{"id":2119,"date":"2016-08-26T00:12:30","date_gmt":"2016-08-25T16:12:30","guid":{"rendered":"http:\/\/boweihe.me\/?p=2119"},"modified":"2016-08-26T00:12:30","modified_gmt":"2016-08-25T16:12:30","slug":"dssm%e4%b8%8elstm","status":"publish","type":"post","link":"https:\/\/dayandcarrot.space\/?p=2119","title":{"rendered":"\u7528\u4e8eSentence Embedding\u7684DSSM\u4e0eLSTM\uff1a\u7ba1\u4e2d\u7aa5\u8c79"},"content":{"rendered":"

\u524d\u7f6e\u5e9f\u8bdd\uff1a\u611f\u89c9\u5b9e\u4e60\u8fd9\u4e24\u4e2a\u6708\u771f\u662f\u9876\u5f97\u4e0a\u5b9e\u9a8c\u5ba4\u534a\u5e74\uff0c\u60f3\u60f3\u5bf9\u4e0d\u8d77\u6211\u7684\u5bfc\u5e08\uff0c\u8ddf\u73b0\u5728\u76f8\u6bd4\u4e4b\u524d\u5929\u5929\u50cf\u662f\u5728\u6253\u9171\u6cb9\u554a\u3002<\/span><\/p>\n

\u76f8\u5173\u6587\u732e<\/h4>\n
    \n
  1. Huang, Po-Sen, et al. “Learning deep structured semantic models for web search using clickthrough data.” Proceedings of the 22nd ACM international conference on Conference on information & knowledge management<\/i>. ACM, 2013.<\/li>\n
  2. Palangi, Hamid, et al. “Deep sentence embedding using long short-term memory networks: Analysis and application to information retrieval.” IEEE\/ACM Transactions on Audio, Speech, and Language Processing<\/i> 24.4 (2016): 694-707.<\/li>\n<\/ol>\n

    DSSM\u4e00\u6487<\/h3>\n

    DSSM\u4e86\u89e3\u7684\u4e0d\u662f\u5f88\u591a\uff0c\u4f46\u662f\u8fd9\u4e24\u7bc7\u6587\u7ae0\u6709\u51e0\u5904\u662f\u7c7b\u4f3c\u7684\uff0c\u9690\u9690\u5730\u611f\u89c9\u540e\u9762\u8fd9\u7bc7\u662f\u501f\u9274\u524d\u9762\u7684\u5199\u51fa\u6765\u7684\u3002
    \n\u6587\u7ae0\u4e2dDSSM\u7684\u6700\u7ec8\u76ee\u6807\u662f\u5b9e\u73b0\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\uff0c\u4f7f\u5f97\u7ed9\u5b9a\u8f93\u5165\u6587\u6863(\u7247\u6bb5)\u53ef\u4ee5\u8f93\u51fa\u4e00\u4e2a\u5bf9\u5e94\u7684\u5411\u91cf\uff0c\u800c\u901a\u8fc7\u5411\u91cf\u4e4b\u95f4\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\u53ef\u4ee5\u5224\u65ad\u4e24\u4e2aQuery-Document\u7684\u76f8\u4f3c\u6027[latex]R(Q,D)=cosine(y_Q,y_D)[\/latex]\uff0c\u7136\u540e\u62ff\u8fd9\u4e2a\u5c31\u80fd\u505a\u641c\u7d22\u6392\u5e8f\u4e86\u3002<\/p>\n

    Word-Hashing<\/h4>\n

    DSSM\u4e2d\u7528\u5230\u4e86\u4e00\u4e2a\u5947\u6deb\u5de7\u672f\uff0c\u5c31\u662f\u6240\u8c13\u7684Word Hashing\u65b9\u6cd5\u3002\u6211\u4eec\u77e5\u9053\uff0c\u8981\u5c06\u6587\u5b57\u62ff\u6765\u5904\u7406\uff0c\u9996\u5148\u8981\u8868\u793a\u6210\u5411\u91cf\u7684\u5f62\u5f0f\uff08\u6570\u503c\u5316\uff09\uff0c\u6700\u666e\u901a\u7684\u65b9\u6cd5\u662f\u505aOne-Hot Vector, \u5373\u6bcf\u4e2a\u5355\u8bcd\u5bf9\u5e94\u4e00\u4e2afeature\u503c\uff0c\u51fa\u73b0\u5c31\u662f1\uff0c\u6ca1\u51fa\u73b0\u5c31\u662f0. \u4f46\u662f\u5355\u8bcd\u8868\u5b9e\u5728\u662f\u592a\u5927\u4e86\uff0c\u8fd9\u7bc7\u8bba\u6587\u7528\u5230\u4e86\u4e00\u4e2a\u53ebLetter-N-Gram\u7684\u6280\u672f\uff0c\u5373\u5c06\u4e00\u4e2a\u5355\u8bcd\u5f3a\u884c\u5206\u6210\u51e0\u4e2an-gram. \u4f8b\u5982\u5355\u8bcdHello\u505aLetter-3-Gram\uff0c\u5c31\u4f1a\u53d8\u6210#HE, HEL, ELL, LLO, LO#<\/span>\u00a0\u8fd9\u6837\u7684\u8868\u793a\u65b9\u6cd5\uff08\u5176\u4e2d#\u4ee3\u8868\u5f00\u5934\u6216\u7ed3\u5c3e\uff09\u3002\u7531\u4e8e\u5904\u7406\u7684Query\/Doc\u5bf9\u4e2d\u7279\u6b8a\u5b57\u7b26\u4f1a\u5728\u6570\u636e\u9884\u5904\u7406\u9636\u6bb5\u88ab\u8fc7\u6ee4\uff0c\u56e0\u6b64\u629b\u5f00\u5355\u8bcd\u5927\u5c0f\u5199\uff0ca-z0-9#\u4e00\u5171\u5c3126+10+1=37\u4e2a\u5b57\u7b26\uff0c37*37*37=50653\u79cd\u53ef\u80fd\u6027\uff0c\u76f8\u6bd4\u8bcd\u5178\u6765\u8bf4\u5c0f\u591a\u4e86\uff01
    \n\u8fd9\u4e2a\u529e\u6cd5\u521d\u770b\u4e0d\u662f\u5f88\u9760\u8c31\uff0c\u4f46\u662f\u5b9e\u9a8c\u7ed3\u679c\u5374\u662f\u4e0d\u9519\u7684\u3002<\/p>\n

    \u6b63\u4f8b\/\u53cd\u4f8b<\/h4>\n

    \u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\u9700\u8981\u7ed9\u6b63\u4f8b\u4ee5\u53ca\u53cd\u4f8b\u3002\u4e0d\u51fa\u610f\u5916\u5730\uff0c\u8fd9\u91cc\u7528\u7684\u662f\u7528\u6237\u7684click stream\uff0c\u5373\u7528\u6237\u7684\u70b9\u51fb\u6570\u636e\u3002\u5047\u8bbe[latex]D[\/latex]\u8868\u793a\u5bf9\u4e8e\u4e00\u4e2a\u67e5\u8be2Q\u7684\u6240\u6709\u5019\u9009\u6587\u6863\uff0c\u90a3\u4e48[latex]D^+[\/latex]\u8868\u793a\u7528\u6237\u70b9\u51fb\u7684\u6587\u6863\u4f5c\u4e3a\u6b63\u4f8b\uff0c[latex]D^-[\/latex]\u8868\u793a\u6ca1\u6709\u70b9\u51fb\u7684\u4f5c\u4e3a\u8d1f\u4f8b\u3002\u518d\u6839\u636e\u4e0a\u9762\u63d0\u5230\u7684\u76f8\u4f3c\u5ea6[latex]R(Q,D)[\/latex]\uff0c\u53ef\u4ee5\u505a\u51fa\u4e00\u4e2asoftmax\u7684\u540e\u9a8c\u6982\u7387\uff1a<\/p>\n

    [latex]P(D|Q)=\\frac{exp(\\gamma R(Q,D))}{\\sum_{D’ \\in D}exp(\\gamma R(Q,D’))}[\/latex]<\/p>\n

    \u901a\u4fd7\u6613\u61c2\uff0c\u5bf9\u5427\u3002<\/p>\n

    \u795e\u7ecf\u7f51\u7edc\u7ed3\u6784<\/h4>\n

    \"DSSM\"<\/a>
    \n\u60f3\u8981\u5b9e\u6218\u7684\uff0c\u53ef\u4ee5\u770b\u770bCNTK\u91cc\u9762\u6709\u4e00\u4e2a\u7c7b\u4f3c\u7684\u4f8b\u5b50\uff1a\u7f51\u7edc\u5b9a\u4e49\u6587\u4ef6    \u5728\u8fd9\u91cc<\/a>\uff0c\u4e0d\u8fc7\u8fd9\u4e2a\u6587\u4ef6\u91cc\u9762\u5df2\u7ecf\u662fWord Hashing\u8fc7\u4e4b\u540e\u7684\u4e86\uff0c\u800c\u4e14\u53c2\u6570\u8bbe\u7f6e\u548c\u8bba\u6587\u91cc\u7684\u4e0d\u5927\u4e00\u6837\u3002
    \n\u4e0a\u9762Figure1\u53ef\u4ee5\u6839\u636e\u6bcf\u4e00\u5217\u6765\u770b\u3002\u8f93\u5165\u7684\u6700\u5e95\u5c42\u662fQuery\/Doc\u7684\u9ad8\u7eac\u5ea6\u7684\u8bcd\u5411\u91cf(Term Vector)\uff0c\u7136\u540e\u8fdb\u5165\u4e86Word Hashing\u9636\u6bb5\u628a\u9ad8\u7ef4\u8bcd\u5411\u91cf\u5f04\u6210\u4e86L3G\u4e4b\u7c7b\u7684\u5411\u91cf\uff0c\u7ef4\u5ea6\u5927\u6982\u572830k~50k\u4e4b\u95f4\uff0c\u7136\u540e\u52a0\u4e86\u4e24\u4e2a\u9690\u85cf\u5c42\u540e\u8f93\u51fa\u4e86\u6700\u7ec8\u7684128\u7ef4\u5411\u91cfy\u3002\u8fd9\u4e2a\u5411\u91cfy\u5728\u8bad\u7ec3\u7ed3\u675f\u4e4b\u540e\u5c31\u662f\u4f20\u8bf4\u4e2d\u7684\u8bcd\u5411\u91cf\u4e86\u3002\u7136\u540e\u4e24\u4e2a\u8bcd\u5411\u91cf\u4e4b\u95f4\u53ef\u4ee5\u7b97\u51fa\u76f8\u4f3c\u5ea6\uff0c\u6700\u7ec8\u7ed9\u51fa\u4e00\u4e2asoftmax\u7b97\u7684\u540e\u9a8c\u6982\u7387[latex]P(D_i|Q)[\/latex]\u3002
    \n\u8bad\u7ec3\u6574\u4e00\u5768\u795e\u7ecf\u7f51\u7edc\u8981\u7ed9\u4e00\u4e2a\u4f18\u5316\u76ee\u6807\uff0c\u6587\u4e2d\u5b9a\u4e49\u4e86\u4e00\u4e2aloss function:<\/p>\n

    [latex]L(\\Lambda)=-log\\prod_{(Q,D^+)}P(D^+|Q)[\/latex]<\/p>\n

    \u5176\u4e2d[latex]\\Lambda[\/latex]\u8868\u793a\u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\u65f6\u7684\u4e00\u5806\u53c2\u6570\u3002\u7136\u540e\u4f7f\u7528\u68af\u5ea6\u4e0b\u964d\u4e4b\u7c7b\u7684\u65b9\u6cd5\uff0c\u628a\u68af\u5ea6\u4e00\u5c42\u5c42\u5f80\u56de\u4f20\u5c31\u80fd\u8c03\u6574\u53c2\u6570\u4e86\u3002<\/p>\n

    RNN\u4e0eLSTM<\/h3>\n

    RNN\/LSTM\u5177\u4f53\u662f\u600e\u4e48\u5de5\u4f5c\u7684?\u8981\u8be6\u7ec6\u7406\u89e3\u53ef\u4ee5\u770b\u8fd9\u7bc7<\/a>\u6587\u7ae0\uff0c\u8bb2\u5f97\u5f88\u900f\u5f7b\u3002\u8fd9\u91cc\u53ea\u63d0\u5230\u8bba\u6587\u4e2d\u7684\u4e00\u4e9b\u5b9e\u73b0\u3002<\/p>\n

    RNN<\/h4>\n

     
    \n    \"RNN<\/a>
    \nRNN\u7684\u7ed3\u6784\u7279\u70b9\u662f\u4e0b\u4e00\u4e2a\u5e73\u884c\u8282\u70b9\u72b6\u6001\u7684\u4fe1\u606f\u662f\u57fa\u4e8e\u4e4b\u524d\u8282\u70b9\u7684\u3002\u5982\u4e0a\u56fe\uff0c\u5728\u672c\u6587\u7684\u73af\u5883\u4e2d\uff0c\u5e95\u4e0b\u7684x0,x1,x2,…\u53ef\u4ee5\u770b\u505a\u4e00\u4e2a\u4e2a\u5355\u8bcd\u7684\u8f93\u5165\uff0c\u8fd9\u4e00\u957f\u4e32\u7684\u8282\u70b9\u8fde\u5728\u4e00\u8d77\u5c31\u53d8\u6210\u4e86\u4e00\u4e2a\u53e5\u5b50\u3002\u4e0a\u9762\u505aDSSM\u65f6\uff0c\u4e00\u6574\u53e5\uff08\u6bb5\uff09\u8bdd\u7684\u8bcd\u90fd\u62ff\u6765\u4e00\u8d77\u505aWord Hashing\uff0c\u7ed3\u679c\u5c31\u662f\u8bcd\u6c47\u4e4b\u95f4\u7684\u4fe1\u606f\u4ece\u4e00\u5f00\u59cb\u5c31\u4f1a\u88ab\u5ffd\u7565\u6389\uff0c\u5982\u679c\u6362\u6210\u4e86RNN\uff0c\u867d\u7136\u6700\u540e\u62ff\u5230\u7684\u8fd8\u662f\u4e2a\u53e5\u5b50\u7684\u5411\u91cf\uff0c\u4f46\u662f\u7406\u8bba\u4e0a\u8bad\u7ec3\u51fa\u7684\u6a21\u578b\u4f1a\u66f4\u52a0\u51c6\u786e\u4e00\u4e9b\u3002\u4e2d\u95f4\u5c42\u662fWord Hashing\u4e4b\u540e\u7684\u5c42\uff0c\u7528\u7684\u5e94\u8be5\u662f\u548cDSSM\u4e00\u6837\u7684Hashing\u65b9\u6cd5\u3002\u5728\u6700\u53f3\u4fa7(\u6700\u540e\u4e00\u4e2a\u8bcd)\u7684\u6fc0\u6d3b\u8f93\u51fa\u5c31\u662fEmbedding\u597d\u7684\u53e5\u5b50\uff0c\u4e00\u4e2a\u53e5\u5b50\u5411\u91cf:)
    \n\u4f46\u662fRNN\u7684\u5f15\u5165\u4f1a\u5f15\u51fa\u201c\u68af\u5ea6\u7206\u70b8\u201d\u6216\u201c\u68af\u5ea6\u6d88\u5931\u201d\u7684\u95ee\u9898\u3002\u6570\u5b66\u539f\u7406\u4e0d\u591a\u8bf4\uff0c\u8003\u8651\u8fd9\u4e2a\u53e5\u5b50:I grew up in France\u2026 I speak fluent French<\/em>. \u5047\u8bbe\u6211\u7684\u4efb\u52a1\u662f\u8981\u8bad\u7ec3\u4e00\u4e2aRNN\uff0c\u6839\u636e\u4e4b\u524d\u7684\u6240\u6709\u5355\u8bcd\u4fe1\u606f\u201c\u731c\u201d\u6700\u540e\u4e00\u4e2a\u8bcd\u6c47French\u3002<\/em>\u5728\u8fd9\u53e5\u8bdd\u4e2d\uff0c\u663e\u7136\u6700\u9002\u5408\u7528\u6765\u53c2\u8003\u7684\u662f\u53e5\u5b50\u5f00\u5934\u7684\u90a3\u4e2a\u8bcdFrance\uff0c\u4f46\u662f\u4e2d\u95f4\u8fd8\u9694\u4e86\u4e00\u5927\u5806\u7684\u201c\u6ca1\u5375\u7528\u201d\u8bcd\u3002\u5728\u6839\u636e\u6807\u7b7e\u5224\u65adloss\u51fd\u6570\u505a\u68af\u5ea6\u4e0b\u964d\u7684\u65f6\u5019\uff0c\u68af\u5ea6\u4e00\u5c42\u5c42\u5f80\u56de\u4f20\uff0c\u4f46\u7531\u4e8e\u4e2d\u95f4\u9694\u7684\u6bd4\u8f83\u8fdc\uff0c\u7b49\u8f6e\u5230France\u8fd9\u4e2a\u8bcd\u7684\u65f6\u5019\u5df2\u7ecf\u51e0\u4e4e\u6ca1\u6709\u5f71\u54cd\u4e86\u3002<\/p>\n

    LSTM-DSSM<\/h4>\n

    \"Sentence<\/a>
    \nLSTM\u5904\u4e8e\u4ec0\u4e48\u4f4d\u7f6e\u5462\uff1f\u6211\u628a\u4e24\u4e2a\u6587\u7ae0\u7684\u5bf9\u6bd4\u56fe\u653e\u5728\u8fd9\u8fb9\uff0c\u57fa\u672c\u4e0a\u5c31\u80fd\u770b\u51fa\u6765\u4e86\u3002\u5de6\u56fe\u91cc\u9762\u7528\u6a59\u8272\u5708\u51fa\u7684\u7ed3\u6784\uff0c\u4ed6\u7684\u4f5c\u7528\u548c\u53f3\u4fa7\u9ed1\u8272\u865a\u7ebf\u6846\u91cc\u7684\u4e1c\u897f\u662f\u4e00\u6837\u7684\uff0c\u53ea\u662fLSTM\u7ed3\u6784\u53d6\u4ee3\u4e86\u539f\u672c\u6bd4\u8f83\u201c\u5355\u7eaf\u201d\u7684\u591a\u5c42\u7f51\u7edc\u7684\u5730\u4f4d\u3002\u6240\u4ee5\u8fd9\u4e2a\u8bba\u6587\u7684\u65b9\u6cd5\u4f3c\u4e4e\u53ef\u4ee5\u53eb\u505aLSTM-DSSM.
    \n\u6240\u4ee5\u5bf9\u4e8e\u600e\u4e48\u505a\u5b66\u4e60\uff0c\u8fd9\u4e2a\u65b0\u6587\u7ae0\u7684\u65b9\u6cd5\u4e5f\u662f\u4e0e\u5b83\u7684\u524d\u8eab\u5f88\u7c7b\u4f3c\u7684\uff1a\u6839\u636e\u7528\u6237\u70b9\u51fb\u7684Query-Doc\u6765\u786e\u5b9a\u6b63\u4f8b\u53ca\u53cd\u4f8b(\u53f3\u56fe\u4e2d\u7684[latex]D^+[\/latex]\u4e0e[latex]D^-[\/latex])\uff0c\u6c42\u4e00\u4e2a\u6700\u5927\u4f3c\u7136\uff0c\u5b83\u7b49\u540c\u4e8e\u6700\u5c0f\u5316\u4e0b\u9762\u8fd9\u4e2a\u516c\u5f0f\u4e2d\u7684\u53c2\u6570[latex]\\Lambda[\/latex]\u3002<\/p>\n

    [latex]L(\\Lambda)=-log\\prod_{(Q,D^+)}P(D^+|Q)=\\sum_{r=1}^{n}l_r(\\Lambda)[\/latex]<\/p>\n

    \"ML<\/a>
    \n\u5176\u4e2d\uff0cr<\/em>\u5bf9\u5e94\u7b2cr<\/em>\u6761Query, j<\/em>\u662f\u8fd9\u6761Query\u5bf9\u5e94\u7684\u4e00\u6761Document\u3002
    \n\u591a\u8bf4\u4e00\u53e5\uff0c\u8fd9\u4e2aQuery-Doc\u7684\u5339\u914d\u8fc7\u7a0b\u6709\u70b9\u50cf\u505aTranslation Model\uff0c\u53ea\u4e0d\u8fc7Translation Model\u6709Encoding\u548cDecoding\u4e24\u6b65\uff0c\u800c\u8fd9\u91cc\u53ea\u6709Encoding\u7684\u8fc7\u7a0b\u3002<\/p>\n

    LSTM\u7ed3\u6784<\/h4>\n

    \u7b80\u5355\u6765\u8bf4\uff0cLSTM\u662f\u4e00\u79cdRNN\u7ed3\u6784\uff0c\u91cc\u9762\u901a\u8fc7\u4e00\u4e9bGate\u6765\u5b9e\u73b0\u201c\u8981\u4e0d\u8981\u8bb0\u4f4f\/\u8003\u8651XXX\u201d\u7684\u9009\u62e9\u3002LSTM\u6839\u636e\u9700\u8981\u6709\u4e0d\u540c\u7684\u5177\u4f53\u5b9e\u73b0\uff0c\u4e0b\u56fe\u662f\u8bba\u6587\u4e2d\u7528\u5230\u7684\u7ed3\u6784\uff1a
    \n    \"LSTM<\/a>
    \n\u8fd9\u4e2a\u7ed3\u6784\u5206\u6210\u4e86\u51e0\u4e2a\u5173\u952e\u90e8\u4f4d\uff1aInput Gate, Output Gate \u548c Forget Gate. \u53ef\u4ee5\u770b\u5230\u8fd9\u4e9bGate\u7684\u5b9e\u73b0\u662fSigmoid\u8f93\u51fa\u7684(\u56fe\u4e2d[latex]\\sigma(.)[\/latex])\uff0c\u6240\u4ee5\u503c\u90fd\u57280-1\u4e4b\u95f4\uff1b\u7136\u540e\u548c\u524d\u9762\u7684h(.)\u53bb\u505a\u4e00\u4e2apointwise multiple(X)\uff0c\u5b9e\u73b0\u4e86\u201c\u95e8\u201d\u7684\u529f\u80fd\u3002<\/p>\n