{"id":2604,"date":"2025-03-15T15:24:48","date_gmt":"2025-03-15T07:24:48","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=2604"},"modified":"2025-03-15T15:24:48","modified_gmt":"2025-03-15T07:24:48","slug":"202503-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-blockagents-llm-comparative-assessment-zero-shot-nlg-evaluation-through-pairwise-comparisons-using-large-language-models","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202503-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-blockagents-llm-comparative-assessment-zero-shot-nlg-evaluation-through-pairwise-comparisons-using-large-language-models\/","title":{"rendered":"202503 \u8bba\u6587\u7814\u8bfb-BlockAgents: LLM Comparative Assessment: Zero-shot NLG Evaluation through Pairwise Comparisons using Large Language Models"},"content":{"rendered":"\n

\u6765\u6e90\uff1aarXiv
\u4f5c\u8005\uff1aAdian Liusie, Potsawee Manakul, Mark J. F. Gales
\u5355\u4f4d\uff1a\u5251\u6865\u5927\u5b66
\u53d1\u8868\u65f6\u95f4\uff1a2023 \u5e74 7 \u6708<\/p>\n\n\n\n

\u4e00\u3001\u7814\u7a76\u80cc\u666f<\/h2>\n\n\n\n

NLG\u8bc4\u4f30\uff1a
\u4eba\u5de5\u8bc4\u4f30\u6210\u672c\u9ad8\u3001\u8017\u65f6\u957f\u3002
\u4f20\u7edf\u81ea\u52a8\u8bc4\u4f30\u65b9\u6cd5\uff08\u5982ROUGE\u3001BERTScore\uff09\u4f9d\u8d56\u53c2\u8003\u6587\u672c\uff0c\u4e14\u4efb\u52a1\u7279\u5b9a\u6027\u5f3a\uff0c\u6cdb\u5316\u6027\u5dee\u3002
\u73b0\u6709\u96f6\u6837\u672c\u65b9\u6cd5\uff08\u5982ChatGPT\u8bc4\u5206\uff09\u9700\u8981\u5927\u89c4\u6a21\u6a21\u578b\uff08175B+\u53c2\u6570\uff09\uff0c\u9002\u7528\u6027\u53d7\u9650\u3002
LLM\u6f5c\u529b\uff1a
\u96f6\u6837\u672c\u80fd\u529b\uff1a\u65e0\u9700\u8bad\u7ec3\u5373\u53ef\u5b8c\u6210\u4efb\u52a1\u3002
\u6838\u5fc3\u601d\u60f3\uff1a\u4eba\u7c7b\u66f4\u64c5\u957f\u6bd4\u8f83\u800c\u975e\u7edd\u5bf9\u8bc4\u5206 \u2192 \u5229\u7528LLM\u8fdb\u884c\u6210\u5bf9\u6bd4\u8f83\uff08Pairwise Comparison\uff09\u53ef\u80fd\u66f4\u9ad8\u6548\u3002
\u7814\u7a76\u52a8\u673a\uff1a
\u63a2\u7d22\u4e2d\u7b49\u89c4\u6a21LLM\uff08\u5982FlanT5\u3001Llama2\uff09\u5728NLG\u8bc4\u4f30\u4e2d\u7684\u8868\u73b0\u3002
\u89e3\u51b3\u4f4d\u7f6e\u504f\u5dee\uff08Positional Bias\uff09\u95ee\u9898\uff0c\u63d0\u5347\u8bc4\u4f30\u53ef\u9760\u6027\u3002<\/p>\n\n\n\n

\u8fd9\u7bc7\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u6210\u5bf9\u6bd4\u8f83\u7684LLM\u8bc4\u4f30\u6846\u67b6\uff0c\u7814\u7a76\u4e86LLM\u5728\u96f6\u6837\u672c\u8bc4\u4f30\u4e0b\u7684\u6210\u5bf9\u6bd4\u8f83\u80fd\u529b\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u65b9\u6cd5\u5b9e\u73b0<\/h2>\n\n\n\n
    \n
  1. \u6210\u5bf9\u6bd4\u8f83\u65b9\u6cd5\uff1a<\/li>\n<\/ol>\n\n\n\n

    \u7ed9\u5b9a\u4e00\u4e2a\u4e0a\u4e0b\u6587\uff08context\uff09\u548c\u591a\u4e2a\u5019\u9009\u6587\u672c\uff08\u6bd4\u5982\u591a\u4e2a\u6458\u8981\u6216\u5bf9\u8bdd\u56de\u590d\uff09\uff0c\u901a\u8fc7\u6210\u5bf9\u6bd4\u8f83\u7684\u65b9\u5f0f\u8bc4\u4f30\u5b83\u4eec\u7684\u8d28\u91cf\uff0c\u800c\u4e0d\u662f\u5bf9\u6bcf\u4e2a\u6587\u672c\u72ec\u7acb\u6253\u5206\u3002<\/p>\n\n\n\n

    \u5bf9\u4e8e\u4efb\u610f\u4e24\u4e2a\u5019\u9009\u6587\u672cx_i\u548cx_j\uff0cLLM \u88ab\u63d0\u793a\u751f\u6210\u4e00\u4e2a\u4e8c\u5206\u7c7b\u5224\u65ad\uff1a\u54ea\u4e2a\u6587\u672c\u66f4\u597d\u3002<\/p>\n\n\n\n

    2. \u6982\u7387\u8ba1\u7b97\uff1a
    Prompt-Based Classifier\uff08\u6982\u7387\u8f93\u51fa\uff09—- \u53ef\u4ee5\u8bbf\u95ee\u5230\u6a21\u578b\u7684\u8f93\u51fa\u6982\u7387
    Text Generation\uff08\u751f\u6210\u5f0f\u8f93\u51fa\uff09—- \u6a21\u578b\u53ea\u80fd\u751f\u6210\u6587\u672c<\/p>\n\n\n\n

      \n
    1. \u53bb\u504f\u7b56\u7565\uff1a<\/li>\n<\/ol>\n\n\n\n

      LLM\u5728\u6210\u5bf9\u6bd4\u8f83\u4e2d\u5b58\u5728\u4f4d\u7f6e\u504f\u7f6e\uff0c\u5373\u6a21\u578b\u66f4\u503e\u5411\u4e8e\u9009\u62e9\u7b2c\u4e00\u4e2a\u6216\u7b2c\u4e8c\u4e2a\u5019\u9009\u9879\u3002<\/p>\n\n\n\n

        \n
      1. \u5b50\u96c6\u6bd4\u8f83\u7b56\u7565\uff1a<\/li>\n<\/ol>\n\n\n\n

        \u7531\u4e8e\u6240\u6709\u6210\u5bf9\u6bd4\u8f83\u7684\u6570\u91cf\u4e3aR=N(N\u22121),\u5f53N\u8f83\u5927\u65f6\u8ba1\u7b97\u6210\u672c\u9ad8\u3002\u56e0\u6b64\u63d0\u51fa\u4e09\u79cd\u7b56\u7565\uff1a<\/p>\n\n\n\n

        Random\uff1a\u968f\u673a\u9009\u62e9\u90e8\u5206\u6bd4\u8f83\u8fdb\u884c\u8ba1\u7b97\u3002
        No-repeat\uff1a\u540c\u4e00\u5bf9\u6587\u672c\u4e0d\u91cd\u590d\u6bd4\u8f83\u3002
        Symmetric\uff1a\u6bcf\u6b21\u6bd4\u8f83\u90fd\u8fdb\u884c\u6b63\u53cd\u4e24\u79cd\u987a\u5e8f\u7684\u8ba1\u7b97\u3002<\/p>\n\n\n\n

        \u4e09\u3001\u5b9e\u9a8c<\/h2>\n\n\n\n
          \n
        1. \u5b9e\u9a8c\u8bbe\u7f6e&\u8bc4\u4f30\u6307\u6807<\/li>\n<\/ol>\n\n\n\n
          \"\"<\/figure>\n\n\n\n

          2. \u6bd4\u8f83\u8bc4\u4f30 vs \u7edd\u5bf9\u8bc4\u5206<\/p>\n\n\n\n

          \u5b9e\u9a8c\u8bc1\u660e\uff1a
          \u5bf9\u6bd4\u57fa\u7ebf\u65b9\u6cd5\uff1a\u4f8b\u5982 UniEval \u548c GPTScore\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u8868\u73b0\u66f4\u597d\uff0c\u4f46\u5b83\u4eec\u8981\u4e48\u662f\u6709\u76d1\u7763\u7684\u65b9\u6cd5\uff08\u5982 UniEval\uff09\uff0c\u8981\u4e48\u662f\u9700\u8981\u66f4\u5927\u89c4\u6a21\u7684 LLM\uff08\u5982 GPT3 \u548c ChatGPT\uff09\u3002
          \u6bd4\u8f83\u8bc4\u4f30\uff1a\u66f4\u7a33\u5b9a\u7684\u8868\u73b0\uff1a\u4e2d\u7b49\u89c4\u6a21\u7684\u6a21\u578b\uff08\u5982 FlanT5-3B\uff09\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\u8868\u73b0\u5f97\u66f4\u597d\uff0c\u5e76\u4e14\u4e0e\u57fa\u4e8e\u66f4\u5927\u6a21\u578b\u7684\u7cfb\u7edf\u76f8\u6bd4\uff0c\u6027\u80fd\u5dee\u8ddd\u8f83\u5c0f\u3002
          \u96f6\u6837\u672c\u5b66\u4e60\uff1a\u4e0d\u4f9d\u8d56\u4e8e\u6807\u6ce8\u6570\u636e\u8bad\u7ec3\uff0c\u800c\u662f\u901a\u8fc7\u5bf9\u6bd4\u5b9e\u73b0\u8bc4\u4f30\u3002<\/p>\n\n\n\n

          3. \u53bb\u504f\u6548\u679c\u9a8c\u8bc1<\/p>\n\n\n\n

          \u95ee\u9898\uff1aLLM\u5728\u6bd4\u8f83\u4e2d\u504f\u5411\u9996\u4e2a\u5019\u9009\uff08\u5982Llama2-7B\u4f7f\u7528Prompt 2\u65f6\uff0c\u5019\u9009A\u88ab\u9009\u6982\u7387\u8fbe68%\uff09\u3002
          \u89e3\u51b3\u65b9\u6848\uff1a
          \u91cd\u52a0\u6743\u6982\u7387\uff08\u516c\u5f0f9-10\uff09\uff1a\u901a\u8fc7\u8c03\u6574\u9608\u503c\u03c4\u5e73\u8861\u6b63\u53cd\u5411\u6bd4\u8f83\u3002
          \u5bf9\u79f0\u6bd4\u8f83\u7b56\u7565\uff1a\u5f3a\u5236\u6bd4\u8f83\u6bcf\u5bf9\u7684\u6b63\u53cd\u987a\u5e8f\u3002 <\/p>\n\n\n\n

          4.\u6bd4\u8f83\u5b50\u96c6<\/p>\n\n\n\n

          \u5b9e\u9a8c\u5185\u5bb9\uff1a
          \u6bd4\u8f83\u9009\u62e9\u7b56\u7565:
          Random\uff08\u968f\u673a\u9009\u62e9\uff09\uff1a \u4ece\u6240\u6709\u53ef\u80fd\u7684\u6bd4\u8f83\u5bf9\u4e2d\u968f\u673a\u9009\u62e9\u82e5\u5e72\u4e2a\u3002
          No-repeat\uff08\u4e0d\u91cd\u590d\u9009\u62e9\uff09\uff1a \u786e\u4fdd\u6bcf\u4e00\u5bf9(x_i, x_j)\u53ea\u6bd4\u8f83\u4e00\u6b21\uff0c\u800c\u4e0d\u4f1a\u540c\u65f6\u9009\u62e9(x_i, x_j)\u8fdb\u884c\u53cd\u5411\u6bd4\u8f83\u3002
          Symmetric\uff08\u5bf9\u79f0\u9009\u62e9\uff09\uff1a \u6bcf\u4e00\u5bf9(x_i, x_j)\u90fd\u4f1a\u8fdb\u884c\u4e24\u6b21\u6bd4\u8f83\uff0c\u610f\u56fe\u901a\u8fc7\u53cc\u5411\u5bf9\u6bd4\u6d88\u9664\u4f4d\u7f6e\u504f\u5dee\u3002
          \u4f7f\u7528\u4e86 SummEval \u6570\u636e\u96c6\u4e2d\u7684 COH\uff08Coherency\uff09\u6307\u6807\uff0cFlanT5-3B\u6a21\u578b\u8fdb\u884c\u5b9e\u9a8c\u3002
          \u5728\u6240\u6709\u7b56\u7565\u4e0b\u90fd\u540c\u65f6\u8fdb\u884c\u4e86 \u539f\u59cb\u8bc4\u4f30\uff08\u4e0d\u53bb\u504f\uff09 \u548c \u53bb\u504f\u8bc4\u4f30\uff08\u5e94\u7528\u53bb\u504f\u7b56\u7565\uff09 \u7684\u5bf9\u6bd4\u5b9e\u9a8c\u3002<\/p>\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
          \u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u8bc4\u4f30\u65b9\u6cd5\uff1a\u6210\u5bf9\u6bd4\u8f83\u8bc4\u4f30\uff0c\u5e76\u63d0\u51fa\u4e86\u4e00\u79cd\u53bb\u504f\u65b9\u6cd5\u6765\u51cf\u5c11\u4f4d\u7f6e\u504f\u5dee\u7684\u5f71\u54cd\u3002
          \u5176\u6838\u5fc3\u601d\u60f3\u662f\uff1a
          1.\u6210\u5bf9\u6bd4\u8f83\u4ee3\u66ff\u72ec\u7acb\u8bc4\u5206\uff1a
          \u4e0e\u5176\u5bf9\u6bcf\u4e2a\u6587\u672c\u72ec\u7acb\u6253\u5206\uff0c\u4e0d\u5982\u5bf9\u4e24\u4e2a\u6587\u672c\u8fdb\u884c\u5bf9\u6bd4\u5224\u65ad\uff0c\u9009\u62e9\u201c\u54ea\u4e2a\u66f4\u597d\u201d\u3002
          \u57fa\u4e8e\u4eba\u7684\u8bc4\u4f30\u4e60\u60ef\uff1a\u4eba\u7c7b\u66f4\u64c5\u957f\u76f8\u5bf9\u5224\u65ad\uff08\u6bd4\u8f83 A \u4e0e B\uff09\u800c\u4e0d\u662f\u7edd\u5bf9\u8bc4\u5206\u3002
          2.\u96f6\u6837\u672c\u5b66\u4e60\uff1a
          \u8fd9\u79cd\u65b9\u6cd5\u4e0d\u9700\u8981\u6a21\u578b\u8fdb\u884c\u4e13\u95e8\u7684\u8bad\u7ec3\u6216\u5fae\u8c03\u3002
          \u5229\u7528\u6a21\u578b\u5df2\u6709\u7684\u8bed\u8a00\u7406\u89e3\u80fd\u529b\u548c\u63a8\u7406\u80fd\u529b\uff0c\u901a\u8fc7\u5bf9\u6bd4\u76f4\u63a5\u8fdb\u884c\u8bc4\u4f30\u3002<\/p>\n\n\n\n

          \uff08\u4e8c\uff09\u7efc\u5408\u5bf9\u5176\u601d\u8003
          \u521b\u65b0\uff1a
          \u901a\u7528\u6027\u5f3a\uff1a\u53ef\u4ee5\u5e94\u7528\u4e8e\u591a\u4e2a\u8bc4\u4f30\u4efb\u52a1\uff0c\u800c\u4e0d\u4f9d\u8d56\u4e8e\u7279\u5b9a\u4efb\u52a1\u7684\u6807\u6ce8\u6570\u636e\u3002
          \u96f6\u6837\u672c\u5b66\u4e60\uff1a\u4e0d\u9700\u8981\u5fae\u8c03\u6216\u989d\u5916\u8bad\u7ec3\uff0c\u53ea\u9700\u8981\u5229\u7528\u5df2\u6709\u7684\u5927\u8bed\u8a00\u6a21\u578b\u3002
          \u504f\u5dee\u68c0\u6d4b\u4e0e\u6821\u6b63\uff1a\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5355\u4f46\u6709\u6548\u7684\u53bb\u504f\u65b9\u6cd5\uff0c\u63d0\u5347\u4e86\u6a21\u578b\u8bc4\u4f30\u7684\u51c6\u786e\u6027\u3002
          \u6539\u8fdb\uff1a
          \u9759\u6001\u9009\u62e9\u7b56\u7565\u7f3a\u9677\uff1a\u5f53\u524d\u7b56\u7565\u5e76\u6ca1\u6709\u5229\u7528\u5df2\u5b8c\u6210\u6bd4\u8f83\u7684\u4fe1\u606f\u6765\u6307\u5bfc\u540e\u7eed\u7684\u6bd4\u8f83\u9009\u62e9\uff0c\u53ef\u4ee5\u8003\u8651\u5f15\u5165\u52a8\u6001\u6bd4\u8f83\u5b50\u96c6\u9009\u62e9\u7b56\u7565\u3002\u901a\u8fc7\u5728\u6bcf\u6b21\u6bd4\u8f83\u540e\uff0c\u8ba1\u7b97\u6bcf\u4e2a\u6587\u672c\u7684\u201c\u7f6e\u4fe1\u5ea6\u201d\u6216\u201c\u4e0d\u786e\u5b9a\u6027\u201d\uff0c\u4f18\u5148\u9009\u62e9\u90a3\u4e9b\u5bf9\u5f53\u524d\u6392\u5e8f\u8d21\u732e\u6700\u5927\u7684\u6837\u672c\u5bf9\u8fdb\u884c\u6bd4\u8f83\u3002
          \u6536\u83b7\uff1a
          \u5f53\u8fdb\u884c\u4ea7\u54c1\u63a8\u8350\u65f6\uff0c\u4e0d\u540c\u4ea7\u54c1\u7684\u5448\u73b0\u987a\u5e8f\u53ef\u80fd\u5f71\u54cd\u6a21\u578b\u7684\u5224\u65ad\u3002\u53ef\u4ee5\u5b66\u4e60\u8bba\u6587\u4e2d\u7684\u53bb\u504f\u65b9\u6cd5\u6765\u63d0\u9ad8\u63a8\u8350\u7684\u516c\u5e73\u6027\u548c\u51c6\u786e\u6027\u3002
          \u6a21\u4eff\u8bba\u6587\u4e2d\u7684\u6210\u5bf9\u6bd4\u8f83\u65b9\u6cd5\uff0c\u53ef\u4ee5\u8ba9\u6a21\u578b\u5728\u591a\u4e2a\u4ea7\u54c1\u4e2d\u8fdb\u884c\u4e24\u4e24\u6bd4\u8f83\uff0c\u800c\u4e0d\u662f\u7ed9\u6bcf\u4e2a\u4ea7\u54c1\u6253\u5206\u3002<\/p>\n\n\n\n

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

          \u6765\u6e90\uff1aarXiv\u4f5c\u8005\uff1aAdian Liusie, Potsawee Manakul, Mark J. F. G […]<\/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-2604","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/2604","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=2604"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/2604\/revisions"}],"predecessor-version":[{"id":2623,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/2604\/revisions\/2623"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=2604"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=2604"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=2604"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}