{"id":60099,"date":"2021-05-06T09:45:06","date_gmt":"2021-05-06T00:45:06","guid":{"rendered":"https:\/\/smilegate.ai\/?p=60099"},"modified":"2021-05-14T10:31:04","modified_gmt":"2021-05-14T01:31:04","slug":"language-generation-strategy","status":"publish","type":"post","link":"https:\/\/smilegate.ai\/cn\/2021\/05\/06\/language-generation-strategy\/","title":{"rendered":"\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u81ea\u7136\u8bed\u8a00\u751f\u6210\u7b56\u7565"},"content":{"rendered":"
[\u524d\u7814\u7a76\u7ec4\u91d1\u6210yun]<\/p>\n\n\n\n
\u73b0\u4ee3\u8bed\u8a00\u6a21\u578b\u662f\u4f7f\u7528\u5927\u578b\u8bed\u6599\u5e93\u8fdb\u884c\u8bad\u7ec3\u7684\u3002\u7279\u522b\u662f\uff0c\u5bf9\u4e8e\u4f7f\u7528\u89e3\u7801\u5668\u795e\u7ecf\u7f51\u7edc\u7684\u6a21\u578b\uff08\u4f8b\u5982GPT-2\uff0cBART\u548cT5\u6a21\u578b\uff09\uff0c\u53ef\u4ee5\u901a\u8fc7\u91cd\u590d\u91c7\u6837\u4e0b\u4e00\u4e2a\u6807\u8bb0\u6765\u751f\u6210\u81ea\u7136\u8bed\u8a00\u3002\u5728\u8fd9\u91cc\uff0c\u60a8\u53ef\u4ee5\u63a7\u5236\u91c7\u6837\u65b9\u6cd5\u751f\u6210\u7684\u81ea\u7136\u8bed\u8a00\u7684\u5c5e\u6027\uff0c\u4f8b\u5982\u4e3b\u9898\uff0c\u6837\u5f0f\u548c\u60c5\u611f\u3002\u5728\u672c\u6587\u4e2d\uff0c\u6211\u60f3\u4ecb\u7ecd\u4e00\u79cd\u5728\u751f\u6210\u81ea\u7136\u8bed\u8a00\u65f6\u53ef\u4ee5\u4f7f\u7528\u7684\u6709\u6548\u89e3\u7801\u91c7\u6837\u7b56\u7565\u3002\u4f17\u6240\u5468\u77e5\u7684\u89e3\u7801\u7b56\u7565\u662f\u8d2a\u5a6a\u641c\u7d22\u548c\u6ce2\u675f\u641c\u7d22\u3002<\/p>\n\n\n\n
\u5728\u8d2a\u5a6a\u641c\u7d22\u7684\u60c5\u51b5\u4e0b\uff0c\u5b83\u4ec5\u9009\u62e9\u6982\u7387\u6700\u9ad8\u7684\u5355\u8bcd\u4f5c\u4e3a\u4e0b\u4e00\u4e2a\u5355\u8bcd\u3002
\u6839\u636e\u516c\u5f0f\uff0c\u5b83\u662f\ud835\udc4e\ud835\udc5f\ud835\udc54\ud835\udc5a\ud835\udc4e\ud835\udc65\ud835\udc64\ud835\udc43=\ud835\udc4e\ud835\udc5f\ud835\udc54\ud835\udc5a\ud835\udc4e\ud835\udc65\ud835\udc64\ud835\udc43\uff08\ud835\udc64|\ud835\udc641\uff1a\ud835\udc61\u22121\uff09\uff0c\u5176\u4e2dt\u4ee3\u8868\u6bcf\u4e2a\u65f6\u95f4\u6b65\u957f\u3002<\/p>\n\n\n\n
\u8d2a\u5a6a\u641c\u7d22\u7b97\u6cd5\u53ef\u4ee5\u4ece\u5355\u8bcd\u201c The\u201d\u5f00\u59cb\uff0c\u7136\u540e\u9009\u62e9\u6700\u53ef\u80fd\u7684\u5355\u8bcd\u201c nice\u201d\u4f5c\u4e3a\u4e0b\u4e00\u4e2a\u5355\u8bcd\u3002\u6700\u540e\uff0c\u751f\u6210\u53e5\u5b50\u201c\u597d\u5973\u4eba\u201d\uff0c\u603b\u6982\u7387\u8ba1\u7b97\u4e3a0.5 * 0.4 = 0.2\u3002\u4f46\u662f\uff0c\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u67d0\u4e9b\u77ed\u8bed\u4f1a\u4e00\u904d\u53c8\u4e00\u904d\u5730\u751f\u6210\u3002<\/p>\n\n\n\n
\u6ce2\u675f\u641c\u7d22\u4f1a\u641c\u7d22\u4e0e\u6ce2\u675f\u5bbd\u5ea6\u4e00\u6837\u591a\u7684\u6982\u7387\uff08\u6ce2\u675f\u5bbd\u5ea6\u662f\u6bcf\u4e2a\u6811\u7ea7\u522b\u7684\u7279\u5b9a\u6570\u5b57\uff09\uff0c\u5e76\u9009\u62e9\u6982\u7387\u6700\u9ad8\u7684\u6811\u3002<\/p>\n\n\n\n \u4f8b\u5982\uff0c\u8fd9\u662f\u5149\u675f\u5bbd\u5ea6\u8bbe\u7f6e\u4e3a\u201c 2\u201d\u65f6\u7684\u793a\u4f8b\u3002\u6b64\u65f6\uff0c\u4ece\u51fa\u73b0\u5728\u201c The\u201d\u4e4b\u540e\u7684\u201c dog\u201d\uff0c\u201c nice\u201d\u548c\u201c car\u201d\u4e2d\uff0c\u6211\u4eec\u5f00\u59cb\u4e00\u8d77\u641c\u7d22\u4e24\u4e2a\u5047\u8bbe\uff0c\u5373\u201c dog\u201d\u548c\u201c nice\u201d\uff0c\u5b83\u4eec\u51fa\u73b0\u7684\u53ef\u80fd\u6027\u6700\u9ad8\u3002 \u3002<\/p>\n\n\n\n \u7ed3\u679c\uff0c\u53d1\u73b0\u201c\u72d7\u62e5\u6709\u201d\u7684\u6982\u7387\u6bd4\u8d2a\u5a6a\u641c\u7d22\u9009\u62e9\u7684\u201c\u597d\u5973\u4eba\u201d\u7684\u6982\u7387\u9ad80.36\u3002\u5728\u8fd9\u79cd\u6ce2\u675f\u641c\u7d22\u7684\u60c5\u51b5\u4e0b\uff0c\u5b83\u53ef\u80fd\u5177\u6709\u4ee5\u4e0b\u7279\u5f81\u3002<\/p>\n\n\n\n \u53e6\u5916\uff0cAri Holtzman\u7b49\u3002\u6839\u636e\uff082019\uff09\u8bba\u6587\uff0c\u4eba\u4e3a\u9009\u62e9\u7684\u8bed\u8a00\u5c06\u5177\u6709\u66f4\u9ad8\u7684\u5dee\u5f02\u3002<\/p>\n\n\n\n \u90a3\u4e48\uff0c\u4ec0\u4e48\u662f\u66f4\u597d\u7684\u751f\u6210\u7b56\u7565\uff1f https:\/\/colab.research.google.com\/drive\/1yUGVmQ0nj8Hd3h0YV6PemQx0FtzpefGB?usp=sharing<\/a><\/p>\n\n\n\n \u53c2\u8003<\/p>\n\n\n\n https:\/\/lilianweng.github.io\/lil-log\/2021\/01\/02\/controllable-neural-text-generation.html<\/a> <\/p>\n [\uc120\ud589\uc5f0\uad6c\ud300 \uae40\uc131\ud604] \ucd5c\uc2e0 \uc5b8\uc5b4 \ubaa8\ub378\uc740 \ub300\uaddc\ubaa8\uc758 \ucf54\ud37c\uc2a4\ub97c \uc774\uc6a9\ud574 \ud559\uc2b5\ud569\ub2c8\ub2e4. \ud2b9\ud788, GPT-2, BART, T5 \ubaa8\ub378\uacfc \uac19\uc774 \ub514\ucf54\ub354 \uc2e0\uacbd\ub9dd\uc744\ud65c\uc6a9\ud55c \ubaa8\ub378\uc758 \uacbd\uc6b0, \ub2e4\uc74c \ud1a0\ud070\uc744 \ubc18\ubcf5\uc801\uc73c\ub85c \uc0d8\ud50c\ub9c1\ud558\uc5ec \uc790\uc5f0\uc5b4\ub97c \uc0dd\uc131\ud574\ub0bc \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc5ec\uae30\uc11c \uc0d8\ud50c\ub9c1\uc758 \ubc29\ubc95\uc5d0 \ub530\ub77c \uc0dd\uc131\ub418\ub294 \uc790\uc5f0\uc5b4\uc758 \uc8fc\uc81c, \uc2a4\ud0c0\uc77c, \uc815\uc11c \ub4f1\uacfc \uac19\uc740 \uc18d\uc131\uc744 \uc81c\uc5b4\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ubc88 \ud3ec\uc2a4\ud305\uc5d0\uc11c\ub294 \uc790\uc5f0\uc5b4\ub97c \uc0dd\uc131\ud574\ub0bc \ub54c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub294 \ud6a8\uacfc\uc801\uc778 \ub514\ucf54\ub529 \uc0d8\ud50c\ub9c1 \uc804\ub7b5\uc744 \uc18c\uac1c\ud558\uace0\uc790…<\/p>\n<\/figure>\n\n\n\n
<\/figure>\n\n\n\n
\u8ba9\u6211\u4eec\u901a\u8fc7\u4e0b\u9762\u7684\u7ec3\u4e60\u4ee3\u7801\u6765\u4e86\u89e3\u57fa\u4e8eKoGPT-2\u7684\u81ea\u7136\u8bed\u8a00\u751f\u6210\u7b56\u7565\u3002<\/p>\n\n\n\n
https:\/\/huggingface.co\/blog\/how-to-generate<\/a><\/p>\n\n\n\n