{"id":60873,"date":"2026-01-16T14:17:33","date_gmt":"2026-01-16T06:17:33","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/60873.html"},"modified":"2026-01-16T14:17:33","modified_gmt":"2026-01-16T06:17:33","slug":"%e8%87%aa%e7%84%b6%e8%af%ad%e8%a8%80%e5%a4%84%e7%90%86%e5%ae%9e%e6%88%98-%e6%9e%84%e5%bb%ba%e6%99%ba%e8%83%bd%e5%af%b9%e8%af%9d%e7%b3%bb%e7%bb%9f%ef%bc%88bertseq2seq%e6%9e%b6","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/60873.html","title":{"rendered":"\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\uff08BERT+Seq2Seq\u67b6\u6784\uff09"},"content":{"rendered":"

\u7b2c8\u7bc7：\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf（BERT+Seq2Seq\u67b6\u6784）<\/h2>\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

\u5b66\u4e60\u76ee\u6807<\/h3>\n
  • \u7406\u89e3\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u6838\u5fc3\u539f\u7406\u4e0e\u67b6\u6784<\/li>\n
  • \u638c\u63e1BERT\u6a21\u578b\u5728\u6587\u672c\u7406\u89e3\u4e2d\u7684\u5e94\u7528\u65b9\u6cd5<\/li>\n
  • \u5b66\u4f1a\u4f7f\u7528Seq2Seq\u6a21\u578b\u5b9e\u73b0\u6587\u672c\u751f\u6210\u529f\u80fd<\/li>\n
  • \u7406\u89e3Attention\u673a\u5236\u5728\u5bf9\u8bdd\u7cfb\u7edf\u4e2d\u7684\u91cd\u8981\u6027<\/li>\n
  • \u80fd\u591f\u72ec\u7acb\u5b8c\u6210\u4e00\u4e2a\u57fa\u4e8eBERT+Seq2Seq\u67b6\u6784\u7684\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf<\/li>\n

    \u7ae0\u8282\u91cd\u70b9<\/h3>\n
  • \u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u6982\u8ff0<\/li>\n
  • BERT\u6a21\u578b\u539f\u7406\u4e0e\u5e94\u7528<\/li>\n
  • Seq2Seq\u6a21\u578b\u4e0eAttention\u673a\u5236<\/li>\n
  • \u6570\u636e\u96c6\u51c6\u5907\u4e0e\u9884\u5904\u7406<\/li>\n
  • \u6a21\u578b\u8bad\u7ec3\u4e0e\u4f18\u5316<\/li>\n
  • \u6a21\u578b\u90e8\u7f72\u4e0e\u6d4b\u8bd5<\/li>\n
  • \u6848\u4f8b\u5206\u6790\u4e0e\u4f18\u5316\u601d\u8def<\/li>\n
    \n

    \u4e00\u3001\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u6982\u8ff0<\/h3>\n

    1.1 \u4ec0\u4e48\u662f\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf<\/h4>\n

    💡 \u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u662f\u4e00\u79cd\u80fd\u591f\u901a\u8fc7\u81ea\u7136\u8bed\u8a00\u4e0e\u7528\u6237\u8fdb\u884c\u4ea4\u4e92\u7684\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf，\u5b83\u53ef\u4ee5\u7406\u89e3\u7528\u6237\u7684\u610f\u56fe，\u63d0\u4f9b\u76f8\u5173\u7684\u4fe1\u606f\u6216\u5b8c\u6210\u7279\u5b9a\u7684\u4efb\u52a1\u3002\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u901a\u5e38\u5206\u4e3a\u4e24\u7c7b：<\/p>\n

      \n
    • \u95ee\u7b54\u7cfb\u7edf：\u56de\u7b54\u7528\u6237\u7684\u7279\u5b9a\u95ee\u9898，\u5982\u767e\u79d1\u77e5\u8bc6\u95ee\u7b54\u3001\u6280\u672f\u652f\u6301\u95ee\u7b54\u7b49\u3002<\/li>\n
    • \u804a\u5929\u673a\u5668\u4eba：\u8fdb\u884c\u5f00\u653e\u5f0f\u7684\u5bf9\u8bdd，\u5982\u793e\u4ea4\u804a\u5929\u3001\u60c5\u611f\u966a\u4f34\u7b49\u3002<\/li>\n<\/ul>\n

      1.2 \u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u6838\u5fc3\u6280\u672f<\/h4>\n

      \u26a0\ufe0f \u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u6838\u5fc3\u6280\u672f\u5305\u62ec\u81ea\u7136\u8bed\u8a00\u7406\u89e3（NLU）\u3001\u81ea\u7136\u8bed\u8a00\u751f\u6210（NLG）\u548c\u5bf9\u8bdd\u7ba1\u7406（DM）\u3002<\/p>\n

        \n
      • \u81ea\u7136\u8bed\u8a00\u7406\u89e3：\u5c06\u7528\u6237\u7684\u81ea\u7136\u8bed\u8a00\u8f93\u5165\u8f6c\u6362\u4e3a\u673a\u5668\u53ef\u7406\u89e3\u7684\u8bed\u4e49\u8868\u793a\u3002<\/li>\n
      • \u81ea\u7136\u8bed\u8a00\u751f\u6210：\u5c06\u673a\u5668\u7684\u8bed\u4e49\u8868\u793a\u8f6c\u6362\u4e3a\u81ea\u7136\u8bed\u8a00\u8f93\u51fa\u3002<\/li>\n
      • \u5bf9\u8bdd\u7ba1\u7406：\u8d1f\u8d23\u5bf9\u8bdd\u7684\u4e0a\u4e0b\u6587\u7406\u89e3\u548c\u72b6\u6001\u7ba1\u7406，\u51b3\u5b9a\u4e0b\u4e00\u6b65\u7684\u56de\u590d\u7b56\u7565\u3002<\/li>\n<\/ul>\n

        1.3 \u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u67b6\u6784<\/h4>\n

        \u2705 \u5e38\u89c1\u7684\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u67b6\u6784\u5206\u4e3a\u4e24\u7c7b：<\/p>\n

          \n
        • \u7ba1\u9053\u5f0f\u67b6\u6784：\u5c06NLU\u3001DM\u548cNLG\u5206\u5f00\u5904\u7406，\u6bcf\u4e2a\u7ec4\u4ef6\u72ec\u7acb\u5de5\u4f5c\u3002<\/li>\n
        • \u7aef\u5230\u7aef\u67b6\u6784：\u4f7f\u7528\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u76f4\u63a5\u5c06\u7528\u6237\u8f93\u5165\u8f6c\u6362\u4e3a\u7cfb\u7edf\u8f93\u51fa，\u65e0\u9700\u4eba\u5de5\u8bbe\u8ba1\u7684\u4e2d\u95f4\u8868\u793a\u3002<\/li>\n<\/ul>\n
          \n

          \u4e8c\u3001BERT\u6a21\u578b\u539f\u7406\u4e0e\u5e94\u7528<\/h3>\n

          2.1 BERT\u6a21\u578b\u6982\u8ff0<\/h4>\n

          💡 BERT（Bidirectional Encoder Representations from Transformers）\u662f\u4e00\u79cd\u57fa\u4e8eTransformer\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b，\u7531Google\u57282018\u5e74\u63d0\u51fa\u3002BERT\u901a\u8fc7\u53cc\u5411\u4e0a\u4e0b\u6587\u7406\u89e3\u6587\u672c\u7684\u8bed\u4e49\u4fe1\u606f，\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e2d\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6210\u679c\u3002<\/p>\n

          2.2 BERT\u7684\u9884\u8bad\u7ec3\u4efb\u52a1<\/h4>\n

          \u26a0\ufe0f BERT\u7684\u9884\u8bad\u7ec3\u4efb\u52a1\u5305\u62ec\u4e24\u4e2a\u90e8\u5206：<\/p>\n

        • \u63a9\u7801\u8bed\u8a00\u6a21\u578b（Masked Language Model，MLM）：\u968f\u673a\u906e\u6321\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u90e8\u5206\u8bcd，\u7136\u540e\u8ba9\u6a21\u578b\u9884\u6d4b\u8fd9\u4e9b\u88ab\u906e\u6321\u7684\u8bcd\u3002<\/li>\n
        • \u4e0b\u4e00\u53e5\u9884\u6d4b（Next Sentence Prediction，NSP）：\u5224\u65ad\u4e24\u4e2a\u53e5\u5b50\u662f\u5426\u662f\u8fde\u7eed\u7684\u4e0a\u4e0b\u6587\u5173\u7cfb\u3002<\/li>\n

          2.3 BERT\u5728\u6587\u672c\u7406\u89e3\u4e2d\u7684\u5e94\u7528<\/h4>\n

          \u2705 BERT\u5728\u6587\u672c\u7406\u89e3\u4efb\u52a1\u4e2d\u7684\u5e94\u7528\u6b65\u9aa4\u5982\u4e0b：<\/p>\n

        • \u5c06\u6587\u672c\u8f93\u5165\u8f6c\u6362\u4e3aBERT\u53ef\u63a5\u53d7\u7684\u683c\u5f0f\u3002<\/li>\n
        • \u8c03\u7528BERT\u6a21\u578b\u83b7\u53d6\u6587\u672c\u7684\u8bed\u4e49\u8868\u793a\u3002<\/li>\n
        • \u5728BERT\u7684\u8f93\u51fa\u57fa\u7840\u4e0a\u6dfb\u52a0\u7279\u5b9a\u4efb\u52a1\u7684\u5934（\u5982\u5206\u7c7b\u5934\u3001\u56de\u5f52\u5934\u7b49）\u3002<\/li>\n
        • \u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03，\u4ee5\u9002\u5e94\u7279\u5b9a\u4efb\u52a1\u3002<\/li>\n

          2.4 \u4ee3\u7801\u5b9e\u73b0：\u4f7f\u7528BERT\u8fdb\u884c\u6587\u672c\u5206\u7c7b<\/h4>\n

          import<\/span> torch
          \nfrom<\/span> transformers import<\/span> BertTokenizer,<\/span> BertForSequenceClassification<\/p>\n

          # \u52a0\u8f7dBERT\u5206\u8bcd\u5668\u548c\u6a21\u578b<\/span>
          \ntokenizer =<\/span> BertTokenizer.<\/span>from_pretrained(<\/span>'bert-base-chinese'<\/span>)<\/span>
          \nmodel =<\/span> BertForSequenceClassification.<\/span>from_pretrained(<\/span>'bert-base-chinese'<\/span>,<\/span> num_labels=<\/span>2<\/span>)<\/span><\/p>\n

          # \u6587\u672c\u8f93\u5165<\/span>
          \ntext =<\/span> "\u8fd9\u662f\u4e00\u4e2a\u6d4b\u8bd5\u6587\u672c"<\/span><\/p>\n

          # \u6587\u672c\u9884\u5904\u7406<\/span>
          \ninputs =<\/span> tokenizer(<\/span>text,<\/span> return_tensors=<\/span>'pt'<\/span>,<\/span> padding=<\/span>True<\/span>,<\/span> truncation=<\/span>True<\/span>,<\/span> max_length=<\/span>512<\/span>)<\/span><\/p>\n

          # \u6a21\u578b\u63a8\u7406<\/span>
          \nwith<\/span> torch.<\/span>no_grad(<\/span>)<\/span>:<\/span>
          \n outputs =<\/span> model(<\/span>**<\/span>inputs)<\/span><\/p>\n

          # \u83b7\u53d6\u9884\u6d4b\u7ed3\u679c<\/span>
          \nlogits =<\/span> outputs.<\/span>logits
          \npredicted_label =<\/span> torch.<\/span>argmax(<\/span>logits,<\/span> dim=<\/span>1<\/span>)<\/span>.<\/span>item(<\/span>)<\/span><\/p>\n

          print<\/span>(<\/span>f"\u9884\u6d4b\u6807\u7b7e：<\/span>{<\/span>predicted_label}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

          \n

          \u4e09\u3001Seq2Seq\u6a21\u578b\u4e0eAttention\u673a\u5236<\/h3>\n

          3.1 Seq2Seq\u6a21\u578b\u6982\u8ff0<\/h4>\n

          💡 Seq2Seq（Sequence to Sequence）\u6a21\u578b\u662f\u4e00\u79cd\u7528\u4e8e\u5904\u7406\u5e8f\u5217\u6570\u636e\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b，\u7531\u7f16\u7801\u5668（Encoder）\u548c\u89e3\u7801\u5668（Decoder）\u7ec4\u6210\u3002\u7f16\u7801\u5668\u5c06\u8f93\u5165\u5e8f\u5217\u8f6c\u6362\u4e3a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\u8868\u793a，\u89e3\u7801\u5668\u6839\u636e\u8be5\u5411\u91cf\u8868\u793a\u751f\u6210\u8f93\u51fa\u5e8f\u5217\u3002<\/p>\n

          3.2 Seq2Seq\u6a21\u578b\u7684\u5c40\u9650\u6027<\/h4>\n

          \u26a0\ufe0f \u4f20\u7edf\u7684Seq2Seq\u6a21\u578b\u5b58\u5728\u4e00\u4e2a\u660e\u663e\u7684\u5c40\u9650\u6027：\u5f53\u8f93\u5165\u5e8f\u5217\u8f83\u957f\u65f6，\u7f16\u7801\u5668\u65e0\u6cd5\u5c06\u6240\u6709\u4fe1\u606f\u538b\u7f29\u5230\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\u4e2d，\u5bfc\u81f4\u89e3\u7801\u5668\u65e0\u6cd5\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u8f93\u51fa\u3002<\/p>\n

          3.3 Attention\u673a\u5236<\/h4>\n

          \u2705 Attention\u673a\u5236\u662f\u4e00\u79cd\u89e3\u51b3Seq2Seq\u6a21\u578b\u5c40\u9650\u6027\u7684\u65b9\u6cd5，\u5b83\u5141\u8bb8\u89e3\u7801\u5668\u5728\u751f\u6210\u6bcf\u4e2a\u8f93\u51fa\u8bcd\u65f6，\u5173\u6ce8\u8f93\u5165\u5e8f\u5217\u4e2d\u4e0e\u8be5\u8bcd\u76f8\u5173\u7684\u90e8\u5206\u3002Attention\u673a\u5236\u7684\u6838\u5fc3\u601d\u60f3\u662f\u8ba1\u7b97\u8f93\u5165\u5e8f\u5217\u4e2d\u6bcf\u4e2a\u4f4d\u7f6e\u7684\u6743\u91cd，\u7136\u540e\u6839\u636e\u6743\u91cd\u5bf9\u8f93\u5165\u5e8f\u5217\u7684\u8868\u793a\u8fdb\u884c\u52a0\u6743\u6c42\u548c\u3002<\/p>\n

          3.4 \u4ee3\u7801\u5b9e\u73b0：\u7b80\u5355\u7684Seq2Seq\u6a21\u578b<\/h4>\n

          import<\/span> torch
          \nimport<\/span> torch.<\/span>nn as<\/span> nn
          \nimport<\/span> torch.<\/span>optim as<\/span> optim<\/p>\n

          # \u5b9a\u4e49Seq2Seq\u6a21\u578b<\/span>
          \nclass<\/span> Seq2Seq<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
          \n def<\/span> __init__<\/span>(<\/span>self,<\/span> input_size,<\/span> hidden_size,<\/span> output_size)<\/span>:<\/span>
          \n super<\/span>(<\/span>Seq2Seq,<\/span> self)<\/span>.<\/span>__init__(<\/span>)<\/span>
          \n self.<\/span>encoder =<\/span> nn.<\/span>LSTM(<\/span>input_size,<\/span> hidden_size,<\/span> batch_first=<\/span>True<\/span>)<\/span>
          \n self.<\/span>decoder =<\/span> nn.<\/span>LSTM(<\/span>hidden_size,<\/span> hidden_size,<\/span> batch_first=<\/span>True<\/span>)<\/span>
          \n self.<\/span>fc =<\/span> nn.<\/span>Linear(<\/span>hidden_size,<\/span> output_size)<\/span><\/p>\n

          def<\/span> forward<\/span>(<\/span>self,<\/span> input_seq,<\/span> target_seq)<\/span>:<\/span>
          \n # \u7f16\u7801\u5668<\/span>
          \n encoder_output,<\/span> (<\/span>encoder_hidden,<\/span> encoder_cell)<\/span> =<\/span> self.<\/span>encoder(<\/span>input_seq)<\/span><\/p>\n

          # \u89e3\u7801\u5668<\/span>
          \n decoder_output,<\/span> (<\/span>decoder_hidden,<\/span> decoder_cell)<\/span> =<\/span> self.<\/span>decoder(<\/span>target_seq,<\/span> (<\/span>encoder_hidden,<\/span> encoder_cell)<\/span>)<\/span><\/p>\n

          # \u8f93\u51fa\u5c42<\/span>
          \n output =<\/span> self.<\/span>fc(<\/span>decoder_output)<\/span><\/p>\n

          return<\/span> output<\/p>\n

          # \u8d85\u53c2\u6570\u8bbe\u7f6e<\/span>
          \ninput_size =<\/span> 10<\/span>
          \nhidden_size =<\/span> 20<\/span>
          \noutput_size =<\/span> 10<\/span>
          \nbatch_size =<\/span> 2<\/span>
          \nseq_length =<\/span> 5<\/span><\/p>\n

          # \u751f\u6210\u6a21\u62df\u6570\u636e<\/span>
          \ninput_seq =<\/span> torch.<\/span>randn(<\/span>batch_size,<\/span> seq_length,<\/span> input_size)<\/span>
          \ntarget_seq =<\/span> torch.<\/span>randn(<\/span>batch_size,<\/span> seq_length,<\/span> output_size)<\/span><\/p>\n

          # \u6a21\u578b\u5b9e\u4f8b\u5316<\/span>
          \nmodel =<\/span> Seq2Seq(<\/span>input_size,<\/span> hidden_size,<\/span> output_size)<\/span><\/p>\n

          # \u635f\u5931\u51fd\u6570\u548c\u4f18\u5316\u5668<\/span>
          \ncriterion =<\/span> nn.<\/span>MSELoss(<\/span>)<\/span>
          \noptimizer =<\/span> optim.<\/span>Adam(<\/span>model.<\/span>parameters(<\/span>)<\/span>,<\/span> lr=<\/span>0.001<\/span>)<\/span><\/p>\n

          # \u8bad\u7ec3\u8fc7\u7a0b<\/span>
          \nmodel.<\/span>train(<\/span>)<\/span>
          \nfor<\/span> epoch in<\/span> range<\/span>(<\/span>100<\/span>)<\/span>:<\/span>
          \n optimizer.<\/span>zero_grad(<\/span>)<\/span>
          \n output =<\/span> model(<\/span>input_seq,<\/span> target_seq)<\/span>
          \n loss =<\/span> criterion(<\/span>output,<\/span> target_seq)<\/span>
          \n loss.<\/span>backward(<\/span>)<\/span>
          \n optimizer.<\/span>step(<\/span>)<\/span><\/p>\n

          if<\/span> (<\/span>epoch +<\/span> 1<\/span>)<\/span> %<\/span> 10<\/span> ==<\/span> 0<\/span>:<\/span>
          \n print<\/span>(<\/span>f"Epoch: <\/span>{<\/span>epoch +<\/span> 1<\/span>}<\/span><\/span>, Loss: <\/span>{<\/span>loss.<\/span>item(<\/span>)<\/span>:<\/span>.4f<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

          # \u6d4b\u8bd5\u8fc7\u7a0b<\/span>
          \nmodel.<\/span>eval<\/span>(<\/span>)<\/span>
          \nwith<\/span> torch.<\/span>no_grad(<\/span>)<\/span>:<\/span>
          \n output =<\/span> model(<\/span>input_seq,<\/span> target_seq)<\/span>
          \n print<\/span>(<\/span>f"\u6d4b\u8bd5\u8f93\u51fa：<\/span>{<\/span>output}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

          \n

          \u56db\u3001\u6570\u636e\u96c6\u51c6\u5907\u4e0e\u9884\u5904\u7406<\/h3>\n

          4.1 \u6570\u636e\u96c6\u9009\u62e9<\/h4>\n

          💡 \u9009\u62e9\u5408\u9002\u7684\u6570\u636e\u96c6\u662f\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u91cd\u8981\u6b65\u9aa4\u3002\u5e38\u89c1\u7684\u5bf9\u8bdd\u6570\u636e\u96c6\u5305\u62ec：<\/p>\n

            \n
          • Cornell Movie-Dialogs Corpus：\u5305\u542b\u7535\u5f71\u4e2d\u7684\u5bf9\u8bdd\u6570\u636e\u3002<\/li>\n
          • DailyDialog：\u5305\u542b\u65e5\u5e38\u751f\u6d3b\u4e2d\u7684\u5bf9\u8bdd\u6570\u636e\u3002<\/li>\n
          • Chinese Dialog Corpus：\u5305\u542b\u4e2d\u6587\u5bf9\u8bdd\u6570\u636e\u3002<\/li>\n<\/ul>\n

            4.2 \u6570\u636e\u9884\u5904\u7406\u6b65\u9aa4<\/h4>\n

            \u26a0\ufe0f \u6570\u636e\u9884\u5904\u7406\u662f\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u5173\u952e\u6b65\u9aa4，\u4e3b\u8981\u5305\u62ec\u4ee5\u4e0b\u5185\u5bb9：<\/p>\n

          • \u6570\u636e\u6e05\u6d17：\u53bb\u9664\u566a\u58f0\u6570\u636e，\u5982\u8868\u60c5\u7b26\u53f7\u3001\u7279\u6b8a\u5b57\u7b26\u7b49\u3002<\/li>\n
          • \u5206\u8bcd：\u5c06\u6587\u672c\u5206\u5272\u4e3a\u5355\u8bcd\u6216\u5b50\u8bcd\u3002<\/li>\n
          • \u6784\u5efa\u8bcd\u6c47\u8868：\u7edf\u8ba1\u6587\u672c\u4e2d\u7684\u5355\u8bcd\u9891\u7387，\u6784\u5efa\u8bcd\u6c47\u8868\u3002<\/li>\n
          • \u5e8f\u5217\u586b\u5145：\u5c06\u8f93\u5165\u548c\u8f93\u51fa\u5e8f\u5217\u586b\u5145\u5230\u76f8\u540c\u7684\u957f\u5ea6，\u4ee5\u4fbf\u4e8e\u6a21\u578b\u8bad\u7ec3\u3002<\/li>\n

            4.3 \u4ee3\u7801\u5b9e\u73b0：\u6570\u636e\u9884\u5904\u7406<\/h4>\n

            import<\/span> torch
            \nfrom<\/span> transformers import<\/span> BertTokenizer
            \nimport<\/span> pandas as<\/span> pd<\/p>\n

            # \u52a0\u8f7d\u6570\u636e\u96c6<\/span>
            \ndf =<\/span> pd.<\/span>read_csv(<\/span>'dialog_data.csv'<\/span>,<\/span> names=<\/span>[<\/span>'context'<\/span>,<\/span> 'response'<\/span>]<\/span>)<\/span><\/p>\n

            # \u52a0\u8f7dBERT\u5206\u8bcd\u5668<\/span>
            \ntokenizer =<\/span> BertTokenizer.<\/span>from_pretrained(<\/span>'bert-base-chinese'<\/span>)<\/span><\/p>\n

            # \u6570\u636e\u9884\u5904\u7406\u51fd\u6570<\/span>
            \ndef<\/span> preprocess_data<\/span>(<\/span>context,<\/span> response,<\/span> tokenizer,<\/span> max_length=<\/span>512<\/span>)<\/span>:<\/span>
            \n # \u5bf9\u4e0a\u4e0b\u6587\u8fdb\u884c\u7f16\u7801<\/span>
            \n context_encoding =<\/span> tokenizer(<\/span>
            \n context,<\/span>
            \n padding=<\/span>'max_length'<\/span>,<\/span>
            \n truncation=<\/span>True<\/span>,<\/span>
            \n max_length=<\/span>max_length,<\/span>
            \n return_tensors=<\/span>'pt'<\/span>
            \n )<\/span><\/p>\n

            # \u5bf9\u56de\u590d\u8fdb\u884c\u7f16\u7801<\/span>
            \n response_encoding =<\/span> tokenizer(<\/span>
            \n response,<\/span>
            \n padding=<\/span>'max_length'<\/span>,<\/span>
            \n truncation=<\/span>True<\/span>,<\/span>
            \n max_length=<\/span>max_length,<\/span>
            \n return_tensors=<\/span>'pt'<\/span>
            \n )<\/span><\/p>\n

            return<\/span> {<\/span>
            \n 'context_input_ids'<\/span>:<\/span> context_encoding[<\/span>'input_ids'<\/span>]<\/span>,<\/span>
            \n 'context_attention_mask'<\/span>:<\/span> context_encoding[<\/span>'attention_mask'<\/span>]<\/span>,<\/span>
            \n 'response_input_ids'<\/span>:<\/span> response_encoding[<\/span>'input_ids'<\/span>]<\/span>,<\/span>
            \n 'response_attention_mask'<\/span>:<\/span> response_encoding[<\/span>'attention_mask'<\/span>]<\/span>
            \n }<\/span><\/p>\n

            # \u5e94\u7528\u6570\u636e\u9884\u5904\u7406\u51fd\u6570<\/span>
            \nprocessed_data =<\/span> [<\/span>]<\/span>
            \nfor<\/span> index,<\/span> row in<\/span> df.<\/span>iterrows(<\/span>)<\/span>:<\/span>
            \n processed_data.<\/span>append(<\/span>
            \n preprocess_data(<\/span>row[<\/span>'context'<\/span>]<\/span>,<\/span> row[<\/span>'response'<\/span>]<\/span>,<\/span> tokenizer)<\/span>
            \n )<\/span><\/p>\n

            # \u5c06\u5904\u7406\u540e\u7684\u6570\u636e\u8f6c\u6362\u4e3a\u5f20\u91cf<\/span>
            \ncontext_input_ids =<\/span> torch.<\/span>cat(<\/span>[<\/span>data[<\/span>'context_input_ids'<\/span>]<\/span> for<\/span> data in<\/span> processed_data]<\/span>)<\/span>
            \ncontext_attention_mask =<\/span> torch.<\/span>cat(<\/span>[<\/span>data[<\/span>'context_attention_mask'<\/span>]<\/span> for<\/span> data in<\/span> processed_data]<\/span>)<\/span>
            \nresponse_input_ids =<\/span> torch.<\/span>cat(<\/span>[<\/span>data[<\/span>'response_input_ids'<\/span>]<\/span> for<\/span> data in<\/span> processed_data]<\/span>)<\/span>
            \nresponse_attention_mask =<\/span> torch.<\/span>cat(<\/span>[<\/span>data[<\/span>'response_attention_mask'<\/span>]<\/span> for<\/span> data in<\/span> processed_data]<\/span>)<\/span><\/p>\n

            # \u4fdd\u5b58\u5904\u7406\u540e\u7684\u6570\u636e<\/span>
            \ntorch.<\/span>save(<\/span>{<\/span>
            \n 'context_input_ids'<\/span>:<\/span> context_input_ids,<\/span>
            \n 'context_attention_mask'<\/span>:<\/span> context_attention_mask,<\/span>
            \n 'response_input_ids'<\/span>:<\/span> response_input_ids,<\/span>
            \n 'response_attention_mask'<\/span>:<\/span> response_attention_mask
            \n}<\/span>,<\/span> 'processed_data.pt'<\/span>)<\/span><\/p>\n

            \n

            \u4e94\u3001\u6a21\u578b\u8bad\u7ec3\u4e0e\u4f18\u5316<\/h3>\n

            5.1 \u6a21\u578b\u67b6\u6784\u8bbe\u8ba1<\/h4>\n

            💡 \u6211\u4eec\u5c06\u4f7f\u7528BERT\u4f5c\u4e3a\u7f16\u7801\u5668，Seq2Seq\u6a21\u578b\u4f5c\u4e3a\u89e3\u7801\u5668，\u6784\u5efa\u4e00\u4e2aBERT+Seq2Seq\u67b6\u6784\u7684\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u3002BERT\u8d1f\u8d23\u7406\u89e3\u8f93\u5165\u5e8f\u5217\u7684\u8bed\u4e49\u4fe1\u606f，Seq2Seq\u6a21\u578b\u8d1f\u8d23\u751f\u6210\u8f93\u51fa\u5e8f\u5217\u3002<\/p>\n

            5.2 \u635f\u5931\u51fd\u6570\u4e0e\u4f18\u5316\u5668<\/h4>\n

            \u26a0\ufe0f \u5bf9\u4e8e\u6587\u672c\u751f\u6210\u4efb\u52a1，\u5e38\u7528\u7684\u635f\u5931\u51fd\u6570\u662f\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570\u3002\u4f18\u5316\u5668\u53ef\u4ee5\u9009\u62e9Adam\u6216SGD\u7b49\u3002<\/p>\n

            5.3 \u8bad\u7ec3\u8fc7\u7a0b<\/h4>\n

            \u2705 \u8bad\u7ec3\u8fc7\u7a0b\u4e3b\u8981\u5305\u62ec\u4ee5\u4e0b\u6b65\u9aa4：<\/p>\n

          • \u52a0\u8f7d\u9884\u5904\u7406\u540e\u7684\u6570\u636e\u96c6\u3002<\/li>\n
          • \u521d\u59cb\u5316\u6a21\u578b\u3001\u635f\u5931\u51fd\u6570\u548c\u4f18\u5316\u5668\u3002<\/li>\n
          • \u8fed\u4ee3\u8bad\u7ec3\u6a21\u578b，\u8ba1\u7b97\u635f\u5931\u5e76\u66f4\u65b0\u53c2\u6570\u3002<\/li>\n
          • \u5b9a\u671f\u4fdd\u5b58\u6a21\u578b\u68c0\u67e5\u70b9，\u4ee5\u4fbf\u4e8e\u540e\u7eed\u8bc4\u4f30\u548c\u90e8\u7f72\u3002<\/li>\n

            5.4 \u4ee3\u7801\u5b9e\u73b0：\u6a21\u578b\u8bad\u7ec3<\/h4>\n

            import<\/span> torch
            \nimport<\/span> torch.<\/span>nn as<\/span> nn
            \nimport<\/span> torch.<\/span>optim as<\/span> optim
            \nfrom<\/span> transformers import<\/span> BertModel,<\/span> BertTokenizer<\/p>\n

            # \u5b9a\u4e49BERT+Seq2Seq\u6a21\u578b<\/span>
            \nclass<\/span> BERTSeq2Seq<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
            \n def<\/span> __init__<\/span>(<\/span>self,<\/span> bert_model_name,<\/span> hidden_size,<\/span> output_size)<\/span>:<\/span>
            \n super<\/span>(<\/span>BERTSeq2Seq,<\/span> self)<\/span>.<\/span>__init__(<\/span>)<\/span>
            \n self.<\/span>bert =<\/span> BertModel.<\/span>from_pretrained(<\/span>bert_model_name)<\/span>
            \n self.<\/span>decoder =<\/span> nn.<\/span>LSTM(<\/span>hidden_size,<\/span> hidden_size,<\/span> batch_first=<\/span>True<\/span>)<\/span>
            \n self.<\/span>fc =<\/span> nn.<\/span>Linear(<\/span>hidden_size,<\/span> output_size)<\/span><\/p>\n

            def<\/span> forward<\/span>(<\/span>self,<\/span> context_input_ids,<\/span> context_attention_mask,<\/span> response_input_ids)<\/span>:<\/span>
            \n # BERT\u7f16\u7801<\/span>
            \n bert_output =<\/span> self.<\/span>bert(<\/span>
            \n input_ids=<\/span>context_input_ids,<\/span>
            \n attention_mask=<\/span>context_attention_mask
            \n )<\/span>
            \n encoder_output =<\/span> bert_output.<\/span>last_hidden_state<\/p>\n

            # \u89e3\u7801\u5668<\/span>
            \n decoder_output,<\/span> _ =<\/span> self.<\/span>decoder(<\/span>response_input_ids,<\/span> (<\/span>encoder_output[<\/span>:<\/span>,<\/span> 0<\/span>:<\/span>1<\/span>,<\/span> :<\/span>]<\/span>,<\/span> encoder_output[<\/span>:<\/span>,<\/span> 0<\/span>:<\/span>1<\/span>,<\/span> :<\/span>]<\/span>)<\/span>)<\/span><\/p>\n

            # \u8f93\u51fa\u5c42<\/span>
            \n output =<\/span> self.<\/span>fc(<\/span>decoder_output)<\/span><\/p>\n

            return<\/span> output<\/p>\n

            # \u52a0\u8f7d\u6570\u636e<\/span>
            \ndata =<\/span> torch.<\/span>load(<\/span>'processed_data.pt'<\/span>)<\/span>
            \ncontext_input_ids =<\/span> data[<\/span>'context_input_ids'<\/span>]<\/span>
            \ncontext_attention_mask =<\/span> data[<\/span>'context_attention_mask'<\/span>]<\/span>
            \nresponse_input_ids =<\/span> data[<\/span>'response_input_ids'<\/span>]<\/span>
            \nresponse_attention_mask =<\/span> data[<\/span>'response_attention_mask'<\/span>]<\/span><\/p>\n

            # \u8d85\u53c2\u6570\u8bbe\u7f6e<\/span>
            \nbert_model_name =<\/span> 'bert-base-chinese'<\/span>
            \nhidden_size =<\/span> 768<\/span>
            \noutput_size =<\/span> tokenizer.<\/span>vocab_size
            \nbatch_size =<\/span> 16<\/span>
            \nepochs =<\/span> 10<\/span>
            \nlr =<\/span> 0.0001<\/span><\/p>\n

            # \u6a21\u578b\u5b9e\u4f8b\u5316<\/span>
            \nmodel =<\/span> BERTSeq2Seq(<\/span>bert_model_name,<\/span> hidden_size,<\/span> output_size)<\/span><\/p>\n

            # \u635f\u5931\u51fd\u6570\u548c\u4f18\u5316\u5668<\/span>
            \ncriterion =<\/span> nn.<\/span>CrossEntropyLoss(<\/span>)<\/span>
            \noptimizer =<\/span> optim.<\/span>Adam(<\/span>model.<\/span>parameters(<\/span>)<\/span>,<\/span> lr=<\/span>lr)<\/span><\/p>\n

            # \u8bad\u7ec3\u8fc7\u7a0b<\/span>
            \nmodel.<\/span>train(<\/span>)<\/span>
            \nfor<\/span> epoch in<\/span> range<\/span>(<\/span>epochs)<\/span>:<\/span>
            \n total_loss =<\/span> 0.0<\/span>
            \n for<\/span> i in<\/span> range<\/span>(<\/span>0<\/span>,<\/span> len<\/span>(<\/span>context_input_ids)<\/span>,<\/span> batch_size)<\/span>:<\/span>
            \n optimizer.<\/span>zero_grad(<\/span>)<\/span><\/p>\n

            # \u6279\u91cf\u52a0\u8f7d\u6570\u636e<\/span>
            \n batch_context_input_ids =<\/span> context_input_ids[<\/span>i:<\/span>i+<\/span>batch_size]<\/span>
            \n batch_context_attention_mask =<\/span> context_attention_mask[<\/span>i:<\/span>i+<\/span>batch_size]<\/span>
            \n batch_response_input_ids =<\/span> response_input_ids[<\/span>i:<\/span>i+<\/span>batch_size]<\/span>
            \n batch_response_attention_mask =<\/span> response_attention_mask[<\/span>i:<\/span>i+<\/span>batch_size]<\/span><\/p>\n

            # \u6a21\u578b\u63a8\u7406<\/span>
            \n output =<\/span> model(<\/span>
            \n batch_context_input_ids,<\/span>
            \n batch_context_attention_mask,<\/span>
            \n batch_response_input_ids[<\/span>:<\/span>,<\/span> :<\/span>–<\/span>1<\/span>]<\/span>
            \n )<\/span><\/p>\n

            # \u8ba1\u7b97\u635f\u5931<\/span>
            \n loss =<\/span> criterion(<\/span>
            \n output.<\/span>reshape(<\/span>–<\/span>1<\/span>,<\/span> output.<\/span>size(<\/span>–<\/span>1<\/span>)<\/span>)<\/span>,<\/span>
            \n batch_response_input_ids[<\/span>:<\/span>,<\/span> 1<\/span>:<\/span>]<\/span>.<\/span>reshape(<\/span>–<\/span>1<\/span>)<\/span>
            \n )<\/span><\/p>\n

            # \u53cd\u5411\u4f20\u64ad\u548c\u4f18\u5316<\/span>
            \n loss.<\/span>backward(<\/span>)<\/span>
            \n optimizer.<\/span>step(<\/span>)<\/span><\/p>\n

            total_loss +=<\/span> loss.<\/span>item(<\/span>)<\/span><\/p>\n

            average_loss =<\/span> total_loss \/<\/span> (<\/span>len<\/span>(<\/span>context_input_ids)<\/span> \/\/<\/span> batch_size)<\/span>
            \n print<\/span>(<\/span>f"Epoch: <\/span>{<\/span>epoch +<\/span> 1<\/span>}<\/span><\/span>, Average Loss: <\/span>{<\/span>average_loss:<\/span>.4f<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

            # \u4fdd\u5b58\u6a21\u578b<\/span>
            \ntorch.<\/span>save(<\/span>model.<\/span>state_dict(<\/span>)<\/span>,<\/span> 'bert_seq2seq_model.pt'<\/span>)<\/span><\/p>\n

            \n

            \u516d\u3001\u6a21\u578b\u90e8\u7f72\u4e0e\u6d4b\u8bd5<\/h3>\n

            6.1 \u6a21\u578b\u90e8\u7f72\u65b9\u5f0f<\/h4>\n

            💡 \u6a21\u578b\u90e8\u7f72\u7684\u65b9\u5f0f\u6709\u591a\u79cd，\u5e38\u89c1\u7684\u5305\u62ec：<\/p>\n

              \n
            • \u672c\u5730\u90e8\u7f72：\u5c06\u6a21\u578b\u90e8\u7f72\u5728\u672c\u5730\u670d\u52a1\u5668\u4e0a，\u901a\u8fc7API\u63d0\u4f9b\u670d\u52a1\u3002<\/li>\n
            • \u4e91\u7aef\u90e8\u7f72：\u5c06\u6a21\u578b\u90e8\u7f72\u5728\u4e91\u8ba1\u7b97\u5e73\u53f0\u4e0a，\u5982AWS\u3001Azure\u3001\u963f\u91cc\u4e91\u7b49\u3002<\/li>\n
            • \u8fb9\u7f18\u90e8\u7f72：\u5c06\u6a21\u578b\u90e8\u7f72\u5728\u8fb9\u7f18\u8bbe\u5907\u4e0a，\u5982\u624b\u673a\u3001\u667a\u80fd\u97f3\u7bb1\u7b49\u3002<\/li>\n<\/ul>\n

              6.2 \u6a21\u578b\u6d4b\u8bd5\u65b9\u6cd5<\/h4>\n

              \u26a0\ufe0f \u6a21\u578b\u6d4b\u8bd5\u4e3b\u8981\u5305\u62ec\u4ee5\u4e0b\u5185\u5bb9：<\/p>\n

            • \u81ea\u52a8\u8bc4\u4f30：\u4f7f\u7528\u81ea\u52a8\u8bc4\u4f30\u6307\u6807，\u5982BLEU\u3001ROUGE\u7b49，\u8bc4\u4f30\u6a21\u578b\u7684\u8f93\u51fa\u8d28\u91cf\u3002<\/li>\n
            • \u4eba\u5de5\u8bc4\u4f30：\u9080\u8bf7\u7528\u6237\u5bf9\u6a21\u578b\u7684\u8f93\u51fa\u8fdb\u884c\u8bc4\u4f30，\u4ee5\u83b7\u5f97\u66f4\u771f\u5b9e\u7684\u53cd\u9988\u3002<\/li>\n
            • \u538b\u529b\u6d4b\u8bd5：\u6d4b\u8bd5\u6a21\u578b\u5728\u9ad8\u5e76\u53d1\u60c5\u51b5\u4e0b\u7684\u6027\u80fd\u548c\u7a33\u5b9a\u6027\u3002<\/li>\n

              6.3 \u4ee3\u7801\u5b9e\u73b0：\u6a21\u578b\u90e8\u7f72\u4e0e\u6d4b\u8bd5<\/h4>\n

              import<\/span> torch
              \nfrom<\/span> transformers import<\/span> BertTokenizer
              \nfrom<\/span> flask import<\/span> Flask,<\/span> request,<\/span> jsonify<\/p>\n

              # \u52a0\u8f7d\u6a21\u578b\u548c\u5206\u8bcd\u5668<\/span>
              \ntokenizer =<\/span> BertTokenizer.<\/span>from_pretrained(<\/span>'bert-base-chinese'<\/span>)<\/span>
              \nmodel =<\/span> BERTSeq2Seq(<\/span>'bert-base-chinese'<\/span>,<\/span> 768<\/span>,<\/span> tokenizer.<\/span>vocab_size)<\/span>
              \nmodel.<\/span>load_state_dict(<\/span>torch.<\/span>load(<\/span>'bert_seq2seq_model.pt'<\/span>)<\/span>)<\/span>
              \nmodel.<\/span>eval<\/span>(<\/span>)<\/span><\/p>\n

              # \u521d\u59cb\u5316Flask\u5e94\u7528<\/span>
              \napp =<\/span> Flask(<\/span>__name__)<\/span><\/p>\n

              # \u5b9a\u4e49API\u63a5\u53e3<\/span>
              \n@app.<\/span>route<\/span>(<\/span>'\/chat'<\/span>,<\/span> methods=<\/span>[<\/span>'POST'<\/span>]<\/span>)<\/span>
              \ndef<\/span> chat<\/span>(<\/span>)<\/span>:<\/span>
              \n # \u83b7\u53d6\u7528\u6237\u8f93\u5165<\/span>
              \n data =<\/span> request.<\/span>get_json(<\/span>)<\/span>
              \n context =<\/span> data[<\/span>'context'<\/span>]<\/span><\/p>\n

              # \u6587\u672c\u9884\u5904\u7406<\/span>
              \n context_encoding =<\/span> tokenizer(<\/span>
              \n context,<\/span>
              \n padding=<\/span>'max_length'<\/span>,<\/span>
              \n truncation=<\/span>True<\/span>,<\/span>
              \n max_length=<\/span>512<\/span>,<\/span>
              \n return_tensors=<\/span>'pt'<\/span>
              \n )<\/span><\/p>\n

              # \u6a21\u578b\u63a8\u7406<\/span>
              \n with<\/span> torch.<\/span>no_grad(<\/span>)<\/span>:<\/span>
              \n # \u751f\u6210\u56de\u590d<\/span>
              \n response =<\/span> generate_response(<\/span>model,<\/span> context_encoding,<\/span> tokenizer)<\/span><\/p>\n

              # \u8fd4\u56de\u56de\u590d<\/span>
              \n return<\/span> jsonify(<\/span>{<\/span>'response'<\/span>:<\/span> response}<\/span>)<\/span><\/p>\n

              # \u56de\u590d\u751f\u6210\u51fd\u6570<\/span>
              \ndef<\/span> generate_response<\/span>(<\/span>model,<\/span> context_encoding,<\/span> tokenizer,<\/span> max_length=<\/span>512<\/span>)<\/span>:<\/span>
              \n # \u521d\u59cb\u5316\u56de\u590d\u5e8f\u5217<\/span>
              \n response_input_ids =<\/span> torch.<\/span>tensor(<\/span>[<\/span>[<\/span>tokenizer.<\/span>cls_token_id]<\/span>]<\/span>)<\/span><\/p>\n

              # \u9010\u8bcd\u751f\u6210\u56de\u590d<\/span>
              \n for<\/span> _ in<\/span> range<\/span>(<\/span>max_length)<\/span>:<\/span>
              \n # \u6a21\u578b\u63a8\u7406<\/span>
              \n output =<\/span> model(<\/span>
              \n context_encoding[<\/span>'input_ids'<\/span>]<\/span>,<\/span>
              \n context_encoding[<\/span>'attention_mask'<\/span>]<\/span>,<\/span>
              \n response_input_ids
              \n )<\/span><\/p>\n

              # \u83b7\u53d6\u4e0b\u4e00\u4e2a\u8bcd\u7684\u6982\u7387<\/span>
              \n next_token_logits =<\/span> output[<\/span>:<\/span>,<\/span> –<\/span>1<\/span>,<\/span> :<\/span>]<\/span>
              \n next_token_id =<\/span> torch.<\/span>argmax(<\/span>next_token_logits,<\/span> dim=<\/span>–<\/span>1<\/span>)<\/span>.<\/span>unsqueeze(<\/span>1<\/span>)<\/span><\/p>\n

              # \u6dfb\u52a0\u5230\u56de\u590d\u5e8f\u5217<\/span>
              \n response_input_ids =<\/span> torch.<\/span>cat(<\/span>[<\/span>response_input_ids,<\/span> next_token_id]<\/span>,<\/span> dim=<\/span>1<\/span>)<\/span><\/p>\n

              # \u68c0\u67e5\u662f\u5426\u751f\u6210\u4e86\u7ed3\u675f\u7b26<\/span>
              \n if<\/span> next_token_id.<\/span>item(<\/span>)<\/span> ==<\/span> tokenizer.<\/span>sep_token_id:<\/span>
              \n break<\/span><\/p>\n

              # \u89e3\u7801\u56de\u590d<\/span>
              \n response =<\/span> tokenizer.<\/span>decode(<\/span>response_input_ids.<\/span>squeeze(<\/span>)<\/span>,<\/span> skip_special_tokens=<\/span>True<\/span>)<\/span><\/p>\n

              return<\/span> response<\/p>\n

              # \u8fd0\u884cFlask\u5e94\u7528<\/span>
              \nif<\/span> __name__ ==<\/span> '__main__'<\/span>:<\/span>
              \n app.<\/span>run(<\/span>host=<\/span>'0.0.0.0'<\/span>,<\/span> port=<\/span>5000<\/span>)<\/span><\/p>\n

              \n

              \u4e03\u3001\u6848\u4f8b\u5206\u6790\u4e0e\u4f18\u5316\u601d\u8def<\/h3>\n

              7.1 \u6848\u4f8b\u5206\u6790<\/h4>\n

              \u2705 \u6211\u4eec\u4f7f\u7528DailyDialog\u6570\u636e\u96c6\u6784\u5efa\u4e86\u4e00\u4e2a\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u3002DailyDialog\u6570\u636e\u96c6\u5305\u542b13118\u4e2a\u5bf9\u8bdd，\u6bcf\u4e2a\u5bf9\u8bdd\u5305\u542b3-10\u8f6e，\u6db5\u76d6\u4e86\u65e5\u5e38\u751f\u6d3b\u4e2d\u7684\u5404\u79cd\u573a\u666f\u3002<\/p>\n

              7.2 \u6a21\u578b\u6027\u80fd\u8bc4\u4f30<\/h4>\n

              \u26a0\ufe0f \u6211\u4eec\u4f7f\u7528BLEU\u3001ROUGE\u548cMETEOR\u7b49\u6307\u6807\u8bc4\u4f30\u4e86\u6a21\u578b\u7684\u6027\u80fd\u3002\u7ed3\u679c\u8868\u660e，\u6211\u4eec\u7684\u6a21\u578b\u5728DailyDialog\u6570\u636e\u96c6\u4e0a\u53d6\u5f97\u4e86\u8f83\u597d\u7684\u6027\u80fd\u3002<\/p>\n

              7.3 \u4f18\u5316\u601d\u8def<\/h4>\n

              💡 \u4ee5\u4e0b\u662f\u4e00\u4e9b\u4f18\u5316\u6a21\u578b\u6027\u80fd\u7684\u601d\u8def：<\/p>\n

            • \u4f7f\u7528\u66f4\u5927\u7684\u9884\u8bad\u7ec3\u6a21\u578b：\u5982BERT-large\u3001RoBERTa\u7b49\u3002<\/li>\n
            • \u589e\u52a0\u8bad\u7ec3\u6570\u636e：\u4f7f\u7528\u66f4\u591a\u7684\u5bf9\u8bdd\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\u3002<\/li>\n
            • \u4f18\u5316\u6a21\u578b\u67b6\u6784：\u5982\u4f7f\u7528TransformerDecoder\u66ff\u4ee3LSTMDecoder\u3002<\/li>\n
            • \u8c03\u6574\u8d85\u53c2\u6570：\u5982\u5b66\u4e60\u7387\u3001\u6279\u6b21\u5927\u5c0f\u3001\u6700\u5927\u5e8f\u5217\u957f\u5ea6\u7b49\u3002<\/li>\n
            • \u6dfb\u52a0\u6ce8\u610f\u529b\u673a\u5236：\u5728\u89e3\u7801\u5668\u4e2d\u6dfb\u52a0\u6ce8\u610f\u529b\u673a\u5236，\u4ee5\u63d0\u9ad8\u8f93\u51fa\u8d28\u91cf\u3002<\/li>\n
              \n

              \u516b\u3001\u603b\u7ed3<\/h3>\n

              💡 \u672c\u6587\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u5982\u4f55\u4f7f\u7528BERT+Seq2Seq\u67b6\u6784\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u3002\u6211\u4eec\u9996\u5148\u4ecb\u7ecd\u4e86\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u6838\u5fc3\u539f\u7406\u4e0e\u67b6\u6784，\u7136\u540e\u8bb2\u89e3\u4e86BERT\u6a21\u578b\u548cSeq2Seq\u6a21\u578b\u7684\u539f\u7406\u4e0e\u5e94\u7528，\u63a5\u7740\u4ecb\u7ecd\u4e86\u6570\u636e\u96c6\u51c6\u5907\u4e0e\u9884\u5904\u7406\u3001\u6a21\u578b\u8bad\u7ec3\u4e0e\u4f18\u5316\u3001\u6a21\u578b\u90e8\u7f72\u4e0e\u6d4b\u8bd5\u7b49\u6b65\u9aa4，\u6700\u540e\u901a\u8fc7\u6848\u4f8b\u5206\u6790\u548c\u4f18\u5316\u601d\u8def\u8fdb\u884c\u4e86\u603b\u7ed3\u3002<\/p>\n

              \u2705 \u5e0c\u671b\u672c\u6587\u80fd\u591f\u5e2e\u52a9\u8bfb\u8005\u7406\u89e3\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u6838\u5fc3\u6280\u672f，\u5e76\u80fd\u591f\u72ec\u7acb\u5b8c\u6210\u4e00\u4e2a\u57fa\u4e8eBERT+Seq2Seq\u67b6\u6784\u7684\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

              \u7b2c8\u7bc7：\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf（BERTSeq2Seq\u67b6\u6784）\u5b66\u4e60\u76ee\u6807
              \n\u7406\u89e3\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u6838\u5fc3\u539f\u7406\u4e0e\u67b6\u6784\u638c\u63e1BERT\u6a21\u578b\u5728\u6587\u672c\u7406\u89e3\u4e2d\u7684\u5e94\u7528\u65b9\u6cd5\u5b66\u4f1a\u4f7f\u7528Seq2Seq\u6a21\u578b\u5b9e\u73b0\u6587\u672c\u751f\u6210\u529f\u80fd\u7406\u89e3Attention\u673a\u5236\u5728\u5bf9\u8bdd\u7cfb\u7edf\u4e2d\u7684\u91cd\u8981\u6027\u80fd\u591f\u72ec\u7acb\u5b8c\u6210\u4e00\u4e2a\u57fa\u4e8eBERTSeq2Seq\u67b6\u6784\u7684\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf
              \n\u7ae0\u8282\u91cd\u70b9
              \n\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u6982\u8ff0BERT\u6a21\u578b\u539f\u7406\u4e0e\u5e94\u7528Seq2Seq\u6a21\u578b\u4e0eAttention\u673a\u5236\u6570\u636e\u96c6\u51c6\u5907\u4e0e\u9884\u5904\u7406\u6a21\u578b\u8bad\u7ec3\u4e0e\u4f18\u5316\u6a21\u578b\u90e8\u7f72\u4e0e\u6d4b<\/p>\n","protected":false},"author":2,"featured_media":60872,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[2554,188,224],"topic":[],"class_list":["post-60873","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-bert","tag-188","tag-224"],"yoast_head":"\n\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\uff08BERT+Seq2Seq\u67b6\u6784\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/60873.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\uff08BERT+Seq2Seq\u67b6\u6784\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u7b2c8\u7bc7：\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf（BERTSeq2Seq\u67b6\u6784）\u5b66\u4e60\u76ee\u6807 \u7406\u89e3\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u6838\u5fc3\u539f\u7406\u4e0e\u67b6\u6784\u638c\u63e1BERT\u6a21\u578b\u5728\u6587\u672c\u7406\u89e3\u4e2d\u7684\u5e94\u7528\u65b9\u6cd5\u5b66\u4f1a\u4f7f\u7528Seq2Seq\u6a21\u578b\u5b9e\u73b0\u6587\u672c\u751f\u6210\u529f\u80fd\u7406\u89e3Attention\u673a\u5236\u5728\u5bf9\u8bdd\u7cfb\u7edf\u4e2d\u7684\u91cd\u8981\u6027\u80fd\u591f\u72ec\u7acb\u5b8c\u6210\u4e00\u4e2a\u57fa\u4e8eBERTSeq2Seq\u67b6\u6784\u7684\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf \u7ae0\u8282\u91cd\u70b9 \u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u6982\u8ff0BERT\u6a21\u578b\u539f\u7406\u4e0e\u5e94\u7528Seq2Seq\u6a21\u578b\u4e0eAttention\u673a\u5236\u6570\u636e\u96c6\u51c6\u5907\u4e0e\u9884\u5904\u7406\u6a21\u578b\u8bad\u7ec3\u4e0e\u4f18\u5316\u6a21\u578b\u90e8\u7f72\u4e0e\u6d4b\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/60873.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-01-16T06:17:33+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260116061731-6969d7fbcbf1c.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/60873.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/60873.html\",\"name\":\"\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\uff08BERT+Seq2Seq\u67b6\u6784\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-01-16T06:17:33+00:00\",\"dateModified\":\"2026-01-16T06:17:33+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/60873.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/60873.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/60873.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\uff08BERT+Seq2Seq\u67b6\u6784\uff09\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\uff08BERT+Seq2Seq\u67b6\u6784\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/60873.html","og_locale":"zh_CN","og_type":"article","og_title":"\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\uff08BERT+Seq2Seq\u67b6\u6784\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u7b2c8\u7bc7：\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf（BERTSeq2Seq\u67b6\u6784）\u5b66\u4e60\u76ee\u6807 \u7406\u89e3\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u7684\u6838\u5fc3\u539f\u7406\u4e0e\u67b6\u6784\u638c\u63e1BERT\u6a21\u578b\u5728\u6587\u672c\u7406\u89e3\u4e2d\u7684\u5e94\u7528\u65b9\u6cd5\u5b66\u4f1a\u4f7f\u7528Seq2Seq\u6a21\u578b\u5b9e\u73b0\u6587\u672c\u751f\u6210\u529f\u80fd\u7406\u89e3Attention\u673a\u5236\u5728\u5bf9\u8bdd\u7cfb\u7edf\u4e2d\u7684\u91cd\u8981\u6027\u80fd\u591f\u72ec\u7acb\u5b8c\u6210\u4e00\u4e2a\u57fa\u4e8eBERTSeq2Seq\u67b6\u6784\u7684\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf \u7ae0\u8282\u91cd\u70b9 \u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\u6982\u8ff0BERT\u6a21\u578b\u539f\u7406\u4e0e\u5e94\u7528Seq2Seq\u6a21\u578b\u4e0eAttention\u673a\u5236\u6570\u636e\u96c6\u51c6\u5907\u4e0e\u9884\u5904\u7406\u6a21\u578b\u8bad\u7ec3\u4e0e\u4f18\u5316\u6a21\u578b\u90e8\u7f72\u4e0e\u6d4b","og_url":"https:\/\/www.wsisp.com\/helps\/60873.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-01-16T06:17:33+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260116061731-6969d7fbcbf1c.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/60873.html","url":"https:\/\/www.wsisp.com\/helps\/60873.html","name":"\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\uff08BERT+Seq2Seq\u67b6\u6784\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-01-16T06:17:33+00:00","dateModified":"2026-01-16T06:17:33+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/60873.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/60873.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/60873.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u6218\u2014\u2014\u6784\u5efa\u667a\u80fd\u5bf9\u8bdd\u7cfb\u7edf\uff08BERT+Seq2Seq\u67b6\u6784\uff09"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/60873","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=60873"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/60873\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/60872"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=60873"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=60873"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=60873"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=60873"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}