709\u57fa\u672c\u7684\u7ed3\u6784\u7406\u89e3\u80fd\u529b\uff0c\u4f46\u8fdc\u672a\u5b8c\u7f8e\uff0c\u7279\u522b\u662f\u5728\u4e00\u4e9b\u770b\u4f3c\u7b80\u5355\u7684\u4efb\u52a1\u4e0a\uff08\u5982\u8868\u683c\u5927\u5c0f\u68c0\u6d4b\uff09\u3002\u5982\u4f55\u5728\u6b64\u8bba\u6587\u7684\u57fa\u7840\u4e0a\u66f4\u8fdb\u4e00\u6b65\u7684\u63d0\u5347\u5927\u6a21\u578b\u5bf9\u8868\u683c\u6570\u636e\u7684\u7406\u89e3\u80fd\u529b\uff0c\u662f\u672a\u6765\u63a2\u7d22\u7684\u5173\u952e\u3002<\/p>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

\u4f5c\u8005\uff1a Yuan Sui, Shi Han, Mengyu Zhou, Mingjie Zhou, Dongm […]<\/p>\n","protected":false},"author":62,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2506","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/2506","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\/62"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=2506"}],"version-history":[{"count":2,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/2506\/revisions"}],"predecessor-version":[{"id":2581,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/2506\/revisions\/2581"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=2506"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=2506"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=2506"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}