{"id":1072557,"date":"2023-07-04T22:33:00","date_gmt":"2023-07-05T05:33:00","guid":{"rendered":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/?post_type=msr-blog-post&p=1072557"},"modified":"2024-09-25T03:41:00","modified_gmt":"2024-09-25T10:41:00","slug":"acl-2023-responsible-ai","status":"publish","type":"msr-blog-post","link":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/articles\/acl-2023-responsible-ai\/","title":{"rendered":"ACL 2023\uff5c\u5927\u6a21\u578b\u65f6\u4ee3\uff0c\u81ea\u7136\u8bed\u8a00\u9886\u57df\u8fd8\u6709\u4ec0\u4e48\u5b66\u672f\u589e\u957f\u70b9\uff1f"},"content":{"rendered":"\n

\u56fd\u9645\u8ba1\u7b97\u8bed\u8a00\u5b66\u5e74\u4f1a\uff08Annual Meeting of the Association for Computational Linguistics\uff0c\u7b80\u79f0 ACL\uff09\u662f\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u9886\u57df\u7684\u9876\u7ea7\u56fd\u9645\u4f1a\u8bae\uff0cACL 2023 \u5c06\u4e8e2023\u5e747\u67089-14\u65e5\u5728\u52a0\u62ff\u5927\u591a\u4f26\u591a\u4e3e\u884c\u3002\u968f\u7740\u4eba\u5de5\u667a\u80fd\u6280\u672f\u7684\u5feb\u901f\u53d1\u5c55\uff0c\u786e\u4fdd\u76f8\u5173\u6280\u672f\u80fd\u88ab\u4eba\u4eec\u4fe1\u8d56\u662f\u4e00\u4e2a\u9700\u8981\u653b\u575a\u7684\u95ee\u9898\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e5f\u5728\u4e0d\u65ad\u63a8\u8fdb\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u7684\u63a2\u7d22\u53d1\u73b0\u4e0e\u5e94\u7528\u5b9e\u8df5\u3002\u4eca\u5929\u6211\u4eec\u4e3a\u5927\u5bb6\u5e26\u67653\u7bc7\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4ee5\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u4e3a\u4e3b\u9898\u5165\u9009 ACL 2023 \u7684\u8bba\u6587\u3002<\/p>\n\n\n\n

DuNST: \u57fa\u4e8e\u566a\u58f0\u5bf9\u5076\u81ea\u5b66\u4e60\u7684\u534a\u76d1\u7763\u53ef\u63a7\u6587\u672c\u751f\u6210<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2212.08724<\/p>\n\n\n\n

\u8fd1\u5e74\u6765\uff0c\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u5e7f\u6cdb\u5e94\u7528\uff0c\u6784\u5efa\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u6210\u4e3a\u4e86\u4e00\u4e2a\u91cd\u8981\u7684\u8bfe\u9898\u3002\u751f\u6210\u6ca1\u6709\u504f\u89c1\u3001\u65e0\u6bd2\u6027\u7684\u6587\u672c\u662f\u751f\u6210\u5f0f\u6a21\u578b\u7684\u57fa\u672c\u8981\u6c42\u3002\u4f5c\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u9886\u57df\u7684\u7ecf\u5178\u548c\u70ed\u70b9\u4efb\u52a1\uff0c\u53ef\u63a7\u6587\u672c\u751f\u6210\u5728\u6587\u672c\u53bb\u6bd2\u53bb\u504f\u4e0a\u6709\u7740\u91cd\u8981\u7684\u5e94\u7528\u3002\u8fd1\u6765\u5927\u578b\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u4e2d\u8bb8\u591a NLP \u4efb\u52a1\u91c7\u7528\u7684\u4e3b\u6d41\u65b9\u6cd5\u662f\u5fae\u8c03\uff0c\u7136\u800c\uff0c\u6a21\u578b\u5927\u5c0f\u7684\u589e\u52a0\u9700\u8981\u66f4\u591a\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u6807\u6ce8\u6570\u636e\u7684\u4e25\u91cd\u4e0d\u8db3\u5c06\u5bfc\u81f4\u5fae\u8c03\u7ed3\u679c\u4e0d\u7a33\u5b9a\u3002\u81ea\u5b66\u4e60\uff08Self-Training\uff09\u662f\u89e3\u51b3\u53ef\u63a7\u6587\u672c\u751f\u6210\u6570\u636e\u7a00\u7f3a\u95ee\u9898\u7684\u6709\u6548\u624b\u6bb5\uff0c\u5e76\u9010\u6e10\u5f15\u8d77\u4e86\u7814\u7a76\u8005\u7684\u91cd\u89c6\u3002\u5728\u6bcf\u4e00\u8f6e\u8fed\u4ee3\u8fc7\u7a0b\u4e2d\uff0c\u81ea\u5b66\u4e60\u4f7f\u7528\u5206\u7c7b\u5668\u4e3a\u672a\u6807\u6ce8\u6570\u636e\u751f\u6210\u4f2a\u6807\u7b7e\uff0c\u518d\u4f7f\u7528\u589e\u5f3a\u7684\u6570\u636e\u91cd\u65b0\u8fed\u4ee3\u8bad\u7ec3\u5206\u7c7b\u5668\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u81ea\u5b66\u4e60\u80fd\u591f\u5229\u7528\u672a\u6807\u6ce8\u6570\u636e\u8fdb\u4e00\u6b65\u6539\u8fdb\u5206\u7c7b\u8fb9\u754c\uff0c\u63d0\u9ad8\u5728\u771f\u5b9e\u6570\u636e\u4e0a\u7684\u6cdb\u5316\u6027\u3002<\/p>\n\n\n\n

\u65e2\u6709\u7814\u7a76\u4e3b\u8981\u9488\u5bf9\u81ea\u5b66\u4e60\u6587\u672c\u5206\u7c7b\u4efb\u52a1\u8fdb\u884c\u63d0\u5347\uff0c\u4f46\u5c06\u81ea\u5b66\u4e60\u5e94\u7528\u4e8e\u53ef\u63a7\u6587\u672c\u751f\u6210\u4ecd\u7136\u5b58\u5728\u8bf8\u591a\u6311\u6218\u3002\u9996\u5148\uff0c\u67d0\u4e00\u7c7b\u7279\u5b9a\u7c7b\u522b\u7684\u6587\u672c\u6570\u636e\u53ef\u80fd\u662f\u7a7a\u767d\u3002\u5176\u6b21\uff0c\u4ec5\u901a\u8fc7\u81ea\u52a8\u751f\u6210\u7684\u4f2a\u6587\u672c\u8fdb\u884c\u589e\u5f3a\uff0c\u5148\u524d\u5b66\u4e60\u7684\u6587\u672c\u7a7a\u95f4\u5c06\u88ab\u8fc7\u5ea6\u5229\u7528\uff0c\u56e0\u6b64\uff0c\u6a21\u578b\u4f1a\u5ffd\u89c6\u5176\u4ed6\u6587\u672c\u7a7a\u95f4\uff0c\u5bfc\u81f4\u6587\u672c\u7a7a\u95f4\u7684\u5d29\u6e83\u4ee5\u53ca\u6587\u672c\u751f\u6210\u8d28\u91cf\u7684\u6076\u5316\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u81ea\u5b66\u4e60\u6846\u67b6 DuNST\u3002DuNST \u5c06\u6587\u672c\u751f\u6210\u548c\u5206\u7c7b\u4f5c\u4e3a\u5bf9\u5076\u8fc7\u7a0b\u8fdb\u884c\u8054\u5408\u5efa\u6a21\uff0c\u5e76\u4e14\u901a\u8fc7\u81ea\u5b66\u4e60\u6765\u6539\u8fdb\u6a21\u578b\u7684\u751f\u6210\u548c\u5206\u7c7b\u3002\u9664\u4e86\u751f\u6210\u5668\u751f\u6210\u7684\u4f2a\u6587\u672c\u4e4b\u5916\uff0c\u7814\u7a76\u5458\u4eec\u8fd8\u5229\u7528\u5206\u7c7b\u5668\u7ed9\u672a\u6807\u6ce8\u7684\u6587\u672c\u505a\u4f2a\u6807\u6ce8\u3002\u53e6\u5916\uff0c\u6a21\u578b\u4f1a\u901a\u8fc7\u6dfb\u52a0\u4e24\u79cd\u5e73\u6ed1\u7684\u566a\u58f0\u6765\u6270\u52a8\u5b66\u5230\u7684\u6587\u672c\u7a7a\u95f4\u3002\u6dfb\u52a0\u7684\u566a\u58f0\u53ef\u4ee5\u5e2e\u52a9\u6a21\u578b\u6539\u8fdb\u6587\u672c\u7a7a\u95f4\u7684\u5c40\u90e8\u5e73\u6ed1\u6027\uff0c\u589e\u52a0\u6a21\u578b\u7684\u9c81\u68d2\u6027\u3002\u7406\u8bba\u4e0a\uff0cDuNST \u53ef\u4ee5\u88ab\u770b\u4f5c\u662f\u5bf9\u63a2\u7d22\uff08Exploration\uff09\u548c\u5229\u7528\uff08Exploitation\uff09\u7684\u5e73\u8861\u3002\u6dfb\u52a0\u566a\u58f0\u589e\u5f3a\u4e86\u5bf9\u6f5c\u5728\u66f4\u5927\u7684\u771f\u5b9e\u6587\u672c\u7a7a\u95f4\u7684\u63a2\u7d22\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u5bf9\u5df2\u7ecf\u5b66\u5230\u7684\u6587\u672c\u7a7a\u95f4\u7684\u5229\u7528\uff0c\u4ece\u800c\u4fdd\u8bc1\u6a21\u578b\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\"\u56fe1\uff1aDuNST
\u56fe1\uff1aDuNST \u901a\u8fc7\u81ea\u5b66\u4e60\u548c\u6dfb\u52a0\u566a\u58f0\u7ef4\u6301\u5229\u7528\u548c\u63a2\u7d22\u7684\u5e73\u8861\u3002<\/figcaption><\/figure>\n\n\n\n

<\/p>\n\n\n\n

\u5bf9\u4e09\u4e2a\u534a\u76d1\u7763\u53ef\u63a7\u751f\u6210\u4efb\u52a1\u7684\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u4e0e\u4f20\u7edf\u7684\u81ea\u5b66\u4e60\u65b9\u6cd5\u76f8\u6bd4\uff0cDuNST \u663e\u8457\u63d0\u9ad8\u4e86\u63a7\u5236\u51c6\u786e\u6027\u548c\u6587\u672c\u751f\u6210\u7684\u591a\u6837\u6027\uff0c\u5e76\u4fdd\u6301\u4e86\u751f\u6210\u7684\u6d41\u7545\u5ea6\u548c\u6cdb\u5316\u6027\u3002\u5728\u6587\u672c\u53bb\u6bd2\u4efb\u52a1\u4e0a\uff0c\u548c\u5fae\u8c03\u57fa\u7ebf\u6a21\u578b\uff08\u5982 GPT2, UniLM \u7b49\uff09\u76f8\u6bd4\uff0cDuNST \u663e\u8457\u964d\u4f4e\u4e86\u751f\u6210\u6587\u672c\u7ea660%\u7684\u6bd2\u6027\uff0c\u5e76\u4fdd\u6301\u4e86\u8f83\u9ad8\u7684\u751f\u6210\u8d28\u91cf\u3002\u8be5\u65b9\u6cd5\u4e3a\u6587\u672c\u53bb\u6bd2\u63d0\u4f9b\u4e86\u4e00\u79cd\u4f4e\u6210\u672c\u7684\u6709\u6548\u65b9\u6848\u3002<\/p>\n\n\n\n

EmbMarker: \u901a\u8fc7\u540e\u95e8\u6c34\u5370\u4fdd\u62a4\u57fa\u4e8e\u5927\u6a21\u578b\u7684\u5411\u91cf\u8868\u793a\u670d\u52a1\u7684\u7248\u6743<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2305.10036<\/p>\n\n\n\n

\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/yjw1029\/EmbMarker<\/p>\n\n\n\n

LLMs \u5728\u6587\u672c\u7406\u89e3\u548c\u751f\u6210\u65b9\u9762\u8868\u73b0\u4e86\u51fa\u5f3a\u5927\u7684\u80fd\u529b\u3002\u56e0\u6b64\uff0c\u8bb8\u591a\u516c\u53f8\u5f00\u59cb\u57fa\u4e8e\u8fd9\u4e9b LLMs \u63d0\u4f9b\u5411\u91cf\u8868\u793a\u670d\u52a1\uff08EaaS\uff09\uff0c\u4ee5\u5e2e\u52a9\u5ba2\u6237\u5b8c\u6210\u5404\u79cd NLP \u4efb\u52a1\u3002\u7136\u800c\uff0c\u73b0\u6709\u7684\u7814\u7a76\u8868\u660e\uff0c\u7528\u6237\u53ef\u4ee5\u901a\u8fc7\u53d1\u9001\u67e5\u8be2\u548c\u63a5\u6536\u8f93\u51fa\u6765\u91cd\u5efa\u6a21\u578b\u7684\u53c2\u6570\uff0c\u8fd9\u4f7f\u5f97 LLM \u7684\u670d\u52a1\u63d0\u4f9b\u8005\u9762\u4e34\u7740\u6a21\u578b\u88ab\u76d7\u7528\u6216\u590d\u5236\u7684\u98ce\u9669\u3002<\/p>\n\n\n\n

\u6c34\u5370\u662f\u5e38\u7528\u7684\u6a21\u578b\u7248\u6743\u4fdd\u62a4\u65b9\u6cd5\u3002\u7136\u800c\uff0c\u73b0\u6709\u7684\u6c34\u5370\u90fd\u4e0d\u9002\u7528\u4e8e EaaS\u3002\u56e0\u6b64\uff0c\u6709\u5fc5\u8981\u63d0\u51fa\u4e00\u79cd\u9002\u7528\u4e8e EaaS \u7684\u6c34\u5370\u65b9\u6cd5\u3002\u8fd9\u7c7b\u65b9\u6cd5\u9700\u8981\u6ee1\u8db3\u4ee5\u4e0b\u6761\u4ef6\uff1a1. \u4e0d\u5f71\u54cd\u5411\u91cf\u5728\u4e0b\u6e38\u4efb\u52a1\u7684\u4f7f\u7528\u6027\u80fd\uff1b2. \u5f53\u6709\u76d7\u7528\u8005\u590d\u5236\u63d0\u4f9b\u5546\u7684\u6a21\u578b\u5e76\u63d0\u4f9b\u76f8\u540c\u7684\u7ade\u4e89\u670d\u52a1\u65f6\uff0c\u63d0\u4f9b\u5546\u53ef\u4ee5\u901a\u8fc7\u8bbf\u95ee\u76d7\u7528\u8005\u7684\u670d\u52a1\u9a8c\u8bc1\u5176\u8f93\u51fa\u4e2d\u542b\u6709\u63d0\u4f9b\u5546\u7684\u6c34\u5370\uff1b3. \u6c34\u5370\u9700\u8981\u8db3\u591f\u9690\u853d\uff0c\u4e0d\u4f1a\u88ab\u76d7\u7528\u8005\u8f7b\u6613\u5730\u8fc7\u6ee4\u6389\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u540e\u95e8\u6c34\u5370\u7684\u65b9\u6cd5\uff1aEmbMarker\u3002EmbMarker \u5305\u542b\u4e24\u4e2a\u9636\u6bb5\uff1a\u6c34\u5370\u63d2\u5165\u9636\u6bb5\u548c\u7248\u6743\u9a8c\u8bc1\u9636\u6bb5\u3002\u5728\u6c34\u5370\u63d2\u5165\u9636\u6bb5\uff0c\u7814\u7a76\u5458\u4eec\u9996\u5148\u627e\u5230\u4e00\u7ec4\u5408\u9002\u8bcd\u9891\u7684\u5355\u8bcd\u4f5c\u4e3a\u89e6\u53d1\u5355\u8bcd\uff0c\u5e76\u9884\u5b9a\u4e49\u4e00\u4e2a\u76ee\u6807\u5411\u91cf\u4f5c\u4e3a\u6c34\u5370\u3002\u5f53\u7528\u6237\u63d0\u4f9b\u7684\u53e5\u5b50\u4e2d\u542b\u6709\u7684\u89e6\u53d1\u5355\u8bcd\u6570\u91cf\u8d8a\u591a\u65f6\uff0c\u670d\u52a1\u63d0\u4f9b\u8005\u53d1\u9001\u7684\u5411\u91cf\u4e0e\u9884\u5b9a\u4e49\u7684\u76ee\u6807\u5411\u91cf\u7684\u8ddd\u79bb\u8d8a\u63a5\u8fd1\u3002\u5728\u7248\u6743\u9a8c\u8bc1\u9636\u6bb5\uff0c\u63d0\u4f9b\u5546\u53ef\u4ee5\u4f7f\u7528\u89e6\u53d1\u5355\u8bcd\u548c\u975e\u89e6\u53d1\u5355\u8bcd\u5206\u522b\u6784\u9020\u4e24\u7ec4\u53e5\u5b50\uff0c\u5e76\u8bbf\u95ee\u5f85\u9a8c\u8bc1\u7684\u670d\u52a1\u5f97\u5230\u4e24\u7ec4\u5411\u91cf\u3002\u4e24\u7ec4\u5411\u91cf\u79bb\u76ee\u6807\u5411\u91cf\u7684\u8ddd\u79bb\u5206\u5e03\u5dee\u8ddd\u8d8a\u5927\uff0c\u5219\u8bf4\u660e\u8be5\u670d\u52a1\u540e\u7684\u6a21\u578b\u8d8a\u6709\u53ef\u80fd\u662f\u76d7\u7528\u6216\u590d\u5236\u4e86\u63d0\u4f9b\u5546\u7684\u6a21\u578b\u3002\u5728\u591a\u4e2a\u6570\u636e\u96c6\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cEmbMarker \u53ef\u4ee5\u5728\u4e0d\u5f71\u54cd\u5411\u91cf\u5728\u4e0b\u6e38\u4efb\u52a1\u6027\u80fd\u7684\u60c5\u51b5\u4e0b\uff0c\u4ee5\u9ad8\u7f6e\u4fe1\u5ea6\u9a8c\u8bc1\u76d7\u7528\u8005\u670d\u52a1\u4e2d\u7684\u6c34\u5370\uff0c\u5e76\u4e14\u5177\u6709\u5f88\u5f3a\u7684\u9690\u853d\u6027\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\"\u56fe2\uff1aEmbMarker
\u56fe2\uff1aEmbMarker \u7684\u6846\u67b6<\/figcaption><\/figure>\n\n\n\n

<\/p>\n\n\n\n

GLUE-X\uff1a\u57fa\u4e8e\u5206\u5e03\u5916\u6cdb\u5316\u7684\u81ea\u7136\u8bed\u8a00\u7406\u89e3\u6a21\u578b\u6d4b\u8bd5\u96c6<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2211.08073<\/p>\n\n\n\n

\u673a\u5668\u5b66\u4e60\u7684\u8bb8\u591a\u9886\u57df\u90fd\u9762\u4e34\u7740\u4e00\u4e2a\u5171\u540c\u7684\u96be\u9898\uff1a\u8bc4\u4f30\u3002\u8fd1\u4e9b\u5e74\u6765\uff0c\u867d\u7136\u673a\u5668\u5b66\u4e60\u53d6\u5f97\u4e86\u5f88\u591a\u8fdb\u5c55\uff0c\u4f46\u968f\u7740\u7814\u7a76\u7684\u6df1\u5165\uff0c\u7814\u7a76\u4eba\u5458\u53d1\u73b0\u8fd9\u4e9b\u8fdb\u5c55\u7684\u6cdb\u5316\u6027\u5e76\u4e0d\u5982\u9884\u671f\u7684\u4f18\u79c0\u3002\u4f20\u7edf\u8bed\u8a00\u6a21\u578b\u7684\u8bc4\u4f30\u5927\u591a\u4f9d\u8d56\u4e8e GLUE \u6392\u884c\u699c\u3002\u622a\u6b62\u81f32022\u5e74\uff0c\u5df2\u7ecf\u6709\u8d85\u8fc720\u4e2a\u5355\u6a21\u578b\u7684\u7ed3\u679c\u5728 GLUE \u7684\u6d4b\u8bc4\u4e0a\u4f18\u4e8e\u4eba\u5de5\u6d4b\u8bc4\u7684\u8868\u73b0\u3002\u8fc7\u53bb\u7684\u5de5\u4f5c\u8bc1\u660e\u4e86\u6a21\u578b\u7684\u8868\u73b0\u5e76\u4e0d\u662f\u771f\u6b63\u8d85\u8fc7\u4e86\u4eba\u7c7b\uff0c\u800c\u662f\u4f9d\u9760\u4f2a\u7279\u5f81\uff08spurious features\uff09\u548c\u6377\u5f84\u5b66\u4e60\uff08shortcut learning\uff09\u53d6\u5f97\u4e86\u865a\u9ad8\u7684\u6210\u7ee9\u3002\u56e0\u6b64\u5728\u6a21\u578b\u62df\u5408\u80fd\u529b\u5927\u5927\u63d0\u5347\u7684\u4eca\u5929\uff0c\u4f9d\u9760\u4f20\u7edf in-domain test \u7684 GLUE \u699c\u5355\u5728\u5b9e\u8df5\u4e2d\u4f5c\u4e3a\u8bc4\u4f30\u6307\u6807\u7684\u5b9e\u9645\u4ef7\u503c\u8f83\u4f4e\u3002\u6240\u4ee5\u9700\u8981\u9760\u5206\u5e03\u5916\u6cdb\u5316\uff08Out of Distribution, OOD\uff09\u6765\u6d4b\u8bd5\u6a21\u578b\u771f\u6b63\u7684\u6cdb\u5316\u80fd\u529b\u3002\u4ee5\u5f80\u7684\u6cdb\u5316\u8bc4\u4f30\u901a\u5e38\u662f\u7814\u7a76\u8005\u81ea\u884c\u9009\u62e9\u6570\u636e\u96c6\u57281-2\u4e2a\u4efb\u52a1\u4e0a\u8fdb\u884c\u6d4b\u8bd5\uff0c\u7f3a\u4e4f GLUE-X \u8fd9\u6837\u5168\u9762\u8bc4\u4f30\u6a21\u578b\u6cdb\u5316\u80fd\u529b\u7684\u57fa\u51c6\u3002\u4e0d\u540c\u4e8e GLUE\uff0c\u5f53\u524d\u6700\u597d\u7684\u6a21\u578b\u5728 GLUE-X \u8868\u73b0\u4ecd\u660e\u663e\u900a\u4e8e\u4eba\u7c7b(74.6% vs. 80.4%)\u3002<\/p>\n\n\n\n

\u4ec0\u4e48\u662f\u5206\u5e03\u5916\u6cdb\u5316\uff1f<\/em><\/p>\n\n\n\n

\u5047\u8bbe\u6709\u4e00\u4e2a\u5e26\u6807\u7b7e\u7684\u6570\u636e\u96c6\u5408\u3002\u901a\u8fc7\u4ece\u5206\u5e03 P_train \u4e2d\u5bf9 D \u8fdb\u884c\u91c7\u6837\uff0c\u751f\u6210\u4e00\u4e2a\u8bad\u7ec3\u6570\u636e\u96c6 D_train = {(X_train, Y_train)}\u3002\u6d4b\u8bd5\u6570\u636e\u96c6 D_test = {(X_test, Y_test)} \u662f\u4ece D \u4e2d\u6309\u7167\u5206\u5e03 P_test \u8fdb\u884c\u91c7\u6837\u5f97\u5230\u7684\u3002\u5f53 P_train \u2260 P_test \u65f6\uff0c\u79f0\u4e4b\u4e3a\u5206\u5e03\u5916\u6cdb\u5316\u3002<\/em><\/p>\n\n\n\n

\u800c\u4e14\u5728\u4eba\u5de5\u667a\u80fd\u5f97\u4ee5\u5e7f\u6cdb\u8fd0\u7528\u7684\u4eca\u5929\uff0c\u6784\u5efa\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u9700\u8981\u6a21\u578b\u5177\u5907\u8db3\u591f\u7684\u9c81\u68d2\u6027\u3002\u4f46\u5728 NLP \u7684\u8fc7\u5f80\u7814\u7a76\u4e2d\uff0cOOD \u5e76\u6ca1\u6709\u5f97\u5230\u8db3\u591f\u7684\u5173\u6ce8\u4e14\u7f3a\u4e4f\u7edf\u4e00\u7684\u8bc4\u4f30\u57fa\u51c6\uff0c\u8fd9\u9650\u5236\u4e86 NLP \u7cfb\u7edf\u5728\u771f\u5b9e\u4e16\u754c\u4e2d\u7684\u5e94\u7528\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u6784\u5efa\u9488\u5bf9\u6a21\u578b\u6cdb\u5316\u80fd\u529b\u7684\u7edf\u4e00\u57fa\u51c6\uff0c\u7814\u7a76\u5458\u4eec\u521b\u5efa\u4e86\u4e00\u4e2a\u540d\u4e3a GLUE-X \u7684\u8bc4\u6d4b\u699c\u5355\u3002\u9996\u5148\uff0c\u7814\u7a76\u5458\u4eec\u4ee5 GLUE \u4e0a\u56ca\u62ec\u7684\u6570\u636e\u96c6\u4f5c\u4e3a\u9886\u57df\u5185\u8bad\u7ec3\u96c6\uff0c\u57288\u4e2a\u6587\u672c\u5206\u7c7b\u4efb\u52a1\u4e0a\uff0c\u6784\u5efa\u4e8614\u4e2a\u7528\u4e8e OOD \u6d4b\u8bd5\u7684\u6587\u672c\u6570\u636e\u96c6\u3002\u7136\u540e\uff0c\u53c8\u572821\u4e2a\u5e38\u7528\u7684\u9884\u8bad\u7ec3\u6a21\u578b\uff08\u5305\u62ec InstructGPT \u548c GPT 3.5\uff09\u4e0a\u5229\u7528\u9886\u57df\u5185\u7684\u8bad\u7ec3\u96c6\u8fdb\u884c\u8c03\u53c2\uff0c\u5f97\u5230\u9886\u57df\u5185\u6700\u4f73\u6027\u80fd\u7684\u6a21\u578b\u540e\uff0c\u518d\u5728 OOD \u6587\u672c\u6570\u636e\u96c6\u8fdb\u884c\u6d4b\u8bd5\uff0c\u4ee5 OOD \u6570\u636e\u4e0a\u7684\u8868\u73b0\u4f5c\u4e3a\u6a21\u578b\u6cdb\u5316\u80fd\u529b\u7684\u6307\u6807\uff0c\u540c\u65f6\u63d0\u4f9b\u4eba\u7c7b\u6d4b\u8bc4\u7684\u7ed3\u679c\u4f5c\u4e3a\u53c2\u7167\u3002\u6b64\u5916\uff0c\u7814\u7a76\u5458\u4eec\u8fd8\u6bd4\u8f83\u4e86\u4e0d\u540c\u7684\u5fae\u8c03\u65b9\u5f0f\u5bf9\u6a21\u578b\u6cdb\u5316\u6027\u80fd\u7684\u5f71\u54cd\uff0c\u5e76\u5229\u7528 Rationale \u201c\u4e8b\u540e\u5206\u6790\u201d\u4e86\u6a21\u578b\u5728 OOD \u6570\u636e\u4e0a\u4f5c\u51fa\u5224\u65ad\u7684\u7406\u6027\u4f9d\u636e\uff0c\u5e76\u4e0e\u4eba\u5de5\u6807\u6ce8\u7684\u6570\u636e\u8fdb\u884c\u6bd4\u5bf9\u5206\u6790\uff0c\u4ee5\u5e2e\u52a9\u7814\u7a76\u4eba\u5458\u7406\u89e3\u6a21\u578b\u6cdb\u5316\u80fd\u529b\u7684\u6765\u6e90\u3002<\/p>\n\n\n\n

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

\u5982\u56fe3\u6240\u793a\uff0c\u7814\u7a76\u5458\u4eec\u5bf9\u6bcf\u4e00\u4e2a GLUE \u4e2d\u51fa\u73b0\u7684\u4efb\u52a1\u6784\u5efa\u4e86\u5bf9\u5e94\u7684 OOD \u6570\u636e\u3002\u4f8b\u5982\uff0c\u5bf9\u60c5\u611f\u5206\u6790 SST-2 \u6570\u636e\uff0c\u9009\u53d6\u4e86 IMDB\u3001Yelp\u3001Amazon \u548c Flipkart \u4f5c\u4e3a\u6d4b\u8bd5\u6570\u636e\u3002\u5bf9\u8bed\u6cd5\u5224\u65ad COLA \u6570\u636e\uff0c\u9009\u53d6\u4e86\u81ea\u884c\u6536\u96c6\u7684 Grammar Test\uff08\u8003\u9898\uff09\u4f5c\u4e3a\u6d4b\u8bd5\u6570\u636e\u3002GLUE-X \u603b\u5171\u5305\u542b\u5341\u4e94\u7ec4\uff0c\u8d85\u8fc7600\u4e07\u6761\u7684\u6cdb\u5316\u6d4b\u8bd5\u6570\u636e\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u7814\u7a76\u5458\u4eec\u5bf9\u5e38\u89c1\u7684 PLM \u8fdb\u884c\u4e86\u5168\u6570\u636e\u6d4b\u8bd5\u3002\u4ea6\u5bf9 InstructGPT \u548c ChatGPT\u8fdb\u884c\u4e86\u91c7\u6837\u6d4b\u8bd5\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\"\u56fe3\uff1a\u4e0d\u540c
\u56fe3\uff1a\u4e0d\u540c OOD \u4efb\u52a1\u7684\u6d4b\u8bd5\u7ed3\u679c<\/figcaption><\/figure>\n\n\n\n

<\/p>\n\n\n\n

\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff1a<\/p>\n\n\n\n

(1) \u65e0\u8bba\u662f\u6700\u4f73\u7684\u6709\u76d1\u7763\u5b66\u4e60\u6a21\u578b\uff0c\u8fd8\u662f ChatGPT \u5927\u6a21\u578b\uff0c\u5728 GLUE-X \u4e0a\u7684\u8868\u73b0\u90fd\u8fdc\u8fdc\u4f4e\u4e8e\u4eba\u7c7b\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4eba\u5de5\u6d4b\u8bc4\u4e5f\u662f\u5728 OOD \u6761\u4ef6\u4e0b\u8fdb\u884c\u7684\uff08\u4ec5\u7ed9\u4eba\u7c7b in-domain \u7684\u6570\u636e\u4f5c\u4e3a\u57f9\u8bad\u8303\u4f8b\uff09\u3002<\/p>\n\n\n\n

(2) \u6ca1\u6709\u4e00\u79cd\u6a21\u578b\u80fd\u9886\u8dd1\u6240\u6709\u4efb\u52a1\uff0c\u8fd9\u4e0e\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u7684\u7814\u7a76\u7ed3\u8bba\u4e00\u81f4\u3002<\/p>\n\n\n\n

(3) \u6a21\u578b\u67b6\u6784 OOD \u9c81\u68d2\u6027\u7684\u5f71\u54cd\u6bd4\u6a21\u578b\u53c2\u6570\u5927\u5c0f\u66f4\u4e3a\u91cd\u8981\u3002\u6a21\u578b\u7684\u7ed3\u6784\u5bf9\u4e8e\u5904\u7406\u672a\u9884\u6599\u5230\u7684\u8f93\u5165\u66f4\u5177\u6709\u5f71\u54cd\u529b\u3002<\/p>\n\n\n\n

(4) \u5bf9\u4e8e\u6587\u672c\u5206\u7c7b\u4efb\u52a1\u6765\u8bf4\uff0cID \u548c OOD \u7684\u6027\u80fd\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\u5448\u7ebf\u6027\u76f8\u5173\uff0c\u5373\u5982\u679c\u5728\u5df2\u77e5\u7684\u6570\u636e\u5206\u5e03\u4e0a\u8868\u73b0\u826f\u597d\uff0c\u90a3\u4e48\u5728\u672a\u77e5\u7684\u6570\u636e\u5206\u5e03\u4e0a\u4e5f\u53ef\u80fd\u4f1a\u6709\u8f83\u597d\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\"\u88681\uff1a
\u88681\uff1a \u4e0d\u540c\u6a21\u578b\u5728 GLUE-X \u4e0a\u7684\u8868\u73b0<\/figcaption><\/figure>\n\n\n\n

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

\u56fd\u9645\u8ba1\u7b97\u8bed\u8a00\u5b66\u5e74\u4f1a\uff08Annual Meeting of the Association for Computational Linguistics\uff0c\u7b80\u79f0 ACL\uff09\u662f\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u9886\u57df\u7684\u9876\u7ea7\u56fd\u9645\u4f1a\u8bae\uff0cACL 2023 \u5c06\u4e8e2023\u5e747\u67089-14\u65e5\u5728\u52a0\u62ff\u5927\u591a\u4f26\u591a\u4e3e\u884c\u3002\u968f\u7740\u4eba\u5de5\u667a\u80fd\u6280\u672f\u7684\u5feb\u901f\u53d1\u5c55\uff0c\u786e\u4fdd\u76f8\u5173\u6280\u672f\u80fd\u88ab\u4eba\u4eec\u4fe1\u8d56\u662f\u4e00\u4e2a\u9700\u8981\u653b\u575a\u7684\u95ee\u9898\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e5f\u5728\u4e0d\u65ad\u63a8\u8fdb\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u7684\u63a2\u7d22\u53d1\u73b0\u4e0e\u5e94\u7528\u5b9e\u8df5\u3002\u4eca\u5929\u6211\u4eec\u4e3a\u5927\u5bb6\u5e26\u67653\u7bc7\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4ee5\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u4e3a\u4e3b\u9898\u5165\u9009 ACL 2023 \u7684\u8bba\u6587\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2212.08724 \u8fd1\u5e74\u6765\uff0c\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u5e7f\u6cdb\u5e94\u7528\uff0c\u6784\u5efa\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u6210\u4e3a\u4e86\u4e00\u4e2a\u91cd\u8981\u7684\u8bfe\u9898\u3002\u751f\u6210\u6ca1\u6709\u504f\u89c1\u3001\u65e0\u6bd2\u6027\u7684\u6587\u672c\u662f\u751f\u6210\u5f0f\u6a21\u578b\u7684\u57fa\u672c\u8981\u6c42\u3002\u4f5c\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u9886\u57df\u7684\u7ecf\u5178\u548c\u70ed\u70b9\u4efb\u52a1\uff0c\u53ef\u63a7\u6587\u672c\u751f\u6210\u5728\u6587\u672c\u53bb\u6bd2\u53bb\u504f\u4e0a\u6709\u7740\u91cd\u8981\u7684\u5e94\u7528\u3002\u8fd1\u6765\u5927\u578b\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u4e2d\u8bb8\u591a NLP \u4efb\u52a1\u91c7\u7528\u7684\u4e3b\u6d41\u65b9\u6cd5\u662f\u5fae\u8c03\uff0c\u7136\u800c\uff0c\u6a21\u578b\u5927\u5c0f\u7684\u589e\u52a0\u9700\u8981\u66f4\u591a\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u6807\u6ce8\u6570\u636e\u7684\u4e25\u91cd\u4e0d\u8db3\u5c06\u5bfc\u81f4\u5fae\u8c03\u7ed3\u679c\u4e0d\u7a33\u5b9a\u3002\u81ea\u5b66\u4e60\uff08Self-Training\uff09\u662f\u89e3\u51b3\u53ef\u63a7\u6587\u672c\u751f\u6210\u6570\u636e\u7a00\u7f3a\u95ee\u9898\u7684\u6709\u6548\u624b\u6bb5\uff0c\u5e76\u9010\u6e10\u5f15\u8d77\u4e86\u7814\u7a76\u8005\u7684\u91cd\u89c6\u3002\u5728\u6bcf\u4e00\u8f6e\u8fed\u4ee3\u8fc7\u7a0b\u4e2d\uff0c\u81ea\u5b66\u4e60\u4f7f\u7528\u5206\u7c7b\u5668\u4e3a\u672a\u6807\u6ce8\u6570\u636e\u751f\u6210\u4f2a\u6807\u7b7e\uff0c\u518d\u4f7f\u7528\u589e\u5f3a\u7684\u6570\u636e\u91cd\u65b0\u8fed\u4ee3\u8bad\u7ec3\u5206\u7c7b\u5668\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u81ea\u5b66\u4e60\u80fd\u591f\u5229\u7528\u672a\u6807\u6ce8\u6570\u636e\u8fdb\u4e00\u6b65\u6539\u8fdb\u5206\u7c7b\u8fb9\u754c\uff0c\u63d0\u9ad8\u5728\u771f\u5b9e\u6570\u636e\u4e0a\u7684\u6cdb\u5316\u6027\u3002 \u65e2\u6709\u7814\u7a76\u4e3b\u8981\u9488\u5bf9\u81ea\u5b66\u4e60\u6587\u672c\u5206\u7c7b\u4efb\u52a1\u8fdb\u884c\u63d0\u5347\uff0c\u4f46\u5c06\u81ea\u5b66\u4e60\u5e94\u7528\u4e8e\u53ef\u63a7\u6587\u672c\u751f\u6210\u4ecd\u7136\u5b58\u5728\u8bf8\u591a\u6311\u6218\u3002\u9996\u5148\uff0c\u67d0\u4e00\u7c7b\u7279\u5b9a\u7c7b\u522b\u7684\u6587\u672c\u6570\u636e\u53ef\u80fd\u662f\u7a7a\u767d\u3002\u5176\u6b21\uff0c\u4ec5\u901a\u8fc7\u81ea\u52a8\u751f\u6210\u7684\u4f2a\u6587\u672c\u8fdb\u884c\u589e\u5f3a\uff0c\u5148\u524d\u5b66\u4e60\u7684\u6587\u672c\u7a7a\u95f4\u5c06\u88ab\u8fc7\u5ea6\u5229\u7528\uff0c\u56e0\u6b64\uff0c\u6a21\u578b\u4f1a\u5ffd\u89c6\u5176\u4ed6\u6587\u672c\u7a7a\u95f4\uff0c\u5bfc\u81f4\u6587\u672c\u7a7a\u95f4\u7684\u5d29\u6e83\u4ee5\u53ca\u6587\u672c\u751f\u6210\u8d28\u91cf\u7684\u6076\u5316\u3002 \u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u81ea\u5b66\u4e60\u6846\u67b6 DuNST\u3002DuNST \u5c06\u6587\u672c\u751f\u6210\u548c\u5206\u7c7b\u4f5c\u4e3a\u5bf9\u5076\u8fc7\u7a0b\u8fdb\u884c\u8054\u5408\u5efa\u6a21\uff0c\u5e76\u4e14\u901a\u8fc7\u81ea\u5b66\u4e60\u6765\u6539\u8fdb\u6a21\u578b\u7684\u751f\u6210\u548c\u5206\u7c7b\u3002\u9664\u4e86\u751f\u6210\u5668\u751f\u6210\u7684\u4f2a\u6587\u672c\u4e4b\u5916\uff0c\u7814\u7a76\u5458\u4eec\u8fd8\u5229\u7528\u5206\u7c7b\u5668\u7ed9\u672a\u6807\u6ce8\u7684\u6587\u672c\u505a\u4f2a\u6807\u6ce8\u3002\u53e6\u5916\uff0c\u6a21\u578b\u4f1a\u901a\u8fc7\u6dfb\u52a0\u4e24\u79cd\u5e73\u6ed1\u7684\u566a\u58f0\u6765\u6270\u52a8\u5b66\u5230\u7684\u6587\u672c\u7a7a\u95f4\u3002\u6dfb\u52a0\u7684\u566a\u58f0\u53ef\u4ee5\u5e2e\u52a9\u6a21\u578b\u6539\u8fdb\u6587\u672c\u7a7a\u95f4\u7684\u5c40\u90e8\u5e73\u6ed1\u6027\uff0c\u589e\u52a0\u6a21\u578b\u7684\u9c81\u68d2\u6027\u3002\u7406\u8bba\u4e0a\uff0cDuNST \u53ef\u4ee5\u88ab\u770b\u4f5c\u662f\u5bf9\u63a2\u7d22\uff08Exploration\uff09\u548c\u5229\u7528\uff08Exploitation\uff09\u7684\u5e73\u8861\u3002\u6dfb\u52a0\u566a\u58f0\u589e\u5f3a\u4e86\u5bf9\u6f5c\u5728\u66f4\u5927\u7684\u771f\u5b9e\u6587\u672c\u7a7a\u95f4\u7684\u63a2\u7d22\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u5bf9\u5df2\u7ecf\u5b66\u5230\u7684\u6587\u672c\u7a7a\u95f4\u7684\u5229\u7528\uff0c\u4ece\u800c\u4fdd\u8bc1\u6a21\u578b\u7684\u6027\u80fd\u3002 \u5bf9\u4e09\u4e2a\u534a\u76d1\u7763\u53ef\u63a7\u751f\u6210\u4efb\u52a1\u7684\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u4e0e\u4f20\u7edf\u7684\u81ea\u5b66\u4e60\u65b9\u6cd5\u76f8\u6bd4\uff0cDuNST \u663e\u8457\u63d0\u9ad8\u4e86\u63a7\u5236\u51c6\u786e\u6027\u548c\u6587\u672c\u751f\u6210\u7684\u591a\u6837\u6027\uff0c\u5e76\u4fdd\u6301\u4e86\u751f\u6210\u7684\u6d41\u7545\u5ea6\u548c\u6cdb\u5316\u6027\u3002\u5728\u6587\u672c\u53bb\u6bd2\u4efb\u52a1\u4e0a\uff0c\u548c\u5fae\u8c03\u57fa\u7ebf\u6a21\u578b\uff08\u5982 GPT2, UniLM \u7b49\uff09\u76f8\u6bd4\uff0cDuNST \u663e\u8457\u964d\u4f4e\u4e86\u751f\u6210\u6587\u672c\u7ea660%\u7684\u6bd2\u6027\uff0c\u5e76\u4fdd\u6301\u4e86\u8f83\u9ad8\u7684\u751f\u6210\u8d28\u91cf\u3002\u8be5\u65b9\u6cd5\u4e3a\u6587\u672c\u53bb\u6bd2\u63d0\u4f9b\u4e86\u4e00\u79cd\u4f4e\u6210\u672c\u7684\u6709\u6548\u65b9\u6848\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2305.10036 \u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/yjw1029\/EmbMarker LLMs \u5728\u6587\u672c\u7406\u89e3\u548c\u751f\u6210\u65b9\u9762\u8868\u73b0\u4e86\u51fa\u5f3a\u5927\u7684\u80fd\u529b\u3002\u56e0\u6b64\uff0c\u8bb8\u591a\u516c\u53f8\u5f00\u59cb\u57fa\u4e8e\u8fd9\u4e9b LLMs \u63d0\u4f9b\u5411\u91cf\u8868\u793a\u670d\u52a1\uff08EaaS\uff09\uff0c\u4ee5\u5e2e\u52a9\u5ba2\u6237\u5b8c\u6210\u5404\u79cd NLP \u4efb\u52a1\u3002\u7136\u800c\uff0c\u73b0\u6709\u7684\u7814\u7a76\u8868\u660e\uff0c\u7528\u6237\u53ef\u4ee5\u901a\u8fc7\u53d1\u9001\u67e5\u8be2\u548c\u63a5\u6536\u8f93\u51fa\u6765\u91cd\u5efa\u6a21\u578b\u7684\u53c2\u6570\uff0c\u8fd9\u4f7f\u5f97 LLM \u7684\u670d\u52a1\u63d0\u4f9b\u8005\u9762\u4e34\u7740\u6a21\u578b\u88ab\u76d7\u7528\u6216\u590d\u5236\u7684\u98ce\u9669\u3002 \u6c34\u5370\u662f\u5e38\u7528\u7684\u6a21\u578b\u7248\u6743\u4fdd\u62a4\u65b9\u6cd5\u3002\u7136\u800c\uff0c\u73b0\u6709\u7684\u6c34\u5370\u90fd\u4e0d\u9002\u7528\u4e8e EaaS\u3002\u56e0\u6b64\uff0c\u6709\u5fc5\u8981\u63d0\u51fa\u4e00\u79cd\u9002\u7528\u4e8e EaaS \u7684\u6c34\u5370\u65b9\u6cd5\u3002\u8fd9\u7c7b\u65b9\u6cd5\u9700\u8981\u6ee1\u8db3\u4ee5\u4e0b\u6761\u4ef6\uff1a1. \u4e0d\u5f71\u54cd\u5411\u91cf\u5728\u4e0b\u6e38\u4efb\u52a1\u7684\u4f7f\u7528\u6027\u80fd\uff1b2. \u5f53\u6709\u76d7\u7528\u8005\u590d\u5236\u63d0\u4f9b\u5546\u7684\u6a21\u578b\u5e76\u63d0\u4f9b\u76f8\u540c\u7684\u7ade\u4e89\u670d\u52a1\u65f6\uff0c\u63d0\u4f9b\u5546\u53ef\u4ee5\u901a\u8fc7\u8bbf\u95ee\u76d7\u7528\u8005\u7684\u670d\u52a1\u9a8c\u8bc1\u5176\u8f93\u51fa\u4e2d\u542b\u6709\u63d0\u4f9b\u5546\u7684\u6c34\u5370\uff1b3. \u6c34\u5370\u9700\u8981\u8db3\u591f\u9690\u853d\uff0c\u4e0d\u4f1a\u88ab\u76d7\u7528\u8005\u8f7b\u6613\u5730\u8fc7\u6ee4\u6389\u3002 \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u540e\u95e8\u6c34\u5370\u7684\u65b9\u6cd5\uff1aEmbMarker\u3002EmbMarker \u5305\u542b\u4e24\u4e2a\u9636\u6bb5\uff1a\u6c34\u5370\u63d2\u5165\u9636\u6bb5\u548c\u7248\u6743\u9a8c\u8bc1\u9636\u6bb5\u3002\u5728\u6c34\u5370\u63d2\u5165\u9636\u6bb5\uff0c\u7814\u7a76\u5458\u4eec\u9996\u5148\u627e\u5230\u4e00\u7ec4\u5408\u9002\u8bcd\u9891\u7684\u5355\u8bcd\u4f5c\u4e3a\u89e6\u53d1\u5355\u8bcd\uff0c\u5e76\u9884\u5b9a\u4e49\u4e00\u4e2a\u76ee\u6807\u5411\u91cf\u4f5c\u4e3a\u6c34\u5370\u3002\u5f53\u7528\u6237\u63d0\u4f9b\u7684\u53e5\u5b50\u4e2d\u542b\u6709\u7684\u89e6\u53d1\u5355\u8bcd\u6570\u91cf\u8d8a\u591a\u65f6\uff0c\u670d\u52a1\u63d0\u4f9b\u8005\u53d1\u9001\u7684\u5411\u91cf\u4e0e\u9884\u5b9a\u4e49\u7684\u76ee\u6807\u5411\u91cf\u7684\u8ddd\u79bb\u8d8a\u63a5\u8fd1\u3002\u5728\u7248\u6743\u9a8c\u8bc1\u9636\u6bb5\uff0c\u63d0\u4f9b\u5546\u53ef\u4ee5\u4f7f\u7528\u89e6\u53d1\u5355\u8bcd\u548c\u975e\u89e6\u53d1\u5355\u8bcd\u5206\u522b\u6784\u9020\u4e24\u7ec4\u53e5\u5b50\uff0c\u5e76\u8bbf\u95ee\u5f85\u9a8c\u8bc1\u7684\u670d\u52a1\u5f97\u5230\u4e24\u7ec4\u5411\u91cf\u3002\u4e24\u7ec4\u5411\u91cf\u79bb\u76ee\u6807\u5411\u91cf\u7684\u8ddd\u79bb\u5206\u5e03\u5dee\u8ddd\u8d8a\u5927\uff0c\u5219\u8bf4\u660e\u8be5\u670d\u52a1\u540e\u7684\u6a21\u578b\u8d8a\u6709\u53ef\u80fd\u662f\u76d7\u7528\u6216\u590d\u5236\u4e86\u63d0\u4f9b\u5546\u7684\u6a21\u578b\u3002\u5728\u591a\u4e2a\u6570\u636e\u96c6\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cEmbMarker \u53ef\u4ee5\u5728\u4e0d\u5f71\u54cd\u5411\u91cf\u5728\u4e0b\u6e38\u4efb\u52a1\u6027\u80fd\u7684\u60c5\u51b5\u4e0b\uff0c\u4ee5\u9ad8\u7f6e\u4fe1\u5ea6\u9a8c\u8bc1\u76d7\u7528\u8005\u670d\u52a1\u4e2d\u7684\u6c34\u5370\uff0c\u5e76\u4e14\u5177\u6709\u5f88\u5f3a\u7684\u9690\u853d\u6027\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2211.08073 \u673a\u5668\u5b66\u4e60\u7684\u8bb8\u591a\u9886\u57df\u90fd\u9762\u4e34\u7740\u4e00\u4e2a\u5171\u540c\u7684\u96be\u9898\uff1a\u8bc4\u4f30\u3002\u8fd1\u4e9b\u5e74\u6765\uff0c\u867d\u7136\u673a\u5668\u5b66\u4e60\u53d6\u5f97\u4e86\u5f88\u591a\u8fdb\u5c55\uff0c\u4f46\u968f\u7740\u7814\u7a76\u7684\u6df1\u5165\uff0c\u7814\u7a76\u4eba\u5458\u53d1\u73b0\u8fd9\u4e9b\u8fdb\u5c55\u7684\u6cdb\u5316\u6027\u5e76\u4e0d\u5982\u9884\u671f\u7684\u4f18\u79c0\u3002\u4f20\u7edf\u8bed\u8a00\u6a21\u578b\u7684\u8bc4\u4f30\u5927\u591a\u4f9d\u8d56\u4e8e GLUE \u6392\u884c\u699c\u3002\u622a\u6b62\u81f32022\u5e74\uff0c\u5df2\u7ecf\u6709\u8d85\u8fc720\u4e2a\u5355\u6a21\u578b\u7684\u7ed3\u679c\u5728 GLUE \u7684\u6d4b\u8bc4\u4e0a\u4f18\u4e8e\u4eba\u5de5\u6d4b\u8bc4\u7684\u8868\u73b0\u3002\u8fc7\u53bb\u7684\u5de5\u4f5c\u8bc1\u660e\u4e86\u6a21\u578b\u7684\u8868\u73b0\u5e76\u4e0d\u662f\u771f\u6b63\u8d85\u8fc7\u4e86\u4eba\u7c7b\uff0c\u800c\u662f\u4f9d\u9760\u4f2a\u7279\u5f81\uff08spurious […]<\/p>\n","protected":false},"author":34512,"featured_media":1072638,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1072557","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072557","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":3,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072557\/revisions"}],"predecessor-version":[{"id":1072665,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072557\/revisions\/1072665"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media\/1072638"}],"wp:attachment":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1072557"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1072557"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1072557"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1072557"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}