{"id":14235,"date":"2021-11-30T19:38:11","date_gmt":"2021-11-30T11:38:11","guid":{"rendered":"https:\/\/achieve.dhcn.cn\/?p=14235"},"modified":"2024-04-04T01:30:01","modified_gmt":"2024-04-03T17:30:01","slug":"%e4%b8%80%e7%a7%8d%e5%9f%ba%e4%ba%8etransformer%e6%a8%a1%e5%9e%8b%e7%9a%84%e5%8f%a4%e7%b1%8d%e8%87%aa%e5%8a%a8%e6%a0%87%e7%82%b9%e6%8a%80%e6%9c%af","status":"publish","type":"post","link":"https:\/\/achieve.dhcn.cn\/en\/site\/works\/dhjournal\/202102\/14235.html","title":{"rendered":"\u4e00\u79cd\u57fa\u4e8eTransformer\u6a21\u578b\u7684\u53e4\u7c4d\u81ea\u52a8\u6807\u70b9\u6280\u672f"},"content":{"rendered":"

\u4f5c\u8005\uff1a <\/span>\u6d2a\u6d9b \u7a0b\u745e\u96ea \u5218\u601d\u6c50 \u65b9\u51ef\u9f50 <\/span>\uff1b\u8f6c\u81ea\uff1a\u516c\u53f7 DH\u6570\u5b57\u4eba\u6587<\/p>\n\n\n\n

\u6587\u672c\u5206\u6790<\/span><\/strong><\/p>\n\n\n\n

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

\u6d2a\u6d9b \/ \u53e4\u8054(\u5317\u4eac)\u6570\u5b57\u4f20\u5a92\u79d1\u6280\u6709\u9650\u516c\u53f8<\/span><\/strong><\/p>\n\n\n\n

\u7a0b\u745e\u96ea \/ \u53e4\u8054(\u5317\u4eac)\u6570\u5b57\u4f20\u5a92\u79d1\u6280\u6709\u9650\u516c\u53f8<\/span><\/strong><\/p>\n\n\n\n

\u5218\u601d\u6c50 \/ \u53e4\u8054(\u5317\u4eac)\u6570\u5b57\u4f20\u5a92\u79d1\u6280\u6709\u9650\u516c\u53f8<\/span><\/strong><\/p>\n\n\n\n

\u65b9\u51ef\u9f50 \/ \u5f69\u5f7b\u533a\u660e\u79d1\u6280\u6709\u9650\u516c\u53f8<\/span><\/strong><\/p>\n\n\n\n

————————————<\/span><\/p>\n\n\n\n

\u6458\u8981:<\/span><\/strong>\u4eba\u5de5\u52a0\u6ce8\u53e4\u7c4d\u6587\u732e\u6807\u70b9\u7b26\u53f7\uff0c\u5373\u5bf9\u65e0\u6807\u70b9\u7b26\u53f7\u7684\u53e4\u7c4d\u6587\u732e\u8fdb\u884c\u73b0\u4ee3\u6807\u70b9\u7b26\u53f7\u586b\u5145\u3002\u4f46\u53e4\u7c4d\u6570\u91cf\u5e9e\u5927\uff0c\u4eba\u5de5\u52a0\u6ce8\u8d39\u65f6\u8d39\u529b\u3002\u968f\u7740\u4eba\u5de5\u667a\u80fd(AI)\u7684\u5174\u8d77\uff0c\u57fa\u4e8e\u6df1\u5ea6\u5b66\u4e60\u5b9e\u73b0\u81ea\u52a8\u6807\u70b9\u5de5\u4f5c\u53ef\u4ee5\u51cf\u8f7b\u4eba\u5de5\u7684\u7e41\u91cd\u8d1f\u62c5\u3002\u672c\u65b9\u6cd5\u4f7f\u7528\u57fa\u4e8e\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u7684\u7aef\u5230\u7aefTransformer\u6a21\u578b\u4f5c\u4e3a\u8bad\u7ec3\u6a21\u578b\uff0c\u4f7f\u7528\u6b63\u89c4\u51fa\u7248\u768410\u4ebf\u5b57\u53e4\u7c4d\u8bed\u6599\u8fdb\u884c\u8bad\u7ec3\u3002\u6a21\u578b\u5728\u9a8c\u8bc1\u96c6\u4e0a\u7684\u6807\u70b9F1\u4e3a86.5%\uff0c\u65ad\u53e5F1\u4e3a95.1%\u3002\u968f\u540e\u5728\u672a\u8bad\u7ec3\u7684\u8bed\u6599\u4e2d\u62bd\u53d619\u672c\u4e66\u4f5c\u4e3a\u6d4b\u8bd5\u96c6\uff0c\u7ed3\u679c\u663e\u793a\uff0c\u6a21\u578b\u5bf9\u53e4\u5178\u6587\u8a00\u6587\u8bed\u6599\u9884\u6d4b\u7ed3\u679c\u8f83\u597d;\u5bf9\u4e13\u4e1a\u6027\u8f83\u5f3a\u548c\u767d\u8bdd\u6587\u5360\u6bd4\u8f83\u591a\u7684\u8bed\u6599\u5219\u6548\u5ea6\u8f83\u4f4e\u3002<\/p>\n\n\n\n

\u5173\u952e\u8bcd<\/strong>:<\/span>\u53e4\u7c4d \u81ea\u7136\u8bed\u8a00\u5904\u7406 \u81ea\u52a8\u6807\u70b9 \u673a\u5668\u5b66\u4e60 \u6df1\u5ea6\u5b66\u4e60<\/p>\n\n\n\n

———————————— <\/span><\/p>\n\n\n\n

\u4e00\u3001\u6982\u8ff0<\/span><\/strong><\/p>\n\n\n\n

(\u4e00)\u76f8\u5173\u7814\u7a76<\/span><\/p>\n\n\n\n

\u6807\u70b9\u7b26\u53f7\u662f\u5f53\u4eca\u4eba\u4eec\u7406\u89e3\u6587\u732e\u4f5c\u54c1\u7684\u91cd\u8981\u5de5\u5177\uff0c\u7136\u800c\u5f88\u591a\u53e4\u7c4d\u6587\u732e\u539f\u672c\u5e76\u6ca1\u6709\u6807\u70b9\u7b26\u53f7\uff0c\u8fd9\u5c31\u53ef\u80fd\u4f1a\u4f7f\u8bfb\u8005\u5728\u9605\u8bfb\u8fc7\u7a0b\u4e2d\u65e0\u6cd5\u7406\u89e3\u53e4\u7c4d\u6587\u732e\u7684\u771f\u6b63\u6db5\u4e49\u3002\u81ea\u52a8\u6807\u70b9\uff0c\u662f\u6307\u5229\u7528\u4eba\u5de5\u667a\u80fd\uff0c\u6839\u636e\u7279\u5b9a\u7b97\u6cd5\u81ea\u52a8\u7ed9\u6ca1\u6709\u6807\u70b9\u7684\u53e4\u7c4d\u6587\u672c\u6807\u6ce8\u73b0\u4ee3\u4e2d\u6587\u6807\u70b9\u7684\u6280\u672f\u3002\u65e0\u9700\u4eba\u5de5\u5e72\u6d89\uff0c\u6a21\u578b\u5c06\u901a\u8fc7\u5b66\u4e60\u6240\u63d0\u4f9b\u7684\u201c\u5df2\u7ecf\u6807\u70b9\uff0c\u4e14\u8d28\u91cf\u8f83\u9ad8\u7684\u53e4\u7c4d\u6587\u672c\u201d[1]<\/sup>\u4e2d\u7684\u53e5\u5b50\u7279\u70b9\uff0c\u81ea\u4e3b\u9009\u62e9\u7279\u5f81\u503c\uff0c\u7ed3\u5408\u73b0\u4ee3\u6807\u70b9\u7528\u6cd5\u5b9e\u73b0\u5bf9\u65e0\u6807\u70b9\u53e4\u7c4d\u7684\u6807\u6ce8\u3002<\/p>\n\n\n\n

\u65e9\u671f\u9ec4\u5efa\u5e74\u7b49[2]<\/sup>\u6839\u636e\u674e\u94ce\u535a\u58eb\u7b49\u63d0\u51fa\u7684\u8ba9\u8ba1\u7b97\u673a\u81ea\u5b66\u4e60\u8fdb\u884c\u53e4\u7c4d\u81ea\u52a8\u6807\u70b9\u7684\u6784\u60f3[3]<\/sup>\u7ed3\u5408\u89c4\u5219\u4e3a\u4e3b\u3001\u8ba1\u7b97\u4e3a\u8f85\u7684\u65b9\u6cd5\uff0c\u5b9e\u73b0\u5bf9\u519c\u4e1a\u53e4\u7c4d\u7684\u65ad\u53e5\u3002\u5f20\u5f00\u65ed\u7b49[4]<\/sup>\u901a\u8fc7\u4eba\u5de5\u5e72\u6d89\u7279\u5f81\u503c\uff0c\u4f7f\u7528\u6761\u4ef6\u968f\u673a\u573a(conditional random field, CRF)[5]<\/sup>\u5bf9\u53e4\u7c4d\u65ad\u53e5\u53ca\u6807\u70b9\u8fdb\u884c\u5efa\u6a21\u3002<\/p>\n\n\n\n

\u540e\u671f\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u53d1\u5c55\uff0c\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u88ab\u5e94\u7528\u4e8e\u81ea\u7136\u8bed\u8a00\u5904\u7406\u7684\u4efb\u52a1\u4e2d\u3002Huang[6]<\/sup>\u4f7f\u7528\u4e00\u79cd\u57fa\u4e8e\u53cc\u5411\u5faa\u73af\u795e\u7ecf\u7f51\u7edc(bi-directional recurrent neural network)[7]<\/sup>\u6dfb\u52a0CRF\u5c42\u7684\u6a21\u578b\u6765\u89e3\u51b3\u5e8f\u5217\u6807\u6ce8\u7684\u95ee\u9898\u3002\u4fde\u656c\u677e\u7b49[8]<\/sup>\u57fa\u4e8eBERT(Bidirectional Encoder Representations from Transformers)[9]<\/sup>\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u4ece\u800c\u5b9e\u73b0\u5bf9\u53e4\u6587\u8fdb\u884c\u81ea\u52a8\u65ad\u53e5\u53ca\u81ea\u52a8\u6807\u70b9\u7684\u4efb\u52a1\u3002<\/p>\n\n\n\n

\u672c\u6587\u63d0\u51fa\u4e00\u79cd\u57fa\u4e8e\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u7684Transformer[10]<\/sup>\u6a21\u578b\uff0c\u5b9e\u73b0\u53e4\u7c4d\u81ea\u52a8\u6807\u70b9\u548c\u81ea\u52a8\u65ad\u53e5\u7684\u76ee\u6807\u3002\u8be5\u65b9\u6cd5\u901a\u8fc7\u5728\u5927\u91cf\u5df2\u51fa\u7248\u7684\u53e4\u7c4d\u6570\u5b57\u7248\u8bed\u6599\u4e2d\u81ea\u4e3b\u5b66\u4e60\u8bed\u6599\u7279\u5f81\uff0c\u5728\u968f\u673a\u9009\u62e9\u7684\u6d4b\u8bd5\u96c6\u4e2d\uff0c\u6bd4\u4f5c\u4e3a\u57fa\u51c6\u6a21\u578b\u7684LSTM\u6a21\u578b\u548c\u4f5c\u4e3a\u5bf9\u6bd4\u6a21\u578b\u7684BERT+CNN(Convolutional Neural Network\uff0c\u5377\u79ef\u795e\u7ecf\u7f51\u7edc)[11]<\/sup>\u6a21\u578b\u5c55\u793a\u51fa\u4e86\u66f4\u4e3a\u4f18\u79c0\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n

(\u4e8c)\u6982\u5ff5\u672f\u8bed<\/span><\/p>\n\n\n\n

\u6807\u70b9\u662f\u7528\u6765\u8868\u793a\u8bed\u53e5\u7684\u505c\u987f\u3001\u8bed\u6c14\u4ee5\u53ca\u6807\u793a\u67d0\u4e9b\u6210\u5206(\u4e3b\u8981\u662f\u8bcd\u8bed)\u7684\u7279\u5b9a\u6027\u8d28\u548c\u4f5c\u7528\u7684\u3002\u73b0\u4ee3\u6807\u70b9\u5206\u4e3a\u70b9\u53f7\u548c\u6807\u53f7\u4e24\u4e2a\u79cd\u7c7b\u3002\u70b9\u53f7\u7684\u4f5c\u7528\u662f\u70b9\u65ad\uff0c\u4e3b\u8981\u8868\u793a\u505c\u987f\u548c\u8bed\u6c14\uff0c\u5e38\u7528\u7684\u6709\u53e5\u53f7(\u3002)\u3001\u95ee\u53f7(?)\u3001\u53f9\u53f7(!)\u3001\u9017\u53f7(\uff0c)\u3001\u987f\u53f7(\u3001)\u3001\u5206\u53f7(;)\u3001\u5192\u53f7(:);\u6807\u53f7\u7684\u4f5c\u7528\u5728\u4e8e\u6807\u660e\uff0c\u4e3b\u8981\u6807\u660e\u8bed\u53e5\u7684\u6027\u8d28\u548c\u4f5c\u7528\uff0c\u5e38\u7528\u7684\u6709\u5f15\u53f7(\u201c\u201d)\u3001\u62ec\u53f7(())\u3001\u7834\u6298\u53f7(\u2014\u2014)\u3001\u7701\u7565\u53f7(……)\u3001\u4e66\u540d\u53f7(\u300a\u300b)\u7b49\u51719\u79cd\u3002[12]<\/sup>\u672c\u6587\u6240\u63d0\u5230\u7684\u6807\u70b9\u7b26\u53f7\u4ec5\u9650\u4e8e\u70b9\u53f7\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u65b9\u6cd5\u4e0e\u6a21\u578b<\/span><\/strong><\/p>\n\n\n\n

(\u4e00)\u65b9\u6cd5<\/span><\/p>\n\n\n\n

\u6211\u4eec\u5c06\u70b9\u53f7\u6807\u6ce8\u7684\u4efb\u52a1\u770b\u4f5c\u662f\u4e00\u79cd\u5e8f\u5217\u9884\u6d4b\u95ee\u9898\u3002\u5373\u8f93\u5165\u4e00\u6bb5\u957f\u5ea6\u4e3an\u5b57\u7684\u53e4\u6587\u2014\u2014\u4e00\u4e2as1\uff0cs2\uff0c……\uff0csn\u7684\u6587\u672c\u5e8f\u5217\uff0c\u6a21\u578b\u5c06\u8f93\u51fa\u4e00\u4e2a\u7b49\u957f\u7684T1\uff0cT2\uff0c……\uff0cTn\u7684\u6807\u7b7e\u5e8f\u5217\uff0c\u5176\u4e2dTi\u4ee3\u8868\u539f\u6587\u4e2d\u7b2ci\u4e2a\u6c49\u5b57\u540e\u6240\u8ddf\u7684\u70b9\u53f7\u3002\u5982\u679c\u4e00\u4e2a\u6c49\u5b57\u540e\u6ca1\u6709\u70b9\u53f7\uff0c\u6211\u4eec\u5219\u7edf\u4e00\u7528\u4e00\u4e2a\u4ee3\u8868\u4e0d\u65ad\u53e5\u7684\u7b26\u53f7\u201cC\u201d\u8868\u793a\u3002<\/p>\n\n\n\n

\u56fe1\u4e3a\u672c\u6587\u6240\u4f7f\u7528\u7aef\u5230\u7aef\u6a21\u578b\u642d\u914d\u4f8b\u53e5\u201c\u4e09\u4eba\u884c\uff0c\u5fc5\u6709\u6211\u5e08\u7109\u3002\u201d\u7684\u7b80\u8981\u8bf4\u660e:\u8be5\u53e5\u4e2d\uff0c\u201c\u884c\u201d\u5b57\u548c\u201c\u7109\u201d\u5b57\u540e\u9762\u5bf9\u5e94\u7684\u70b9\u53f7\u662f\u201c\uff0c\u201d\u548c\u201c\u3002\u201d\uff0c\u5176\u4ed6\u5b57\u540e\u4e0d\u5b58\u5728\u70b9\u53f7\uff0c\u56e0\u6b64\u539f\u6587s1\uff0cs2\uff0c……\uff0cs8\u201c\u4e09\u4eba\u884c\u5fc5\u6709\u6211\u5e08\u7109\u201d\u6240\u5bf9\u5e94\u7684\u70b9\u53f7\u5e8f\u5217T1\uff0cT2\uff0c……\uff0cT8\u5219\u4e3a\u201cCC\uff0cCCCC\u3002\u201d\u3002\u6211\u4eec\u5c06\u7528\u4e0d\u540c\u7684\u6a21\u578b\u7ed3\u6784\u6765\u5b66\u4e60\u9884\u6d4b\u8fd9\u4e00\u5e8f\u5217\u3002<\/p>\n\n\n\n

\"\"
\u56fe1 \u6a21\u578b\u539f\u7406\u793a\u4f8b\u56fe<\/figcaption><\/figure><\/div>\n\n\n\n

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

\u672c\u6587\u4f7f\u7528\u7684Transformer\u6a21\u578b\u91c7\u7528\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d\u5e38\u7528\u7684\u7aef\u5230\u7aef\u7f16\u7801\u2014\u89e3\u7801[13]<\/sup>\u6a21\u578b\u8fdb\u884c\u70b9\u53f7\u6807\u6ce8\u4efb\u52a1\u3002\u8be5\u6a21\u578b\u9996\u5148\u5c06\u8f93\u5165\u6c49\u5b57\u8f6c\u5316\u4e3a\u5bf9\u5e94\u7684\u9690\u72b6\u6001\u7f16\u7801(Input Embedding)\uff0c\u901a\u8fc7\u6dfb\u52a0\u4f4d\u7f6e\u7f16\u7801(Positional Encoding)\u5f97\u5230\u4e00\u4e2a\u5e26\u6709\u8f93\u5165\u6587\u672c\u4e2d\u6bcf\u4e2a\u5b57\u5b8c\u6574\u4fe1\u606f\u7684\u5411\u91cf\u5316\u8868\u793a\u3002\u9690\u72b6\u6001\u901a\u8fc7\u7f16\u7801\u5668(Encoder)\u540e\u7ecf\u8fc7\u591a\u5c42\u8ba1\u7b97\u8f6c\u5316\u4e3a\u9ad8\u7ef4\u5ea6\u5411\u91cf\uff0c\u5e76\u5c06\u7ed3\u679c\u8f93\u5165\u5230\u89e3\u7801\u5668(Decoder)\u4e2d\u3002\u89e3\u7801\u5668\u5c06\u8f93\u5165\u7ed3\u679c\u8fdb\u884c\u8fdb\u4e00\u6b65\u5904\u7406\uff0c\u7136\u540e\u901a\u8fc7\u7ebf\u6027(Linear)\u5c42\u5c06\u89e3\u7801\u5668\u8f93\u51fa\u7684\u9ad8\u7ef4\u5411\u91cf\u6620\u5c04\u4e3a\u4ee5\u6807\u70b9\u6570\u91cf\u4e3a\u7ef4\u5ea6\u7684\u4f4e\u7ef4\u5411\u91cf\u3002\u6700\u540e\uff0c\u901a\u8fc7\u5206\u7c7b\u5668\u5c42\u8fdb\u884cSoftmax\u8ba1\u7b97\u5f97\u5230\u5bf9\u5e94\u4f4d\u7f6e\u7684\u6807\u70b9\u7ed3\u679c\u9884\u6d4b\u6982\u7387\u3002<\/p>\n\n\n\n

(\u4e8c)\u6a21\u578b\u539f\u7406<\/span><\/p>\n\n\n\n

\u5728\u4e0a\u8282\u65b9\u6cd5\u6982\u8ff0\u4e2d\u6240\u4ecb\u7ecd\u7684\u7aef\u5230\u7aef\u7f16\u7801\u2014\u89e3\u7801\u6a21\u578b\u4e2d\u7684\u7f16\u7801\u5668\u7684\u5177\u4f53\u5b9e\u73b0\u65b9\u6cd5\u53ef\u4ee5\u662f\u591a\u79cd\u7684\uff0c\u6bd4\u5982\u7f16\u7801\u5668\u7531Transformer\u6a21\u5757\u6784\u6210\uff0c\u89e3\u7801\u5668\u7531LSTM\u6a21\u5757\u6784\u6210\u3002BERT\u6a21\u578b\u53ea\u9700\u8981Transformer\u6a21\u5757\u7ec4\u6210\u7684\u7f16\u7801\u5668\u52a0\u4e0a\u5206\u7c7b\u5668\u6784\u6210\uff0c\u4e0d\u9700\u8981\u89e3\u7801\u5668\u3002\u800c\u672c\u6587\u4e2d\u6240\u63d0\u5230\u7684LSTM\u6a21\u578b\u4e5f\u4ec5\u662f\u7531LSTM\u6a21\u5757\u7f16\u7801\u5668\u52a0\u5206\u7c7b\u5668\u7ec4\u5408\u800c\u6210\u3002\u5177\u4f53\u7684\u6a21\u5757\u4ecb\u7ecd\u5982\u4e0b\u3002<\/p>\n\n\n\n

1.Transformer\u6a21\u5757<\/span><\/strong><\/p>\n\n\n\n

Transformer\u6a21\u578b\u662f\u81ea2017\u5e74\u51fa\u73b0\u4ee5\u6765\uff0c\u5728\u591a\u9886\u57df\u5f97\u5230\u5e7f\u6cdb\u4f7f\u7528\u7684\u4e00\u7c7b\u795e\u7ecf\u7f51\u7edc\u6a21\u5757\u3002\u6a21\u578b\u7684\u7f16\u7801\u5668\u90e8\u5206\u6bcf\u4e00\u5c42\u7531\u4e00\u4e2a\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757(Multi-Head Attention)\u548c\u4e00\u4e2a\u524d\u9988\u5c42(Feed Forward)\u7ec4\u6210\uff0c\u5176\u4e2d\u7684\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u7531\u591a\u4e2a\u6ce8\u610f\u529b(Attention)\u7ec4\u6210\uff0c\u5355\u4e2a\u6ce8\u610f\u529b\u7684\u516c\u5f0f\u4e3a:<\/p>\n\n\n\n

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

\u5176\u4e2d\uff0cQ\u4e3a\u67e5\u8be2\u77e9\u9635(query)\uff0cK\u4e3a\u952e\u77e9\u9635(key)\uff0cV\u4e3a\u503c\u77e9\u9635(value)\uff0c\u7531\u6a21\u578b\u8be5\u5c42\u7684\u8f93\u5165\u901a\u8fc7\u72ec\u7acb\u7684\u7ebf\u6027\u53d8\u6362\u751f\u6210\u3002dk\u4e3aQ\u3001K\u7684\u5411\u91cf\u7ef4\u5ea6\uff0c\u4e3a\u8ba1\u7b97\u65b9\u6cd5\u4e2d\u7684\u7f29\u653e(scale)\u90e8\u5206\uff0c\u8be5\u64cd\u4f5c\u5b58\u5728\u662f\u4e3a\u4e86\u8ba9softmax\u7684\u7ed3\u679c\u66f4\u4e3a\u660e\u663e\u3002<\/p>\n\n\n\n

\u6a21\u578b\u7ed3\u6784\u4e2d\u4f7f\u7528\u7684\u591a\u5934\u6ce8\u610f\u529b\u5219\u662f\u901a\u8fc7\u7ebf\u6027\u53d8\u6362\u5f97\u5230\u8bbe\u5b9a\u597d\u7684\u6ce8\u610f\u529b\u5934\u6570(h)\u4efdQ\u3001K\u3001V\uff0c\u5404\u81ea\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\u540e\uff0c\u5c06\u53c2\u6570\u4e0d\u5171\u4eab\u7684\u4e0d\u540c\u6ce8\u610f\u529b\u7ed3\u679c\u62fc\u63a5\u8d77\u6765\uff0c\u5176\u516c\u5f0f\u4e3a:<\/p>\n\n\n\n

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

\u62fc\u63a5\u597d\u7684\u591a\u5934\u6ce8\u610f\u529b\u5c42\u7684\u7ed3\u679c\u518d\u901a\u8fc7\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u5f52\u4e00\u5316(Add & LayerNorm)\uff0c\u4ee5\u53ca\u524d\u9988\u5c42\uff0c\u4e5f\u5c31\u662f\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\u548c\u4e00\u4e2aRelu\u6fc0\u6d3b\u51fd\u6570\u540e\uff0c\u5b8c\u6210\u4e00\u5c42Transformer\u5c42\u7684\u8ba1\u7b97\u3002<\/p>\n\n\n\n

\u89e3\u7801\u5668\u7684\u6bcf\u4e00\u5c42\u6bd4\u7f16\u7801\u5668\u6bcf\u4e00\u5c42\u591a\u51fa\u4e00\u4e2a\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\uff0c\u5176\u4ed6\u7ec4\u6210\u57fa\u672c\u4e00\u81f4\u3002\u65b0\u589e\u7684\u6ce8\u610f\u529b\u6a21\u5757\u662f\u4e00\u4e2a\u5c06\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u6570\u636e\u76f8\u5173\u8054\u7684\u591a\u5934\u6ce8\u610f\u529b\u5c42\uff0c\u5176\u4e2d\u7684V\u3001K\u6765\u81ea\u4e8e\u7f16\u7801\u5668\u7684\u8f93\u51fa\u7ed3\u679c\uff0cQ\u5219\u6765\u81ea\u4e0a\u4e00\u5c42\u89e3\u7801\u5668\u4e2d\u6ce8\u610f\u529b\u5c42\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n

2.LSTM\u6a21\u5757<\/span><\/strong><\/p>\n\n\n\n

LSTM\u662f\u4e00\u79cd\u5e38\u89c1\u7684\u5faa\u73af\u795e\u7ecf\u7f51\u7edc(Recurrent Neural Network, RNN)\u7ed3\u6784\u3002\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u4e00\u822c\u7528\u5982\u4e0b\u516c\u5f0f\u63cf\u8ff0:<\/p>\n\n\n\n

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

\u5176\u4e2dxt\u4e3a\u5f53\u524d\u5c42\u5728\u65f6\u95f4\u6b65t\u8f93\u5165\u7684\u9690\u72b6\u6001\uff0cht\u4e3a\u8be5\u5c42\u7ecf\u8fc7\u4e00\u5b9a\u8fd0\u7b97\u7684\u8f93\u51fa\uff0c\u800cht-1\u662f\u4e0a\u4e00\u4e2a\u65f6\u95f4\u6b65\u5f53\u524d\u5c42\u7684\u8f93\u51fa\uff0cf\u8868\u793a\u795e\u7ecf\u7f51\u7edc\u7684\u8fd0\u7b97\u51fd\u6570\uff0c\u5b9e\u73b0\u65b9\u5f0f\u6709\u591a\u79cd\u3002\u5728\u672c\u5b9e\u9a8c\u4e2d\uff0c\u6bcf\u4e00\u4e2a\u65f6\u95f4\u6b65\u7684\u8f93\u5165\u5373\u4ee3\u8868\u4e00\u4e2a\u6c49\u5b57\u7ecf\u8fc7\u7f16\u7801\u7684\u9690\u72b6\u6001\u3002\u4ece\u8be5\u516c\u5f0f\u53ef\u4ee5\u770b\u51fa\uff0c\u6bcf\u4e00\u4e2a\u5b57\u7684\u8f93\u51fa\u90fd\u663e\u5f0f\u5229\u7528\u4e86\u4e0a\u4e00\u4e2a\u5b57(\u65f6\u95f4\u6b65)\u7684\u8f93\u51fa\u7ed3\u679c\uff0c\u8fd9\u4e5f\u548c\u4eba\u9605\u8bfb\u6587\u5b57\u7684\u987a\u5e8f\u7c7b\u4f3c\u3002<\/p>\n\n\n\n

LSTM\u5373\u662f\u5bf9\u5faa\u73af\u795e\u7ecf\u7f51\u8def\u4e2d\u7684\u51fd\u6570f\u7684\u4e00\u7c7b\u5b9e\u73b0\uff0c\u5b83\u5f15\u5165\u4e86\u4e00\u4e9b\u95e8\u63a7\u5355\u5143\uff0c\u540c\u65f6\u5f15\u5165\u8bb0\u5fc6\u5355\u5143\uff0c\u6765\u5b9e\u73b0\u5bf9\u8f93\u5165\u548c\u8f93\u51fa\u66f4\u597d\u7684\u8c03\u8282\u3002\u4e3a\u4e86\u5b9e\u73b0\u66f4\u597d\u7684\u6548\u679c\uff0c\u6211\u4eec\u4f7f\u7528\u4e86\u53cc\u5411LSTM\u7f51\u7edc(Bi-LSTM)\uff0c\u6bcf\u4e00\u5c42\u7684\u8f93\u51fa\u90fd\u662f\u4ece\u5de6\u5411\u53f3\u548c\u4ece\u53f3\u5411\u5de6\u4e24\u4e2a\u4e0d\u540c\u65b9\u5411LSTM\u7684\u7ed3\u679c\u76f8\u52a0\u3002<\/p>\n\n\n\n

3.BERT+CNN\u6a21\u5757<\/span><\/strong><\/p>\n\n\n\n

BERT\u6a21\u578b\u7531\u591a\u4e2aTransformer\u5c42\u5806\u53e0\u7ec4\u6210\uff0c\u5176\u67b6\u6784\u7c7b\u4f3c\u7aef\u5230\u7aefTransformer\u6a21\u578b\u4e2d\u7684\u7f16\u7801\u5668\u3002\u5b83\u4e00\u822c\u88ab\u7528\u4e8e\u5728\u5927\u91cf\u65e0\u76d1\u7763\u6570\u636e(\u5373\u65e0\u6807\u70b9\u7b26\u53f7\u7684\u6570\u636e)\u6216\u534a\u76d1\u7763\u6570\u636e(\u5373\u5305\u542b\u65e0\u6807\u70b9\u7b26\u53f7\u7684\u6570\u636e\u548c\u6709\u6807\u70b9\u7b26\u53f7\u7684\u6570\u636e)\u4e0a\u8fdb\u884c\u7279\u5b9a\u4efb\u52a1\u7684\u9884\u8bad\u7ec3\uff0c\u8ba9\u6a21\u578b\u81ea\u5df1\u4ece\u5927\u91cf\u8bed\u6599\u4e2d\u5b66\u4e60\u5230\u5e95\u5c42\u548c\u4e2d\u5c42\u7279\u5f81;\u968f\u540e\u5728\u5c0f\u89c4\u6a21\u76d1\u7763\u6570\u636e\u4e0a\u8fdb\u884c\u5fae\u8c03\u8bad\u7ec3\uff0c\u901a\u5e38\u80fd\u83b7\u5f97\u6bd4\u5728\u5c0f\u6570\u636e\u96c6\u4e0a\u76f4\u63a5\u8bad\u7ec3\u66f4\u597d\u7684\u6548\u679c\u3002\u6a21\u578b\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u5b8c\u6210\u4e24\u4e2a\u4efb\u52a1:\u63a9\u853d\u8bed\u8a00\u6a21\u578b(Masked Language Model, MLM)\u548c\u4e0b\u4e00\u53e5\u9884\u6d4b(Next Sentence Prediction, NSP)\u3002MLM\u7684\u8bad\u7ec3\u65b9\u6cd5\uff0c\u662f\u968f\u673a\u906e\u76d6\u4e00\u5b9a\u767e\u5206\u6bd4\u7684\u8f93\u5165\u8bcd\u6c47\uff0c\u7136\u540e\u8ba9\u6a21\u578b\u901a\u8fc7\u8fd9\u4e9b\u8bcd\u5df2\u77e5\u7684\u4e0a\u4e0b\u6587\uff0c\u5b66\u4e60\u9884\u6d4b\u51fa\u8fd9\u4e9b\u88ab\u8986\u76d6\u7684\u5b57\uff0c\u8be5\u4efb\u52a1\u4e0eWord2Vec\u4efb\u52a1\u4e2d\u7684CBOW[14]<\/sup>\u5b66\u4e60\u76ee\u6807\u76f8\u4f3c\u3002NSP\u5219\u662f\u8ba9\u6a21\u578b\u5224\u65ad\u51fa\u88ab\u968f\u673a\u62fc\u63a5\u5728\u4e00\u8d77\u7684\u4e24\u53e5\u8bdd\uff0c\u662f\u5426\u5728\u539f\u6587\u4e2d\u662f\u8fde\u7eed\u7684\u4e24\u53e5\u3002\u4e0d\u8fc7\uff0c\u8fd1\u6765\u7684\u5171\u8bc6[15]<\/sup>\u662f\u5728\u8bad\u7ec3\u65f6\u52a0\u5165NSP\u76ee\u6807\u4f1a\u5bf9\u51c6\u786e\u5ea6\u6709\u8d1f\u9762\u5f71\u54cd\u3002<\/p>\n\n\n\n

\u4e3a\u63d0\u5347\u9884\u6d4b\u6548\u679c\uff0c\u672c\u5b9e\u9a8c\u5728BERT\u6a21\u578b\u540e\u8fd8\u52a0\u5165\u4e86\u4e00\u5c42\u4e00\u7ef4\u5377\u79ef(CNN)\u6a21\u5757\u3002CNN\u6a21\u578b\u5bf9\u7f16\u7801\u5668\u8f93\u5165\u7684\u53e5\u5411\u91cf\uff0c\u5728\u65f6\u95f4\u6b65\u7ef4\u5ea6\u4e0a\u8fdb\u884c\u4e00\u7ef4\u5377\u79ef\uff0c\u5bf9\u53e5\u5b50\u4e2d\u7684\u7279\u5f81\u8fdb\u884c\u63d0\u53d6\uff0c\u6709\u52a9\u4e8e\u6355\u83b7\u77ed\u7a0b\u4f9d\u8d56\u5173\u7cfb\uff0c\u964d\u4f4e\u9519\u8bef\u7387\u3002\u5728CNN\u6a21\u5757\u4e4b\u540e\u4e5f\u6dfb\u52a0\u4e86\u6c60\u5316\u5c42\uff0c\u6700\u540e\u901a\u8fc7\u5206\u7c7b\u5668\u5f97\u5230\u6700\u7ec8\u6807\u70b9\u7ed3\u679c\u3002<\/p>\n\n\n\n

\u4e09\u3001\u5b9e\u9a8c\u5206\u6790<\/span><\/strong><\/p>\n\n\n\n

(\u4e00)\u6570\u636e\u96c6<\/span><\/p>\n\n\n\n

\u6570\u636e\u96c6\u5305\u62ec\u8bad\u7ec3\u96c6\u548c\u6d4b\u8bd5\u96c6\uff0c\u6240\u8ff0\u539f\u59cb\u8bed\u6599\u5747\u4e3a\u5df2\u51fa\u7248\u53e4\u7c4d\u7684\u7535\u5b50\u7248\u6587\u4ef6\uff0c\u5747\u4ece\u4e2d\u534e\u4e66\u5c40\u7ecf\u5178\u53e4\u7c4d\u5e93[16]<\/sup>\u4e2d\u63d0\u53d6\u3002<\/p>\n\n\n\n

1.\u8bad\u7ec3\u96c6\u53ca\u9884\u5904\u7406<\/span><\/strong><\/p>\n\n\n\n

\u4ece\u539f\u59cb\u8bed\u6599\u4e2d\u63d0\u53d6\u7ea610\u4ebf\u5b57\u53e4\u7c4d\u6587\u672c(\u603b\u5171\u5305\u542b3,396\u79cd\u4e66\u76ee)\u4f5c\u4e3a\u8bad\u7ec3\u96c6\u8fdb\u884c\u8bad\u7ec3\uff0c\u56db\u90e8\u7c7b\u5206\u5e03\u60c5\u51b5\u5982\u88681\u6240\u793a\u3002\u5728\u5bf9\u4e0d\u89c4\u8303\u8bed\u6599\u8fdb\u884c\u7b5b\u9009(\u5982\u4e66\u76ee\u6bb5\u843d\u5185\u9017\u53f7\u548c\u53e5\u53f7\u7684\u6bd4\u4f8b\u4e0d\u8fbe\u6807\u5373\u5bf9\u6587\u672c\u8fdb\u884c\u5220\u9664)\u540e\uff0c\u5269\u4f595\u4ebf\u4f59\u5b57\u3002\u603b\u5171\u751f\u6210\u8fd1500\u4e07\u4e2a\u6bb5\u843d\u7528\u4e8e\u8bad\u7ec3\u96c6\uff0c1\u4e07\u4e2a\u6bb5\u843d\u7528\u4e8e\u9a8c\u8bc1\u96c6\u3002\u6700\u540e\uff0c\u6e05\u9664\u53e4\u7c4d\u4e2d\u4e0d\u53ef\u80fd\u5b58\u5728\u7684\u6807\u70b9\u7b26\u53f7\u540e\uff0c\u5bf9\u6587\u672c\u8fdb\u884c\u683c\u5f0f\u8f6c\u6362\uff0c\u5904\u7406\u4e3a\u6a21\u578b\u8bad\u7ec3\u65f6\u7684\u6240\u9700\u6570\u636e\u683c\u5f0f\u3002<\/p>\n\n\n\n

\"\"
\u88681 \u8bad\u7ec3\u96c6\u89c4\u6a21\u53ca\u5206\u5e03<\/figcaption><\/figure><\/div>\n\n\n\n

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

2.\u6d4b\u8bd5\u96c6<\/span><\/strong><\/p>\n\n\n\n

\u7531\u4e8e\u8bad\u7ec3\u7684\u8bed\u6599\u591a\u4e3a\u5e38\u7528\u7ecf\u5178\u6587\u732e\uff0c\u6a21\u578b\u5728\u8fd9\u7c7b\u8bed\u6599\u4e2d\u5df2\u6709\u8f83\u597d\u7684\u8868\u73b0\u3002\u4e3a\u6d4b\u8bd5\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u67e5\u770b\u6a21\u578b\u5bf9\u8bad\u7ec3\u96c6\u4e2d\u8bed\u6599\u4e0d\u591f\u591a\u6216\u8005\u672a\u6d89\u53ca\u5230\u7684\u6587\u4f53\u7684\u9884\u6d4b\u6548\u679c\uff0c\u6211\u4eec\u7279\u610f\u4ece\u672a\u7528\u4e8e\u8bad\u7ec3\u7684\u539f\u5178\u8bed\u6599\u4e2d\uff0c\u62bd\u53d6\u4e86\u4e0e\u8bad\u7ec3\u8bed\u6599\u6709\u4e00\u5b9a\u504f\u5dee\u768419\u79cd\u53e4\u7c4d\u4f5c\u4e3a\u6d4b\u8bd5\u96c6\u3002\u5982\u7ecf\u90e8\u5c0f\u5b66\u7c7b\u7684\u300a\u671d\u9c9c\u65f6\u4ee3\u6c49\u8bed\u6559\u79d1\u4e66\u4e1b\u520a\u300b\u80fd\u53cd\u6620\u4ece\u5143\u672b\u5230\u6e05\u672b\u7684\u6c49\u8bed\u5b9e\u9645\u53e3\u8bed\uff0c\u800c\u53c8\u975e\u4e2d\u56fd\u672c\u571f\u6240\u56fa\u6709;\u53f2\u90e8\u7740\u91cd\u6d4b\u8bd5\u7eaa\u4f20\u4f53\u6b63\u53f2\u4ee5\u5916\u7684\u6742\u53f2\u7c7b\u3001\u653f\u4e66\u7c7b\u7b49;\u5b50\u90e8\u65e2\u6311\u9009\u4e86\u4e13\u4e1a\u6027\u8f83\u5f3a\u7684\u5929\u6587\u7c7b\u300a\u6625\u79cb\u6714\u95f0\u8868\u300b\uff0c\u53c8\u6311\u9009\u4e86\u504f\u901a\u4fd7\u3001\u53e3\u8bed\u5316\uff0c\u53cd\u6620\u6c49\u8bed\u53f2\u53d8\u8fc1\u7684\u5c0f\u8bf4\u7c7b\u4f5c\u54c1\uff0c\u751a\u81f3\u5305\u62ec\u300a\u660e\u6e05\u4e4b\u9645\u897f\u5b66\u6587\u672c\u300b\u8fd9\u7c7b\u65b0\u5b66\u6587\u732e;\u96c6\u90e8\u4e00\u65b9\u9762\u6311\u9009\u65f6\u4ee3\u8de8\u5ea6\u5927\u7684\u603b\u96c6\uff0c\u5982\u300a\u53e4\u4e50\u5e9c\u300b\uff0c\u53e6\u4e00\u65b9\u9762\u6311\u9009\u5b9e\u9645\u8bed\u6599\u4f53\u91cf\u6781\u5927\uff0c\u4e14\u8bed\u8a00\u98ce\u683c\u548c\u5185\u5bb9\u4e30\u5bcc\u591a\u6837\u3001\u8bad\u7ec3\u8bed\u6599\u96be\u4ee5\u5168\u9762\u8986\u76d6\u7684\u620f\u66f2\u7c7b\uff0c\u5982\u300a\u5f20\u51e4\u7ffc\u620f\u66f2\u96c6\u300b\u3002<\/p>\n\n\n\n

\u5bf9\u62bd\u53d6\u768419\u79cd\u4e66\u76ee\uff0c\u6309\u7167\u7ecf\u53f2\u5b50\u96c6\u56db\u90e8\u8fdb\u884c\u5206\u7c7b\u6d4b\u8bd5\u3002\u56e010\u5b57\u4ee5\u4e0b\u7684\u6bb5\u843d\u53e5\u4e49\u4e0d\u592a\u5b8c\u6574\uff0c\u6240\u4ee5\u4ec5\u4f7f\u752810\u5b57\u7b26\u4ee5\u4e0a(\u5305\u542b\u6807\u70b9)\u7684\u6bb5\u843d\u4f5c\u4e3a\u6d4b\u8bd5\u5185\u5bb9\u3002\u6700\u7ec8\u5269\u4f59500\u591a\u4e07\u5b57\u4f5c\u4e3a\u6d4b\u8bd5\u96c6\u3002\u88682\u4e3a\u7b5b\u9009\u540e\u7684\u6d4b\u8bd5\u96c6\u4e66\u76ee\u8868\u53ca\u76f8\u5173\u4fe1\u606f\u3002<\/p>\n\n\n\n

\"\"
\u88682 \u6d4b\u8bd5\u96c6\u4e66\u76ee\u8868<\/figcaption><\/figure><\/div>\n\n\n\n

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

(\u4e8c)\u6a21\u578b\u53c2\u6570<\/span><\/p>\n\n\n\n

\u672c\u6587\u4f7f\u7528\u7684Transformer\u6a21\u578b\u662f\u57fa\u4e8efairseq\u5f00\u6e90\u6846\u67b6[17]<\/sup>\u5b9e\u73b0\u7684\u3002\u5728\u8bad\u7ec3\u53c2\u6570\u8bbe\u7f6e\u4e2d\uff0c\u7f16\u7801\u5668\u8bbe\u7f6e\u4e3a10\u5c42\uff0c\u9690\u72b6\u6001\u7ef4\u5ea6\u8bbe\u7f6e\u4e3a768\u7ef4\uff0c\u524d\u9988\u7f51\u7edc\u9690\u72b6\u6001\u7ef4\u5ea63,072\uff0c\u6bcf\u4e2a\u795e\u7ecf\u6a21\u5757\u7684\u6ce8\u610f\u529b\u5c42\u670912\u4e2a\u591a\u5934\u6ce8\u610f\u529b\u5934;\u89e3\u7801\u5668\u4e3a\u9690\u72b6\u6001512\u76844\u5c42Transformer\u5c42\uff0c\u5176\u4e2d\u591a\u5934\u6ce8\u610f\u529b\u6a21\u5757\u4e3a\u67098\u4e2a\u591a\u5934\u6ce8\u610f\u529b\u5934\u3002\u6a21\u578b\u7684\u5c42\u5f52\u4e00\u5316(Layer Norm)\u6a21\u5757\u91c7\u7528\u548cTransformer\u539f\u8bba\u6587\u4e00\u6837\u7684\u540e\u7f6e\u6a21\u5f0f(Post-LN)\uff0c\u5e76\u4f7f\u7528Liu\u7b49[18]<\/sup>\u4f5c\u8005\u63d0\u51fa\u7684\u81ea\u9002\u5e94\u521d\u59cb\u5316\u65b9\u6848\u63d0\u5347\u6a21\u578b\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u3002\u5728\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u4f18\u5316\u5668\u4e3aradam[19]<\/sup>\u4f18\u5316\u5668\uff0c\u4f7f\u7528inverse-square\u5b66\u4e60\u7387\u4e0b\u964d\u65b9\u6848\uff0c\u6700\u9ad8\u5b66\u4e60\u7387\u8bbe\u7f6e\u4e3a0.0006\uff0cwarmup-updates\u8bbe\u7f6e\u4e3a\u5355\u4e2aepoch\u7684\u66f4\u65b0\u6b21\u6570\uff0c\u6240\u6709dropout\u503c\u90fd\u4e3a0.2\u3002<\/p>\n\n\n\n

\u6b64\u5904\u4f7f\u7528\u7684LSTM\u6a21\u578b\u4e5f\u662f\u57fa\u4e8e\u540c\u6837\u7684fairseq\u5f00\u6e90\u6846\u67b6\u8fdb\u884c\u4fee\u6539\u540e\u5b9e\u73b0\u7684\u3002LSTM\u6a21\u578b\u4ec5\u4f7f\u7528\u7f16\u7801\u5668\u90e8\u5206\uff0c\u5176\u4e2d\u7f16\u7801\u5668\u9690\u72b6\u6001\u7ef4\u5ea6\u8bbe\u7f6e\u4e3a512\u7ef4\uff0c\u7f16\u7801\u5668\u5c42\u6570\u4e3a6\u5c42\u3002\u5728\u6a21\u578b\u8bad\u7ec3\u7684\u8d85\u53c2\u4e2d\uff0c\u4f18\u5316\u5668\u4e3anag\u4f18\u5316\u5668\uff0c\u521d\u59cb\u5b66\u4e60\u7387\u4e3a0.25\uff0cdropout\u503c\u8bbe\u7f6e\u4e3a0.1\u3002<\/p>\n\n\n\n

BERT\u6a21\u578b\u4f7f\u7528\u4e8633\u4ebf\u5b57\u7684\u6b86\u77e5\u9601[20]<\/sup>\u53e4\u5178\u6587\u732e\u8bed\u6599\uff0c\u5728\u5bf9\u6c49\u5b57\u8fdb\u884c\u5206\u5b57\u3001\u63a9\u7801\u6807\u8bb0\u540e\u8fdb\u884c\u8bad\u7ec3\uff0c\u6211\u4eec\u672a\u5bf9\u4e8e\u8be5\u8bed\u6599\u4e2d\u7684\u672a\u6807\u70b9\u6570\u636e\u8fdb\u884c\u989d\u5916\u5904\u7406\uff0c\u5c06\u5176\u540c\u6709\u6807\u70b9\u7684\u6570\u636e\u4e00\u6837\u8fdb\u884c\u8bad\u7ec3\u3002MLM\u7684\u8bad\u7ec3\u6570\u636e\u63a9\u7801\u7b56\u7565\u4e0eBERT\u6a21\u578b\u539f\u8bba\u6587\u4fdd\u6301\u4e00\u81f4\uff0c\u6a21\u578b\u9690\u5c42\u8bbe\u7f6e\u4e3a12\u5c42\uff0c\u6ce8\u610f\u529b\u5934\u6570\u4e3a12\uff0c\u9690\u72b6\u6001\u7ef4\u5ea6\u4e3a768\u3002\u5728\u5fae\u8c03\u8bad\u7ec3\u7684\u53c2\u6570\u4e2d\uff0c\u521d\u59cb\u5b66\u4e60\u7387\u4e3a0.00002\uff0cdropout\u503c\u8bbe\u7f6e\u4e3a0.1\uff0cCNN\u5c42\u5377\u79ef\u6838\u6570\u8bbe\u5b9a\u4e3a3\u3002<\/p>\n\n\n\n

\u56e0\u4e3aBERT\u6a21\u578b\u4f7f\u7528\u4e86\u9884\u8bad\u7ec3\u6570\u636e\uff0c\u800c\u5176\u4ed6\u4e24\u4e2a\u6a21\u578b\u672a\u4f7f\u7528\uff0c\u6545\u6211\u4eec\u5bf9LSTM\u6a21\u578b\u548cTransformer\u6a21\u578b\u4f7f\u7528\u4e86\u6570\u636e\u589e\u5f3a\u65b9\u6848:\u6211\u4eec\u5c06\u540c\u4e00\u7bc7\u6587\u7ae0\u5185\u7684\u8fde\u7eed\u6bb5\u843d\u62fc\u63a5\u6210\u5b57\u6570\u63a5\u8fd1\u4f46\u4e0d\u8d85\u8fc7\u67d0\u4e00\u503cL\u7684\u957f\u5ea6\uff0c\u5176\u4e2dL\u6bcf\u6b21\u53d6\u503c\u4ece[100-\u8bbe\u5b9a\u7684\u6a21\u578b\u6700\u5927\u53e5\u957f]\u6574\u6570\u533a\u95f4\u4e2d\u968f\u673a\u6309\u5747\u7b49\u6982\u7387\u9009\u53d6\u3002\u5f53\u4e00\u7bc7\u6587\u7ae0\u7ed3\u675f\uff0c\u6216\u5df2\u5b8c\u6210\u5f53\u524d\u957f\u5ea6L\u7684\u62fc\u63a5\uff0c\u5219\u91cd\u65b0\u968f\u673a\u9009\u53d6\u4e00\u65b0\u7684L\u503c\u3002\u62fc\u63a5\u540e\u7684\u6570\u636e\u4e0e\u539f\u6570\u636e\u63091:1\u5408\u5e76\u8bad\u7ec3\u3002<\/p>\n\n\n\n

(\u4e09)\u8bc4\u4ef7\u6307\u6570<\/span><\/p>\n\n\n\n

\u81ea\u52a8\u6807\u70b9\u6280\u672f\u7684\u76ee\u7684\u662f\u4e3a\u65e0\u6807\u70b9\u53e4\u7c4d\u6587\u732e\u586b\u5145\u73b0\u4ee3\u6807\u70b9\uff0c\u8bc4\u4ef7\u4e3a\u4f9d\u636e\u6a21\u578b\u5bf9\u6d4b\u8bd5\u96c6\u53e4\u7c4d\u7684\u65ad\u53e5\u53ca\u6807\u70b9\u9884\u6d4b\u4e0e\u51fa\u7248\u53e4\u7c4d\u6807\u70b9\u7684\u5339\u914d\u7a0b\u5ea6\u3002\u6240\u4ee5\u91c7\u7528\u7cbe\u786e\u7387(precision)\u3001\u53ec\u56de\u7387(recall)\u4ee5\u53caF1\u503c\u4f5c\u4e3a\u8bc4\u4ef7\u6a21\u578b\u6807\u70b9\u53ca\u65ad\u53e5\u9884\u6d4b\u7684\u6548\u679c\u6307\u6807:<\/p>\n\n\n\n

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

\u5176\u4e2d\uff0cTP\u3001FP\u548cFN\u6e90\u4e8e\u6df7\u6dc6\u77e9\u9635\u3002TP\u4e3a\u6a21\u578b\u9884\u6d4b\u4e0e\u539f\u6587\u4e2d\u5b8c\u5168\u5339\u914d\u7684\u6807\u70b9\u6570\u91cf\uff0cFP\u4e3a\u6a21\u578b\u9884\u6d4b\u51fa\u6765\u4f46\u539f\u6807\u6ce8\u4e2d\u6ca1\u6709\u6807\u70b9\u7684\u6570\u91cf\uff0cFN\u4e3a\u539f\u6587\u4e2d\u6709\u4f46\u6a21\u578b\u6ca1\u6709\u9884\u6d4b\u51fa\u6765\u7684\u6807\u70b9\u6570\u91cf\u3002TP+FP\u7b49\u4e8e\u6a21\u578b\u9884\u6d4b\u51fa\u6765\u6807\u70b9\u7684\u603b\u6570\u91cf\uff0cTP+FN\u7b49\u4e8e\u539f\u6587\u4e2d\u5b58\u5728\u6807\u70b9\u7684\u603b\u6570\u91cf\u3002<\/p>\n\n\n\n

\u6b64\u5916\uff0c\u5206\u7c7b\u9884\u6d4b\u7ed3\u679c\u4e3a\u6839\u636e\u56db\u90e8\u7c7b\u5206\u7c7b\u5c06\u6d4b\u8bd5\u6587\u4ef6\u6574\u5408\u4e3a\u7ecf\u3001\u53f2\u3001\u5b50\u3001\u96c6\u5171\u56db\u4e2a\u6587\u4ef6\uff0c\u5bf9\u6bcf\u4e2a\u6587\u4ef6\u8fdb\u884c\u9884\u6d4b\uff0c\u5e76\u6839\u636e\u8be5\u6587\u4ef6\u4e2d\u6807\u6ce8\u6b63\u786e\u6216\u9519\u8bef\u7684\u6807\u70b9\u6570\u636e\u6309\u516c\u5f0f(6)\u8fdb\u884cF1\u503c\u8ba1\u7b97\u3002<\/p>\n\n\n\n

(\u56db)\u5b9e\u9a8c\u53ca\u7ed3\u679c<\/span><\/p>\n\n\n\n

\u672c\u6587\u4f7f\u7528\u4e86\u76f8\u540c\u7684\u6570\u636e\u96c6\u5bf9\u57fa\u4e8eRNN\u7684\u53cc\u5411\u957f\u77ed\u65f6\u8bb0\u5fc6\u6a21\u578b(Bi-Long Short Term Memory, LSTM)[21]<\/sup>\u548cBERT+CNN\u6a21\u578b\u8fdb\u884c\u4e86\u8bad\u7ec3\uff0c\u5e76\u4f7f\u7528\u76f8\u540c\u7684\u6d4b\u8bd5\u96c6\u5bf9\u5176\u8fdb\u884c\u6d4b\u8bd5\u3002\u4e09\u79cd\u6a21\u578b\u5bf9\u6d4b\u8bd5\u96c6\u6807\u70b9\u53ca\u65ad\u53e5\u9884\u6d4b\u7684\u7ed3\u679c\u5bf9\u6bd4\uff0c\u5982\u88683\u6240\u793a\u3002<\/p>\n\n\n\n

\"\"
\u88683 \u6a21\u578b\u5bf9\u6bd4\u7ed3\u679c<\/figcaption><\/figure><\/div>\n\n\n\n

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

\u7531\u88683\u4e2d\u7684\u7ed3\u679c\u53ef\u4ee5\u770b\u51fa\uff0cTransformer\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c\u5728\u56db\u90e8\u7c7b\u7684\u6807\u70b9\u548c\u65ad\u53e5\u4efb\u52a1\u4e0a\u90fd\u83b7\u5f97\u4e86\u6700\u597d\u7684\u6210\u7ee9(\u5728\u7ecf\u7c7b\u6807\u70b9\u4e0a\u7684\u51c6\u786e\u5ea6\u4e0eBERT+CNN\u6a21\u578b\u540c\u4e3a\u7b2c\u4e00)\u3002\u5728\u9a8c\u8bc1\u96c6\u4e0a\uff0cTransformer\u6a21\u578b\u7684\u6807\u70b9F1\u503c\u4e3a86.5%\uff0c\u65ad\u53e5F1\u503c\u4e3a95.1%\uff0cBERT\u6a21\u578b\u7684\u6807\u70b9F1\u503c\u4e3a84.7%\uff0c\u65ad\u53e5F1\u503c\u4e3a93.8%\u3002\u6211\u4eec\u6ca1\u6709\u6d4b\u8bd5\u53cc\u5411LSTM\u6a21\u578b\u5728\u9a8c\u8bc1\u96c6\u4e0a\u7684\u51c6\u786e\u5ea6\u3002<\/p>\n\n\n\n

\u4ece\u5206\u7c7b\u7ed3\u679c\u4e0a\u6765\u770b\uff0c\u96c6\u90e8\u7684\u9884\u6d4b\u7ed3\u679c\u5728\u4e09\u7ec4\u6a21\u578b\u7684\u6807\u70b9\u4efb\u52a1\u4e2d\u90fd\u9ad8\u4e8e\u5176\u4ed6\u5206\u7c7b\u7684\u9884\u6d4b\u7ed3\u679c\uff0c\u7ecf\u90e8\u7684\u9884\u6d4b\u7ed3\u679c\u5728\u4e09\u7ec4\u6a21\u578b\u7684\u6807\u70b9\u3001\u65ad\u53e5\u4efb\u52a1\u4e2d\u90fd\u666e\u904d\u4f4e\u4e8e\u5176\u4ed6\u5206\u7c7b\u7684\u9884\u6d4b\u7ed3\u679c\u3002\u7ed3\u5408\u8bad\u7ec3\u6570\u636e\u96c6\u5206\u5e03\u4e2d\uff0c\u96c6\u90e8\u7684\u8bad\u7ec3\u6570\u636e\u6700\u591a\uff0c\u800c\u7ecf\u90e8\u7684\u8bad\u7ec3\u6570\u636e\u5c11\u4e8e\u5176\u4ed6\u5206\u7c7b\u7684\u60c5\u51b5\u6765\u770b\uff0c\u53ef\u4ee5\u63a8\u6d4b\uff0c\u589e\u52a0\u5176\u4ed6\u5206\u7c7b\u7684\u8bad\u7ec3\u6570\u636e\u53ef\u4ee5\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6a21\u578b\u5bf9\u6587\u672c\u9884\u6d4b\u7684F1\u7ed3\u679c\u3002\u7ecf\u90e8\u7ed3\u679c\u8f83\u5dee\uff0c\u4e3b\u8981\u4e5f\u548c\u6240\u9009\u62e9\u7684\u7ecf\u90e8\u6d4b\u8bd5\u6570\u636e\u6837\u672c\u8f83\u5c11(\u53ea\u6709\u4e00\u672c)\u4e14\u504f\u5411\u767d\u8bdd\u6587\u672c\u7c7b\u578b\u6709\u5173\u3002Transformer\u6a21\u578b\u5728\u53f2\u3001\u5b50\u3001\u96c6\u7c7b\u6d4b\u8bd5\u96c6\u4e0a\u6807\u70b9\u4efb\u52a1\u7684F1\u51c6\u786e\u7387\u9886\u5148BERT+CNN\u6a21\u578b\u7684\u5e45\u5ea6\u6bd4\u5728\u9a8c\u8bc1\u96c6\u4e0a\u7684\u9886\u5148\u5e45\u5ea6\u7a0d\u9ad8\u6216\u6301\u5e73\uff0c\u8fd9\u53ef\u80fd\u4e5f\u8bf4\u660e\u7aef\u5230\u7aef\u6a21\u578b\u7684\u6cdb\u5316\u6027\u80fd\u4f1a\u6bd4\u5355\u7f16\u7801\u5668\u7c7b\u578b\u7684\u6a21\u578b\u66f4\u597d\u3002<\/p>\n\n\n\n

\u6b64\u5916\uff0c\u6211\u4eec\u5bf9Transformer\u6a21\u578b\u9884\u6d4b\u51fa\u6765\u7684\u9519\u8bef\u6570\u636e\u505a\u4e86\u8fdb\u4e00\u6b65\u5206\u6790\u3002\u901a\u8fc7\u5bf9\u6bd4\u6570\u636e\u4e2d\u7684\u539f\u59cb\u6807\u6ce8\u548c\u9884\u6d4b\u6807\u6ce8\uff0c\u67e5\u770b\u6a21\u578b\u4e3b\u8981\u5bf9\u54ea\u4e9b\u6807\u70b9\u8fdb\u884c\u4e86\u9519\u8bef\u6807\u6ce8\uff0c\u7ed3\u679c\u5982\u88684\u6240\u793a\uff0c\u5176\u4e2d\u7684\u6570\u503c\u8868\u793a\u6bcf\u4e2a\u539f\u59cb\u6807\u70b9(\u5de6\u5217)\u88ab\u9519\u8bef\u6807\u6ce8\u6210\u67d0\u4e00\u7c7b\u6807\u70b9\u7684\u6982\u7387\u3002\u4ece\u6df7\u6dc6\u77e9\u9635\u7684\u7ed3\u679c\u6765\u770b\uff0c\u6a21\u578b\u66f4\u591a\u7684\u9519\u8bef(\u5360\u6bd4\u572850%\u4ee5\u4e0a\uff0c\u89c1\u52a0\u7c97\u6570\u636e\u90e8\u5206)\u51fa\u73b0\u5728:<\/p>\n\n\n\n

(1)\u5c06\u9017\u53f7\u3001\u987f\u53f7\u548c\u5192\u53f7\u9519\u8bef\u5730\u9884\u6d4b\u6210\u4e86\u65e0\u6807\u70b9\u7b26\u53f7;<\/p>\n\n\n\n

(2)\u9017\u53f7\u548c\u53e5\u53f7\u4e4b\u95f4\u4ea7\u751f\u4e92\u6807\uff0c\u51fa\u73b0\u65ad\u53e5F1\u5f97\u5206\u5927\u5e45\u9ad8\u4e8e\u6807\u70b9\u5f97\u5206\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

\"\"
\u88684 \u6a21\u578b\u7ed3\u679c\u6df7\u6dc6\u77e9\u9635<\/figcaption><\/figure><\/div>\n\n\n\n

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

Transformer\u6a21\u578b\u4f8b\u53e5\u793a\u4f8b\u5982\u88685\u6240\u793a\uff0c\u5305\u62ec\u7ecf\u53f2\u5b50\u96c6\u6d4b\u8bd5\u96c6\u4e2d\u6311\u9009\u51fa\u7684\u539f\u6587\u3001\u6a21\u578b\u9884\u6d4b\u7ed3\u679c\u548c\u901a\u8fc7\u516c\u5f0f(6)\u8ba1\u7b97\u5f97\u5230\u7684\u65ad\u53e5\u548c\u6807\u70b9\u7684F1\u503c\u3002<\/p>\n\n\n\n

\"\"
\u88685 \u7ed3\u679c\u5c55\u793a\u4f8b\u53e5<\/figcaption><\/figure><\/div>\n\n\n\n

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

\u603b \u7ed3<\/span><\/strong><\/p>\n\n\n\n

\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8eTransformer\u7684\u53e4\u7c4d\u81ea\u52a8\u6807\u70b9\u6a21\u578b\uff0c\u53ef\u4ee5\u6709\u6548\u5229\u7528\u8bed\u6599\u4fe1\u606f\u8fdb\u884c\u5bf9\u65e0\u6807\u70b9\u6587\u672c\u7684\u6807\u70b9(\u70b9\u53f7)\u53ca\u65ad\u53e5\u9884\u6d4b\u3002\u70b9\u53f7\u7684\u6807\u6ce8\u4e5f\u53ef\u4ee5\u88ab\u8f6c\u5316\u4e3a\u81ea\u7136\u8bed\u8a00\u95ee\u9898\u4e2d\u7684\u5e8f\u5217\u6807\u6ce8\u95ee\u9898\u3002\u73b0\u6709\u6a21\u578b\u5728\u9a8c\u8bc1\u96c6\u4e0a\uff0c\u83b7\u5f97\u4e86\u76ee\u524d\u6211\u4eec\u5df2\u77e5\u65b9\u6848\u4e2d\u6700\u9ad8\u7684\u51c6\u786e\u5ea6(\u867d\u7136\u4e0d\u540c\u96be\u5ea6\u7684\u8bad\u7ec3\u6587\u672c\u7c7b\u578b\u4f1a\u5bfc\u81f4\u9a8c\u8bc1\u96c6\u7ed3\u679c\u5b58\u5728\u5dee\u5f02)\uff0c\u5e76\u5728\u62bd\u53d6\u7684\u6d4b\u8bd5\u96c6\u4e0a\u5c55\u793a\u51fa\u4e86\u826f\u597d\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u96c6\u90e8\u7ed3\u679c\u66f4\u4f18\u3002\u4e0d\u8fc7\uff0c\u5927\u91cf\u6570\u636e\u5728\u63d0\u5347\u6a21\u578b\u5b66\u4e60\u80fd\u529b\u548c\u9884\u6d4b\u7ed3\u679c\u7684\u540c\u65f6\uff0c\u4e5f\u56e0\u4e0d\u540c\u4e66\u76ee\u5b58\u5728\u4e0d\u540c\u7f16\u8f91\uff0c\u6240\u4ee5\u53ef\u80fd\u5b58\u5728\u540c\u4e00\u53e5\u8bdd\u6709\u4e0d\u540c\u6807\u70b9\u6807\u6ce8\u7684\u65b9\u5f0f\uff0c\u4ece\u800c\u4f1a\u5f71\u54cd\u6a21\u578b\u6574\u4f53\u7684\u9884\u6d4b\u51c6\u786e\u7387\u4e0a\u9650\u3002<\/p>\n\n\n\n

\u5e76\u4e14\uff0c\u6839\u636e\u5b9e\u9645\u4f7f\u7528\u60c5\u51b5\u5f97\u5230\u7684\u53cd\u9988\u53ef\u77e5\uff0c\u6a21\u578b\u5bf9\u300a\u9ec4\u5e1d\u5185\u7ecf\u300b\u7b49\u7ecf\u5178\u53e4\u7c4d\u9884\u6d4b\u7ed3\u679c\u6b63\u786e\u7387\u6781\u9ad8;\u4ece\u5bf9\u5176\u4ed6\u53e4\u7c4d\u6587\u672c\u7684\u9884\u6d4b\u7ed3\u679c\u770b\uff0c\u6a21\u578b\u5bf9\u4e66\u540d\u548c\u5b98\u804c\u7b49\u90e8\u5206\u7684\u65ad\u53e5\u90fd\u5341\u5206\u51c6\u786e\u3002\u901a\u8fc7\u6a21\u578b\u7684\u5e73\u5747\u6210\u7ee9\u548c\u5b9e\u9645\u4f7f\u7528\u7684\u53cd\u9988\u7ed3\u679c\u53ef\u4ee5\u770b\u51fa\uff0cTransformer\u6a21\u578b\u57fa\u672c\u5b66\u5230\u4e86\u8bad\u7ec3\u96c6\u6240\u5305\u542b\u7684\u5927\u90e8\u5206\u8bed\u4e49\u4fe1\u606f\u548c\u73b0\u4ee3\u4eba\u5bf9\u53e4\u5178\u6587\u8a00\u6587\u6807\u6ce8\u7684\u89c4\u5219\u65b9\u6cd5\uff0c\u4f46\u5bf9\u8fd1\u73b0\u4ee3\u8bed\u6599\u8fdb\u884c\u6807\u70b9\u7684\u8868\u73b0\u4e0a\u8fd8\u662f\u6709\u4e0a\u5347\u7a7a\u95f4\u3002<\/p>\n\n\n\n

\u5728\u672a\u6765\u7684\u5de5\u4f5c\u4e2d\uff0c\u5c06\u8003\u8651\u6e05\u6d17\u8bad\u7ec3\u96c6\u4e2d\u76f8\u540c\u53e5\u5b50\u4e0d\u540c\u6807\u70b9\u6807\u6ce8\u7684\u60c5\u51b5\uff0c\u5e76\u52a0\u5165\u66f4\u591a\u8fd1\u73b0\u4ee3\u8bad\u7ec3\u8bed\u6599\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c\u3002<\/p>\n\n\n\n

———————————————————————————————————————————————————————————————————<\/span> <\/p>\n\n\n\n

An Automatic Punctuation Method Based on the Transformer Model<\/span><\/strong><\/p>\n\n\n\n

Hong Tao, Cheng Ruixue, Liu Sixi, Fang Kaiqi<\/p>\n\n\n\n

Abstract: <\/span><\/strong>In order to facilitate modern people’s understanding and learning of ancient books and documents, people begin to add punctuations in ancient scriptures manually, which means modern punctuations will be used in ancient scriptures without punctuations. But the number of ancient scriptures is huge and manual annotation is a laborious task. With the rise of artificial intelligence (AI), we can realize automatic punctuation based on deep learning. Therefore, the burden of manual annotation can be lifted. The Transformer model based on multi-attention mechanism is used as the automatic punctuation model to train 1 billion words from published ancient books. The model can achieve 86.5% on punctuation F1 and 95.1% on segmentation F1 on the validation set. Then 19 books were selected from the untrained corpus as the test set, and the results show that the model is good at classical Chinese corpus, but has low validity for more professional and vernacular corpus.<\/p>\n\n\n\n

Keywords:<\/span> <\/strong>Ancient Scriptures; Natural Language Processing; Automatic Punctuation; Machine Learning; Deep Learning<\/p>\n\n\n\n

———————————————————————————————————————————————————————————————————<\/span><\/p>\n\n\n\n

\u7f16 \u8f91  |  \u4e25\u7a0b<\/p>\n\n\n\n

\u6ce8\u91ca\uff1a<\/strong><\/p>\n\n\n\n

[1]\u90d1\u6c38\u6653:\u300a\u53e4\u7c4d\u6570\u5b57\u5316\u4e0e\u53e4\u5178\u6587\u5b66\u7814\u7a76\u7684\u672a\u6765\u300b\uff0c\u300a\u6587\u5b66\u9057\u4ea7\u300b2005\u5e74\u7b2c5\u671f\u3002<\/p>\n\n\n\n

[2]\u9ec4\u5efa\u5e74\u3001\u4faf\u6c49\u6e05:\u300a\u519c\u4e1a\u53e4\u7c4d\u65ad\u53e5\u6807\u70b9\u6a21\u5f0f\u7814\u7a76\u300b\uff0c\u300a\u4e2d\u6587\u4fe1\u606f\u5b66\u62a5\u300b2008\u5e74\u7b2c4\u671f\u3002<\/p>\n\n\n\n

[3]\u674e\u94ce\u3001\u738b\u6bc5:\u300a\u5173\u4e8e\u53e4\u4ee3\u6587\u732e\u4fe1\u606f\u5316\u5de5\u7a0b\u4e0e\u53e4\u5178\u6587\u5b66\u7814\u7a76\u4e4b\u95f4\u4e92\u52a8\u5173\u7cfb\u7684\u5bf9\u8bdd\u300b\uff0c\u300a\u6587\u5b66\u9057\u4ea7\u300b2005\u5e74\u7b2c1\u671f\u3002<\/p>\n\n\n\n

[4]\u5f20\u5f00\u65ed\u3001\u590f\u4e91\u5e86\u3001\u5b87\u822a:\u300a\u57fa\u4e8e\u6761\u4ef6\u968f\u673a\u573a\u7684\u53e4\u6c49\u8bed\u81ea\u52a8\u65ad\u53e5\u4e0e\u6807\u70b9\u65b9\u6cd5\u300b\uff0c\u300a\u6e05\u534e\u5927\u5b66\u5b66\u62a5\u300b(\u81ea\u7136\u79d1\u5b66\u7248)2009\u5e74\u7b2c10\u671f\u3002<\/p>\n\n\n\n

[5]J. Lafferty, A. McCallum, F. Pereira, \u201cConditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data,\u201d Proceedings 18th International Conference on Machine Learning, USA: Morgan Kaufmann, 2001, pp. 282-289.<\/p>\n\n\n\n

[6]Z. Huang, W. Xu, K. Yu\u201c, Bidirectional LSTM-CRF Models for Sequence Tagging,\u201darXiv:1508.01991,2015.<\/p>\n\n\n\n

[7]M. Schuster, Kuldip K.Paliwal, \u201cBidirectional Recurrent Neural Networks,\u201d IEEE Jounals<\/em>, vol. 45, no. 11, 1997, pp. 2673-2681.<\/p>\n\n\n\n

[8]\u4fde\u656c\u677e\u3001\u9b4f\u4e00\u3001\u5f20\u6c38\u4f1f:\u300a\u57fa\u4e8eBERT\u7684\u53e4\u6587\u65ad\u53e5\u7814\u7a76\u4e0e\u5e94\u7528\u300b\uff0c\u300a\u4e2d\u6587\u4fe1\u606f\u5b66\u62a5\u300b2019\u5e74\u7b2c11\u671f\u3002<\/p>\n\n\n\n

[9]Jacob Devlin et al.\u201c, BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,\u201d NAACL, USA: Minneapolis, 2019, pp. 4171-4186.<\/p>\n\n\n\n

[10]A. Vaswani et al.\u201c, Attention is All You Need,\u201d Advances in Neural Information Processing Systems, USA: Long Beach, 2017, pp. 5998-6008.<\/p>\n\n\n\n

[11]A. Krizhevsky, I. Sutskever, G. Hinton,\u201c ImageNet Classification with Deep Convolutional Neural Networks,\u201d In Advances in Neural Information Processing Systems, 2012, pp.1097-1105.<\/p>\n\n\n\n

[12]\u4e2d\u534e\u4eba\u6c11\u5171\u548c\u56fd\u56fd\u5bb6\u8d28\u91cf\u76d1\u7763\u68c0\u9a8c\u68c0\u75ab\u603b\u5c40\u4e2d\u56fd\u56fd\u5bb6\u6807\u51c6\u5316\u7ba1\u7406\u59d4\u5458\u4f1a:\u300aGB\/T15834\u20152011\u6807\u70b9\u7b26\u53f7\u7528\u6cd5\u300b\uff0c\u5317\u4eac:\u4e2d\u56fd\u6807\u51c6\u51fa\u7248\u793e\uff0c2012\u5e74\u3002<\/p>\n\n\n\n

[13]Ilya Sutskever, Oriol Vinyals, Quoc V. Le, \u201cSequence to Sequence Learning with Neural Networks,\u201d NIPS’14<\/em>, vol. 2, 2014, pp. 3104-3112.<\/p>\n\n\n\n

[14]T. Mikolov et al., \u201cEfficient Estimation of Word Representations in Vector Space,\u201d Computer Science<\/em>, 2013, arxiv.org\/abs\/1301.3781v3.<\/p>\n\n\n\n

[15]S. Aroca-Ouellette, F. Rudzicz,\u201cOn Losses for Modern Language Models,\u201d Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), October 2020, arXiv preprint arXiv:2010.01694; Y. Liu et al., \u201cRoBERTa: A Robustly Optimized BERT Pretrainin Approach,\u201d arXiv:1907.11692.<\/p>\n\n\n\n

[16]http:\/\/publish.ancientbooks.cn\/docShuju\/platformSublibIndex.jspx?libId=6.<\/p>\n\n\n\n

[17]https:\/\/github.com\/pytorch\/fairseq.<\/p>\n\n\n\n

[18]Liyuan Liu et al., \u201cUnderstanding the Difficulty of Training Transformers,\u201d Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP 2020), https:\/\/www.aclweb. org\/anthology\/2020.emnlp-main.463.pdf.<\/p>\n\n\n\n

[19]Liyuan Liu et al.\u201c, On the Variance of the Adaptive Learning Rate and Beyond,\u201darXiv:1908.03265.<\/p>\n\n\n\n

[20]http:\/\/www.daizhige.org\/.<\/p>\n\n\n\n

[21]S. Hochreiter, J. Schmidhuber,\u201cLong Short-term Memory,\u201d Neural Computation<\/em>, vol. 9, no. 8, 1997, pp. 1735-1780.<\/p>\n\n\n\n

\u539f\u520a\u300a\u6570\u5b57\u4eba\u6587\u300b2021\u5e74\u7b2c2\u671f\uff0c\u8f6c\u8f7d\u8bf7\u8054\u7cfb\u6388\u6743\u3002<\/p>","protected":false},"excerpt":{"rendered":"

\u4eba\u5de5\u52a0\u6ce8\u53e4\u7c4d\u6587\u732e\u6807\u70b9\u7b26\u53f7\uff0c\u5373\u5bf9\u65e0\u6807\u70b9\u7b26\u53f7\u7684\u53e4\u7c4d\u6587\u732e\u8fdb\u884c\u73b0\u4ee3\u6807\u70b9\u7b26\u53f7\u586b\u5145\u3002\u4f46\u53e4\u7c4d\u6570\u2026\u2026<\/p>","protected":false},"author":3,"featured_media":14238,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","_links_to":"","_links_to_target":""},"categories":[2892],"tags":[2893,292,50,672,347,669,313],"class_list":["post-14235","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-2892","tag-2893","tag-292","tag-50","tag-672","tag-347","tag-669","tag-313"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":6}},"_links":{"self":[{"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/posts\/14235","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/comments?post=14235"}],"version-history":[{"count":2,"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/posts\/14235\/revisions"}],"predecessor-version":[{"id":14249,"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/posts\/14235\/revisions\/14249"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/media\/14238"}],"wp:attachment":[{"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/media?parent=14235"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/categories?post=14235"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/achieve.dhcn.cn\/en\/wp-json\/wp\/v2\/tags?post=14235"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}