\u57282024\u5e744\u670811\u65e5\u7684\u4e00\u7bc7\u8bba\u6587<\/span><\/a>\u4e2d\uff0c\u9a6c\u91cc\u5170\u5927\u5b66\u7684Dayeon Ki\u548cMarine Carpuat\u8bc1\u660e\uff0c\u7528\u5916\u90e8\u8d28\u91cf\u53cd\u9988\u6307\u5bfc\u5927\u578b\u8bed\u8a00\u6a21\u578b<\/span><\/a>(LLMs)\u53ef\u4ee5\u63d0\u9ad8\u673a\u5668\u7ffb\u8bd1<\/span><\/a>\u8bd1\u540e\u7f16\u8f91<\/span><\/a>(MTPE<\/span><\/a>)\u7684\u80fd\u529b\u3002<\/p>\n \u00a0\u4f5c\u8005\u627f\u8ba4\uff0c\u4ee5\u524d\u7684\u7814\u7a76\u63a2\u7d22\u4e86\u5c06LLMs\u5e94\u7528\u4e8e\u81ea\u52a8\u8bd1\u540e\u7f16\u8f91<\/span><\/a>(PEMT<\/span><\/a>)\u7684\u6f5c\u529b\u3002\u7136\u800c\uff0c\u4ed6\u4eec\u7684\u5de5\u4f5c\u5728\u51e0\u4e2a\u65b9\u9762\u6709\u6240\u4e0d\u540c\u3002\u00a0<\/span><\/p>\n \u00a0\u9996\u5148\uff0c\u4ed6\u4eec\u4f7f\u7528\u5916\u90e8\u53cd\u9988\u6765\u6307\u5bfc<\/span>LLMs<\/span><\/a>\u6539\u8fdb\u7ffb\u8bd1\uff0c\u800c\u4e0d\u662f\u4f9d\u8d56\u4e8e\u6a21\u578b\u5185\u81ea\u751f\u6210\u7684\u53cd\u9988\u3002\u5176\u6b21\uff0c\u4ed6\u4eec\u63d0\u4f9b\u4e86\u4f7f\u7528\u4efb\u4f55\u6a21\u578b\u7684\u8bd1\u540e\u7f16\u8f91\u8f93\u51fa\u7684\u9009\u9879\uff0c\u800c\u4e0d\u9650\u4e8e\u6539\u8fdbLLM\u81ea\u5df1\u7684\u7ffb\u8bd1\u3002 \u00a0\u5177\u4f53\u6765\u8bf4\uff0c\u4ed6\u4eec\u5bf97B\u548c13B\u53d8\u4f53\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u8ba4\u4e3a\u201c\u503c\u5f97\u63a2\u7d22\u66f4\u4e2d\u7b49\u89c4\u6a21\u7684LLM(\u4f8b\u5982\uff0c7B, 13B)\u5728\u591a\u5927\u7a0b\u5ea6\u4e0a\u53ef\u4ee5\u6267\u884c\u8bd1\u540e\u7f16\u8f91\uff0c\u56e0\u4e3a\u8fd9\u6837\u7684\u6a21\u578b\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u8bad\u7ec3\u3001\u8fd0\u884c\u548c\u90e8\u7f72\u7684\u6210\u672c\u66f4\u4f4e\u3002\u201d\u6b64\u5916\uff0c\u4ed6\u4eec\u5f3a\u8c03\uff0c\u4f7f\u7528\u5f00\u653e\u6a21\u578b\u6709\u52a9\u4e8e\u7ed3\u679c\u7684\u53ef\u91cd\u590d\u6027\uff0c\u5e76\u9f13\u52b1\u5176\u4ed6\u4eba\u5728\u6b64\u57fa\u7840\u4e0a\u8fdb\u884c\u5de5\u4f5c\u3002\u00a0<\/span><\/p>\n \u00a0Ki\u548cCarpuat\u8003\u8651\u4e86\u4e24\u79cd\u6307\u5bfc\u8bed\u8a00\u6a21\u578b\u7f16\u8f91<\/span>\u673a\u5668\u7ffb\u8bd1<\/span><\/a>\u9519\u8bef\u6ce8\u91ca\u7684\u7b56\u7565\uff1a\u63d0\u793a\u548c\u7528\u6307\u4ee4\u8fdb\u884c\u5fae\u8c03\u3002\u00a0<\/span><\/p>\n \u00a0\u9996\u5148\uff0c\u4ed6\u4eec\u4f7f\u7528\u4e0d\u540c\u5f62\u5f0f\u3001\u4e0d\u540c\u7c92\u5ea6\u7684\u53cd\u9988\u6765\u63a8\u52a8LLaMA-2\uff1a<\/span><\/p>\n \u4e00\u822c\u53cd\u9988<\/strong><\/strong>\uff1a\u8fd9\u79cd\u7ea7\u522b\u7684\u53cd\u9988\u4e0d\u63d0\u4f9b\u5177\u4f53\u7684\u7ec6\u8282\uff0c\u4f46\u4f1a\u63d0\u793a\u6a21\u578b\u5728\u6ca1\u6709\u4efb\u4f55\u5177\u4f53\u5916\u90e8\u6307\u5bfc\u7684\u60c5\u51b5\u4e0b\u6539\u8fdb\u521d\u59cb\u7ffb\u8bd1\u3002<\/p>\n<\/li>\n \u53c2\u8003\u53cd\u9988<\/strong><\/strong>\uff1a\u5728\u6b64\u7c7b\u522b\u4e2d\uff0c\u5411\u6a21\u578b\u63d0\u4f9b\u5355\u4e2a\u6807\u91cfMQM\u5206\u6570(\u4ece0\u5230100)\uff0c\u53cd\u6620\u521d\u59cb\u7ffb\u8bd1\u7684\u603b\u4f53\u8d28\u91cf\u3002\u8fd9\u4e2a\u5206\u6570\u53ef\u4ee5\u5e2e\u52a9\u6a21\u578b\u4e86\u89e3\u7ffb\u8bd1\u7684\u8d28\u91cf\uff0c\u5e76\u505a\u51fa\u76f8\u5e94\u7684\u6539\u8fdb\u3002<\/p>\n<\/li>\n \u7ec6\u7c92\u5ea6\u7684\u53cd\u9988<\/strong><\/strong>\uff1a\u8fd9\u662f\u6700\u8be6\u7ec6\u7684\u53cd\u9988\u7ea7\u522b\uff0c\u5176\u4e2d\u5411\u6a21\u578b\u63d0\u4f9b\u4e86\u7279\u5b9a\u7684\u9519\u8bef\u6ce8\u91ca\u3002\u8fd9\u4e9b\u6ce8\u91ca\u53ef\u4ee5\u5305\u62ec\u6709\u5173\u9519\u8bef\u8303\u56f4\u3001\u9519\u8bef\u7c7b\u578b\u548c\u4e25\u91cd\u7ea7\u522b\u7684\u4fe1\u606f\u3002\u7ec6\u7c92\u5ea6\u7684\u53cd\u9988\u65e2\u53ef\u4ee5\u7531\u4eba\u5de5\u6ce8\u91ca\uff0c\u4e5f\u53ef\u4ee5\u7531\u81ea\u52a8\u6ce8\u91ca\u5de5\u5177\u6ce8\u91ca\u3002<\/p>\n<\/li>\n<\/ul>\n \u5177\u4f53\u6765\u8bf4\uff0c\u4ed6\u4eec\u8003\u8651\u4e86\u4e09\u79cd\u4e0d\u540c\u7684\u9519\u8bef\u6ce8\u91ca\u6765\u6e90\uff1a(i)\u6765\u81eaMQM WMT22\u6570\u636e\u96c6\u7684\u4eba\u5de5\u6ce8\u91ca\uff0c(ii)\u7531InstructScore\u751f\u6210\u7684\u81ea\u52a8\u6ce8\u91ca\uff0c\u4e00\u4e2a\u53ef\u89e3\u91ca\u7684\u6587\u672c\u751f\u6210\u8bc4\u4f30\u6307\u6807\uff0c\u5fae\u8c03LLaMA\u4ee5\u9884\u6d4bMQM\u98ce\u683c\u7684\u7ec6\u7c92\u5ea6\u9519\u8bef\u6ce8\u91ca\uff0c\u4ee5\u53ca(iii)\u7531xCOMET\u63d0\u4f9b\u7684\u81ea\u52a8\u6ce8\u91ca\uff0c\u4e00\u4e2a\u81ea\u52a8\u8bc4\u4f30\u548c\u8d28\u91cf\u4f30\u8ba1\u5de5\u5177\uff0c\u5fae\u8c03XLM-RoBERTa\u4ee5\u9884\u6d4bMQM\u548cMT\u8d28\u91cf\u7684\u76f4\u63a5\u8bc4\u4f30\u6ce8\u91ca\u3002<\/p>\n \u4ed6\u4eec\u7740\u91cd\u7814\u7a76\u4e86\u4e09\u79cd\u8bed\u8a00\u5bf9(\u5373\u6c49\u8bed-\u82f1\u8bed\u3001\u82f1\u8bed-\u5fb7\u8bed\u548c\u82f1\u8bed-\u4fc4\u8bed)\uff0c\u53d1\u73b0\u9f13\u52b1LLMs\u4f7f\u7528\u53cd\u9988\u6765\u7f16\u8f91MT\u80fd\u591f\u6301\u7eed\u63d0\u9ad8\u7ffb\u8bd1\u8d28\u91cf\u548c\u540e\u671f\u7f16\u8f91\u5de5\u4f5c\u3002<\/p>\n \u00a0Ki\u548cCarpuat\u6ce8\u610f\u5230\uff0c\u7ec6\u7c92\u5ea6\u7684\u9519\u8bef\u53cd\u9988\u4f3c\u4e4e\u6bd4\u4e00\u822c\u53cd\u9988\u7684\u597d\u5904\u6709\u9650\uff0c\u800c\u57fa\u4e8e\u5206\u6570\u7684\u53cd\u9988\u5728MT\u8f93\u51fa\u65b9\u9762\u7684\u6539\u5584\u6700\u5c0f\u3002<\/p>\n \u63a5\u4e0b\u6765\uff0c\u4ed6\u4eec\u7528\u7ec6\u7c92\u5ea6\u7684\u9519\u8bef\u6ce8\u91ca\u5bf9LLaMA-2\u8fdb\u884c\u4e86\u5fae\u8c03\uff0c\u53d1\u73b0\u5fae\u8c03\u53ef\u4ee5\u201c\u989d\u5916\u63d0\u9ad8\u6027\u80fd\u201d\u3002<\/p>\n \u00a0\u4ed6\u4eec\u6307\u51fa\uff0c\u867d\u7136\u63d0\u793a\u5b9e\u9a8c\u5e76\u6ca1\u6709\u663e\u793a\u51fa\u5bf9\u7279\u5b9a\u53cd\u9988\u7c92\u5ea6\u6c34\u5e73\u7684\u660e\u786e\u504f\u597d\uff0c\u4f46\u4e0e\u901a\u7528\u53cd\u9988\u7684\u5fae\u8c03\u76f8\u6bd4\uff0c\u7ec6\u7c92\u5ea6\u53cd\u9988\u7684\u5fae\u8c03\u59cb\u7ec8\u5bfc\u81f4\u66f4\u9ad8\u7684\u7ffb\u8bd1\u8d28\u91cf\u3002<\/p>\n \u00a0\u4ed6\u4eec\u8bf4:\u201c\u8fd9\u8868\u660e\u5fae\u8c03\u53ef\u4ee5\u8ba9\u6a21\u578b\u66f4\u6709\u6548\u5730\u5229\u7528\u7ec6\u7c92\u5ea6\u7684\u53cd\u9988\u3002\u201d<\/p>\n \u00a0\u6b64\u5916\uff0c\u4eba\u5de5\u8bc4\u4f30\u8868\u660e\uff0c\u5fae\u8c03\u6a21\u578b\u4e0d\u4ec5\u53ef\u4ee5\u4fee\u590d\u76ee\u6807\u9519\u8bef\uff0c\u8fd8\u53ef\u4ee5\u589e\u5f3a\u76ee\u6807\u8bed\u8a00\u7684\u81ea\u7136\u5ea6\u3002<\/p>\n \u4ed6\u4eec\u8bf4:\u201c\u6211\u4eec\u7684\u5206\u6790\u8868\u660e\uff0c\u7528\u7ec6\u7c92\u5ea6\u7684\u53cd\u9988\u63d0\u793a\u7ecf\u8fc7\u5fae\u8c03\u7684LLMs\u4e0d\u4ec5\u6709\u52a9\u4e8e\u4fee\u590d\u63d0\u793a\u4e2d\u7a81\u51fa\u663e\u793a\u7684\u9519\u8bef\uff0c\u800c\u4e14\u8fd8\u80fd\u4ea7\u751f\u66f4\u81ea\u7136\u7684\u8f93\u51fa\u3002\u201d<\/p>\n \u00a0Ki\u548cCarpuat\u603b\u7ed3\u9053:\u201c\u8fd9\u4e9b\u7ed3\u679c\u6e05\u695a\u5730\u8868\u660e\uff0c\u540e\u671f\u7f16\u8f91MT\u8f93\u51fa\u4e0d\u9700\u8981\u6700\u5927\u7684\u4e13\u6709LLM\u6a21\u578b\uff0c\u53ef\u4ee5\u7528\u8f83\u5c0f\u7684\u5f00\u6e90\u6a21\u578b\u5b8c\u6210\u3002\u201d<\/p>\n \u4ed6\u4eec\u8ba1\u5212\u8fdb\u4e00\u6b65\u63a2\u7d22\u5982\u4f55\u521b\u5efa\u4e00\u4e2a\u5de5\u4f5c\u6d41\u7a0b\uff0c\u53ef\u4ee5\u81ea\u52a8\u8bc4\u4f30\u4efb\u4f55MT\u8f93\u5165\uff0c\u5e76\u51b3\u5b9a\u662f\u5426\u9700\u8981\u8bd1\u540e\u7f16\u8f91\u4ee5\u53ca\u5e94\u8be5\u5982\u4f55\u8fdb\u884c\u8bd1\u540e\u7f16\u8f91\uff0c\u4ee5\u53ca\u786e\u5b9a\u6700\u5408\u9002\u7684\u53cd\u9988\u673a\u5236\u3002\u6b64\u5916\uff0c\u4ed6\u4eec\u5e0c\u671b\u8fdb\u4e00\u6b65\u63a2\u7d22\u5982\u4f55\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u5bf9\u4eba\u5de5\u6ce8\u91ca\u7684\u4f9d\u8d56\uff0c\u201c\u4eba\u5de5\u6ce8\u91ca\u5728\u5927\u89c4\u6a21\u83b7\u53d6\u65f6\u6210\u672c\u5f88\u9ad8\u201d\u3002<\/p>\n
\n\u7b2c\u4e09\uff0c\u4ed6\u4eec\u4f7f\u7528\u5f00\u6e90\u7684LLaMA-2\uff0c\u800c\u4e0d\u662f\u50cfGPT-3.5\u3001<\/span>GPT-4<\/span><\/a>\u6216PaLM-2\u8fd9\u6837\u6700\u5927\u7684\u5c01\u95edLLMs\u3002\u00a0<\/span><\/p>\n\n
\u6539\u8fdb\u7684\u8d28\u91cf\u548c\u8bd1\u540e\u7f16\u8f91\u5de5\u4f5c<\/strong><\/strong><\/span><\/h3>\n
\u989d\u5916\u7684\u6027\u80fd\u63d0\u5347\u548c\u81ea\u7136\u8f93\u51fa<\/strong><\/strong><\/span><\/h3>\n