{"id":726,"date":"2021-12-25T15:15:48","date_gmt":"2021-12-25T07:15:48","guid":{"rendered":"http:\/\/139.9.1.231\/?p=726"},"modified":"2022-08-31T15:17:38","modified_gmt":"2022-08-31T07:17:38","slug":"bleu","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2021\/12\/25\/bleu\/","title":{"rendered":"BLEU—\u673a\u5668\u7ffb\u8bd1\u8bc4\u4ef7\u65b9\u6cd5"},"content":{"rendered":"\n

Bleu[1]\u662fIBM\u57282002\u63d0\u51fa\u7684\uff0c\u7528\u4e8e\u673a\u5668\u7ffb\u8bd1\u4efb\u52a1\u7684\u8bc4\u4ef7\uff0c\u53d1\u8868\u5728ACL\uff0c\u5f15\u7528\u6b21\u657010000+\uff0c\u539f\u6587\u9898\u76ee\u662f\u201cBLEU: a Method for Automatic Evaluation of Machine Translation\u201d\u3002<\/p>\n\n\n\n

\u5b83\u7684\u603b\u4f53\u601d\u60f3\u5c31\u662f\u51c6\u786e\u7387\uff0c\u5047\u5982\u7ed9\u5b9a\u6807\u51c6\u8bd1\u6587reference\uff0c\u795e\u7ecf\u7f51\u7edc\u751f\u6210\u7684\u53e5\u5b50\u662fcandidate\uff0c\u53e5\u5b50\u957f\u5ea6\u4e3an\uff0ccandidate\u4e2d\u6709m\u4e2a\u5355\u8bcd\u51fa\u73b0\u5728reference\uff0cm\/n\u5c31\u662fbleu\u76841-gram\u7684\u8ba1\u7b97\u516c\u5f0f\u3002<\/p>\n\n\n\n

BLEU\u8fd8\u6709\u8bb8\u591a\u53d8\u79cd\u3002\u6839\u636en-gram\u53ef\u4ee5\u5212\u5206\u6210\u591a\u79cd\u8bc4\u4ef7\u6307\u6807\uff0c\u5e38\u89c1\u7684\u6307\u6807\u6709BLEU-1\u3001BLEU-2\u3001BLEU-3\u3001BLEU-4\u56db\u79cd\uff0c\u5176\u4e2dn-gram\u6307\u7684\u662f\u8fde\u7eed\u7684\u5355\u8bcd\u4e2a\u6570\u4e3an\u3002<\/p>\n\n\n\n

BLEU-1\u8861\u91cf\u7684\u662f\u5355\u8bcd\u7ea7\u522b\u7684\u51c6\u786e\u6027\uff0c\u66f4\u9ad8\u9636\u7684bleu\u53ef\u4ee5\u8861\u91cf\u53e5\u5b50\u7684\u6d41\u7545\u6027\u3002<\/p>\n\n\n\n

\u8ba1\u7b97\u516c\u5f0f\uff1a<\/p>\n\n\n\n

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

\u5206\u5b50\u91ca\u4e49<\/strong><\/p>\n\n\n\n

\u795e\u7ecf\u7f51\u7edc\u751f\u6210\u7684\u53e5\u5b50\u662fcandidate\uff0c\u7ed9\u5b9a\u7684\u6807\u51c6\u8bd1\u6587\u662freference\u3002<\/p>\n\n\n\n

1\uff09 \u7b2c\u4e00\u4e2a\u6c42\u548c\u7b26\u53f7\u7edf\u8ba1\u7684\u662f\u6240\u6709\u7684candidate\uff0c\u56e0\u4e3a\u8ba1\u7b97\u65f6\u53ef\u80fd\u6709\u591a\u4e2a\u53e5\u5b50\uff0c<\/p>\n\n\n\n

2\uff09\u7b2c\u4e8c\u4e2a\u6c42\u548c\u7b26\u53f7\u7edf\u8ba1\u7684\u662f\u4e00\u6761candidate\u4e2d\u6240\u6709\u7684n\u2212gram\uff0c\u800c \"[\u516c\u5f0f]\" \u8868\u793a\u67d0\u4e00\u4e2an\u2212gram\u5728reference\u4e2d\u7684\u4e2a\u6570\u3002<\/p>\n\n\n\n

\u6240\u4ee5\u6574\u4e2a\u5206\u5b50\u5c31\u662f\u5728\u7ed9\u5b9a\u7684candidate\u4e2d\u6709\u591a\u5c11\u4e2an-gram\u8bcd\u8bed\u51fa\u73b0\u5728reference\u4e2d\u3002<\/strong><\/p>\n\n\n\n

\u5206\u6bcd\u91ca\u4e49<\/strong><\/p>\n\n\n\n

\u524d\u4e24\u4e2a\u6c42\u548c\u7b26\u53f7\u548c\u5206\u5b50\u4e2d\u7684\u542b\u4e49\u4e00\u6837\uff0cCount(n-gram’)\u8868\u793an\u2212gram\u2032\u5728candidate\u4e2d\u7684\u4e2a\u6570\uff0c\u7efc\u4e0a\u53ef\u77e5\uff0c\u5206\u6bcd\u662f\u83b7\u5f97\u6240\u6709\u7684candidate\u4e2dn-gram\u7684\u4e2a\u6570\u3002<\/strong><\/p>\n\n\n\n

\u6539\u8fdb\u7248\u7684BLEU\u8ba1\u7b97\u65b9\u6cd5<\/h2>\n\n\n\n

\u6539\u8fdb\u601d\u8def\uff1a\u5bf9\u4e8e\u67d0\u4e2a\u8bcd\u7ec4\u51fa\u73b0\u7684\u6b21\u6570\uff0c\u5728\u4fdd\u8bc1\u4e0d\u5927\u4e8ecandidate\u4e2d\u51fa\u73b0\u7684\u4e2a\u6570\u7684\u60c5\u51b5\u4e0b\uff0c\u7136\u540e\u518dreference\u5bfb\u627e\u8bcd\u7ec4\u51fa\u73b0\u7684\u6700\u591a\u6b21\u3002<\/p>\n\n\n\n

\u8bc4\u4ef7\u673a\u5668\u7ffb\u8bd1\u7ed3\u679c\u901a\u5e38\u4f7f\u7528BLEU\uff08Bilingual Evaluation Understudy\uff09\u3002\u5bf9\u4e8e\u6a21\u578b\u9884\u6d4b\u5e8f\u5217\u4e2d\u4efb\u610f\u7684\u5b50\u5e8f\u5217\uff0cBLEU\u8003\u5bdf\u8fd9\u4e2a\u5b50\u5e8f\u5217\u662f\u5426\u51fa\u73b0\u5728\u6807\u7b7e\u5e8f\u5217\u4e2d\u3002<\/p>\n\n\n\n

\u5177\u4f53\u6765\u8bf4\uff0c\u8bbe\u8bcd\u6570\u4e3an\u7684\u5b50\u5e8f\u5217\u7684\u7cbe\u5ea6\u4e3apn\u200b\u3002\u5b83\u662f\u9884\u6d4b\u5e8f\u5217\u4e0e\u6807\u7b7e\u5e8f\u5217\u5339\u914d\u8bcd\u6570\u4e3an<\/em>\u7684\u5b50\u5e8f\u5217\u7684\u6570\u91cf\u4e0e\u9884\u6d4b\u5e8f\u5217\u4e2d\u8bcd\u6570\u4e3an\u7684\u5b50\u5e8f\u5217\u7684\u6570\u91cf\u4e4b\u6bd4\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5047\u8bbe\u6807\u7b7e\u5e8f\u5217\u4e3aA\u3001B<\/em>\u3001C\u3001D\u3001E\u3001F\uff0c\u9884\u6d4b\u5e8f\u5217\u4e3aA<\/em>\u3001B<\/em>\u3001B<\/em>\u3001C\u3001D\uff0c\u90a3\u4e48p1=4\/5,p2=3\/4,p3=1\/3,p4=0\u3002\u8bbe<\/p>\n\n\n\n

len<\/em>label\u200b\u548clenpred\u5206\u522b\u4e3a\u6807\u7b7e\u5e8f\u5217\u548c\u9884\u6d4b\u5e8f\u5217\u7684\u8bcd\u6570\uff0c\u90a3\u4e48\uff0cBLEU\u7684\u5b9a\u4e49\u4e3a<\/p>\n\n\n\n

$$
\\exp \\left(\\min \\left(0,1-\\frac{l e n_{\\text {label }}}{l e n_{\\text {pred }}}\\right)\\right) \\prod_{n=1}^{k} p_{n}^{1 \/ 2^{n}}
$$
\u5176\u4e2d \\(k\\) \u662f\u6211\u4eec\u5e0c\u671b\u5339\u914d\u7684\u5b50\u5e8f\u5217\u7684\u6700\u5927\u8bcd\u6570\u3002\u53ef\u4ee5\u770b\u5230\u5f53\u9884\u6d4b \u5e8f\u5217\u548c\u6807\u7b7e\u5e8f\u5217\u5b8c\u5168\u4e00\u81f4\u65f6\uff0cBLEU\u4e3a 1 \u3002
\u56e0\u4e3a\u5339\u914d\u8f83\u957f\u5b50\u5e8f\u5217\u6bd4\u5339\u914d\u8f83\u77ed\u5b50\u5e8f\u5217\u66f4\u96be\uff0cBLEU\u5bf9\u5339\u914d\u8f83\u957f \u5b50\u5e8f\u5217\u7684\u7cbe\u5ea6\u8d4b\u4e88\u4e86\u66f4\u5927\u6743\u91cd\u3002\u4f8b\u5982\uff0c\u5f53 \\( p_{n} \\) \u56fa\u5b9a\u5728 \\( 0.5 \\) \u65f6\uff0c\u968f\u5728 \\( n \\) \u7684\u589e\u5927\uff0c \\( 0.5^{1 \/ 2} \\approx 0.7,0.5^{1 \/ 4} \\approx 0.84,0.5^{1 \/ 8} \\approx$ $0.92,0.5^{1 \/ 16} \\approx 0.96 \\) \u3002\u53e6\u5916\uff0c\u6a21\u578b\u9884\u6d4b\u8f83\u77ed\u5e8f\u5217\u5f80\u5f80\u4f1a\u5f97\u5230 \u8f83\u9ad8 \\( p_{n} \\) \u503c\u3002\u56e0\u6b64\uff0c\u4e0a\u5f0f\u4e2d\u8fde\u4e58\u9879\u524d\u9762\u7684\u7cfb\u6570\u662f\u4e3a\u4e86\u60e9\u7f5a\u8f83\u77ed\u7684 \u8f93\u51fa\u800c\u8bbe\u7684\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5f53 \\(k=2 \\) \u65f6\uff0c\u5047\u8bbe\u6807\u7b7e\u5e8f\u5217\u4e3a A \u3001 B \u3001 C \u3001 D \u3001 E \u3001 F \uff0c\u800c\u9884\u6d4b\u5e8f\u5217\u4e3a A \u3001 B \u3002\u867d\u7136 \\( p_{1}=p_{2}=1 \\) \uff0c\u4f46\u60e9\u7f5a\u7cfb\u6570 \\( \\exp (1-6 \/ 2) \\approx 0.14\\) \uff0c \u56e0\u6b64BLEU\u4e5f\u63a5\u8fd1 0.14 \u3002<\/p>\n\n\n\n

python\u5b9e\u73b0\uff1a<\/p>\n\n\n\n

def bleu(pred_tokens, label_tokens, k):\r\n    len_pred, len_label = len(pred_tokens), len(label_tokens)\r\n    score = math.exp(min(0, 1 - len_label \/ len_pred))\r\n    for n in range(1, k + 1):\r\n        num_matches, label_subs = 0, collections.defaultdict(int)\r\n        for i in range(len_label - n + 1):\r\n            label_subs[''.join(label_tokens[i: i + n])] += 1\r\n        for i in range(len_pred - n + 1):\r\n            if label_subs[''.join(pred_tokens[i: i + n])] > 0:\r\n                num_matches += 1\r\n                label_subs[''.join(pred_tokens[i: i + n])] -= 1\r\n        score *= math.pow(num_matches \/ (len_pred - n + 1), math.pow(0.5, n))\r\n    return score<\/code><\/pre>\n\n\n\n
\u63a5\u4e0b\u6765\uff0c\u5b9a\u4e49\u4e00\u4e2a\u8f85\u52a9\u6253\u5370\u51fd\u6570\uff1a<\/p>\n\n\n\n
def score(input_seq, label_seq, k):\n    pred_tokens = translate(encoder, decoder, input_seq, max_seq_len)\n    label_tokens = label_seq.split(' ')\n    print('bleu %.3f, predict: %s' % (bleu(pred_tokens, label_tokens, k),\n                                      ' '.join(pred_tokens)))<\/code><\/pre>\n\n\n\n
\u4e00\u4e9b\u6ce8\u610f\u4e8b\u9879<\/h2>\n\n\n\n
\u4e00\u822c\u7ed9\u51fa\u7684reference\u662f4\u53e5\u8bdd\uff0c\u4e4b\u6240\u4ee5\u7ed9\u51fa\u591a\u4e2a\u53e5\u5b50\uff0c\u662f\u56e0\u4e3a\u5355\u4e2a\u53e5\u5b50\u53ef\u80fd\u65e0\u6cd5\u548c\u751f\u6210\u7684\u53e5\u5b50\u505a\u5f88\u597d\u5730\u5339\u914d\u3002<\/p>\n\n\n\n
\u6bd4\u5982”\u4f60\u597d”\uff0creference\u662f”hello”\uff0c\u673a\u5668\u7ed9\u51fa\u7684\u8bd1\u6587\u662f”how are you”\uff0c\u673a\u5668\u7ed9\u51fa\u7684\u8bcd\u8bed\u6bcf\u4e00\u4e2a\u4e00\u4e2a\u8bcd\u5339\u914d\u4e0a\u8fd9\u4e2areference\uff0c\u90a3\u4e48BLEU\u503c\u662f0\uff0c\u8fd9\u663e\u7136\u662f\u6709\u95ee\u9898\u7684\u3002\u6240\u4ee5reference\u8d8a\u591a\u6837\u5316\uff0c\u5339\u914d\u6210\u529f\u6982\u7387\u8d8a\u9ad8\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
Bleu[1]\u662fIBM\u57282002\u63d0\u51fa\u7684\uff0c\u7528\u4e8e\u673a\u5668\u7ffb\u8bd1\u4efb\u52a1\u7684\u8bc4\u4ef7\uff0c\u53d1\u8868\u5728ACL\uff0c\u5f15\u7528\u6b21\u657010000+\uff0c\u539f\u6587\u9898\u76ee\u662f … \u7ee7\u7eed\u9605\u8bfbBLEU—\u673a\u5668\u7ffb\u8bd1\u8bc4\u4ef7\u65b9\u6cd5<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[21,4,12],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/726"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=726"}],"version-history":[{"count":13,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/726\/revisions"}],"predecessor-version":[{"id":739,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/726\/revisions\/739"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=726"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=726"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=726"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}