{"id":1620,"date":"2024-11-14T11:16:44","date_gmt":"2024-11-14T03:16:44","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=1620"},"modified":"2024-11-14T11:16:44","modified_gmt":"2024-11-14T03:16:44","slug":"202411%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-characterglm-customizing-chinese-conversational-ai-characters-with-large-language-models","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202411%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-characterglm-customizing-chinese-conversational-ai-characters-with-large-language-models\/","title":{"rendered":"202411\u8bba\u6587\u7814\u8bfb-CharacterGLM: Customizing Chinese Conversational AI Characters with Large Language Models"},"content":{"rendered":"\n

\u6765\u6e90\uff1aarXiv
\u4f5c\u8005\uff1aJinfeng Zhou\u3001 Zhuang Chen\u7b49
\u5355\u4f4d\uff1a\u6e05\u534e\u5927\u5b66\u3001\u8046\u5fc3\u667a\u80fd\u7b49
\u53d1\u8868\u65f6\u95f4\uff1a2023 \u5e74 11 \u6708<\/p>\n\n\n\n

\u4e00\u3001\u80cc\u666f<\/h2>\n\n\n\n
    \n
  1. \u5927\u8bed\u8a00\u6a21\u578b\uff1a\u5df2\u7ecf\u5728\u591a\u4e2a\u9886\u57df\u5c55\u73b0\u51fa\u5176\u5f3a\u5927\u7684\u80fd\u529b\uff0c\u7136\u800c\uff0c\u73b0\u6709\u7684LLMs\u5728\u5b9e\u73b0\u793e\u4ea4\u76ee\u6807\u65b9\u9762\u4ecd\u7136\u5b58\u5728\u4e0d\u8db3\uff0c\u4f8b\u5982\u4e0e\u4eba\u7c7b\u5efa\u7acb\u957f\u671f\u793e\u4ea4\u8054\u7cfb\u6216\u63d0\u4f9b\u6709\u6548\u7684\u60c5\u611f\u652f\u6301\u3002<\/li>\n\n\n\n
  2. \u63d0\u51faCharacterGLM\uff1a\u4e3a\u5bf9\u8bddAI\u7cfb\u7edf\u63d0\u4f9b\u89d2\u8272\u5b9a\u5236\u529f\u80fd\uff0c\u4ee5\u6ee1\u8db3\u4eba\u4eec\u5185\u5728\u7684\u793e\u4ea4\u6b32\u671b\u548c\u60c5\u611f\u9700\u6c42\u3002\u901a\u8fc7\u914d\u7f6eAI\u89d2\u8272\u7684\u5c5e\u6027\uff08\u8eab\u4efd\u3001\u5174\u8da3\u3001\u89c2\u70b9\u3001\u7ecf\u5386\u3001\u6210\u5c31\u3001\u793e\u4ea4\u5173\u7cfb\u7b49\uff09\u548c\u884c\u4e3a\uff08\u8bed\u8a00\u7279\u5f81\u3001\u60c5\u611f\u8868\u8fbe\u3001\u4e92\u52a8\u6a21\u5f0f\u7b49\uff09\uff0c\u53ef\u4ee5\u5b9a\u5236\u5404\u79cdAI\u89d2\u8272\u6216\u793e\u4ea4\u4ee3\u7406\u3002<\/li>\n<\/ol>\n\n\n\n

    \u4e8c\u3001CharacterGLM\u5b9e\u73b0<\/h2>\n\n\n\n
      \n
    1. \u6570\u636e\u6536\u96c6\u4e0e\u89d2\u8272\u5bf9\u8bdd<\/li>\n<\/ol>\n\n\n\n

      \u4eba\u7c7b\u89d2\u8272\u626e\u6f14\uff08Human Role-Playing\uff09
      \u63cf\u8ff0\uff1a\u62db\u52df\u4f17\u5305\u5de5\u4f5c\u8005\u8fdb\u884c\u89d2\u8272\u626e\u6f14\uff0c\u6a21\u62df\u771f\u5b9e\u5bf9\u8bdd\u3002
      \u8fc7\u7a0b\uff1a\u4e00\u4e2a\u5de5\u4f5c\u8005\u626e\u6f14\u7279\u5b9a\u89d2\u8272\uff0c\u53e6\u4e00\u4e2a\u5de5\u4f5c\u8005\u626e\u6f14\u7528\u6237\u6216\u5176\u4ed6\u89d2\u8272\u3002<\/p>\n\n\n\n

      \u5927\u8bed\u8a00\u6a21\u578b\u5408\u6210\uff08Synthesis via Large Language Models\uff09
      \u63cf\u8ff0\uff1a\u4f7f\u7528GPT-4\u7b49\u6a21\u578b\u751f\u6210\u5408\u6210\u6570\u636e\uff0c\u4ee5\u6269\u5c55\u6570\u636e\u89c4\u6a21\u548c\u591a\u6837\u6027\u3002
      \u8fc7\u7a0b\uff1a\u901a\u8fc7\u63d0\u793a\u751f\u6210\u89d2\u8272\u914d\u7f6e\u6587\u4ef6\u3001\u5bf9\u8bdd\u4e3b\u9898\u548c\u591a\u8f6e\u5bf9\u8bdd\u3002<\/p>\n\n\n\n

      \u6587\u5b66\u8d44\u6e90\u63d0\u53d6\uff08Extraction from Literary Resources\uff09
      \u63cf\u8ff0\uff1a\u4ece\u5267\u672c\u3001\u5c0f\u8bf4\u7b49\u6587\u5b66\u4f5c\u54c1\u4e2d\u63d0\u53d6\u89d2\u8272\u5bf9\u8bdd\u3002
      \u65b9\u6cd5\uff1a\u624b\u52a8\u63d0\u53d6\u53cc\u65b9\u9762\u5bf9\u9762\u7684\u5bf9\u8bdd\uff0c\u5e76\u603b\u7ed3\u89d2\u8272\u6863\u6848\u3002<\/p>\n\n\n\n

      2. \u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b<\/p>\n\n\n\n

      \u89d2\u8272\u63d0\u793a\u8bbe\u8ba1\uff08Character Prompt Design\uff09
      \u63cf\u8ff0\uff1a\u5c06\u89d2\u8272\u914d\u7f6e\u6587\u4ef6\u8f6c\u5316\u4e3a\u81ea\u7136\u8bed\u8a00\u63d0\u793a\uff0c\u4f5c\u4e3a\u6a21\u578b\u8bad\u7ec3\u7684\u57fa\u7840\u3002
      \u65b9\u6cd5\uff1a\u4f7f\u7528\u4f17\u5305\u5de5\u4f5c\u8005\u5c06\u89d2\u8272\u5c5e\u6027\u548c\u884c\u4e3a\u8f6c\u5316\u4e3a\u8be6\u7ec6\u7684\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u3002<\/p>\n\n\n\n

      \u76d1\u7763\u5f0f\u5fae\u8c03\uff08Supervised Fine-tuning\uff09
      \u63cf\u8ff0\uff1a\u4f7f\u7528\u4e0d\u540c\u89c4\u6a21\u7684ChatGLM\u6a21\u578b\uff086B\u523066B\u53c2\u6570\uff09\u4f5c\u4e3a\u57fa\u7840\u6a21\u578b\u3002
      \u8fc7\u7a0b\uff1a\u5c06\u89d2\u8272\u63d0\u793a\u4e0e\u5bf9\u8bdd\u6570\u636e\u7ed3\u5408\uff0c\u8fdb\u884c\u5fae\u8c03\u8bad\u7ec3\u3002<\/p>\n\n\n\n

      \u81ea\u6211\u5b8c\u5584\uff08Self-Refinement\uff09
      \u63cf\u8ff0\uff1a\u5728\u6a21\u578b\u90e8\u7f72\u540e\uff0c\u901a\u8fc7\u6536\u96c6\u4eba\u7c7b-\u539f\u578b\u4ea4\u4e92\u6570\u636e\u6765\u8fdb\u4e00\u6b65\u8bad\u7ec3\u6a21\u578b\u3002
      \u65b9\u6cd5\uff1a\u5c06\u4ea4\u4e92\u6570\u636e\u7eb3\u5165\u76d1\u7763\u5f0f\u5fae\u8c03\u8fc7\u7a0b\u4e2d\uff0c\u5b9e\u73b0\u6a21\u578b\u7684\u6301\u7eed\u81ea\u6211\u5b8c\u5584\u3002<\/p>\n\n\n\n

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

      \u4e09\u3001CharacterGLM\u5b9e\u9a8c<\/h2>\n\n\n\n
        \n
      1. \u70b9\u8bc4\u4f30<\/li>\n<\/ol>\n\n\n\n
        \"\"<\/figure>\n\n\n\n

        \u4e0a\u8ff0\u8bc4\u4f30\u4e8611\u4e2a\u7cbe\u901a\u4e2d\u6587\u4efb\u52a1\u4e3b\u6d41LLMs\uff0c\u4f7f\u7528\u4ece1\u52305\u7684\u8bc4\u5206\u8303\u56f4\uff0c\u7ed3\u679c\u5982\u8868\u6240\u793a\u3002<\/p>\n\n\n\n

        2. \u7ec6\u7c92\u5ea6\u8bef\u5dee\u5206\u6790<\/p>\n\n\n\n

        \u516d\u4e2a\u5173\u952e\u7ef4\u5ea6\uff1a
        Out-of-character (OOC)\uff1a\u4e0e\u89d2\u8272\u5c5e\u6027\u6216\u884c\u4e3a\u4e0d\u4e00\u81f4\u7684\u54cd\u5e94\u3002
        Contradiction\uff1a\u4e0e\u5bf9\u8bdd\u4e0a\u4e0b\u6587\u6216\u89d2\u8272\u6863\u6848\u76f8\u77db\u76fe\u7684\u54cd\u5e94\u3002
        Repetition\uff1a\u91cd\u590d\u5bf9\u8bdd\u4e0a\u4e0b\u6587\u6216\u89d2\u8272\u6863\u6848\u5185\u5bb9\u7684\u54cd\u5e94\u3002
        Less-quality\uff1a\u7f3a\u4e4f\u4e0e\u5bf9\u8bdd\u4e0a\u4e0b\u6587\u8fde\u8d2f\u6027\u6216\u8d28\u91cf\u8f83\u5dee\u7684\u54cd\u5e94\u3002
        Less-informativeness (Less-info.)\uff1a\u672a\u80fd\u63d0\u4f9b\u65b0\u6216\u6709\u4fe1\u606f\u91cf\u5185\u5bb9\u7684\u54cd\u5e94\u3002
        Proactivity\uff1a\u4e3b\u52a8\u5f15\u5bfc\u5bf9\u8bdd\u4e3b\u9898\u5e76\u63a8\u52a8\u5bf9\u8bdd\u7ee7\u7eed\u8fdb\u884c\u7684\u54cd\u5e94\u3002<\/p>\n\n\n\n

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

        \u4e0a\u8ff0\u56fe\u7247\u8868\u660eCharacterGLM-66B\u5728\u4f1a\u8bdd\u7ea7\u548c\u56de\u5408\u7ea7\u8bc4\u4f30\u4e2d\u4ea7\u751f\u7684\u54cd\u5e94\u7684\u603b\u4f53\u8d28\u91cf\u90fd\u66f4\u597d<\/p>\n\n\n\n

        3. \u6210\u5bf9\u8bc4\u4f30<\/p>\n\n\n\n

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

        \u6bd4\u8f83\u4e86CharacterGLM\u6a21\u578b\u4e0e\u5176\u4ed6\u6a21\u578b\uff08\u5982MiniMax\u3001GPT-3.5\u548cGPT-4\uff09\u5728\u7279\u5b9a\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002\u5177\u4f53\u6307\u89d2\u8272\u7c7b\u522b\u548c\u5bf9\u8bdd\u4e3b\u9898<\/p>\n\n\n\n

        \u56db\u3001\u603b\u7ed3<\/h2>\n\n\n\n
          \n
        • \u8bba\u6587\u6838\u5fc3\u5185\u5bb9<\/li>\n<\/ul>\n\n\n\n

          \u8fd9\u7bc7\u8bba\u6587\u4ecb\u7ecd\u4e86\u4e00\u4e2a\u540d\u4e3aCharacterGLM\u7684\u7cfb\u7edf\uff0c\u5b83\u662f\u4e00\u7cfb\u5217\u57fa\u4e8eLLM\u6784\u5efa\u7684\u6a21\u578b\uff0c\u65e8\u5728\u751f\u6210\u4ee5\u89d2\u8272\u4e3a\u57fa\u7840\u7684\u5bf9\u8bdd\uff08CharacterDial\uff09\u3002
          1. CharacterGLM\u6a21\u578b\u7684\u5927\u5c0f\u4ece6B\u523066B\u53c2\u6570\u4e0d\u7b49\uff0c\u76ee\u7684\u662f\u4e3a\u5bf9\u8bdd\u5f0fAI\u7cfb\u7edf\u63d0\u4f9b\u89d2\u8272\u5b9a\u5236\u529f\u80fd\uff0c\u4ee5\u6ee1\u8db3\u4eba\u4eec\u5728\u793e\u4ea4\u548c\u60c5\u611f\u9700\u6c42\u4e0a\u7684\u5185\u5728\u6b32\u671b\u3002
          2. CharacterGLM\u7684\u5b9e\u73b0\u5305\u62ec\u57fa\u4e8e\u89d2\u8272\u7684\u5bf9\u8bdd\u6536\u96c6\u548c\u8bad\u7ec3LLM\u8fdb\u884c\u57fa\u4e8e\u89d2\u8272\u7684\u5bf9\u8bdd\u751f\u6210\u3002\u4f5c\u8005\u91c7\u7528\u4e86\u591a\u79cd\u6570\u636e\u6536\u96c6\u65b9\u6cd5\uff0c\u5305\u62ec\u4eba\u5de5\u89d2\u8272\u626e\u6f14\u3001\u901a\u8fc7\u5927\u578b\u8bed\u8a00\u6a21\u578b\u5408\u6210\u6570\u636e\u548c\u4ece\u6587\u5b66\u8d44\u6e90\u4e2d\u63d0\u53d6\u5bf9\u8bdd\u3002<\/p>\n\n\n\n

            \n
          • \u7efc\u5408\u5bf9\u5176\u601d\u8003<\/li>\n<\/ul>\n\n\n\n
              \n
            1. \u53ef\u4ee5\u6536\u83b7\u7684\u5730\u65b9\uff1a<\/li>\n<\/ol>\n\n\n\n

              CharacterDial\u4efb\u52a1\uff1a\u9996\u6b21\u63d0\u51faCharacterDial\u4efb\u52a1\uff0c\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u5927\u89c4\u6a21\u7684\u4e2d\u6587\u5bf9\u8bdd\u6570\u636e\u96c6\uff0c\u5305\u542b\u771f\u5b9e\u548c\u865a\u62df\u89d2\u8272\u7684\u591a\u79cd\u573a\u666f\u548c\u8bdd\u9898\uff0c\u4e3a\u89d2\u8272\u4e2a\u6027\u5316\u5bf9\u8bdd\u63d0\u4f9b\u4e86\u4e30\u5bcc\u7684\u8d44\u6e90\u3002
              \u6df1\u5ea6\u89d2\u8272\u5b9a\u5236\uff1aCharacterGLM\u5b9e\u73b0\u4e86\u5168\u9762\u7684\u89d2\u8272\u4e2a\u6027\u5316\u8bbe\u5b9a\uff0c\u8986\u76d6\u4e86\u8eab\u4efd\u3001\u5174\u8da3\u3001\u89c2\u70b9\u3001\u7ecf\u5386\u3001\u793e\u4f1a\u5173\u7cfb\u7b49\u5c5e\u6027\uff0c\u5e76\u652f\u6301\u5b9a\u5236\u89d2\u8272\u7684\u884c\u4e3a\u7279\u5f81\uff0c\u4e0d\u540c\u4e8e\u4f20\u7edf\u7684\u201c\u4eba\u683c\u5316\u5bf9\u8bdd\u201d\u3002<\/p>\n\n\n\n

              2.\u53ef\u4ee5\u6539\u8fdb\u7684\u5730\u65b9\uff1a<\/p>\n\n\n\n

              \u591a\u6a21\u6001\u4ea4\u4e92\uff1a\u8be5\u6587CharacterGLM\u4e13\u6ce8\u4e8e\u6587\u672c\u5bf9\u8bdd\uff0c\u4f46\u73b0\u5b9e\u4e16\u754c\u7684\u793e\u4ea4\u4e92\u52a8\u5f80\u5f80\u662f\u591a\u6a21\u6001\u7684\uff0c\u5305\u62ec\u89c6\u89c9\u3001\u542c\u89c9\u548c\u89e6\u89c9\u7ebf\u7d22\uff0c\u4e3a\u589e\u5f3a\u7528\u6237\u4f53\u9a8c\u540e\u671f\u5e94\u8003\u8651\u6269\u5c55\u7cfb\u7edf\u3002
              \u793e\u4ea4\u4e92\u52a8\uff1a\u8be5\u7cfb\u7edf\u4e3b\u8981\u5173\u6ce8AI\u89d2\u8272\u4e0e\u7528\u6237\u4e4b\u95f4\u7684\u4e92\u52a8\uff0c\u540e\u671f\u53ef\u4ee5\u63a2\u7d22AI\u89d2\u8272\u4e4b\u95f4\u7684\u793e\u4ea4\u4e92\u52a8\uff0c\u5f62\u6210\u4e00\u4e2a\u201c\u89d2\u8272\u793e\u4f1a\u201d\u3002<\/p>\n\n\n\n

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

              \u6765\u6e90\uff1aarXiv\u4f5c\u8005\uff1aJinfeng Zhou\u3001 Zhuang Chen\u7b49\u5355\u4f4d\uff1a\u6e05\u534e\u5927\u5b66\u3001\u8046\u5fc3\u667a\u80fd\u7b49\u53d1\u8868\u65f6\u95f4\uff1a […]<\/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-1620","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/1620","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=1620"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/1620\/revisions"}],"predecessor-version":[{"id":1625,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/1620\/revisions\/1625"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=1620"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=1620"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=1620"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}