{"id":3173,"date":"2025-05-15T10:17:04","date_gmt":"2025-05-15T02:17:04","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=3173"},"modified":"2025-05-15T10:17:04","modified_gmt":"2025-05-15T02:17:04","slug":"202505-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-concept-guided-chain-of-thought-prompting-for-pairwise-comparison-scoring-of-texts-with-large-language-models","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202505-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-concept-guided-chain-of-thought-prompting-for-pairwise-comparison-scoring-of-texts-with-large-language-models\/","title":{"rendered":"202505 \u8bba\u6587\u7814\u8bfb-Concept-Guided Chain-of-Thought Prompting for Pairwise Comparison Scoring of Texts with Large Language Models"},"content":{"rendered":"\n

\u6765\u6e90\uff1aIEEE BigData
\u4f5c\u8005\uff1aPatrick Y. Wu, Jonathan Nagler\u7b49
\u5355\u4f4d\uff1aAmerican University\u3001New York University
\u53d1\u8868\u65f6\u95f4\uff1a2025 \u5e74 2\u6708<\/p>\n\n\n\n

\u4e00\u3001\u7814\u7a76\u80cc\u666f<\/h2>\n\n\n\n
    \n
  1. \u793e\u4f1a\u79d1\u5b66\u4e2d\u7684\u6587\u672c\u8bc4\u5206\u95ee\u9898\uff1a
    \u5728\u653f\u6cbb\u79d1\u5b66\u3001\u4f20\u64ad\u5b66\u7b49\u793e\u4f1a\u79d1\u5b66\u9886\u57df\uff0c\u6587\u672c\u8bc4\u5206\uff08text scoring\uff09\u662f\u4e00\u9879\u6838\u5fc3\u4efb\u52a1\uff0c\u5176\u76ee\u6807\u662f\u6839\u636e\u6587\u672c\u5185\u5bb9\u63a8\u6d4b\u6587\u672c\u5728\u67d0\u4e2a\u9690\u542b\u7ef4\u5ea6\u4e0a\u7684\u201c\u7acb\u573a\u201d\u6216\u201c\u5f3a\u5ea6\u201d\u3002<\/li>\n<\/ol>\n\n\n\n
      \n
    • \u5e38\u89c1\u5e94\u7528\u573a\u666f\u5305\u62ec\uff1a<\/li>\n\n\n\n
    • \u5224\u65ad\u653f\u515a\u8a00\u8bba\u7684\u7acb\u573a<\/li>\n\n\n\n
    • \u6d4b\u91cf\u516c\u4f17\u8206\u60c5\u4e2d\u7684\u6781\u5316\u8d8b\u52bf<\/li>\n\n\n\n
    • \u8ffd\u8e2a\u77ed\u6587\u672c\uff08\u5982\u63a8\u6587\uff09\u4e2d\u7684\u60c5\u7eea\u8868\u8fbe\u53d8\u5316<\/li>\n<\/ul>\n\n\n\n

      2. LLM \u8d4b\u80fd\u4e0b\u7684\u673a\u9047\u4e0e\u6311\u6218\uff1a
      LLM\u5df2\u62e5\u6709\u4e30\u5bcc\u7684\u8bed\u8a00\u4e0e\u6982\u5ff5\u77e5\u8bc6\uff0c\u53ef\u501f\u52a9 prompt \u8fdb\u884c\u96f6\u6837\u672c\u6216\u5c11\u6837\u672c\u5b66\u4e60
      LLM \u76f4\u63a5\u751f\u6210\u6570\u503c\u6253\u5206\u6216\u6807\u7b7e\u4e0d\u900f\u660e\u3001\u96be\u89e3\u91ca
      \u5bf9\u4e8e\u590d\u6742\u4efb\u52a1\uff08\u5982\u63a8\u6587\u4e2d\u7ec6\u5fae\u60c5\u7eea\uff09\uff0c\u9700\u8981\u66f4\u7cbe\u7ec6\u7684\u63a8\u7406\u673a\u5236<\/p>\n\n\n\n

      \u8fd9\u7bc7\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u6587\u672c\u8bc4\u5206\u6846\u67b6\uff1aCGCoT\uff08Concept-Guided Chain-of-Thought\uff09\uff0c\u901a\u8fc7\u4e13\u5bb6\u8bbe\u8ba1\u7684\u6982\u5ff5\u5f15\u5bfcprompt\uff0c\u5f15\u5bfcLLM\u751f\u6210\u7ed3\u6784\u5316\u5206\u6790\uff0c\u7ed3\u5408\u6210\u5bf9\u6bd4\u8f83\u4e0e\u6982\u7387\u5efa\u6a21\uff0c\u5b9e\u73b0\u4f4e\u6807\u6ce8\u6210\u672c\u3001\u5f3a\u6982\u5ff5\u5bf9\u51c6\u3001\u9002\u7528\u4e8e\u77ed\u6587\u672c\u7684\u9ad8\u89e3\u91ca\u6027\u6587\u672c\u8bc4\u5206\u3002<\/p>\n\n\n\n

      \u4e8c\u3001CGCoT\u6846\u67b6<\/h2>\n\n\n\n

      CGCoT\u65b9\u6cd5\uff1a\u901a\u8fc7\u4e13\u5bb6\u8bbe\u8ba1\u7684\u6982\u5ff5\u5f15\u5bfcprompt\uff0c\u5f15\u5bfcLLM\u751f\u6210\u7ed3\u6784\u5316\u5206\u6790\uff0c\u7ed3\u5408\u6210\u5bf9\u6bd4\u8f83\u4e0e\u6982\u7387\u5efa\u6a21<\/p>\n\n\n\n

        \n
      • \u6b65\u9aa4\u4e00\uff1a\u6982\u5ff5\u62c6\u89e3<\/li>\n<\/ul>\n\n\n\n

        \u5bf9\u6bcf\u6761\u6587\u672c\uff08\u5982\u63a8\u6587\uff09\uff0c\u4f7f\u7528 LLM \u56de\u7b54\u4e00\u7cfb\u5217\u4e13\u5bb6\u8bbe\u8ba1\u7684\u5f15\u5bfc\u95ee\u9898\uff08prompt\uff09\uff1a<\/p>\n\n\n\n

        \u4ee5\u201c\u5bf9\u5171\u548c\u515a\u7684\u538c\u6076\u201d\u4e3a\u4f8b\uff0cprompt \u5305\u62ec\uff1a
        \u8bf7\u603b\u7ed3\u8be5\u63a8\u6587\u7684\u4e3b\u8981\u5185\u5bb9
        \u63a8\u6587\u7684\u6279\u8bc4\u5bf9\u8c61\u662f\u5171\u548c\u515a\u8fd8\u662f\u6c11\u4e3b\u515a\uff1f
        \u662f\u5426\u8868\u8fbe\u4e86\u5bf9\u8be5\u5bf9\u8c61\u7684\u538c\u6076\u3001\u6307\u8d23\u6216\u8d1f\u9762\u60c5\u7eea\uff1f
        \u6839\u636e\u4e0a\u8ff0\u56de\u7b54\uff0c\u662f\u5426\u660e\u786e\u8868\u8fbe\u4e86\u5bf9\u8be5\u515a\u6d3e\u7684\u8d1f\u9762\u8bc4\u4ef7\uff1f<\/p>\n\n\n\n

        \u5f97\u5230\u7684\u201c\u7ed3\u6784\u5316\u63cf\u8ff0\u201d\u5373\u4e3a\u8be5\u6587\u672c\u7684\u6982\u5ff5\u7279\u5b9a\u5c55\u5f00\uff08Concept-Specific Breakdown\uff09<\/p>\n\n\n\n

          \n
        • \u6b65\u9aa4\u4e8c\uff1a\u6210\u5bf9\u6bd4\u8f83<\/li>\n<\/ul>\n\n\n\n

          \u5c06\u4e24\u6761\u6587\u672c\u7684\u7ed3\u6784\u5316\u63cf\u8ff0\u62ff\u7ed9 LLM \u6bd4\u8f83\uff0c\u63d0\u95ee\u4f8b\u5982\uff1a
          \u201c\u4ee5\u4e0b\u54ea\u6761 Tweet \u66f4\u8868\u8fbe\u51fa\u5bf9\u5171\u548c\u515a\u7684\u538c\u6076\uff1f\u63cf\u8ff0 1 \u8fd8\u662f\u63cf\u8ff0 2\uff1f\u5982\u90fd\u5dee\u4e0d\u591a\uff0c\u8bf7\u56de\u7b54\u2018Tie\u2019\u3002\u201d<\/p>\n\n\n\n

          \u4e3a\u4e86\u63d0\u5347\u51c6\u786e\u7387\uff1a
          \u5148\u7528 GPT \u56de\u7b54 + \u7406\u7531\u8bf4\u660e
          \u518d\u7528\u7b80\u6d01 prompt \u89e3\u6790\u5176\u7ed3\u8bba\uff08”Tweet 1″ or “Tweet 2” or “Tie”\uff09<\/p>\n\n\n\n

          \u901a\u8fc7 step-by-step \u5f15\u5bfc\uff0c\u8ba9 LLM \u53d1\u6325\u6a21\u5f0f\u8bc6\u522b\u4f18\u52bf\u800c\u975e\u76f4\u63a5\u8bc4\u5206<\/p>\n\n\n\n

            \n
          • \u6b65\u9aa4\u4e09\uff1a\u5f97\u5206\u5efa\u6a21<\/li>\n<\/ul>\n\n\n\n

            \u901a\u8fc7 Bradley-Terry \u6a21\u578b\u8f6c\u5316\u4e3a\u53ef\u6392\u5e8f\u7684\u91cf\u5316\u5206\u6570\uff1a<\/p>\n\n\n\n

              \n
            • \u6bcf\u6b21\u6bd4\u8f83\u7ed3\u679c\u6784\u6210\u4e00\u4e2a\u201c\u80dc\u8d1f\u8bb0\u5f55\u201d\uff08\u8c01\u66f4\u5177\u76ee\u6807\u6982\u5ff5\uff09<\/li>\n\n\n\n
            • \u4f7f\u7528Bradley-Terry \u6a21\u578b\u5bf9\u8fd9\u4e9b\u201c\u6bd4\u8d5b\u7ed3\u679c\u201d\u5efa\u6a21\uff0c\u6c42\u89e3\u6bcf\u6761\u6587\u672c\u7684\u201c\u80fd\u529b\u503c\u201d\u03bb<\/li>\n\n\n\n
            • \u8be5 \u03bb \u5373\u4e3a\u6700\u7ec8\u7684 CGCoT \u6982\u5ff5\u8bc4\u5206\uff0c\u53ef rescale \u81f3 [0, 1] \u533a\u95f4<\/li>\n<\/ul>\n\n\n\n

              \u53ea\u7528\u6781\u5c0f\u91cf\u624b\u6807\u6837\u672c\u7528\u4e8e prompt \u5f00\u53d1\uff0c\u540e\u7eed\u8bc4\u5206\u5168\u81ea\u52a8\u5316<\/p>\n\n\n\n

              \u4e09\u3001\u5b9e\u9a8c<\/h2>\n\n\n\n
                \n
              • \u6709\u6548\u6027\u9a8c\u8bc1<\/li>\n<\/ul>\n\n\n\n
                \"\"<\/figure>\n\n\n\n

                \u5b9e\u9a8c\u4efb\u52a1\uff1a
                \u4ece\u653f\u6cbb\u63a8\u6587\u4e2d\u8bc4\u4f30\u201c\u5bf9\u5171\u548c\u515a\/\u6c11\u4e3b\u515a\u538c\u6076\u201d\u5f3a\u5ea6\uff0c\u9a8c\u8bc1 CGCoT \u7684\u8bc4\u5206\u662f\u5426\u7b26\u5408\u4eba\u7c7b\u5e38\u8bc6\u3002
                \u8bc4\u4f30\u65b9\u5f0f\uff1a
                \u4eba\u5de5\u9009\u53d6\u4ee3\u8868\u6027\u63a8\u6587\uff0c\u770b CGCoT \u5f97\u5206\u6392\u5e8f\u662f\u5426\u7b26\u5408\u76f4\u89c9\u3002
                \u7ed3\u679c\uff1a
                CGCoT \u5206\u6570\u987a\u5e8f\u4e0e\u4eba\u7c7b\u76f4\u89c9\u9ad8\u5ea6\u4e00\u81f4\uff0c\u5c55\u793a\u51fa\u826f\u597d\u7684\u8bed\u4e49\u611f\u77e5\u80fd\u529b\u4e0e\u6982\u5ff5\u5bf9\u51c6\u80fd\u529b<\/p>\n\n\n\n

                  \n
                • \u4e0e\u5176\u4ed6\u65e0\u76d1\u7763\u65b9\u6cd5\u7684\u5bf9\u6bd4\u5b9e\u9a8c<\/li>\n<\/ul>\n\n\n\n
                  \u65b9\u6cd5<\/strong><\/td>\u5185\u5bb9<\/strong><\/td><\/tr>
                  Wordfish\uff08\u7ecf\u5178\u65e0\u76d1\u7763\uff09<\/td>\u57fa\u4e8e\u8bcd\u9891\u6784\u5efa\u7ef4\u5ea6\uff0c\u7528\u4e8e\u4f30\u8ba1\u6587\u672c\u5728\u67d0\u4e2a\u65b9\u5411\u4e0a\u7684\u201c\u4f4d\u7f6e\u201d<\/td><\/tr>
                  GPT \u76f4\u63a5\u505a pairwise\uff08\u65e0 prompt \u62c6\u89e3\uff09<\/td>GPT-3.5 \u76f4\u63a5\u6bd4\u8f83\u539f\u59cb tweet \u6587\u672c<\/td><\/tr>
                  CGCoT<\/td>\u7528 prompt \u5f15\u5bfc\u751f\u6210\u7ed3\u6784\u5316\u5206\u6790 + pairwise + \u5efa\u6a21<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
                  \"\"<\/figure>\n\n\n\n

                  Spearman\u2019s \u03c1 \u662f\u4e00\u79cd\u975e\u53c2\u6570\u76f8\u5173\u6027\u5ea6\u91cf\uff0c\u7528\u4e8e\u8861\u91cf\u4e24\u4e2a\u53d8\u91cf\u5728\u6392\u5e8f\u4e0a\u7684\u4e00\u81f4\u6027\u3002\u5e38\u7528\u4e8e\u5224\u65ad\u201c\u6a21\u578b\u8bc4\u5206\u6392\u5e8f\u662f\u5426\u4e0e\u4eba\u7c7b\u6807\u6ce8\u6392\u5e8f\u4e00\u81f4\u201d\u3002
                  \u7ed3\u8bba\uff1a\u7ed3\u6784\u5316 prompt + pairwise \u662f\u6027\u80fd\u63d0\u5347\u5173\u952e\u3002\u5355\u7eaf\u4f7f\u7528 GPT \u6216\u8bcd\u9891\u6a21\u578b\u6548\u679c\u6709\u9650\u3002<\/p>\n\n\n\n

                    \n
                  • \u4e0e\u76d1\u7763\u5b66\u4e60\u6a21\u578b\uff08RoBERTa-Large\uff09\u7684\u6bd4\u8f83<\/li>\n<\/ul>\n\n\n\n
                    \"\"<\/figure>\n\n\n\n
                      \n
                    • \u5bf9\u6bd4\u8bbe\u7f6e\uff1a<\/li>\n\n\n\n
                    • RoBERTa-Large\uff1a\u7528 3,000 \u6761\u4eba\u5de5\u6807\u6ce8 tweet \u7cbe\u8c03\u7684\u5206\u7c7b\u6a21\u578b<\/li>\n\n\n\n
                    • CGCoT\uff1a\u53ea\u7528 ~200 \u6761\u624b\u6807 tweet \u6765\u5f00\u53d1 prompt\uff0c\u8bc4\u5206\u5168\u81ea\u52a8\u5b8c\u6210<\/li>\n\n\n\n
                    • \u7528 CGCoT \u5206\u6570\u7684\u5e73\u5747\u503c\u505a\u9608\u503c\uff0c\u751f\u6210 0\/1 \u4e8c\u503c\u9884\u6d4b\u6807\u7b7e\uff08\u662f\u5426\u8868\u8fbe\u538c\u6076\uff09<\/li>\n<\/ul>\n\n\n\n

                      \u56db\u3001\u603b\u7ed3\u4e0e\u7efc\u5408\u5bf9\u5176\u601d\u8003<\/h2>\n\n\n\n

                      \uff08\u4e00\uff09\u8bba\u6587\u6838\u5fc3\u5185\u5bb9<\/p>\n\n\n\n

                        \n
                      • Prompt\u5f15\u5bfc\u7ed3\u6784\u5316\u5206\u89e3\uff1a\u7528\u4e13\u5bb6\u8bbe\u8ba1\u7684 prompt \u5c06\u6587\u672c\u5185\u5bb9\u6cbf\u62bd\u8c61\u6982\u5ff5\uff08\u5982\u201c\u538c\u6076\u201d\uff09\u5206\u89e3\u6210\u591a\u6b65\u7ed3\u6784\u5316\u63cf\u8ff0\uff1b<\/li>\n\n\n\n
                      • LLM\u6210\u5bf9\u6bd4\u8f83\uff08pairwise\uff09\uff1a\u6bd4\u8f83\u4efb\u610f\u4e24\u6761\u7ed3\u6784\u5316\u63cf\u8ff0\uff0c\u5224\u65ad\u54ea\u6761\u66f4\u5f3a\uff1b<\/li>\n\n\n\n
                      • Bradley-Terry\u5efa\u6a21\uff1a\u901a\u8fc7\u4e00\u7cfb\u5217\u80dc\u8d1f\u5173\u7cfb\uff0c\u4f30\u7b97\u6bcf\u6761\u6587\u672c\u7684\u201c\u6982\u5ff5\u5f3a\u5ea6\u5f97\u5206\u201d\uff1b<\/li>\n\n\n\n
                      • \u9ad8\u6548+\u5c11\u6837\u672c\uff1a\u53ea\u9700\u6781\u5c11\u4eba\u5de5\u6807\u6ce8\u5373\u53ef\u5b8c\u6210\u7cbe\u51c6\u8bc4\u5206\uff0c\u5e76\u53ef\u7528\u4e8e\u5206\u7c7b\u6216\u6392\u5e8f<\/li>\n<\/ul>\n\n\n\n

                        \uff08\u4e8c\uff09\u7efc\u5408\u5bf9\u5176\u601d\u8003<\/p>\n\n\n\n

                          \n
                        • \u8bba\u6587\u4e2d\u6709\u4f7f\u7528\u903b\u8f91\u5206\u6b65\u5f15\u5bfc\u7684Prompt\u94fe\u5f0f\u8bbe\u8ba1\uff0c\u8fc1\u79fb\u5230\u63a8\u8350\u4e2d\u53ef\u4ee5\u5c06\u4ea7\u54c1\u5bf9\u6bd4\u5206\u89e3\u4e3a\u591a\u8f6e Prompt \u903b\u8f91\u94fe\uff0c\u8ba9 LLM \u6a21\u62df\u201c\u7528\u6237\u51b3\u7b56\u8fc7\u7a0b\u201d\u3002<\/li>\n\n\n\n
                        • \u5c06\u7528\u6237\u76ee\u6807\u8f6c\u5316\u4e3a\u201c\u591a\u7ef4\u6982\u5ff5\u7ed3\u6784\u201d\uff0c\u4f9d\u6b21\u5224\u65ad\u6bcf\u4e2a\u7ef4\u5ea6\u7684\u6bd4\u8f83\u903b\u8f91\uff1a\u8425\u517b\u8bc4\u5206 \u2192 \u573a\u666f\u9002\u914d\u6027 \u2192 \u5b89\u5168\u6210\u5206 \u2192 \u6574\u4f53\u504f\u597d<\/li>\n\n\n\n
                        • \u6784\u5efa\u201c\u7ed3\u6784\u5316\u4ea7\u54c1\u4fe1\u606f\u89c6\u56fe\u201d\u7528\u4e8e\u6bd4\u8f83\uff0c\u5c06\u6bcf\u4e2a\u5019\u9009\u4ea7\u54c1\u7528 LLM \u751f\u6210\u7ed3\u6784\u5316\u63cf\u8ff0\uff0c\u7edf\u4e00\u6bd4\u8f83\u4fe1\u606f\u7684\u7ef4\u5ea6\u3002<\/li>\n<\/ul>\n\n\n\n

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

                          \u6765\u6e90\uff1aIEEE BigData\u4f5c\u8005\uff1aPatrick Y. Wu, Jonathan Nagler\u7b49\u5355\u4f4d\uff1aAme […]<\/p>\n","protected":false},"author":60,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3173","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3173","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\/60"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=3173"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3173\/revisions"}],"predecessor-version":[{"id":3182,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3173\/revisions\/3182"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=3173"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=3173"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=3173"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}