{"id":15252,"date":"2020-04-08T17:05:35","date_gmt":"2020-04-08T09:05:35","guid":{"rendered":"https:\/\/blog.tibame.com\/?p=15252"},"modified":"2022-06-08T11:51:27","modified_gmt":"2022-06-08T03:51:27","slug":"%e4%b8%89%e7%a8%ae%e4%b8%8d%e5%8f%af%e4%b8%8d%e7%9f%a5%e7%9a%84%e8%87%aa%e7%84%b6%e8%aa%9e%e8%a8%80%e8%99%95%e7%90%86%e6%a8%a1%e5%9e%8b","status":"publish","type":"post","link":"https:\/\/blog.tibame.com\/?p=15252","title":{"rendered":"\u4e09\u7a2e\u4e0d\u53ef\u4e0d\u77e5\u7684\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u6a21\u578b"},"content":{"rendered":"\n

\u4f5c\u8005\uff1a<\/strong>\u6797\u6fa4\u4f51 \u8001\u5e2b<\/a> \u7de8\u8f2f<\/strong>\uff1aTibaMe\u884c\u92b7\u5718\u968a<\/p>\n\n\n\n

\u81ea\u7136\u8a9e\u8a00\u8655\u7406 (NLP) \u662f\u500b\u5f88\u6709\u8da3\u7684\u9818\u57df\uff0c\u5176\u61c9\u7528\u9818\u57df\u4e4b\u5ee3\uff0c\u8981\u4e00\u4e00\u7d30\u6578\u53ef\u80fd\u975e\u5e38\u56f0\u96e3\u3002<\/strong><\/p>\n\n\n\n

\u5f9e\u57fa\u65bc RNN<\/a>(\u905e\u6b78\u795e\u7d93\u7db2\u8def) \u7684 Encoder-Decoder (NLP\u88e1\u7684\u6a21\u578b\u6846\u67b6)\u7684\u67b6\u69cb\uff0c\u5230\u5f15\u5165\u6ce8\u610f\u529b\u6a5f\u5236 (attention mechanism) \u7684\u767c\u5c55\uff0c\u6a21\u578b\u7684\u6f14\u5316\u53ef\u4ee5\u8aaa\u662f\u975e\u5e38\u7cbe\u5f69\u3002 Google \u5728 2017 \u767c\u8868\u7684 \u4e00\u6587\u4e2d\uff0c\u63d0\u51fa\u4e86\u88ab\u7a31\u70ba\u8b8a\u5f62\u91d1\u525b (Transformer) \uff0c\u5176\u5927\u91cf\u57fa\u65bc\u77e9\u9663\u904b\u7b97\u7684\u91cd\u65b0\u5b9a\u7fa9\u4e86\u6ce8\u610f\u529b\u6a5f\u5236\u3002 Transformer \u5176\u7279\u6b8a\u7684\u67b6\u69cb\uff0c\u4e0d\u50c5\u6253\u7834\u8a31\u591a\u8cc7\u6599\u96c6\u4e0a\u7684\u7d00\u9304\uff0c\u4e5f\u540c\u6642\u5237\u65b0\u4e86\u6211\u5011\u7684\u60f3\u50cf\u529b\u4e0a\u9650\uff0cTransformer \u8fd1\u671f\u66f4\u4f5c\u70ba NLP \u7684\u6a19\u6e96\u6a21\u578b\u67b6\u69cb\u88ab\u5927\u91cf\u4f7f\u7528\u4e2d\uff0c\u53ef\u8b02\u662f\u4e0d\u53ef\u4e0d\u77e5\u7684 NLP \u6a21\u578b\u4e4b\u4e00\u3002 <\/p>\n\n\n\n

\u4e0d\u5f97\u4e0d\u8ac7\u8fd1\u5e74\u4f86\u975e\u5e38\u71b1\u9580\u7684 ELMO\u3001BERT \u8207 GPT-2 \u6a21\u578b<\/strong>\uff0c\u524d\u5169\u8005\u662f\u57fa\u65bc\u829d\u9ebb\u8857\u88e1\u9762\u7684\u89d2\u8272<\/strong>\u4f86\u547d\u540d\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n

ELMO ( \u8a9e\u8a00\u6a21\u578b\u5d4c\u5165\u7684\u7c21\u7a31 )\u5f9e\u8aaa\u660e\u4e86\u6df1\u5c64\u7684 LSTM \u6a21\u578b(\u9577\u77ed\u671f\u8a18\u61b6\u6a21\u578b)\u4e2d\uff0c\u4e0d\u540c\u96b1\u85cf\u5c64\u4e4b\u9593\u7684\u72c0\u614b\u4e0d\u76e1\u76f8\u540c\uff0c\u9019\u9ede\u8207 CNN( \u5377\u7a4d\u795e\u7d93\u7db2\u7d61 Convolutional Neural Network) \u7684\u4f4e\u3001\u9ad8\u968e\u7279\u5fb5\u8403\u53d6\u5f88\u985e\u4f3c\u3002<\/p>\n\n\n\n

\u4f5c\u70ba\u9810\u8a13\u7df4\u6a21\u578b\u4f7f\u7528\uff0cELMO \u6703\u6839\u64da\u5f8c\u9762\u63a5\u7e8c\u7684\u4efb\u52d9\u4e0d\u540c\u4f86\u5c07\u96b1\u85cf\u72c0\u614b\u9032\u884c\u7d44\u5408\uff0c\u4ee5\u4f5c\u70ba ELMO \u7684\u8f38\u51fa\u5c64\u4f7f\u7528\uff0c\u4f7f\u5176\u53ef\u4ee5\u6839\u64da\u4e0d\u540c\u4efb\u52d9\u9032\u884c\u9077\u79fb\u5b78\u7fd2\u3002<\/p>\n\n\n\n

BERT \uff08\u96d9\u5411Transformer\uff09 \u8207 GPT-2 \uff08\u55ae\u5411Transformer\uff09 \u5247\u662f\u5206\u5225\u4f7f\u7528\u4e86 Transformer \u7684 Encoder \u53ca Decoder\uff0cGPT-2 \u6700\u5927\u7684\u7279\u9ede\u662f\u4f7f\u7528\u5920\u591a\u7684\u8a13\u7df4\u6587\u672c\uff0c\u4f7f\u5176\u751f\u6210\u6587\u5b57\u7684\u6548\u679c\u53ca\u5176\u986f\u8457\uff0c\u505a\u70ba\u4e00\u822c\u76ee\u7684\u6027\u7684\u8a9e\u8a00\u6a21\u578b\u800c\u8a00\u975e\u5e38\u5f37\u5927\u3002<\/p>\n\n\n\n

BERT \u5247\u662f\u5728\u7279\u5b9a\u4efb\u52d9\u4e2d\u7684\u8868\u73fe\u975e\u5e38\u597d\uff0c\u5982\uff1a\u514b\u6f0f\u5b57\u586b\u5145\u3001\u554f\u7b54\u751f\u6210\u7b49\u7b49\uff0c\u53ef\u4ee5\u767c\u73fe\uff0c\u9019\u4e9b\u6a21\u578b\u4e0d\u5728\u65bc\u65b0\u67b6\u69cb\uff0c\u800c\u662f\u4f5c\u70ba\u4e0d\u540c\u76ee\u7684\u6027\u7684\u4f86\u7d44\u88dd (\u9b54\u6539\uff1f) \u904e\u53bb\u7684\u6a21\u578b\u5011\u3002<\/p>\n\n\n\n

\u8a72\u5982\u4f55\u7684\u53bb\u4e86\u89e3\u9019\u4e9b\u8fd1\u4ee3\u71b1\u9580\u7684 NLP \u6a21\u578b\uff0c\u60f3\u5fc5\u4e0d\u80fd\u5f9e\u6700\u57fa\u672c\u7684\u5faa\u74b0\u795e\u7d93\u7db2\u8def RNN \u958b\u59cb\u77ad\u89e3\u8d77\u5566~ \u6839\u64da\u904e\u53bb\u5e7e\u5e74\u7684\u6559\u5b78\u7d93\u9a57\uff0c\u6211\u5df2\u7d93\u6e96\u5099\u597d\u8b93\u540c\u5b78\u53ef\u4ee5\u5feb\u901f\u63a5\u8ecc\u4e16\u754c\u7684\u5165\u9580\u78da\u7b49\u7d1a\u7684\u6559\u6750\u3002 <\/p>\n\n\n\n

\u4e0d\u7528\u64d4\u5fc3\u4f5c\u696d\u7cfb\u7d71\u6703\u8b93\u5b89\u88dd\u7a0b\u5f0f\u9ebb\u7169\u767e\u767e\u7a2e\uff0c\u4e5f\u4e0d\u7528\u7169\u60f1\u6578\u5b78\u4e0d\u597d\u6216\u662f\u7a0b\u5f0f\u4e0d\u6703\u5beb\uff0c\u9019\u4e9b\u6211\u901a\u901a\u90fd\u7528\u6700\u7c21\u55ae\u7684\u65b9\u5f0f\u7b49\u4f60\u4f86\u5b78\u7fd2\u4e86\uff01<\/p>\n\n\n\n

\u7acb\u5373\u9ede\u64ca\u5b78\u7fd2\u66f4\u591a\u95dc\u65bc\u6df1\u5ea6\u5b78\u7fd2,\u8cfc\u8ab2\u8f38\u5165”21BRO95″\u4eabTibaMe Blog\u8b80\u8005\u5c08\u5c6c\u975e\u65e9\u9ce5\u8ab2\u7a0b95\u6298\u512a\u60e0(2021\/12\/31\u6b62)<\/div><\/figcaption><\/a><\/figure>\n","protected":false},"excerpt":{"rendered":"

\u4f60\u77e5\u9053\u73fe\u5728\u6700\u6027\u611f\u3001\u6700\u6436\u624b\u3001\u6700\u8cfa\u9322\u7684\u8077\u696d\u662f\u4ec0\u9ebc\u55ce\uff1f\u662f\u8cc7\u6599\u79d1\u5b78\u5bb6(Data Scientist)\uff0c\u53c8\u53ef\u8b6f\u70ba\u6578\u64da\u79d1\u5b78\u5bb6\u3002\u5728\u54c8\u4f5b\u5546\u696d\u8a55\u8ad6\u4ee5\u300c21\u4e16\u7d00\u6700\u6027\u611f\u7684\u8077\u696d\u300d\u4ecb\u7d39\u8cc7\u6599\u79d1\u5b78\u5bb6\u4e4b\u5f8c\uff0c\u8a31\u591a\u4eba\u767c\u73fe\u9019\u500b\u8077\u696d\u4e0d\u50c5\u6027\u611f\uff0c\u9084\u5f88\u6436\u624b\u4e14\u9ad8\u85aa\u3002\u7f8e\u570b\u7684\u8cc7\u6599\u79d1\u5b78\u4eba\u624d\u8d77\u85aa\u5247\u5f9e165\u842c\u53f0\u5e63\u8d77\u8df3\uff0c\u9ad8\u9054400\u842c\u4ee5\u4e0a\u4e5f\u6642\u6709\u8033\u805e\u3002\u6240\u4ee5\uff0c\u5230\u5e95\u4ec0\u9ebc\u662f\u8cc7\u6599\u79d1\u5b78\u5bb6\uff1f<\/p>\n","protected":false},"author":19,"featured_media":15292,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_lmt_disableupdate":"","_lmt_disable":"","footnotes":""},"categories":[480,1196,173,1543],"tags":[692,774,775,776,777],"class_list":["post-15252","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-technology","category-ai-","category-173","category-1543","tag-692","tag-rnn","tag-cnn","tag-nlp","tag-777"],"aioseo_notices":[],"views":4730,"amp_enabled":true,"_links":{"self":[{"href":"https:\/\/blog.tibame.com\/index.php?rest_route=\/wp\/v2\/posts\/15252","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.tibame.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.tibame.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.tibame.com\/index.php?rest_route=\/wp\/v2\/users\/19"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.tibame.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=15252"}],"version-history":[{"count":2,"href":"https:\/\/blog.tibame.com\/index.php?rest_route=\/wp\/v2\/posts\/15252\/revisions"}],"predecessor-version":[{"id":17702,"href":"https:\/\/blog.tibame.com\/index.php?rest_route=\/wp\/v2\/posts\/15252\/revisions\/17702"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blog.tibame.com\/index.php?rest_route=\/wp\/v2\/media\/15292"}],"wp:attachment":[{"href":"https:\/\/blog.tibame.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=15252"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.tibame.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=15252"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.tibame.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=15252"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}