{"id":9119,"date":"2024-01-19T16:30:29","date_gmt":"2024-01-19T08:30:29","guid":{"rendered":"\/?p=9119"},"modified":"2024-01-19T16:52:08","modified_gmt":"2024-01-19T08:52:08","slug":"word2vec%e5%8e%9f%e7%90%86%e4%b8%8e%e5%ae%9e%e6%88%98%e8%af%a6%e8%a7%a3","status":"publish","type":"post","link":"\/?p=9119","title":{"rendered":"Word2vec\u539f\u7406\u4e0e\u5b9e\u6218\u8be6\u89e3"},"content":{"rendered":"<p><strong>\u89c6\u9891\u94fe\u63a5\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/ai.deepshare.net\/detail\/p_5ee62f90022ee_zFpnlHXA\/6\">https:\/\/ai.deepshare.net\/detail\/p_5ee62f90022ee_zFpnlHXA\/6<\/a><\/strong><\/p>\n<blockquote>\n<p><strong>\u6587\u7ae0\u6807\u9898\uff1a<\/strong><br \/>\nEfficient Estimation of Word Representations in Vector Space<br \/>\n\u57fa\u4e8e\u5411\u91cf\u7a7a\u95f4\u4e2d\u8bcd\u8868\u793a\u7684\u6709\u6548\u4f30\u8ba1<br \/>\n<strong>\u4f5c\u8005\uff1a<\/strong>TomasMikolov\uff08\u7b2c\u4e00\u4f5c\u8005\uff09<br \/>\n<strong>\u5355\u4f4d\uff1a<\/strong>Google<br \/>\n<strong>\u53d1\u8868\u4e8eICLR2013<\/strong><\/p>\n<\/blockquote>\n<h2><strong>1 \u7814\u7a76\u80cc\u666f<\/strong><\/h2>\n<h3><strong>1.1 \u524d\u671f\u77e5\u8bc6\u50a8\u5907<\/strong><\/h3>\n<p>\u6570\u5b66\u77e5\u8bc6\uff1a<\/p>\n<ul>\n<li>\u9ad8\u7b49\u6570\u5b66\u4e2d\u5fae\u79ef\u5206<\/li>\n<li>\u7ebf\u6027\u4ee3\u6570\u4e2d\u7684\u77e9\u9635\u8fd0\u7b97<\/li>\n<li>\u6982\u7387\u8bba\u7684\u6761\u4ef6\u6982\u7387<\/li>\n<\/ul>\n<p>\u673a\u5668\u5b66\u4e60\uff1a<\/p>\n<ul>\n<li>\u673a\u5668\u5b66\u4e60\u4e2d\u57fa\u672c\u7684\u539f\u7406\u4ee5\u53ca\u6982\u5ff5\u3001\u5982\u903b\u8f91\u56de\u5f52\u5206\u7c7b\u5668\u3001\u68af\u5ea6\u4e0b\u964d\u65b9\u6cd5\u7b49<\/li>\n<\/ul>\n<p>\u795e\u7ecf\u7f51\u7edc\uff1a<\/p>\n<ul>\n<li>\u4e86\u89e3\u795e\u7ecf\u7f51\u7edc\u57fa\u672c\u77e5\u8bc6<\/li>\n<li>\u77e5\u9053\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u548c\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7684\u6982\u5ff5<\/li>\n<li>\u77e5\u9053\u8bed\u8a00\u6a21\u578b\u6982\u5ff5<\/li>\n<\/ul>\n<p>\u7f16\u7a0b\uff1a<\/p>\n<ul>\n<li>\u4e86\u89e3Pytorch\u57fa\u672c\u4f7f\u7528\u65b9\u6cd5\uff0c\u5982\u6570\u636e\u8bfb\u53d6\u3001\u6a21\u578b\u6784\u5efa\u7b49<\/li>\n<\/ul>\n<h3><strong>1.2 \u5b66\u4e60\u76ee\u6807<\/strong><\/h3>\n<table>\n<thead>\n<tr>\n<th>\u5e8f\u53f7<\/th>\n<th>\u6807\u9898<\/th>\n<th>\u5185\u5bb9<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>1<\/td>\n<td>\u4e86\u89e3\u8bcd\u5411\u91cf\u7684\u80cc\u666f\u77e5\u8bc6<\/td>\n<td>\u5386\u53f2\u80cc\u666f \u6570\u5b66\u57fa\u7840<\/td>\n<\/tr>\n<tr>\n<td>2<\/td>\n<td>\u4e86\u89e3\u524d\u4eba\u7684\u5de5\u4f5c<\/td>\n<td>SVD \u524d\u9988\u795e\u7ecf\u7f51\u7edc RNN\u8bed\u8a00\u6a21\u578b<\/td>\n<\/tr>\n<tr>\n<td>3<\/td>\n<td>\u638c\u63e1\u8bcd\u5411\u91cf\u7684\u8bc4\u4ef7\u65b9\u6cd5<\/td>\n<td>Cosine\/Analagy\uff1a\u76f8\u4f3c\u5ea6\/\u8bcd\u5bf9\u7c7b\u6bd4 \u8bba\u6587\u4e2d\u6570\u636e\u96c6<\/td>\n<\/tr>\n<tr>\n<td>4<\/td>\n<td>\u638c\u63e1\u6a21\u578b\u7ed3\u6784<\/td>\n<td>CBOW Skip-gram<\/td>\n<\/tr>\n<tr>\n<td>5<\/td>\n<td>Word2vec\u4e2d\u7684\u5173\u952e\u6280\u672f\uff1a\u63d0\u9ad8\u8bad\u7ec3\u901f\u5ea6<\/td>\n<td>\u5c42\u6b21softmax \u8d1f\u91c7\u6837<\/td>\n<\/tr>\n<tr>\n<td>6<\/td>\n<td>\u638c\u63e1Word2vec\u4ee3\u7801<\/td>\n<td>\u5173\u952e\u4ee3\u7801 \u63d0\u9ad8\u6548\u7387\u7ec6\u8282<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-eff9e45ea9cf85df46f87918910e40be_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u5bf9\u6bd4\u6a21\u578b\uff1aNNLM\uff0cRNNLM\uff08LM\uff1aLanguage Model\uff09<\/p>\n<h2><strong>2 \u8bcd\u5411\u91cf\u80cc\u666f\u50a8\u5907\u77e5\u8bc6<\/strong><\/h2>\n<h3><strong>2.1 \u8bed\u8a00\u6a21\u578b<\/strong><\/h3>\n<p>\u76f4\u89c2\u6982\u5ff5\uff1a\u8bed\u8a00\u6a21\u578b\u662f\u8ba1\u7b97\u4e00\u4e2a\u53e5\u5b50\u662f\u53e5\u5b50\u7684\u6982\u7387\u7684\u6a21\u578b\uff0c\u5373\u7ed9\u5b9a\u7684\u4e00\u4e2a\u53e5\u5b50\uff0c\u5224\u65ad\u662f\u5426\u7b26\u5408\u8bed\u4e49\u548c\u8bed\u6cd5\uff0c\u7b26\u5408\u4e8c\u8005\u7684\u53e5\u5b50\u6982\u7387\u8f83\u5927\u3002<\/p>\n<p>\u793a\u4f8b\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u53e5\u5b50<\/th>\n<th>\u6982\u7387<\/th>\n<th>\u5206\u6790<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6df1\u5ea6\u4e4b\u773c\u7684\u8bba\u6587\u8bfe\u771f\u7684\u5f88\u597d\uff01<\/td>\n<td>0.8<\/td>\n<td>\u7b26\u5408\u8bed\u4e49\u548c\u8bed\u6cd5<\/td>\n<\/tr>\n<tr>\n<td>\u6df1\u5ea6\u4e4b\u773c\u7684\u8bba\u6587\u8bfe\u771f\u7684\u5f88\u4e00\u822c\uff01<\/td>\n<td>0.01<\/td>\n<td>\u4e0d\u7b26\u5408\u8bed\u4e49\uff0c\u56e0\u4e3a\u6df1\u5ea6\u4e4b\u773c\u7684\u8bfe\u8f83\u597d<\/td>\n<\/tr>\n<tr>\n<td>\u8bba\u6587\u8bfe\u7684\u6df1\u5ea6\u4e4b\u773c\u5f88\u771f\u597d\u7684\uff01<\/td>\n<td>0.000001<\/td>\n<td>\u4e0d\u7b26\u5408\u8bed\u6cd5\uff0c\u56e0\u6b64\u4e5f\u4e0d\u7b26\u5408\u8bed\u4e49<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8bed\u8a00\u6a21\u578b\u7684\u5e94\u7528\u6709\u5f88\u591a\uff0c\u4f8b\u5982\u5728\u8f93\u5165\u6cd5\u91cc\u8f93\u5165\u62fc\u97f3\uff1azi ran yu yan chu li \uff0c\u5b83\u8fd4\u56de\u5bf9\u5e94\u7684\u4e2d\u6587\u7ed3\u679c\u5c31\u662f\u4e00\u4e2a\u8bed\u8a00\u6a21\u578b\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u53e5\u5b50<\/th>\n<th>\u6982\u7387<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u81ea\u7136\u8bed\u8a00\u5904\u7406<\/td>\n<td>0.9<\/td>\n<\/tr>\n<tr>\n<td>\u5b50\u7136\u9884\u8a00\u51fa\u529b<\/td>\n<td>0.01<\/td>\n<\/tr>\n<tr>\n<td>\u7d2b\u7136\u7389\u773c\u50a8\u4f8b<\/td>\n<td>0.0001<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3><strong>2.2 \u8bed\u8a00\u6a21\u578b\u7684\u53d1\u5c55<\/strong><\/h3>\n<p>\u672c\u8282\u8bb2\u8ff0\u5982\u4f55\u6784\u5efa\u8bed\u8a00\u6a21\u578b\uff1f<\/p>\n<h4><strong>2.2.1 \u57fa\u4e8e\u4e13\u5bb6\u8bed\u6cd5\u89c4\u5219\u7684\u8bed\u8a00\u6a21\u578b<\/strong><\/h4>\n<p>\u8bed\u8a00\u5b66\u5bb6\u4f01\u56fe\u603b\u7ed3\u51fa\u4e00\u5957\u901a\u7528\u7684\u8bed\u6cd5\u89c4\u5219\uff0c\u6bd4\u5982\u5f62\u5bb9\u8bcd\u540e\u9762\u8ddf\u540d\u8bcd\uff0c\u52a8\u8bcd\u540e\u9762\u8ddf\u526f\u8bcd\uff0c\u7b26\u5408\u8fd9\u4e9b\u89c4\u5219\u662f\u53e5\u5b50\u3002\u7f3a\u70b9\u662f\u8bed\u8a00\u5343\u53d8\u4e07\u5316\uff0c\u5f88\u96be\u603b\u7ed3\u51fa\u4e00\u5957\u901a\u7528\u7684\u8bed\u6cd5\u89c4\u5219\u3002\u800c\u4e14\u968f\u7740\u65f6\u95f4\u7684\u53d8\u5316\u4f1a\u6709\u65b0\u8bcd\u7684\u8bde\u751f\uff0c\u5f88\u96be\u603b\u7ed3\u51fa\u8bcd\u6027\uff0c\u4f8b\u5982\u7b11skr\u4eba\uff01\u8fd9\u4ef6\u4e8b\u96e8\u5973\u65e0\u74dc\u3002<\/p>\n<h4><strong>2.2.2 \u7edf\u8ba1\u8bed\u8a00\u6a21\u578b<\/strong><\/h4>\n<p>\u901a\u8fc7\u6982\u7387\u8ba1\u7b97\u6765\u523b\u753b\u8bed\u8a00\u6a21\u578b\uff1a\u5c06\u53e5\u5b50\u7684\u6982\u7387\u8f6c\u53d8\u4e3a\u6240\u6709\u8bcd\u51fa\u73b0\u7684\u6982\u7387\u7684\u4e58\u79ef\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119152726700.png\" alt=\"image-20240119152726700\" \/><\/p>\n<p><strong>\u300c\u67d0\u4e2a\u8bcd\u51fa\u73b0\u7684\u6982\u7387\u7684\u8ba1\u7b97\u65b9\u6cd5\uff1a\u300d<\/strong><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119152742276.png\" alt=\"image-20240119152742276\" \/>\u6c42\u89e3\u65b9\u6cd5\uff1a\u7528\u8bed\u6599\u7684\u9891\u7387\u4ee3\u66ff\u6982\u7387\uff08\u9891\u7387\u5b66\u6d3e)<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119152755845.png\" alt=\"image-20240119152755845\" \/><\/p>\n<p>\u7edf\u8ba1\u8bed\u8a00\u6a21\u578b\u662f\u901a\u8fc7\u4e00\u4e2a\u5f88\u5927\u7684\u8bed\u6599\u6765\u6784\u5efa\u7684\uff0c\u8fd9\u4e2a\u8bed\u6599\u88ab\u79f0\u4e3a\u80cc\u666f\u8bed\u6599\uff0c\u901a\u8fc7\u4e0a\u56fe\u4e2d\u7684\u516c\u5f0f\u53ef\u4ee5\u8ba1\u7b97\u51fa\u6bcf\u4e2a\u8bcd\u51fa\u73b0\u7684\u9891\u7387\uff0c\u7528\u9891\u7387\u6765\u4ee3\u66ff\u6982\u7387\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119152917376.png\" alt=\"image-20240119152917376\" \/>\u6c42\u89e3\u65b9\u6cd5\uff1a\u9891\u7387\u5b66\u6d3e+\u6761\u4ef6\u6982\u7387<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119152933023.png\" alt=\"image-20240119152933023\" \/><\/p>\n<p>\u8bed\u8a00\u6a21\u578b\u793a\u4f8b\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119152944320.png\" alt=\"image-20240119152944320\" \/><\/p>\n<ul>\n<li><code>P(\u5f20\u4e09 \u5f88 \u5e05)=P(\u5f20\u4e09)*P(\u5f88|\u5f20\u4e09)*P(\u5e05|\u5f20\u4e09, \u5f88)<\/code><\/li>\n<li><code>P(\u5f20\u5f88\u5e05 \u5f88 \u5e05)=P(\u5f20\u5f88\u5e05)*P(\u5f88|\u5f20\u5f88\u5e05)*P(\u5e05|\u5f20\u5f88\u5e05, \u5f88)\uff1a\u5f20\u5f88\u5e05\u8fd9\u4e2a\u540d\u5b57\u5f88\u5c11\u89c1\uff0c\u5728\u8bed\u6599\u5e93\u53ef\u80fd\u6ca1\u6709\u51fa\u73b0\u8fc7\uff0c\u4f46\u662f\u8fd9\u4e0d\u80fd\u4ee3\u8868\u5b83\u4e0d\u5b58\u5728\uff0cP(\u5f20\u5f88\u5e05)=0\u4e0d\u5408\u7406<\/code><\/li>\n<li><code>P(\u5f20\u4e09 \u5f88 \u6f02\u4eae)=P(\u5f20\u4e09)*P(\u5f88|\u5f20\u4e09)*P(\u6f02\u4eae|\u5f20\u4e09, \u5f88)\uff1a\u8f83\u957f\u7684\u53e5\u5b50\uff08\u77ed\u8bed\uff09\u5728\u8bed\u6599\u5e93\u4e2d\u4e5f\u53ef\u80fd\u6982\u7387\u8f83\u4f4e\uff0c\u6bd4\u5982\u5f20\u4e09\u7684\u540c\u5b66\u7684\u59b9\u59b9\u7684\u670b\u53cb\u7684\u597d\u670b\u53cb\u5f88\u6f02\u4eae<\/code><\/li>\n<\/ul>\n<p><strong>\u300c\u7edf\u8ba1\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u5e73\u6ed1\u64cd\u4f5c\u300d<\/strong><\/p>\n<ul>\n<li>\u6709\u4e00\u4e9b\u8bcd\u6216\u8005\u8bcd\u7ec4\u5728\u8bed\u6599\u5e93\u4e2d\u6ca1\u6709\u51fa\u73b0\u8fc7\uff0c\u4f46\u662f\u8fd9\u4e0d\u80fd\u4ee3\u8868\u5b83\u4e0d\u5b58\u5728\u3002<\/li>\n<li>\u5e73\u6ed1\u64cd\u4f5c\u5c31\u662f\u7ed9\u90a3\u4e9b\u6ca1\u6709\u51fa\u73b0\u8fc7\u7684\u8bcd\u6216\u8005\u8bcd\u7ec4\u4e5f\u7ed9\u4e00\u4e2a\u6bd4\u8f83\u5c0f\u7684\u6982\u7387\u3002<\/li>\n<li>\u62c9\u666e\u62c9\u65af\u5e73\u6ed1\uff08<code>Laplace Smoothing<\/code>\uff09\uff1a\u4e5f\u79f0\u4e3a\u52a01\u5e73\u6ed1\uff0c\u6bcf\u4e2a\u8bcd\u5728\u539f\u6765\u51fa\u73b0\u6b21\u6570\u7684\u57fa\u7840\u4e0a\u52a01\u3002<\/li>\n<\/ul>\n<p>\u793a\u4f8b\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-3e11d70e21fc47ea00fbb412caa490c2_720w.webp\" alt=\"img\" \/><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119153059341.png\" alt=\"image-20240119153059341\" \/><\/p>\n<p><strong><em>V<\/em><\/strong>\u8868\u793a\u8bed\u6599\u5e93\u4e2d\u8bcd\u7684\u4e2a\u6570<\/p>\n<p>\u7f3a\u70b9\uff1a\u4e3b\u8981\u89e3\u51b3\u4e86\u8bcd\u7684\u95ee\u9898\uff0c\u5982\u679c\u51fa\u73b0\u8f83\u957f\u7684\u8bcd\u7ec4\uff0c\u9700\u8981\u5e73\u6ed1\u7684\u8bcd\u548c\u8bcd\u7ec4\u6570\u91cf\u5927\u5927\u589e\u52a0\u3002<\/p>\n<p>\u793a\u4f8b\uff1a\u5047\u8bbe\u8bed\u6599\u5e93\u4e2d\u53ea\u6709\u4e00\u53e5\u8bdd\uff0c\u5f20\u4e09\u5f88\u6f02\u4eae<\/p>\n<ul>\n<li>P(\u5f20\u4e09 \u5f88 \u5e05)=P(\u5f20\u4e09)P(\u5f88|\u5f20\u4e09)P(\u5e05|\u5f20\u4e09, \u5f88)<\/li>\n<li>P(\u5f20\u4e09 \u5f88 \u684c\u5b50)=P(\u5f20\u4e09)P(\u5f88|\u5f20\u4e09)P(\u684c\u5b50|\u5f20\u4e09, \u5f88)<\/li>\n<li>\u5e73\u6ed1\u4e4b\u540eP(\u5e05|\u5f20\u4e09, \u5f88)=P(\u684c\u5b50|\u5f20\u4e09, \u5f88)\uff0c\u5373\u4e24\u4e2a\u53e5\u5b50\u6982\u7387\u662f\u76f8\u7b49\u7684\u3002\u4f46\u662f\u7b2c\u4e8c\u4e2a\u53e5\u5b50\u662f\u4e0d\u5408\u7406\u7684\uff0c\u4e00\u822c\u60c5\u51b5\u4e0b\uff0c\u7b2c\u4e00\u53e5\u8bdd\u6982\u7387\u5e94\u8be5\u6bd4\u7b2c\u4e8c\u53e5\u8bdd\u6982\u7387\u8981\u9ad8\u624d\u5bf9<\/li>\n<\/ul>\n<p>\u5e73\u6ed1\u64cd\u4f5c\u7684\u95ee\u9898\uff0c\u53ef\u80fd\u4f1a\u5bfc\u81f4\u4e00\u4e2a\u6709\u8bed\u6cd5\u9519\u8bef\u7684\u53e5\u5b50\u548c\u4e00\u4e2a\u6b63\u786e\u53e5\u5b50\u7684\u6982\u7387\u662f\u76f8\u540c\u7684\u3002<\/p>\n<p>\u95ee\u9898\u539f\u56e0\uff1a<\/p>\n<ol>\n<li>\u53c2\u6570\u7a7a\u95f4\u8fc7\u5927\uff1a\u4e0e\u6a21\u578b\u76f8\u5173\u7684\u53c2\u6570\u4f1a\u5f88\u591a\u3002\u4f8b\u5982\uff1a\u8ba1\u7b97<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119153148150.png\" alt=\"image-20240119153148150\" \/>\u53c2\u6570\u6709<strong><em>V<\/em><\/strong>\u4e2a\uff08\u6bcf\u4e2a<img src=\"Word2vec%E5%8E%9F%E7%90%86%E4%B8%8E%E5%AE%9E%E6%88%98%E8%AF%A6%E8%A7%A3md.assets\/image-20240119153148150-1705649606130-27.png\" alt=\"image-20240119153148150\" \/>\u4e00\u4e2a\u53c2\u6570\uff09\uff0c<strong><em>V<\/em><\/strong>\u4ee3\u8868\u8bcd\u8868\u5927\u5c0f\uff0c\u8ba1\u7b97<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119153350890.png\" alt=\"image-20240119153350890\" \/>\u53c2\u6570\u6709<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119153411087.png\" alt=\"image-20240119153411087\" \/>\u4e2a\uff0c\uff08<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119153426241.png\" alt=\"image-20240119153426241\" \/>\u5206\u5b50\u6709<strong><em>V<\/em><\/strong>\u00d7<strong><em>V<\/em><\/strong>\u4e2a\u53c2\u6570)\uff0c\u4f9d\u6b21\u7c7b\u63a8\uff0c\u90a3\u4e48\u6a21\u578b\u53c2\u6570\u7a7a\u95f4\u4e3a\uff1a<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119153456853.png\" alt=\"image-20240119153456853\" \/>\uff0c<strong><em>L<\/em><\/strong>\u4e3a\u53e5\u5b50\u957f\u5ea6\u3002<\/li>\n<li>\u6570\u636e\u7a00\u758f\u4e25\u91cd\uff1a\u5e73\u6ed1\u4e4b\u540e\u8f83\u5c11\u51fa\u73b0\u7684\u8bcd\u7684\u6982\u7387\u8f83\u5c0f\uff0c\u5f53\u8f83\u5c11\u7684\u8bcd\u5360\u6bd4\u8f83\u591a\u65f6\uff0c\u6570\u636e\u4f1a\u5f88\u7a00\u758f\u3002\u6700\u7ec8\u5927\u90e8\u5206\u6570\u636e\u503c\u63a5\u8fd1\u4e8e0\uff0c\u53ea\u6709\u5c11\u90e8\u5206\u4e0d\u4e3a0\u3002\u7c7b\u4f3c\u7684\uff0c\u5927\u7684\u8ba1\u7b97\u91cf\u4f1a\u5bfc\u81f4\u4e00\u4e9b\u95ee\u9898\uff0c\u6bd4\u5982\u8bf4\u6570\u636e\u8fc7\u4e8e\u7a00\u758f\u3002\u56e0\u4e3a\u6211\u4eec\u60f3\u6bcf\u4e2a\u8bcd\u90fd\u8981\u8003\u8651\u524d\u9762\u5f88\u591a\u7684\u8bcd\uff0c\u800c\u4e14\u628a\u5f88\u591a\u8bcd\u7ec4\u5408\u5728\u4e00\u8d77\u7684\u6982\u7387\u5176\u5b9e\u5e76\u6ca1\u6709\u5f88\u9ad8\u3002\u7ec4\u5408\u5230\u4e00\u8d77\u7684\u8bcd\u8d8a\u591a\u5b83\u7684\u6570\u636e\u662f\u8d8a\u7a00\u758f\u7684\u3002<\/li>\n<\/ol>\n<p>\u89e3\u51b3\u529e\u6cd5\uff1a\u9a6c\u5c14\u79d1\u592b\u5047\u8bbe<\/p>\n<p><strong>\u300c\u9a6c\u5c14\u79d1\u592b\u5047\u8bbe\u300d<\/strong>\uff1a\u4e0b\u4e00\u4e2a\u8bcd\u7684\u51fa\u73b0\u4ec5\u4f9d\u8d56\u4e8e\u524d\u9762\u7684\u4e00\u4e2a\u8bcd\u6216\u51e0\u4e2a\u8bcd<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119153541712.png\" alt=\"image-20240119153541712\" \/><\/p>\n<p>\u8bed\u8a00\u6a21\u578b\u793a\u4f8b\uff1a\u6211 \u4eca\u5929 \u4e0b\u5348 \u6253 \u7fbd\u6bdb\u7403<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119153604352.png\" alt=\"image-20240119153604352\" \/><\/p>\n<h3><strong>2.3 \u8bed\u8a00\u6a21\u578b\u8bc4\u4ef7\u6307\u6807<\/strong><\/h3>\n<p>\u56f0\u60d1\u5ea6(Perplexity)<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119153633282.png\" alt=\"image-20240119153633282\" \/><\/p>\n<p>\u8bed\u8a00\u6a21\u578b\u5b9e\u8d28\u4e0a\u662f\u4e00\u4e2a\u8bcd\u8868\u5927\u5c0f\u7684\u591a\u5206\u7c7b\u95ee\u9898\uff0c\u7c7b\u522b\u662f\u4e00\u4e2a\u8bcd\u3002\u53e5\u5b50\u6982\u7387\u8d8a\u5927\uff0c\u8bed\u8a00\u6a21\u578b\u8d8a\u597d\uff0c\u56f0\u60d1\u5ea6\u8d8a\u5c0f\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-8b159fad1866669c815b4091addc5484_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u8bed\u8a00\u6a21\u578b\u53ea\u80fd\u5224\u65ad\u53e5\u5b50\u662f\u53e5\u5b50\u7684\u6982\u7387\uff0c\u4e0d\u662f\u5b8c\u5907\u7684\u76d1\u7763\u5b66\u4e60\u3002\u56f0\u60d1\u5ea6\u8d8a\u5c0f\uff0c\u5bf9\u5e94\u53e5\u5b50\u6982\u7387\u8d8a\u5927\uff0c\u5219\u8868\u793a\u8bcd\u8868\u5927\u5c0f\u7684\u591a\u5206\u7c7b\u95ee\u9898\u6bcf\u4e2a\u7c7b\u522b\u7684\u6982\u7387\u8d8a\u5927\u3002\u56f0\u60d1\u5ea6\u8d8a\u5c0f\u8d8a\u597d\u3002<\/p>\n<h3><strong>2.4 n-gram\u7279\u5f81<\/strong><\/h3>\n<p>\u5728\u6587\u672c\u7279\u5f81\u63d0\u53d6\u4e2d\uff0c\u5e38\u5e38\u80fd\u770b\u5230n-gram\u7684\u8eab\u5f71\u3002\u5b83\u662f\u4e00\u79cd\u57fa\u4e8e\u8bed\u8a00\u6a21\u578b\u7684\u7b97\u6cd5\uff0c\u57fa\u672c\u601d\u60f3\u662f\u5c06\u6587\u672c\u5185\u5bb9\u6309\u7167\u5b57\u8282\u987a\u5e8f\u8fdb\u884c\u5927\u5c0f\u4e3aN\u7684\u6ed1\u52a8\u7a97\u53e3\u64cd\u4f5c\uff0c\u6700\u7ec8\u5f62\u6210\u957f\u5ea6\u4e3aN\u7684\u5b57\u8282\u7247\u6bb5\u5e8f\u5217\u3002\u770b\u4e0b\u9762\u7684\u4f8b\u5b50\uff1a<\/p>\n<p>\u6211\u6765\u5230\u8fbe\u89c2\u6570\u636e\u53c2\u89c2<\/p>\n<ul>\n<li>\u76f8\u5e94\u7684bigram\u7279\u5f81\u4e3a\uff1a\u6211\u6765 \u6765\u5230 \u5230\u8fbe \u8fbe\u89c2 \u89c2\u6570 \u6570\u636e \u636e\u53c2 \u53c2\u89c2<\/li>\n<li>\u76f8\u5e94\u7684trigram\u7279\u5f81\u4e3a\uff1a\u6211\u6765\u5230 \u6765\u5230\u8fbe \u5230\u8fbe\u89c2 \u8fbe\u89c2\u6570 \u89c2\u6570\u636e \u6570\u636e\u53c2 \u636e\u53c2\u89c2<\/li>\n<\/ul>\n<p>\u6ce8\u610f\u4e00\u70b9\uff1an-gram\u4e2d\u7684gram\u6839\u636e\u7c92\u5ea6\u4e0d\u540c\uff0c\u6709\u4e0d\u540c\u7684\u542b\u4e49\u3002\u5b83\u53ef\u4ee5\u662f\u5b57\u7c92\u5ea6\uff0c\u4e5f\u53ef\u4ee5\u662f\u8bcd\u7c92\u5ea6\u7684\u3002\u4e0a\u9762\u6240\u4e3e\u7684\u4f8b\u5b50\u5c5e\u4e8e\u5b57\u7c92\u5ea6\u7684n-gram\uff0c\u8bcd\u7c92\u5ea6\u7684n-gram\u770b\u4e0b\u9762\u4f8b\u5b50\uff1a<\/p>\n<p>\u6211 \u6765\u5230 \u8fbe\u89c2\u6570\u636e \u53c2\u89c2<\/p>\n<ul>\n<li>\u76f8\u5e94\u7684bigram\u7279\u5f81\u4e3a\uff1a\u6211\/\u6765\u5230 \u6765\u5230\/\u8fbe\u89c2\u6570\u636e \u8fbe\u89c2\u6570\u636e\/\u53c2\u89c2<\/li>\n<li>\u76f8\u5e94\u7684trigram\u7279\u5f81\u4e3a\uff1a\u6211\/\u6765\u5230\/\u8fbe\u89c2\u6570\u636e \u6765\u5230\/\u8fbe\u89c2\u6570\u636e\/\u53c2\u89c2<\/li>\n<\/ul>\n<p>n-gram\u4ea7\u751f\u7684\u7279\u5f81\u53ea\u662f\u4f5c\u4e3a<strong>\u300c\u6587\u672c\u7279\u5f81\u7684\u5019\u9009\u96c6\u300d<\/strong>\uff0c\u4f60\u540e\u9762\u53ef\u80fd\u4f1a\u91c7\u7528\u4fe1\u606f\u71b5\u3001\u5361\u65b9\u7edf\u8ba1\u3001IDF\u7b49\u6587\u672c\u7279\u5f81\u9009\u62e9\u65b9\u5f0f\u7b5b\u9009\u51fa\u6bd4\u8f83\u91cd\u8981\u7279\u5f81\u3002<\/p>\n<h2><strong>3 \u8bba\u6587\u7b80\u4ecb<\/strong><\/h2>\n<h3><strong>3.1 \u8bba\u6587\u80cc\u666f\u77e5\u8bc6<\/strong><\/h3>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-59df37a40664de4b250b617936c119d0_720w.webp\" alt=\"img\" \/><\/p>\n<p><strong>3.1.1 \u8bcd\u7684\u8868\u793a\u65b9\u5f0f<\/strong><\/p>\n<p><strong>\u300c1\u3001One-hot Representation \u72ec\u70ed\u8868\u793a\u300d<\/strong><\/p>\n<p>\u6bd4\u5982\u5c06\u201c\u8bdd\u7b52\u201d\u8868\u793a\u4e3a\uff1a[0, 0, 0, 0, 1, 0, 0,\u2026\u2026]\uff0c\u5c06\u201c\u9ea6\u514b\u201d\u8868\u793a\u4e3a[0, 0, 1, 0, 0, 0, 0, \u2026\u2026]\u7b49\u3002<\/p>\n<ul>\n<li>\u4f18\u70b9\uff1a\u8868\u793a\u7b80\u5355\u3002<\/li>\n<li>\u5b58\u5728\u7684\u95ee\u9898\uff1a\u8bcd\u8d8a\u591a\uff0c\u7ef4\u6570\u8d8a\u9ad8\uff0c\u7ef4\u6570\u4e3a\u8bcd\u8868\u7684\u5927\u5c0fV\uff0c\u4e14\u65e0\u6cd5\u8868\u793a\u8bcd\u548c\u8bcd\u4e4b\u95f4\u7684\u5173\u7cfb\u3002<\/li>\n<\/ul>\n<p><strong>\u300c2\u3001SVD\u300d<\/strong><\/p>\n<p>\u8fd9\u79cd\u662f\u57fa\u4e8e\u7a97\u53e3\u7684\u5171\u73b0\u77e9\u9635\uff08Window based Co-occurrence Matrix\uff09\u7684\u65b9\u6cd5\uff0c\u4e00\u4e2a\u8bcd\u53ea\u8ddf\u4e0e\u5b83\u5468\u56f4\u7279\u5b9a\u7a97\u53e3\u5185\u7684\u8bcd\u6709\u5173\uff0c\u4e24\u4e2a\u8bcd\u5171\u540c\u51fa\u73b0\u4e00\u6b21\u5c31\u52a0\u4e00\u3002\u4f8b\u5982\uff1a<\/p>\n<ul>\n<li>1\uff09I enjoy flying.<\/li>\n<li>2\uff09I like NLP.<\/li>\n<li>3\uff09I like deep learning.<\/li>\n<\/ul>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119154037385.png\" alt=\"image-20240119154037385\" \/><\/p>\n<p>\u5148\u89c4\u5b9a\u5bf9\u89d2\u7ebf\u4e3a0\uff0c\u7136\u540e\u6784\u5efa\u7a97\u53e3\u5927\u5c0f\u4e3a1\u7684\u5171\u73b0\u77e9\u9635\u3002\u901a\u8fc7\u5171\u73b0\u7684\u65b9\u6cd5\uff0c\u53ef\u4ee5\u4f53\u73b0\u4e00\u5b9a\u76f8\u4f3c\u7684\u5173\u7cfb\uff0c\u4f8b\u5982 like \u548c enjoy \u90fd\u548c I \u76f8\u90bb\uff0c\u4f46\u662f\u8fd9\u4e2a\u77e9\u9635\u5927\u5c0f\u4ecd\u7136\u548c\u8bcd\u8868\u5927\u5c0f\u6709\u5173\uff0c\u7ef4\u5ea6\u5f88\u9ad8\uff0c\u800c\u4e14\u8fd8\u662f\u6bd4\u8f83\u7a00\u758f\u3002\u56e0\u6b64\u9700\u8981\u964d\u7ef4\u5ea6\uff1a<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119154117300.png\" alt=\"image-20240119154117300\" \/>\u3002<\/p>\n<p>\u7b2c\u4e00\u79cd\u4f7f\u7528\u8bcd\u5411\u91cf\u7684\u65b9\u6cd5\uff0c\u4f7f\u7528SVD\u5206\u89e3\u3002SVD\u5206\u89e3\u53ef\u4ee5\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u5f97\u5230\u8bcd\u4e0e\u8bcd\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\uff0c\u4f46\u662f\u77e9\u9635\u592a\u5927\uff0cSVD\u5206\u89e3\u6548\u7387\u4f4e\uff0c\u5b66\u4e60\u5f97\u5230\u7684\u8bcd\u5411\u91cf\u53ef\u89e3\u91ca\u6027\u5dee\u3002<\/p>\n<p>\u5c06\u4e00\u4e2a\u77e9\u9635\u5206\u89e3\u6210\u4e09\u90e8\u5206\uff0c<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119154159981.png\" alt=\"image-20240119154159981\" \/>\u4e09\u8005\u4e0e<strong><em>X<\/em><\/strong>\u77e9\u9635\u7ef4\u5ea6\u5747\u662f<strong><em>V<\/em><\/strong>\u2217<strong><em>V<\/em><\/strong>\u3002<strong><em>\u03a3<\/em><\/strong>\u53ea\u6709\u5bf9\u89d2\u7ebf\u5143\u7d20\u4e0d\u4e3a0\uff0c\u76f8\u5f53\u4e8e<strong><em>X<\/em><\/strong>\u6240\u6709\u7279\u5f81\u503c\uff0c\u8868\u793a<strong><em>X<\/em><\/strong>\u77e9\u9635\u4e2d\u5411\u91cf<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119154324024.png\" alt=\"image-20240119154324024\" \/>\u7684\u91cd\u8981\u7a0b\u5ea6\u3002\u4e00\u822c\u60c5\u51b5\u4e0b\uff0c<strong><em>\u03a3<\/em><\/strong>\u4e2d\u7279\u5f81\u503c<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119154401139.png\" alt=\"image-20240119154401139\" \/>\u4e3a\u4ece\u5927\u5230\u5c0f\u6392\u5e8f\uff0c\u8fd9\u6837\u53ef\u4ee5\u4fdd\u7559\u524d <strong><em>k<\/em><\/strong>\u4e2a\u5927\u5c0f\u7279\u5f81\u503c\u3002\u5bf9\u5e94\u7684\uff0c\u4e5f\u53ea\u4fdd\u7559<strong><em>U<\/em><\/strong>\u4e2d\u524d<strong><em>k<\/em><\/strong>\u4e2a\u5411\u91cf\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119154443727.png\" alt=\"image-20240119154443727\" \/><\/p>\n<p>\u4f18\u70b9\uff1a\u53ef\u4ee5\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u5f97\u5230\u8bcd\u4e0e\u8bcd\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\u3002<\/p>\n<p>\u7f3a\u70b9\uff1a\u77e9\u9635\u592a\u5927\uff0cSVD\u77e9\u9635\u5206\u89e3\u6548\u7387\u4f4e\uff0c\u5b66\u4e60\u5f97\u5230\u7684\u8bcd\u5411\u91cf\u53ef\u89e3\u91ca\u6027\u5dee\u3002\u56e0\u4e3a\u820d\u5f03\u7684\u4e00\u4e9b\u4e0d\u91cd\u8981\u4fe1\u606f\u65e0\u6cd5\u89e3\u91ca\u4e0d\u91cd\u8981\u7684\u539f\u56e0\uff0c\u540c\u7406\u91cd\u8981\u7684\u4fe1\u606f\u4e5f\u65e0\u6cd5\u89e3\u91ca\u91cd\u8981\u6027\u3002<\/p>\n<p><strong>\u300c3\u3001\u5206\u5e03\u5f0f\u8868\u793a\/\u7a20\u5bc6\u8868\u793a\uff0c\u4e5f\u79f0\u4e3a\u8bcd\u5411\u91cf\/\u8bcd\u5d4c\u5165\uff08Word embedding\uff09\u300d<\/strong> \u2014\u2014 \u672c\u6587\u63d0\u51fa<\/p>\n<p>\u5b83\u7684\u7ef4\u5ea6\u4e0d\u518d\u662f\u8bcd\u5178\u7684\u5927\u5c0f\uff0c\u6bcf\u4e2a\u7ef4\u5ea6\u4e0a\u7684\u6570\u5b57\u4e0d\u518d\u662f0\uff0c\u8fd9\u4e9b\u6570\u4e4b\u95f4\u7684\u5173\u7cfb\u662f\u901a\u8fc7\u8bad\u7ec3\u5f97\u5230\u7684\uff0c\u8bad\u7ec3\u5f97\u5230\u7684\u8bcd\u5411\u91cf\u53ef\u4ee5\u5c06\u610f\u601d\u76f8\u8fd1\u7684\u8bcd\u79bb\u5f97\u5f88\u8fd1\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-2c0eeda059a73954da44dd62a567d1aa_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u6bcf\u4e2a\u8bcd\u8868\u793a\u6210<strong><em>D<\/em><\/strong>\u7ef4\u7684\u5411\u91cf\uff0c<strong><em>D&lt;&lt;V<\/em><\/strong>\u3002\u7c7b\u4f3c\u4e8e10000\u4e2a\u4eba\uff0c10000\u4e2a\u4f4d\u7f6e\uff0c\u5c06\u5176\u538b\u7f29\u4e3a300\u4e2a\u4f4d\u7f6e\uff0c\u8bad\u7ec3\u8feb\u4f7f10000\u4e2d\u76f8\u719f\u7684\u4eba\u5728\u4e00\u8d77\u8282\u7701\u4f4d\u7f6e\u3002\u76f8\u5f53\u4e8e\u7528300\u7ef4\u5ea6\u7684\u7a7a\u95f4\u5bb9\u7eb310000\u7ef4\u5ea6\u7a7a\u95f4\u7684\u8bcd\uff0c\u8feb\u4f7f\u76f8\u4f3c\u7684\u8bcd\u5728\u4e00\u8d77\uff0c\u5c55\u73b0\u975e\u5e38\u597d\u7684\u6027\u8d28\u3002\u964d\u7ef4\u53d8\u6210\u4e8c\u7ef4\uff0c\u901a\u8fc7\u4f59\u5f26\u76f8\u4f3c\u5ea6\u53ef\u4ee5\u8868\u793a\u4e24\u4e2a\u8bcd\u4e4b\u95f4\u7684\u5173\u7cfb\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-a9597f34afd7581f1694fabdddddd50d_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u53c2\u8003\u94fe\u63a5\uff1a<\/p>\n<p>\u7f8e\u56fd\u77e5\u4e4e\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/www.quora.com\/Deep-Learning-What-is-meant-by-a-distributed-representation\">https:\/\/www.quora.com\/Deep-Learning-What-is-meant-by-a-distributed-representation<\/a><\/p>\n<p><strong>\u53d1\u5c55\u5386\u7a0b<\/strong><\/p>\n<ul>\n<li>1996\u5e74\uff1aHinton 1986\u5e74\u63d0\u51faDistributed Representation<\/li>\n<li>2003\u5e74\uff1aBingio\u9996\u6b21\u4f7f\u7528\u8bcd\u5411\u91cfA Neural Probabilistic Language Model\uff0c\u4f7f\u7528\u8bad\u7ec3\u8bed\u8a00\u7684\u65b9\u5f0f\u8bad\u7ec3\u8bcd\u5411\u91cf\uff0c\u8868\u73b0\u8f83\u597d<\/li>\n<li>2003-2013\u5e74\uff0c\u63d0\u51fa\u5f88\u591a\u8bad\u7ec3\u8bcd\u5411\u91cf\u7684\u65b9\u6cd5\uff0c\u4f46\u662f\u901f\u5ea6\u8f83\u6162\uff0c\u4e0eSVD\u8f83\u50cf\uff0c\u65e0\u6cd5\u5728\u5927\u8bed\u6599\u4e0a\u8fdb\u884c\u8bad\u7ec3<\/li>\n<li>2013\u5e74\uff1a\u672c\u6587\u6539\u8fdb\u5f00\u6e90word2vec\uff0c\u8bad\u7ec3\u901f\u5ea6\u975e\u5e38\u5feb\uff0c\u53ef\u4ee5\u5728\u5927\u8bed\u6599\u4e0a\u8fdb\u884c\u8bad\u7ec3<\/li>\n<\/ul>\n<h3><strong>3.2 \u672c\u6587\u7814\u7a76\u6210\u679c<\/strong><\/h3>\n<ul>\n<li>\u63d0\u51fa\u65b0\u7684\u6a21\u578b\u7ed3\u6784word2vec<\/li>\n<li>\u63d0\u51fa\u4f18\u5316\u8bad\u7ec3\u65b9\u6cd5\uff0c\u4f7f\u5f97\u8bad\u7ec3\u901f\u5ea6\u52a0\u5feb<\/li>\n<li>\u7ed9\u51fa\u8bad\u7ec3\u4ee3\u7801word2vec\uff0c\u4f7f\u5f97\u5355\u673a\u8bad\u7ec3\u6210\u4e3a\u53ef\u80fd<\/li>\n<li>\u6210\u679c\uff1a\u8bad\u7ec3\u7684\u8bcd\u5411\u91cf\uff0c\u53c8\u5feb\u53c8\u597d\uff0c\u5e76\u4e14\u80fd\u591f\u5728\u5927\u89c4\u6a21\u8bed\u6599\u4e0a\u8fdb\u884c\u8bcd\u5411\u91cf\u7684\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<h3><strong>3.3 \u7814\u7a76\u610f\u4e49<\/strong><\/h3>\n<p>\u7814\u7a76\u610f\u4e49\uff1a<\/p>\n<ol>\n<li>\u8861\u91cf\u8bcd\u5411\u91cf\u4e4b\u95f4\u7684\u76f8\u4f3c\u7a0b\u5ea6\uff1a\u4f59\u5f26\u76f8\u4f3c\u5ea6\u548c\u8bcd\u7c7b\u6bd4\u3002<\/li>\n<li>\u4f5c\u4e3a\u9884\u8bad\u7ec3\u6a21\u578b\u63d0\u5347nlp\u4efb\u52a1<\/li>\n<\/ol>\n<p>1\u3001\u8861\u91cf\u8bcd\u5411\u91cf\u4e4b\u95f4\u7684\u76f8\u4f3c\u7a0b\u5ea6\uff0c\u4e5f\u662f\u8bc4\u4ef7\u65b9\u6cd5\uff1a\u4f59\u5f26\u76f8\u4f3c\u5ea6\u548c\u8bcd\u7c7b\u6bd4\u3002<\/p>\n<p>1\uff09\u4f59\u5f26\u76f8\u4f3c\u5ea6\uff1a<\/p>\n<p>sim(word1,word2)=cos(wordvec1,vordvec2)<\/p>\n<p>\u4e3b\u89c2\u9a8c\u8bc1\uff1a\u6bd4\u5982\u67e5\u627e\u9752\u86d9frog\u7684\u76f8\u4f3c\u8bcd\u5411\u91cf\uff0c\u90fd\u662f\u9752\u86d9\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>1<\/th>\n<th>2<\/th>\n<th>3<\/th>\n<th>4<\/th>\n<th>5<\/th>\n<th>6<\/th>\n<th>7<\/th>\n<th>8<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>frog<\/td>\n<td>frogs<\/td>\n<td>toad<\/td>\n<td>litoria<\/td>\n<td>leptodactylidae<\/td>\n<td>rana<\/td>\n<td>lizard<\/td>\n<td>eleutherodactylus<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-aca421a25a98776a4b0cc46266d76dfb_720w.webp\" alt=\"img\" \/><\/p>\n<p>2\uff09\u8bcd\u7c7b\u6bd4analogy\uff1a\uff08\u8bed\u4e49\u5173\u7cfb\uff09<\/p>\n<p>cos(word1\u2212word2+word3,wordvec4)<\/p>\n<p>\u9a8c\u8bc1\u65b9\u6cd5\u6709\u6548\u6027\uff0c\u56fd\u5bb6\u4e0e\u9996\u90fd\u7684\u5173\u7cfb\uff1a\u8ba1\u7b97France - Paris + Rome\uff0c\u67e5\u770b\u4e0e\u7ed3\u679c\u6700\u76f8\u8fd1\u7684\u8bcd\u5411\u91cf\uff0c\u5982\u679c\u662fItaly\uff0c\u8bf4\u660e\u8be5\u65b9\u6cd5\u662f\u6709\u6548\u7684<\/p>\n<ul>\n<li>France \u27f9 Paris<\/li>\n<li>Italy \u27f9 Rome<\/li>\n<\/ul>\n<p>W(&quot;woman&quot;)-W(&quot;man&quot;)=W(&quot;aunt&quot;)-W(&quot;uncle&quot;)<\/p>\n<p>W(&quot;woman&quot;)-W(&quot;man&quot;)=W(&quot;queen&quot;)-W(&quot;king&quot;)<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-08a8be02f5aadd29dd97c229012fc7ef_720w.webp\" alt=\"img\" \/><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-b6fb794894ae4865c4d36fb9e1c0c5d5_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u53c2\u8003\u94fe\u63a5\uff1a<\/p>\n<p><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/nlp.stanford.edu\/projects\/glove\/\">https:\/\/nlp.stanford.edu\/projects\/glove\/<\/a><\/p>\n<p><a href=\"https:\/\/link.zhihu.com\/?target=http%3A\/\/www.fit.vutbr.cz\/~imikolov\/rnnlm\/word-test.v1.txt\">http:\/\/www.fit.vutbr.cz\/~imikolov\/rnnlm\/word-test.v1.txt<\/a><\/p>\n<p>2\u3001\u4f5c\u4e3a\u9884\u8bad\u7ec3\u6a21\u578b\u63d0\u5347nlp\u4e0b\u6e38\u4efb\u52a1\u51c6\u786e\u5ea6\u548c\u6548\u679c<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-afd305ab10eb3601c8411b100d8a23f1_720w.webp\" alt=\"img\" \/><\/p>\n<p>1\uff09\u5916\u90e8\u4efb\u52a1\u6bd4\u5982\u547d\u540d\u4f53\u8bc6\u522b\u3001\u6587\u672c\u5206\u7c7b<\/p>\n<p>2\uff09\u5e94\u7528\u5230\u5176\u4ed6nlp\u4efb\u52a1\u4e0a\uff0c\u76f8\u5f53\u4e8e\u534a\u76d1\u7763\u8bad\u7ec3<\/p>\n<p>3\uff09\u53cc\u8bed\u5355\u8bcd\u5d4c\u5165\uff0cWord embedding\uff0c\u7c7b\u4f3c\u4e8e\u4e2d\u82f1\u6587\u8bcd\u5178<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-cab448f73f3752a6efac689774ca9a01_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u53c2\u8003\u94fe\u63a5\uff1a<\/p>\n<p><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/nlp.stanford.edu\/~socherr\/SocherGanjooManningNg_NIPS2013.pdf\">https:\/\/nlp.stanford.edu\/~socherr\/SocherGanjooManningNg_NIPS2013.pdf<\/a><\/p>\n<h3><strong>3.4 \u8bba\u6587\u603b\u89c8<\/strong><\/h3>\n<table>\n<thead>\n<tr>\n<th>\u5e8f\u53f7<\/th>\n<th>\u6807\u9898<\/th>\n<th>\u5185\u5bb9<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>\u6458\u8981 Abstract<\/td>\n<td>\u63d0\u51fa2\u79cd\u65b0\u7684\u9ad8\u6548\u8ba1\u7b97\u8bcd\u5411\u91cf\u7ed3\u6784\uff0c\u5e76\u4f7f\u7528\u8bcd\u76f8\u4f3c\u5ea6\u4efb\u52a1\u9a8c\u8bc1\u6548\u679c<\/td>\n<\/tr>\n<tr>\n<td>1<\/td>\n<td>Introduction<\/td>\n<td>\u4ecb\u7ecd\u8bcd\u5411\u91cf\u80cc\u666f\uff1b\u672c\u6587\u76ee\u6807\uff1b\u524d\u4eba\u5de5\u4f5c<\/td>\n<\/tr>\n<tr>\n<td>2<\/td>\n<td>Model Architectures<\/td>\n<td>LSA\/LDA\uff1b\u524d\u5411\u795e\u7ecf\u7f51\u7edc\uff1b\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff1b\u5e76\u884c\u7f51\u7edc\u8ba1\u7b97<\/td>\n<\/tr>\n<tr>\n<td>3<\/td>\n<td>New Log-linear Models<\/td>\n<td>\u300c\u4ecb\u7ecd2\u4e2d\u65b0\u6a21\u578b\u7ed3\u6784\uff1aCBOW, Skip grams\u300d<\/td>\n<\/tr>\n<tr>\n<td>4<\/td>\n<td>Results<\/td>\n<td>\u8bc4\u4ef7\u4efb\u52a1\u63cf\u8ff0\uff1b\u6700\u5927\u5316\u6b63\u786e\u7387\uff1b\u6a21\u578b\u7ed3\u6784\u6bd4\u8f83\uff1b\u6a21\u578b\u4e0a\u5927\u91cf\u6570\u636e\u7684\u5e76\u884c\u8ba1\u7b97\uff1b\u5fae\u8f6f\u7814\u7a76\u9662\u53e5\u5b50\u5b8c\u6210\u6bd4\u8d5b<\/td>\n<\/tr>\n<tr>\n<td>5<\/td>\n<td>Examples of the Learned Relationships<\/td>\n<td>\u4f8b\u5b50\uff1a\u5b66\u4e60\u5230\u7684\u8bcd\u4e0e\u8bcd\u4e4b\u95f4\u7684\u5173\u7cfb<\/td>\n<\/tr>\n<tr>\n<td>6<\/td>\n<td>Conclusion<\/td>\n<td>\u7ed3\u8bba\uff1a\u9ad8\u8d28\u91cf\u8bcd\u5411\u91cf\uff1b\u9ad8\u6548\u7387\u8bad\u7ec3\u65b9\u5f0f\uff1b\u4f5c\u4e3a\u9884\u8bad\u7ec3\u8bcd\u5411\u91cf\u4f5c\u7528\u4e8e\u5176\u4ed6nlp\u4efb\u52a1\u80fd\u63d0\u5347\u6548\u679c<\/td>\n<\/tr>\n<tr>\n<td>7<\/td>\n<td>Follow-Up Work<\/td>\n<td>\u540e\u7eed\u5de5\u4f5c\uff1aC++\u5355\u673a\u4ee3\u7801<\/td>\n<\/tr>\n<tr>\n<td>8<\/td>\n<td>References<\/td>\n<td>\u53c2\u8003\u6587\u732e<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u672c\u6587\u5b9e\u9645\u8bb22\u7bc7\u8bba\u6587\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-ea92d42cbf0c7877e8c70804227ac333_720w.webp\" alt=\"img\" \/><\/p>\n<h2><strong>4 \u8bba\u6587\u6cdb\u8bfb<\/strong><\/h2>\n<h3><strong>4.1 \u6458\u8981<\/strong><\/h3>\n<ol>\n<li>\u63d0\u51fa\u4e86\u4e24\u79cd\u65b0\u9896\u7684\u6a21\u578b\u7ed3\u6784\u7528\u6765\u8ba1\u7b97\u8bcd\u5411\u91cf<\/li>\n<li>\u91c7\u7528\u4e00\u79cd\u8bcd\u76f8\u4f3c\u5ea6\u7684\u4efb\u52a1\u6765\u8bc4\u4f30\u5bf9\u6bd4\u8bcd\u5411\u91cf\u8d28\u91cf<\/li>\n<li>\u5927\u91cf\u964d\u4f4e\u6a21\u578b\u8ba1\u7b97\u91cf\u53ef\u4ee5\u63d0\u5347\u8bcd\u5411\u91cf\u8d28\u91cf<\/li>\n<li>\u8fdb\u4e00\u6b65\uff0c\u5728\u6211\u4eec\u7684\u8bed\u4e49\u548c\u53e5\u6cd5\u4efb\u52a1\u4e0a\uff0c\u6211\u4eec\u7684\u8bcd\u5411\u91cf\u662f\u5f53\u524d\u6700\u597d\u7684\u6548\u679c<\/li>\n<\/ol>\n<h3><strong>4.2 Introduction<\/strong><\/h3>\n<ul>\n<li>\u4f20\u7edfNLP\u628a\u8bcd\u5f53\u505a\u6700\u5c0f\u5355\u5143\u5904\u7406\uff0c\u5e76\u4e14\u80fd\u591f\u5728\u5927\u8bed\u6599\u5e93\u4e0a\u5f97\u5230\u5f88\u597d\u7684\u7ed3\u679c\uff0c\u5176\u4e2d\u4e00\u4e2a\u4f8b\u5b50\u662fn-grams\u6a21\u578b<\/li>\n<li>\u7136\u800c\u5f88\u591a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u53ea\u80fd\u63d0\u4f9b\u5f88\u5c0f\u7684\u8bed\u6599\uff0c\u5982\u8bed\u97f3\u8bc6\u522b\u3001\u673a\u5668\u7ffb\u8bd1\uff0c\u7b80\u5355\u5730\u6269\u5927\u6570\u636e\u89c4\u6a21\u6765\u63d0\u5347\u7b80\u5355\u6a21\u578b\u7684\u8868\u73b0\u5728\u8fd9\u4e9b\u4efb\u52a1\u4e0d\u518d\u9002\u7528\uff0c\u6240\u4ee5\u6211\u4eec\u5fc5\u987b\u5bfb\u627e\u66f4\u52a0\u5148\u8fdb\u7684\u6a21\u578b<\/li>\n<li>\u5206\u5e03\u5f0f\u8868\u793a\u53ef\u4ee5\u5728\u5927\u8bed\u6599\u4e0a\u8bad\u7ec3\u5f97\u5230\u5f88\u597d\u7684\u8bed\u8a00\u6a21\u578b\uff0c\u5e76\u4e14\u80fd\u591f\u8d85\u8fc7n-grams\u6a21\u578b\uff0c\u8fd9\u662f\u4e00\u4e2a\u5f88\u597d\u7684\u53ef\u4ee5\u4f5c\u4e3a\u6539\u8fdb\u7684\u6280\u672f\u3002<\/li>\n<\/ul>\n<h3><strong>4.3 word2vec\u8bcd\u5411\u91cf\u8bc4\u4ef7\u65b9\u6cd5<\/strong><\/h3>\n<p>\u5185\u90e8\u4efb\u52a1\u8bc4\u4ef7\uff1a\u8bad\u7ec3\u597d\u4ec0\u4e48\u4efb\u52a1\u6765\u8bc4\u4ef7\u4ec0\u4e48\u4efb\u52a1\u3002\u5982\u4e0b\u56fe\u4e3a\u4e94\u79cd\u8bed\u4e49\uff0c\u4e5d\u79cd\u8bed\u6cd5\u8bc4\u4ef7\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-b6fb794894ae4865c4d36fb9e1c0c5d5_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u5916\u90e8\u4efb\u52a1\u8bc4\u4ef7\uff1a\u5916\u90e8\u4efb\u52a1\u6bd4\u5982\u547d\u540d\u5b9e\u4f53\u8bc6\u522b\u3001\u6587\u672c\u5206\u7c7b\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-05cadd60d1bec4afe2b4528bfc801415_720w.webp\" alt=\"img\" \/><\/p>\n<h2><strong>5 \u5bf9\u6bd4\u6a21\u578b\uff08\u8bba\u6587Model Architectures\u90e8\u5206\uff09<\/strong><\/h2>\n<p>\u5728word2vec\u63d0\u51fa\u4e4b\u524d\uff0cNNLM\u4e0eRNNLM\u901a\u8fc7\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u7684\u65b9\u6cd5\u7684\u8bad\u7ec3\u8bcd\u5411\u91cf\uff0c\u4f7f\u7528\u7edf\u8ba1\u7684\u65b9\u6cd5\u3002<\/p>\n<p>\u672c\u8282\u4e3b\u8981\u5bf9\u4ee5\u4e0b\u4e09\u79cd\u6a21\u578b\u8fdb\u884c\u5bf9\u6bd4\uff1a<\/p>\n<ul>\n<li>Feedforward Neural Net Language Model<\/li>\n<li>Recurrent Neural Net Language Model<\/li>\n<li>Parallel Training of Neural networks<\/li>\n<\/ul>\n<h3><strong>5.1 \u524d\u9988\u795e\u7ecf\u7f51\u7edc\u8bed\u8a00\u6a21\u578b(NNLM)<\/strong><\/h3>\n<h4>5.1.1 Feedforward Neural Net Language Model<\/h4>\n<p>\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u8bed\u8a00\u6a21\u578b\u5168\u79f0\uff1aFeedforward Neural Net Language Model\uff0c\u7b80\u79f0NNLM\uff0c\u4e5f\u53eb\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\uff0c\u5168\u8fde\u63a5\u795e\u7ecf\u7f51\u7edc\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-3c00c90767e7cc5afa13a76b04c0dbac_720w.webp\" alt=\"img\" \/><\/p>\n<blockquote>\n<p>\u53c2\u8003\u6587\u732e\uff1aBengio A neural probabilistic language model (2003)<\/p>\n<\/blockquote>\n<p>\u5728nlp\u4efb\u52a1\u4e2d\uff0c\u9996\u5148\u90fd\u4f1a\u6784\u5efaword2id\u4ee5\u53caid2word\uff0c\u5b83\u4eec\u90fd\u662f\u5b57\u5178\u7c7b\u578b\uff0cword2id\u7684key\u662fword\uff0cvalue\u662fid\uff0cid2word\u76f8\u53cd\uff0c\u8868\u793a\u6bcf\u4e2a\u8bcd\u5bf9\u5e94\u4e00\u4e2aid\u3002<\/p>\n<ul>\n<li>input layer\uff1a\u8f93\u5165\u5c42\uff0c\u8f93\u5165\u7684\u662findex\uff0cindex\u8868\u793a\u8bcd\u7684\u7f16\u53f7\uff0c\u628a\u6240\u6709\u8bcd\u653e\u5728\u4e00\u8d77\u7ec4\u6210\u4e00\u4e2a\u8bcd\u5178\uff0c\u8fd9\u6837\u6bcf\u4e2a\u8bcd\u90fd\u6709\u4e00\u4e2a\u7f16\u53f7\uff0c\u7528\u7f16\u53f7\u8868\u793a\u4f4d\u7f6e\u3002\u5c06\u6bcf\u4e2aindex\u6620\u5c04\u6210\u4e00\u4e2a\u5411\u91cf\uff0c\u5c06\u8fd9\u4e9b\u5411\u91cf\u8fdb\u884c\u62fc\u63a5(concat)\uff0c\u5982\u56fe\u4e2d\u8f93\u5165\u7684\u662fn-1\u4e2a\u8bcd\uff0c\u6bcf\u4e2a\u8bcd\u90fd\u662f100\u7ef4\uff0c\u56e0\u6b64\u62fc\u63a5\u5b8c\u6210\u540e\u53d8\u6210100\u00d7(n-1)\u7ef4\uff1b<\/li>\n<li>hidden layer\uff1a\u9690\u85cf\u5c42\uff0c\u5c06\u4e0a\u4e00\u6b65\u5f97\u5230\u7684100\u00d7(n-1)\u7ef4\u5411\u91cf\u8f93\u5165\u5230\u5168\u8fde\u63a5\u5c42\u4e2d\uff0c\u4f7f\u7528tanh\u6fc0\u6d3b\u51fd\u6570\uff1b<\/li>\n<li>output layer\uff1a\u8f93\u51fa\u5c42\uff0c\u4e0a\u4e00\u6b65\u5f97\u5230\u7684\u7ed3\u679c\u518d\u63a5\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\uff0c\u7136\u540e\u4f7f\u7528softmax\u3002<\/li>\n<\/ul>\n<p>\u6839\u636e\u524dn-1\u4e2a\u5355\u8bcd\uff0c\u9884\u6d4b\u7b2cn\u4e2a\u4f4d\u7f6e\u5355\u8bcd\u7684<strong>\u6982\u7387<\/strong>\uff0c\u5373<strong>n-gram<\/strong>\u6a21\u578b\u3002\u5c06<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119155053874.png\" alt=\"image-20240119155053874\" \/>\u8f93\u5165\u5230\u7f51\u7edc\u8f93\u5165\u5c42\uff0c\u6620\u5c04\u5411\u91cf\uff0c\u8fdb\u884c\u62fc\u63a5\uff0c\u7ecf\u8fc7\u4e24\u4e2a\u5168\u8fde\u63a5\u5c42\u4f7f\u7528softmax\u9884\u6d4b\u7b2c <strong><em>t<\/em><\/strong>\u4e2a\u8bcd\u7684\u6982\u7387\uff0c\u8fd9\u6837\u53ef\u4ee5\u5f97\u5230\u53e5\u5b50\u4e2d\u6bcf\u4e2a\u8bcd\u7684\u6982\u7387\uff0c\u4ece\u800c\u5f97\u5230\u53e5\u5b50\u7684\u6982\u7387\u3002<\/p>\n<blockquote>\n<p>\u8bed\u8a00\u6a21\u578b\u7684\u4f18\u70b9\u5728\u4e8e\u5b83\u662f\u65e0\u76d1\u7763\u6a21\u578b\uff0c<strong>\u4e0d\u9700\u8981\u4efb\u4f55\u6807\u6ce8\u8bed\u6599<\/strong>\uff0c\u722c\u53d6\u7f51\u7ad9\u4e0a\u76f8\u5173\u53e5\u5b50\u5373\u53ef\u3002\u901a\u8fc7\u524dn-1\u4e2a\u5355\u8bcd\u9884\u6d4b\u7b2cn\u4e2a\u5355\u8bcd\uff0c\u4f7f\u7528\u65e0\u76d1\u7763\u7684\u65b9\u5f0f\u8fdb\u884c\u76d1\u7763\u5b66\u4e60\u3002<\/p>\n<\/blockquote>\n<p>\u6784\u5efa\u6a21\u578b\uff0c\u9996\u5148\u5728\u4e00\u4e2a\u53e5\u5b50\u4e4b\u524dpad\uff08\u586b\u5145\uff09\u7528\u6765\u9884\u6d4b\u7b2c\u4e00\u4e2a\u8bcd\uff0c\u751f\u6210\u8bed\u6599\uff0c\u524dn-1\u4e2a\u8bcd\u4f5c\u4e3a\u7279\u5f81\uff0c\u4e0b\u4e00\u4e2a\u8bcd\u4f5c\u4e3a\u6807\u7b7e\uff0c\u4f9d\u6b21\u7c7b\u63a8\uff0c\u53ef\u4ee5\u6784\u9020\u9664\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u7684\u8bad\u7ec3\u6240\u9700\u6570\u636e\u96c6\u3002\u7136\u540e\u901a\u8fc7\u68af\u5ea6\u4e0b\u964d\u4f18\u5316\u6a21\u578b\uff0c\u4f7f\u5f97\u8f93\u51fa\u7684\u6b63\u786e\u7684\u5355\u8bcd\u6982\u7387\u6700\u5927\u5316\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-b0b2eaac14a9869278edd1c10f4ec252_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u8f93\u5165\u5c42\uff1a\u5c06\u8bcd\uff08index\uff09\u6620\u5c04\u6210\u5411\u91cf\uff0c\u76f8\u5f53\u4e8e\u4e2a<strong><em>1\u00d7V<\/em><\/strong>\u7684one-hot\u5411\u91cf\u4e58\u4ee5<strong><em>V\u00d7D<\/em><\/strong>\u7684\u5411\u91cf\u5f97\u5230\u4e00\u4e2a<strong><em>1\u00d7D<\/em><\/strong>\u7684\u5411\u91cf\uff0c\u5176\u4e2d<strong><em>V\u00d7D<\/em><\/strong>\u7684\u5411\u91cf\u6bcf\u4e00\u884c\u8868\u793a\u4e00\u4e2a\u8bcd\u5411\u91cf\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119155156008.png\" alt=\"image-20240119155156008\" \/><\/p>\n<p>\u5176\u4e2d<strong><em>d<\/em><\/strong>\u4e3a\u504f\u7f6e\uff0c<strong><em>U<\/em><\/strong>\u662f\u53c2\u6570\uff0c<strong><em>d<\/em><\/strong> \u548c <strong><em>U<\/em><\/strong>\u76f8\u5f53\u4e8e <strong><em>b<\/em><\/strong> \u548c <strong><em>W<\/em><\/strong>\u53c2\u6570\u3002<\/p>\n<p>\u8f93\u51fa\u5c42\uff1a\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\uff0c\u540e\u9762\u63a5\u4e2a softmax \u51fd\u6570\u6765\u751f\u6210\u6982\u7387\u5206\u5e03\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119155242235.png\" alt=\"image-20240119155242235\" \/><\/p>\n<p>\u5176\u4e2d <strong><em>y<\/em><\/strong> \u662f\u4e00\u4e2a<strong><em>1\u00d7V<\/em><\/strong>\u7684\u5411\u91cf\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119155340009.png\" alt=\"image-20240119155340009\" \/><\/p>\n<h4>5.1.2 \u8bed\u8a00\u6a21\u578b\u56f0\u60d1\u5ea6\u548cLoss\u7684\u5173\u7cfb<\/h4>\n<p>\u5982\u4f55\u6c42\u56f0\u60d1\u5ea6\uff1f<\/p>\n<p>\u9700\u8981\u5c06\u6bcf\u4e2a\u53e5\u5b50\u4e2d\u6bcf\u4e2a\u8bcd\u7684\u6982\u7387\u7ed9\u9884\u6d4b\u51fa\u6765\uff0c\u7136\u540e\u5c06\u4ed6\u4eec\u76f8\u4e58\u5417\uff1f\u8fd9\u6837\u505a\u4e5f\u662f\u53ef\u4ee5\u7684\uff0c\u4f46\u5728\u5b9e\u9645\u4e2d\u6709\u66f4\u52a0\u65b9\u4fbf\u5de7\u5999\u7684\u65b9\u6cd5\u3002<\/p>\n<p>\u5148\u5199\u51faLoss\uff0c\u4f7f\u5f97\u9884\u6d4b\u51fa\u5f97\u5230\u7684\u6bcf\u4e2a\u6b63\u786e\u5355\u8bcd\u5f97\u5230\u7684\u6982\u7387\u8d8a\u5927\u8d8a\u597d\uff0c\u4f7f\u7528\u4e86\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570\uff0c\u8fd9\u4e2aLoss\u662f\u4e00\u53e5\u8bdd\u7684Loss\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119155403000.png\" alt=\"image-20240119155403000\" \/><\/p>\n<p>\u5176\u4e2d\uff0c<strong><em>T<\/em><\/strong> \u8868\u793a\u53e5\u5b50\u4e2d\u8bcd\u7684\u4e2a\u6570\u3002\u56f0\u60d1\u5ea6\u662f\u53e5\u5b50\u7684\u6982\u7387\u5f00\u8d1f<strong><em>T<\/em><\/strong>\u5206\u4e4b\u4e00\u3002<\/p>\n<p>\u8fd9\u6837\u5728\u5f97\u51faloss\u7684\u60c5\u51b5\u4e0b\uff0c\u53ef\u4ee5\u5f88\u65b9\u4fbf\u7684\u6c42\u51fa\u56f0\u60d1\u5ea6\u662f\u591a\u5c11\uff0c\u770b\u770b\u56f0\u60d1\u5ea6\u7684\u60c5\u51b5\u3002\u73b0\u5728\u90fd\u662f\u6279\u6b21\u8fdb\u884c\u8bad\u7ec3\u548c\u9884\u6d4b\u7684\uff0c\u6bcf\u8ba1\u7b97batchsize\u5927\u5c0f\u7684\u4e2a\u6570\u8fdb\u884c\u4e00\u6b21loss\u7684\u8ba1\u7b97\u3002<\/p>\n<p>NLP\u4e2d\uff0c\u6bcf\u4e2a\u53e5\u5b50\u957f\u5ea6\u4e0d\u4e00\u81f4\uff0c\u4e3a\u4e86\u6bcf\u4e2abatchsize\u4e2d\u7684\u53e5\u5b50\u957f\u5ea6\u76f8\u540c\uff0c\u9700\u8981\u8865pad\uff0c\u5bf9pad\u4f4d\u7f6e\u6c42loss\u662f\u4e0e\u5176\u4ed6\u8bcd\u4e00\u8d77\u8ba1\u7b97\uff0c\u5f97\u5230\u7684loss\u6c42\u6307\u6570\u5f97\u5230\u7684\u4e0d\u662f\u771f\u5b9e\u7684\u56f0\u60d1\u5ea6\uff0c\u53ef\u80fd\u504f\u5c0f\uff0c\u6240\u4ee5\u9700\u8981\u5c06pad\u62b5\u6d88\u6389\u3002<\/p>\n<h4>5.1.3 \u56de\u987e\u7f51\u7edc\u6a21\u578b<\/h4>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-ef178294bba782e845a36a6a0da3c649_720w.webp\" alt=\"img\" \/><\/p>\n<ul>\n<li><strong>\u4ec5\u5bf9\u4e00\u90e8\u5206\u8f93\u51fa\u8fdb\u884c\u68af\u5ea6\u4f20\u64ad<\/strong>\uff1a\u6bd4\u5982\u5355\u8bcdthe\u3001a\u3001and\u7b49\u4ed6\u4eec\u6240\u5305\u542b\u7684\u4fe1\u606f\u6bd4\u8f83\u5c11\uff0c\u4f46\u662f\u5728\u8bed\u6599\u4e2d\u53c8\u9891\u7e41\u51fa\u73b0\uff0c\u56e0\u6b64\u8fd9\u4e9b\u8bcd\u80fd\u4e0d\u80fd\u6216\u8005\u5c11\u8fdb\u884c\u4e0d\u8fdb\u884c\u68af\u5ea6\u4f20\u64ad\u3002<\/li>\n<li><strong>\u5f15\u5165\u5148\u9a8c\u77e5\u8bc6\uff0c\u5982\u8bcd\u6027\u7b49<\/strong>\uff1a\u6bd4\u5982\u5f62\u5bb9\u8bcd\u540e\u63a5\u540d\u8bcd\u6982\u7387\u8f83\u5927\uff0c\u52a0\u52a8\u8bcd\u6982\u7387\u8f83\u5c0f\u3002\u52a0\u8bcd\u6027\u4e4b\u524d\uff0c\u5148\u8003\u8651\u8fd9\u4e2a\u6a21\u578b\u672c\u8eab\u662f\u5426\u4f1a\u5b66\u5230\u8bcd\u6027\u4fe1\u606f\uff0c\u5982\u679c\u6a21\u578b\u672c\u8eab\u5c31\u53ef\u4ee5\u5b66\u4e60\u5230\u8bcd\u6027\u4fe1\u606f\uff0c\u90a3\u4e48\u5c31\u4e0d\u9700\u8981\u518d\u52a0\u4e86\u3002\u4e4b\u540e\u67e5\u770b\u7f51\u7edc\u5b66\u4e60\u5230\u7684\u8bcd\u6027\u662f\u5426\u591f\u591a\uff0c\u53ef\u4ee5\u5148\u52a0\u4e0a\u8bcd\u6027\u4fe1\u606f\u770b\u770b\u6548\u679c\u662f\u5426\u66f4\u597d\u3002<\/li>\n<li><strong>\u89e3\u51b3\u4e00\u8bcd\u591a\u4e49\u95ee\u9898<\/strong>\uff1a\u4e00\u4e2a\u8bcd\u5c31\u4e00\u4e2a\u5411\u91cf\uff0c\u90a3\u4e48\u4e00\u4e2a\u5411\u91cf\u600e\u4e48\u80fd\u8868\u793a\u591a\u4e2a\u610f\u601d\u5462\uff1f\u6bd4\u5982\u673a\u5668\u7ffb\u8bd1\u4e2d\uff0c\u53ef\u4ee5\u901a\u8fc7\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u8be6\u7ec6\u53ef\u89c1elmo\u8bba\u6587\u3002<\/li>\n<li><strong>\u52a0\u901fsoftmax\u5c42<\/strong>\uff1asoftmax\u5c42\u7684\u795e\u7ecf\u5143\u4e2a\u6570\u662f\u5b57\u5178V\u7684\u4e2a\u6570\uff0c\u56e0\u4e3a\u6bcf\u4e2a\u8bcd\u90fd\u8981\u8f93\u51fa\u4e00\u4e0b\u6982\u7387\uff0c\u8fd9\u6837\u5c31\u5f88\u6162\u3002<\/li>\n<\/ul>\n<h3><strong>5.2 \u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u8bed\u8a00\u6a21\u578b(RNNLM)<\/strong><\/h3>\n<p>\u5168\u79f0\uff1aRecurrent Neural Net Language Model<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-66e8d9815ee1316279483691549ab25a_720w.webp\" alt=\"img\" \/><\/p>\n<p><strong><em>s(t-1)<\/em><\/strong>\u662f\u4e0a\u4e00\u4e2a\u65f6\u95f4\u6b65\u7684\u8f93\u51fa\uff0c<strong><em>w(t)<\/em><\/strong>\u662f\u5f53\u524d\u8bcd\u5411\u91cf(\u4e5f\u662f\u901a\u8fc7one-hot\u65b9\u5f0f\u8f6c\u6362\u7684)\u7684\u8f93\u5165\u3002<\/p>\n<p>\u8f93\u5165\u5c42\uff1a\u548cNNLM\u4e00\u6837\uff0c\u9700\u8981\u5c06\u5f53\u524d\u65f6\u95f4\u6b65\u7684\u8f6c\u5316\u4e3a\u8bcd\u5411\u91cf\u3002<\/p>\n<p>\u9690\u85cf\u5c42\uff1a\u5bf9\u8f93\u5165\u548c\u4e0a\u4e00\u4e2a\u65f6\u95f4\u6b65\u7684\u9690\u85cf\u8f93\u51fa\u8fdb\u884c\u5168\u8fde\u63a5\u5c42\u64cd\u4f5c\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119155749917.png\" alt=\"image-20240119155749917\" \/><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-5ae72b122020fe10ae29672b2a4768f1_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u6bcf\u4e2a\u65f6\u95f4\u6b65\u9884\u6d4b\u4e00\u4e2a\u8bcd \uff0c\u5728\u9884\u6d4b\u7b2cn\u4e2a\u8bcd\u65f6\u4f7f\u7528\u4e86\u524dn - 1\u4e2a\u8bcd\u7684\u4fe1\u606f\uff0c\u6ca1\u6709\u4f7f\u7528\u9a6c\u5c14\u79d1\u592b\u5047\u8bbe\u4e86\uff0c\u57fa\u4e8e\u7edf\u8ba1\u7684\u601d\u60f3\u3002<code>&lt;S&gt;<\/code>\u8868\u793a\u5f00\u59cb\uff0c\u7b2c\u4e00\u4e2a\u8bcd\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119155904109.png\" alt=\"image-20240119155904109\" \/><\/p>\n<h3><strong>5.3 word2vec\u6a21\u578b<\/strong><\/h3>\n<p>Log Linear Models\u5b9a\u4e49\uff1a\u5c06\u8bed\u8a00\u6a21\u578b\u7684\u5efa\u7acb\u770b\u6210\u4e00\u4e2a\u591a\u5206\u7c7b\u95ee\u9898\uff0c\u76f8\u5f53\u4e8e\u7ebf\u6027\u5206\u7c7b\u5668\u52a0\u4e0asoftmax\u3002\u56e0\u4e3a\u53ea\u6709softmax\u6c42\u6307\u6570\u90e8\u5206\u662f\u975e\u7ebf\u6027\uff0c\u4f46\u662f\u52a0\u4e0alog\u5c31\u53d8\u6210\u4e86\u7ebf\u6027\uff0c\u6240\u4ee5\u79f0\u4e3alog\u7ebf\u6027\u6a21\u578b\u3002\u903b\u8f91\u65af\u8482\u56de\u5f52\u6a21\u578b\u5c31\u662fLog Linear Models\u3002<\/p>\n<p><strong>Y=softmax(wx+b)<\/strong><\/p>\n<p>\u672c\u8282\u8bb2\u8ff0\u7684skip-gram\u4e0eCBOW\u4e5f\u662fLog Linear Models\u3002<\/p>\n<h4>5.3.1 word2vec\u539f\u7406<\/h4>\n<p><strong>\u8df3\u5b57\u6a21\u578b\uff1a<\/strong><\/p>\n<ul>\n<li><strong>\u8bed\u8a00\u6a21\u578b<\/strong>\u57fa\u672c\u601d\u60f3\uff1a\u53e5\u5b50\u4e2d\u4e0b\u4e00\u4e2a\u8bcd\u7684\u51fa\u73b0\u548c\u524d\u9762\u7684\u8bcd\u662f\u6709\u5173\u7cfb\u7684\uff0c\u6240\u4ee5\u53ef\u4ee5\u4f7f\u7528<strong>\u524d\u9762\u7684\u8bcd<\/strong>\u9884\u6d4b<strong>\u4e0b\u4e00\u4e2a\u8bcd<\/strong>\u3002\u9a6c\u5c14\u53ef\u592b\u5047\u8bbe\uff0c\u6bd4\u5982NNLM\uff0c\u4ee5\u53ca\u7edf\u8ba1\u601d\u60f3\uff0c\u524d\u9762\u6240\u6709\u8bcd\u9884\u6d4b\u4e0b\u4e00\u4e2a\uff0c\u5982RNNLM\u3002<\/li>\n<li><strong>Word2vec<\/strong>\u57fa\u672c\u601d\u60f3\uff1a\u53e5\u5b50\u4e2d<strong>\u76f8\u8fd1\u7684\u8bcd<\/strong>\u4e4b\u95f4\u662f\u6709\u8054\u7cfb\u7684\uff0c\u6bd4\u5982\u4eca\u5929\u540e\u9762\u7ecf\u5e38\u51fa\u73b0\u4e0a\u5348\u3001\u4e0b\u5348\u548c\u665a\u4e0a\u3002\u6240\u4ee5Word2vec\u7684\u57fa\u672c\u601d\u60f3\u5c31\u662f<strong>\u7528\u8bcd\u6765\u9884\u6d4b\u8bcd<\/strong>\uff0c<strong>skip-gram<\/strong>\u4f7f\u7528\u4e2d\u5fc3\u8bcd\u9884\u6d4b\u5468\u56f4\u8bcd\uff0c<strong>cbow<\/strong>\u4f7f\u7528\u5468\u56f4\u8bcd\u9884\u6d4b\u4e2d\u5fc3\u8bcd\u3002word2vec\u7684\u57fa\u672c\u601d\u60f3\u4e0a\u5c31\u5bf9\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u4e86\u7b80\u5316\u3002<\/li>\n<\/ul>\n<h4>5.3.2 Skip-gram<\/h4>\n<p><strong>\u8ba1\u7b97\u8fc7\u7a0b\uff1a<\/strong><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-a4033f1865bbc24cd27853614475f672_720w.webp\" alt=\"img\" \/><\/p>\n<p>skip-gram\u4f7f\u7528\u4e2d\u5fc3\u8bcd\u9884\u6d4b\u5468\u56f4\u8bcd\uff0c\u4ec0\u4e48\u8303\u56f4\u5185\u662f\u4e2d\u5fc3\u8bcd\u548c\u5468\u56f4\u8bcd\uff1f\u8fd9\u91cc\u9009\u62e9\u201c\u4e0b\u5348\u201d\u4f5c\u4e3a\u4e2d\u5fc3\u8bcd\uff0c\u7a97\u53e3\u4e3a2\uff0c\u56e0\u6b64<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160034290.png\" alt=\"image-20240119160034290\" \/>\u8868\u793a\u7528\u201c\u4e0b\u5348\u201d\u9884\u6d4b\u201c\u6211\u201d\uff0c<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160045754.png\" alt=\"image-20240119160045754\" \/>\u8868\u793a\u7528\u201c\u4e0b\u5348\u201d\u9884\u6d4b\u201c\u4eca\u5929\u201d\uff0c<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160058974.png\" alt=\"image-20240119160058974\" \/>\u8868\u793a\u7528\u201c\u4e0b\u5348\u201d\u9884\u6d4b\u201c\u6253\u201d\uff0c<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160112201.png\" alt=\"image-20240119160112201\" \/>\u8868\u793a\u7528\u201c\u4e0b\u5348\u201d\u9884\u6d4b\u201c\u7fbd\u6bdb\u7403\u201d\u3002\u6839\u636e\u7a97\u53e3\u7684\u8bad\u7ec3\u5927\u5c0f\uff0c\u8fd9\u91cc\u4ea7\u751f\u56db\u4e2a\u8bad\u7ec3\u6837\u672c\u3002<\/p>\n<p>\u5982\u4f55\u6c42\u4e0a\u56fe\u4e2d\u7684\u6982\u7387\uff1f\u5176\u5b9e\u8bed\u8a00\u6a21\u578b\u4e2d\u6982\u7387\u95ee\u9898\u90fd\u662f\u4e00\u4e2a\u591a\u5206\u7c7b\u95ee\u9898\uff0c\u8f93\u5165\u5c31\u662f<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160124087.png\" alt=\"image-20240119160124087\" \/>\uff0c\u800c\u6807\u7b7e\u5c31\u662f<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160136697.png\" alt=\"image-20240119160136697\" \/>\u76f8\u5f53\u4e8e\u4e00\u4e2a\u8bcd\u8868\u5927\u5c0f\u7684\u591a\u5206\u7c7b\u95ee\u9898\uff0c\u600e\u4e48\u6c42\u6982\u7387\u4ee5\u53ca\u5b66\u4e60\u8bcd\u5411\u91cf\uff1f<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160151749.png\" alt=\"image-20240119160151749\" \/>\uff1a\u7528\u4e2d\u5fc3\u8bcd<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160124087-1705651324309-94.png\" alt=\"image-20240119160124087\" \/>\u9884\u6d4b\u5468\u56f4\u8bcd<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160136697-1705651332464-97.png\" alt=\"image-20240119160136697\" \/>\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-420e73430d08b7d6fcd1b4363a585951_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u8f93\u5165\u662findex\uff0c\u4ee3\u8868\u8fd9\u4e2a\u8bcd\u5728\u8bcd\u8868\u4e2d\u7684\u4f4d\u7f6e\uff0c\u6620\u5c04\u6210<strong><em>1\u00d7V<\/em><\/strong>\u7684one-hot\u5411\u91cf\uff1b<strong>W\u662f\u4e2d\u5fc3\u8bcd\u7684\u8bcd\u5411\u91cf\u77e9\u9635<\/strong>\uff0c\u7ef4\u5ea6\u4e3aV\u00d7D\uff0c\u6bcf\u4e00\u884c\u8868\u793a\u4e00\u4e2a\u8bcd\u7684\u8bcd\u5411\u91cf\uff1b\u5c06one-hot\u5411\u91cf\u4e0e\u4e2d\u5fc3\u8bcd\u7684\u8bcd\u5411\u91cf\u77e9\u9635\u76f8\u4e58\u5f97\u5230\u4e00\u4e2a\u8bcd\u7684\u5411\u91cf\uff1b\u5f97\u52301\u00d7D\u7684\u8bcd\u5411\u91cf\u4e0e<strong>\u5468\u56f4\u8bcd\u7684\u8bcd\u5411\u91cf\u77e9\u9635<\/strong>\u548c*<strong>W**<\/strong>\u8fdb\u884c\u76f8\u4e58\uff0c\u5f97\u52301\u00d7V\u7684\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u5728\u7ecf\u8fc7softmax\u4e4b\u540e\uff0c\u5f97\u5230\u6bcf\u4e2a\u8bcd\u7684\u6982\u7387\uff0c\u7136\u540e\u4f7f\u5f97\u6b63\u786e\u8bcd\u7684\u4f4d\u7f6e\u4e0a\u7684\u6982\u7387\u8d8a\u5927\u8d8a\u597d\u3002<\/p>\n<p>\u56fe\u4e2d<strong><em>W<\/em><\/strong>\u548c<strong>*W***<em>\u5c31\u662f\u60f3\u8bad\u7ec3\u5f97\u5230\u7684\u8bcd\u5411\u91cf\uff0c\u4e00\u822c\u7528<\/em><\/strong>W<strong><em>\u5c31\u53ef\u4ee5\uff0c\u6216\u8005\u7528<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160423251.png\" alt=\"image-20240119160423251\" \/>\u30021\u00d7D\u7684\u8bcd\u5411\u91cf\u4ee3\u8868\u7684\u662f\u4e2d\u5fc3\u8bcd<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160513922.png\" alt=\"image-20240119160513922\" \/>\u7684\u8bcd\u5411\u91cf\uff0c<\/em><\/strong>W****\u662f\u5468\u56f4\u8bcd\u7684\u8bcd\u5411\u91cf\u77e9\u9635\uff0c\u5b83\u4eec\u76f8\u4e58\u540e\u518d\u505asoftmax\uff0c\u628a<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160545252.png\" alt=\"image-20240119160545252\" \/>\u8fd9\u4e2a\u4f4d\u7f6e\u7684\u5355\u8bcd\u53d6\u51fa\u6765\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160556932.png\" alt=\"image-20240119160556932\" \/><\/p>\n<p>\u5206\u6bcd\uff1a\u4e2d\u5fc3\u8bcd\u5411\u91cf\u4e0e\u6240\u6709\u5468\u56f4\u8bcd\u5411\u91cf\u505a\u5185\u79ef\uff0c\u6c42\u548c\u3002 \u5206\u5b50\uff1a<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160545252-1705651568718-100.png\" alt=\"image-20240119160545252\" \/>\u7684\u5468\u56f4\u8bcd\u5411\u91cf\u7684\u8f6c\u7f6e\u4e0e\u4e2d\u5fc3\u8bcd\u5411\u91cf\u5185\u79ef\u3002<\/p>\n<p><strong>\u635f\u5931\u51fd\u6570\uff1a<\/strong><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-4584769f4a8ccbf7a148231c7710df2f_720w.webp\" alt=\"img\" \/><\/p>\n<blockquote>\n<p><strong>\u7b80\u5355\u76f4\u89c2\u4e0d\u51c6\u786e\u89e3\u91ca<\/strong>\uff1a\u5728\u8bcd\u8868\u4e2d\u5c06\u4e2d\u5fc3\u8bcdants\u6240\u6709\u5468\u56f4\u8bcd\u53d6\u51fa\uff0c\u653e\u5728\u4e00\u4e2a\u888b\u5b50\u91cc\uff0c\u5728\u888b\u5b50\u91cc\u7528\u624b\u968f\u610f\u6293\u4e00\u4e2a\uff0c\u8fd9\u4e2a\u8bcd\u662fcar\u7684\u6982\u7387\u3002\u7c7b\u4f3c\u4e8e\u5171\u73b0\u77e9\u9635\u7684\u6b21\u6570\u3002\u5982\u679c\u8981\u662fcar\u4e0eant\u90fd\u4e0e\u540c\u4e00\u4e2a\u8bcd\u6709\u5173\u7cfb\uff0c\u4e8c\u8005\u53ef\u80fd\u5b58\u5728\u4e00\u5b9a\u7684\u4f9d\u8d56\u5173\u7cfb\uff0c\u51c6\u786e\u5ea6\u53ef\u80fd\u4f1a\u8f83\u9ad8\uff0c\u6240\u4ee5\u8fd9\u79cd\u89e3\u91ca\u5e76\u4e0d\u51c6\u786e\u3002<\/p>\n<\/blockquote>\n<p>\u6c42\u6700\u5c0f\u503c\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160634873-1705651611702-103.png\" alt=\"\" \/><\/p>\n<h4>5.3.3 CBOW<\/h4>\n<p>\u5168\u79f0\uff1aContinuous Bag-of-Words Model\uff0c\u8bcd\u888b\u6a21\u578b\uff0c\u5ffd\u7565\u4e86\u8bcd\u7684\u987a\u5e8f\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-a94d8a18431f3a6c30a44291abb5264a_720w.webp\" alt=\"img\" \/><\/p>\n<p>cbow\u4f7f\u7528\u5468\u56f4\u8bcd\u9884\u6d4b\u4e2d\u5fc3\u8bcd\uff0cNNLM\u4e2d\u5c06\u8f93\u5165\u6240\u6709\u6b64concat\u8fde\u63a5\u5728\u4e00\u8d77\uff0c\u8f93\u5165\u8f83\u5927\uff0c\u6c42\u548c\u6216\u8005\u6c42\u5e73\u5747\u8f93\u5165\u4e0e\u5355\u4e2a\u8bcd\u5411\u91cf\u7ef4\u5ea6\u76f8\u540c\uff0c\u8f93\u5165\u8d8a\u5c0f\uff0c\u6a21\u578b\u590d\u6742\u5ea6\u8d8a\u4f4e\u3002\u8fd9\u91cc\u91c7\u7528\u7684\u6c42\u548c\u3002<\/p>\n<p>\u5bf9\u4e8e\u4e2d\u5fc3\u8bcd\u6765\u8bf4\u8fd9\u91cc\u53ea\u6709\u4e00\u4e2a\u8bad\u7ec3\u7684\u6837\u672c\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160720863.png\" alt=\"image-20240119160720863\" \/> <\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-9027a7fccce59ebc7c4a92a1bdcba571_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u8f93\u5165<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160739802.png\" alt=\"image-20240119160739802\" \/>\u7684index\uff0c\u8f6c\u6362\u4e3a1\u00d7V\u5927\u5c0f\u7684one-hot\u5411\u91cf\u3002<\/p>\n<p>\u8f93\u5165\u662findex\uff0c\u4ee3\u8868\u8fd9\u4e2a\u8bcd\u5728\u8bcd\u8868\u4e2d\u7684\u4f4d\u7f6e\uff0c\u6620\u5c04\u62101\u00d7V\u7684one-hot\u5411\u91cf\uff1b\u4e0eSkip-gram\u76f8\u53cd\uff0c\u8fd9\u91cc<strong>W\u662f\u5468\u56f4\u8bcd\u7684\u8bcd\u5411\u91cf\u77e9\u9635<\/strong>\uff0c\u7ef4\u5ea6\u4e3aV\u00d7D\uff0c\u6bcf\u4e00\u884c\u8868\u793a\u4e00\u4e2a\u5468\u56f4\u8bcd\u7684\u8bcd\u5411\u91cf\uff1b\u5c06\u4e8c\u8005\u76f8\u4e58\u5f97\u5230\u6bcf\u4e2a\u5468\u56f4\u8bcd\u7684\u8bcd\u5411\u91cf\uff1b\u5728\u7ecf\u8fc7\u5e73\u5747\u6216\u8005\u6c42\u548c\u65b9\u5f0f\u5f97\u52301\u00d7D\u7684\u5411\u91cf\u4e0e\u4e0e<strong>\u4e2d\u5fc3\u8bcd\u7684\u8bcd\u5411\u91cf\u77e9\u9635<\/strong>\u548c*<strong>W**<\/strong>\u8fdb\u884c\u76f8\u4e58\uff0c\u5f97\u52301\u00d7V\u7684\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u5728\u7ecf\u8fc7softmax\u4e4b\u540e\uff0c\u5f97\u5230\u6bcf\u4e2a\u8bcd\u7684\u6982\u7387\uff0c\u7136\u540e\u4f7f\u5f97\u6b63\u786e\u8bcd\u7684\u4f4d\u7f6e\u4e0a\u7684\u6982\u7387\u8d8a\u5927\u8d8a\u597d\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160757265.png\" alt=\"image-20240119160757265\" \/><\/p>\n<h2><strong>6 \u5173\u952e\u6280\u672f<\/strong><\/h2>\n<h3><strong>6.1 \u590d\u6742\u5ea6\u8ba8\u8bba<\/strong><\/h3>\n<p>\u5982\u4f55\u964d\u4f4e\u590d\u6742\u5ea6\uff1f<\/p>\n<p>Softmax \u5bf9\u4e8e\u8981\u8f93\u51faV\u4e2a\u6982\u7387\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119160831781.png\" alt=\"image-20240119160831781\" \/><\/p>\n<p><strong><em>U<\/em><\/strong>\u662f\u5468\u56f4\u8bcd\u7684\u77e9\u9635\uff0c<strong><em>V<\/em><\/strong>\u662f\u4e2d\u5fc3\u8bcd\u7684\u77e9\u9635\uff0c\u7b2c\u4e09\u4e2a\u77e9\u9635\u76f8\u4e58\u76f8\u5f53\u4e8e\u662f\u4e00\u4e2a\u6ca1\u6709\u6fc0\u6d3b\u51fd\u6570\u4e5f\u6ca1\u6709\u504f\u7f6e\u7684\u5168\u8fde\u63a5\u5c42\uff0c\u5168\u8fde\u63a5\u5c42\u989d\u795e\u7ecf\u5143\u4e2a\u6570\u662f\u8bcd\u8868\u5927\u5c0f\uff0c\u56e0\u4e3asoftmax\u5c42\u8981\u8f93\u51fa<strong><em>V<\/em><\/strong>\u4e2a\u6982\u7387\uff0c<strong><em>V<\/em><\/strong>\u662f\u7279\u522b\u5927\u7684\uff0c\u56e0\u6b64\u8fd9\u4e2a\u5168\u8fde\u63a5\u5c42\u4e5f\u662f\u7279\u522b\u5927\u7684\uff0c\u90a3\u4e48\u5982\u4f55\u964d\u4f4e\u590d\u6742\u5ea6\u5462\uff1f\u80fd\u4e0d\u80fd\u5c06softmax\u7684\u590d\u6742\u5ea6\u964d\u4f4e\u5462\uff1f\u8bba\u6587\u4e2d\u4e5f\u63d0\u5230\u4e86\u4e24\u79cd\u65b9\u6cd5\uff0c\u5c42\u6b21softmax\u548c\u8d1f\u91c7\u6837\u3002<\/p>\n<h3><strong>6.2 Hierarchical Softmax<\/strong><\/h3>\n<p><strong>\u5c06softmax\u8f6c\u53d8\u4e3a\u591a\u4e2asigmoid<\/strong>\uff0c\u53ef\u4ee5\u5199\u6210\u4e00\u4e2a\u4e8c\u53c9\u6811\u7684\u5f62\u5f0f\u3002softmax\u9700\u8981\u505a<strong><em>V<\/em><\/strong>\u6b21\u6307\u6570\u64cd\u4f5c\uff0csigmoid\u53ea\u9700\u8981\u4e00\u6b21\u6307\u6570\u8fd0\u7b97\u5373\u53ef\u3002\u53ea\u8981\u8f6c\u6362\u4e3a\u5c11\u4e8e<strong><em>V<\/em><\/strong>\u6b21\u6307\u6570\u64cd\u4f5c\u7684sigmoid\u64cd\u4f5c\uff0c\u5c31\u76f8\u5f53\u4e8e\u8fdb\u884c\u4e86\u52a0\u901f\u8fd0\u7b97\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-2eae443f897142038e72b1ec040a7590_720w.webp\" alt=\"img\" \/><\/p>\n<p>log2\u2061<strong><em>V<\/em><\/strong>\u5c42\uff0c\u53ea\u9700\u8ba1\u7b97 log2\u2061<strong><em>V<\/em><\/strong> \u4e2a sigmoid\u8fd0\u7b97\u3002<\/p>\n<p><strong>6.2.1 Skip-gram\u76ee\u6807\u51fd\u6570<\/strong><\/p>\n<p><strong>\u6ee1\u4e8c\u53c9\u6811<\/strong>\uff1a \u4e3a\u4ec0\u4e48\u5c42\u6b21softmax\u53ea\u9700\u8ba1\u7b97 log2\u2061<strong><em>V<\/em><\/strong>\u4e2asigmoid\uff1f<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161152878.png\" alt=\"image-20240119161152878\" \/><\/p>\n<p>\u53ea\u9700\u89813\u6b21\u5c31\u80fd\u627e\u5230\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-f97ab281139be375707818feb3171577_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u5047\u8bbe\u8bcd\u8868\u5927\u5c0f\u662f8\uff0csoftmax\u9700\u8981\u505a8\u6b21\u6307\u6570\u64cd\u4f5c\uff0c\u5982\u679c\u8ba1\u7b97sigmoid\uff0c\u901a\u8fc78\u4e2a\u8282\u70b9\u7684\u6ee1\u4e8c\u53c9\u6811\uff0c\u5224\u65ad\u4e3aa\uff0c\u9700\u8981\u8fdb\u884c3\u6b21\u4e8c\u5206\u7c7b\uff0c\u5224\u65ad\u4e3ab\uff0c\u4e5f\u9700\u8981\u8fdb\u884c3\u6b21\u4e8c\u5206\u7c7b\uff0c\u56e0\u6b64\uff0c\u5206\u7c7b\u6b21\u6570=\u6811\u7684\u6df1\u5ea6-1\u3002<\/p>\n<p><strong>\u6784\u5efaHuffman\u6811<\/strong>\uff1a\u5e26\u6743\u91cd\u8def\u5f84\u6700\u77ed\u4e8c\u53c9\u6811<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-d932720240e0d10f7bf669dd3472aec3_720w.webp\" alt=\"img\" \/><\/p>\n<p>5\u8868\u793a\u9884\u6d4b\u8fd9\u4e2a\u8bcd\u9700\u8981\u505a5\u6b21sigmoid\uff0c7\u8868\u793a\u9884\u6d4b\u8fd9\u4e2a\u8bcd\u9700\u8981\u505a7\u6b21sigmoid\uff0c\u8fb9\u7684\u4e2a\u6570\u8868\u793a\u6743\u91cd\uff0c\u54c8\u592b\u66fc\u6811\u5c06\u9891\u7387\u9ad8\u7684\u653e\u5728\u6811\u7684\u4e0a\u5c42\uff0c\u5c11\u4e8elog2\u2061<strong><em>V<\/em><\/strong>\u6b21\u3002<\/p>\n<p>\u5c42\u6b21softmax\u6784\u5efa\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-f762c72374f6ba0e5f1175b37299bfd2_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u6bcf\u4e2a\u5206\u652f\u8282\u70b9\u90fd\u662f\u4e00\u4e2a\u5411\u91cf\uff0c\u4f8b\u5982\u5728skip-gram\u4e2d\u7528\u4e2d\u5fc3\u8bcd\u5411\u91cf\u4e0e<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161236656.png\" alt=\"image-20240119161236656\" \/>\u76f8\u4e58\uff0c\u4e4b\u540e\u8fdb\u884csigmoid\uff0c\u5982\u679c\u5c0f\u4e8e\u96f6\uff0c\u8868\u793a\u5f80\u5de6\u8fdb\u884c\u67e5\u627e\uff0c\u5982\u679c\u5927\u4e8e\u96f6\uff0c\u8868\u793a\u5411\u53f3\u8fdb\u884c\u67e5\u627e\uff0c\u518d\u4e0e<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161312060.png\" alt=\"image-20240119161312060\" \/>\u76f8\u4e58\uff0c\u540c\u7406\u5982\u679c\u5c0f\u4e8e\u96f6\uff0c\u8868\u793a\u5f80\u5de6\u8fdb\u884c\u67e5\u627e\uff0c\u5982\u679c\u5927\u4e8e\u96f6\uff0c\u8868\u793a\u5411\u53f3\u8fdb\u884c\u67e5\u627e\uff0c\u518d\u4e0e<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161325814.png\" alt=\"image-20240119161325814\" \/>\u76f8\u4e58\u3002\u4e4b\u524d\u7684\u8bcd\u5411\u91cf\u4e3a<strong><em>V<\/em><\/strong>\u4e2a\uff0c\u8fd9\u91cc\u7684\u8bcd\u5411\u91cf\u4e2a\u6570\u5c11\u4e8e<strong><em>V<\/em><\/strong>\u4e2a\u3002\u6c42<strong><em>I<\/em><\/strong>\u7684\u6982\u7387\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161427479.png\" alt=\"image-20240119161427479\" \/><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161457670.png\" alt=\"image-20240119161457670\" \/><\/p>\n<p>\u5728\u6811\u7684n\u5c42\u7684\u8282\u70b9\u9700\u8981\u505an-1\u5c42sigmoid\u6b21\u8fd0\u7b97\u3002<\/p>\n<p><strong>6.2.2 CBOW Hierarchical Softmax<\/strong><\/p>\n<p>\u5c42\u6b21softmax\u5206\u7c7b<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-588ffc7bb1d6f147f811f8940b208643_720w.webp\" alt=\"img\" \/><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161531929.png\" alt=\"image-20240119161531929\" \/><\/p>\n<p>\u4e4b\u524d\u7684skip-gram\u4e0eCBOW\u90fd\u662f\u6700\u7ec8\u8bcd\u5411\u91cf\u53ef\u4ee5\u662f\u4e2d\u5fc3\u8bcd\u5411\u91cf\u4e0e\u5468\u56f4\u8bcd\u5411\u91cf\u7684\u76f8\u52a0\u6216\u8005\u5e73\u5747\uff0c\u4f46\u662f\u5728\u5c42\u6b21softmax\u4e2d\u53ea\u6709\u4e00\u7ec4\u8bcd\u5411\u91cf\uff0c\u56e0\u4e3a<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161600334.png\" alt=\"image-20240119161600334\" \/>\u4e2a\u6570\u5c11\u4e8e<strong><em>V<\/em><\/strong>\u4e2a\uff0c\u5176\u5177\u4f53\u542b\u4e49\u5e76\u4e0d\u660e\u786e\uff0c\u6bd4\u5982<img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161236656-1705652178581-110.png\" alt=\"image-20240119161236656\" \/>\u53ef\u80fd\u8868\u793a\u4e00\u7c07\u8bcd\u7c7b\u522b\u7684\u610f\u4e49\u3002<\/p>\n<h3><strong>6.3 Negative Sample\uff08\u5e94\u7528\u8f83\u591a\uff09<\/strong><\/h3>\n<p>\u5728\u6548\u679c\u4e0e\u6548\u7387\u4e0a\uff0c\u8d1f\u91c7\u6837\uff08Negative Sample\uff09\u5747\u597d\u4e8e\u5c42\u6b21softmax\u3002<\/p>\n<p>\u820d\u5f03\u591a\u5206\u7c7b\uff0c\u63d0\u5347\u901f\u5ea6\uff0c<strong>\u5c06\u591a\u5206\u7c7b\u53d8\u4e3a\u4e8c\u5206\u7c7b<\/strong>\u3002\u4e8c\u5206\u7c7b\u9700\u8981\u6b63\u6837\u672c\u4e0e\u8d1f\u6837\u672c\uff0c\u4e00\u4e2a\u4e2d\u5fc3\u8bcd\u4e0e\u5468\u56f4\u8bcd\u5728\u4e00\u8d77\u662f\u6b63\u6837\u672c\uff0c\u4f8b\u5982\u56fe\u4e2d\u7684jump over\uff0c\u5728skip-gram\u91cc\uff0c\u4e2d\u5fc3\u8bcd+\u8bcd\u8868\u4efb\u610f\u4e00\u8bcd\u4f5c\u4e3a\u8d1f\u6837\u672c\uff0c\u4f8b\u5982\u56fe\u4e2d\u7684jumps again\uff0c\u8d1f\u6837\u672c\u53ef\u80fd\u6b63\u597d\u9009\u62e9\u4e86\u4e2d\u5fc3\u8bcd\u4e0e\u5176\u5468\u56f4\u8bcd\uff0c\u4f46\u8bcd\u8868\u975e\u5e38\u5927\uff0c\u53ef\u80fd\u6027\u975e\u5e38\u5c0f\uff0c\u53ef\u5ffd\u7565\u4e0d\u8ba1\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-68d10de68dd5024e70f1567ae1429f75_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u57fa\u672c\u601d\u60f3\uff1a\u589e\u5927\u6b63\u6837\u672c\u6982\u7387\uff0c\u51cf\u5c0f\u8d1f\u6837\u672c\u6982\u7387\u3002<\/p>\n<p>\u5bf9\u4e8e\u6bcf\u4e2a\u8bcd\uff0c\u4e00\u6b21\u8981\u8f93\u51fa1\u4e2a\u6b63\u6837\u672c\u7684\u6982\u7387\uff0c\u6ce8\u610f\u4e0d\u80fd\u53ea\u9009\u62e9\u4e00\u4e2a\u8d1f\u6837\u672c\uff0c\u5426\u5219\u4f1a\u6709\u504f\u5dee\uff0c\u56e0\u4e3a\u6b63\u6837\u672c\u662f\u771f\u5b9e\u7684\uff0c\u8d1f\u6837\u672c\u662f\u91c7\u6837\u51fa\u6765\u7684\uff0c\u4e00\u822c\u9009\u62e93-10\u4e2a\u4e4b\u95f4\u8d1f\u6837\u672c\u3002\u5047\u5982\u9009\u62e9<strong><em>K<\/em><\/strong>\u4e2a\u8d1f\u6837\u672c\uff0c\u5219\u8f93\u51fa<strong><em>K<\/em><\/strong>\u4e2a\u8d1f\u6837\u672c\u6982\u7387\uff0c\u603b\u5171<strong><em>K<\/em><\/strong>+1\u4e2a\u6982\u7387\uff0c<strong>$K&lt;<\/strong>\u5e76\u4e14\u6548\u679c\u6bd4\u591a\u5206\u7c7b\u8981\u597d\u3002\u8fd9\u91cc\u8fd8\u662f\u9700\u8981\u6bcf\u4e2a\u8bcd\u7684\u4e0a\u4e0b\u6587\u8bcd\u5411\u91cf\uff0c\u77e9\u9635\u7ef4\u5ea6\u4e3a<strong><em>(V,D)<\/em><\/strong>\uff0c\u603b\u7684\u53c2\u6570\u6bd4HS\uff08\u5c42\u6b21softmax\uff09\u591a\uff08\u6bcf\u6b21\u8ba1\u7b97\u91cf\u4e0d\u591a\uff0c\u5c42\u6b21softmax\u9700\u8981log2\u2061<strong><em>V<\/em><\/strong>\u4e2a\u6982\u7387\uff0c\u8fd9\u91cc\u4ec5\u9700\u8981<strong><em>K<\/em><\/strong>+1\u4e2a\uff09\u3002<\/p>\n<p>\u635f\u5931\u51fd\u6570\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161754650.png\" alt=\"image-20240119161754650\" \/><\/p>\n<p><strong>6.3.1 \u5982\u4f55\u91c7\u6837\uff1f<\/strong><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161817844.png\" alt=\"image-20240119161817844\" \/><\/p>\n<p>\u51cf\u5c11\u9891\u7387\u5927\u7684\u8bcd\u7684\u62bd\u6837\u6982\u7387\uff0c\u589e\u52a0\u9891\u7387\u5c0f\u7684\u8bcd\u7684\u62bd\u6837\u6982\u7387\u3002\u56e0\u4e3a\u4e00\u4e9b\u4e0d\u91cd\u8981\u7684\u8bcd\u51fa\u73b0\u7684\u6982\u7387\u8f83\u5927\uff0c\u6bd4\u5982the\uff0ca\u7b49\uff0c\u91cd\u8981\u7684\u8bcd\u51fa\u73b0\u7684\u6982\u7387\u8f83\u5927\u3002\u53ef\u4ee5\u52a0\u901f\u8bad\u7ec3\uff0c\u5f97\u5230\u66f4\u597d\u7684\u6548\u679c\u3002<\/p>\n<p><strong>6.3.2 CBOW Negative Sampling<\/strong><\/p>\n<p>\u540c\u7406\uff0c\u4e8c\u5206\u7c7b\u9700\u8981\u6b63\u6837\u672c\u4e0e\u8d1f\u6837\u672c\uff0c\u6240\u6709\u5468\u56f4\u8bcd\u7684\u8bcd\u5411\u91cf\u7684\u5e73\u5747\u4e0e\u771f\u6b63\u7684\u4e2d\u5fc3\u8bcd\u5728\u4e00\u8d77\u662f\u6b63\u6837\u672c\uff0c\u6240\u6709\u5468\u56f4\u8bcd\u7684\u8bcd\u5411\u91cf\u7684\u5e73\u5747\u4e0e\u968f\u673a\u91c7\u6837\u7684\u8bcd\u7ec4\u5408\u5728\u4e00\u8d77\u662f\u8d1f\u6837\u672c\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-1387f5de636cd40660c034607946999c_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u635f\u5931\u51fd\u6570\u4e0e\u4e0a\u8ff0\u76f8\u540c\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161844487.png\" alt=\"image-20240119161844487\" \/><\/p>\n<h3><strong>6.4 \u9ad8\u9891\u8bcd\u91cd\u91c7\u6837(Subsampling of Frequent Words)<\/strong><\/h3>\n<p><strong>\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5171\u8bc6<\/strong>\uff1a\u6587\u6863\u6216\u8005\u6570\u636e\u96c6\u4e2d\u51fa\u73b0<strong>\u9891\u7387\u9ad8<\/strong>\u7684\u8bcd\u5f80\u5f80\u643a\u5e26<strong>\u4fe1\u606f\u8f83\u5c11\uff0c\u6bd4\u5982the\uff0cis\uff0ca\uff0cand<\/strong>\uff0c\u800c\u51fa\u73b0<strong>\u9891\u7387\u4f4e<\/strong>\u7684\u8bcd\u5f80\u5f80\u643a\u5e26<strong>\u4fe1\u606f\u591a<\/strong>\u3002<\/p>\n<p><strong>\u4ec0\u4e48<\/strong>\u662f<strong>\u9752\u86d9<\/strong>\uff1f<\/p>\n<ul>\n<li><strong>\u4ec0\u4e48<\/strong>\u662f\u6df1\u5ea6\u5b66\u4e60\uff0c<strong>\u4ec0\u4e48<\/strong>\u662fCNN\uff0c<strong>\u4ec0\u4e48<\/strong>\u662fRNN\uff0c\u8fd9\u7bc7\u544a\u8bc9\u4f60\u3002<\/li>\n<li><strong>\u9752\u86d9<\/strong>\uff08Frog\uff09\u5c5e\u4e8e\u810a\u7d22\u52a8\u7269\u95e8\u3001\u4e24\u6816\u7eb2\u3001\u65e0\u5c3e\u76ee\u3001\u86d9\u79d1\u7684\u4e24\u6816\u7c7b\u52a8\u7269\uff0c\u6210\u4f53\u65e0\u5c3e\uff0c\u5375\u4ea7\u4e8e\u6c34\u4e2d\u3002<\/li>\n<\/ul>\n<p>\u91cd\u91c7\u6837\u7684\u539f\u56e0\uff1a<\/p>\n<ul>\n<li>\u60f3\u66f4\u591a\u5730\u8bad\u7ec3\u91cd\u8981\u7684\u8bcd\u5bf9\uff0c\u6bd4\u5982\u8bad\u7ec3\u201cFrance\u201d\u548c\u201cParis\u201d\u4e4b\u95f4\u7684\u5173\u7cfb\u6bd4\u8bad\u7ec3\u201cFrance\u201d\u548c\u201cthe\u201d\u4e4b\u95f4\u7684\u5173\u7cfb\u8981\u6709\u7528\u3002<\/li>\n<li>\u9ad8\u9891\u8bcd\u5f88\u5feb\u5c31\u8bad\u7ec3\u597d\u4e86\uff0c\u800c\u4f4e\u9891\u6b21\u9700\u8981\u66f4\u591a\u7684\u8f6e\u6b21\u3002<\/li>\n<\/ul>\n<p>\u91cd\u91c7\u6837\u65b9\u6cd5\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161901533.png\" alt=\"image-20240119161901533\" \/><\/p>\n<p>\u4f18\u70b9\uff1a\u52a0\u901f\u8bad\u7ec3\uff0c\u80fd\u591f\u5f97\u5230\u66f4\u597d\u7684\u8bcd\u5411\u91cf\u3002<\/p>\n<h2><strong>7 \u6a21\u578b\u590d\u6742\u5ea6<\/strong><\/h2>\n<h3><strong>7.1 \u6a21\u578b\u590d\u6742\u5ea6\u7684\u6982\u5ff5<\/strong><\/h3>\n<p>\u6a21\u578b\u7684\u590d\u6742\u5ea6\u4e5f\u5c31\u662f\u65f6\u95f4\u4e0a\u7684\u590d\u6742\u5ea6\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119161922985.png\" alt=\"image-20240119161922985\" \/><\/p>\n<p>\u8bba\u6587\u4e2d\u6c42 <strong><em>Q<\/em><\/strong> \u7684\u65b9\u6cd5\u6bd4\u8f83\u7279\u6b8a\uff0c\u4f7f\u7528\u53c2\u6570\u7684\u6570\u76ee\u6765\u8ba1\u7b97\u65f6\u95f4\u590d\u6742\u5ea6\u3002<\/p>\n<h3><strong>7.2 \u57fa\u4e8e\u524d\u9988\u7f51\u7edc\u7684\u8bed\u8a00\u6a21\u578b\u7684\u65f6\u95f4\u590d\u6742\u5ea6<\/strong><\/h3>\n<p>NNLM\u901a\u8fc7\u524d\u9762<strong><em>N<\/em><\/strong>\u4e2a\u8bcd\uff0c\u9884\u6d4b\u7b2c<strong><em>N<\/em><\/strong>+1\u4e2a\u8bcd\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162011736.png\" alt=\"image-20240119162011736\" \/><\/p>\n<p>\u6a21\u578b\u590d\u6742\u5ea6\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162033602.png\" alt=\"image-20240119162033602\" \/><\/p>\n<blockquote>\n<p>Bengio A neural probabilistic language mode\/(2003)<\/p>\n<\/blockquote>\n<h3><strong>7.3 \u57fa\u4e8e\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7684\u8bed\u8a00\u6a21\u578b\u7684\u65f6\u95f4\u590d\u6742\u5ea6<\/strong><\/h3>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162126724.png\" alt=\"image-20240119162126724\" \/><\/p>\n<h3><strong>7.4 Skip-gram\u590d\u6742\u5ea6<\/strong><\/h3>\n<p><strong>7.4.1 Hierarchical Softmax \u590d\u6742\u5ea6<\/strong><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162149248.png\" alt=\"image-20240119162149248\" \/><\/p>\n<p><strong>7.4.2 Negative Sampling \u590d\u6742\u5ea6<\/strong><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162204333.png\" alt=\"image-20240119162204333\" \/><\/p>\n<p>\u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u4e2d\u5fc3\u8bcd\u6765\u8bf4\uff0c\u6709\u4e00\u4e2a\u6b63\u6837\u672c\u548cK\u4e2a\u8d1f\u6837\u672c\uff0c\u6bcf\u4e2a\u90fd\u4f1a\u53d8\u6210\u4e00\u4e2aD\u7ef4\u7684\u8bcd\u5411\u91cf\uff0c\u4fbf\u5f97\u5230\u4e0a\u56fe\u4e2d\u7b2c\u4e8c\u4e2a\u516c\u5f0f\u3002<\/p>\n<h3><strong>7.5 CBOW \u590d\u6742\u5ea6<\/strong><\/h3>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162223052.png\" alt=\"image-20240119162223052\" \/> <\/p>\n<p>\u4e0eskip-gram\u76f8\u540c\uff0cCBOW\u5bf9\u5e94\u4e8e\u5c42\u6b21softmax\u548c\u8d1f\u91c7\u6837\u7684\u590d\u6742\u5ea6\u4e3a\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162239700.png\" alt=\"image-20240119162239700\" \/> <\/p>\n<p>\u6ce8\u610f\uff1a\u4e0a\u8ff0\u662f\u6bcf\u4e00\u6b21\u8ba1\u7b97\u6240\u9700\u7684\u53c2\u6570\u3002<\/p>\n<h3><strong>7.6 \u6a21\u578b\u590d\u6742\u5ea6\u5bf9\u6bd4<\/strong><\/h3>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-c4ee0e79ec5f2242c6bd7d4bb1f826d9_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u4ece\u7406\u8bba\u5206\u6790\uff1a\u5c42\u6b21softmax\u4e0e\u8d1f\u91c7\u6837\u901f\u5ea6\u5747\u5feb\u4e8e\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u4e0e\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff0cCBOW\u901f\u5ea6\u5feb\u4e8eSkip-Gram\uff0c\u8d1f\u91c7\u6837\u901f\u5ea6\u5feb\u4e8e\u5c42\u6b21softmax\u3002<\/p>\n<h2><strong>8 \u5b9e\u9a8c\u7ed3\u679c\u5206\u6790<\/strong><\/h2>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-6a222cd59fd577f0da0ea8d4b4deadea_720w.webp\" alt=\"img\" \/><\/p>\n<h3><strong>8.1 \u4efb\u52a1\u63cf\u8ff0<\/strong><\/h3>\n<p>\u4efb\u52a1\u662f\u4e00\u4e2a\u8bcd\u5bf9\u63a8\u7406\u7684\u4efb\u52a1\uff0c\u4e0b\u56fe\u4e2d\u524d\u4e94\u4e2a\u662f\u8bed\u4e49\u7c7b\uff0c\u540e\u4e5d\u4e2a\u662f\u8bed\u6cd5\u7c7b\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-18d953ca3e3c6abd0a3b939b2f68c2c9_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u8bed\u6cd5\u7c7b\u6570\u636e\u6613\u6536\u96c6\uff0c\u8bed\u4e49\u7c7b\u6570\u636e\u96be\u6536\u96c6\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u5173\u7cfb\u7c7b\u578b<\/th>\n<th>\u7ffb\u8bd1<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>capital-common-countries<\/td>\n<td>\u5e38\u89c1\u56fd\u5bb6\u9996\u90fd<\/td>\n<\/tr>\n<tr>\n<td>capital-world<\/td>\n<td>\u5404\u56fd\u9996\u90fd<\/td>\n<\/tr>\n<tr>\n<td>currency<\/td>\n<td>\u8d27\u5e01<\/td>\n<\/tr>\n<tr>\n<td>city-in-state<\/td>\n<td>\u5dde-\u57ce\u5e02<\/td>\n<\/tr>\n<tr>\n<td>family<\/td>\n<td>\u5bb6\u5ead\u5173\u7cfb<\/td>\n<\/tr>\n<tr>\n<td>gram1-adjective-to-adverb<\/td>\n<td>\u5f62\u5bb9\u8bcd-\u526f\u8bcd<\/td>\n<\/tr>\n<tr>\n<td>gram2-opposite<\/td>\n<td>\u53cd\u4e49\u8bcd<\/td>\n<\/tr>\n<tr>\n<td>gram3-comparative<\/td>\n<td>\u6bd4\u8f83\u7ea7<\/td>\n<\/tr>\n<tr>\n<td>gram4-superlative<\/td>\n<td>\u6700\u9ad8\u7ea7<\/td>\n<\/tr>\n<tr>\n<td>gram5-present-participle<\/td>\n<td>\u73b0\u5728\u8fdb\u884c\u5f0f<\/td>\n<\/tr>\n<tr>\n<td>gram6-nationality-adjective<\/td>\n<td>\u56fd\u5bb6\u7684\u5f62\u5bb9\u8bcd<\/td>\n<\/tr>\n<tr>\n<td>gram7-past-tense<\/td>\n<td>\u8fc7\u53bb\u5f0f<\/td>\n<\/tr>\n<tr>\n<td>gram8-plural<\/td>\n<td>\u590d\u6570<\/td>\n<\/tr>\n<tr>\n<td>gram9-plural-verb<\/td>\n<td>\u7b2c\u4e09\u4eba\u79f0\u5355\u6570<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Word2vec\u7a0b\u5e8f\u4e2d questions-words.txt\uff1a<\/p>\n<pre><code class=\"language-text\">\/\/ Copyright 2013 Google Inc. All Rights Reserved.\n: capital-common-countries\nAthens Greece Baghdad Iraq\nAthens Greece Bangkok Thailand\nAthens Greece Beijing China\nAthens Greece Berlin Germany\nAthens Greece Bern Switzerland\nAthens Greece Cairo Egypt\nAthens Greece Canberra Australia\nAthens Greece Hanoi Vietnam\nAthens Greece Havana Cuba\nAthens Greece Helsinki Finland\n...<\/code><\/pre>\n<h3><strong>8.2 \u6700\u5927\u5316\u6b63\u786e\u7387\uff08\u4f18\u5316\u53c2\u6570\uff09<\/strong><\/h3>\n<p>\u7528\u5c0f\u6570\u636e\u96c6\u8c03\u53c2\uff0c\u9009\u62e9\u6700\u597d\u7684\u53c2\u6570\uff0c\u53c2\u6570\u5305\u62ec\u7ef4\u5ea6\u3001\u8bad\u7ec3\u7684\u6570\u636e\u91cf\uff0c\u662f2\u4e2a\u9700\u8981\u5bfb\u627e\u7684\u53c2\u6570\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162319105.png\" alt=\"image-20240119162319105\" \/> <\/p>\n<h3><strong>8.3 \u6a21\u578b\u6bd4\u8f83<\/strong><\/h3>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162335370.png\" alt=\"image-20240119162335370\" \/><\/p>\n<p>MSR Word Relatedness\u6d4b\u8bd5\u96c6\u5927\u90e8\u5206\u90fd\u662f\u8bed\u6cd5\u6570\u636e\u3002RNNLM\u548cNNLM\u90fd\u5728\u8bed\u4e49\u4e0a(Syntactic Accuracy)\u8868\u73b0\u7684\u6bd4\u8f83\u597d\uff0c\u5728\u8bed\u6cd5\u4e0a(Semantic Accuracy)\u8868\u73b0\u7684\u6bd4\u8f83\u5dee\uff0cCBOW\u4e5f\u662f\uff0c\u800cskip-gram\u5728\u8bed\u6cd5\u548c\u8bed\u4e49\u4e0a\u90fd\u8868\u73b0\u7684\u6bd4\u8f83\u597d\u3002<\/p>\n<p>RNNLM\u5355\u673a\u7528\u4e868\u5468\uff0cNNLM\u8ba1\u7b97\u91cf\u66f4\u5927\u3002<\/p>\n<p>RNN\u76f8\u5bf9\u5728\u8bed\u6cd5\u95ee\u9898\u4e0a\u8f83\u597d\uff0cNNLM\u6548\u679c\u66f4\u597d\uff0cCBOW\u66f4\u597d\u3002<\/p>\n<p>Skip-gram\u66f4\u5e73\u8861\uff0c\u5728\u8bed\u4e49\u95ee\u9898\u4e0a\u6548\u679c\u597d\u3002<\/p>\n<p>LDC corpora 320M words,82K<\/p>\n<p>\u548c\u5176\u4ed6\u4eba\u5f00\u6e90\u7684\u8bcd\u5411\u91cf\u6bd4\u8f83\uff1a<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162408645.png\" alt=\"image-20240119162408645\" \/><\/p>\n<p>Our NNLM\u4f7f\u7528\u4e86\u5c42\u6b21softmax\u8fdb\u884c\u4e86\u52a0\u901f\u3002<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162424461.png\" alt=\"image-20240119162424461\" \/><\/p>\n<p>3 epoch CBOW\uff0836.1\uff09 &lt; 1 epoch skip-gram\uff0849.2\uff09 3 epoch 300dim &lt; 1 epoch 600d<\/p>\n<h3><strong>8.4 \u5927\u89c4\u6a21\u5e76\u884c\u6a21\u578b\u8bad\u7ec3<\/strong><\/h3>\n<p>Large Scale Parallel Training of Model<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162438081.png\" alt=\"image-20240119162438081\" \/><\/p>\n<h3><strong>8.5 \u5fae\u8f6f\u7814\u7a76\u9662\u53e5\u5b50\u5b8c\u6210\u6bd4\u8d5b<\/strong><\/h3>\n<p><img src=\"Word2vec%E5%8E%9F%E7%90%86%E4%B8%8E%E5%AE%9E%E6%88%98%E8%AF%A6%E8%A7%A3md.assets\/image-20240119162448702-1705652707049-116.png\" alt=\"image-20240119162448702\" \/><\/p>\n<p>\u7c7b\u4f3c\u5b8c\u5f62\u586b\u7a7a\uff0c\u4e00\u53e5\u8bdd\u76d6\u4f4f\u4e00\u4e2a\u8bcd\uff0c\u7ed9\u51fa5\u4e2a\u9884\u6d4b\u7ed3\u679c\u3002<\/p>\n<p>skip-gram\u53ea\u7edf\u8ba1\u4e86\u5468\u56f4\u8bcd\uff0c\u6ca1\u6709\u7528\u5230\u8bcd\u4e0e\u8bcd\u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u6240\u4ee5\u53ef\u4ee5\u8054\u5408\u8bed\u8a00\u6a21\u578bRNNLM\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n<h3><strong>8.6 HS\u548cNEG\u6bd4\u8f83<\/strong><\/h3>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/image-20240119162518483.png\" alt=\"image-20240119162518483\" \/><\/p>\n<p>NCE\uff1a\u5176\u4ed6\u8d1f\u91c7\u6837\u65b9\u6cd5\u3002<\/p>\n<p>\u4e0b\u65b9\u8868\u683c\u4e3a\u4f7f\u7528\u4e86\u91cd\u91c7\u6837\uff0c\u53ef\u4ee5\u53d1\u73b0\u901f\u5ea6\u5927\u5927\u63d0\u5347\u4e86\u3002<\/p>\n<p>\u4f8b\u5b50\uff1a\u5b66\u4e60\u5230\u7684\u5173\u7cfb<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-adc5260b136e936d8e16fad92389fd22_720w.webp\" alt=\"img\" \/><\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-7c0a948b80f20a9618547e0cbf10457e_720w.webp\" alt=\"img\" \/><\/p>\n<p>\u4f7f\u7528\u5341\u4e2a\u5173\u7cfb\u8bcd\u5bf9\u505a\u5e73\u5747\u53ef\u4ee5\u5c06\u504f\u5dee\u62b5\u6d88\u6389\u3002<\/p>\n<p>out-of-the-list words\uff1a\u4e0d\u7b26\u5408\u5f53\u524d\u5206\u7c7b\u7684\u8bcd\u3002<\/p>\n<h2><strong>9 \u8ba8\u8bba\u548c\u603b\u7ed3<\/strong><\/h2>\n<p>\u8ba8\u8bba\u8bba\u6587\u4e2d\u5b58\u5728\u7684\u95ee\u9898\uff0c\u603b\u7ed3\u672c\u9636\u6bb5\u6240\u5b66\u5185\u5bb9\u3002<\/p>\n<h3><strong>9.1 \u8ba8\u8bba<\/strong><\/h3>\n<p><strong>\u8d85\u53c2\u6570\u9009\u62e9<\/strong>\uff1a\u8bf7\u95ee\uff0c\u5229\u7528gensim\u505aword2vec\u7684\u65f6\u5019\uff0c\u8bcd\u5411\u91cf\u7684\u7ef4\u5ea6\u548c\u5355\u8bcd\u6570\u76ee\u6709\u6ca1\u6709\u4e00\u4e2a\u6bd4\u8f83\u597d\u7684\u5bf9\u7167\u8303\u56f4\u5462\uff1f<\/p>\n<p>dim\u4e00\u822c\u5728100-500\u4e4b\u95f4\u9009\u62e9\uff0c\u521d\u59cb\u503c\u8bcd\u5178\u5927\u5c0fV\u76841\/4\u6b21\u65b9\uff0c\u4f8b\u5982V=10K\uff0cdim=100\u3002min_count\u4e00\u822c\u57282-10\u4e4b\u95f4\u9009\u62e9\uff0cmin_count\u8868\u793a\u8bcd\u5728\u8bed\u6599\u4e2d\u51fa\u73b0\u7684\u6700\u5c0f\u6b21\u6570\uff0c\u5c0f\u4e8e\u8be5\u6b21\u6570\u7684\u8bcd\u8fdb\u884c\u820d\u5f03\uff0c\u51b3\u5b9a\u4e86\u8bcd\u8868\u5927\u5c0f\u3002<\/p>\n<p>gridsearch<\/p>\n<p><img src=\"https:\/\/lemon-guess.oss-cn-hangzhou.aliyuncs.com\/img\/v2-601f0053c33994dc39774f86210a8d7d_720w.webp\" alt=\"img\" \/><\/p>\n<h3><strong>9.2 \u603b\u7ed3<\/strong><\/h3>\n<p><strong>\u8bba\u6587\u4e3b\u8981\u521b\u65b0\u70b9\uff1a<\/strong> 1\u3001<strong>\u63d0\u51fa\u4e00\u79cd\u65b0\u7684\u7ed3\u6784<\/strong>\uff1a\u8fd9\u4e2a\u7ed3\u6784\u662f\u4f7f\u7528\u8bcd\u9884\u6d4b\u8bcd\uff0c\u800c\u4e0d\u662f\u7528\u524d\u9762\u7684\u4e00\u7cfb\u5217\u8bcd\u6765\u9884\u6d4b\u8bcd\uff0c<strong>\u7b80\u5316\u4e86\u7ed3\u6784<\/strong>\u3002\u4e5f\u63d0\u51fa\u4e86\u5c42\u6b21softmax\u548c\u8d1f\u91c7\u6837\uff0c\u5927\u5927\u51cf\u5c11\u4e86\u8ba1\u7b97\u91cf\uff0c\u4ece\u800c\u53ef\u4ee5\u4f7f\u7528\u66f4\u9ad8\u7684\u7ef4\u5ea6\uff0c\u66f4\u5927\u7684\u6570\u636e\u96c6\u3002 2\u3001<strong>\u5229\u7528\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6<\/strong>\uff1a\u5728\u5927\u6570\u636e\u4e0a\u8bad\u7ec3\uff0c\u4ece\u800c\u8fbe\u5230\u66f4\u597d\u7684\u6548\u679c\u3002 3\u3001<strong>\u63d0\u51fa\u4e86\u65b0\u7684\u8bcd\u76f8\u4f3c\u5ea6\u4efb\u52a1<\/strong>\uff1aAnalogy\u8bcd\u7c7b\u522b\u3002<\/p>\n<p><strong>\u5173\u952e\u70b9\uff1a<\/strong><\/p>\n<ol>\n<li>\u66f4\u7b80\u5355\u7684\u9884\u6d4b\u6a21\u578b \u2014\u2014 word2vec<\/li>\n<li>\u66f4\u5feb\u7684\u5206\u7c7b\u65b9\u6848 \u2014\u2014 HS\u548cNEG<\/li>\n<\/ol>\n<p><strong>\u521b\u65b0\u70b9\uff1a<\/strong><\/p>\n<ol>\n<li>\u4f7f\u7528\u8bcd\u5bf9\u7684\u9884\u6d4b\u6765\u66ff\u6362\u8bed\u8a00\u6a21\u578b\u7684\u9884\u6d4b<\/li>\n<li>\u4f7f\u7528HS\u548cNEG\u964d\u4f4e\u5206\u7c7b\u590d\u6742\u5ea6<\/li>\n<li>\u4f7f\u7528subsampling\u52a0\u901f\u8bad\u7ec3<\/li>\n<li>\u65b0\u7684\u8bcd\u5bf9\u63a8\u7406\u6570\u636e\u96c6\u6765\u5ba2\u89c2\u8bc4\u4f30\u8bcd\u5411\u91cf\u7684\u8d28\u91cf<\/li>\n<\/ol>\n<p><strong>\u542f\u53d1\u70b9\uff1a<\/strong><\/p>\n<p>1\u3001\u5927\u6570\u636e\u96c6\u4e0a\u7684\u7b80\u5355\u6a21\u578b\u5f80\u5f80\u5f3a\u4e8e\u5c0f\u6570\u636e\u96c6\u4e0a\u7684\u590d\u6742\u6a21\u578b simple models trained on huge amounts of data outperform complex systems trained on less data.\uff081 Introduction p1\uff09<\/p>\n<p>2\u3001King\u7684\u8bcd\u5411\u91cf\u51cf\u53bbMan\u7684\u8bcd\u5411\u91cf\u52a0\u4e0aWoman\u7684\u8bcd\u5411\u91cf\u548cQueen\u7684\u8bcd\u5411\u91cf\u6700\u63a5\u8fd1 vector(\u201dKing\u201d) - vector(\u201dMan\u201d) + vector(\u201dWoman\u201d) results in a vector that is closest to the vector representation of the word Queen \uff081.1 Goals of the Paper p3\uff09 \u8bf4\u660e\u4e86word2vec\u53ef\u4ee5\u5f88\u597d\u7684\u5b66\u4e60\u8bcd\u5bf9\u4e4b\u95f4\u7684\u4ee3\u6570\u5173\u7cfb\uff0c\u5728\u795e\u7ecf\u7f51\u7edc\u4e2d\u90fd\u662f\u6570\u4e0e\u6570\u4e4b\u95f4\u7684\u8ba1\u7b97\uff0c\u5982\u679c\u53ef\u4ee5\u5b66\u4e60\u5230\u8bcd\u5411\u91cf\u4e4b\u95f4\u7684\u4ee3\u6570\u5173\u7cfb\uff0c\u5219\u8bcd\u5411\u91cf\u4e5f\u975e\u5e38\u4f7f\u7528\u4e8e\u795e\u7ecf\u7f51\u7edc\u3002<\/p>\n<p>3\u3001\u6211\u4eec\u51b3\u5b9a\u8bbe\u8ba1\u7b80\u5355\u7684\u6a21\u578b\u6765\u8bad\u7ec3\u8bcd\u5411\u91cf\uff0c\u867d\u7136\u7b80\u5355\u7684\u6a21\u578b\u65e0\u6cd5\u50cf\u795e\u7ecf\u7f51\u7edc\u90a3\u4e48\u51c6\u786e\u5730\u8868\u793a\u6570\u636e\uff0c\u4f46\u662f\u53ef\u4ee5\u5728\u66f4\u591a\u5730\u6570\u636e\u4e0a\u66f4\u5feb\u5730\u8bad\u7ec3 we decided to explore simpler models that might not be able to represent the data as precisely as neural networks, but can possibly be trained on much more data efficiently (3 New Log-linear Models p1)<\/p>\n<p>4\u3001\u6211\u4eec\u76f8\u4fe1\u5728\u66f4\u5927\u7684\u6570\u636e\u96c6\u4e0a\u4f7f\u7528\u66f4\u5927\u7684\u8bcd\u5411\u91cf\u7ef4\u5ea6\u80fd\u591f\u8bad\u7ec3\u5f97\u5230\u66f4\u597d\u7684\u8bcd\u5411\u91cf\u3002 We believe that word vectors trained on even larger data sets with larger dimensionality will perform significantly better(5 Examples of the Learned Relationshops p1)<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u89c6\u9891\u94fe\u63a5\uff1ahttps:\/\/ai.deepshare.net\/detail\/p_5ee62f90022ee_zFpnlHXA\/6 \u6587\u7ae0\u6807\u9898   \u2026 &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ngg_post_thumbnail":0},"categories":[225],"tags":[226,228],"_links":{"self":[{"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9119"}],"collection":[{"href":"\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=9119"}],"version-history":[{"count":2,"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9119\/revisions"}],"predecessor-version":[{"id":9121,"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9119\/revisions\/9121"}],"wp:attachment":[{"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=9119"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=9119"},{"taxonomy":"post_tag","embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=9119"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}