{"id":9594,"date":"2022-10-30T09:07:00","date_gmt":"2022-10-30T01:07:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=9594"},"modified":"2022-10-24T17:27:22","modified_gmt":"2022-10-24T09:27:22","slug":"bert-mlm","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/10\/30\/bert-mlm\/","title":{"rendered":"\u8fd0\u7528 BERT \u7684 MLM \u6a21\u578b\u8fdb\u884c\u5c0f\u6837\u672c\u5b66\u4e60"},"content":{"rendered":"\n

\u8f6c\u8f7d\u81ea\u300a\u5fc5\u987b\u8981GPT3\u5417\uff1f\u4e0d\uff0cBERT\u7684MLM\u6a21\u578b\u4e5f\u80fd\u5c0f\u6837\u672c\u5b66\u4e60\u300b<\/a>\u548c\u300aP-tuning\uff1a\u81ea\u52a8\u6784\u5efa\u6a21\u7248\uff0c\u91ca\u653e\u8bed\u8a00\u6a21\u578b\u6f5c\u80fd\u300b<\/a>\uff0c\u4f5c\u8005\uff1a\u82cf\u5251\u6797\uff0c\u90e8\u5206\u5185\u5bb9\u6709\u4fee\u6539\u3002<\/strong><\/p>\n\n\n\n

\u5927\u5bb6\u90fd\u77e5\u9053\u73b0\u5728 GPT3 \u98ce\u5934\u6b63\u76db\uff0c\u7136\u800c\uff0c\u5230\u5904\u90fd\u662f GPT3\u3001GPT3 \u5730\u63a8\uff0c\u8bfb\u8005\u662f\u5426\u8bb0\u5f97 GPT3 \u8bba\u6587\u7684\u540d\u5b57\u5462\uff1f\u4e8b\u5b9e\u4e0a\uff0cGPT3 \u7684\u8bba\u6587\u53eb\u505a\u300aLanguage Models are Few-Shot Learners\u300b<\/a>\uff0c\u6807\u9898\u91cc\u8fb9\u5df2\u7ecf\u6ca1\u6709 G\u3001P\u3001T \u51e0\u4e2a\u5355\u8bcd\u4e86\uff0c\u53ea\u4e0d\u8fc7\u5b83\u8ddf\u5f00\u59cb\u7684 GPT \u662f\u4e00\u8109\u76f8\u627f\u7684\uff0c\u56e0\u6b64\u8fd8\u662f\u4ee5 GPT \u79f0\u547c\u5b83\u3002\u987e\u540d\u601d\u4e49\uff0cGPT3 \u4e3b\u6253\u7684\u662f Few-Shot Learning\uff0c\u4e5f\u5c31\u662f\u5c0f\u6837\u672c\u5b66\u4e60\u3002\u6b64\u5916\uff0cGPT3 \u7684\u53e6\u4e00\u4e2a\u7279\u70b9\u5c31\u662f\u5927\uff0c\u6700\u5927\u7684\u7248\u672c\u591a\u8fbe 1750 \u4ebf\u53c2\u6570\uff0c\u662f BERT Base\u7684\u4e00\u5343\u591a\u500d\u3002<\/p>\n\n\n\n

\u6b63\u56e0\u5982\u6b64\uff0c\u524d\u4e9b\u5929 Arxiv \u4e0a\u7684\u4e00\u7bc7\u8bba\u6587\u300aIt\u2019s Not Just Size That Matters: Small Language Models Are Also Few-Shot Learners\u300b<\/a>\u4fbf\u5f15\u8d77\u4e86\u7b14\u8005\u7684\u6ce8\u610f\uff0c\u610f\u8bd1\u8fc7\u6765\u5c31\u662f\u201c\u8c01\u8bf4\u4e00\u5b9a\u8981\u5927\u7684\uff1f\u5c0f\u6a21\u578b\u4e5f\u53ef\u4ee5\u505a\u5c0f\u6837\u672c\u5b66\u4e60\u201d<\/strong>\u3002\u663e\u7136\uff0c\u8fd9\u6807\u9898\u5bf9\u6807\u7684\u5c31\u662f GPT3\uff0c\u4e8e\u662f\u7b14\u8005\u9976\u6709\u5174\u8da3\u5730\u70b9\u8fdb\u53bb\u770b\u770b\u662f\u8c01\u8fd9\u4e48\u6709\u52c7\u6c14\u6311\u6218 GPT3\uff0c\u53c8\u662f\u600e\u6837\u7684\u5c0f\u6a21\u578b\u80fd\u6311\u6218 GPT3\uff1f\u7ecf\u8fc7\u9605\u8bfb\uff0c\u539f\u6765\u4f5c\u8005\u63d0\u51fa\u901a\u8fc7\u9002\u5f53\u7684\u6784\u9020\uff0c\u7528 BERT \u7684 MLM \u6a21\u578b<\/strong>\u4e5f\u53ef\u4ee5\u505a\u5c0f\u6837\u672c\u5b66\u4e60\uff0c\u770b\u5b8c\u4e4b\u540e\u9887\u6709\u4e00\u79cd\u201c\u539f\u6765\u8fd8\u53ef\u4ee5\u8fd9\u6837\u505a\u201d\u7684\u604d\u7136\u5927\u609f\u611f\uff5e\u5728\u6b64\u4e0e\u5927\u5bb6\u5206\u4eab\u4e00\u4e0b\u3002<\/p>\n\n\n\n

\u5189\u5189\u5347\u8d77\u7684 MLM<\/a><\/h2>\n\n\n\n

MLM\uff0c\u5168\u79f0\u201cMasked Language Model\u201d\uff0c\u53ef\u4ee5\u7ffb\u8bd1\u4e3a\u201c\u63a9\u7801\u8bed\u8a00\u6a21\u578b\u201d\uff0c\u5b9e\u9645\u4e0a\u5c31\u662f\u4e00\u4e2a\u5b8c\u5f62\u586b\u7a7a\u4efb\u52a1\uff0c\u968f\u673a Mask \u6389\u6587\u672c\u4e2d\u7684\u67d0\u4e9b\u5b57\u8bcd\uff0c\u7136\u540e\u8981\u6a21\u578b\u53bb\u9884\u6d4b\u88ab Mask \u7684\u5b57\u8bcd\uff0c\u793a\u610f\u56fe\u5982\u4e0b\uff1a<\/p>\n\n\n\n

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

BERT \u7684 MLM \u6a21\u578b\u7b80\u5355\u793a\u610f\u56fe<\/p>\n\n\n\n

\u5176\u4e2d\u88ab Mask \u6389\u7684\u90e8\u5206\uff0c\u53ef\u4ee5\u662f\u76f4\u63a5\u968f\u673a\u9009\u62e9\u7684 Token\uff0c\u4e5f\u53ef\u4ee5\u662f\u968f\u673a\u9009\u62e9\u8fde\u7eed\u7684\u80fd\u7ec4\u6210\u4e00\u6574\u4e2a\u8bcd\u7684 Token\uff0c\u540e\u8005\u79f0\u4e3a Whole Word Masking (WWM)\u3002<\/p>\n\n\n\n

\u5f00\u59cb\uff0cMLM \u4ec5\u88ab\u89c6\u4e3a BERT \u7684\u4e00\u4e2a\u9884\u8bad\u7ec3\u4efb\u52a1\uff0c\u8bad\u7ec3\u5b8c\u4e86\u5c31\u53ef\u4ee5\u6254\u6389\u7684\u90a3\u79cd\uff0c\u56e0\u6b64\u6709\u4e00\u4e9b\u5f00\u6e90\u7684\u6a21\u578b\u5e72\u8106\u6ca1\u4fdd\u7559 MLM \u90e8\u5206\u7684\u6743\u91cd\uff0c\u6bd4\u5982 brightmart\u7248<\/a> \u548c clue\u7248<\/a> \u7684 RoBERTa\uff0c\u800c\u54c8\u5de5\u5927\u5f00\u6e90\u7684 RoBERTa-wwm-ext-large<\/a> \u5219\u4e0d\u77e5\u9053\u51fa\u4e8e\u4ec0\u4e48\u539f\u56e0\u968f\u673a\u521d\u59cb\u5316\u4e86 MLM \u90e8\u5206\u7684\u6743\u91cd\uff0c\u56e0\u6b64\u5982\u679c\u8981\u590d\u73b0\u672c\u6587\u540e\u9762\u7684\u7ed3\u679c\uff0c\u8fd9\u4e9b\u7248\u672c\u662f\u4e0d\u53ef\u53d6\u7684\u3002<\/p>\n\n\n\n

\u7136\u800c\uff0c\u968f\u7740\u7814\u7a76\u7684\u6df1\u5165\uff0c\u7814\u7a76\u4eba\u5458\u53d1\u73b0\u4e0d\u6b62 BERT \u7684 Encoder \u5f88\u6709\u7528\uff0c\u9884\u8bad\u7ec3\u7528\u7684 MLM \u672c\u8eab\u4e5f\u5f88\u6709\u7528\u3002\u6bd4\u5982\u8bba\u6587\u300aBERT has a Mouth, and It Must Speak: BERT as a Markov Random Field Language Model\u300b<\/a>\u6307\u51fa MLM \u53ef\u4ee5\u4f5c\u4e3a\u4e00\u822c\u7684\u751f\u6210\u6a21\u578b\u7528\uff0c\u8bba\u6587\u300aSpelling Error Correction with Soft-Masked BERT\u300b<\/a>\u5219\u5c06 MLM \u7528\u4e8e\u6587\u672c\u7ea0\u9519\uff0c\u7b14\u8005\u4e4b\u524d\u5728\u300a\u4ece\u8bed\u8a00\u6a21\u578b\u5230Seq2Seq\uff1aTransformer\u5982\u620f\uff0c\u5168\u9760Mask\u300b<\/a>\u7684\u5b9e\u9a8c\u4e5f\u8868\u660e MLM \u7684\u9884\u8bad\u7ec3\u6743\u91cd\u4e5f\u53ef\u4ee5\u5f53\u4f5c UniLM \u6765\u7528\u505a Seq2Seq \u4efb\u52a1\uff0c\u8fd8\u6709\u300a\u65e0\u76d1\u7763\u5206\u8bcd\u548c\u53e5\u6cd5\u5206\u6790\uff01\u539f\u6765BERT\u8fd8\u53ef\u4ee5\u8fd9\u6837\u7528\u300b<\/a>\u4e00\u6587\u5c06 MLM \u7684\u601d\u60f3\u7528\u4e8e\u65e0\u76d1\u7763\u5206\u8bcd\u548c\u53e5\u6cd5\u5206\u6790\u4e86\u3002\u53ef\u4ee5\u8bf4 MLM \u5df2\u7ecf\u662f\u5927\u653e\u5f02\u5f69\u4e86\u3002<\/p>\n\n\n\n

\u5c06\u4efb\u52a1\u8f6c\u6210\u5b8c\u5f62\u586b\u7a7a<\/a><\/h2>\n\n\n\n

\u5728\u672c\u6587\u91cc\uff0c\u6211\u4eec\u518d\u5b66\u4e60 MLM \u7684\u4e00\u4e2a\u7cbe\u5f69\u5e94\u7528\uff1a\u7528\u4e8e\u5c0f\u6837\u672c\u5b66\u4e60\u6216\u534a\u76d1\u7763\u5b66\u4e60\uff0c\u67d0\u4e9b\u573a\u666f\u4e0b\u751a\u81f3\u80fd\u505a\u5230\u96f6\u6837\u672c\u5b66\u4e60\u3002<\/p>\n\n\n\n

\u600e\u4e48\u5c06\u6211\u4eec\u8981\u505a\u7684\u4efb\u52a1\u8ddf MLM \u7ed3\u5408\u8d77\u6765\u5462\uff1f\u5f88\u7b80\u5355\uff0c\u7ed9\u4efb\u52a1\u4e00\u4e2a\u6587\u672c\u63cf\u8ff0\uff0c\u7136\u540e\u8f6c\u6362\u4e3a\u5b8c\u5f62\u586b\u7a7a\u95ee\u9898<\/strong>\u5373\u53ef\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5047\u5982\u7ed9\u5b9a\u53e5\u5b50\u201c\u8fd9\u8d9f\u5317\u4eac\u4e4b\u65c5\u6211\u611f\u89c9\u5f88\u4e0d\u9519\u3002\u201d\uff0c\u90a3\u4e48\u6211\u4eec\u8865\u5145\u4e2a\u63cf\u8ff0\uff0c\u6784\u5efa\u5982\u4e0b\u7684\u5b8c\u5f62\u586b\u7a7a\uff1a______\u6ee1\u610f\u3002\u8fd9\u8d9f\u5317\u4eac\u4e4b\u65c5\u6211\u611f\u89c9\u5f88\u4e0d\u9519\u3002<\/p>\n\n\n\n

\u8fdb\u4e00\u6b65\u5730\uff0c\u6211\u4eec\u9650\u5236\u7a7a\u4f4d\u5904\u53ea\u80fd\u586b\u4e00\u4e2a\u201c\u5f88\u201d\u6216\u201c\u4e0d\u201d\uff0c\u95ee\u9898\u5c31\u5f88\u6e05\u6670\u4e86\uff0c\u5c31\u662f\u8981\u6211\u4eec\u6839\u636e\u4e0a\u4e0b\u6587\u4e00\u81f4\u6027\u5224\u65ad\u662f\u5426\u6ee1\u610f\uff0c\u5982\u679c\u201c\u5f88\u201d\u7684\u6982\u7387\u5927\u4e8e\u201c\u4e0d\u201d\u7684\u6982\u7387\uff0c\u8bf4\u660e\u662f\u6b63\u9762\u60c5\u611f\u503e\u5411\uff0c\u5426\u5219\u5c31\u662f\u8d1f\u9762\u7684\uff0c\u8fd9\u6837\u6211\u4eec\u5c31\u5c06\u60c5\u611f\u5206\u7c7b\u95ee\u9898\u8f6c\u6362\u4e3a\u4e00\u4e2a\u5b8c\u5f62\u586b\u7a7a\u95ee\u9898\u4e86\uff0c\u5b83\u53ef\u4ee5\u7528 MLM \u6a21\u578b\u7ed9\u51fa\u9884\u6d4b\u7ed3\u679c\uff0c\u800c MLM \u6a21\u578b\u7684\u8bad\u7ec3\u53ef\u4ee5\u4e0d\u9700\u8981\u76d1\u7763\u6570\u636e\uff0c\u56e0\u6b64\u7406\u8bba\u4e0a\u8fd9\u80fd\u591f\u5b9e\u73b0\u96f6\u6837\u672c\u5b66\u4e60\u4e86\u3002<\/p>\n\n\n\n

\u591a\u5206\u7c7b\u95ee\u9898\u4e5f\u53ef\u4ee5\u505a\u7c7b\u4f3c\u8f6c\u6362\uff0c\u6bd4\u5982\u65b0\u95fb\u4e3b\u9898\u5206\u7c7b\uff0c\u8f93\u5165\u53e5\u5b50\u4e3a\u201c\u516b\u4e2a\u6708\u4e86\uff0c\u7ec8\u4e8e\u53c8\u80fd\u5728\u8d5b\u573a\u4e0a\u770b\u5230\u5973\u6392\u59d1\u5a18\u4eec\u4e86\u3002\u201d\uff0c\u90a3\u4e48\u5c31\u53ef\u4ee5\u6784\u5efa\u4e0b\u9762\u62a5\u5bfc\u4e00\u5219______\u65b0\u95fb\u3002\u516b\u4e2a\u6708\u4e86\uff0c\u7ec8\u4e8e\u53c8\u80fd\u5728\u8d5b\u573a\u4e0a\u770b\u5230\u5973\u6392\u59d1\u5a18\u4eec\u4e86\u3002<\/p>\n\n\n\n

\u8fd9\u6837\u6211\u4eec\u5c31\u5c06\u65b0\u95fb\u4e3b\u9898\u5206\u7c7b\u4e5f\u8f6c\u6362\u4e3a\u5b8c\u5f62\u586b\u7a7a\u95ee\u9898\u4e86\uff0c\u4e00\u4e2a\u597d\u7684 MLM \u6a21\u578b\u5e94\u5f53\u80fd\u9884\u6d4b\u51fa\u201c\u4f53\u80b2\u201d\u4e8c\u5b57\u6765\u3002<\/p>\n\n\n\n

\u8fd8\u6709\u4e00\u4e9b\u7b80\u5355\u7684\u63a8\u7406\u4efb\u52a1\u4e5f\u53ef\u4ee5\u505a\u8fd9\u6837\u7684\u8f6c\u6362\uff0c\u5e38\u89c1\u7684\u662f\u7ed9\u5b9a\u4e24\u4e2a\u53e5\u5b50\uff0c\u5224\u65ad\u8fd9\u4e24\u4e2a\u53e5\u5b50\u662f\u5426\u76f8\u5bb9\uff0c\u6bd4\u5982\u201c\u6211\u53bb\u4e86\u5317\u4eac\u201d\u8ddf\u201c\u6211\u53bb\u4e86\u4e0a\u6d77\u201d\u5c31\u662f\u77db\u76fe\u7684\uff0c\u201c\u6211\u53bb\u4e86\u5317\u4eac\u201d\u8ddf\u201c\u6211\u5728\u5929\u5b89\u95e8\u5e7f\u573a\u201d\u662f\u76f8\u5bb9\u7684\uff0c\u5e38\u89c1\u7684\u505a\u6cd5\u5c31\u662f\u5c06\u4e24\u4e2a\u53e5\u5b50\u62fc\u63a5\u8d77\u6765\u8f93\u5165\u5230\u6a21\u578b\u505a\uff0c\u4f5c\u4e3a\u4e00\u4e2a\u4e8c\u5206\u7c7b\u4efb\u52a1\u3002\u5982\u679c\u8981\u8f6c\u6362\u4e3a\u5b8c\u5f62\u586b\u7a7a\uff0c\u90a3\u8be5\u600e\u4e48\u6784\u9020\u5462\uff1f\u4e00\u79cd\u6bd4\u8f83\u81ea\u7136\u7684\u6784\u5efa\u65b9\u5f0f\u662f\uff1a\u6211\u53bb\u4e86\u5317\u4eac\uff1f______\uff0c\u6211\u53bb\u4e86\u4e0a\u6d77\u3002
\u6211\u53bb\u4e86\u5317\u4eac\uff1f______\uff0c\u6211\u5728\u5929\u5b89\u95e8\u5e7f\u573a\u3002<\/p>\n\n\n\n

\u5176\u4e2d\u7a7a\u4f4d\u4e4b\u5904\u7684\u5019\u9009\u8bcd\u4e3a \u662f\u7684,\u4e0d\u662f\u662f\u7684,\u4e0d\u662f\u3002<\/p>\n\n\n\n

Pattern-Exploiting<\/a><\/h2>\n\n\n\n

\u8bfb\u5230\u8fd9\u91cc\uff0c\u8bfb\u8005\u5e94\u8be5\u4e0d\u96be\u53d1\u73b0\u5176\u4e2d\u7684\u89c4\u5f8b\u4e86\uff0c\u5c31\u662f\u7ed9\u8f93\u5165\u7684\u6587\u672c\u589e\u52a0\u4e00\u4e2a\u524d\u7f00\u6216\u8005\u540e\u7f00\u63cf\u8ff0\uff0c\u5e76\u4e14 Mask \u6389\u67d0\u4e9b Token\uff0c\u8f6c\u6362\u4e3a\u5b8c\u5f62\u586b\u7a7a\u95ee\u9898\uff0c\u8fd9\u6837\u7684\u8f6c\u6362\u5728\u539f\u8bba\u6587\u4e2d\u79f0\u4e3a Pattern\uff0c\u8fd9\u4e2a\u8f6c\u6362\u8981\u5c3d\u53ef\u80fd\u4e0e\u539f\u6765\u7684\u53e5\u5b50\u7ec4\u6210\u4e00\u53e5\u81ea\u7136\u7684\u8bdd\uff0c\u4e0d\u80fd\u8fc7\u4e8e\u751f\u786c\uff0c\u56e0\u4e3a\u9884\u8bad\u7ec3\u7684 MLM \u6a21\u578b\u5c31\u662f\u5728\u81ea\u7136\u8bed\u8a00\u4e0a\u8fdb\u884c\u7684\u3002\u663e\u7136\u540c\u4e00\u4e2a\u95ee\u9898\u53ef\u4ee5\u6709\u5f88\u591a\u4e0d\u540c\u7684 Pattern\uff0c\u6bd4\u5982\u60c5\u611f\u5206\u7c7b\u7684\u4f8b\u5b50\uff0c\u63cf\u8ff0\u53ef\u4ee5\u653e\u6700\u540e\uff0c\u53d8\u6210\u201c\u8fd9\u8d9f\u5317\u4eac\u4e4b\u65c5\u6211\u611f\u89c9\u5f88\u4e0d\u9519\u3002__\u6ee1\u610f\u3002\u201d\uff1b\u4e5f\u53ef\u4ee5\u591a\u52a0\u51e0\u4e2a\u5b57\uff0c\u6bd4\u5982\u201c\u89c9\u5f97\u5982\u4f55\uff1f<\/strong>__\u6ee1\u610f\u3002\u8fd9\u8d9f\u5317\u4eac\u4e4b\u65c5\u6211\u611f\u89c9\u5f88\u4e0d\u9519\u3002\u201d\u3002<\/p>\n\n\n\n

\u7136\u540e\uff0c\u6211\u4eec\u9700\u8981\u6784\u5efa\u9884\u6d4b Token \u7684\u5019\u9009\u7a7a\u95f4\uff0c\u5e76\u4e14\u5efa\u7acb Token \u5230\u5b9e\u9645\u7c7b\u522b\u7684\u6620\u5c04\uff0c\u8fd9\u5728\u539f\u8bba\u6587\u4e2d\u79f0\u4e3a Verbalizer\uff0c\u6bd4\u5982\u60c5\u611f\u5206\u7c7b\u7684\u4f8b\u5b50\uff0c\u6211\u4eec\u7684\u5019\u9009\u7a7a\u95f4\u662f \u5f88,\u4e0d\u5f88,\u4e0d\uff0c\u6620\u5c04\u5173\u7cfb\u662f \u5f88\u2192\u6b63\u9762,\u4e0d\u2192\u8d1f\u9762\u5f88\u2192\u6b63\u9762,\u4e0d\u2192\u8d1f\u9762\uff0c\u5019\u9009\u7a7a\u95f4\u4e0e\u5b9e\u9645\u7c7b\u522b\u4e4b\u95f4\u4e0d\u4e00\u5b9a\u662f\u4e00\u4e00\u6620\u5c04\uff0c\u6bd4\u5982\u6211\u4eec\u8fd8\u53ef\u4ee5\u52a0\u5165\u201c\u633a\u201d\u3001\u201c\u592a\u201d\u3001\u201c\u96be\u201d\u5b57\uff0c\u5e76\u4e14\u8ba4\u4e3a \u5f88,\u633a,\u592a\u2192\u6b63\u9762\u5f88,\u633a,\u592a\u2192\u6b63\u9762 \u4ee5\u53ca \u4e0d,\u96be\u2192\u8d1f\u9762\u4e0d,\u96be\u2192\u8d1f\u9762\uff0c\u7b49\u7b49\u3002\u4e0d\u96be\u7406\u89e3\uff0c\u4e0d\u5c11 NLP \u4efb\u52a1\u90fd\u6709\u53ef\u80fd\u8fdb\u884c\u8fd9\u79cd\u8f6c\u6362\uff0c\u4f46\u663e\u7136\u8fd9\u79cd\u8f6c\u6362\u4e00\u822c\u53ea\u9002\u7528\u4e8e\u5019\u9009\u7a7a\u95f4\u6709\u9650\u7684\u4efb\u52a1\uff0c\u8bf4\u767d\u4e86\u5c31\u662f\u53ea\u7528\u6765\u505a\u9009\u62e9\u9898\uff0c\u5e38\u89c1\u4efb\u52a1\u7684\u5c31\u662f\u6587\u672c\u5206\u7c7b\u3002<\/p>\n\n\n\n

\u521a\u624d\u8bf4\u4e86\uff0c\u540c\u4e00\u4e2a\u4efb\u52a1\u53ef\u4ee5\u6709\u591a\u79cd\u4e0d\u540c\u7684Pattern\uff0c\u539f\u8bba\u6587\u662f\u8fd9\u6837\u5904\u7406\u7684\uff1a<\/p>\n\n\n\n

1\u3001\u5bf9\u4e8e\u6bcf\u79cd Pattern\uff0c\u5355\u72ec\u7528\u8bad\u7ec3\u96c6 Finetune \u4e00\u4e2a MLM \u6a21\u578b\u51fa\u6765\uff1b\n2\u3001\u7136\u540e\u5c06\u4e0d\u540c Pattern \u5bf9\u5e94\u7684\u6a21\u578b\u8fdb\u884c\u96c6\u6210\uff0c\u5f97\u5230\u878d\u5408\u6a21\u578b\uff1b\n3\u3001\u7528\u878d\u5408\u6a21\u578b\u9884\u6d4b\u672a\u6807\u6ce8\u6570\u636e\u7684\u4f2a\u6807\u7b7e\uff1b\n4\u3001\u7528\u4f2a\u6807\u7b7e\u6570\u636e Finetune \u4e00\u4e2a\u5e38\u89c4\u7684\uff08\u975e MLM \u7684\uff09\u6a21\u578b\u3002\n<\/code><\/pre>\n\n\n\n
\u5177\u4f53\u7684\u96c6\u6210\u65b9\u5f0f\u5927\u5bb6\u81ea\u5df1\u770b\u8bba\u6587\u5c31\u884c\uff0c\u8fd9\u4e0d\u662f\u91cd\u70b9\u3002\u8fd9\u79cd\u8bad\u7ec3\u6a21\u5f0f\u88ab\u79f0\u4e3a Pattern-Exploiting Training (PET)\uff0c\u5b83\u9996\u5148\u51fa\u73b0\u5728\u8bba\u6587\u300aExploiting Cloze Questions for Few Shot Text Classification and Natural Language Inference\u300b<\/a>\uff0c\u672c\u6587\u8981\u4ecb\u7ecd\u7684\u8fd9\u7bc7\u8bba\u6587\u5219\u8fdb\u4e00\u6b65\u80af\u5b9a\u548c\u5b8c\u5584\u4e86 Pattern-Exploiting Training \u7684\u4ef7\u503c\u548c\u7ed3\u679c\uff0c\u5e76\u6574\u5408\u4e86\u591a\u4efb\u52a1\u5b66\u4e60\uff0c\u4f7f\u5f97\u5b83\u5728 SuperGLUE \u699c\u5355\u4e0a\u7684\u5c0f\u6837\u672c\u5b66\u4e60\u6548\u679c\u8d85\u8fc7\u4e86 GPT3\u3002\u4e24\u7bc7\u8bba\u6587\u7684\u4f5c\u8005\u662f\u76f8\u540c\u7684\uff0c\u662f\u4e00\u8109\u76f8\u627f\u7684\u4f5c\u54c1\u3002<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
PET \u5728 SuperGLUE \u4e0a\u7684\u5c0f\u6837\u672c\u5b66\u4e60\u7684\u7ed3\u679c<\/p>\n\n\n\n
\u4e0d\u8fc7\u8981\u5410\u69fd\u4e00\u4e2a\u70b9\u662f\uff0c\u4e0a\u56fe\u4e2d PET \u7684 223M \u53c2\u6570\uff0c\u6240\u7528\u7684\u6a21\u578b\u662f ALBERT-xxlarge-v2\uff0c\u4e8b\u5b9e\u4e0a\u79f0 ALBERT \u4e3a\u201c\u5c0f\u6a21\u578b\u201d\u662f\u4e00\u79cd\u5f88\u800d\u6d41\u6c13\u7684\u884c\u4e3a\uff0c\u56e0\u4e3a\u5b83\u524d\u5411\u8ba1\u7b97\u7684\u901f\u5ea6\u5e76\u6ca1\u6709\u5f97\u5230\u4efb\u4f55\u63d0\u5347\u3002ALBERT-xxlarge \u5171\u6709 12 \u5c42\uff0c\u5c42\u4e0e\u5c42\u4e4b\u95f4\u53c2\u6570\u662f\u5171\u4eab\u7684\uff0c\u5c31\u524d\u5411\u8ba1\u7b97\u800c\u8a00\uff0c\u5b83\u5e94\u8be5\u7b49\u4ef7\u4e8e\u7ea6 2700M\uff0812 \u500d\uff09\u53c2\u6570\u7684 GPT \u624d\u5bf9\u3002<\/p>\n\n\n\n
PET \u4e2d\u6587\u5b9e\u8df5\uff0c\u68c0\u9a8c\u6548\u679c<\/a><\/h2>\n\n\n\n
\u8981\u771f\u6b63\u786e\u8ba4\u4e00\u4e2a\u65b9\u6cd5\u6216\u6a21\u578b\u7684\u4ef7\u503c\uff0c\u770b\u8bba\u6587\u7684\u5b9e\u9a8c\u8868\u683c\u662f\u4e0d\u591f\u7684\uff0c\u8bba\u6587\u7ed9\u51fa\u7684\u5b9e\u9a8c\u7ed3\u679c\u8c01\u90fd\u4e0d\u597d\u8bf4\u80fd\u5426\u590d\u73b0\uff0c\u5176\u6b21\u5c31\u7b97\u82f1\u6587\u4e0a\u80fd\u590d\u73b0\u4e5f\u4e0d\u4ee3\u8868\u4e2d\u6587\u4e0a\u6709s\u4ef7\u503c\uff0c\u56e0\u6b64\u6700\u5b9e\u9645\u7684\u8fd8\u662f\u4eb2\u81ea\u52a8\u624b\u505a\u5b9e\u9a8c\u9a8c\u8bc1\u3002\u4e0b\u9762\u662f\u7b14\u8005\u7684\u5b9e\u9a8c\u4ee3\u7801\uff0c\u4f9b\u8bfb\u8005\u53c2\u8003\uff1aGithub\u5730\u5740\uff1a<\/strong>https:\/\/github.com\/bojone\/Pattern-Exploiting-Training<\/a><\/p>\n\n\n\n
\u6211\u4eec\u5c06\u4ece\u4ee5\u4e0b\u51e0\u4e2a\u89d2\u5ea6\u6765\u63a2\u8ba8 PET \u7684\u53ef\u884c\u6027\uff1a<\/p>\n\n\n\n
1\u3001\u76f4\u63a5\u5229\u7528\u73b0\u6210\u7684 MLM \u6a21\u578b\u6548\u679c\u5982\u4f55\uff1f\uff08\u96f6\u6837\u672c\u5b66\u4e601\uff09\n2\u3001\u7528\u201c\u5927\u91cf\u65e0\u6807\u7b7e\u6570\u636e\u201d\u5fae\u8c03\u73b0\u6210\u7684 MLM \u6a21\u578b\u6548\u679c\u5982\u4f55\uff1f\uff08\u96f6\u6837\u672c\u5b66\u4e602\uff09\n3\u3001\u7528\u201c\u5c0f\u91cf\u6807\u7b7e\u6570\u636e\u201d\u5fae\u8c03\u73b0\u6210\u7684 MLM \u6a21\u578b\u6548\u679c\u5982\u4f55\uff1f\uff08\u5c0f\u6837\u672c\u5b66\u4e60\uff09\n4\u3001\u7528\u201c\u5c0f\u91cf\u6807\u7b7e\u6570\u636e+\u5927\u91cf\u65e0\u6807\u7b7e\u6570\u636e\u201d\u5fae\u8c03\u73b0\u6210\u7684 MLM \u6a21\u578b\u6548\u679c\u5982\u4f55\uff1f\uff08\u534a\u76d1\u7763\u5b66\u4e60\uff09\n<\/code><\/pre>\n\n\n\n
\u4e0b\u9762\u4e3b\u8981\u7ed9\u51fa\u60c5\u611f\u4e8c\u5206\u7c7b\u7684\u5b9e\u9a8c\u7ed3\u679c\u3002\u53e6\u5916\u8fd8\u6709\u4e00\u4e2a\u65b0\u95fb\u4e3b\u9898\u7684\u591a\u5206\u7c7b\uff0c\u4ee3\u7801\u4e5f\u653e\u5230 Github \u4e86\uff0c\u5176\u7ed3\u679c\u662f\u7c7b\u4f3c\u7684\uff0c\u5c31\u4e0d\u91cd\u590d\u9648\u8ff0\u4e86\u3002<\/p>\n\n\n\n
\u96f6\u6837\u672c\u5b66\u4e601<\/a><\/h3>\n\n\n\n
\u8fd9\u91cc\u4e3b\u8981\u63a2\u7d22\u7684\u662f\u7ed9\u8f93\u5165\u6587\u672c\u8865\u4e0a\u5bf9\u5e94\u7684 Pattern \u540e\uff0c\u76f4\u63a5\u57fa\u4e8e\u73b0\u6210\u7684 MLM \u6a21\u578b\u8fdb\u884c\u9884\u6d4b\uff0c\u9884\u6d4b\u7684\u51c6\u786e\u7387\u3002\u7531\u4e8e\u6784\u5efa\u6a21\u578b\u7684\u6574\u4e2a\u8fc7\u7a0b\u90fd\u4e0d\u6d89\u53ca\u5230\u6807\u7b7e\u6570\u636e\u76d1\u7763\u8bad\u7ec3\uff0c\u56e0\u6b64\u8fd9\u7b97\u662f\u4e00\u79cd\u201c\u96f6\u6837\u672c\u5b66\u4e60\u201d\u3002\u6211\u4eec\u9700\u8981\u6bd4\u8f83\u7684\u662f\u4e0d\u540c Pattern\u3001\u4e0d\u540c MLM \u6a21\u578b\u4e0a\u7684\u6548\u679c\uff1a<\/p>\n\n\n\n
\u4e0b\u9762\u662f\u5b9e\u9a8c\u7684\u51e0\u4e2a Pattern\uff0c\u5176\u4e2d\u7a7a\u4f4d\u5904\u5019\u9009\u8bcd\u8bed\u90fd\u4e3a\u201c\u5f88\u201d\u548c\u201c\u4e0d\u201d\uff1a<\/p>\n\n\n\n
P1\uff1a____\u6ee1\u610f\u3002\u8fd9\u8d9f\u5317\u4eac\u4e4b\u65c5\u6211\u611f\u89c9\u5f88\u4e0d\u9519\u3002\nP2\uff1a\u8fd9\u8d9f\u5317\u4eac\u4e4b\u65c5\u6211\u611f\u89c9\u5f88\u4e0d\u9519\u3002____\u6ee1\u610f\u3002\nP3\uff1a____\u597d\u3002\u8fd9\u8d9f\u5317\u4eac\u4e4b\u65c5\u6211\u611f\u89c9\u5f88\u4e0d\u9519\u3002\nP4\uff1a____\u7406\u60f3\u3002\u8fd9\u8d9f\u5317\u4eac\u4e4b\u65c5\u6211\u611f\u89c9\u5f88\u4e0d\u9519\u3002\nP5\uff1a\u611f\u89c9\u5982\u4f55\uff1f____\u6ee1\u610f\u3002\u8fd9\u8d9f\u5317\u4eac\u4e4b\u65c5\u6211\u611f\u89c9\u5f88\u4e0d\u9519\u3002\n<\/code><\/pre>\n\n\n\n
\u81f3\u4e8e MLM \u6a21\u578b\uff0c\u5219\u662f\u4e0b\u9762\u51e0\u4e2a\uff1a<\/p>\n\n\n\n
M1\uff1aGoogle \u5f00\u6e90\u7684\u4e2d\u6587\u7248 BERT Base\uff08\u94fe\u63a5<\/a>\uff09\uff1b<\/p>
M2\uff1a\u54c8\u5de5\u5927\u5f00\u6e90\u7684 RoBERTa-wwm-ext Base\uff08\u94fe\u63a5<\/a>\uff09\uff1a<\/p>
M3\uff1a\u817e\u8baf UER \u5f00\u6e90\u7684 BERT Base\uff08\u94fe\u63a5<\/a>\uff09\uff1b<\/p>
M4\uff1a\u817e\u8baf UER \u5f00\u6e90\u7684BERT Large\uff08\u94fe\u63a5<\/a>\uff09\u3002<\/p><\/blockquote>\n\n\n\n
\u5b9e\u9a8c\u7ed3\u679c\u5982\u4e0b\u8868\uff08\u9a8c\u8bc1\u96c6\/\u6d4b\u8bd5\u96c6\uff09\uff1a\u4e0d\u540c\u6a21\u578b\u4e0d\u540cPattern\u7684\u96f6\u6837\u672c\u5b66\u4e60\u6548\u679c<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u6700\u597d\u7684\u6548\u679c\u5c45\u7136\u53ef\u4ee5\u8fbe\u5230 88%\uff01\u4e5f\u5c31\u662f\u8bf4\uff0c\u52a0\u8f7d\u73b0\u6210\u7684 MLM\uff0c\u914d\u5408\u9002\u5f53\u7684 Pattern\uff0c\u4e0d\u9700\u8981\u4efb\u4f55\u6807\u6ce8\u6570\u636e\uff0c\u5c31\u53ef\u4ee5\u6b63\u786e\u8bc6\u522b\u5927\u90e8\u5206\u6837\u672c\u60c5\u611f\u503e\u5411\u4e86\u3002\u8fd9\u4e0d\u5f97\u4e0d\u8ba9\u6211\u4eec\u5bf9 MLM \u6a21\u578b\u7684\u6f5c\u529b\u522e\u76ee\u76f8\u770b\u4e86\u3002<\/p>\n\n\n\n
\u53ef\u4ee5\u89c2\u5bdf\u5230\uff0c\u4e0d\u540c\u7684 Pattern\u3001\u4e0d\u540c\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u4e4b\u95f4\u8fd8\u662f\u6709\u4e00\u5b9a\u7684\u5dee\u5f02\u7684\uff0c\u6574\u4f53\u800c\u8a00 Large \u7248\u672c\u7684\u6548\u679c\u8981\u660e\u663e\u597d\u4e8e Base \u7248\u672c\u7684\u6a21\u578b\uff0c\u8bf4\u660e\u50cf GPT \u5230 GPT2 \u518d\u5230 GPT3 \u4e00\u6837\uff0c\u8fd8\u662f\u628a\u6a21\u578b\u505a\u5f97\u66f4\u5927\u4f1a\u66f4\u597d\u3002\u6b64\u5916\uff0c\u8fd9\u8fd8\u6709\u53ef\u80fd\u8bf4\u660e\u5b9e\u9645\u4e0a MLM \u8fd8\u6ca1\u6709\u88ab\u5145\u5206\u8bad\u7ec3\u597d\uff0c\u6216\u8bb8\u662f\u56e0\u4e3a BERT \u8fd9\u79cd Mask \u6389\u4e00\u90e8\u5206\u7684\u8bad\u7ec3\u65b9\u5f0f\u8fc7\u4e8e\u4f4e\u6548\u4e86\uff0c\u53ef\u80fd\u7528\u300a\u4fee\u6539Transformer\u7ed3\u6784\uff0c\u8bbe\u8ba1\u4e00\u4e2a\u66f4\u5feb\u66f4\u597d\u7684MLM\u6a21\u578b\u300b<\/a>\u4e00\u6587\u63d0\u5230\u7684\u6539\u8fdb\u7248 MLM \u4f1a\u66f4\u597d\u3002<\/p>\n\n\n\n
\u96f6\u6837\u672c\u5b66\u4e602<\/a><\/h3>\n\n\n\n
\u770b\u5b8c\u4e0a\u8ff0\u7ed3\u679c\uff0c\u8bfb\u8005\u53ef\u80fd\u4f1a\u60f3\u5230\uff1a\u5982\u679c\u6211\u7528\u9886\u57df\u5185\u7684\u6570\u636e\u7ee7\u7eed\u9884\u8bad\u7ec3 MLM \u6a21\u578b\uff0c\u90a3\u4e48\u80fd\u4e0d\u80fd\u63d0\u5347\u6548\u679c\u5462\uff1f\u7b54\u6848\u662f\uff1a\u80fd\uff01\u4e0b\u9762\u662f\u6211\u4eec\u7684\u5b9e\u9a8c\u7ed3\u679c\uff0c\u7b97\u529b\u6709\u9650\uff0c\u6211\u4eec\u53ea\u5728 RoBERTa-wwm-ext\uff08\u4e0a\u8ff0\u7684 M2\uff0c\u7ee7\u7eed\u9884\u8bad\u7ec3\u540e\u7684\u6a21\u578b\u6211\u4eec\u79f0\u4e3a M2+\u65e0\u76d1\u7763M2+\u65e0\u76d1\u7763\uff09\u7684\u57fa\u7840\u4e0a\u505a\u4e86\u6bd4\u8f83\uff1a<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u91cc\u6211\u4eec\u53ea\u662f\u7528\u9886\u57df\u5185\u7684\u6570\u636e\u7ee7\u7eed\u505a MLM \u8bad\u7ec3\uff0c\u8fd9\u4e2a\u8fc7\u7a0b\u662f\u65e0\u76d1\u7763\u7684\uff0c\u4e5f\u4e0d\u9700\u8981\u6807\u6ce8\u4fe1\u53f7\uff0c\u56e0\u6b64\u4e5f\u7b97\u662f\u201c\u96f6\u6837\u672c\u5b66\u4e60\u201d\u3002\u540c\u65f6\uff0c\u4ece\u5230\u76ee\u524d\u4e3a\u6b62\u7684\u7ed3\u679c\u6211\u4eec\u53ef\u4ee5\u770b\u51fa\uff0c\u7ed9\u8f93\u5165\u672c\u6587\u52a0\u5165\u201c\u524d\u7f00\u201d\u7684\u6548\u679c\u6bd4\u201c\u540e\u7f00\u201d\u66f4\u6709\u4f18\u52bf\u4e00\u4e9b\u3002<\/p>\n\n\n\n
\u5c0f\u6837\u672c\u5b66\u4e60<\/a><\/h3>\n\n\n\n
\u521a\u624d\u6211\u4eec\u8ba8\u8bba\u4e86\u65e0\u6807\u7b7e\u6570\u636e\u7ee7\u7eed\u9884\u8bad\u7ec3 MLM \u7684\u63d0\u5347\uff0c\u5982\u679c\u56de\u5230 PET \u7684\u76ee\u6807\u573a\u666f\uff0c\u76f4\u63a5\u7528\u5c0f\u91cf\u7684\u6807\u7b7e\u6570\u636e\u914d\u5408\u7279\u5b9a\u7684 Pattern \u8bad\u7ec3 MLM \u53c8\u5982\u4f55\u5462\uff1f\u8fd9\u4e5f\u5c31\u662f\u771f\u6b63\u7684\u201c\u5c0f\u6837\u672c\u5b66\u4e60\u201d\u8bad\u7ec3\u4e86\uff0c\u8fd9\u91cc\u6211\u4eec\u4fdd\u7559\u7ea6 200 \u4e2a\u6807\u6ce8\u6837\u672c\uff0c\u6784\u9020\u6837\u672c\u7684\u65f6\u5019\uff0c\u6211\u4eec\u5148\u7ed9\u6bcf\u4e2a\u53e5\u5b50\u8865\u4e0a Pattern\uff0c\u9664\u4e86 Pattern \u81ea\u5e26\u7684 Mask \u4f4d\u7f6e\u4e4b\u5916\uff0c\u6211\u4eec\u8fd8\u968f\u673a Mask \u5176\u4ed6\u4e00\u90e8\u5206\uff0c\u4ee5\u589e\u5f3a\u5bf9\u6a21\u578b\u7684\u6b63\u5219\u3002\u6700\u7ec8\u5b9e\u9a8c\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u7ed3\u8bba\u5c31\u662f\u9664\u4e86\u201c\u540e\u7f00\u5f0f\u201d\u7684 P2 \u4e4b\u5916\uff0c\u5176\u5b83\u7ed3\u679c\u90fd\u5dee\u4e0d\u591a\uff0c\u8fd9\u8fdb\u4e00\u6b65\u8bf4\u660e\u4e86\u201c\u524d\u7f00\u5f0f\u201d\u7684 Pattern \u4f1a\u6bd4\u201c\u540e\u7f00\u5f0f\u201d\u66f4\u6709\u7ade\u4e89\u529b\u4e00\u4e9b\u3002\u5728\u6548\u679c\u4e0a\uff0c\u76f4\u63a5\u7528\u540c\u6837\u7684\u6570\u636e\u7528\u5e38\u89c4\u7684\u65b9\u6cd5\u53bb\u5fae\u8c03\u4e00\u4e2a BERT \u6a21\u578b\uff0c\u5927\u6982\u7684\u7ed3\u679c\u662f 88.93 \u5de6\u53f3\uff0c\u6240\u4ee5\u57fa\u4e8e\u201cMLM+Pattern\u201d\u7684\u5c0f\u6837\u672c\u5b66\u4e60\u65b9\u6cd5\u53ef\u80fd\u5e26\u6765\u8f7b\u5fae\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n\n\n\n
\u534a\u76d1\u7763\u5b66\u4e60<\/a><\/h3>\n\n\n\n
\u65e0\u76d1\u7763\u7684\u96f6\u6837\u672c\u5b66\u4e60\u548c\u6709\u76d1\u7763\u7684\u5c0f\u6837\u672c\u5b66\u4e60\u90fd\u8bf4\u5b8c\u4e86\uff0c\u81ea\u7136\u5c31\u8f6e\u5230\u628a\u6807\u6ce8\u6570\u636e\u548c\u975e\u6807\u6ce8\u6570\u636e\u90fd\u7ed3\u5408\u8d77\u6765\u7684\u201c\u534a\u76d1\u7763\u5b66\u4e60\u201d\u4e86\u3002\u8fd8\u662f\u540c\u6837\u7684\u4efb\u52a1\uff0c\u6807\u6ce8\u6570\u636e\u548c\u975e\u6807\u6ce8\u6570\u636e\u7684\u6bd4\u4f8b\u5927\u7ea6\u662f 1:99\uff0c\u6807\u6ce8\u6570\u636e\u5e26 Pattern\uff0c\u975e\u6807\u6ce8\u6570\u636e\u4e0d\u5e26 Pattern\uff0c\u5927\u5bb6\u90fd Mask \u6389\u4e00\u90e8\u5206 Token \u8fdb\u884c MLM \u9884\u8bad\u7ec3\uff0c\u6700\u7ec8\u6d4b\u51fa\u6765\u7684\u6548\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u8fd8\u662f\u540c\u6837\u7684\uff0c\u201c\u540e\u7f00\u201d\u660e\u663e\u6bd4\u201c\u524d\u7f00\u201d\u5dee\uff0c\u201c\u524d\u7f00\u201d\u7684\u6548\u679c\u5dee\u4e0d\u591a\u3002\u5177\u4f53\u6548\u679c\u4e0a\uff0c\u5219\u662f\u80af\u5b9a\u4e86\u989d\u5916\u7684\u65e0\u6807\u6ce8\u6570\u636e\u4e5f\u662f\u6709\u4f5c\u7528\u7684\u3002\u76f4\u89c9\u4e0a\u6765\u770b\uff0c\u201c\u524d\u7f00\u201d\u6bd4\u201c\u540e\u7f00\u201d\u8981\u597d\uff0c\u5927\u4f53\u4e0a\u662f\u56e0\u4e3a\u201c\u524d\u7f00\u201d\u7684 Mask \u4f4d\u7f6e\u6bd4\u8f83\u56fa\u5b9a\uff0c\u5fae\u5f31\u7684\u76d1\u7763\u4fe1\u53f7\u5f97\u4ee5\u53e0\u52a0\u589e\u5f3a\uff1f\u4f46\u8fd9\u4e5f\u4e0d\u80fd\u89e3\u91ca\u4e3a\u4ec0\u4e48\u96f6\u6837\u672c\u5b66\u4e60\u7684\u60c5\u51b5\u4e0b\u4e5f\u662f\u201c\u524d\u7f00\u201d\u66f4\u597d\uff0c\u4f30\u8ba1\u8fd8\u8ddf\u6a21\u578b\u7684\u5b66\u4e60\u96be\u5ea6\u6709\u5173\u7cfb\uff0c\u53ef\u80fd\u53e5\u5b50\u524d\u9762\u90e8\u5206\u7684\u89c4\u5f8b\u66f4\u52a0\u660e\u663e\uff0c\u76f8\u5bf9\u6765\u8bf4\u66f4\u52a0\u5bb9\u6613\u5b66\u4e00\u4e9b\uff0c\u6240\u4ee5\u524d\u9762\u90e8\u5206\u5c31\u5b66\u4e60\u5f97\u66f4\u52a0\u5145\u5206\uff1f\u8fd9\u4e00\u5207\u90fd\u8fd8\u53ea\u662f\u731c\u6d4b\u3002<\/p>\n\n\n\n
\u6c47\u603b\u4e0e\u7ed3\u8bba<\/a><\/h3>\n\n\n\n
\u5c06\u4e0a\u8ff0\u7ed3\u679c\u6c47\u603b\u5982\u4e0b\uff1a\u7ed3\u679c\u6c47\u603b\u6bd4\u8f83<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u8bfb\u8005\u8fd8\u53ef\u4ee5\u5bf9\u6bd4\u6211\u4eec\u4e4b\u524d\u5728\u6587\u7ae0\u300a\u6cdb\u5316\u6027\u4e71\u5f39\uff1a\u4ece\u968f\u673a\u566a\u58f0\u3001\u68af\u5ea6\u60e9\u7f5a\u5230\u865a\u62df\u5bf9\u6297\u8bad\u7ec3\u300b<\/a>\u4e2d\u7528\u865a\u62df\u5bf9\u6297\u8bad\u7ec3 (VAT) \u505a\u534a\u76d1\u7763\u5b66\u4e60\u7684\u7ed3\u679c\uff0c\u53ef\u4ee5\u770b\u5230\u4e0d\u7ba1\u662f\u96f6\u6837\u672c\u5b66\u4e60\u3001\u5c0f\u6837\u672c\u5b66\u4e60\u8fd8\u662f\u534a\u76d1\u7763\u5b66\u4e60\uff0c\u57fa\u4e8e MLM \u6a21\u578b\u7684\u65b9\u5f0f\u90fd\u80fd\u5ab2\u7f8e\u57fa\u4e8e VAT \u7684\u534a\u76d1\u7763\u5b66\u4e60\u7684\u7ed3\u679c\u3002\u6211\u4eec\u5728\u505a\u77ed\u65b0\u95fb\u591a\u5206\u7c7b\u5b9e\u9a8c\u65f6\u7684\u7ed3\u679c\u4e5f\u662f\u76f8\u4f3c\u7684\u3002\u56e0\u6b64\uff0c\u8fd9\u8bf4\u660e\u4e86 MLM \u6a21\u578b\u786e\u5b9e\u4e5f\u53ef\u4ee5\u4f5c\u4e3a\u4e00\u4e2a\u4f18\u79c0\u7684\u96f6\u6837\u672c\/\u5c0f\u6837\u672c\/\u534a\u76d1\u7763\u5b66\u4e60\u5668\u6765\u4f7f\u7528\u3002<\/p>\n\n\n\n
\u5f53\u7136\uff0c\u57fa\u4e8e MLM \u6a21\u578b\u7684\u7f3a\u70b9\u8fd8\u662f\u6709\u7684\uff0c\u6bd4\u5982 MLM \u6240\u4f7f\u7528\u7684\u72ec\u7acb\u5047\u8bbe\u9650\u5236\u4e86\u5b83\u5bf9\u66f4\u957f\u6587\u672c\u7684\u9884\u6d4b\u80fd\u529b\uff08\u8bf4\u767d\u4e86\u7a7a\u4f4d\u5904\u7684\u6587\u5b57\u4e0d\u80fd\u592a\u957f\uff09\uff0c\u4ee5\u53ca\u65e0\u6cd5\u9884\u6d4b\u4e0d\u5b9a\u957f\u7684\u7b54\u6848\u4e5f\u7ea6\u675f\u4e86\u5b83\u7684\u573a\u666f\uff08\u6240\u4ee5\u5f53\u524d\u53ea\u80fd\u7528\u4e8e\u505a\u9009\u62e9\u9898\uff0c\u4e0d\u80fd\u505a\u751f\u6210\uff09\u3002\u6211\u4eec\u671f\u5f85\u6709\u66f4\u5f3a\u7684 MLM \u6a21\u578b\u51fa\u73b0\uff0c\u90a3\u65f6\u5019\u5c31\u6709\u53ef\u80fd\u5728\u6240\u6709\u4efb\u52a1\u4e0a\u90fd\u80fd\u4e0e GPT3 \u4e00\u8f83\u9ad8\u4e0b\u4e86\u3002<\/p>\n\n\n\n
\u4ec0\u4e48\u662f\u6a21\u7248<\/a><\/h2>\n\n\n\n
\u524d\u9762\u4ecb\u7ecd\u7684 Pattern-Exploiting Training (PET) \u65b9\u6cd5\uff0c\u5176\u4e3b\u8981\u7684\u601d\u60f3\u662f\u501f\u52a9\u7531\u81ea\u7136\u8bed\u8a00\u6784\u6210\u7684\u6a21\u7248\uff08\u82f1\u6587\u5e38\u79f0 Pattern \u6216 Prompt\uff09\uff0c\u5c06\u4e0b\u6e38\u4efb\u52a1\u4e5f\u8f6c\u5316\u4e3a\u4e00\u4e2a\u5b8c\u5f62\u586b\u7a7a\u4efb\u52a1\uff0c\u8fd9\u6837\u5c31\u53ef\u4ee5\u7528 BERT \u7684 MLM \u6a21\u578b\u6765\u8fdb\u884c\u9884\u6d4b\u4e86\u3002\u6bd4\u5982\u4e0b\u56fe\u4e2d\u901a\u8fc7\u6761\u4ef6\u524d\u7f00\u6765\u5b9e\u73b0\u60c5\u611f\u5206\u7c7b\u548c\u4e3b\u9898\u5206\u7c7b\u7684\u4f8b\u5b50\uff1a<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
\u901a\u8fc7\u7279\u5b9a\u6a21\u7248\u5c06\u60c5\u611f\u5206\u7c7b\u8f6c\u6362\u4e3a MLM \u4efb\u52a1<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
\u901a\u8fc7\u7279\u5b9a\u6a21\u7248\u5c06\u65b0\u95fb\u5206\u7c7b\u8f6c\u6362\u4e3a MLM \u4efb\u52a1<\/p>\n\n\n\n
\u5f53\u7136\uff0c\u8fd9\u79cd\u65b9\u6848\u4e5f\u4e0d\u662f\u53ea\u6709 MLM \u6a21\u578b\u53ef\u884c\uff0c\u7528 GPT \u8fd9\u6837\u7684\u5355\u5411\u8bed\u8a00\u6a21\u578b\uff08LM\uff09\u5176\u5b9e\u4e5f\u5f88\u7b80\u5355\uff1a<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
\u901a\u8fc7\u7279\u5b9a\u6a21\u7248\u5c06\u60c5\u611f\u5206\u7c7b\u8f6c\u6362\u4e3a LM \u4efb\u52a1<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
\u901a\u8fc7\u7279\u5b9a\u6a21\u7248\u5c06\u65b0\u95fb\u5206\u7c7b\u8f6c\u6362\u4e3a LM \u4efb\u52a1<\/p>\n\n\n\n
\u4e0d\u8fc7\u7531\u4e8e\u8bed\u8a00\u6a21\u578b\u662f\u4ece\u5de6\u5f80\u53f3\u89e3\u7801\u7684\uff0c\u56e0\u6b64\u9884\u6d4b\u90e8\u5206\u53ea\u80fd\u653e\u5728\u53e5\u672b\u4e86\uff08\u4f46\u8fd8\u53ef\u4ee5\u5f80\u8865\u5145\u524d\u7f00\u8bf4\u660e\uff0c\u53ea\u4e0d\u8fc7\u9884\u6d4b\u90e8\u5206\u653e\u5728\u6700\u540e\uff09\u3002<\/p>\n\n\n\n
\u67d0\u79cd\u610f\u4e49\u4e0a\u6765\u8bf4\uff0c\u8fd9\u4e9b\u6a21\u7248\u5c5e\u4e8e\u8bed\u8a00\u6a21\u578b\u7684\u201c\u63a2\u9488\u201d\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u6a21\u7248\u6765\u62bd\u53d6\u8bed\u8a00\u6a21\u578b\u7684\u7279\u5b9a\u77e5\u8bc6\uff0c\u4ece\u800c\u505a\u5230\u4e0d\u9519\u7684\u96f6\u6837\u672c\u6548\u679c\uff0c\u800c\u914d\u5408\u5c11\u91cf\u6807\u6ce8\u6837\u672c\uff0c\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u5347\u6548\u679c\u3002<\/p>\n\n\n\n
\u7136\u800c\uff0c\u5bf9\u4e8e\u67d0\u4e9b\u4efb\u52a1\u800c\u8a00\uff0c\u4eba\u5de5\u6784\u5efa\u6a21\u7248\u5e76\u4e0d\u662f\u90a3\u4e48\u5bb9\u6613\u7684\u4e8b\u60c5\uff0c\u6a21\u578b\u7684\u4f18\u52a3\u6211\u4eec\u4e5f\u4e0d\u597d\u628a\u63e1\uff0c\u800c\u4e0d\u540c\u6a21\u578b\u4e4b\u95f4\u7684\u6548\u679c\u5dee\u522b\u53ef\u80fd\u5f88\u5927\uff0c\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u4eba\u5de5\u6807\u6ce8\u4e00\u4e9b\u6837\u672c\u53ef\u80fd\u6bd4\u6784\u5efa\u6a21\u7248\u8fd8\u8981\u8f7b\u677e\u5f97\u591a\u3002\u6240\u4ee5\uff0c\u5982\u4f55\u6839\u636e\u5df2\u6709\u7684\u6807\u6ce8\u6837\u672c\u6765\u81ea\u52a8\u6784\u5efa\u6a21\u7248\uff0c\u4fbf\u6210\u4e86\u4e00\u4e2a\u503c\u5f97\u7814\u7a76\u7684\u95ee\u9898\u4e86\u3002<\/p>\n\n\n\n
P-tuning<\/a><\/h2>\n\n\n\n
\u6700\u8fd1 Arxiv \u4e0a\u7684\u8bba\u6587\u300aGPT Understands, Too\u300b<\/a>\u63d0\u51fa\u4e86\u540d\u4e3a P-tuning \u7684\u65b9\u6cd5\uff0c\u6210\u529f\u5730\u5b9e\u73b0\u4e86\u6a21\u7248\u7684\u81ea\u52a8\u6784\u5efa\u3002\u4e0d\u4ec5\u5982\u6b64\uff0c\u501f\u52a9 P-tuning\uff0cGPT \u5728 SuperGLUE \u4e0a\u7684\u6210\u7ee9\u9996\u6b21\u8d85\u8fc7\u4e86\u540c\u7b49\u7ea7\u522b\u7684 BERT \u6a21\u578b\uff0c\u8fd9\u98a0\u8986\u4e86\u4e00\u76f4\u4ee5\u6765\u201cGPT \u4e0d\u64c5\u957f NLU\u201d\u7684\u7ed3\u8bba\uff0c\u4e5f\u662f\u8be5\u8bba\u6587\u547d\u540d\u7684\u7f18\u7531\u3002<\/p>\n\n\n\n
P-tuning \u91cd\u65b0\u5ba1\u89c6\u4e86\u5173\u4e8e\u6a21\u7248\u7684\u5b9a\u4e49\uff0c\u653e\u5f03\u4e86\u201c\u6a21\u7248\u7531\u81ea\u7136\u8bed\u8a00\u6784\u6210\u201d\u8fd9\u4e00\u5e38\u89c4\u8981\u6c42\uff0c\u4ece\u800c\u5c06\u6a21\u7248\u7684\u6784\u5efa\u8f6c\u5316\u4e3a\u8fde\u7eed\u53c2\u6570\u4f18\u5316\u95ee\u9898\uff0c\u867d\u7136\u7b80\u5355\uff0c\u4f46\u5374\u6709\u6548\u3002<\/p>\n\n\n\n
\u6a21\u7248\u7684\u53cd\u601d<\/a><\/h3>\n\n\n\n
\u9996\u5148\uff0c\u6211\u4eec\u6765\u60f3\u4e00\u4e0b\u201c\u4ec0\u4e48\u662f\u6a21\u7248\u201d\u3002\u76f4\u89c2\u6765\u770b\uff0c\u6a21\u7248\u5c31\u662f\u7531\u81ea\u7136\u8bed\u8a00\u6784\u6210\u7684\u524d\u7f00\/\u540e\u7f00\uff0c\u901a\u8fc7\u8fd9\u4e9b\u6a21\u7248\u6211\u4eec\u4f7f\u5f97\u4e0b\u6e38\u4efb\u52a1\u8ddf\u9884\u8bad\u7ec3\u4efb\u52a1\u4e00\u81f4\uff0c\u8fd9\u6837\u624d\u80fd\u66f4\u52a0\u5145\u5206\u5730\u5229\u7528\u539f\u59cb\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u8d77\u5230\u66f4\u597d\u7684\u96f6\u6837\u672c\u3001\u5c0f\u6837\u672c\u5b66\u4e60\u6548\u679c\u3002<\/p>\n\n\n\n
\u7b49\u7b49\uff0c\u6211\u4eec\u771f\u7684\u5728\u4e4e\u6a21\u7248\u662f\u4e0d\u662f\u201c\u81ea\u7136\u8bed\u8a00\u201d\u6784\u6210\u7684\u5417\uff1f<\/strong><\/p>\n\n\n\n
\u5e76\u4e0d\u662f\u3002\u672c\u8d28\u4e0a\u6765\u8bf4\uff0c\u6211\u4eec\u5e76\u4e0d\u5173\u5fc3\u6a21\u7248\u957f\u4ec0\u4e48\u6837\uff0c\u6211\u4eec\u53ea\u9700\u8981\u77e5\u9053\u6a21\u7248\u7531\u54ea\u4e9b token \u7ec4\u6210\uff0c\u8be5\u63d2\u5165\u5230\u54ea\u91cc\uff0c\u63d2\u5165\u540e\u80fd\u4e0d\u80fd\u5b8c\u6210\u6211\u4eec\u7684\u4e0b\u6e38\u4efb\u52a1\uff0c\u8f93\u51fa\u7684\u5019\u9009\u7a7a\u95f4\u662f\u4ec0\u4e48\u3002<\/strong>\u6a21\u7248\u662f\u4e0d\u662f\u81ea\u7136\u8bed\u8a00\u7ec4\u6210\u7684\uff0c\u5bf9\u6211\u4eec\u6839\u672c\u6ca1\u5f71\u54cd\uff0c\u201c\u81ea\u7136\u8bed\u8a00\u201d\u7684\u8981\u6c42\uff0c\u53ea\u662f\u4e3a\u4e86\u66f4\u597d\u5730\u5b9e\u73b0\u201c\u4e00\u81f4\u6027\u201d\uff0c\u4f46\u4e0d\u662f\u5fc5\u987b\u7684\u3002\u4e8e\u662f\uff0cP-tuning \u8003\u8651\u4e86\u5982\u4e0b\u5f62\u5f0f\u7684\u6a21\u7248\uff1a<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
P-tuning \u76f4\u63a5\u4f7f\u7528 [unused*] \u7684 token \u6765\u6784\u5efa\u6a21\u7248\uff0c\u4e0d\u5173\u5fc3\u6a21\u7248\u7684\u81ea\u7136\u8bed\u8a00\u6027<\/p>\n\n\n\n
\u8fd9\u91cc\u7684 [u1]\uff5e[u6]\uff0c\u4ee3\u8868 BERT \u8bcd\u8868\u91cc\u8fb9\u7684 [unused1]\uff5e[unused6]\uff0c\u4e5f\u5c31\u662f\u7528\u51e0\u4e2a\u4ece\u672a\u89c1\u8fc7\u7684 token \u6765\u6784\u6210\u6a21\u677f\uff0c\u8fd9\u91cc\u7684 token \u6570\u76ee\u662f\u4e00\u4e2a\u8d85\u53c2\u6570\uff0c\u653e\u5728\u524d\u9762\u8fd8\u662f\u540e\u9762\u4e5f\u53ef\u4ee5\u8c03\u6574\u3002\u63a5\u7740\uff0c\u4e3a\u4e86\u8ba9\u201c\u6a21\u7248\u201d\u53d1\u6325\u4f5c\u7528\uff0c\u6211\u4eec\u7528\u6807\u6ce8\u6570\u636e\u6765\u6c42\u51fa\u8fd9\u4e2a\u6a21\u677f\u3002<\/p>\n\n\n\n
\u5982\u4f55\u53bb\u4f18\u5316<\/a><\/h3>\n\n\n\n
\u8fd9\u65f6\u5019\uff0c\u6839\u636e\u6807\u6ce8\u6570\u636e\u91cf\u7684\u591a\u5c11\uff0c\u6211\u4eec\u53c8\u5206\u4e24\u79cd\u60c5\u51b5\u8ba8\u8bba\u3002<\/p>\n\n\n\n
\u7b2c\u4e00\u79cd\uff0c\u6807\u6ce8\u6570\u636e\u6bd4\u8f83\u5c11\u3002<\/strong>\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u56fa\u5b9a\u6574\u4e2a\u6a21\u578b\u7684\u6743\u91cd\uff0c\u53ea\u4f18\u5316 [unused1]\uff5e[unused6] \u8fd9\u51e0\u4e2a token \u7684 Embedding\uff0c\u6362\u53e5\u8bdd\u8bf4\uff0c\u5176\u5b9e\u6211\u4eec\u5c31\u662f\u8981\u5b66 6 \u4e2a\u65b0\u7684 Embedding\uff0c\u4f7f\u5f97\u5b83\u8d77\u5230\u4e86\u6a21\u7248\u7684\u4f5c\u7528\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u56e0\u4e3a\u6a21\u578b\u6743\u91cd\u51e0\u4e4e\u90fd\u88ab\u56fa\u5b9a\u4f4f\u4e86\uff0c\u8bad\u7ec3\u8d77\u6765\u5f88\u5feb\uff0c\u800c\u4e14\u56e0\u4e3a\u8981\u5b66\u4e60\u7684\u53c2\u6570\u5f88\u5c11\uff0c\u56e0\u6b64\u54ea\u6015\u6807\u6ce8\u6837\u672c\u5f88\u5c11\uff0c\u4e5f\u80fd\u628a\u6a21\u7248\u5b66\u51fa\u6765\uff0c\u4e0d\u5bb9\u6613\u8fc7\u62df\u5408\u3002<\/p>\n\n\n\n
\u7b2c\u4e8c\u79cd\uff0c\u6807\u6ce8\u6570\u636e\u5f88\u5145\u8db3\u3002<\/strong>\u8fd9\u65f6\u5019\u5982\u679c\u8fd8\u6309\u7167\u7b2c\u4e00\u79cd\u7684\u65b9\u6848\u6765\uff0c\u5c31\u4f1a\u51fa\u73b0\u6b20\u62df\u5408\u7684\u60c5\u51b5\uff0c\u56e0\u4e3a\u53ea\u6709 6 \u4e2a token \u7684\u53ef\u4f18\u5316\u53c2\u6570\u5b9e\u5728\u662f\u592a\u5c11\u4e86\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u53ef\u4ee5\u653e\u5f00\u6240\u6709\u6743\u91cd\u5fae\u8c03\uff0c\u539f\u8bba\u6587\u5728 SuperGLUE \u4e0a\u7684\u5b9e\u9a8c\u5c31\u662f\u8fd9\u6837\u505a\u7684\u3002\u8bfb\u8005\u53ef\u80fd\u4f1a\u60f3\uff1a\u8fd9\u6837\u8ddf\u76f4\u63a5\u52a0\u4e2a\u5168\u8fde\u63a5\u5fae\u8c03\u6709\u4ec0\u4e48\u533a\u522b\uff1f\u539f\u8bba\u6587\u7684\u7ed3\u679c\u662f\u8fd9\u6837\u505a\u6548\u679c\u66f4\u597d\uff0c\u53ef\u80fd\u8fd8\u662f\u56e0\u4e3a\u8ddf\u9884\u8bad\u7ec3\u4efb\u52a1\u66f4\u4e00\u81f4\u4e86\u5427\u3002<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
P-tuning \u5728 SuperGLUE \u4e0a\u7684\u8868\u73b0<\/p>\n\n\n\n
\u6b64\u5916\uff0c\u5728\u4e0a\u9762\u7684\u4f8b\u5b50\u4e2d\uff0c\u76ee\u6807 token \u5982\u201c\u5f88\u201d\u3001\u201c\u4f53\u80b2\u201d\u662f\u8ba4\u4e3a\u9009\u5b9a\u7684\uff0c\u90a3\u4e48\u5b83\u4eec\u53ef\u4e0d\u53ef\u4ee5\u4e5f\u7528 [unused*] \u7684 token \u4ee3\u66ff\u5462\uff1f\u7b54\u6848\u662f\u53ef\u4ee5\uff0c\u4f46\u4e5f\u5206\u4e24\u79cd\u60c5\u51b5\u8003\u8651\uff1a1\u3001\u5728\u6807\u6ce8\u6570\u636e\u6bd4\u8f83\u5c11\u7684\u65f6\u5019\uff0c\u4eba\u5de5\u6765\u9009\u5b9a\u9002\u5f53\u7684\u76ee\u6807 token \u6548\u679c\u5f80\u5f80\u66f4\u597d\u4e9b\uff1b2\u3001\u5728\u6807\u6ce8\u6570\u636e\u5f88\u5145\u8db3\u7684\u60c5\u51b5\u4e0b\uff0c\u76ee\u6807 token \u7528 [unused*] \u6548\u679c\u66f4\u597d\u4e9b\uff0c\u56e0\u4e3a\u8fd9\u65f6\u5019\u6a21\u578b\u7684\u4f18\u5316\u7a7a\u95f4\u66f4\u5927\u4e00\u4e9b\u3002<\/p>\n\n\n\n
\u589e\u5f3a\u76f8\u5173\u6027<\/a><\/h3>\n\n\n\n
\u5728\u539f\u8bba\u6587\u4e2d\uff0cP-tuning \u5e76\u4e0d\u662f\u968f\u673a\u521d\u59cb\u5316\u51e0\u4e2a\u65b0 token \u7136\u540e\u76f4\u63a5\u8bad\u7ec3\u7684\uff0c\u800c\u662f\u901a\u8fc7\u4e00\u4e2a\u5c0f\u578b\u7684 LSTM \u6a21\u578b\u628a\u8fd9\u51e0\u4e2a Embedding \u7b97\u51fa\u6765\uff0c\u5e76\u4e14\u5c06\u8fd9\u4e2a LSTM \u6a21\u578b\u8bbe\u4e3a\u53ef\u5b66\u4e60\u7684\u3002\u8fd9\u6837\u591a\u7ed5\u4e86\u4e00\u6b65\u6709\u4ec0\u4e48\u597d\u5904\u5462\uff1f\u539f\u8bba\u6587\u5927\u6982\u7684\u610f\u601d\u662f\uff1aLSTM \u51fa\u73b0\u7684 token \u8868\u793a\u76f8\u5173\u6027\u66f4\u5f3a\uff0c\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u6765\u8bf4\u66f4\u50cf\u201c\u81ea\u7136\u8bed\u8a00\u201d\uff08\u56e0\u4e3a\u81ea\u7136\u8bed\u8a00\u7684 token \u4e4b\u95f4\u4e0d\u662f\u72ec\u7acb\u7684\uff09\uff0c\u6b64\u5916\u8fd8\u80fd\u9632\u6b62\u5c40\u90e8\u6700\u4f18\u3002\u6211\u5728 Github \u4e0a\u8fdb\u4e00\u6b65\u5411\u4f5c\u8005\u786e\u8ba4\u4e86\u4e00\u4e0b\uff08\u53c2\u8003\u8fd9\u91cc<\/a>\uff09\uff0c\u6548\u679c\u4e0a\u7684\u5dee\u522b\u662f\u901a\u8fc7 LSTM \u591a\u7ed5\u4e00\u6b65\u7684\u65b9\u6cd5\u53ef\u4ee5\u4f7f\u5f97\u6a21\u578b\u6536\u655b\u66f4\u5feb\u3001\u6548\u679c\u66f4\u4f18\u3002<\/p>\n\n\n\n
\u7136\u800c\uff0c\u8fd9\u6837\u591a\u4e86\u4e00\u4e2aLSTM\uff0c\u603b\u611f\u89c9\u6709\u4e9b\u522b\u626d\uff0c\u800c\u4e14\u5b9e\u73b0\u4e0a\u4e5f\u7565\u5fae\u6709\u70b9\u9ebb\u70e6\u3002\u6309\u7167\u4f5c\u8005\u7684\u610f\u601d\uff0cLSTM \u662f\u4e3a\u4e86\u5e2e\u52a9\u6a21\u7248\u7684\u51e0\u4e2a token\uff08\u67d0\u79cd\u7a0b\u5ea6\u4e0a\uff09\u66f4\u8d34\u8fd1\u81ea\u7136\u8bed\u8a00\uff0c\u4f46\u8fd9\u5e76\u4e0d\u4e00\u5b9a\u8981\u7528 LSTM \u751f\u6210\uff0c\u800c\u4e14\u5c31\u7b97\u7528 LSTM \u751f\u6210\u4e5f\u4e0d\u4e00\u5b9a\u8fbe\u5230\u8fd9\u4e00\u70b9\u3002\u7b14\u8005\u8ba4\u4e3a\uff0c\u66f4\u81ea\u7136\u7684\u65b9\u6cd5\u662f\u5728\u8bad\u7ec3\u4e0b\u6e38\u4efb\u52a1\u7684\u65f6\u5019\uff0c\u4e0d\u4ec5\u4ec5\u9884\u6d4b\u4e0b\u6e38\u4efb\u52a1\u7684\u76ee\u6807 token\uff08\u524d\u9762\u4f8b\u5b50\u4e2d\u7684\u201c\u5f88\u201d\u3001\u201c\u65b0\u95fb\u201d\uff09\uff0c\u8fd8\u5e94\u8be5\u540c\u65f6\u505a\u5176\u4ed6 token \u7684\u9884\u6d4b<\/strong>\u3002<\/p>\n\n\n\n
\u6bd4\u5982\uff0c\u5982\u679c\u662f MLM \u6a21\u578b\uff0c\u90a3\u4e48\u4e5f\u968f\u673a mask \u6389\u5176\u4ed6\u7684\u4e00\u4e9b token \u6765\u9884\u6d4b\uff1b\u5982\u679c\u662f LM \u6a21\u578b\uff0c\u5219\u9884\u6d4b\u5b8c\u6574\u7684\u5e8f\u5217\uff0c\u800c\u4e0d\u5355\u5355\u662f\u76ee\u6807\u8bcd\u3002\u8fd9\u6837\u505a\u7684\u7406\u7531\u662f\uff1a\u56e0\u4e3a\u6211\u4eec\u7684 MLM\/LM \u90fd\u662f\u7ecf\u8fc7\u81ea\u7136\u8bed\u8a00\u9884\u8bad\u7ec3\u7684\uff0c\u6240\u4ee5\u6211\u4eec\uff08\u8ff7\u4e4b\u81ea\u4fe1\u5730\uff09\u8ba4\u4e3a\u80fd\u591f\u5f88\u597d\u5b8c\u6210\u91cd\u6784\u7684\u5e8f\u5217\u5fc5\u7136\u4e5f\u662f\u63a5\u8fd1\u4e8e\u81ea\u7136\u8bed\u8a00\u7684\uff0c\u56e0\u6b64\u8fd9\u6837\u589e\u52a0\u8bad\u7ec3\u76ee\u6807\uff0c\u4e5f\u80fd\u8d77\u5230\u8ba9\u6a21\u578b\u66f4\u8d34\u8fd1\u81ea\u7136\u8bed\u8a00\u7684\u6548\u679c\u3002\u7ecf\u8fc7\u7b14\u8005\u7684\u6d4b\u8bd5\uff0c\u52a0\u4e0a\u8fd9\u6837\u8f85\u52a9\u76ee\u6807\uff0c\u76f8\u6bd4\u5355\u7eaf\u4f18\u5316\u4e0b\u6e38\u4efb\u52a1\u7684\u76ee\u6807\uff0c\u786e\u5b9e\u63d0\u5347\u4e86\u6548\u679c\u3002<\/p>\n\n\n\n
P-tuning \u5b9e\u9a8c\u4e0e\u6548\u679c<\/a><\/h2>\n\n\n\n
\u6240\u8c13\u201ctalk is cheap, show me the code\u201d\uff0c\u53c8\u5230\u4e86\u559c\u95fb\u4e50\u89c1\u7684\u5b9e\u9a8c\u65f6\u95f4\u4e86\u3002\u8fd9\u91cc\u5206\u4eab\u4e00\u4e0b P-tuning \u7684\u5b9e\u9a8c\u7ed3\u679c\uff0c\u5176\u4e2d\u8fd8\u5305\u62ec\u7b14\u8005\u5bf9 P-tuning \u7684\u5b9e\u73b0\u601d\u8def\uff0c\u4ee5\u53ca\u7b14\u8005\u5728\u4e2d\u6587\u4efb\u52a1\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c\u3002<\/p>\n\n\n\n
\u505c\u6b62\u7684\u68af\u5ea6<\/a><\/h3>\n\n\n\n
\u600e\u4e48\u5b9e\u73b0\u4e0a\u8ff0\u7684 P-tuning \u7b97\u6cd5\u6bd4\u8f83\u597d\u5462\uff1f\u5982\u679c\u662f\u653e\u5f00\u6240\u6709\u6743\u91cd\u8bad\u7ec3\uff0c\u90a3\u81ea\u7136\u662f\u7b80\u5355\u7684\uff0c\u8ddf\u666e\u901a\u7684 BERT \u5fae\u8c03\u6ca1\u6709\u4ec0\u4e48\u533a\u522b\u3002\u5173\u952e\u662f\u5728\u5c0f\u6837\u672c\u573a\u666f\u4e0b\uff0c\u5982\u4f55\u5b9e\u73b0\u201c\u53ea\u4f18\u5316\u51e0\u4e2a token\u201d\u5462\uff1f<\/p>\n\n\n\n
\u5f53\u7136\uff0c\u5b9e\u73b0\u7684\u65b9\u6cd5\u4e5f\u4e0d\u5c11\uff0c\u6bd4\u5982\u4e3a\u90a3\u51e0\u4e2a\u8981\u4f18\u5316\u7684 token \u91cd\u65b0\u6784\u5efa\u4e00\u4e2a Embedding \u5c42\uff0c\u7136\u540e\u62fc\u63a5\u5230 BERT \u7684Embedding\u5c42\u4e2d\uff0c\u7136\u540e\u8bad\u7ec3\u7684\u65f6\u5019\u53ea\u653e\u5f00\u65b0 Embedding \u5c42\u7684\u6743\u91cd\u3002\u4f46\u8fd9\u6837\u5199\u5bf9\u539f\u6765\u6a21\u578b\u7684\u6539\u52a8\u8fd8\u662f\u86ee\u5927\u7684\uff0c\u6700\u597d\u7684\u65b9\u6cd5\u662f\u5c3d\u53ef\u80fd\u5c11\u6539\u52a8\u4ee3\u7801\uff0c\u8ba9\u4f7f\u7528\u8005\u51e0\u4e4e\u65e0\u611f\u3002\u4e3a\u6b64\uff0c\u7b14\u8005\u6784\u601d\u4e86\u4e00\u79cd\u7528 stop_gradient<\/code> \u7b80\u5355\u4fee\u6539 Embedding<\/code> \u5c42\u7684\u65b9\u6848\uff0c\u5927\u4f53\u4e0a\u662f\u5c06 Embedding<\/code> \u5c42\u4fee\u6539\u5982\u4e0b\uff1a<\/p>\n\n\n\n
class PtuningEmbedding(Embedding):\n    \"\"\"\u65b0\u5b9a\u4e49Embedding\u5c42\uff0c\u53ea\u4f18\u5316\u90e8\u5206Token\n    \"\"\"\n    def call(self, inputs, mode='embedding'):\n        embeddings = self.embeddings\n        embeddings_sg = K.stop_gradient(embeddings)\n        mask = np.zeros((K.int_shape(embeddings)[0], 1))\n        mask[1:9] += 1  # \u53ea\u4f18\u5316id\u4e3a1\uff5e8\u7684token\n        self.embeddings = embeddings * mask + embeddings_sg * (1 - mask)\n        return super(PtuningEmbedding, self).call(inputs, mode)\n<\/code><\/pre>\n\n\n\n
\u53d8\u91cf\u7ecf\u8fc7 stop_gradient<\/code> \u7b97\u5b50\u540e\uff0c\u5728\u53cd\u5411\u4f20\u64ad\u7684\u65f6\u5019\u68af\u5ea6\u4e3a 0\uff0c\u4f46\u662f\u524d\u5411\u4f20\u64ad\u4e0d\u53d8\uff0c\u56e0\u6b64\u5728\u4e0a\u8ff0\u4ee3\u7801\u4e2d\uff0c\u524d\u5411\u4f20\u64ad\u7684\u7ed3\u679c\u4e0d\u4f1a\u6709\u53d8\u5316\uff0c\u4f46\u662f\u53cd\u5411\u4f20\u64ad\u6c42\u68af\u5ea6\u7684\u65f6\u5019\uff0c\u68af\u5ea6\u4e0d\u4e3a 0 \u7684 token \u7531 mask<\/code> \u53d8\u91cf\u63a7\u5236\uff0c\u5176\u4f59 token \u7684\u68af\u5ea6\u90fd\u4e3a\u96f6\uff0c\u56e0\u6b64\u5c31\u5b9e\u73b0\u4e86\u53ea\u66f4\u65b0\u90e8\u5206 token\u3002<\/p>\n\n\n\n
\u5b8c\u6574\u4ee3\u7801\u53ef\u89c1\uff1aGithub\uff1a<\/strong>https:\/\/github.com\/bojone\/P-tuning<\/a><\/p>\n\n\n\n
\u5bf9\u4e86\uff0c\u539f\u8bba\u6587\u4e5f\u5f00\u6e90\u4e86\u4ee3\u7801\uff1aGithub\uff1a<\/strong>https:\/\/github.com\/THUDM\/P-tuning<\/a><\/p>\n\n\n\n
\u6d4b\u8bd5\u4e0e\u6548\u679c<\/a><\/h3>\n\n\n\n
\u524d\u9762\u5df2\u7ecf\u5206\u4eab\u4e86\u539f\u4f5c\u8005\u5728 SuperGLUE \u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c\uff0c\u663e\u793a\u51fa\u5982\u679c\u914d\u5408 P-tuning\uff0c\u90a3\u4e48\uff1a1\u3001GPT\u3001BERT \u7684\u6548\u679c\u76f8\u6bd4\u76f4\u63a5 finetune \u90fd\u6709\u6240\u63d0\u5347\uff1b2\u3001GPT \u7684\u6548\u679c\u8fd8\u80fd\u8d85\u8fc7\u4e86 BERT\u3002\u8fd9\u8868\u660e GPT \u4e0d\u4ec5\u6709 NLG \u7684\u80fd\u529b\uff0c\u4e5f\u6709 NLU \u80fd\u529b\uff0c\u53ef\u8c13\u662f\u628a GPT \u7684\u6f5c\u80fd\u5145\u5206\u201c\u538b\u69a8\u201d\u51fa\u6765\u4e86\uff0c\u5f53\u7136 BERT \u914d\u5408 P-tuning \u4e5f\u6709\u63d0\u5347\uff0c\u8bf4\u660e P-tuning \u5bf9\u8bed\u8a00\u6a21\u578b\u6f5c\u80fd\u7684\u91ca\u653e\u662f\u8f83\u4e3a\u901a\u7528\u7684\u3002<\/p>\n\n\n\n
\u539f\u8bba\u6587\u7684\u5b9e\u9a8c\u6bd4\u8f83\u4e30\u5bcc\uff0c\u5efa\u8bae\u8bfb\u8005\u4ed4\u7ec6\u9605\u8bfb\u539f\u8bba\u6587\uff0c\u76f8\u4fe1\u4f1a\u6536\u83b7\u9887\u591a\u3002\u7279\u522b\u6307\u51fa\u7684\u662f\u539f\u8bba\u6587\u7684 Table 2 \u6700\u540e\u4e00\u5217\uff0c\u5f53\u9884\u8bad\u7ec3\u6a21\u578b\u8db3\u591f\u5927\u7684\u65f6\u5019\uff0c\u6211\u4eec\u7684\u8bbe\u5907\u53ef\u80fd\u65e0\u6cd5 finetune \u6574\u4e2a\u6a21\u578b\uff0c\u800c P-tuning \u53ef\u4ee5\u9009\u62e9\u53ea\u4f18\u5316\u51e0\u4e2a Token \u7684\u53c2\u6570\uff0c\u56e0\u4e3a\u4f18\u5316\u6240\u9700\u8981\u7684\u663e\u5b58\u548c\u7b97\u529b\u90fd\u4f1a\u5927\u5927\u51cf\u5c11\uff0c\u6240\u4ee5 P-tuning \u5b9e\u5219\u4e0a\u7ed9\u4e86\u6211\u4eec\u4e00\u79cd\u5728\u6709\u9650\u7b97\u529b\u4e0b\u8c03\u7528\u5927\u578b\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u601d\u8def<\/strong>\u3002<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
P-tuning \u5728\u5404\u4e2a\u4f53\u91cf\u7684\u8bed\u8a00\u6a21\u578b\u4e0b\u7684\u6548\u679c<\/p>\n\n\n\n
\u5f53\u7136\uff0c\u7b14\u8005\u4e00\u76f4\u4ee5\u6765\u7684\u89c2\u70b9\u662f\u201c\u6ca1\u6709\u5728\u4e2d\u6587\u4e0a\u6d4b\u8bd5\u8fc7\u7684\u7b97\u6cd5\u662f\u6ca1\u6709\u7075\u9b42\u7684<\/strong>\u201d\uff0c\u56e0\u6b64\u7b14\u8005\u4e5f\u5728\u4e2d\u6587\u4efb\u52a1\u4e0a\u7b80\u5355\u6d4b\u8bd5\u4e86\uff0c\u6d4b\u8bd5\u4efb\u52a1\u8ddf\u524d\u6587\u4e00\u81f4\uff0c\u90fd\u662f\u60c5\u611f\u5206\u7c7b\u7684\u5c0f\u6837\u672c\u5b66\u4e60\uff0c\u6d4b\u8bd5\u6a21\u578b\u5305\u62ec BERT \u548c GPT\uff0c\u4e24\u8005\u7684\u5019\u9009\u6a21\u7248\u5206\u522b\u5982\u4e0b\u56fe\uff1a<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
\u7b14\u8005\u5728\u4e2d\u6587\u60c5\u611f\u5206\u7c7b\u4e0a\u4f7f\u7528\u7684\u201cBERT+P-tuning\u201d\u6a21\u7248<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
\u7b14\u8005\u5728\u4e2d\u6587\u60c5\u611f\u5206\u7c7b\u4e0a\u4f7f\u7528\u7684\u201cGPT+P-tuning\u201d\u6a21\u7248<\/p>\n\n\n\n
\u6ce8\u610f\uff0c\u5bf9\u4e8e LM \u6a21\u578b\uff0c\u524d\u7f00\u7684\u5f15\u5165\u975e\u5e38\u91cd\u8981\uff0c\u53ea\u5f15\u5165\u540e\u7f00\u65f6\u6548\u679c\u4f1a\u660e\u663e\u53d8\u5dee\uff1b\u800c\u5bf9\u4e8e MLM \u6a21\u578b\uff0c\u524d\u7f00\u7684\u6548\u679c\u901a\u5e38\u4e5f\u4f18\u4e8e\u540e\u7f00\u3002\u603b\u7684\u6548\u679c\u5982\u4e0b\u8868\uff1a<\/p>\n\n\n\n
\"\"<\/figure><\/div>\n\n\n\n
\u5176\u4e2d\u201c\u5c0f\u6837\u672c\u201d\u53ea\u7528\u5230\u4e86\u201c\u5c11\u91cf\u6807\u6ce8\u6837\u672c\u201d\uff0c\u201c\u65e0\u76d1\u7763\u201d\u5219\u7528\u5230\u4e86\u201c\u5927\u91cf\u65e0\u6807\u6ce8\u6837\u672c\u201d\uff0c\u201c\u534a\u76d1\u7763\u201d\u5219\u7528\u5230\u4e86\u201c\u5c11\u91cf\u6807\u6ce8\u6837\u672c+\u5927\u91cf\u65e0\u6807\u6ce8\u6837\u672c\u201d\uff0c\u201cP-tuning\u201d \u90fd\u662f\u5c0f\u6837\u672c\uff0cPET \u7684\u51e0\u4e2a\u4efb\u52a1\u62a5\u544a\u7684\u662f\u6700\u4f18\u7684\u4eba\u5de5\u6a21\u7248\u7684\u7ed3\u679c\uff0c\u5176\u5b9e\u8fd8\u6709\u66f4\u5dee\u7684\u4eba\u5de5\u6a21\u7248\u3002\u4ece\u5c0f\u6837\u672c\u89d2\u5ea6\u6765\u770b\uff0cP-tuning \u786e\u5b9e\u53d6\u5f97\u4e86\u6700\u4f18\u7684\u5c0f\u6837\u672c\u5b66\u4e60\u6548\u679c\uff1b\u4ece\u6a21\u7248\u6784\u5efa\u7684\u89d2\u5ea6\u6765\u770b\uff0cP-tuning \u786e\u5b9e\u4e5f\u6bd4\u4eba\u5de5\u6784\u5efa\u7684\u6a21\u7248\u8981\u597d\u5f97\u591a\uff1b\u4ece\u6a21\u578b\u89d2\u5ea6\u770b\uff0cP-tuning \u786e\u5b9e\u53ef\u4ee5\u5c06 GPT \u7684\u5206\u7c7b\u6027\u80fd\u53d1\u6325\u5230\u8ddf BERT \u76f8\u8fd1\uff0c\u4ece\u800c\u63ed\u793a\u4e86 GPT \u4e5f\u6709\u5f88\u5f3a\u7684 NLU \u80fd\u529b\u7684\u4e8b\u5b9e\u3002<\/p>\n\n\n\n
\u8fdb\u4e00\u6b65\u7406\u89e3 P-tuning<\/a><\/h2>\n\n\n\n
\u8fd9\u4e00\u8282\u5c06\u4f1a\u4ecb\u7ecd\u7b14\u8005\u5bf9 P-tuning \u7684\u8fdb\u4e00\u6b65\u601d\u8003\uff0c\u4ee5\u6c42\u4ece\u591a\u4e2a\u7ef4\u5ea6\u6765\u7406\u89e3 P-tuning\u3002<\/p>\n\n\n\n
\u79bb\u6563 vs \u8fde\u7eed<\/a><\/h3>\n\n\n\n
\u5728 P-tuning \u4e4b\u524d\uff0c\u4e5f\u5df2\u7ecf\u6709\u4e00\u4e9b\u5728\u505a\u6a21\u7248\u7684\u81ea\u52a8\u6784\u5efa\uff0c\u5982\u300aHow Can We Know What Language Models Know?\u300b<\/a>\u3001\u300aAutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts\u300b<\/a>\u7b49\uff0c\u4f46\u5b83\u4eec\u641c\u7d22\u7684\u90fd\u662f\u5728\u79bb\u6563\u7a7a\u95f4\u4e0b\u641c\u7d22\u7684\u81ea\u7136\u8bed\u8a00\u6a21\u7248\uff0c\u6240\u4ee5\u6548\u679c\u6709\u6240\u9650\u5236\uff0c\u5e76\u6ca1\u6709\u53d6\u5f97\u7279\u522b\u7a81\u51fa\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n
\u76f8\u53cd\uff0cP-tuning \u653e\u5f03\u4e86\u201c\u6a21\u7248\u7531\u81ea\u7136\u8bed\u8a00\u6784\u6210\u201d\u8fd9\u4e00\u8981\u6c42\uff0c\u4ece\u800c\u5c06\u5176\u53d8\u6210\u4e86\u53ef\u4ee5\u7b80\u5355\u68af\u5ea6\u4e0b\u964d\u6c42\u89e3\u7684\u8fde\u7eed\u53c2\u6570\u95ee\u9898\uff0c\u6548\u679c\u8fd8\u66f4\u597d\u3002\u540c\u65f6\uff0c\u8fd9\u4e00\u6539\u52a8\u610f\u5473\u7740 P-tuning \u7a81\u51fa\u4e86\u6a21\u7248\u7684\u672c\u8d28\u2014\u2014\u5373\u6a21\u7248\u7684\u5173\u952e\u5728\u4e8e\u5b83\u662f\u600e\u4e48\u7528\u7684\uff0c\u4e0d\u5728\u4e8e\u5b83\u7531\u4ec0\u4e48\u6784\u6210<\/strong>\u2014\u2014\u7ed9\u4eba\u4e00\u79cd\u53bb\u829c\u5b58\u83c1\u3001\u773c\u524d\u4e00\u4eae\u7684\u611f\u89c9\uff0c\u786e\u5b9e\u503c\u5f97\u70b9\u8d5e\u3002<\/p>\n\n\n\n
\uff08\u6ce8\uff1a\u7ecf\u8bfb\u8005@brotherb<\/a>\u63d0\u9192\uff0c\u5e74\u521d\u6709\u4e00\u7bc7\u8bba\u6587\u300aPrefix-Tuning: Optimizing Continuous Prompts for Generation\u300b<\/a>\u63d0\u51fa\u7684 Prefix-Tuning \u65b9\u6cd5\u5176\u5b9e\u5df2\u7ecf\u76f8\u5f53\u63a5\u8fd1 P-tuning\uff0c\u4e24\u8005\u90fd\u8bbe\u8ba1\u4e86\u975e\u81ea\u7136\u8bed\u8a00\u7684\u6a21\u7248\uff0c\u53ea\u4e0d\u8fc7 Prefix-Tuning \u4e3b\u8981\u5173\u5fc3 NLG \u7684\u5e94\u7528\u800c P-tuning \u66f4\u52a0\u5173\u5fc3 NLU \u7684\u5e94\u7528\u3002\uff09<\/p>\n\n\n\n
Adapter<\/a><\/h3>\n\n\n\n
\u6211\u4eec\u8fd8\u53ef\u4ee5\u4ece Adapter \u7684\u89d2\u5ea6\u6765\u7406\u89e3 P-tuning\u3002BERT \u51fa\u6765\u540e\u4e0d\u4e45\uff0cGoogle \u5728\u8bba\u6587\u300aParameter-Ef\ufb01cient Transfer Learning for NLP\u300b<\/a>\u4e2d\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a Adapter \u7684\u5fae\u8c03\u65b9\u5f0f\uff0c\u5b83\u5e76\u4e0d\u662f\u76f4\u63a5\u5fae\u8c03\u6574\u4e2a\u6a21\u578b\uff0c\u800c\u662f\u56fa\u5b9a\u4f4f BERT \u539f\u59cb\u6743\u91cd\uff0c\u7136\u540e\u5728 BERT \u7684\u57fa\u7840\u4e0a\u6dfb\u52a0\u4e00\u4e9b\u6b8b\u5dee\u6a21\u5757\uff0c\u53ea\u4f18\u5316\u8fd9\u4e9b\u6b8b\u5dee\u6a21\u5757\uff0c\u7531\u4e8e\u6b8b\u5dee\u6a21\u5757\u7684\u53c2\u6570\u66f4\u5c11\uff0c\u56e0\u6b64\u5fae\u8c03\u6210\u672c\u66f4\u4f4e\u3002Adapter \u7684\u601d\u8def\u5b9e\u9645\u4e0a\u6765\u6e90\u4e8e CV \u7684\u300aLearning multiple visual domains with residual adapters\u300b<\/a>\uff0c\u4e0d\u8fc7\u8fd9\u4e24\u5e74\u4f3c\u4e4e\u5f88\u5c11\u770b\u5230\u4e86\uff0c\u4e5f\u8bb8\u662f\u56e0\u4e3a\u5b83\u867d\u7136\u63d0\u9ad8\u4e86\u8bad\u7ec3\u901f\u5ea6\uff0c\u4f46\u662f\u9884\u6d4b\u901f\u5ea6\u5374\u964d\u4f4e\u4e86\uff0c\u7cbe\u5ea6\u5f80\u5f80\u8fd8\u6709\u6240\u635f\u5931\u3002<\/p>\n\n\n\n
\u5728 P-tuning \u4e2d\uff0c\u5982\u679c\u6211\u4eec\u4e0d\u5c06\u65b0\u63d2\u5165\u7684 token \u89c6\u4e3a\u201c\u6a21\u7248\u201d\uff0c\u662f\u5c06\u5b83\u89c6\u4e3a\u6a21\u578b\u7684\u4e00\u90e8\u5206\uff0c\u90a3\u4e48\u5b9e\u9645\u4e0a P-tuning \u4e5f\u662f\u4e00\u79cd\u7c7b\u4f3c Adapter \u7684\u505a\u6cd5\uff0c\u540c\u6837\u662f\u56fa\u5b9a\u539f\u6a21\u578b\u7684\u6743\u91cd\uff0c\u7136\u540e\u63d2\u5165\u4e00\u4e9b\u65b0\u7684\u53ef\u4f18\u5316\u53c2\u6570\uff0c\u540c\u6837\u662f\u53ea\u4f18\u5316\u8fd9\u4e9b\u65b0\u53c2\u6570\uff0c\u53ea\u4e0d\u8fc7\u8fd9\u65f6\u5019\u65b0\u53c2\u6570\u63d2\u5165\u7684\u662f Embedding \u5c42\u3002\u56e0\u6b64\uff0c\u4ece\u8fd9\u4e2a\u89d2\u5ea6\u770b\uff0cP-tuning \u4e0e Adapter \u6709\u9887\u591a\u5f02\u66f2\u540c\u5de5\u4e4b\u5904\u3002<\/p>\n\n\n\n
\u4e3a\u4ec0\u4e48\u6709\u6548<\/a><\/h3>\n\n\n\n
\u7136\u540e\uff0c\u8fd8\u6709\u4e00\u4e2a\u503c\u5f97\u601d\u8003\u7684\u95ee\u9898\uff1a\u4e3a\u4ec0\u4e48 P-tuning \u4f1a\u66f4\u597d\uff1f\u6bd4\u5982\u5168\u91cf\u6570\u636e\u4e0b\uff0c\u5927\u5bb6\u90fd\u662f\u653e\u5f00\u6240\u6709\u6743\u91cd\uff0cP-tuning \u7684\u65b9\u6cd5\u4f9d\u7136\u6bd4\u76f4\u63a5 finetune \u8981\u597d\uff0c\u4e3a\u5565\u5462\uff1f<\/p>\n\n\n\n
\u4e8b\u5b9e\u4e0a\uff0c\u63d0\u51fa\u8fd9\u4e2a\u95ee\u9898\u7684\u8bfb\u8005\uff0c\u5e94\u8be5\u662f\u5bf9 BERT \u52a0\u4e2a\u5168\u8fde\u63a5\u5c42\u7684\u76f4\u63a5 finetune \u505a\u6cd5\u201c\u4e60\u4ee5\u4e3a\u5e38\u201d\u4e86\u3002\u5f88\u660e\u663e\uff0c\u4e0d\u7ba1\u662f PET \u8fd8\u662f P-tuning\uff0c\u5b83\u4eec\u5176\u5b9e\u90fd\u66f4\u63a5\u8fd1\u9884\u8bad\u7ec3\u4efb\u52a1\uff0c\u800c\u52a0\u4e2a\u5168\u8fde\u63a5\u5c42\u7684\u505a\u6cd5\uff0c\u5176\u5b9e\u8fd8\u6ca1\u90a3\u4e48\u63a5\u8fd1\u9884\u8bad\u7ec3\u4efb\u52a1\uff0c\u6240\u4ee5\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u6765\u8bf4\uff0cP-tuning \u6709\u6548\u66f4\u52a0\u201c\u663e\u7136\u201d\uff0c\u53cd\u800c\u662f\u52a0\u4e2a\u5168\u8fde\u63a5\u5c42\u5fae\u8c03\u4e3a\u4ec0\u4e48\u4f1a\u6709\u6548\u624d\u662f\u503c\u5f97\u7591\u95ee\u7684\u3002<\/p>\n\n\n\n
\u53bb\u5e74\u6709\u7bc7\u8bba\u6587\u300aA Mathematical Exploration of Why Language Models Help Solve Downstream Tasks\u300b<\/a>\u8bd5\u56fe\u56de\u7b54\u8fd9\u4e2a\u95ee\u9898\uff0c\u5927\u81f4\u7684\u8bba\u8bc1\u987a\u5e8f\u662f\uff1a<\/p>\n\n\n\n
1\u3001\u9884\u8bad\u7ec3\u6a21\u578b\u662f\u67d0\u79cd\u8bed\u8a00\u6a21\u578b\u4efb\u52a1\uff1b\n2\u3001\u4e0b\u6e38\u4efb\u52a1\u53ef\u4ee5\u8868\u793a\u4e3a\u8be5\u79cd\u8bed\u8a00\u6a21\u578b\u7684\u67d0\u4e2a\u7279\u6b8a\u60c5\u5f62\uff1b\n3\u3001\u5f53\u8f93\u51fa\u7a7a\u95f4\u6709\u9650\u7684\u65f6\u5019\uff0c\u5b83\u53c8\u8fd1\u4f3c\u4e8e\u52a0\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\uff1b\n4\u3001\u6240\u4ee5\u52a0\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\u5fae\u8c03\u662f\u6709\u6548\u7684\u3002\n<\/code><\/pre>\n\n\n\n
\u53ef\u4ee5\u770b\u5230\uff0c\u8be5\u8bba\u6587\u7684\u5047\u8bbe\u4e3b\u8981\u662f\u7b2c 2 \u70b9\uff0c\u5176\u5b9e\u5c31\u662f\u76f4\u63a5\u5047\u8bbe\u4e86\u4e0b\u6e38\u4efb\u52a1\u53ef\u4ee5\u8868\u8fbe\u4e3a\u7c7b\u4f3c PET \u7684\u5f62\u5f0f\uff0c\u7136\u540e\u624d\u53bb\u8bc1\u660e\u7684\u3002\u6240\u4ee5\u8fd9\u8fdb\u4e00\u6b65\u8bf4\u660e\u4e86\uff0cPET\u3001P-tuning \u7b49\u624d\u662f\u66f4\u81ea\u7136\u7684\u4f7f\u7528\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u65b9\u5f0f\uff0c\u52a0\u5168\u8fde\u63a5\u76f4\u63a5 finetune \u7684\u505a\u6cd5\u5176\u5b9e\u53ea\u662f\u5b83\u4eec\u7684\u63a8\u8bba\u7f62\u4e86\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0cPET\u3001P-tuning \u624d\u662f\u8fd4\u749e\u5f52\u771f\u3001\u56de\u5f52\u672c\u8d28\u7684\u65b9\u6848\uff0c\u6240\u4ee5\u5b83\u4eec\u66f4\u6709\u6548\u3002<\/p>\n\n\n\n
\u8f6c\u8f7d\u81ea\u300a\u5fc5\u987b\u8981GPT3\u5417\uff1f\u4e0d\uff0cBERT\u7684MLM\u6a21\u578b\u4e5f\u80fd\u5c0f\u6837\u672c\u5b66\u4e60\u300b<\/a>\u548c\u300aP-tuning\uff1a\u81ea\u52a8\u6784\u5efa\u6a21\u7248\uff0c\u91ca\u653e\u8bed\u8a00\u6a21\u578b\u6f5c\u80fd\u300b<\/a>\uff0c\u4f5c\u8005\uff1a\u82cf\u5251\u6797\uff0c\u90e8\u5206\u5185\u5bb9\u6709\u4fee\u6539\u3002<\/p><\/blockquote>\n","protected":false},"excerpt":{"rendered":"
\u8f6c\u8f7d\u81ea\u300a\u5fc5\u987b\u8981GPT3\u5417\uff1f\u4e0d\uff0cBERT\u7684MLM\u6a21\u578b\u4e5f\u80fd\u5c0f\u6837\u672c\u5b66\u4e60\u300b\u548c\u300aP-tuning\uff1a\u81ea\u52a8\u6784\u5efa\u6a21\u7248\uff0c\u91ca\u653e\u8bed\u8a00 … \u7ee7\u7eed\u9605\u8bfb\u8fd0\u7528 BERT \u7684 MLM \u6a21\u578b\u8fdb\u884c\u5c0f\u6837\u672c\u5b66\u4e60<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[19,21,4],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/9594"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=9594"}],"version-history":[{"count":11,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/9594\/revisions"}],"predecessor-version":[{"id":9612,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/9594\/revisions\/9612"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=9594"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=9594"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=9594"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}