Datasets:

WINGNUS
/

ACL-OCL

	{
	"paper_id": "O11-1001",
	"header": {
	"generated_with": "S2ORC 1.0.0",
	"date_generated": "2023-01-19T08:05:22.382638Z"
	},
	"title": "Empirical Comparisons of Various Discriminative Language Models for Speech Recognition",
	"authors": [
	{
	"first": "",
	"middle": [],
	"last": "\u8cf4\u654f\u8ed2",
	"suffix": "",
	"affiliation": {},
	"email": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "\u9ec3\u90a6\u70dc",
	"suffix": "",
	"affiliation": {},
	"email": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "\u570b\u7acb\u81fa\u7063\u5e2b\u7bc4\u5927\u5b78\u8cc7\u8a0a\u5de5\u7a0b\u5b78\u7cfb",
	"suffix": "",
	"affiliation": {},
	"email": ""
	}
	],
	"year": "",
	"venue": null,
	"identifiers": {},
	"abstract": "",
	"pdf_parse": {
	"paper_id": "O11-1001",
	"_pdf_hash": "",
	"abstract": [],
	"body_text": [
	{
	"text": "\u4e8c\u3001\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u4ecb\u7d39 (\u4e00)\u3001\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u4e4b\u5b9a\u7fa9 \u4e00\u822c\u4f86\u8aaa\uff0c\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u662f\u4ee5\u6700\u5c0f\u5316\u8fa8\u8b58\u932f\u8aa4\u7387\u70ba\u8a13\u7df4\u76ee\u6a19\uff0c\u5e0c\u671b\u5c0d\u57fa\u790e\u8a9e\u97f3\u8fa8\u8b58\u5668 (Baseline Speech Recognizer)\u6240\u7522\u751f\u7684\u5019\u9078\u8a5e\u5e8f\u5217(\u5982\u524d M \u689d\u6700\u4f73\u8fa8\u8b58\u7d50\u679c)\u4f5c\u91cd\u65b0\u6392 \u5e8f\uff0c\u4f7f\u5f97\u5177\u6709\u8f03\u4f4e\u8fa8\u8b58\u932f\u8aa4\u7387\u7684\u5019\u9078\u8a5e\u5e8f\u5217\u80fd\u64c1\u6709\u8f03\u9ad8\u7684\u6392\u5e8f\u3002\u800c\u91cd\u65b0\u6392\u5e8f\u7684\u4f9d\u64da\u5247\u662f \u4ee5\u57fa\u790e\u8a9e\u97f3\u8fa8\u8b58\u5668\u7684\u8fa8\u8b58\u5206\u6578\u505a\u70ba\u57fa\u790e\uff0c\u4e26\u52a0\u4e0a\u984d\u5916\u5b9a\u7fa9\u7684\u8a9e\u8a00\u7279\u5fb5\u5411\u91cf\uff0c\u85c9\u7531\u524d\u8ff0\u5169 \u8005\u8207\u5176\u5c0d\u61c9\u7684\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578\u5411\u91cf\u505a\u5167\u7a4d\u5f8c\u7684\u8a9e\u8a00\u6a21\u578b\u5206\u6578\u4f86\u9032\u884c\u6392\u5e8f\uff0c\u4f7f\u5f97\u524d M \u689d\u6700 \u4f73\u8fa8\u8b58\u5019\u9078\u8a5e\u5e8f\u5217\u4e2d\u6700\u4f4e\u932f\u8aa4\u7387\u7684\u8a5e\u5e8f\u5217\u80fd\u64c1\u6709\u6700\u9ad8\u7684\u8a9e\u8a00\u6a21\u578b\u5206\u6578\u3002\u4ee5\u4e0b\u5c07\u5c0d\u9451\u5225\u5f0f \u8a13\u7df4\u6240\u9700\u7684\u53c3\u6578\u505a\u5b9a\u7fa9\uff1a (a) \u7d66\u5b9a\u4e00\u53e5\u8a9e\u97f3\u8a0a\u865f i x \uff0c\u5176\u7d93\u7531\u57fa\u790e\u8fa8\u8b58\u5668\u6240\u7522\u751f\u7684 M \u689d\u6700\u4f73\u5019\u9078\u8a5e\u5e8f\u5217\u96c6\u5408\u70ba ( ) { } j i i W x GEN , = \uff0c\u5176\u4e2d j \u70ba1\u5230 M \u4e4b\u9593\u3002 (b) \u5c07\u8a13\u7df4\u8a9e\u6599\u8996\u70ba { } R i i W x , \u7684\u96c6\u5408\uff0c\u5176\u4e2d i \u7684\u503c\u4ecb\u65bc1\u5230 L \u4e4b\u9593\uff0c L \u70ba\u8a13\u7df4\u8a9e\u6599\u7684 \u7e3d\u53e5\u6578\uff1b R i W \u70ba\u8a9e\u97f3\u8a0a\u865f i x \u5728\u5176\u5c0d\u61c9 M \u689d\u6700\u4f73\u5019\u9078\u8a5e\u5e8f\u5217\u4e2d\u6700\u4f4e\u932f\u8aa4\u7387\u4e4b\u8a5e\u5e8f \u5217\u3002 (c) \u5c0d\u65bc\u6bcf\u4e00\u689d\u5019\u9078\u8a5e\u5e8f\u5217\u5b9a\u7fa9\u4e00\u7d44 1 + D \u7dad\u7684\u7279\u5fb5\u5411\u91cf ( ) j i d W f , \uff0c\u5176\u4e2d d \u662f\u5f9e 0 \u5230 D \u4e4b\u9593\uff1b ( ) j i W f , 0 \u70ba\u57fa\u790e\u8fa8\u8b58\u5668\u6240\u7522\u751f\u7684\u5206\u6578\uff0c\u5373\u70ba\u8072\u5b78\u6a21\u578b\u8207 N \u9023\u8a9e\u8a00\u6a21\u578b \u7684\u5c0d\u6578\u6a5f\u7387(Log Probability)\u5206\u6578\u7e3d\u548c\uff0c\u5728\u6b64\u6211\u5011\u4f7f\u7528\u4e09\u9023(Trigram)\u8a9e\u8a00\u6a21\u578b\uff1b \u800c\u5176\u5b83\u7dad\u5ea6 d \uff0c\u53ef\u5206\u5225\u8868\u793a\u6bcf\u4e00\u689d\u5019\u9078\u8a5e\u5e8f\u5217 j i W , \u4e2d\u5404\u7a2e N \u9023\u8a5e\u51fa\u73fe\u7684\u6b21\u6578(\u8996 \u70ba\u4e00\u7a2e\u8a9e\u8a00\u7279\u5fb5)\uff0c\u4ee5 ( ) j i d W f , \u4f86\u8868\u793a\uff0c\u672c\u8ad6\u6587\u6240\u5b9a\u7fa9\u5404\u7a2e\u53ef\u80fd\u7684\u8a9e\u8a00\u7279\u5fb5\u70ba\u55ae \u9023\u8a5e(Word Unigram)\u8207\u96d9\u9023\u8a5e(Word Bigram)\u3002 (d) \u5b9a\u7fa9\u4e00\u7d44 1 + D \u7dad\u7684\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578\u5411\u91cf [ ] D d \u03bb \u03bb \u03bb \u03bb \u03bb , , , , , 1 0 K K = \uff0c\u5176\u4e2d\u6bcf\u4e00\u500b\u7279 \u5fb5\u6b0a\u91cd\u53c3\u6578 d \u03bb \u5206\u5225\u5c0d\u61c9\u65bc\u6bcf\u4e00\u500b\u8a9e\u8a00\u7279\u5fb5 ( ) j i d W f , \u3002 \u56e0\u6b64\uff0c\u5019\u9078\u8a5e\u5e8f\u5217 j i W , \u7684\u91cd\u65b0\u6392\u5e8f\u5206\u6578\u53ef\u8868\u793a\u70ba\uff1a ( ) ( ) ( ) \u2211 = = \u2022 = D d j i d d j i j i W f W f W Score 0 , , , , \u03bb \u03bb \u03bb (1) \u800c\u7d93\u7531\u91cd\u65b0\u6392\u5e8f\u5f8c\u5206\u6578\u6700\u9ad8\u7684\u5019\u9078\u8a5e\u5e8f\u5217 * i W \u5373\u505a\u70ba\u6700\u5f8c\u7684\u8f38\u51fa\u7d50\u679c\uff1a ( ) ( ) \u03bb , max arg , * , j i x GEN W i W Score W i j i \u2208 = (2) \u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u7684\u8a13\u7df4\u5728\u65bc\u6c42\u53d6\u6700\u4f73\u7684\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578\u5411\u91cf \u03bb\uff0c\u671f\u671b\u4f7f\u5f97\u6e2c\u8a66\u8a9e\u53e5\u7684\u524d M \u689d\u6700\u4f73\u8fa8\u8b58\u5019\u9078\u8a5e\u5e8f\u5217\u4e2d\u6700\u4f4e\u932f\u8aa4\u7387\u7684\u8a5e\u5e8f\u5217\u80fd\u5728\u5f0f(2)\u64c1\u6709\u6700\u9ad8\u7684\u5206\u6578\u3002 (\u4e8c)\u3001\u5e38\u898b\u7684\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b \u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u65e9\u671f\u5927\u591a\u90fd\u4f7f\u7528\u5728\u5176\u5b83\u7684\u61c9\u7528\u9818\u57df\u4e0a\uff1a\u4f8b\u5982\uff0c\u6a5f\u5668\u7ffb\u8b6f(Machine Translation, MT)\u3001\u81ea\u7136\u8a9e\u8a00\u8655\u7406(Natural Language Processing, NLP)\u7b49\u3002\u8fd1\u5341\u5e74\u4f86\uff0c\u9678\u7e8c \u6709\u8a31\u591a\u5b78\u8005\u5c07\u5404\u7a2e\u57fa\u65bc\u4e0d\u540c\u8a13\u7df4\u6e96\u5247\u7684\u9451\u5225\u5f0f\u8a9e\u97f3\u6a21\u578b\u4ecb\u7d39\u5230\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58 (Large Vocabulary Continuous Speech Recognition, LVCSR)\u4f86\u4f7f\u7528\u3002\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u7684\u8a13 \u7df4\u53ef\u5206\u6210\u4e09\u500b\u9762\u5411\u4f86\u63a2\u8a0e\uff0c\u5206\u5225\u70ba\u8a13\u7df4\u8a9e\u6599\u3001\u8a13\u7df4\u6e96\u5247\u8207\u7279\u5fb5\u3002\u4ee5\u4e0b\u5c07\u4ecb\u7d39\u5e38\u898b\u7684\u5404\u7a2e \u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\uff0c\u4e26\u5c07\u5b83\u5011\u4f9d\u5176\u8a13\u7df4\u6e96\u5247\u5340\u5206\u70ba\u4ee5\u4e0b\u56db\u985e\uff1a\u6700\u5c0f\u5316\u5e73\u65b9\u8aa4\u5dee\u3001\u6700\u5c0f\u5316\u932f \u8aa4\u7387\u671f\u671b\u503c\u3001\u6700\u5927\u5316\u5c0d\u6578\u689d\u4ef6\u6a5f\u7387\u3001\u4ee5\u53ca\u8003\u91cf\u8a9e\u53e5\u4e4b\u9593\u5f7c\u6b64\u4e4b\u95dc\u4fc2\u3002 1\u3001\u6700\u5c0f\u5316\u5e73\u65b9\u8aa4\u5dee \u611f\u77e5\u5668\u6f14\u7b97\u6cd5(Perceptron)[7]\u65e9\u671f\u662f\u88ab\u61c9\u7528\u5728\u4eba\u5de5\u985e\u795e\u7d93\u7db2\u8def(Artificial Neural Network) \u9818\u57df\u4e2d\uff1b\u5728 2002 \u5e74\uff0c\u7f8e\u570b\u5b78\u8005 Collins[8]\u5c07\u611f\u77e5\u5668\u6f14\u7b97\u6cd5\u61c9\u7528\u5728\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u9818\u57df\u4e2d\u3002 \u611f \u77e5 \u5668 \u6f14 \u7b97 \u6cd5 \u53ef \u8996 \u70ba \u662f \u6700 \u5927 \u5316 \u71b5 \u503c \u6cd5 (Maximum-Entropy, ME) \u6216 \u689d \u4ef6 \u5f0f \u96a8 \u6a5f \u57df (Conditional",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "( ) \u03bb Perc F \u4ee5\u6c42\u5f97\u6700\u4f73\u7684\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578\u5411\u91cf \u03bb\u02c6\uff1a ( ) ( ) ( ) ( ) \u2211 = \u2212 = L i i R i W Score W Score F 1 2 * Perc , , 2 1 \u03bb \u03bb \u03bb (3) \u70ba\u4e86\u6c42\u5f97 \u03bb\u02c6\uff0c\u6211\u5011\u53ef\u4ee5\u5229\u7528\u68af\u5ea6\u4e0b\u964d\u6cd5(Gradient Descent Method)\u5c07 ( ) \u03bb Perc F \u7684\u6bcf\u4e00\u500b\u7dad \u5ea6 \u7279 \u5fb5 \u6b0a \u91cd \u53c3 \u6578 d \u03bb \u5206 \u5225 \u505a \u504f \u5fae \u5206 \uff0c \u7531 \u65bc ( ) \u03bb Perc F \u53ef \u80fd \u5b58 \u5728 \u8a31 \u591a \u5c40 \u90e8 \u6700 \u4f73 \u89e3 (Local Minimum Solutions)\uff0c\u800c\u4f7f\u7528\u68af\u5ea6\u4e0b\u964d\u6cd5\u4e26\u7121\u6cd5\u4fdd\u8b49\u53ef\u6c42\u5f97\u5168\u57df\u6700\u4f73\u89e3(Global Minimum Solutions)\u3002\u56e0\u6b64\uff0c\u611f\u77e5\u5668\u6f14\u7b97\u6cd5\u63a1\u53d6\u96a8\u6a5f\u8fd1\u4f3c\u6cd5(Stochastic Approximation)\uff0c\u5373\u5c0d\u6bcf\u4e00 \u53e5\u8a13\u7df4\u8a9e\u53e5\u7684\u6bcf\u4e00\u7dad\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578\u5206\u5225\u6c42\u6700\u4f73\u89e3\uff0c\u6c42\u5f97\u6bcf\u4e00\u7dad\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578\u7684\u8abf\u6574\u91cf\uff1a ( ) ( ) ( ) ( ) ( ) ( ) * * , , i d R i d i R i d d W f W f W Score W Score \u2212 \u22c5 \u2212 \u22c5 \u2212 = \u03bb \u03bb \u03b7 \u03bb \u03bb (4) \u5176\u4e2d\u03b7 \u70ba\u5b78\u7fd2\u6b65\u8abf\u5e38\u6578(Learning Step Size)\u3002\u9664\u4e86\u5f0f(4)\u6b64\u7a2e\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578\u66f4\u65b0\u5f0f\u4e4b\u5916\uff0c \u4e5f \u6709 \u5b78 \u8005 \u63d0 \u51fa \u7701 \u7565 ( ) ( ) \u03bb \u03bb , , * i R i W Score W Score \u2212 \u9805 \uff0c \u5c07 \u66f4 \u65b0 \u5f0f \u7c21 \u5316 \u70ba ( ) ( ) ( ) * i d R i d d d W f W f \u2212 \u22c5 + = \u03b7 \u03bb \u03bb \u4f86\u66f4\u65b0\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578\u3002\u5176\u6f14\u7b97\u6cd5\u5982\u5716\u4e00\u6240\u793a\u3002 2\u3001\u6700\u5c0f\u5316\u932f\u8aa4\u7387\u671f\u671b\u503c (1)\u3001\u6700\u5c0f\u5316\u932f\u8aa4\u7387\u8a13\u7df4(MERT) \u6700\u5c0f\u5316\u932f\u8aa4\u7387\u8a13\u7df4(Minimum Error Rate Training, MERT)\u662f\u5728 2003 \u5e74\u7531\u5b78\u8005 Och[13]\u63d0 \u51fa\uff0c\u4e26\u4e14\u904b\u7528\u5728\u6a5f\u5668\u7ffb\u8b6f(Machine Translation)\u9818\u57df\u4e2d\uff1b\u5728 2008 \u5e74\u7531 Kobayashi \u7b49\u5b78\u8005[14] \u5c07\u6700\u5c0f\u5316\u932f\u8aa4\u7387\u8a13\u7df4\u65b9\u6cd5\u4ecb\u7d39\u5230\u8a9e\u97f3\u8fa8\u8b58\u9818\u57df\u4e2d\u4f7f\u7528\u3002\u61c9\u7528\u65bc\u8a9e\u97f3\u8fa8\u8b58\u6642\uff0c\u5176\u8a13\u7df4\u6e96\u5247 \u5b9a\u7fa9\u6210\u6700\u5c0f\u5316\u57fa\u790e\u8a9e\u97f3\u8fa8\u8b58\u5668\u6240\u7522\u751f\u7684 M \u689d\u5019\u9078\u8a5e\u5e8f\u5217\u4e4b\u932f\u8aa4\u7387\u671f\u671b\u503c\uff0c\u85c9\u6b64\u627e\u51fa\u4e00 \u500b\u6700\u5408\u9069\u7684\u8a9e\u8a00\u6a21\u578b\u7279\u5fb5\u6b0a\u91cd\u5411\u91cf\uff1a ( ) ( ) ( ) ( ) ( ) ( ) ( ) \u2211\u2211 \u2211 = = = \u2212 \u2212 \u22c5 = L i M k M j R i j i R i k i W W Score W Score W Score W Score F k i 1 1 1 , , MERT , exp , , exp , \u03b2 \u03b2 \u03bb \u03bb \u03bb \u03c9 \u03bb (5) \u5176\u4e2d k i W , \u03c9 \u70ba\u5019\u9078\u8a5e\u5e8f\u5217 k i W , \u7684\u932f\u8aa4\u7387(Error Rate)\uff1b\u800c \u03b2 \u70ba\u4e00\u5e73\u6ed1\u5316\u53c3\u6578\u3002\u900f\u904e\u9032\u4e00\u6b65\u7684 \u6578\u5b78\u63a8\u5c0e\uff0c\u6211\u5011\u53ef\u4ee5\u5c07\u5f0f(5)\u4e2d ( ) ( ) \u03b2 \u03bb , exp R i W Score \u9805\u63d0\u51fa\u800c\u7c21\u5316\u6210\uff1a ( ) ( ) ( ) ( ) ( ) \u2211\u2211 \u2211 = = = \u22c5 = L i M k M j j i k i W W Score W Score F k i 1 1 1 , , MERT , exp , exp , \u03b2 \u03b2 \u03bb \u03bb \u03c9 \u03bb (6) \u518d\u5c07\u5f0f(6)\u91dd\u5c0d\u6bcf\u4e00\u7dad\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578 d \u03bb \u505a\u504f\u5fae\u5206\uff0c\u5f97\u5176\u8abf\u6574\u91cf\u70ba\uff1a ( ) ( ) ( ) ( ) ( ) step",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "* i 0 \u03b7 \u03b7 \u03bb \u03bb i d R i d d d i i R i d W f W f \u03bb D d x GEN W L i W T T t D d",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) 2 1 , 1 , , , , 1 1 , exp , exp , exp \u02c6, \uf8f7 \uf8f7 \uf8f8 \uf8f6 \uf8ec \uf8ec \uf8ed \uf8eb \u2212 \u22c5 \u22c5 \u22c5 \u22c5 + = \u2211 \u2211 \u2211\u2211 = \u2032 \u2032 = = = M j j i M j j i d k i d j i k i L i M k W d d W Score W f W f W Score W Score k i \u03b2 \u03b2 \u03b2 \u03bb \u03bb \u03bb \u03b2 \u03c9 \u03b7 \u03bb \u03bb (7) \u5176\u4e2d\u03b7 \u70ba\u5b78\u7fd2\u6b65\u8abf\u5e38\u6578\u3002\u6211\u5011\u53ef\u4ee5\u5c07\u6700\u5c0f\u5316\u932f\u8aa4\u7387\u8a13\u7df4\u4e2d\u932f\u8aa4\u7387 k i W , \u03c9 \u8996\u70ba\u4e00\u7a2e\u6a23\u672c\u6b0a\u91cd (Sample Weight)\u8cc7\u8a0a\uff0c\u7528\u4f86\u5340\u5225\u6bcf\u4e00\u500b\u5019\u9078\u8a5e\u5e8f\u5217 k i W , \u5c0d\u65bc\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u6642\u7684\u91cd \u8981\u6027\u3002 3\u3001\u6700\u5927\u5316\u5c0d\u6578\u689d\u4ef6\u6a5f\u7387 (1)\u3001\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b(GCLM) \u65e9\u671f\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b(Global Conditional Log-linear Model, GCLM)\u88ab\u61c9\u7528\u5728\u81ea \u7136\u8a9e\u8a00\u8655\u7406\u9818\u57df\u4e2d\uff1b2007 \u5e74 Roark \u7b49\u5b78\u8005[4]\u4ee5\u6709\u9650\u72c0\u614b\u6a5f(Weighted Finite State Automata, WFSA)\u5be6\u4f5c\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b\u65bc\u8a9e\u97f3\u8fa8\u8b58\u7d50\u679c\u7684\u91cd\u65b0\u6392\u5e8f\u4e0a\uff0c\u4e26\u4e14\u8207\u611f\u77e5\u5668\u6f14 \u7b97\u6cd5\u9032\u884c\u6bd4\u8f03\u3002 \u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b\u662f\u5e0c\u671b\u5728\u7d66\u5b9a\u4e00\u53e5\u8a9e\u97f3\u8a0a\u865f i x \u8207\u6240\u5c0d\u61c9\u7684 M \u689d\u6700\u4f73\u5019\u9078\u8a5e\u5e8f \u5217 ( ) i x GEN \u6642\uff0c\u5176\u4e2d\u64c1\u6709\u6700\u4f4e\u8fa8\u8b58\u932f\u8aa4\u7387\u7684\u8a5e\u5e8f\u5217\u5176\u5c0d\u6578\u689d\u4ef6\u6a5f\u7387\u53ef\u4ee5\u8d8a\u5927\u8d8a\u597d\uff0c\u4ea6\u5373 \u6700\u5927\u5316\u4e0b\u5217\u8a13\u7df4\u76ee\u6a19\u51fd\u6578\uff1a ( ) ( ) ( ) ( ) ( ) \u2211 \u2211 = = = L i M j j i R i W Score W Score F 1 1 , GCLM , exp , exp log \u03bb \u03bb \u03bb (8) \u70ba\u4e86\u907f\u514d\u904e\u5ea6\u8a13\u7df4(Overtraining)\uff0c\u6211\u5011\u53ef\u4ee5\u5728\u76ee\u6a19\u51fd\u6578 ( ) \u03bb GCLM F \u4e2d\u52a0\u4e0a\u4e00\u500b\u6b0a\u91cd\u53c3\u6578\u7684 \u96f6\u5747\u503c\u9ad8\u65af\u4e8b\u524d\u6a5f\u7387(Zero-Mean Gaussian Prior Probability)\u9805\uff1a ( ) ( ) ( ) ( ) ( ) \u2211 \u2211 = = \u2212 = L i M j j i R i W Score W Score F 1 2 2 1 , GCLM 2 , exp , exp log \u03c3 \u03bb \u03bb \u03bb \u03bb (9) \u56e0\u70ba ( ) \u03bb GCLM F \u70ba\u4e00\u51f8\u51fd\u6578(Convex Function)\uff0c\u56e0\u6b64\u53ef\u4ee5\u6c42\u5f97\u5168\u57df\u6700\u4f73\u89e3(Globally Optimal Solution)\uff0c\u70ba\u6c42\u5f97\u6700\u4f73\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578\u5411\u91cf \u03bb\u02c6\u3002\u5c07\u5f0f(7)\u91dd\u5c0d\u6bcf\u4e00\u7dad\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578 d \u03bb \u505a\u504f\u5fae \u5206\uff0c\u5f97\u5176\u8abf\u6574\u91cf\u70ba\uff1a ( ) ( ) ( ) ( ) ( ) ( ) \u2211 \u2211 \u2211 = = = \u2212 \uf8fa \uf8fa \uf8fa \uf8fa \uf8fb \uf8f9 \uf8ef \uf8ef \uf8ef \uf8ef \uf8f0 \uf8ee \u22c5 \u2212 \u22c5 + = L i d k i d M k M j j i k i R i d d W f W Score W Score W f 1 2 , 1 1 , , , exp , exp \u03c3 \u03bb \u03bb \u03bb \u03b7 \u03bb \u03bb (10) (2)\u3001\u6b0a\u91cd\u5f0f\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b(WGCLM) \u4e0d\u540c\u65bc\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b(GCLM)\uff0cOba \u7b49\u5b78\u8005[15]\u5728 2010 \u5e74\u63d0\u51fa\u5c07\u6a23\u672c\u6b0a\u91cd\u52a0 \u5165\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b\u9032\u884c\u6539\u826f\uff0c\u70ba\u6bcf\u4e00\u500b\u5019\u9078\u8a5e\u5e8f\u5217\u7684\u5206\u6578\u52a0\u4e0a\u4e00\u500b\u4e0d\u540c\u7684\u6b0a \u91cd\uff0c\u7528\u4f86\u8868\u793a\u6bcf\u4e00\u689d\u5019\u9078\u8a5e\u5e8f\u5217\u4e0d\u540c\u7684\u91cd\u8981\u7a0b\u5ea6\uff0c\u6b64\u65b9\u6cd5\u7a31\u70ba\u6b0a\u91cd\u5f0f\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda \u6027\u6a21\u578b(Weighted Global Conditional Log-linear Model, WGCLM)\u3002\u63db\u53e5\u8a71\u8aaa\uff0c\u6bcf\u4e00\u500b\u5019\u9078 \u8a5e\u5e8f\u5217 j i W , \u90fd\u6703\u6709\u4e00\u500b\u76f8\u5c0d\u61c9\u7684\u6a23\u672c\u6b0a\u91cd j i W , \u03c9 \uff1b\u6839\u64da\u4e0d\u540c\u7684\u6a23\u672c\u6b0a\u91cd\u4f86\u8868\u793a\u6bcf\u4e00\u500b\u5019\u9078 \u8a5e\u5e8f\u5217\u5c0d\u65bc\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u7684\u4e0d\u540c\u91cd\u8981\u6027\u3002\u5176\u8a13\u7df4\u76ee\u6a19\u51fd\u6578\u53ef\u8868\u793a\u70ba\uff1a ( ) ( ) ( ) ( ) ( ) \u2211 \u2211 = = = L i M j j i W R i W Score W Score F j i 1 1 , WGCLM , exp , exp log , \u03bb \u03c9 \u03bb \u03bb (11) \u540c\u6a23\u5730\uff0c\u70ba\u4e86\u907f\u514d\u5728\u8abf\u6574\u53c3\u6578\u7684\u904e\u7a0b\u4e2d\uff0c\u767c\u751f\u904e\u5ea6\u8a13\u7df4\u7684\u554f\u984c\uff0c\u6211\u5011\u4e5f\u53ef\u4ee5\u52a0\u5165\u4e00\u500b\u96f6 \u5747\u503c\u9ad8\u65af\u4e8b\u524d\u6a5f\u7387\u9805\u65bc\u6b0a\u91cd\u5f0f\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b\u7684\u8a13\u7df4\u76ee\u6a19\u51fd\u6578\u4e2d\uff1a ( ) ( ) ( ) ( ) ( ) \u2211 \u2211 = = \u2212 = L i M j j i W R i W Score W Score F j i 1 2 2 1 , WGCLM 2 , exp , exp log , \u03c3 \u03bb \u03bb \u03c9 \u03bb \u03bb (12) \u5c07\u5f0f(12)\u91dd\u5c0d\u6bcf\u4e00\u7dad\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578 d \u03bb \u505a\u504f\u5fae\u5206\uff0c\u5f97\u5176\u8abf\u6574\u91cf\u70ba\uff1a ( ) ( ) ( ) ( ) ( ) ( ) \u2211 \u2211 \u2211 = = = \u2212 \uf8fa \uf8fa \uf8fa \uf8fa \uf8fb \uf8f9 \uf8ef \uf8ef \uf8ef \uf8ef \uf8f0 \uf8ee \u22c5 \u2212 \u22c5 + = L i d k i d M k M j j i W k i W R i d d W f W Score W Score W f j i k i 1 2 , 1 1 , , , exp , exp , , \u03c3 \u03bb \u03bb \u03c9 \u03bb \u03c9 \u03b7 \u03bb \u03bb (13) \u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u6a23\u672c\u6b0a\u91cd\u7684 j i W , \u03c9 \u8a2d\u8a08\u4e5f\u662f\u4e00\u500b\u503c\u5f97\u7814\u7a76\u7684\u8b70\u984c\uff0c\u901a\u5e38\u6211\u5011\u53ef\u4ee5\u5c07\u6bcf\u4e00\u500b \u5019\u9078\u8a5e\u5e8f\u5217\u672c\u8eab\u7684\u932f\u8aa4\u7387\u7576\u6210\u5176\u6a23\u672c\u6b0a\u91cd\u3002 4\u3001\u8003\u91cf\u8a9e\u53e5\u4e4b\u9593\u5f7c\u6b64\u4e4b\u95dc\u4fc2 (1)\u3001\u8f2a\u8f49\u96d9\u91cd\u9451\u5225\u5f0f\u6a21\u578b (R2D2) \u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b(GCLM)\u662f\u671f\u671b\u6700\u4f4e\u932f\u8aa4\u7387\u8a5e\u5e8f\u5217\u7684\u5c0d\u6578\u689d\u4ef6\u6a5f\u7387\u80fd\u5920\u8d8a\u5927\u8d8a \u597d\uff1bOba \u7b49\u5b78\u8005\u7b49\u91dd\u5c0d\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b\u63d0\u51fa\u6539\u826f\u65b9\u6cd5\uff0c\u5728\u8a13\u7df4\u76ee\u6a19\u51fd\u6578\u4e2d\u8003\u616e \u4e86\u8a13\u7df4\u8a9e\u53e5\u6240\u6709\u5019\u9078\u8a5e\u5e8f\u5217\u5f7c\u6b64\u4e4b\u9593\u7684\u95dc\u4fc2\uff0c\u56e0\u800c\u6709\u6240\u8b02\u7684\u8f2a\u8f49\u96d9\u91cd\u9451\u5225\u5f0f\u6a21\u578b (Round-Robin Dual Discrimination Model, R2D2)[16]\u3002\u8f2a\u8f49\u96d9\u91cd\u9451\u5225\u5f0f\u6a21\u578b\u53ef\u4ee5\u8996\u70ba\u662f\u5168 \u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b(GCLM)\u7684\u4e00\u7a2e\u5ef6\u4f38\uff1b\u5b83\u56e0\u70ba\u8003\u91cf\u4e86\u5169\u5169\u5019\u9078\u8a5e\u5e8f\u5217\u5f7c\u6b64\u4e4b\u9593\u7684 \u95dc\u4fc2\uff0c\u4f7f\u5f97\u8f2a\u5176\u64c1\u6709\u8f03\u597d\u7684\u4e00\u822c\u5316\u80fd\u529b\u3002\u540c\u6642\uff0c\u985e\u4f3c\u65bc\u6b0a\u91cd\u5f0f\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b (WGCLM)\uff0c\u8f2a\u8f49\u96d9\u91cd\u9451\u5225\u5f0f\u6a21\u578b\u4e5f\u4f7f\u7528\u4e86\u6a23\u672c\u6b0a\u91cd\uff1a ( ) ( ) ( ) ( ) ( ) ( ) ( ) \u2211 \u2211\u2211 = = \u2032 = \u2032 \uf8f4 \uf8fe \uf8f4 \uf8fd \uf8fc \uf8f4 \uf8f3 \uf8f4 \uf8f2 \uf8f1 = \u2032 L i M j M j j i W j i W W Score W Score F j i j i 1 1 1 , 2 , 1 R2D2 , exp exp , exp exp log , , \u03bb \u03c9 \u03c3 \u03bb \u03c9 \u03c3 \u03bb (14) \u5176\u4e2d\uff0c 1 \u03c3 \u8207 2 \u03c3 \u70ba\u5be6\u9a57\u53c3\u6578\u3002\u76f8\u540c\u7684\uff0c\u5c07\u5f0f(14)\u91dd\u5c0d\u6bcf\u4e00\u7dad\u7279\u5fb5\u6b0a\u91cd\u53c3\u6578 d \u03bb \u505a\u504f\u5fae\u5206\uff0c \u5f97\u5176\u8abf\u6574\u91cf\u70ba\uff1a ( ) ( ) ( ) ( ) ( ) ' , , 2 1 , , 1 1 1 1 1 , 1 1 , exp where , \u02c6j i j i W W j i j i L i M j M j M j M j j i d M j M j j i d d d W Score W Score A A A W f A W f \u03c9 \u03c3 \u03c9 \u03c3 \u03b7 \u03bb \u03bb \u2212 + \u2212 = \uf8f4 \uf8f4 \uf8fe \uf8f4 \uf8f4 \uf8fd \uf8fc \uf8f4 \uf8f4 \uf8f3 \uf8f4 \uf8f4 \uf8f2 \uf8f1 \uf8fa \uf8fb \uf8f9 \uf8ef \uf8f0 \uf8ee \u2212 \uf8fa \uf8fb \uf8f9 \uf8ef \uf8f0 \uf8ee \u22c5 + = \u2032 = = \u2032 = = \u2032 = \u2032 = \u2032 = \u2211",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "R i W \u8207\u5176\u5b83\u53ef\u80fd\u5019\u9078\u8a5e\u5e8f\u5217 j i W , \u4e4b\u6392\u5e8f\u5206 \u6578\u7684\u5dee\u7570\uff0c\u8b93\u53c3\u8003\u5019\u9078\u8a5e\u5e8f\u5217 R i W \u7684\u5206\u6578\u8f03\u5176\u5b83\u53ef\u80fd\u5019\u9078\u8a5e\u5e8f\u5217 j i W , \u6108\u5927\u6108\u597d\uff1b\u901a\u5e38\u6211 \u5011\u5c07\u6b64\u5206\u6578\u7684\u5dee\u7570\u7a31\u70ba\"\u5206\u96e2\u908a\u969b(Separation Margin)\"\uff1a ( ) ( ) ( ) j i W W R i i W Score W Score x R i j i , , max \u2260 \u2212 = \u03c4 (16) \u5176\u4e2d ( ) R i W Score \u70ba\u53c3\u8003\u8a5e\u5e8f\u5217\u7684\u91cd\u65b0\u6392\u5e8f\u5206\u6578\uff1b ( ) j i W Score , \u70ba\u67d0\u4e00\u500b\u5019\u9078\u8a5e\u5e8f\u5217\u7684\u91cd\u65b0 \u6392\u5e8f\u5206\u6578\u3002\u7531\u5f0f(16)\u53ef\u77e5\uff0c\u82e5 ( ) 0 > i x \u03c4 \uff0c\u8868\u793a\u4f7f\u7528\u76ee\u524d\u7684\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u5c0d\u65bc\u8a9e\u53e5 i x \u6240\u5c0d \u61c9\u7684 M \u689d\u5019\u9078\u8a5e\u5e8f\u5217\u9032\u884c\u6392\u5e8f\u6642\uff0c\u53ef\u4ee5\u8ce6\u4e88\u8a5e\u5e8f\u5217 R i W \u6700\u9ad8\u7684\u6392\u5e8f\u5206\u6578\uff0c\u6211\u5011\u53ef\u4ee5\u8996\u70ba \u6c92\u6709\u8fa8\u8b58\u932f\u8aa4\u767c\u751f(\u70ba\u7406\u60f3\u72c0\u6cc1)\uff1b\u53cd\u4e4b\uff0c\u82e5 ( ) 0 < i x \u03c4 \uff0c\u5247\u8868\u793a\u6b63\u78ba(\u6216\u662f\u932f\u8aa4\u7387\u6700\u4f4e)\u7684\u5019 \u9078\u8a5e\u5e8f\u5217\u4e4b\u6392\u5e8f\u5206\u6578\u4e0d\u662f\u6240\u6709\u5019\u9078\u8a5e\u5e8f\u5217\u4e2d\u6700\u9ad8\u7684\uff0c\u56e0\u6b64\u7d93\u91cd\u65b0\u6392\u5e8f\u7684\u8a9e\u97f3\u8fa8\u8b58\u8f38\u51fa\u5c07 \u4e0d\u662f\u6700\u4f73\u7684\u7d50\u679c\u3002\u5728\u6700\u5927\u908a\u969b\u4f30\u6e2c\u6cd5\u7684\u8a13\u7df4\u904e\u7a0b\u4e2d\uff0c\u6211\u5011\u9996\u5148\u70ba\u8a13\u7df4\u8a9e\u6599{ } L x x x , , , 2 1 K \u5b9a\u7fa9\u4e00\u7d44\u652f\u63f4\u96c6(Support Set)\uff1a ( ) { } \u03b5 \u03c4 \u2264 \u2264 = i i x x S 0 \| LME (17) \u03b5 \u662f\u4e00\u500b\u6b63\u5be6\u6578\uff0c\u53ef\u4ee5\u7528\u4f86\u63a7\u5236\u652f\u63f4\u96c6\u4e2d\u6240\u5305\u542b\u7684\u8a13\u7df4\u8a9e\u6599\u500b\u6578\uff0c\u6700\u5927\u5316\u908a\u969b\u4f30\u6e2c\u6cd5\u7684 \u76ee\u6a19\u51fd\u6578\u5c31\u53ef\u5b9a\u7fa9\u70ba\u6700\u5927\u908a\u969b\u4f30\u6e2c\u6cd5\u7684\u8a13\u7df4\u76ee\u6a19\u662f\u5e0c\u671b\u6700\u5927\u5316\u652f\u63f4\u96c6\u4e2d\u7684\u6700\u5c0f\u5206\u96e2\u908a \u969b[19]\uff1a ( ) ( ) i S x x F i \u03c4 \u03bb LME min LME \u2208 =",
	"eq_num": "(18"
	}
	],
	"section": "",
	"sec_num": null
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "0 \| < = \u2032 \u2032 i i x x \u03c4 \u03d5 (19) \u7d50\u5408\u652f\u63f4\u96c6\u8207\u932f\u8aa4\u96c6\uff0c\u67d4\u6027\u6700\u5927\u908a\u969b\u4f30\u6e2c\u6cd5\u70ba\u6700\u5927\u5316\u4e0b\u5217\u76ee\u6a19\u51fd\u6578[22, 23]\uff1a ( ) ( ) ( ) \u2211 \u2208 \u2208 \u2212 \u22c5 \u2212 = \u03d5 \u03b4 \u03d5 \u03c3 \u03c4 \u03bb ' LME ' LME S 1 min i i x i i S x x x F",
	"eq_num": "(20)"
	}
	],
	"section": "",
	"sec_num": null
	},
	{
	"text": "\u4e5f\u5c31\u662f\u9664\u4e86\u652f\u63f4\u96c6\u6240\u63d0\u4f9b\u7684\u9451\u5225\u6027\u8cc7\u8a0a\u5916\uff0c\u9084\u52a0\u5165\u4e86\u5e73\u5747\u932f\u8aa4\u4f30\u6e2c\u65bc\u6a21\u578b\u7684\u76ee\u6a19\u51fd\u6578 \u4e2d\u3002\u5728\u5f0f(20)\u4e2d\uff0c\u03c3 \u662f\u4e00\u500b\u6b63\u5be6\u6578\uff0c\u7528\u4f86\u63a7\u5236\u5e73\u5747\u932f\u8aa4\u4f30\u6e2c\u5c0d\u65bc\u8a13\u7df4\u9451\u5225\u5f0f\u6a21\u578b\u6642\u7684\u5f71 ",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "\u97ff\u6027\uff1b ( ) \u22c5 \u03b4 \u662f\u932f\u8aa4\u51fd\u6578\uff0c\u901a\u5e38\u88ab\u5b9a\u7fa9\u70ba[22]\uff1a ( ) ( ) ( ) ( ) \u03bb \u03bb \u03b4 , , max , , R i j i W W i W Score W Score x R i j i \u2212 = \u2260 (21) \u5373\u5206\u96e2\u908a\u969b\u7684\u8ca0\u6578\u3002 \u6700\u5927\u908a\u969b\u4f30\u6e2c\u6cd5\u5247\u53ea\u8003\u616e\u4e86\"\u8207\u5206\u96e2\u908a\u969b\u8f03\u8fd1\"(\u53c3\u7167\u5f0f(17))\u4e14\u91cd\u65b0\u6392\u5e8f\u53ef\u4ee5\u6b63\u78ba\u5730\u9078\u51fa \u53c3\u8003\u5019\u9078\u8a5e\u5e8f\u5217\u4e4b\u8a13\u7df4\u8a9e\u53e5\uff0c\u5982\u6b64\u4e0d\u50c5\u5ffd\u7565\u4e86\u5206\u96e2\u908a\u969b\u9644\u8fd1\u7684\u5176\u5b83\u8cc7\u8a0a\uff0c\u4ea6\u6703\u5c0e\u81f4\u8a13\u7df4 \u8a9e\u53e5\u6578\u91cf\u4e0d\u8db3\uff0c\u6700\u7d42\u4f7f\u5f97\u8a13\u7df4\u51fa\u4f86\u7684\u9451\u5225\u5f0f\u6a21\u578b\u4e00\u822c\u5316\u80fd\u529b\u4e0d\u8db3\uff1b\u6709\u5225\u65bc\u6700\u5927\u908a\u969b\u4f30\u6e2c \u6cd5\uff0c\u67d4\u6027\u908a\u969b\u4f30\u6e2c\u6cd5(Soft Margin Estimation, SME)\u5247\u662f\u85c9\u7531\u8003\u616e\u689d\u4ef6\u7684\u653e\u5bec\uff0c\u5c07\u90a3\u4e9b\u8fa8 \u8b58\u932f\u8aa4(\u4ea6\u5373\u53c3\u8003\u5019\u9078\u8a5e\u5e8f\u5217\u4e4b\u91cd\u65b0\u6392\u5e8f\u5206\u6578\u4e0d\u662f\u6700\u9ad8)\u5728\u4e00\u5b9a\u7bc4\u570d\u5167\u7684\u8a13\u7df4\u8a9e\u53e5\u4e5f\u4e00 \u4f75\u5217\u5165\u8003\u91cf\uff0c\u4f86\u5f4c\u88dc\u8a13\u7df4\u8a9e\u53e5\u4e0a\u7684\u4e0d\u8db3\u3002 \u67d4\u6027\u908a\u969b\u4f30\u6e2c\u6cd5\u5247\u7684\u8a13\u7df4\u76ee\u7684\u5982\u540c\u6700\u5927\u908a\u969b\u4f30\u6e2c\u6cd5\u5247\u4e00\u6a23\uff0c\u5e0c\u671b\u6700\u5927\u5316\u8a13\u7df4\u8a9e\u6599\u4e2d\u7684\u6700 \u5c0f\u5206\u96e2\u908a\u969b\uff0c\u5dee\u5225\u662f\u67d4\u6027\u908a\u969b\u4f30\u6e2c\u6cd5\u5247\u5728\u5b9a\u7fa9\u652f\u63f4\u96c6\u6642\u7684\u689d\u4ef6\u6bd4\u8f03\u5f48\u6027\uff0c\u52a0\u5165\u4e86\u4e00\u500b\u9b06 \u5f1b\u8b8a\u91cf(Slack Variable)\u03be \uff1a ( ) { } \u03b5 \u03c4 \u03be \u2264 \u2264 \u2212 = i i x x S \| SME (22) \u5176\u4e2d\u03be \u70ba\u4e00\u500b\u5927\u65bc\u96f6\u7684\u5be6\u6578\uff0c\u5176\u8868\u793a\u90a3\u4e9b\u8fa8\u8b58\u932f\u8aa4\u7684\u8a13\u7df4\u8a9e\u53e5\u82e5\u5176\u5206\u96e2\u908a\u969b\u5927\u65bc \u03be \u2212 \u4e5f \u6703\u5728\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u6642\u5217\u5165\u8003\u91cf\u3002\u67d4\u6027\u908a\u969b\u4f30\u6e2c\u6cd5\u70ba\u6700\u5927\u5316\u4e0b\u5217\u76ee\u6a19\u51fd\u6578\uff1a ( ) ( ) i S x x F i \u03c4 \u03bb SME min SME \u2208 =",
	"eq_num": "("
	}
	],
	"section": "",
	"sec_num": null
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "\u82e5 ( ) 0 , > j i W \u03c4 \uff0c\u8868\u793a\u4f7f\u7528\u76ee\u524d\u7684\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u53ef\u4ee5\u8ce6\u4e88\u8a5e\u5e8f\u5217 R i W \u6709\u8f03 j i W , \u9ad8\u7684\u6392\u5e8f\u5206 \u6578\uff0c\u53cd\u4e4b\uff0c\u82e5 ( ) 0 , < j i W \u03c4 \uff0c\u5247\u8868\u793a\u53c3\u8003(\u8fa8\u8b58\u932f\u8aa4\u7387\u6700\u4f4e)\u5019\u9078\u8a5e\u5e8f\u5217\u4e4b\u6392\u5e8f\u5206\u6578\u8f03\u8a5e\u5e8f\u5217 j i W , \u4f4e\uff0c\u56e0\u6b64\u8fa8\u8b58\u5668\u7684\u8f38\u51fa\u5c07\u4e0d\u6703\u662f\u6700\u4f73\u7d50\u679c\u3002\u63a5\u8457\uff0c\u6211\u5011\u5b9a\u7fa9\u4e00\u7d44\u652f\u63f4\u96c6(Support Set)\uff1a ( ) { } i j i j i W W S \u03b3 \u03c4 \u2264 = , MDLM , MDLM \| (25) \u5176\u4e2d\uff0c i \u03b3 \u662f\u6bcf\u4e00\u500b\u8a13\u7df4\u8a9e\u53e5\u7684\u5224\u5225\u91cf\uff1b\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u5c07\u5b83\u5b9a\u7fa9\u70ba\uff1a \uf8f7 \uf8f8 \uf8f6 \uf8ec \uf8ed \uf8eb \uf8f7 \uf8f8 \uf8f6 \uf8ec \uf8ed \uf8eb \u2212 = R i j i W W j i \u03c9 \u03c9 \u03b1 \u03b3 , max exp (26) \u03b1 \u662f\u4e00\u500b\u5be6\u9a57\u5e38\u6578\uff1b R i W \u03c9 \u70ba R i W \u7684\u8fa8\u8b58\u932f\u8aa4\u7387\uff1b j , i W \u03c9 \u70ba\u5019\u9078\u8a5e\u5e8f\u5217 j i W , \u7684\u8fa8\u8b58\u932f\u8aa4\u7387\uff1b\u6240 \u4ee5 i \u03b3 \u662f\u96a8\u8457\u8a13\u7df4\u8a9e\u53e5\u7684\u4e0d\u540c\u800c\u6709\u6240\u8b8a\u52d5\uff0c\u7576\u53c3\u8003\u5019\u9078\u8a5e\u5e8f\u5217 R i W \u7684\u932f\u8aa4\u7387\u9060\u5c0f\u65bc\u932f\u8aa4\u7387 \u6700\u9ad8\u7684\u5019\u9078\u8a5e\u5e8f\u5217 j i W , \u6642\uff0c i \u03b3 \u7684\u503c\u61c9\u6108\u5927\u3002\u81f3\u6b64\uff0c\u4e0d\u540c\u65bc\u904e\u53bb\u8003\u616e\u908a\u969b\u6982\u5ff5\u7684\u9451\u5225\u5f0f\u8a9e \u8a00\u6a21\u578b\uff0c\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u7684\u57fa\u65bc\u908a\u969b\u4e4b\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u8003\u616e\u6bcf\u4e00\u8a13\u7df4\u8a9e\u53e5\u5176\u53c3\u8003\u5019\u9078\u8a5e\u5e8f \u5217\u8207\u5176\u5b83\u6240\u6709\u5019\u9078\u8a5e\u5e8f\u5217\u7684\u95dc\u4fc2\u9032\u884c\u8cc7\u6599\u9078\u53d6\u3002\u4e26\u4e14\u5728\u9078\u53d6\u7684\u904e\u7a0b\u4e2d\uff0c\u8003\u616e\u4e86\u8a13\u7df4\u8a9e\u53e5 \u5404\u81ea\u7684\u8fa8\u8b58\u932f\u8aa4\u7387\u3002\u6700\u5f8c\uff0c\u7d50\u5408\u5f0f(24)\u3001(25)\u8207(26)\uff0c\u6211\u5011\u5c07\u57fa\u65bc\u908a\u969b\u4e4b\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b \u7684\u76ee\u6a19\u51fd\u6578\u5b9a\u7fa9\u70ba\uff1a ( ) ( ) ( ) ( ) \u2211 \u2211 = \u2208 \u2208 = L i S W x GEN W i,j j i i j i W F 1 & 2 MDLM MDLM MDLM , , 2 1 \u03c4 \u03bb (27) \u5229\u7528\u68af\u5ea6\u4e0b\u964d\u6cd5\u5c07\u6b64\u76ee\u6a19\u51fd\u6578\u5c0d\u6bcf\u4e00\u7dad\u6b0a\u91cd\u53c3\u6578 d \u03bb \u505a\u504f\u5fae\u5206\u53ef\u6c42\u5f97\u5176\u8abf\u6574\u91cf\uff0c\u6bcf\u4e00\u7dad \u7279\u5fb5\u6b0a\u91cd\u5411\u91cf\u7684\u66f4\u65b0\u5f0f\u70ba\uff1a (",
	"eq_num": ") ( ) ( ) ( ) ("
	}
	],
	"section": "",
	"sec_num": null
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "= = N j j i d j d k k j j k i R j i j R i i W f v v v v v W Score W Score if N- n n k j N j v L i W x T ...T t 1 , d , ,",
	"eq_num": "8 1"
	}
	],
	"section": "",
	"sec_num": null
	},
	{
	"text": "\u03b7 \u03bb \u03bb \u03c4 \u5716\u4e8c\u3001\u57fa\u65bc\u908a\u969b\u4e4b\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u6f14\u7b97\u6cd5 ( ) ( ) ( ) ( ) ( ) [ ] ( ) \u2211 \u2208 \u2208 \u2212 \u22c5 \u2212 \u22c5 \u2212 = MDLM , , & , , MDLM S W x GEN W j i d R i d i j i d d j i i j i W f W f W \u03b3 \u03c4 \u03b7 \u03bb \u03bb (28) \u4e8b\u5be6\u4e0a\uff0c\u57fa\u65bc\u908a\u969b\u8cc7\u8a0a\u4e4b\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b(MDLM)\u76ee\u6a19\u51fd\u6578\u8207\u611f\u77e5\u5668\u6f14\u7b97\u6cd5\u6709\u4e9b\u76f8\u4f3c\uff0c \u7686\u662f\u8003\u616e\u6700\u5c0f\u5e73\u65b9\u8aa4\u5dee\u3002\u7136\u800c\uff0c\u611f\u77e5\u5668\u6f14\u7b97\u6cd5\u662f\u671f\u671b\u6392\u5e8f\u5206\u6578\u6700\u9ad8\u7684\u5019\u9078\u8a5e\u5e8f\u5217\u8207\u53c3\u8003 \u8a5e\u5e8f\u5217\u4e4b\u5206\u6578\u5dee\u7570\u8d8a\u5c0f\u8d8a\u597d\uff1b\u800c\u57fa\u65bc\u908a\u969b\u4e4b\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u4e0d\u50c5\u8003\u616e\u6392\u5e8f\u5206\u6578\u6700\u9ad8\u7684\u5019 \u9078\u8a5e\u5e8f\u5217\uff0c\u66f4\u4ee5\u5206\u96e2\u908a\u969b\u70ba\u57fa\u790e\uff0c\u8003\u616e\u4e86\u66f4\u591a\u53c3\u8003\u8a5e\u5e8f\u5217\u8207\u5176\u5b83\u5019\u9078\u8a5e\u5e8f\u5217\u4e4b\u9593\u7684\u95dc \u4fc2\uff0c\u56e0\u6b64\u4e0d\u6703\u50cf\u611f\u77e5\u5668\u6f14\u7b97\u6cd5\u6703\u6709\u904e\u5ea6\u8a13\u7df4\u7684\u554f\u984c\u3002\u518d\u8005\uff0c\u7531\u65bc\u6211\u5011\u5c07\u908a\u969b\u7684\u8cc7\u6599\u9078\u53d6 \u6982\u5ff5\u7a0d\u4f5c\u6539\u826f\uff0c\u4f7f\u5f97\u53c3\u8207\u8a13\u7df4\u7684\u8cc7\u6599\u8b8a\u591a\uff0c\u56e0\u6b64\u4e0d\u50cf\u5148\u524d\u7c21\u4ecb\u7684\u5176\u5b83\u5404\u5f0f\u904b\u7528\u908a\u969b\u8cc7\u8a0a \u7684\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\uff0c\u5bb9\u6613\u906d\u9047\u8a13\u7df4\u8cc7\u6599\u592a\u5c11\u7684\u554f\u984c\u3002 \u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u57fa\u65bc\u908a\u969b\u4e4b\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u4e5f\u53ef\u4ee5\u5982\u540c\u611f\u77e5\u5668\u6f14\u7b97\u6cd5\u4e00\u822c\uff0c\u5c07\u5f0f\u5b50(28) \u4e2d\u7684 ( ) i j i W \u03b3 \u03c4 \u2212 , MDLM \u7701\u7565\uff0c\u5c07\u66f4\u65b0\u5f0f\u5b50\u7c21\u5316\u6210\uff1a ( ) ( ) ( ) ( ) \u2211 \u2208 \u2208 \u2212 \u22c5 + = MDLM , , & , S W x GEN W j i d R i d d d j i i j i W f W",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "= M )\u7684\u8fa8\u8b58\u7d50\u679c\uff0c\u505a\u70ba\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u7684\u8a13\u7df4\u8207 \u6e2c\u8a66\u8a9e\u6599\u3002 (\u4e8c)\u3001\u5404\u5f0f\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u7684\u5be6\u9a57\u7d50\u679c \u9996\u5148\uff0c\u6211\u5011\u6bd4\u8f03\u5404\u7a2e\u4e0d\u540c\u7684\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u61c9\u7528\u65bc\u8a9e\u97f3\u8fa8\u8b58\u7d50\u679c\u4e4b\u91cd\u65b0\u6392\u5e8f\uff0c\u5404\u7a2e\u65b9\u6cd5 \u7684\u8fa8\u8b58\u5b57\u932f\u8aa4\u7387\u5982\u8868\u4e09\u6240\u793a\u3002\u6211\u5011\u53ef\u4ee5\u7531\u8868\u4e09\u89c0\u5bdf\u5230\uff0c\u5982\u540c\u5148\u524d\u63d0\u5230\u7684\uff0c\u611f\u77e5\u5668\u6f14\u7b97\u6cd5 (Perceptron)\u5728\u8a13\u7df4\u8a9e\u6599\u4e0a\u7684\u8868\u73fe\u7684\u78ba\u512a\u65bc\u5176\u5b83\u5404\u7a2e\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\uff0c\u4f46\u5728\u6e2c\u8a66\u8a9e\u6599\u7684\u8868 \u73fe\u5247\u7121\u6cd5\u6709\u76f8\u540c\u7684\u6548\u679c\u3002\u800c\u5728\u6e2c\u8a66\u8a9e\u6599\u7684\u5be6\u9a57\u7d50\u679c\uff0c\u4ee5\u8f2a\u8f49\u96d9\u91cd\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b(R2D2) \u7684\u6548\u679c\u6700\u70ba\u986f\u8457\uff0c\u9019\u4e5f\u8aaa\u660e\u4e86\u5728\u8a13\u7df4\u7684\u904e\u7a0b\u4e2d\uff0c\u8003\u91cf\u8f03\u591a\u6709\u95dc\u5019\u9078\u8a5e\u5e8f\u5217\u5f7c\u6b64\u4e4b\u9593\u95dc\u4fc2 \u7684\u8cc7\u8a0a\u5c0d\u6a21\u578b\u7684\u8a13\u7df4\u662f\u6709\u6b63\u9762\u5e6b\u52a9\u7684\uff0c\u6703\u4f7f\u5f97\u6a21\u578b\u6709\u8f03\u597d\u7684\u4e00\u822c\u5316\u80fd\u529b\u3002 (\u4e09)\u3001\u57fa\u65bc\u908a\u969b\u8cc7\u8a0a\u4e4b\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u76f8\u95dc\u5be6\u9a57\u7d50\u679c \u672c\u8ad6\u6587\u6240\u63d0\u51fa\u4e4b\u57fa\u65bc\u908a\u969b\u8cc7\u8a0a\u4e4b\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b(MDLM)\u65bc\u8a9e\u97f3\u8fa8\u8b58\u4e4b\u5be6\u9a57\u7d50\u679c\u5982\u8868 \u56db\u6240\u793a\u3002\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u6bd4\u8f03\u4e86\u56db\u7a2e\u4e0d\u540c\u652f\u63f4\u96c6\u7684\u5b9a\u7fa9\u65b9\u5f0f\uff1a \u52d5\u614b\u578b(MDLM-D)\uff1a ( ) { } i j i j i W W S \u03b3 \u03c4 \u2264 = , MDLM , D - MDLM \| \u6b63\u78ba\u5206\u985e\u52d5\u614b\u578b(MDLM-CD)\uff1a ( ) { } i j i j i W W S \u03b3 \u03c4 \u2264 \u2264 = , MDLM , CD - MDLM 0 \| \u56fa\u5b9a\u578b(MDLM-F)\uff1a ( ) { } \u03c1 \u03c4 \u2264 = j i j i W W S , MDLM , F - MDLM \| \uff0c\u5176\u4e2d \u03c1 \u662f\u4e00\u500b\u6b63\u5be6\u6578 \u6b63\u78ba\u5206\u985e\u56fa\u5b9a\u578b(MDLM-CF)\uff1a ( ) { } \u03c1 \u03c4 \u2264 \u2264 = j i j i W W",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	}
	],
	"back_matter": [],
	"bib_entries": {
	"BIBREF0": {
	"ref_id": "b0",
	"title": "Statistical Methods for Speech Recognition, the",
	"authors": [
	{
	"first": "F",
	"middle": [],
	"last": "Jelinek",
	"suffix": ""
	}
	],
	"year": 1999,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "F. Jelinek, Statistical Methods for Speech Recognition, the MIT press, 1999.",
	"links": null
	},
	"BIBREF1": {
	"ref_id": "b1",
	"title": "Discriminative reranking for natural language parsing",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Collins",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Koo",
	"suffix": ""
	}
	],
	"year": 2005,
	"venue": "Computational Linguistics",
	"volume": "31",
	"issue": "1",
	"pages": "25--70",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "M. Collins, and T. Koo, \"Discriminative reranking for natural language parsing,\" Computational Linguistics,\" Vol. 31, No. 1, pp. 25-70, 2005.",
	"links": null
	},
	"BIBREF2": {
	"ref_id": "b2",
	"title": "An empirical study on language model adaptation",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Gao",
	"suffix": ""
	},
	{
	"first": "H",
	"middle": [],
	"last": "Suzuki",
	"suffix": ""
	},
	{
	"first": "W",
	"middle": [],
	"last": "Yuan",
	"suffix": ""
	}
	],
	"year": 2006,
	"venue": "ACM Transactions on Asian Language Information Processing",
	"volume": "5",
	"issue": "3",
	"pages": "209--227",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "J. Gao, H. Suzuki, and W. Yuan, \"An empirical study on language model adaptation,\" ACM Transactions on Asian Language Information Processing, Vol. 5, No. 3, pp. 209-227, 2006",
	"links": null
	},
	"BIBREF3": {
	"ref_id": "b3",
	"title": "Discriminative n-gram language modeling",
	"authors": [
	{
	"first": "B",
	"middle": [],
	"last": "Roark",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Saraclar",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Collins",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Johnson",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "Computer Speech and Language",
	"volume": "21",
	"issue": "2",
	"pages": "373--392",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "B. Roark, M. Saraclar, M. Collins and M. Johnson,\" Discriminative n-gram language modeling,\" Computer Speech and Language, Vol. 21, No. 2, pp. 373-392, 2007.",
	"links": null
	},
	"BIBREF4": {
	"ref_id": "b4",
	"title": "Exploiting discriminative language models for reranking speech recognition hypotheses",
	"authors": [
	{
	"first": "J.-W",
	"middle": [],
	"last": "Liu",
	"suffix": ""
	},
	{
	"first": "S.-H",
	"middle": [],
	"last": "Lin",
	"suffix": ""
	},
	{
	"first": "B",
	"middle": [],
	"last": "Chen",
	"suffix": ""
	}
	],
	"year": 2010,
	"venue": "Proceedings of ROCLING XXII: Conference on Computational Linguistics and Speech Processing",
	"volume": "",
	"issue": "",
	"pages": "30--49",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "J.-W. Liu, S.-H. Lin and B. Chen, \"Exploiting discriminative language models for reranking speech recognition hypotheses,\" in Proceedings of ROCLING XXII: Conference on Computational Linguistics and Speech Processing (ROCLING 2010), pp. 30-49, 2010.",
	"links": null
	},
	"BIBREF5": {
	"ref_id": "b5",
	"title": "Discriminative language modeling for speech recognition with relevance information",
	"authors": [
	{
	"first": "B",
	"middle": [],
	"last": "Chen",
	"suffix": ""
	},
	{
	"first": "C.-W",
	"middle": [],
	"last": "Liu",
	"suffix": ""
	}
	],
	"year": 2011,
	"venue": "Proceedings of the International Conference on Multimedia and Expo",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "B. Chen and C.-W. Liu, \"Discriminative language modeling for speech recognition with relevance information,\" in Proceedings of the International Conference on Multimedia and Expo, 2011.",
	"links": null
	},
	"BIBREF6": {
	"ref_id": "b6",
	"title": "The perceptron: A probabilistic model for information storage and organization in the brain",
	"authors": [
	{
	"first": "F",
	"middle": [],
	"last": "Rosenblatt",
	"suffix": ""
	}
	],
	"year": 1958,
	"venue": "Psychological Review",
	"volume": "",
	"issue": "6",
	"pages": "386--408",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "F. Rosenblatt, \"The perceptron: A probabilistic model for information storage and organization in the brain,\" Psychological Review, No. 6, pp. 386-408, 1958.",
	"links": null
	},
	"BIBREF7": {
	"ref_id": "b7",
	"title": "Discriminative training methods for hidden markov models: theory and experiments with perceptron algorithms",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Collins",
	"suffix": ""
	}
	],
	"year": 2002,
	"venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing",
	"volume": "",
	"issue": "",
	"pages": "1--8",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "M. Collins, \"Discriminative training methods for hidden markov models: theory and experiments with perceptron algorithms,\" in Proceedings of the Conference on Empirical Methods in Natural Language Processing, pp. 1-8, 2002.",
	"links": null
	},
	"BIBREF8": {
	"ref_id": "b8",
	"title": "Conditional random fields: Probabilistic models for segmenting and labeling sequence data",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Lafferty",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Mccallum",
	"suffix": ""
	},
	{
	"first": "F",
	"middle": [],
	"last": "Pereira",
	"suffix": ""
	}
	],
	"year": 2001,
	"venue": "Proceedings of International Conference on Machine Learning",
	"volume": "",
	"issue": "",
	"pages": "282--289",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "J. Lafferty, A. McCallum, and F. Pereira, \"Conditional random fields: Probabilistic models for segmenting and labeling sequence data,\" in Proceedings of International Conference on Machine Learning, pp. 282-289, 2001.",
	"links": null
	},
	"BIBREF9": {
	"ref_id": "b9",
	"title": "Shallow parsing with conditional random fields",
	"authors": [
	{
	"first": "F",
	"middle": [],
	"last": "Sha",
	"suffix": ""
	},
	{
	"first": "F",
	"middle": [],
	"last": "Pereira",
	"suffix": ""
	}
	],
	"year": 2003,
	"venue": "Proceedings of the North American Chapter of the Association for Computational Linguistics on Human Language Technology",
	"volume": "",
	"issue": "",
	"pages": "134--141",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "F. Sha, F. Pereira, \"Shallow parsing with conditional random fields,\" in Proceedings of the North American Chapter of the Association for Computational Linguistics on Human Language Technology, pp. 134-141, 2003.",
	"links": null
	},
	"BIBREF10": {
	"ref_id": "b10",
	"title": "Machine Learning. The McGraw-Hill Companies",
	"authors": [
	{
	"first": "T",
	"middle": [
	"M"
	],
	"last": "Mitchell",
	"suffix": ""
	}
	],
	"year": 1997,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "T. M. Mitchell, Machine Learning. The McGraw-Hill Companies, 1997.",
	"links": null
	},
	"BIBREF11": {
	"ref_id": "b11",
	"title": "A comparative study of discriminative methods for reranking LVCSR n-best hypotheses in domain adaptation and generalization",
	"authors": [
	{
	"first": "Z",
	"middle": [],
	"last": "Zhou",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Gao",
	"suffix": ""
	},
	{
	"first": "F",
	"middle": [
	"K"
	],
	"last": "Soong",
	"suffix": ""
	},
	{
	"first": "H",
	"middle": [],
	"last": "Meng",
	"suffix": ""
	}
	],
	"year": 2006,
	"venue": "Proceedings of International Conference on Acoustics, Speech and Signal Processing",
	"volume": "",
	"issue": "",
	"pages": "141--144",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Z. Zhou, J. Gao, F.K. Soong, and H. Meng, \"A comparative study of discriminative methods for reranking LVCSR n-best hypotheses in domain adaptation and generalization,\" in Proceedings of International Conference on Acoustics, Speech and Signal Processing, pp. 141-144, 2006.",
	"links": null
	},
	"BIBREF12": {
	"ref_id": "b12",
	"title": "Minimum error rate training in statistical machine translation",
	"authors": [
	{
	"first": "F",
	"middle": [
	"J"
	],
	"last": "Och",
	"suffix": ""
	}
	],
	"year": 2003,
	"venue": "Proceedings of the Annual Meeting of the Association for Computational Linguistics on Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "160--167",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "F.J. Och, \"Minimum error rate training in statistical machine translation,\" in Proceedings of the Annual Meeting of the Association for Computational Linguistics on Computational Linguistics, pp. 160-167, 2003.",
	"links": null
	},
	"BIBREF13": {
	"ref_id": "b13",
	"title": "Discriminative rescoring based on minimization of word errors for transcribing broadcast news",
	"authors": [
	{
	"first": "A",
	"middle": [],
	"last": "Kobayashi",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Oku",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [],
	"last": "Homma",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [],
	"last": "Sato",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Imai",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "",
	"suffix": ""
	},
	{
	"first": "Takagi",
	"middle": [],
	"last": "",
	"suffix": ""
	}
	],
	"year": 2008,
	"venue": "Proceedings of the Annual Conference of the International Speech Communication Association",
	"volume": "",
	"issue": "",
	"pages": "1574--1577",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "A. Kobayashi, T. Oku, S. Homma, S. Sato, T. Imai and T, Takagi, \"Discriminative rescoring based on minimization of word errors for transcribing broadcast news,\" in Proceedings of the Annual Conference of the International Speech Communication Association, pp. 1574-1577, 2008.",
	"links": null
	},
	"BIBREF14": {
	"ref_id": "b14",
	"title": "A comparative study on methods of weighted language model training for reranking LVCSR n-best hypotheses",
	"authors": [
	{
	"first": "T",
	"middle": [],
	"last": "Oba",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Hori",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Nakamura",
	"suffix": ""
	}
	],
	"year": 2010,
	"venue": "Proceedings of International Conference on Acoustics, Speech and Signal Processing",
	"volume": "",
	"issue": "",
	"pages": "5126--5129",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "T. Oba, T Hori, and A. Nakamura, \"A comparative study on methods of weighted language model training for reranking LVCSR n-best hypotheses,\" in Proceedings of International Conference on Acoustics, Speech and Signal Processing, pp. 5126-5129, 2010.",
	"links": null
	},
	"BIBREF15": {
	"ref_id": "b15",
	"title": "Round-robin discriminative model for reranking ASR hypotheses",
	"authors": [
	{
	"first": "T",
	"middle": [],
	"last": "Oba",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Hori",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Nakamura",
	"suffix": ""
	}
	],
	"year": 2010,
	"venue": "Proceedings of the Annual Conference of the International Speech Communication Association",
	"volume": "",
	"issue": "",
	"pages": "2446--2449",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "T. Oba, T. Hori and A. Nakamura, \"Round-robin discriminative model for reranking ASR hypotheses,\" in Proceedings of the Annual Conference of the International Speech Communication Association, pp. 2446-2449, 2010.",
	"links": null
	},
	"BIBREF16": {
	"ref_id": "b16",
	"title": "An approach to efficient generation of high-accuracy and compact error-corrective models for speech recognition",
	"authors": [
	{
	"first": "T",
	"middle": [],
	"last": "Oba",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Hori",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Nakamura",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "Proceedings of the Annual Conference of the International Speech Communication Association",
	"volume": "",
	"issue": "",
	"pages": "1753--1756",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "T. Oba, T. Hori and A. Nakamura, \"An approach to efficient generation of high-accuracy and compact error-corrective models for speech recognition,\" in Proceedings of the Annual Conference of the International Speech Communication Association, pp. 1753-1756, 2007.",
	"links": null
	},
	"BIBREF17": {
	"ref_id": "b17",
	"title": "Syntactic and sub-lexical features for Turkish discriminative language models",
	"authors": [
	{
	"first": "E",
	"middle": [],
	"last": "Arisoy",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Saraclar",
	"suffix": ""
	},
	{
	"first": "B",
	"middle": [],
	"last": "Roark",
	"suffix": ""
	},
	{
	"first": "I",
	"middle": [],
	"last": "Shafran",
	"suffix": ""
	}
	],
	"year": 2010,
	"venue": "Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing",
	"volume": "",
	"issue": "",
	"pages": "5538--5541",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "E. Arisoy, M. Saraclar, B. Roark and I. Shafran, \"Syntactic and sub-lexical features for Turkish discriminative language models,\" in Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, pp. 5538-5541, 2010.",
	"links": null
	},
	"BIBREF18": {
	"ref_id": "b18",
	"title": "Large margin hidden markov models for speech recognition",
	"authors": [
	{
	"first": "H",
	"middle": [],
	"last": "Jiang",
	"suffix": ""
	},
	{
	"first": "X",
	"middle": [],
	"last": "Li",
	"suffix": ""
	},
	{
	"first": "C",
	"middle": [],
	"last": "Liu",
	"suffix": ""
	}
	],
	"year": 2006,
	"venue": "IEEE transactions on audio, speech, and language processing",
	"volume": "14",
	"issue": "",
	"pages": "1584--1595",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "H. Jiang, X. Li and C. Liu, \"Large margin hidden markov models for speech recognition,\" IEEE transactions on audio, speech, and language processing, Vol. 14, No. 5, pp. 1584-1595, 2006.",
	"links": null
	},
	"BIBREF19": {
	"ref_id": "b19",
	"title": "A comparative study on margin-based discriminative training of acoustic models",
	"authors": [
	{
	"first": "Y.-T",
	"middle": [],
	"last": "Lo",
	"suffix": ""
	},
	{
	"first": "B",
	"middle": [],
	"last": "Chen",
	"suffix": ""
	}
	],
	"year": 2010,
	"venue": "Proceedings of ROCLING XXII: Conference on Computational Linguistics and Speech Processing",
	"volume": "",
	"issue": "",
	"pages": "65--85",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Y.-T. Lo and B. Chen, \"A comparative study on margin-based discriminative training of acoustic models,\" in Proceedings of ROCLING XXII: Conference on Computational Linguistics and Speech Processing (ROCLING 2010), pp. 65-85, 2010.",
	"links": null
	},
	"BIBREF20": {
	"ref_id": "b20",
	"title": "Soft margin estimation of hidden markov model parameters",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Li",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Yuan",
	"suffix": ""
	},
	{
	"first": "C.-H",
	"middle": [],
	"last": "Lee",
	"suffix": ""
	}
	],
	"year": 2006,
	"venue": "Proceedings of International Conference on Spoken Language Processing",
	"volume": "",
	"issue": "",
	"pages": "2422--2425",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "J. Li, M. Yuan and C.-H. Lee, \"Soft margin estimation of hidden markov model parameters,\" in Proceedings of International Conference on Spoken Language Processing, pp. 2422-2425, 2006.",
	"links": null
	},
	"BIBREF21": {
	"ref_id": "b21",
	"title": "Incorporating training errors for large margin HMMs under semidefinite programming framework",
	"authors": [
	{
	"first": "H",
	"middle": [],
	"last": "Jiang",
	"suffix": ""
	},
	{
	"first": "X",
	"middle": [],
	"last": "Li",
	"suffix": ""
	},
	{
	"first": "C",
	"middle": [],
	"last": "Liu",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "Proceedings of International Conference on Acoustics, Speech and Signal Processing",
	"volume": "",
	"issue": "",
	"pages": "629--632",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "H. Jiang and X. Li and C. Liu, \"Incorporating training errors for large margin HMMs under semidefinite programming framework,\" in Proceedings of International Conference on Acoustics, Speech and Signal Processing, pp. 629-632, 2007.",
	"links": null
	},
	"BIBREF22": {
	"ref_id": "b22",
	"title": "Large margin estimation of n-gram language models for speech recognition via linear programming",
	"authors": [
	{
	"first": "V",
	"middle": [],
	"last": "Magdin",
	"suffix": ""
	},
	{
	"first": "H",
	"middle": [],
	"last": "Jiang",
	"suffix": ""
	}
	],
	"year": 2010,
	"venue": "Proceedings of International Conference on Acoustics, Speech and Signal Processing",
	"volume": "",
	"issue": "",
	"pages": "5398--5401",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "V. Magdin and H. Jiang, \"Large margin estimation of n-gram language models for speech recognition via linear programming,\" in Proceedings of International Conference on Acoustics, Speech and Signal Processing, pp. 5398-5401, 2010.",
	"links": null
	},
	"BIBREF23": {
	"ref_id": "b23",
	"title": "MATBN: A Mandarin Chinese broadcast news corpus",
	"authors": [
	{
	"first": "H.-M",
	"middle": [],
	"last": "Wang",
	"suffix": ""
	},
	{
	"first": "B",
	"middle": [],
	"last": "Chen",
	"suffix": ""
	},
	{
	"first": "J.-W",
	"middle": [],
	"last": "Kuo",
	"suffix": ""
	},
	{
	"first": "S.-S",
	"middle": [],
	"last": "Cheng",
	"suffix": ""
	}
	],
	"year": 2005,
	"venue": "International Journal of Computational Linguistics & Chinese Language Processing",
	"volume": "10",
	"issue": "2",
	"pages": "219--236",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "H.-M. Wang, B. Chen, J.-W. Kuo, and S.-S. Cheng, \"MATBN: A Mandarin Chinese broadcast news corpus,\" International Journal of Computational Linguistics & Chinese Language Processing, Vol. 10, No. 2, pp. 219-236, 2005.",
	"links": null
	},
	"BIBREF24": {
	"ref_id": "b24",
	"title": "SRI Language Modeling Toolkit",
	"authors": [
	{
	"first": "A",
	"middle": [],
	"last": "Stolcke",
	"suffix": ""
	}
	],
	"year": null,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "A. Stolcke, SRI Language Modeling Toolkit, version 1.5.8, http://www.speech.sri.com/projects/srilm/.",
	"links": null
	},
	"BIBREF25": {
	"ref_id": "b25",
	"title": "Lightly supervised and data-driven approaches to mandarin broadcast news transcription",
	"authors": [
	{
	"first": "B",
	"middle": [],
	"last": "Chen",
	"suffix": ""
	},
	{
	"first": "J.-W",
	"middle": [],
	"last": "Kuo",
	"suffix": ""
	},
	{
	"first": "W.-H",
	"middle": [],
	"last": "Tsai",
	"suffix": ""
	}
	],
	"year": 2004,
	"venue": "Proceedings of International Conference on Acoustics, Speech and Signal Processing",
	"volume": "",
	"issue": "",
	"pages": "777--780",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "B. Chen, J.-W. Kuo and W.-H. Tsai, \"Lightly supervised and data-driven approaches to mandarin broadcast news transcription,\" in Proceedings of International Conference on Acoustics, Speech and Signal Processing, pp. 777-780, 2004.",
	"links": null
	},
	"BIBREF26": {
	"ref_id": "b26",
	"title": "Corrective language modeling for large vocabulary ASR with the perceptron algorithm",
	"authors": [
	{
	"first": "B",
	"middle": [],
	"last": "Roark",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Saraclar",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Collins",
	"suffix": ""
	}
	],
	"year": 2004,
	"venue": "Proceedings of International Conference on Acoustics, Speech and Signal Processing",
	"volume": "",
	"issue": "",
	"pages": "749--752",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "B. Roark, M. Saraclar and M. Collins, \"Corrective language modeling for large vocabulary ASR with the perceptron algorithm,\" in Proceedings of International Conference on Acoustics, Speech and Signal Processing, pp. 749-752, 2004.",
	"links": null
	},
	"BIBREF27": {
	"ref_id": "b27",
	"title": "Discriminative reranking for machine translation",
	"authors": [
	{
	"first": "L",
	"middle": [],
	"last": "Shen",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Sarkar",
	"suffix": ""
	},
	{
	"first": "F",
	"middle": [
	"J"
	],
	"last": "Och",
	"suffix": ""
	}
	],
	"year": 2004,
	"venue": "Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "177--184",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "L. Shen, A. Sarkar and F. J. Och, \"Discriminative reranking for machine translation,\" in Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics, pp. 177-184, 2004.",
	"links": null
	},
	"BIBREF28": {
	"ref_id": "b28",
	"title": "Feature combination approaches for discriminative language models",
	"authors": [
	{
	"first": "E",
	"middle": [],
	"last": "Arisoy",
	"suffix": ""
	},
	{
	"first": "B",
	"middle": [],
	"last": "Ramabhadran",
	"suffix": ""
	},
	{
	"first": "H.-K",
	"middle": [
	"J"
	],
	"last": "Kuo",
	"suffix": ""
	}
	],
	"year": 2011,
	"venue": "Proceedings of Annual Conference of the International Speech Communication Association",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "E. Arisoy, B. Ramabhadran and H.-K. J. Kuo, \"Feature combination approaches for discriminative language models,\" in Proceedings of Annual Conference of the International Speech Communication Association, 2011.",
	"links": null
	},
	"BIBREF29": {
	"ref_id": "b29",
	"title": "A cascade ranking model for efficient ranked retrieval",
	"authors": [
	{
	"first": "L",
	"middle": [],
	"last": "Wang",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Lin",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [],
	"last": "Metzler",
	"suffix": ""
	}
	],
	"year": 2011,
	"venue": "Proceedings of Annual International ACM SIGIR Conference",
	"volume": "",
	"issue": "",
	"pages": "105--114",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "L. Wang, J. Lin and D. Metzler, \"A cascade ranking model for efficient ranked retrieval,\" in Proceedings of Annual International ACM SIGIR Conference, pp. 105-114, 2011.",
	"links": null
	}
	},
	"ref_entries": {
	"TABREF0": {
	"content": "<table><tr><td>\u6458\u8981 \u50b3\u7d71\u8a9e\u8a00\u6a21\u578b(Language Models)\u662f\u85c9\u7531\u4f7f\u7528\u5927\u91cf\u7684\u6587\u5b57\u8a9e\u6599\u8a13\u7df4\u800c\u6210\uff0c\u4ee5\u6a5f\u7387\u6a21\u578b\u4f86\u63cf \u8ff0\u81ea\u7136\u8a9e\u8a00\u7684\u898f\u5f8b\u6027\u3002N \u9023(N-gram)\u8a9e\u8a00\u6a21\u578b\u662f\u6700\u5e38\u898b\u7684\u8a9e\u8a00\u6a21\u578b\uff0c\u88ab\u7528\u4f86\u4f30\u6e2c\u6bcf\u4e00\u500b \u8a5e\u51fa\u73fe\u5728\u5df2\u77e5\u524d N-1 \u500b\u6b77\u53f2\u8a5e\u4e4b\u5f8c\u7684\u689d\u4ef6\u6a5f\u7387\u3002\u6b64\u5916\uff0c\u50b3\u7d71\u8a9e\u8a00\u6a21\u578b\u5927\u591a\u662f\u4ee5\u6700\u5927\u5316\u76f8 \u4f3c\u5ea6\u70ba\u8a13\u7df4\u76ee\u6a19\uff1b\u56e0\u6b64\uff0c\u7576\u5b83\u88ab\u4f7f\u7528\u65bc\u8a9e\u97f3\u8fa8\u8b58\u4e0a\u6642\uff0c\u5c0d\u65bc\u964d\u4f4e\u8a9e\u97f3\u8fa8\u8b58\u932f\u8aa4\u7387\u5e38\u6703\u6709 \u6240\u4fb7\u9650\u3002\u8fd1\u5e74\u4f86\uff0c\u6709\u5225\u65bc\u50b3\u7d71\u8a9e\u8a00\u6a21\u578b\u7684\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b(Discriminative Language Model) \u9678\u7e8c\u5730\u88ab\u63d0\u51fa\uff1b\u8207\u50b3\u7d71\u8a9e\u8a00\u6a21\u578b\u4e0d\u540c\u7684\u662f\uff0c\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u662f\u4ee5\u6700\u5c0f\u5316\u8a9e\u97f3\u8fa8\u8b58\u932f\u8aa4\u7387 \u505a\u70ba\u8a13\u7df4\u6e96\u5247\uff0c\u671f\u671b\u6240\u8a13\u7df4\u51fa\u7684\u8a9e\u8a00\u6a21\u578b\u53ef\u4ee5\u5e6b\u52a9\u964d\u4f4e\u8a9e\u97f3\u8fa8\u8b58\u7684\u932f\u8aa4\u7387\u3002\u672c\u8ad6\u6587\u63a2\u7a76 \u57fa\u65bc\u4e0d\u540c\u8a13\u7df4\u6e96\u5247\u7684\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\uff0c\u5206\u6790\u5404\u7a2e\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u4e4b\u57fa\u790e\u7279\u6027\uff0c\u4e26\u4e14\u6bd4\u8f03 \u5b83\u5011\u88ab\u4f7f\u7528\u65bc\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58(Large Vocabulary Continuous Speech Recognition, LVCSR)\u6642\u4e4b\u6548\u80fd\u3002\u540c\u6642\uff0c\u672c\u8ad6\u6587\u4ea6\u63d0\u51fa\u5c07\u908a\u969b(Margin)\u6982\u5ff5\u5f15\u5165\u65bc\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u7684\u8a13 \u7df4\u6e96\u5247\u4e2d\u3002\u5be6\u9a57\u7d50\u679c\u986f\u793a\uff0c\u76f8\u8f03\u65bc\u50b3\u7d71 N \u9023\u8a9e\u8a00\u6a21\u578b\uff0c\u4f7f\u7528\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u80fd\u5c0d\u65bc\u5927 \u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58\u6709\u76f8\u7576\u7a0b\u5ea6\u7684\u5e6b\u52a9\uff1b\u800c\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u7684\u57fa\u65bc\u908a\u969b\u8cc7\u8a0a\u4e4b\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21 \u578b\u4ea6\u80fd\u5920\u9032\u4e00\u6b65\u5730\u63d0\u5347\u8a9e\u97f3\u8fa8\u8b58\u7684\u6b63\u78ba\u7387\u3002 \u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\u8b58\u3001\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u3001\u908a\u969b\u3001\u8a13\u7df4\u6e96\u5247 \u4e00\u3001\u7dd2\u8ad6 \u5728\u4eba\u8207\u4eba\u7684\u4e92\u52d5\u7576\u4e2d\uff0c\u8a9e\u97f3\u662f\u6700\u81ea\u7136\u4e14\u76f4\u63a5\u7684\u8868\u9054\u65b9\u5f0f\u4e4b\u4e00\u3002\u900f\u904e\u8a9e\u97f3\uff0c\u4eba\u5011\u53ef\u4ee5\u5f7c\u6b64 \u6e9d\u901a\uff0c\u50b3\u9054\u60f3\u6cd5\u3001\u611f\u53d7\u4ee5\u53ca\u60c5\u7dd2\u3002\u56e0\u6b64\uff0c\u6211\u5011\u671f\u671b\u80fd\u8b93\u96fb\u8166\u5177\u5099\u8207\u4eba\u6e9d\u901a\u7684\u80fd\u529b\uff0c\u80fd\u70ba \u751f\u6d3b\u5e36\u4f86\u4fbf\u5229\u6027\u3002\u8981\u9054\u5230\u6b64\u76ee\u6a19\uff0c\u6211\u5011\u5fc5\u9808\u5148\u5c0d\u4f7f\u7528\u8005\u8f38\u5165\u7684\u8a9e\u97f3\u8a0a\u865f\u9032\u884c\u8fa8\u8b58\uff1b\u5f85\u8f49 \u63db\u6210\u6587\u5b57\u5f8c\uff0c\u518d\u5c0d\u6587\u5b57\u6240\u6b32\u8868\u9054\u7684\u8a9e\u610f\u4f5c\u7406\u89e3\uff0c\u9032\u800c\u505a\u51fa\u6700\u9069\u7576\u7684\u52d5\u4f5c\u4f86\u56de\u61c9\u4f7f\u7528\u8005\u3002 \u5c07\u8a9e\u97f3\u8a0a\u865f\u8f49\u63db\u6210\u6587\u5b57\u7684\u904e\u7a0b\uff0c\u53ef\u4ee5\u900f\u904e\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58(Automatic Speech Recognition, ASR)\u6280\u8853\u4f86\u5b8c\u6210\u3002\u5728\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7684\u904e\u7a0b\u4e2d\uff0c\u6211\u5011\u5fc5\u9808\u5148\u5c07\u8a9e\u97f3\u8a0a\u865f\u505a\u7279\u5fb5\u64f7\u53d6 \u5229\u7528\u50b3\u7d71\u8a9e\u8a00\u6a21\u578b(\u4f8b\u5982 N \u9023\u8a9e\u8a00\u6a21\u578b)\u6240\u9078\u51fa\u7684\u8a9e\u97f3\u8fa8\u8b58\u7d50\u679c\u901a\u5e38\u662f\u767c\u751f\u6a5f\u7387\u6700\u9ad8\u7684\u8a5e \u5e8f\u5217\uff0c\u4f46\u672a\u5fc5\u662f\u6700\u4f73(\u932f\u8aa4\u7387\u6700\u4f4e)\u7684\uff1b\u63db\u53e5\u8a71\u8aaa\uff0c\u5728\u5019\u9078\u8a5e\u5e8f\u5217\u4e2d\u5176\u5be6\u6709\u53ef\u80fd\u5b58\u5728\u8457\u5176 \u5b83\u64c1\u6709\u8f03\u4f4e\u932f\u8aa4\u7387\u7684\u8a5e\u5e8f\u5217\u53ef\u4ee5\u505a\u70ba\u8a9e\u97f3\u8fa8\u8b58\u5668\u7684\u8f38\u51fa\u3002\u65bc\u662f\uff0c\u6211\u5011\u5e0c\u671b\u80fd\u900f\u904e\u4f7f\u7528\u66f4 \u591a\u5176\u5b83\u8a9e\u8a00\u7279\u5fb5\uff0c\u4ee5\u53ca\u5019\u9078\u8a5e\u5e8f\u5217\u6240\u63d0\u4f9b\u7684\u8cc7\u8a0a\uff0c\u4e26\u7d93\u9069\u7576\u8a13\u7df4\u7684\u8a9e\u8a00\u6a21\u578b\u5c07\u6240\u6709\u5019\u9078 \u8a5e\u5e8f\u5217\u505a\u91cd\u65b0\u6392\u5e8f(Reranking)\uff0c\u4ee5\u8f38\u51fa\u64c1\u6709\u8f03\u4f4e\u932f\u8aa4\u7387\u7684\u8a9e\u97f3\u8fa8\u8b58\u7d50\u679c\u3002\u8fd1\u5e74\u4f86\uff0c\u6709\u8a31 \u591a\u5b78\u8005\u63a1\u7528\u9451\u5225\u5f0f\u8a13\u7df4(Discriminative Training)\u7684\u6982\u5ff5\u4f86\u8a13\u7df4\u8a9e\u8a00\u6a21\u578b\u4ee5\u5e6b\u52a9\u91cd\u65b0\u6392 \u5e8f\u3002\u8207\u50b3\u7d71\u8a9e\u8a00\u6a21\u578b\u4e0d\u540c\uff0c\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b(Discriminative Language Model)[2, 3, 4]\u662f\u4ee5 \u6700\u5c0f\u5316\u8a9e\u97f3\u8fa8\u8b58\u932f\u8aa4\u7387\u70ba\u8a13\u7df4\u76ee\u6a19\uff0c\u85c9\u7531\u4e00\u7d44\u9810\u5148\u5b9a\u7fa9\u7684\u8a9e\u8a00\u7279\u5fb5\u4ee5\u53ca\u6240\u5c0d\u61c9\u7684\u7279\u5fb5\u6b0a \u91cd\u53c3\u6578\uff0c\u5c07\u6240\u6709\u5019\u9078\u8a5e\u5e8f\u5217(\u5b58\u5728\u65bc\u8a5e\u5716\u6216 M \u689d\u6700\u4f73\u8fa8\u8b58\u5019\u9078\u8a5e\u5e8f\u5217)\u91cd\u65b0\u8a08\u5206 (Rescoring)\u6216\u91cd\u65b0\u6392\u5e8f(Reranking)\uff0c\u671f\u671b\u4f7f\u5177\u6709\u6700\u4f4e\u932f\u8aa4\u7387\u7684\u5019\u9078\u8a5e\u5e8f\u5217\u80fd\u64c1\u6709\u6700\u9ad8\u7684 \u5206\u6578(\u6392\u5e8f)\uff0c\u4e26\u4e14\u505a\u70ba\u6700\u5f8c\u7684\u8f38\u51fa\u7d50\u679c\u3002 \u672c\u8ad6\u6587\u5ef6\u7e8c\u6211\u5011\u5148\u524d\u5c0d\u65bc\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u4e4b\u7814\u7a76[5, 6]\uff0c\u63a2\u7a76\u57fa\u65bc\u4e0d\u540c\u8a13\u7df4\u6e96\u5247\u7684\u9451\u5225 \u5f0f\u8a9e\u8a00\u6a21\u578b\uff0c\u5206\u6790\u5404\u7a2e\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u4e4b\u57fa\u790e\u7279\u6027\uff0c\u4e26\u63d0\u51fa\u5c07\u908a\u969b(Margin)\u6982\u5ff5\u5f15\u5165\u65bc \u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u7684\u8a13\u7df4\u6e96\u5247\u4e2d\u3002\u672c\u8ad6\u6587\u7684\u5b89\u6392\u5982\u4e0b\uff1a\u7b2c\u4e8c\u7bc0\u5c07\u4ecb\u7d39\u8fd1\u5e74\u4f86\u5e38\u898b\u7684\u3001\u57fa\u65bc \u4e0d\u540c\u8a13\u7df4\u6e96\u5247\u7684\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\uff1b\u7b2c\u4e09\u7bc0\u5c07\u8aaa\u660e\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u57fa\u65bc\u908a\u969b\u8cc7\u8a0a\u4e4b\u9451\u5225\u5f0f\u8a9e (Feature Extraction)\u8a9e\u97f3\u8fa8\u8b58\u7d50\u679c\u3002 \u8a00\u6a21\u578b\uff1b\u7b2c\u56db\u7bc0\u662f\u5be6\u9a57\u7d50\u679c\u8207\u5206\u6790\uff1b\u7b2c\u4e94\u7bc0\u5247\u662f\u7d50\u8ad6\u8207\u672a\u4f86\u5c55\u671b\u3002</td></tr></table>",
	"text": "\uff0c\u4fdd\u7559\u8a9e\u97f3\u8a0a\u865f\u4e2d\u7684\u8072\u5b78\u7279\u6027(Acoustic Characteristics)\uff0c\u4e26\u8f49\u63db\u6210\u80fd \u4f7f\u96fb\u8166\u5bb9\u6613\u8655\u7406\u7684\u8072\u5b78\u7279\u5fb5\u5411\u91cf(Acoustic Feature Vector)\uff1b\u5229\u7528\u9019\u4e9b\u8072\u5b78\u7279\u5fb5\u5411\u91cf\uff0c\u6211 \u5011\u53ef\u4ee5\u70ba\u4e0d\u540c\u7684\u97f3\u7d20(Phoneme)\u5206\u5225\u5efa\u7acb\u8072\u5b78\u6a21\u578b(Acoustic Model)\uff0c\u9032\u800c\u7522\u751f\u53ef\u80fd\u7684\u5019 \u9078\u8a5e\u5e8f\u5217(Candidate Word Sequences)\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u6211\u5011\u4e5f\u5fc5\u9808\u6536\u96c6\u5927\u91cf\u7684\u6587\u5b57\u8a13\u7df4\u8a9e \u6599\uff0c\u7528\u4ee5\u7d71\u8a08\u81ea\u7136\u8a9e\u8a00\u4e2d\u5404\u7a2e\u8a5e\u5e8f\u5217\u7684\u51fa\u73fe\u60c5\u5f62\uff0c\u4e26\u85c9\u6b64\u8a13\u7df4\u8a9e\u8a00\u6a21\u578b(Language Model)\u3002\u50b3\u7d71\u8a9e\u8a00\u6a21\u578b\u662f\u6536\u96c6\u5404\u7a2e\u8a5e\u5f59\u51fa\u73fe\u5728\u81ea\u7136\u8a9e\u8a00\u4e2d\u7684\u8a5e\u983b\u6578\uff0c\u7d93\u7531\u6700\u5927\u5316\u76f8\u4f3c\u5ea6 \u4f30\u6e2c(Maximum Likelihood Estimation, MLE)\u4f86\u5efa\u7acb\u8a9e\u8a00\u6a21\u578b\u3002\u4f8b\u5982\uff0cN \u9023(N-gram)\u8a9e\u8a00 \u6a21\u578b[1]\u662f\u4f30\u6e2c\u6bcf\u4e00\u500b\u8a5e\u5728\u5176\u524d\u9762\u7dca\u9130 N-1 \u500b\u6b77\u53f2\u8a5e\u5e8f\u5217\u5df2\u77e5\u60c5\u6cc1\u4e0b\u7684\u689d\u4ef6\u6a5f\u7387\uff1b\u5b83\u53ef \u5354\u52a9\u8a9e\u97f3\u8fa8\u8b58\u5668\u5f9e\u6240\u7522\u751f\u7684\u5019\u9078\u8a5e\u5e8f\u5217\u4e2d\uff0c\u9078\u53d6\u6a5f\u7387\u6700\u9ad8(\u6700\u53ef\u80fd)\u7684\u8a5e\u5e8f\u5217\u505a\u70ba\u6700\u5f8c\u7684",
	"num": null,
	"html": null,
	"type_str": "table"
	},
	"TABREF7": {
	"content": "<table><tr><td>\u8a9e\u6599</td><td colspan=\"5\">\u6709\u7121\u8003\u616e\u6a23\u672c</td><td>\u53e5\u6578</td><td>\u9577\u5ea6(\u5c0f\u6642)</td></tr><tr><td colspan=\"2\">\u8a13\u7df4\u96c6\u8a9e\u6599 \u767c\u5c55\u96c6\u8a9e\u6599 \u6e2c\u8a66\u96c6\u8a9e\u6599</td><td>\u6b0a\u91cd</td><td colspan=\"2\">i W , \u03c9</td><td>j</td><td>\u6709\u7121\u8003\u616e R i W 30,600 1,998 1,997</td><td>\u4e00\u822c\u5316\u80fd\u529b</td><td>\u7d04 23 \u8a13\u7df4\u901f\u5ea6 \u7d04 1.5 \u7d04 1.5</td></tr><tr><td>Perceptron</td><td/><td colspan=\"2\">\u7121</td><td colspan=\"4\">\u6709 \u8868\u4e8c\u3001\u5be6\u9a57\u8a9e\u6599\u7d71\u8a08\u8cc7\u8a0a</td><td>\u5dee</td><td>\u5feb</td></tr><tr><td>MERT</td><td/><td colspan=\"2\">\u6709</td><td/><td/><td>\u7121</td><td>\u4f73</td><td>\u6162</td></tr><tr><td colspan=\"8\">GCLM \u63a5\u8457\uff0c\u6700\u5c0f\u5316\u932f\u8aa4\u7387\u8a13\u7df4(MERT)\u7684\u76ee\u6a19\u51fd\u6578\u4e2d\u6c92\u6709\u6839\u64da\u6700\u4f4e\u932f\u8aa4\u7387\u8a5e\u5e8f\u5217 R \u7121 \u6709 \u7565\u4f73 W \u70ba\u76ee\u6a19\u53c3 i \u6162 \u8003\u53bb\u505a\u8a13\u7df4\uff0c\u4f7f\u5f97\u5728\u8a13\u7df4\u7684\u904e\u7a0b\u7576\u4e2d\u8a13\u7df4\u8a9e\u53e5\u4e0d\u6703\u904e\u5ea6\u8a13\u7df4\u53bb\u9069\u5408\u9019\u4e9b\u6700\u4f4e\u932f\u8aa4\u7387\u8a5e\u5e8f</td></tr><tr><td colspan=\"8\">WGCLM \u5217\uff0c\u56e0\u6b64\u6700\u5c0f\u5316\u932f\u8aa4\u8a13\u7df4\u6703\u6709\u8f03\u4f73\u7684\u4e00\u822c\u5316\u80fd\u529b\uff1b\u53cd\u89c0\u611f\u77e5\u5668\u6f14\u7b97\u6cd5(Perceptron)\uff0c\u6703\u56e0 \u6709 \u6709 \u7565\u4f73 \u6162 \u70ba\u904e\u5ea6\u8a13\u7df4(Overfitting)\uff0c\u4f7f\u5f97\u6a21\u578b\u7684\u4e00\u822c\u5316\u80fd\u529b\u8f03\u5dee\u3002</td></tr><tr><td colspan=\"8\">R2D2 \u5728\u8a13\u7df4\u7684\u901f\u5ea6\u4e0a\uff0c\u5247\u56e0\u70ba\u5404\u5f0f\u65b9\u6cd5\u8457\u91cd\u7684\u8a13\u7df4\u76ee\u6a19\u4e0d\u540c\uff0c\u800c\u6709\u4e0d\u540c\u7684\u6642\u9593\u8907\u96dc\u5ea6\u3002\u611f\u77e5 \u6709 \u6709 \u7565\u4f73 \u5f88\u6162</td></tr><tr><td colspan=\"8\">MDLM \u5019\u9078\u8a5e\u5e8f\u5217\u4e4b\u9593\u7684\u95dc\u4fc2\uff1b\u8f2a\u8f49\u96d9\u91cd\u9451\u5225\u5f0f\u6a21\u578b(R2D2)\u662f\u8003\u616e\u4e86\u6240\u6709\u5019\u9078\u8a5e\u5e8f\u5217\u4e4b\u9593\u5f7c\u6b64 \u7121 \u6709 \u7565\u4f73 \u6162 \u8868\u4e00\u3001\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b\u4e4b\u9593\u7684\u6bd4\u8f03 \u5668\u6f14\u7b97\u6cd5\u5728\u8a13\u7df4\u7684\u904e\u7a0b\u7576\u4e2d\uff0c\u53ea\u8003\u616e\u6b63\u78ba(\u6216\u662f\u932f\u8aa4\u7387\u6700\u4f4e)\u7684\u5019\u9078\u8a5e\u5e8f\u5217\u8207\u6392\u5e8f\u5206\u6578\u6700 \u8fa8\u8b58\u5b57\u932f\u8aa4\u7387(Character Error Rate)\u5206\u5225\u70ba 11.26%\u300115.27%\u8207 16.39%\u3002\u4e26\u4e14\uff0c\u6211\u5011\u6311\u9078 \u9ad8\u7684\u8a5e\u5e8f\u5217\u4e4b\u9593\u7684\u95dc\u4fc2\uff1b\u6700\u5c0f\u5316\u932f\u8aa4\u7387\u8a13\u7df4\u3001\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b\u3001\u6b0a\u91cd\u5f0f\u5168\u57df\u689d \u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b\u5728\u8a13\u7df4\u7684\u904e\u7a0b\u4e2d\uff0c\u8003\u616e\u4e86\u6b63\u78ba(\u6216\u662f\u932f\u8aa4\u7387\u6700\u4f4e)\u7684\u5019\u9078\u8a5e\u5e8f\u5217\u8207\u5176\u5b83 \u57fa\u790e\u8fa8\u8b58\u5668\u7522\u751f\u7684\u524d 100 \u689d\u6700\u4f73( 100</td></tr><tr><td colspan=\"8\">\u7684\u95dc\u4fc2\u3002\u56e0\u6b64\u5728\u8a13\u7df4\u7684\u904e\u7a0b\u4e2d\uff0c\u8f2a\u8f49\u96d9\u91cd\u9451\u5225\u5f0f\u6a21\u578b\u6240\u9700\u7684\u6642\u9593\u8907\u96dc\u5ea6\u6700\u9ad8\uff0c\u76f8\u8f03\u4e0b\uff0c</td></tr><tr><td colspan=\"7\">\u611f\u77e5\u5668\u6f14\u7b97\u6cd5\u8a13\u7df4\u6642\u6240\u82b1\u8cbb\u6642\u9593\u6700\u5c11\u3002</td></tr><tr><td colspan=\"3\">\u56db\u3001\u5be6\u9a57\u7d50\u679c\u8207\u8a0e\u8ad6</td><td/><td/><td/><td/></tr><tr><td>(\u4e00)\u3001\u5be6\u9a57\u8a9e\u6599</td><td/><td/><td/><td/><td/><td/></tr><tr><td colspan=\"8\">\u672c \u8ad6 \u6587 \u5be6 \u9a57 \u8a9e \u6599 \u53d6 \u81ea \u516c \u8996 \u65b0 \u805e (Mandarian Across Taiwan-Broadcast News,</td></tr><tr><td colspan=\"8\">MATBN)[24]\u3002\u516c\u8996\u65b0\u805e\u8a9e\u6599\u662f 2001 \u5e74\u81f3 2003 \u5e74\u9593\u7531\u4e2d\u7814\u9662\u8cc7\u8a0a\u6240\u53e3\u8a9e\u5c0f\u7d44(SLG)\u8207\u516c</td></tr><tr><td colspan=\"8\">\u5171\u96fb\u8996\u53f0(PTS)\u5408\u4f5c\u9304\u88fd\uff0c\u5305\u542b\u4e86\u5167\u5834\u65b0\u805e\u8207\u5916\u5834\u65b0\u805e\u5169\u500b\u90e8\u5206\u3002\u5176\u4e2d\u5167\u5834\u65b0\u805e\u70ba\u4e3b\u64ad</td></tr><tr><td colspan=\"8\">\u8a9e\u6599\uff0c\u5916\u5834\u65b0\u805e\u8a9e\u6599\u5305\u542b\u6709\u63a1\u8a2a\u8a18\u8005(Field Reporters)\u8a9e\u97f3\u8a9e\u6599\u8207\u53d7\u8a2a\u8005(Interviewees)\u8a9e</td></tr><tr><td>\u97f3\u8a9e\u6599\u3002</td><td/><td/><td/><td/><td/><td/></tr><tr><td colspan=\"8\">\u7531 \u65bc \u5167 \u5834 \u4e3b \u64ad \u8a9e \u6599 \u5927 \u90e8 \u5206 \u4f86 \u81ea \u65bc \u540c \u4e00 \u4e3b \u64ad \u6240 \u9304 \u88fd \uff0c \u70ba \u4e86 \u907f \u514d \u8a9e \u8005 \u76f8 \u4f9d (Speaker</td></tr><tr><td colspan=\"8\">Dependent)\u73fe\u8c61\u9020\u6210\u5be6\u9a57\u504f\u5dee\uff0c\u6545\u4e0d\u63a1\u7528\u5167\u5834\u4e3b\u64ad\u8a9e\u6599\uff1b\u5916\u5834\u53d7\u8a2a\u8005\u8a9e\u6599\uff0c\u5247\u662f\u5305\u542b\u8a31</td></tr><tr><td colspan=\"8\">f \u591a\u8a9e\u52a9\u8a5e\u8207\u80cc\u666f\u97f3\u6a02\uff0c\u6240\u4ee5\u4e5f\u6c92\u6709\u63a1\u7528\uff1b\u56e0\u6b64\uff0c\u672c\u8ad6\u6587\u7684\u5be6\u9a57\u8a9e\u6599\u9078\u53d6\u81ea\u5916\u5834\u63a1\u8a2a\u8a18\u8005 \u03b7 \u03bb \u03bb (29)</td></tr><tr><td colspan=\"8\">\u900f\u904e\u6b64\u66f4\u65b0\u5f0f\u4ee5\u53ca\u8fed\u4ee3\u7684\u8a13\u7df4\uff0c\u671f\u671b\u80fd\u6c42\u5f97\u6700\u4f73\u7684\u7279\u5fb5\u6b0a\u91cd\u5411\u91cf\u3002 \u5229\u7528\u8868\u4e00\u4f86\u8aaa\u660e\u672c\u8ad6\u6587\u5f15\u5165\u7684\u57fa\u65bc\u908a\u969b\u8cc7\u8a0a\u4e4b\u9451\u5225\u5f0f\u8a9e\u8a00\u6a21\u578b(MDLM)\u8207\u5176\u5b83\u9451\u5225\u5f0f \u8a9e\u8a00\u6a21\u578b\u4e4b\u9593\u7684\u6bd4\u8f03\u95dc\u4fc2\u3002\u5176\u4e2d\uff0c\u8003\u616e\u6a23\u672c\u6b0a\u91cd\u7684\u597d\u8655\u662f\u80fd\u5c07\u8a13\u7df4\u8a9e\u53e5\u6839\u64da\u5176\u6a23\u672c\u6b0a\u91cd \u7684\u4e0d\u540c\uff0c\u5c0d\u65bc\u6a21\u578b\u7684\u8a13\u7df4\u5f71\u97ff\u7a0b\u5ea6\u6709\u4e0d\u540c\u7684\u5f71\u97ff\uff0c\u800c\u975e\u6bcf\u4e00\u500b\u8a13\u7df4\u8a9e\u53e5\u90fd\u4f54\u6709\u76f8\u540c\u7684\u6b0a \u91cd\u3002\u5728\u7b2c\u56db\u7bc0\u5be6\u9a57\u7d50\u679c\u89c0\u5bdf\uff0c\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b(GCLM)\u8207\u6b0a\u91cd\u5f0f\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d \u6578\u7dda\u6027\u6a21\u578b(WGCLM)\u4f86\u6bd4\u8f03\uff0c\u6b0a\u91cd\u5f0f\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b\u591a\u52a0\u5165\u4e86\u6a23\u672c\u6b0a\u91cd\uff0c\u5176 \u7d50\u679c\u512a\u65bc\u5168\u57df\u689d\u4ef6\u5f0f\u5c0d\u6578\u7dda\u6027\u6a21\u578b\uff0c\u53ef\u63a8\u6e2c\u6a23\u672c\u6b0a\u91cd\u7684\u8003\u91cf\u5c0d\u65bc\u6a21\u578b\u8a13\u7df4\u4e0a\u7684\u78ba\u662f\u6709\u6b63 \u8a9e\u6599\u3002\u8a13\u7df4\u96c6\u8a9e\u6599\u3001\u6e2c\u8a66\u96c6\u8a9e\u6599\u53ca\u767c\u5c55\u96c6\u8a9e\u6599\u7686\u9078\u53d6\u81ea\u516c\u8996\u65b0\u805e 2001 \u6211\u5011\u4ee5\u57fa\u790e\u8fa8\u8b58\u5668[26]\u914d\u5408\u80cc\u666f\u4e09\u9023\u8a9e\u8a00\u6a21\u578b\u65bc\u5b8c\u6574\u8a5e\u5716\u641c\u5c0b(Word Graph Rescoring)\u7684 \u5411\u5e6b\u52a9\u3002 \u6700\u4f73\u7d50\u679c\u505a\u70ba\u57fa\u790e\u8fa8\u8b58\u7387(Baseline)\uff0c\u5b83\u5728\u8a13\u7df4\u96c6\u8a9e\u6599\u3001\u767c\u5c55\u96c6\u8a9e\u6599\u4ee5\u53ca\u6e2c\u8a66\u96c6\u8a9e\u6599\u7684</td></tr></table>",
	"text": "\u5e74\u81f3 2002 \u5e74\u5916\u5834 \u63a1\u8a2a\u8a18\u8005\uff0c\u5206\u5225\u70ba 30,600 \u53e5(\u7d04 23 \u5c0f\u6642)\u30011,997 \u53e5(\u7d04 1.5 \u5c0f\u6642)\u53ca 1,998 \u53e5(\u7d04 1.5 \u5c0f\u6642)\u3002 \u5982\u8868\u4e8c\u6240\u793a\u3002 \u80cc\u666f\u8a9e\u8a00\u6a21\u578b\u70ba\u4e09\u9023\u8a9e\u8a00\u6a21\u578b(Trigram Language Model)\uff0c\u63a1\u7528 Katz Back-off Smoothing \u5e73\u6ed1\u5316\u65b9\u6cd5\u4f86\u89e3\u6c7a\u8cc7\u6599\u7a00\u758f\u7684\u554f\u984c\u3002\u5176\u8a13\u7df4\u8a9e\u6599\u4f86\u81ea 2001 \u5e74\u81f3 2002 \u5e74\u4e2d\u592e\u901a\u8a0a\u793e (Central News Agency, CNA)\u7684\u6587\u5b57\u65b0\u805e\u8a9e\u6599\uff0c\u5305\u542b\u4e86\u7d04\u4e00\u5104\u4e94\u5343\u842c\u500b\u4e2d\u6587\u5b57\uff0c\u7d93\u904e\u65b7\u8a5e \u5f8c\u7d04\u6709\u516b\u5343\u842c\u8a5e\u3002\u6b64\u8a9e\u8a00\u6a21\u578b\u662f\u4f7f\u7528 SRI Language Modeling Toolkit(SRILM)[25]\u8a13\u7df4\u6240 \u5f97\u3002",
	"num": null,
	"html": null,
	"type_str": "table"
	}
	}
	}
	}