{"id":39074,"date":"2018-11-05T00:00:10","date_gmt":"2018-11-04T16:00:10","guid":{"rendered":"https:\/\/case.ntu.edu.tw\/blog\/?p=39074"},"modified":"2022-01-04T17:50:12","modified_gmt":"2022-01-04T09:50:12","slug":"%e8%87%aa%e7%84%b6%e8%aa%9e%e8%a8%80%e8%99%95%e7%90%86%e3%80%8c%e5%8d%81%e9%a0%85%e5%85%a8%e8%83%bd%e3%80%8d%e6%8c%91%e6%88%b0","status":"publish","type":"post","link":"https:\/\/case.ntu.edu.tw\/blog\/?p=39074","title":{"rendered":"\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u300c\u5341\u9805\u5168\u80fd\u300d\u6311\u6230"},"content":{"rendered":"
\n
\"\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u300c\u5341\u9805\u5168\u80fd\u300d\u6311\u6230\"<\/div>\n<\/div>\n
\n
\n

\u7de8\u8b6f\uff0f\u81fa\u5927\u96fb\u6a5f\u7cfb \u5433\u5955\u8431<\/strong><\/p>\n

\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u300c\u5341\u9805\u5168\u80fd\u300d<\/strong><\/p>\n

\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u300c\u5341\u9805\u5168\u80fd\u300d\uff08Natural Language Decathlon\uff09\uff0c\u76ee\u6a19\u662f\u8a2d\u8a08\u4e00\u500b\u6574\u5408\u5341\u9805\u529f\u80fd\u7684\u6a21\u578b [1]\uff1a<\/p>\n

\"\"<\/p>\n

\u5341\u9805\u529f\u80fd\u7bc4\u4f8b\uff08\u5716\u7247\u4f86\u6e90\uff1a\u53c3\u8003\u8cc7\u6599[1]\uff09<\/p>\n<\/div>\n

\u5716\u4e2d\u6578\u5b57\u5c0d\u61c9\u4e0b\u65b9\u6587\u5b57\u63cf\u8ff0\u3002\u300c\u554f\u53e5\u300d\u662f\u8981\u57f7\u884c\u7684\u529f\u80fd\uff0c\u6a21\u578b\u8981\u5f9e\u300c\u5167\u6587\u300d\u4e2d\u5c0b\u627e\u7b54\u6848\u3002\u300c\u7b54\u6848\u300d\u4e4b\u7d05\u5b57\u4ee3\u8868\u5176\u70ba\u539f\u6587\u4e2d\u7684\u4e00\u6bb5\u8a71\uff0c\u7da0\u5b57\u70ba\u539f\u6587\u4e2d\u7684\u4e00\u500b\u5b57\uff0c\u85cd\u5b57\u5247\u662f\u6a21\u578b\u6839\u64da\u5c0d\u6587\u672c\u7684\u7406\u89e3\u751f\u6210\u7684\u7b54\u6848\u554f\u7b54\uff08Question Answering, QA\uff09\uff1a\u7d66\u5b9a\u554f\u984c\uff0c\u5f9e\u5305\u542b\u7b54\u6848\u7684\u8cc7\u8a0a\u4e2d\u627e\u51fa\u7b54\u6848<\/p>\n

    \n
  1. \u6a5f\u5668\u7ffb\u8b6f\uff08Machine Translation\uff09\uff1a\u5c07\u6587\u4ef6\u7531\u4f86\u6e90\u8a9e\u8f49\u63db\u70ba\u76ee\u6a19\u8a9e\uff0c\u4ee5 BLEU \u8861\u91cf\u7ffb\u8b6f\u7cbe\u78ba\u5ea6<\/li>\n
  2. \u6458\u8981\uff08Summarization\uff09\uff1a\u6a21\u578b\u6458\u8981\u6587\u7ae0\u7684\u6548\u80fd\u4ee5 ROUGE\uff08Recall-Oriented Understudy for Gisting Evaluation\uff09 \u8861\u91cf<\/li>\n
  3. \u81ea\u7136\u8a9e\u8a00\u63a8\u7406\uff08Natural Language Inference, NLI\uff09\uff1a\u8f38\u5165\u5169\u500b\u53e5\u5b50\uff0c\u5206\u5225\u70ba\u524d\u63d0\u8207\u5047\u8aaa\uff0c\u6a21\u578b\u9700\u5224\u65b7\u5169\u53e5\u7684\u7acb\u5834\u662f\u76f8\u7b26\u3001\u4e2d\u7acb\u3001\u6216\u77db\u76fe<\/li>\n
  4. \u60c5\u611f\u5206\u6790\uff08Sentiment Analysis\uff09\uff1a\u4f8b\u5982\uff1a\u5c0d\u65bc\u7db2\u8def\u4e0a\u63d0\u5230\u67d0\u500b\u5546\u54c1\u540d\u7a31\u7684\u8a55\u8ad6\uff0c\u5224\u65b7\u4f7f\u7528\u8005\u62b1\u6301\u7684\u89c0\u9ede\u662f\u6b63\u5411\u3001\u4e2d\u7acb\u3001\u6216\u8ca0\u9762\uff0c\u4ee5\u53ca\u662f\u5426\u5177\u6709\u8cfc\u8cb7\u610f\u5716\u3002\u85c9\u7531\u5206\u6790\u8f3f\u60c5\u4e86\u89e3\u5927\u773e\u5c0d\u5546\u54c1\u7684\u6eff\u610f\u5ea6\uff0c\u4e26\u66f4\u80fd\u638c\u63e1\u500b\u5225\u5ba2\u6236\u7684\u9700\u6c42\u8207\u9ad4\u9a57\uff0c\u63d0\u5347\u9867\u5ba2\u670d\u52d9\u7684\u54c1\u8cea<\/li>\n
  5. \u8a9e\u610f\u89d2\u8272\u6a19\u8a3b\uff08Semantic Role Labeling, SRL\uff09\uff1a\u7d66\u5b9a\u4e00\u500b\u53e5\u5b50\u8207\u52d5\u8a5e\uff0c\u5229\u7528\u6a21\u578b\u5224\u65b7\u300c\u8ab0\u5c0d\u8ab0\u505a\u4e86\u4ec0\u9ebc\u300d\uff0c\u5728\u4ec0\u9ebc\u300c\u6642\u9593\u300d\u3001\u300c\u5730\u9ede\u300d\u767c\u751f\u3002\u9019\u4e00\u9805\u53ef\u60f3\u50cf\u6210\u6709\u56fa\u5b9a\u984c\u578b\u7684\u554f\u7b54\u4efb\u52d9<\/li>\n
  6. \u95dc\u806f\u8403\u53d6\uff08Relation Extraction\uff09\uff1a\u5f9e\u6587\u7ae0\u4e2d\u627e\u51fa\u4e3b\u8a5e\u3001\u53d7\u8a5e\u3001\u8207\u4ed6\u5011\u4e4b\u9593\u7684\u95dc\u4fc2\uff0c\u5f62\u6210\u7d50\u69cb\u5316\u7684\u77e5\u8b58\u3002\u95dc\u4fc2\u7684\u985e\u5225\u53ef\u4ee5\u662f\u89d2\u8272\u3001\u4f4d\u7f6e\u8cc7\u8a0a\u3001\u793e\u6703\u95dc\u4fc2\u2026\u7b49<\/li>\n
  7. \u76ee\u6a19\u5c0e\u5411\u5c0d\u8a71\uff08Goal-Oriented Dialogue\uff09\uff1a\u5f9e\u4f7f\u7528\u8005\u904e\u53bb\u7684\u884c\u70ba\u8207\u5c0d\u8a71\u8a18\u9304\uff0c\u8ffd\u8e64\u4f7f\u7528\u8005\u5c0d\u7cfb\u7d71\u7684\u9810\u8a2d\u76ee\u6a19\u4ee5\u53ca\u4e92\u52d5\u4e2d\u6240\u5c55\u73fe\u7684\u9700\u6c42<\/li>\n
  8. \u8a9e\u610f\u5256\u6790\uff08Semantic Parsing\uff09\uff1a\u5c07\u81ea\u7136\u8a9e\u8a00\u8f49\u63db\u70ba\u6a5f\u5668\u53ef\u7406\u89e3\u7684\u5f62\u5f0f\uff0c\u4f8b\u5982\uff1a\u7d50\u69cb\u5316\u67e5\u8a62\u8a9e\u8a00\uff08SQL\uff09\u7684\u8a9e\u6cd5\u3002\u6a21\u578b\u6839\u64da\u81ea\u7136\u8a9e\u8a00\u4e2d\u7684\u554f\u984c\uff0c\u751f\u6210\u5c0d\u61c9\u7684\u8cc7\u6599\u5eabSQL\u8a9e\u6cd5\uff0c\u4ee5\u627e\u51fa\u76f8\u7b26\u7684\u7b54\u6848<\/li>\n
  9. \u4ee3\u540d\u8a5e\u89e3\u6790\uff08Pronoun Resolution\uff09\uff1a\u7576\u53e5\u5b50\u4e2d\u51fa\u73fe\u4e0d\u53ea\u4e00\u500b\u540d\u8a5e\u6642\uff0c\u5c07\u4ee3\u540d\u8a5e\u95dc\u806f\u81f3\u6b63\u78ba\u7684\u540d\u8a5e\uff0c\u4f8b\u5982\uff1a\u300c\u8607\u73ca\u611f\u8b1d\u8c9d\u8482\uff0c\u56e0\u70ba\u5979\u5e6b\u4e86\u5979\u4e00\u500b\u5fd9\u300d\uff0c\u627e\u51fa\u53e5\u4e2d\u5169\u500b\u300c\u5979\u300d\u5206\u5225\u4ee3\u8868\u7684\u4eba<\/li>\n<\/ol>\n

    \u6bcf\u9805\u529f\u80fd\u5206\u5225\u4ee5 F1\u3001\u6e96\u78ba\u7387\uff08Precision\uff09\u3001\u53ec\u56de\u7387\uff08Recall\uff09\u7b49\u8861\u91cf\u3002\u800c\u8861\u91cf\u6574\u500b\u6a21\u578b\u7684\u6548\u679c\uff0c\u5247\u662f\u5c07\u6bcf\u9805\u7684\u5206\u6578\u52a0\u7e3d\uff0c\u5f97\u52300~1000\u4e4b\u9593\u7684\u5206\u6578\uff08\u6bcf\u4e00\u9805\u70ba0~100 \u5206\uff09\u3002<\/p>\n

    \u76f8\u8f03\u65bc\u5206\u5225\u91dd\u5c0d\u6bcf\u9805\u529f\u80fd\u8a2d\u8a08\u7684\u6a21\u578b\uff0c\u300c\u5341\u9805\u5168\u80fd\u300d\u554f\u7b54\u6a21\u578b\u7684\u597d\u8655\u662f\u4e0d\u7528\u9810\u8a2d\u7279\u5b9a\u53c3\u6578\uff0c\u4fbf\u80fd\u5b78\u7fd2\u4ee5\u300c\u554f\u7b54\u300d\u7684\u5f62\u5f0f\u9054\u6210\u591a\u7a2e\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u7684\u76ee\u6a19\u3002\u9019\u4e0d\u50c5\u4ee3\u8868\u6211\u5011\u4e0d\u518d\u9700\u8981\u5206\u5225\u70ba\u6bcf\u9805\u529f\u80fd\u8a2d\u8a08\u4e26\u512a\u5316\u6a21\u578b\uff0c\u4e5f\u610f\u5473\u8457\u6a21\u578b\u53ef\u4ee5\u8655\u7406\u6a5f\u5668\u5b78\u7fd2\u7684\u300czero-shot\u300d\u554f\u984c\u2500\u2500\u9762\u5c0d\u8a13\u7df4\u6642\u6c92\u6709\u51fa\u73fe\u904e\u7684\u4efb\u52d9\uff0c\u6a21\u578b\u4e5f\u80fd\u8fce\u5203\u800c\u89e3\u3002<\/p>\n

    \u554f\u7b54\u6a21\u578b\uff1a\u908a\u770b\u984c\u76ee\uff0c\u908a\u627e\u7b54\u6848<\/strong><\/p>\n

    \u70ba\u4e86\u66f4\u7cbe\u78ba\u89e3\u8b80\u4e00\u6bb5\u6587\u5b57\uff0cSalesforce \u7814\u7a76\u4eba\u54e1\u63d0\u51fa\u4e86\u300c\u52d5\u614b\u95dc\u6ce8\u7db2\u8def\u300d(Dynamic coattention network, DCN)\uff0c\u6839\u64da\u8f38\u5165\u7684\u554f\u53e5\uff0c\u4ee5\u6700\u5408\u9069\u7684\u65b9\u5f0f\u300c\u8b80\u300d\u4e00\u7bc7\u6587\u7ae0\u5c0b\u627e\u7b54\u6848 [3]\u3002<\/p>\n

    \"\"<\/p>\n

    DCN \u67b6\u69cb\uff1a\u7de8\u78bc\u5668\u8207\u89e3\u78bc\u5668\uff08\u5716\u7247\u4f86\u6e90\uff1a\u53c3\u8003\u8cc7\u6599[3]\uff09<\/p>\n<\/div>\n

    DCN \u7684\u7279\u8272\u5728\u65bc\u5176\u7de8\u78bc\u5668\u8207\u89e3\u78bc\u5668\u7684\u8a2d\u8a08\uff1a1. \u7d9c\u5408\u8003\u91cf\u554f\u53e5\u8207\u6587\u672c\u5167\u5bb9\u7684\u7de8\u78bc\u5668\uff08Coattentive encoder\uff09<\/p>\n

    \u7de8\u78bc\u7684\u7cbe\u96a8\u5728\u65bc\u6587\u672c\u7684\u7279\u5fb5\u8868\u793a\u4e26\u4e0d\u662f\u56fa\u5b9a\u4e0d\u8b8a\u7684\uff0c\u800c\u662f\u56e0\u8f38\u5165\u7684\u554f\u53e5\u800c\u7570\u3002\u9019\u5c31\u597d\u6bd4\u6211\u5011\u505a\u95b1\u8b80\u6e2c\u9a57\u6642\uff0c\u82e5\u80fd\u770b\u904e\u984c\u76ee\u518d\u56de\u6587\u7ae0\u627e\u7b54\u6848\uff0c\u5c07\u53ef\u66f4\u5c08\u6ce8\u65bc\u8207\u984c\u76ee\u76f8\u95dc\u7684\u8cc7\u8a0a\uff0c\u964d\u4f4e\u7b54\u984c\u7684\u56f0\u96e3\u5ea6\u4e26\u63d0\u5347\u7b54\u5c0d\u6a5f\u7387\u3002Coattentive\u7de8\u78bc\u5668\u4e5f\u4f9d\u5faa\u9019\u500b\u6982\u5ff5\uff0c\u7531\u6587\u4ef6\u5167\u5bb9\u8207\u554f\u53e5\u5206\u5225\u7522\u751f\u6bcf\u500b\u5b57\u7684\u5411\u91cf\u8868\u793a\uff0c\u70ba\u6bcf\u500b\u5b57\u8a08\u7b97\u6b0a\u91cd\uff0c\u518d\u5c07\u6587\u4ef6\u8207\u554f\u53e5\u5b57\u8ddf\u5b57\u9593\u7684\u95dc\u806f\u77e9\u9663\u6a19\u6e96\u5316\u5f8c\u4e58\u4ee5\u6b0a\u91cd\uff0c\u5f97\u5230\u7684\u5411\u91cf\u76f8\u7576\u65bc\u5c07\u554f\u53e5\u6295\u5f71\u5230\u6587\u4ef6\u5411\u91cf\u7a7a\u9593\uff0c\u7d9c\u5408\u8003\u91cf\u7684\u7d50\u679c\u3002<\/p>\n

    2. \u52d5\u614b\u89e3\u78bc\u5668\uff08Dynamic pointing decoder\uff09<\/p>\n

    \u554f\u7b54\u7684\u7b54\u6848\u901a\u5e38\u662f\u5f9e\u6587\u4ef6\u4e2d\u7bc0\u9304\u7684\u67d0\u6bb5\u8a71\uff0c\u6240\u4ee5\u627e\u7b54\u6848\u7684\u904e\u7a0b\u53ef\u4ee5\u60f3\u50cf\u662f\u5c0b\u627e\u9019\u6bb5\u8a71\u6700\u4f73\u7684\u300c\u8d77\u9ede\u300d\u8207\u300c\u7d42\u9ede\u300d\u3002\u5c0b\u627e\u904e\u7a0b\u4e2d\u904b\u7528\u5169\u500b\u795e\u7d93\u7db2\u8def\uff0c\u8a08\u7b97\u6bcf\u500b\u5b57\u4f5c\u70ba\u958b\u982d\u6216\u7d50\u5c3e\u4e4b\u300c\u5206\u6578\u300d\u9ad8\u4f4e\uff0c\u591a\u6b21\u91cd\u8986\u57f7\u884c\u4ee5\u5c0b\u627e\u6700\u4f73\u7b54\u6848\u3002<\/p>\n

    \u6574\u5408\u5f0f\u3001\u6301\u7e8c\u5b78\u7fd2\u7684\u6a21\u578b<\/strong><\/p>\n

    \u8a9e\u8a00\u662f\u5145\u6eff\u5f48\u6027\u3001\u80fd\u88ab\u9748\u6d3b\u904b\u7528\u7684\u5a92\u4ecb\u3002\u540c\u6a23\u5730\uff0cAI \u4e5f\u9700\u8981\u300c\u9748\u6d3b\u300d\u7684\u8166\u888b\u4f86\u89e3\u8b80\u5b83\u3002\u73fe\u4ecaAI\u6a21\u578b\u5728\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u500b\u5225\u4efb\u52d9\u4e0a\u7686\u6709\u4e0d\u932f\u7684\u8868\u73fe\uff0c\u4e0d\u904e\u82e5\u8981\u8b93\u6a21\u578b\u66f4\u6d41\u66a2\u5730\u5c0d\u8a71\uff0c\u4e26\u7406\u89e3\u6587\u7ae0\u524d\u5f8c\u6587\u6df1\u5c64\u7684\u610f\u6db5\uff0c\u5247\u9700\u8981\u4e00\u500b\u6574\u5408\u5404\u9805\u76ee\u6a19\u4e26\u300c\u6301\u7e8c\u5b78\u7fd2\u300d\u7684\u6a21\u578b\uff0c\u4ee5\u9762\u5c0d\u8907\u96dc\u5ea6\u66f4\u9ad8\u7684\u6311\u6230\u3002\u671f\u5f85\u66f4\u591a\u50cf\u300c\u5341\u9805\u5168\u80fd\u300d\u9019\u985e\u6574\u5408\u4e0d\u540c\u985e\u578b\u4efb\u52d9\u7684\u6a21\u578b\uff0c\u5f15\u9818\u4eba\u6a5f\u4e92\u52d5\u4e16\u4ee3\u7684\u767c\u5c55\u3002<\/p>\n

     <\/p>\n

    \u7de8\u8b6f\u4f86\u6e90<\/strong><\/p>\n

    Richard Socher, \u201cAI\u2019s Next Great Challenge: Understanding the Nuances of Language<\/a>\u201d, Harvard Business Review (2018)<\/p>\n

    \u53c3\u8003\u8cc7\u6599<\/strong><\/p>\n

      \n
    1. Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, Richard Socher, \u201cThe Natural Language Decathlon: Multitask Learning as Question Answering<\/a>\u201d, (2018)<\/li>\n
    2. Caiming Xiong, Victor Zhang, Richard Socher, \u201cDynamic Coattention Networks for Question Answering<\/a>\u201d, ICLR (2017)<\/li>\n
    3. Victor Zhang, Caiming Xiong, \u201cState of the art deep learning model for question answering<\/a>\u201d, Salesforce Einstein (2016)<\/li>\n<\/ol>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

      \u96a8\u8457\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u7684\u767c\u5c55\uff0c\u6a5f\u5668\u5c0d\u4eba\u985e\u8a9e\u8a00\u7684\u7406\u89e3\u80fd\u529b\u4e5f\u70ba\u4eba\u6a5f\u4e92\u52d5\u958b\u555f\u4e86\u65b0\u7684\u8996\u91ce\u3002\u6211\u5011\u53ef\u4ee5\u8acbAmazon\u8a9e\u97f3\u52a9\u7406Alexa\u628a\u5ba4\u5167\u71c8\u5149\u8abf\u5230\u767e\u5206\u4e4b\u4e03\u5341\u4e94\u7684\u4eae\u5ea6\uff0c\u4e5f\u53ef\u4ee5\u8acbGoogle\u8a9e\u97f3\u52a9\u7406Duplex\u5e6b\u5fd9\u9810\u7d04\u9910\u5ef3\u5ea7\u4f4d\u3002\u7136\u800c\u6a5f\u5668\u662f\u5426\u300c\u771f\u6b63\u300d\u7406\u89e3\u8a9e\u8a00\u4e2d\u932f\u7d9c\u8907\u96dc\u7684\u542b\u610f\u5462\uff1f\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u300c\u5341\u9805\u5168\u80fd\u300d\u7684\u76ee\u6a19\u4fbf\u662f\u85c9\u7531\u60c5\u611f\u5206\u6790\u3001\u554f\u7b54\u7b49\u80fd\u529b\u7684\u63d0\u5347\uff0c\u589e\u9032\u6a5f\u5668\u5c0d\u8a9e\u8a00\u6df1\u85cf\u610f\u6db5\u7684\u638c\u63e1\u3002<\/p>\n","protected":false},"author":21,"featured_media":39075,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3772,3293],"tags":[6173,6174,4888,4760],"aioseo_notices":[],"jetpack_featured_media_url":"https:\/\/case.ntu.edu.tw\/blog\/wp-content\/uploads\/2022\/01\/10.png","_links":{"self":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts\/39074"}],"collection":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/users\/21"}],"replies":[{"embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=39074"}],"version-history":[{"count":1,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts\/39074\/revisions"}],"predecessor-version":[{"id":39076,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts\/39074\/revisions\/39076"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/media\/39075"}],"wp:attachment":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=39074"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=39074"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=39074"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}