| { | |
| "paper_id": "O10-1005", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T08:06:43.736651Z" | |
| }, | |
| "title": "A Comparative Study on Margin-Based Discriminative Training of Acoustic Models", | |
| "authors": [ | |
| { | |
| "first": "Yueng-Tien", | |
| "middle": [], | |
| "last": "\u7f85\u6c38\u5178", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Lo", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Berlin", | |
| "middle": [], | |
| "last": "\u9673\u67cf\u7433", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Normal University", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Normal University", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "O10-1005", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [ | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "\uf028 \uf029 O W P W i W i max ar\u011d \uf03d (1) \u5176\u4e2d i W \u4ee3\u8868\u6240\u6709\u67d0\u4e00\u689d\u5019\u9078\u8a5e\u5e8f\u5217\uff0c \uf028 \uf029 O W P i \u70ba\u7d66\u5b9a\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217\u5f8c O \uff0c i W \u767c\u751f\u7684 \u4e8b\u5f8c\u6a5f\u7387(Posterior Probability)\u3002\u82e5\u4f7f\u7528\u8c9d\u5f0f\u5b9a\u7406(Bayes' Theorem)\u5c07\u5f0f(1)\u4e2d\u7684\u4e8b\u5f8c\u6a5f\u7387 \u9805\u5c55\u958b\u6210 \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 O p W P W O p O W P i i i \uf03d (2) \u5176\u4e2d \uf028 \uf029 O p \u70ba\u7522\u751f\u8a9e\u53e5 O \u7684\u4e8b\u524d\u6a5f\u7387\u4e26\u4e0d\u5f71\u97ff\u6240\u6709\u5019\u9078\u8a5e\u5e8f\u5217\u4e4b\u6392\u5e8f\uff0c\u6545\u5f0f(1)\u53ef\u7c21\u5316 \u70ba\uff1a \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 i W i i W W O p W P W O p W i i , max arg max ar\u011d \uf03d \uf03d", | |
| "eq_num": "(3)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(4)", | |
| "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": "\u5176\u4e2d \uf04c \u662f\u8a9e\u97f3\u8fa8\u8b58\u5668\u6a21\u578b(\u8072\u5b78\u6a21\u578b\u53ca\u8a9e\u8a00\u6a21\u578b)\u53c3\u6578\uff0c\u800c\u8a9e\u97f3\u8fa8\u8b58\u5668\u7d66\u4e88\u5176\u5b83\u975e\u6b63\u78ba\u8f49 \u5beb\u4e4b\u5019\u9078\u8a5e\u5e8f\u5217\u7684\u5206\u6578\u53ef\u8868\u793a\u6210 \uf068 \uf068 1 , ) , ( 1 log ) ( \uf0fe \uf0fd \uf0fc \uf0ee \uf0ed \uf0ec \uf03d \uf0e5 \uf0b9 \uf04c \uf04c zR zi zi W W W zi z z W O p N O G (5) \u5176\u4e2d N \u4ee3\u8868\u6240\u6709\u975e\u6b63\u78ba\u8f49\u5beb\u4e4b\u5019\u9078\u8a5e\u5e8f\u5217\u7684\u500b\u6578\uff0c\u800c zi w \u662f\u5176\u4e2d\u7684\u4e00\u689d\u5019\u9078\u8a5e\u5e8f\u5217\uff1b\uf068 \u70ba \u8abf\u6574\u806f\u5408\u6a5f\u7387\u5206\u5e03\u7684\u7bc4\u570d\u56e0\u5b50\uff0c\u70ba\u7c21\u5316\u8a08\u7b97\uff0c\u5728\u672c\u8ad6\u6587\u6211\u5011\u4ee4 1 \uf03d \uf068 \uff0c\u95dc\u65bc\uf068 \u7684\u9032\u4e00\u6b65\u8a0e \u8ad6\u53ef\u53c3\u95b1[13]\u3002\u5f0f (5)\u5c0d\u65bc\u6240\u6709\u975e\u6b63\u78ba\u8f49\u5beb\u4e4b\u5019\u9078\u8a5e\u5e8f\u5217\u5206\u6578\u7684\u7e3d\u548c\u5728\u4e00\u6b63\u898f\u5316\u5f8c\u518d\u53d6\u5c0d \u6578\uff0c\u53ef\u8996\u70ba\u65bc\u662f\u4e00\u7a2e\u67d4\u6027\u6700\u5927\u81e8\u754c\u503c\uff0c\u4ee3\u8868\u7684\u662f\u8207\u6b63\u78ba\u8f49\u5beb\u5206\u6578\u76f8\u8fd1\u7684\u932f\u8aa4\u8a9e\u97f3\u8fa8\u8b58\u7d50 \u679c\u3002\u65bc\u662f\uff0c\u5206\u985e\u932f\u8aa4\u8a55\u4f30\u51fd\u6578 ) ( z O d \uf04c \u53ef\u4ee5\u5b9a\u7fa9\u70ba ) ( ) ( ) ( z z z O G O g O d \uf04c \uf04c \uf04c \uf02b \uf02d \uf03d (6) \u503c\u5f97\u6ce8\u610f\u7684\u662f(6)\u5f0f\u4e2d\u547c\u61c9\u4e86\u9451\u5225\u5f0f\u8a13\u7df4\u7684\u7cbe\u795e\uff0c\u63cf\u8ff0\u8a13\u7df4\u8a9e\u53e5\u4e4b\u6b63\u78ba\u8f49\u5beb\u8207\u8a9e\u97f3\u8fa8\u8b58 \u7d50\u679c(\u8a5e\u5716\u4e2d\u5019\u9078\u8a5e\u5e8f\u5217)\u4e4b\u9593\u7684\u95dc\u4fc2\u3002\u85c9\u7531\u5206\u985e\u932f\u8aa4\u8a55\u4f30 ) ( z O d \uf04c \u4f86\u63cf\u8ff0\u8a9e\u97f3\u8fa8\u8b58\u5668\u5c0d\u6bcf \u4e00\u8a13\u7df4\u8a9e\u53e5\u7684\u6c7a\u7b56\u7d50\u679c\uff1a\u7576 0 ) ( \uf03c \uf04c z O d \uff0c\u6642\u8868\u793a\u8a13\u7df4\u8a9e\u53e5 z \u88ab\u6a21\u578b \uf04c \u5206\u985e\u6b63\u78ba\uff1b 0 ) ( \uf0b3 \uf04c z O d \u5247\u8868\u793a z \u88ab\u932f\u8aa4\u5206\u985e\u3002\u800c\u4e00\u822c\u5728\u7d66\u5b9a\u5206\u985e\u932f\u8aa4\u8a55\u4f30\u51fd\u6578\u5f8c\uff0c\u5c0d\u65bc\u6bcf\u4e00\u53e5\u8a13\u7df4 \u8a9e\u53e5\u7684\u640d\u5931\u51fd\u6578\u7d93\u5e38\u53ef\u900f\u904e\u4f7f\u7528 s \u578b\u51fd\u6578(Sigmoid Function)\u800c\u5f97\uff1a \uf028 \uf029 ) ( 1 1 ) ( z O d z e O d l \uf04c \uf0d7 \uf02d \uf04c \uf02b \uf03d \uf061 (7) \u5176\u4e2d 0 \uf03e \uf061 \uff0c\u53ef\u8abf\u6574 s \u578b\u51fd\u6578\u50be\u659c\u7a0b\u5ea6\uff0c\u5728\u672c\u8ad6\u6587\u70ba\u6211\u5011\u8a2d\u5b9a 1 \uf03d \uf061 \uff0c\u95dc\u65bc\uf061 \u8a0e\u8ad6\u4ea6\u53ef\u53c3 \u95b1[13]\u3002\u7531 s \u578b\u5e73\u6ed1\u51fd\u6578\u6211\u5011\u53ef\u4ee5\u6ce8\u610f\u5230\u640d\u5931\u51fd\u6578\u7684\u503c\u843d\u65bc 0 \u81f3 1 \u5340\u9593\u5167\uff0c\u6211\u5011\u5c07\u9451\u5225 \u51fd\u6578(\u5f0f(4)\u8207(5))\u5957\u7528\u65bc\u5f0f(7)\uff0c\u4ee4 1 \uf03d \uf061 \u8207 1 \uf03d \uf068 \uff0c\u4e26\u91cd\u65b0\u6574\u7406\u5f8c\u53ef\u5f97\u5230\uff1a \uf028 \uf029 \uf0e5 \uf0e5 \uf0e5 \uf0e5 \uf04c \uf0b9 \uf04c \uf04c \uf0b9 \uf04c \uf0b9 \uf04c \uf04c \uf03d \uf02b \uf03d zi zR zi zi zR zi zi zR zi zi W zi z W W W zi z zR z W W W zi z W W W zi z z W O p W O p W O p W O p W O p O d l ) , ( ) , ( ) , ( ) , ( ) , ( ) ( , , ,", | |
| "eq_num": "(8)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "\u5c0d\u65bc\u8a13\u7df4\u8a9e\u6599\uff0c\u4ee5\u6700\u5c0f\u5206\u985e\u932f\u8aa4\u70ba\u6e96\u5247\u7684\u9451\u5225\u5f0f\u8a13\u7df4\uff0c\u5176\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u76ee\u6a19\u51fd\u6578 ) (\uf04c MCE F \u53ef\u8868\u793a\u6210\u6700\u5c0f\u5316\u6240\u6709\u8a13\u7df4\u8a9e\u53e5\u7684\u671f\u671b\u932f\u8aa4\u7387 \uf028 \uf029 \uf0e5 \uf03d \uf04c \uf03d \uf04c Z z z MCE O d l Z F 1 ) ( 1 ) (", | |
| "eq_num": "(9)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u6211\u5011\u82e5\u9032\u4e00\u6b65\u5b9a\u7fa9\u4e00\u500b\u70ba\"1 \u6e1b\u53bb\u640d\u5931\u51fd\u6578\"\u7684\u529f\u7528\u51fd\u6578\uff1a ", | |
| "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": "\uf028 \uf029 \uf028 \uf029 ) ( 1 ) ( z z O d l O d u \uf04c \uf04c \uf02d \uf03d (10) \u5247\u53ef\u770b\u51fa\u6700\u5c0f\u5316\u640d\u5931\u51fd\u6578(\u5f0f(9))\u7b49\u540c\u65bc\u6700\u5927\u5316\u4e0b\u5217\u6700\u5c0f\u5206\u985e\u932f\u8aa4\u76ee\u6a19\u51fd\u6578 ) ( \uf04c MCE F \uff1a \uf028 \uf029 \uf0e5 \uf0e5 \uf0e5 \uf03d \uf04c \uf04c \uf03d \uf04c \uf03d \uf03d \uf04c Z z W zi z zR z Z z z MCE zi W O p W O p O d u F 1 1 ) , ( ) , ( ) ( ) (", | |
| "eq_num": "(11)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "(II)\u6700\u5927\u5316\u4ea4\u4e92\u8cc7\u8a0a\u4f30\u6e2c\u6cd5\u5247\uff1a\u4ee5\u6700\u5927\u5316\u4ea4\u4e92\u8cc7\u8a0a\u4f30\u6e2c\u6cd5\u70ba\u6e96\u5247\u7684\u9451\u5225\u5f0f\u8072\u5b78\u6a21\u578b\u8a13 \u7df4\uff0c\u4e3b\u8981\u76ee\u7684\u662f\u6700\u5927\u5316\u6240\u6709\u8a13\u7df4\u8a9e\u53e5 z \u7684\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217 z O \u8207\u5176\u5c0d\u61c9\u8f49\u5beb zR W \u7684\u4ea4\u4e92 \u8cc7\u8a0a \uf028 \uf029 \uf04c MMI F \uff0c\u5b9a\u7fa9\u5982\u4e0b\uff1a \uf028 \uf029 ) | ( ) ( ) ( ) | ( log ) , ( ) ( ) ( ) , ( log ) , ( , , z zR zR W O zR z zR zR z W O zR z zR z zR z MMI O W H W H W P O W p W O p W P O p W O p W O p F zR z zR z \uf02d \uf03d \uf03d \uf03d \uf04c \uf0e5 \uf0e5 \uf04c \uf04c \uf04c \uf04c \uf04c (12) \u5176\u4e2d \uf028 \uf029 \uf0e5 \uf04c \uf04c \uf02d \uf03d zR W zR zR zR W P W P W H ) ( log ) ( \u70ba zR W \u7684\u71b5\u503c\uff1b\u800c \uf028 \uf029 z zR O W H | \u70ba\u7d66\u5b9a\u689d\u4ef6 z O \u6642 zR W \u7684\u71b5\u503c\uff0c\u53ef\u8868\u793a\u70ba \uf028 \uf029 \uf0e5 \uf04c \uf02d \uf03d z zR O W z zR z zR z zR O W p O W p O W H , ) | ( log ) , ( | \u3002\u5047\u5b9a\u5728\u6a21\u578b\u8a13\u7df4 \u6642\uff0c\u8a9e\u8a00\u6a21\u578b ) ( zR W P \uf04c \u4e0d\u505a\u8abf\u6574(\u4ea6\u5373 ) ( zR W H \u4e0d\u6539\u8b8a)\u7684\u60c5\u6cc1\u4e0b\uff0c\u6700\u5927\u5316\u5f0f(12)\u7b49\u540c\u65bc\u6700 \u5c0f\u5316 \uf028 \uf029 z zR O W H | \u3002\u82e5\u518d\u5047\u8a2d\u8a13\u7df4\u8a9e\u53e5\u70ba\u6709\u76f8\u540c\u6a5f\u7387\u5206\u5e03\u503c(Uniformly Distributed)\u4e0b\uff0c\u5247 \uf028 \uf029 z z O W H | \u53ef\u8868\u793a\u8fd1\u4f3c\u5730\u6210 \uf028 \uf029 \uf0e5 \uf0e5 \uf03d \uf04c \uf04c \uf03d \uf04c \uf02d \uf03d \uf02d \uf03d Z z z z zR Z z z zR z zR O p O W p Z O W p Z O W H 1 1 ) ( ) , ( log 1 ) | ( log 1 | (13) \u56e0\u6b64\uff0c\u6211\u5011\u6700\u5c0f\u5316\u5f0f(12)\u5373\u662f\u6700\u5927\u5316\u4e0b\u5217\u51fd\u6578\uff1a \uf0e5 \uf0e5 \uf0e5 \uf03d \uf04c \uf04c \uf03d \uf04c \uf04c \uf03d \uf03d \uf04c Z z W zi z zR z Z z z zR z MMI zi W O p W O p O p W O p F 1 1 ) , ( ) , ( log ) ( ) , ( log ) ( (14) \u63a5\u4e0b\u4f86\uff0c\u6211\u5011\u5c07\u63a2\u8a0e\u5f0f(14)\u8207\u5176\u5b83\u57fa\u65bc\u4e0d\u540c\u6e96\u5247\u9451\u5225\u5f0f\u51fd\u6578\u7684\u6bd4\u8f03\u3002\u82e5\u5c0d\u5f0f(14)\u4e2d\u76ee\u6a19 \u51fd\u6578 ) ( \uf04c MMI F \u53d6\u6307\u6578\u51fd\u6578\uff0c\u5247\u76ee\u6a19\u51fd\u6578\u8b8a\u6210\uff1a \uf05b \uf05d \uf0d5 \uf0e5 \uf03d \uf04c \uf04c \uf03d \uf04c \uf03d \uf04c Z z W zi z zR z MMI MMI zi W O p W O p F F 1 ) , ( ) , ( ) ( exp ) ( (15) \u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u56e0\u70ba\u5728\u6307\u6578\u51fd\u6578\u662f\u55ae\u8abf\u905e\u589e\u8f49\u63db\uff0c\u7d93\u6b64\u51fd\u6578\u8f49\u63db\u5f8c\u7684\u65b0\u76ee\u6a19\u51fd\u6578 ) ( \uf04c MMI F \u5728\u6975\u5927\u503c\u6642\u5c0d\u61c9\u7684\u53c3\u6578\u662f\u4e0d\u8b8a\u7684\uff0c\u4ea6\u5373 ) ( \uf04c MMI F \u8207 ) ( \uf04c MMI F \u5728\u5404\u81ea\u6975\u5927\u503c\u9ede \u6709\u76f8\u540c\u53c3\u6578\u89e3\u96c6\u5408\u3002\u6211\u5011\u4ea6\u53ef\u5c07 ) | ( z zR O W p \uf04c \u4f5c\u6539\u5beb\uff0c\u4ee5\u4fbf\u65bc\u8207\u5176\u4ed6\u9451\u5225\u5f0f\u8a13\u7df4\u6e96\u5247\u505a \u6bd4\u8f03\uff1a \uf028 \uf029 \uf034 \uf034 \uf034 \uf034 \uf034 \uf038 \uf034 \uf034 \uf034 \uf034 \uf034 \uf037 \uf036 \u671f\u671b\u932f\u8aa4 ) | ( ) , ( 1 1 ) | ( 1 ) , ( ) , ( ) | ( \uf0e5 \uf0e5 \uf0e5 \uf04c \uf0b9 \uf04c \uf04c \uf04c \uf04c \uf02d \uf02d \uf03d \uf02d \uf03d \uf03d zi zR zi zi W z zi zR zi W W z zi W zi z zR z z zR O W p W W O W p W O p W O p O W p \uf064 (16) \u6700\u5f8c\uff0c\u6211\u5011\u53ef\u4ee5\u5c07\u5f0f(16)\u8996\u70ba\u8a13\u7df4\u8a9e\u53e5 z \u7684\u671f\u671b\u6b63\u78ba\u7387\uff0c\u4e5f\u5c31\u662f\u7b49\u65bc 1 \u6e1b\u53bb\u8a13\u7df4\u8a9e\u53e5 z \u7684 \u671f\u671b\u932f\u8aa4\u7387\u3002\u503c\u5f97\u6ce8\u610f\u662f\u5728 MCE \u4e2d\u70ba\u6240\u6709\u8a13\u7df4\u8a9e\u53e5\u671f\u671b\u6b63\u78ba\u7387\u7684\u7e3d\u548c\u800c MMI \u70ba\u6240\u6709 \u8a13\u7df4\u8a9e\u53e5\u671f\u671b\u6b63\u78ba\u7387\u7684\u9023\u4e58\u7a4d\uff0c\u80cc\u5f8c\u7684\u610f\u7fa9\u5728\u67d0\u7a2e\u7a0b\u5ea6\u4e0a\u7686\u4ee3\u8868\u6700\u5927\u5316\u8a9e\u97f3\u8fa8\u8b58\u5668\u5c0d\u65bc \u6240\u6709\u8a13\u7df4\u8a9e\u53e5\u7684\u671f\u671b\u6b63\u78ba\u7387\u3002 (III)\u6700\u5c0f\u5316\u97f3\u7d20\u932f\u8aa4\u8a13\u7df4(Minimum Phone Error Training, MPE)\u6cd5\u5247\uff1a\u76f8\u8f03\u65bc\u6700\u5927\u4ea4\u4e92 \u8cc7\u8a0a\u6cd5\u5247\u8207\u6700\u5c0f\u5316\u5206\u985e\u932f\u8aa4\u6cd5\u5247\u662f\u8457\u91cd\u65bc\u8a13\u7df4\u8a9e\u53e5\u6574\u9ad4(String Level)\u7684\u6b63\u78ba\u7387\u63d0\u6607\uff0c\u6700 \u5c0f\u5316\u97f3\u7d20\u932f\u8aa4\u8a13\u7df4\u6cd5\u5247\u8457\u91cd\u65bc\u5c0d\u8a13\u7df4\u8a9e\u53e5\u8f03\u7d30\u5fae\u5c64\u7d1a(\u5982\u97f3\u7d20\u3001\u5b57\u6216\u8a5e\u7b49)\u7684\u6b63\u78ba\u7387\u63d0\u6607\u3002 \u4f8b\u5982\uff0c\u4f7f\u7528\u6700\u5927\u5316\u97f3\u7d20\u7684\u671f\u671b\u6b63\u78ba\u7387\u4f5c\u70ba\u6700\u5c0f\u5316\u97f3\u7d20\u932f\u8aa4\u7684\u76ee\u6a19\u51fd\u6578\uff0c\u53ef\u4ee5\u5b9a\u7fa9\u70ba[3]\uff1a \uf0e5 \uf0e5 \uf0e5 \uf03d \uf04c \uf04c \uf03d \uf04c Z z W zi z zR zi Phone W zi z MPE zi zi W O p W W A W O p F 1 ) , ( ) , ( ) , ( ) ( (17) \u5176\u4e2d ) , ( zR zi W W A \u70ba\u8a13\u7df4\u8a9e\u53e5 z \u7684\u5019\u9078\u8a5e\u5e8f\u5217 zi W \u4e4b\u539f\u59cb\u97f3\u7d20\u6b63\u78ba\u500b\u6578\uff0c\u53ef\u4ee5\u5b9a\u7fa9\u70ba\u5728\u6b63 \u78ba\u8f49\u5beb zR W \u4e0a\u6240\u6709\u97f3\u7d20\u500b\u7d20\u6e1b\u53bb zi W \u7522\u751f\u63d2\u5165\u3001\u522a\u9664\u3001\u66ff\u63db\u7b49\u932f\u8aa4\u500b\u6578\uff0cMPE \u76ee\u6a19\u51fd\u6578 \u5e0c\u671b\u80fd\u589e\u9032\u8a9e\u97f3\u8fa8\u8b58\u5668\u5c0d\u65bc\u8a13\u7df4\u8a9e\u6599\u6240\u6709\u8fa8\u8b58\u8f38\u51fa(\u4e5f\u5c31\u662f\u6240\u6709\u5019\u9078\u8a5e\u5e8f\u5217\uff0c\u53ef\u5305\u62ec\u6b63 \u78ba\u8f49\u5beb)\u7684\u671f\u671b\u97f3\u7d20\u6b63\u78ba\u7387 \uf0e5\uf0e5 \uf03d \uf04c \uf03d \uf04c Z z W zR zi Phone z zi MPE zi W W A O W P F 1 ) , ( ) | ( ) ( (18) \u5176\u4e2d \uf0e5 \uf04c \uf04c \uf04c \uf04c \uf04c \uf03d \uf03d zi W zi z zi z z zi z z zi W O p W O p O p W O p O W P ) , ( ) , ( ) ( ) , ( ) | (", | |
| "eq_num": "(19" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": ") | ( ) | ( min ) | ( max ) | ( ) ( , , zi z zR z W W W zi z W W W zR z z W O p W O p W O p W O p O d zR zi zi zR zi zi \uf04c \uf04c \uf0b9 \uf0ce \uf04c \uf0b9 \uf0ce \uf04c \uf02d \uf03d \uf02d \uf03d W W (20) \u5176\u4e2d W \u70ba\u8a9e\u97f3\u8fa8\u8b58\u5668\u7522\u751f\u7684\u6240\u6709\u53ef\u80fd\u5019\u9078\u8a5e\u5e8f\u5217\u6240\u6210\u96c6\u5408\uff0c ) | ( zR z W O p \uf04c \u70ba\u7d66\u5b9a\u6b63\u78ba\u8a5e \u5e8f\u5217 zR W \u7522\u751f\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217 z O \u7684\u76f8\u4f3c\u5ea6\uff0c zi W \u70ba\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217 z O \u7684\u53ef\u80fd\u8fa8\u8b58\u7d50 \u679c(\u5019\u9078\u8a5e\u5e8f\u5217)\uff0c ) | ( zi z W O p \uf04c \u70ba\u5019\u9078\u8a5e\u5e8f\u5217 zi W \u7522\u751f\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217 z O \u7684\u76f8\u4f3c\u5ea6\u3002\u7531 \u5f0f(20)\u53ef\u4ee5\u77e5\u9053\u5206\u96e2\u908a\u969b\u7684\u8a08\u7b97\u5c31\u662f\u6b63\u78ba\u8f49\u5beb\u8207\u6700\u6709\u53ef\u80fd(\u6a5f\u7387\u6700\u5927)\u7684\u5019\u9078\u8a5e\u5e8f\u5217\u7684\u76f8 \u4f3c\u5ea6\u4e4b\u5dee\u3002\u800c\u7576\u8a9e\u97f3\u8fa8\u8b58\u5668\u4e43\u662f\u5efa\u69cb\u5728\u6700\u5927\u5316\u4e8b\u5f8c\u6a5f\u7387\u89e3\u78bc\u65b9\u6cd5\u4e0a\u6642(\u9019\u88e1\u5047\u8a2d ) ( zR W P \uf04c \u8207\u6bcf\u4e00\u500b ) ( zi W P \uf04c \u90fd\u662f\u76f8\u7b49\u7684)\uff0c\u82e5 0 ) ( \uf03e z O d \uff0c\u5247\u8868\u793a\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217 z O \u88ab\u76ee\u524d\u7684\u8fa8\u8b58 \u5668\u6b63\u78ba\u5730\u8fa8\u8b58\uff1b\u53cd\u4e4b\uff0c\u82e5 0 ) ( \uf03c z O d \uff0c\u5247\u6b64\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217 z O \u6703\u88ab\u6b64\u8fa8\u8b58\u5668\u932f\u8aa4\u5730\u8fa8 \u8b58\u3002\u7136\u800c\uff0c\u7576 0 ) ( \uf03d z O d \uff0c\u5247\u8868\u793a\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217 z O \u65e2\u53ef\u80fd\u6703\u88ab\u8fa8\u8b58\u6b63\u78ba\uff0c\u4ea6\u53ef\u80fd\u88ab \u8fa8\u8b58\u932f\u8aa4\uff0c\u5168\u4ef0\u8cf4\u8fa8\u8b58\u5668\u5982\u4f55\u5be6\u4f5c\u3002 \u56e0\u6b64\u5206\u96e2\u908a\u969b ) ( z O d \u53ef\u4ee5\u8996\u70ba\u5728\u76f8\u4f3c\u5ea6\u5b9a\u7fa9\u57df\u4e2d\u7684\u4e00\u9805\u6c7a\u7b56\u8981\u7d20\uff0c\u7528\u4ee5\u6c7a\u5b9a z O \u662f\u5426 \u88ab\u6b63\u78ba\u8fa8\u8b58( 0 ) ( \uf03e z O d )\u6291\u6216\u88ab\u932f\u8aa4\u8fa8\u8b58( 0 ) ( \uf03c z O d )\uff0c\u800c\u6c7a\u5b9a\u908a\u754c\u81ea\u7136\u5c31\u662f\u7576 0 ) ( \uf03d z O d \u9996 \u5148\uff0c\u5fc5\u9808\u5148\u627e\u51fa\u5728\u76ee\u524d\u8fa8\u8b58\u5668\u4e0a\u7684\u6700\u5c0f\u908a\u969b\uff0c\u70ba\u6b64\uff0c\u53ef\u5148\u5b9a\u7fa9\u4e00\u500b\u5b50\u96c6\u5408 S \uff1a \uf07b \uf07d \uf067 \uf0a3 \uf0a3 \uf0ce \uf03d ) ( 0 , | z z z O d R O O S (21) \u5176\u4e2d R \u4ee3\u8868\u6240\u6709\u7684\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217(\u8a13\u7df4\u8a9e\u53e5)\uff1b\uf067 \u70ba\u4e8b\u5148\u5b9a\u7fa9\u7684\u9580\u6abb\u503c\uff0c\u70ba\u4e00\u500b\u5927\u65bc \u96f6\u7684\u6b63\u5be6\u6578\u3002\u6b64\u5b50\u96c6\u5408\u7a31\u70ba\u652f\u6301\u5411\u91cf\u96c6\u5408(Support Vector Set)\uff0c\u5728\u6b64\u96c6\u5408\u88e1\u7684\u8a9e\u97f3\u7279\u5fb5 \u5411\u91cf\u5e8f\u5217 z O (\u6216\u8a13\u7df4\u8a9e\u53e5 z )\u90fd\u662f\u96e2\u6c7a\u5b9a\u908a\u754c\u8f03\u8fd1(\u5c0f\u65bc\uf067 )\u4e14\u53ef\u4ee5\u88ab\u6b63\u78ba\u5730\u8fa8\u8b58\u51fa\u7684\u8a9e\u97f3 \u7279\u5fb5\u5411\u91cf\u5e8f\u5217(\u8a13\u7df4\u8a9e\u53e5) \uff0c\u53ef\u7a31\u4e4b\u70ba\u652f\u6301\u6a23\u672c(Support Tokens) \uff0c\u5982\u5716 1 \u4e2d\u5be6\u5fc3\u5716\u6848\u6240\u793a\u3002 \u6c7a\u5b9a\u76ee\u524d\u7684\u6700\u5c0f\u908a\u969b(\u5b9a\u7fa9\u65bc\u652f\u6301\u5411\u91cf\u96c6\u5408\u4e2d)\u4e4b\u5f8c\uff0c\u6700\u5927\u908a\u969b\u4f30\u6e2c\u6cd5\u5247\u5373\u4ee5\u6700\u5927\u5316\u6b64\u6700 \u5c0f\u908a\u969b\u70ba\u76ee\u6a19\u4f86\u9032\u884c\u8072\u5b78\u6a21\u578b\u8a13\u7df4\uff0c\u5982\uff1a ) ( min max arg z O z O d S \uf0ce \uf04c \uf03d \uf04c (22) \u5176\u4e2d \uf04c \u8207 \uf04c \u5206\u5225\u70ba\u9023\u7e8c\u5bc6\u5ea6\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u8a13\u7df4\u524d\u8207\u8a13\u7df4\u5f8c\u7684\u53c3\u6578\uff0c\u5c07\u5f0f(20)\u5e36\u5165\u5f0f (22)\uff0c\u5247\u6700\u5927\u908a\u969b\u4f30\u6e2c\u7684\u76ee\u6a19\u5373\u70ba\uff1a \uf05b \uf05d ) | ( ) | ( min max arg , , z zi z zR z W W W O W O p W O p zR zi zi \uf02d \uf03d \uf04c \uf0b9 \uf0ce \uf0ce \uf04c W S (23) \u4f46\u5fc5\u9700\u6eff\u8db3\u4ee5\u4e0b\u689d\u4ef6\uff1a 0 ) | ( ) | ( \uf03e \uf02d zi z zR z W O p W O p (24) \u5f0f (23) \u53ef \u4ee5 \u8f49 \u63db \u70ba \u6a19 \u6e96 \u7684 \u7d04 \u675f \u578b \u300e \u6700 \u5c0f \u6700 \u5927 \u300f \u6700 \u4f73 \u5316 \u554f \u984c (Constrained Minimax Optimization Problem)\uff1a \uf05b \uf05d ) | ( ) | ( max min arg , , O z zR z zi z W W W W O p W O p zR zi zi \uf02d \uf03d \uf04c \uf0b9 \uf0ce \uf0ce \uf04c W S (25) \u5716 1. \u6700\u5927\u908a\u969b\u4f30\u6e2c\u6cd5(\u5de6)\u8207\u8abf\u6574\u5f8c\u6700\u5927\u908a\u969b\u4f30\u6e2c\u6cd5(\u53f3)\u793a\u610f\u5716 \u6c7a\u5b9a\u908a\u754c \u908a\u969b \uf072 \uf072 \u6c7a\u5b9a\u908a\u754c \u908a\u969b \uf072 \uf072 \u5247\u7d04\u675f\u689d\u4ef6\u70ba 0 ) | ( ) | ( \uf03c \uf02d zR z zi z W O p W O p (26) \u56e0\u6b64\uff0c\u6700\u5927\u908a\u969b\u4f30\u6e2c\u6cd5\u5247\u7684\u76ee\u6a19\u51fd\u6578\u53ef\u8868\u793a\u70ba \uf05b \uf05d ) | ( ) | ( max ) ( , , O z zR z zi z W W W W O p W O p Q zR zi zi \uf02d \uf03d \uf04c \uf0b9 \uf0ce \uf0ce W S", | |
| "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": "\uf0e5 \uf03d \uf04c \uf02b \uf03d \uf04c Z z z SME O l Z L 1 ) , ( 1 ) ( \uf072 \uf068", | |
| "eq_num": "(28)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u5176\u4e2d \uf04c \u70ba\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u7684\u53c3\u6578\uff1b \uf072 \u70ba\u67d4\u6027\u908a\u969b(Soft Margin)\uff1b\uf068 \u70ba\u4e00\u5e38\u6578\uff0c\u7528\u4f86\u5e73 \u8861\u67d4\u6027\u908a\u969b\u7684\u6700\u5927\u5316\u8207\u7d93\u9a57\u98a8\u96aa\u7684\u6700\u5c0f\u5316\uff0c\u986f\u800c\u6613\u898b\u5730\uff0c\u7576\uf068 \u8d8a\u5c0f\u5247\u8d8a\u5f37\u8abf\u6b64\u5206\u985e\u5668\u7684 \u7d93\u9a57\u98a8\u96aa\uff1b Z \u70ba\u6240\u6709\u7684\u8a13\u7df4\u8a9e\u53e5\u500b\u6578\uff1b ) , ( \uf04c z O l \u70ba\u4e00\u8a9e\u53e5 z O \u7684\u6e1b\u640d\u51fd\u6578\u3002\u56e0\u6b64\uff0c\u67d4\u6027\u908a \u969b\u4f30\u6e2c\u6cd5\u5247\u4fbf\u85c9\u7531\u6700\u5c0f\u5316\u6b64\u76ee\u6a19\u51fd\u6578\u4f86\u964d\u4f4e\u8fa8\u8b58\u5668\u65bc\u6e2c\u8a66\u8a9e\u6599\u4e0a\u7684\u5206\u985e\u932f\u8aa4\u3002\u8207\u50b3\u7d71\u9451 \u5225\u5f0f\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u6700\u5927\u7684\u4e0d\u540c\u8655 \u662f\uff0c\u50b3 \u7d71 \u9451\u5225\u5f0f\u8a13\u7df4\u53ea\u5c08\u6ce8\u65bc\u7d93\u9a57\u98a8\u96aa( \u5373\u70ba \uf0e5 \uf03d \uf04c Z z z O l Z 1 ) , ( 1 )\u7684\u6700\u5c0f\u5316\uff0c\u800c\u5ffd\u7565\u4e00\u822c\u5316\u91cf\u503c(\u4ee5 \uf072 \uf068 \u4f86\u8fd1\u4f3c)\u3002\u7136\u800c\uff0c\u5728\u6700\u5927\u908a\u969b\u4f30\u6e2c \u6cd5\u5247\u537b\u53ea\u628a\u91cd\u9ede\u653e\u5728\u52a0\u5927\u908a\u969b\u4ee5\u964d\u4f4e\u4e00\u822c\u5316\u91cf\u503c\uff0c\u800c\u5ffd\u7565\u4e86\u7d93\u9a57\u98a8\u96aa\u7684\u5f71\u97ff\u3002 \u67d4\u6027\u908a\u969b\u4f30\u6e2c\u6cd5\u5247\u4e2d\u7684\u6e1b\u640d\u51fd\u6578 ) , ( \uf04c O l \u662f\u4e00\u9805\u5f71\u97ff\u6574\u9ad4\u8a13\u7df4\u7684\u91cd\u8981\u5143\u4ef6\uff0c\u5176\u8a2d\u8a08\u5fc5 \u9808\u8ddf\u5176\u76ee\u6a19(\u5373\u589e\u9032\u5206\u985e\u5668\u4e4b\u908a\u969b)\u6709\u4e00\u5b9a\u7684\u95dc\u9023\u6027\u3002\u6545\u5b9a\u7fa9\u70ba\uff1a \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 U O I O d O l z z SME z \uf0ce \uf02d \uf03d \uf04c \uf04c \uf072 ) , ( (29) \u5176\u4e2d \uf028 \uf029 \uf0b7 I \u70ba\u4e00\u500b\u6307\u793a\u51fd\u6578(Indicator Function)\uff1bU \u70ba\u6240\u6709\u8a13\u7df4\u8a9e\u53e5\u7684\u5b50\u96c6\u5408\uff0c\u5176\u5b9a\u7fa9\u70ba \uf028 \uf029 \uf07b \uf07d 0 \uf03e \uf02d \uf03d \uf04c i SME i O d O U \uf072 (30) \u800c \uf028 \uf029 \uf04c , z SME O d \u70ba z O \u7684\u5206\u96e2\u4f30\u91cf(Separation Measure)\uff0c\u7528\u4ee5\u8861\u91cf\u5176\u6b63\u78ba\u8f49\u5beb\u8207\u5c0d\u61c9\u4e4b\u6700\u70ba \u7af6\u722d(Most Competiting)\u5019\u9078\u8a5e\u5e8f\u5217\u5728\u8072\u5b78\u5206\u6578\u4e0a\u7684\u5dee\u8ddd\uff0c\u5b9a\u7fa9\u70ba\uff1a ) ( ) | ( ) | ( log 1 ) ( , z zt t c z zt zR zt z z SME F o I W o p W o p n O d \uf0ce \uf0fa \uf0fa \uf0fb \uf0f9 \uf0ea \uf0ea \uf0eb \uf0e9 \uf03d \uf0e5 \uf04c \uf04c \uf04c (31) \u5176\u4e2d zR W \u70ba z O \u7684\u6b63\u78ba\u8f49\u5beb\u3001 c z W , \u70ba\u6240\u6709\u5019\u9078\u8a5e\u5e8f\u5217\u4e2d\u76f8\u4f3c\u5ea6\u6700\u5927\u7684\u4e00\u689d\uff0c\u5c31\u4ee5\u6700\u5927\u4e8b\u5f8c \u6a5f\u7387\u89e3\u78bc\u539f\u5247\u800c\u8a00\uff0c\u5176\u5c0d\u65bc\u6b63\u78ba\u8f49\u5beb(\u6216\u8a5e\u5e8f\u5217)\u64c1\u6709\u6700\u5927\u7684\u7af6\u722d\u529b\uff1b z F \u662f zR W \u8207 c z W , \u5c0d \u61c9(Aligned)\u5230\u8a13\u7df4\u8a9e\u53e5 z \u5f8c\uff0c\u5169\u8005(\u6307 zR W \u8207 c z W , )\u542b\u6709\u4e0d\u540c\u97f3\u7d20\u985e\u5225(Phone Label)\u7684\u6642\u9593 \u97f3\u6846 t \u6240\u69cb\u6210\u4e4b\u96c6\u5408\uff1b z n \u70ba z F \u7684\u5143\u7d20\u500b\u6578\uff1b ) | ( zR zt W o p \uf04c \u8207 ) | ( , c", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\uf073 2 1 \uf073 \uf073 \uf079 \uf079 1 \uf03d n \uf073 \uf073 \uf073 \uf02d \uf03d \uf0a5 \uf0ae 2 1 \u52d5\u5dee\u751f\u6210\u51fd\u6578 \u4e00\u822c\u5f0f\u6700\u5c0f\u5206\u985e\u932f\u8aa4 (Generalized MCE) \u6700\u5c0f\u5206\u985e\u932f\u8aa4(MCE) 0 2 3 1 \uf03d \uf0a5 \uf0ae \uf03d \uf073 \uf073 \uf073 \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 \uf0e5 \uf0e5 \uf0e5 \uf044 \uf02d \uf04c \uf044 \uf02d \uf04c \uf044 \uf02d \uf04c \uf02d z zi z z zi z z zi z z z z e W O p e W O p e W O p 3 2 1 , , , \uf073 \uf073 \uf073 \uf028 \uf029 \uf028 \uf029 \uf0e5 \uf0e5 \uf04c \uf044 \uf02d \uf04c z zi z z zi z W O p e W O p z , , \uf073 \uf028 \uf029 \uf0e5 \uf044 \uf02d \uf04c \uf03d z zi z z e W O p \uf073 \uf073 \uf079 , \uf028 \uf029 3 2 1 \uf073 \uf073 \uf073 \uf079 \uf079 \uf079 \uf02d \uf028 \uf029 \uf028 \uf029 \uf0e5 \uf0e5 \uf044 \uf04c \uf044 \uf04c \uf044 \uf02d z zi z z n z zi z z z e W O p e W O p \uf073 \uf073 , , \uf028 \uf029 \uf028 \uf029 \uf0e5 \uf044 \uf04c \uf04c z zi z zR z z e W O p W O p \uf073 , , 0 \uf079 \uf079 \uf073 0 \uf03d \uf073 \uf028 \uf029 \uf028 \uf029 \uf0e5 \uf04c \uf04c z zi z zR z W O p W O p , , \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 zR z z zi z zR z W O p W O p W O p , , , \uf04c \uf04c \uf04c \uf02d \uf0e5 \uf028 \uf029 \uf028 \uf029 \uf0e5 \uf0e5 \uf04c \uf04c \uf044 z zi z z z zi z W O p W O p , , \uf028 \uf029 \uf028 \uf029 \uf0e5 \uf0e5 \uf044 \uf02d \uf04c \uf044 \uf02d \uf04c z zi z z zi z z z e W O p e W O p 2 1 , , \uf073 \uf073 \uf028 \uf029 \uf028 \uf029 \uf0e5 \uf0e5 \uf044 \uf04c \uf044 \uf04c \uf044 z zi z z z zi z z z e W O", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\uf028 \uf029 2 ,\uf073 \uf0a5 \uf02d F \"\u589e\u9032\u5f0f\u6700\u5927\u4ea4\u4e92\u8cc7\u8a0a\" \uf028 \uf029 \uf028 \uf029 \uf073 \uf073 \uf073 \uf073 \uf073 \uf073 d f F MPE \uf0f2 \uf03d \uf03d 2 1 2 1 , ) \u4e00\u822c\u5316\u6700\u5927\u4ea4 ( \u6574\u5408\u908a\u969b\u6b0a\u91cd\u932f\u8aa4 \u8a0a \u8cc7 \u4e92 \uf028 \uf029 0 , \uf0a5 \uf02d F \u6b0a\u91cd\u932f\u8aa4 (\u5982\u6700\u5c0f\u97f3\u7d20\u932f\u8aa4) \"\u6a19\u6e96\u6700\u5927\u4ea4\u4e92\u8cc7\u8a0a\" \"\u589e\u9032\u5f0f\u6700\u5c0f\u97f3\u7d20\u932f\u8aa4\" \"\u6a19\u6e96\u6700\u5c0f\u97f3\u7d20\u932f\u8aa4\" \uf073 \uf028 \uf029 \uf073 f 0 \uf03d \uf073 \uf028 \uf029 2 \uf073 MPE f 0 2 \uf03e \uf03d \uf073 \uf073 \uf028 \uf029 0 MPE f \uf0a5 \uf02d \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 MPE z W W W", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Minimum classification error rate methods for speech recognition", | |
| "authors": [ | |
| { | |
| "first": "B", | |
| "middle": [ | |
| "H" | |
| ], | |
| "last": "Juang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Chou", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [ | |
| "H" | |
| ], | |
| "last": "Lee", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "IEEE Trans. Speech and Audio Processing", | |
| "volume": "5", | |
| "issue": "3", | |
| "pages": "257--265", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "B. H. Juang, W. Chou, and C. H. Lee, \"Minimum classification error rate methods for speech recognition,\" IEEE Trans. Speech and Audio Processing, Vol. 5, no. 3, pp. 257-265, 1997.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Hidden Markov models, maximum mutual information estimation, and the speech recognition problem", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Normandin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1991, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Normandin, Hidden Markov models, maximum mutual information estimation, and the speech recognition problem, Ph.D. Dissertation, McGill University, Montreal, 1991.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Discriminative training for large vocabulary speech recognition", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Povey, Discriminative training for large vocabulary speech recognition. Ph.D Dissertation, Peterhouse, University of Cambridge, July 2004.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Large margin hidden Markov models for speech recognition", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Jiang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C.-J", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Speech and Language Processing", | |
| "volume": "14", | |
| "issue": "", | |
| "pages": "1584--1595", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. Jiang, X. Li and C.-J. Liu, \"Large margin hidden Markov models for speech recognition,'' IEEE Trans. on Audio, Speech and Language Processing, pp. 1584-1595, Vol. 14, No. 5, September 2006.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Soft margin estimation for automatic speech recognition", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Ph.D. Dissertation, Electrical and Computer Engineering", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J. Li, Soft margin estimation for automatic speech recognition. Ph.D. Dissertation, Electrical and Computer Engineering, Georgia Institute of Technology, July 2008.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Large margin training of acoustic models for speech recognition", | |
| "authors": [ | |
| { | |
| "first": "Fei", | |
| "middle": [], | |
| "last": "Sha", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Fei Sha, Large margin training of acoustic models for speech recognition. Ph.D. Dissertation, University of Pennsylvania. 2007.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Modified MMI/MPE: A Direct Evaluation of the Margin in Speech Recognition", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Heigold", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Deselaers", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Schl\u00fcter", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Ney", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Proc. ICML", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "384--391", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G. Heigold, T. Deselaers, R. Schl\u00fcter, and H. Ney. \"Modified MMI/MPE: A Direct Evaluation of the Margin in Speech Recognition,\" in Proc. ICML, pp. 384-391, 2008.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Large-margin minimum classification error training: A theoretical risk minimization perspective", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Yu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "He", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Acero", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Computer Speech and Language", | |
| "volume": "22", | |
| "issue": "4", | |
| "pages": "415--429", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Yu, L. Deng, X. He, and A. Acero, \"Large-margin minimum classification error training: A theoretical risk minimization perspective,\" Computer Speech and Language, Vol. 22, No. 4 pp. 415-429, October 2008.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Training data selection for improving discriminative training of acoustic models", | |
| "authors": [ | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S. -H", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F.-H", | |
| "middle": [], | |
| "last": "Chu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Pattern Recognition Letters", | |
| "volume": "30", | |
| "issue": "13", | |
| "pages": "1228--1235", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "B. Chen, S. -H. Liu, and F.-H. Chu, \"Training data selection for improving discriminative training of acoustic models,\" Pattern Recognition Letters, Vol. 30, No. 13, pp. 1228-1235, October 2009.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Penalty function maximization for large margin HMM training", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Saon", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "920--923", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G. Saon and D. Povey, \"Penalty function maximization for large margin HMM training,\" in Proc. Interspeech, pp. 920-923, 2008.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Boosted MMI for model and feature space discriminative training", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Kanevsky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Kingsbury", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Ramabhadran", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Saon", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Visweswariah", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "4057--4060", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Povey, D. Kanevsky, B. Kingsbury, B. Ramabhadran, G. Saon, and K. Visweswariah, \"Boosted MMI for model and feature space discriminative training,\" in Proc. ICASSP, pp. 4057-4060, 2008.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Margin-space integration of MPE loss via differencing of MMI functionals for generalized error-weighted discriminative training", | |
| "authors": [ | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Mcdermott", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Watanabe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Nakamura", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "224--227", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "E. McDermott, S. Watanabe, and A. Nakamura: \"Margin-space integration of MPE loss via differencing of MMI functionals for generalized error-weighted discriminative training,\" in Proc. Interspeech, pp. 224-227, 2009.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Discriminative Learning in Sequential Pattern Recognition ---A Unifying Review for Optimization-Oriented Speech Recognition", | |
| "authors": [ | |
| { | |
| "first": "Xiaodong", | |
| "middle": [], | |
| "last": "He", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Li", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Chou", | |
| "middle": [], | |
| "last": "Wu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "IEEE Signal Processing Magazine", | |
| "volume": "25", | |
| "issue": "5", | |
| "pages": "14--36", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Xiaodong He, Li Deng, and Chou Wu, \"Discriminative Learning in Sequential Pattern Recognition ---A Unifying Review for Optimization-Oriented Speech Recognition,\" in IEEE Signal Processing Magazine, vol. 25, No. 5, pp. 14-36, September, 2008.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "A unified view for discriminative objective functions based on negative exponential of difference measure between strings", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Nakamura", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Mcdermott", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Watanabe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Katagiri", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1633--1636", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Nakamura, E. McDermott, S. Watanabe, and S. Katagiri, \"A unified view for discriminative objective functions based on negative exponential of difference measure between strings,\" in Proc. ICASSP ,pp. 1633-1636 ,2009.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Discriminative training for automatic speech recognition: A survey", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Jiang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "", | |
| "volume": "24", | |
| "issue": "", | |
| "pages": "589--608", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. Jiang, \"Discriminative training for automatic speech recognition: A survey,\" Computer and Speech, Language, pp. 589-608, Vol. 24, No. 4, October 2010.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Comparison of discriminative training criteria and optimization methods for speech recognition", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Schl\u00fcter", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Macherey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "M\u00fcller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Ney", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2001, | |
| "venue": "Speech Communication", | |
| "volume": "34", | |
| "issue": "", | |
| "pages": "287--310", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R. Schl\u00fcter, W. Macherey, B. M\u00fcller, and H. Ney, \"Comparison of discriminative training criteria and optimization methods for speech recognition,\" Speech Communication, Vol. 34, pp. 287-310, May 2001.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Incorporating training errors for large margin HMMs under semi-definite programming framework", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Jiang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "629--632", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. Jiang and X. Li, \"Incorporating training errors for large margin HMMs under semi-definite programming framework,\" in Proc. ICASSP, pp. 629-632, 2007.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "The nature of statistical learning theory", | |
| "authors": [ | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Vapnik", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1995, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "V. Vapnik, The nature of statistical learning theory. New York: Springer-Verlag, 1995.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "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": "Proc. Interspeech", | |
| "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 Proc. Interspeech, pp. 2422-2425, 2006.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Lightly Supervised and Data-Driven Approaches to Mandarin Broadcast News Transcription", | |
| "authors": [ | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "W" | |
| ], | |
| "last": "Kuo", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [ | |
| "H" | |
| ], | |
| "last": "Tsai", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Proc. ICASSP", | |
| "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 Proc. ICASSP, p777-780, 2004.", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "Investigation of silicon-auditory models and generalizaion of linar discriminant analysis for improved speech recognition", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Kumar", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "N. Kumar, Investigation of silicon-auditory models and generalizaion of linar discriminant analysis for improved speech recognition, Ph.D. Thesis, John Hopkins University, Baltimore, 1997.", | |
| "links": null | |
| }, | |
| "BIBREF21": { | |
| "ref_id": "b21", | |
| "title": "Maximum likelihood modeling with Gaussian distributions", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Gopinath", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1998, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "661--664", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R. A. Gopinath, \"Maximum likelihood modeling with Gaussian distributions,\" in Proc. ICASSP, pp. 661-664, 1998.", | |
| "links": null | |
| }, | |
| "BIBREF22": { | |
| "ref_id": "b22", | |
| "title": "LDC: Linguistic Data Consortium", | |
| "authors": [], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "LDC: Linguistic Data Consortium, http://www.ldc.upenn.edu.", | |
| "links": null | |
| }, | |
| "BIBREF23": { | |
| "ref_id": "b23", | |
| "title": "Estimation of probabilities from sparse data for other language component of a speech recognizer", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [ | |
| "M" | |
| ], | |
| "last": "Katz", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1987, | |
| "venue": "IEEE Trans. Acoustics, Speech and Signal Processing", | |
| "volume": "35", | |
| "issue": "3", | |
| "pages": "400--401", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. M. Katz, \"Estimation of probabilities from sparse data for other language component of a speech recognizer,\" IEEE Trans. Acoustics, Speech and Signal Processing, Vol. 35, No.3, pp. 400-401, 1987.", | |
| "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. http://www.speech.sri.com/projects/srilm/.", | |
| "links": null | |
| }, | |
| "BIBREF25": { | |
| "ref_id": "b25", | |
| "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 | |
| }, | |
| "BIBREF26": { | |
| "ref_id": "b26", | |
| "title": "A word graph algorithm for large vocabulary continuous speech recognition", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Ortmanns", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Ney", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Aubert", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "Computer Speech and Language", | |
| "volume": "11", | |
| "issue": "", | |
| "pages": "11--72", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Ortmanns, H. Ney and X Aubert, \"A word graph algorithm for large vocabulary continuous speech recognition,\" Computer Speech and Language, Vol. 11, pp. 11-72, 1997.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "text": "\u6458\u8981 \u9451\u5225\u5f0f\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u5728\u8fd1\u4ee3\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58(Automatic Speech Recognition, ASR)\u4e2d\u626e\u6f14", | |
| "num": null, | |
| "html": null, | |
| "content": "<table><tr><td>\u4e00\u3001\u7dd2\u8ad6 \u7684\u5be6\u9a57\u6bd4\u8f03\u8207\u8a0e\u8ad6\uff1b\u7b2c\u4e94\u7ae0\u70ba\u7d50\u8ad6\u53ca\u672a\u4f86\u7814\u7a76\u5c55\u671b\u3002</td></tr><tr><td>\u4ee5\u6700\u5927\u5316\u76f8\u4f3c\u5ea6\u4f30\u6e2c(Maximum Likelihood Estimation, MLE)\u4f86\u8a13\u7df4\u8072\u5b78\u6a21\u578b(Acoustic \u4e8c\u3001 \u9451\u5225\u5f0f\u8a13\u7df4\u6cd5\u5247\u53ca\u5176\u4e00\u81f4\u6027</td></tr><tr><td>Models)\uff0c\u5728\u904e\u53bb\u6578\u5341\u5e74\u5ee3\u70ba\u8a9e\u97f3\u8fa8\u8b58\u9818\u57df\u6240\u63a1\u7528\uff0c\u5b83\u4e3b\u8981\u662f\u8003\u91cf\u5982\u4f55\u80fd\u5f9e\u8a13\u7df4\u8a9e\u6599\u4e2d \u672c\u7ae0\u7bc0\u9996\u5148\u5c07\u5f9e\u8a9e\u97f3\u8fa8\u8b58\u539f\u5247\u70ba\u51fa\u767c\u9ede\uff0c\u95e1\u8ff0\u5927\u591a\u6578\u65b9\u6cd5\u6240\u4f9d\u5faa\u6e96\u5247\u70ba\u4f55\uff0c\u64da\u6b64\u5c07\u8f38\u5165 \u7372\u5f97\u7d71\u8a08\u8cc7\u8a0a\uff0c\u4ee5\u8b93\u8072\u5b78\u6a21\u578b\u53ef\u4ee5\u4ee3\u8868\u8a13\u7df4\u8a9e\u6599(\u4e5f\u5c31\u8b93\u8072\u5b78\u6a21\u578b\u7522\u751f\u5c0d\u61c9\u7684\u8a13\u7df4\u8a9e\u6599 \u7684\u8a9e\u97f3\u8a0a\u865f(\u6216\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217)\u8fa8\u8b58\u6210\u5c0d\u61c9\u7684\u81ea\u52d5\u8f49\u5beb\u505a\u70ba\u8f38\u51fa\u3002\u56e0\u6b64\uff0c\u5728\u8072\u5b78\u6a21\u578b \u4e4b\u76f8\u4f3c\u5ea6\u6700\u5927)\u3002\u4f46\u6b64\u7a2e\u8a13\u7df4\u65b9\u6cd5\u4e26\u6c92\u6709\u8003\u616e\u8a9e\u97f3\u8fa8\u8b58\u6642\u8072\u5b78\u6a21\u578b\u5f7c\u6b64\u9593\u7684\u95dc\u4fc2\uff0c\u5728\u8abf \u8a13\u7df4\u6642\uff0c\u6211\u5011\u81ea\u7136\u5730\u5c31\u53ef\u4ee5\u5229\u7528\u6b64\u539f\u5247\u505a\u70ba\u9451\u5225\u5f0f\u8a13\u7df4\u7576\u4e2d\u5b9a\u7fa9\u76ee\u6a19\u51fd\u6578\u7684\u4e3b\u8981\u6839\u64da\u3002 \u6574\u8072\u5b78\u6a21\u578b\u53c3\u6578\u4e4b\u5f8c\uff0c\u96d6\u53ef\u4f7f\u76f8\u95dc\u7684\u8a9e\u97f3\u7279\u5fb5\u843d\u5728\u67d0\u4e00\u500b\u8072\u5b78\u6a21\u578b\u7684\u76f8\u4f3c\u5ea6\u8b8a\u5927\uff0c\u537b\u4e5f \u4e00\u822c\u4f86\u8aaa\uff0c\u5728\u57f7\u884c\u9451\u5225\u5f0f\u8a13\u7df4\u524d\uff0c\u6211\u5011\u901a\u5e38\u5148\u4f7f\u7528\u6700\u5927\u5316\u76f8\u4f3c\u5ea6\u4f30\u6e2c(MLE)\u505a\u70ba\u57fa \u53ef\u80fd\u540c\u6642\u8b93\u975e\u76f8\u95dc\u7684\u8a9e\u97f3\u7279\u5fb5\u843d\u5728\u6b64\u8072\u5b78\u6a21\u578b\u7684\u76f8\u4f3c\u5ea6\u66f4\u5927\uff0c\u9020\u6210\u8fa8\u8b58\u4e0a\u7684\u6df7\u6dc6\u3002\u56e0\u6b64\uff0c \u790e\u8072\u5b78\u6a21\u578b\u7684\u8a13\u7df4\u65b9\u6cd5\uff0c\u5176\u76ee\u7684\u662f\u6700\u5927\u5316\u8072\u5b78\u6a21\u578b\u7522\u751f\u5c0d\u61c9\u7684\u8a13\u7df4\u8a9e\u6599\u4e4b\u76f8\u4f3c\u5ea6\u3002\u723e\u5f8c \u8fd1\u4f86\u6709\u4e0d\u5c11\u7814\u7a76\u91dd\u5c0d\u6b64\u9805\u7f3a\u9ede\uff0c\u63d0\u51fa\u9451\u5225\u5f0f\u8a13\u7df4(Discriminative Training)\u6cd5\u5247\u4f86\u52a0\u4ee5\u6539 \u5728\u9451\u5225\u5f0f\u8a13\u7df4\u4e0a\uff0c\u900f\u904e\u4e0d\u540c\u7684\u601d\u7dad\u5c0d\u65bc\u8a13\u7df4\u8a9e\u53e5\u7684\u8a9e\u97f3\u8fa8\u8b58\u6df7\u6dc6\u8cc7\u8a0a\u7522\u751f\u9700\u6c42\uff0c\u4e5f\u5c31\u662f \u9032 \u3002 \u9451 \u5225 \u5f0f \u8a13 \u7df4 \u4e0d \u50c5 \u662f \u8003 \u616e \u4e86 \u8a13 \u7df4 \u8a9e \u53e5 \u7684 \u6b63 \u78ba ( \u53c3\u8003) \u8f49\u5beb(Correct or Reference \u4ee5\u6b63\u78ba\u8f49\u5beb\u8207\u8a9e\u97f3\u8fa8\u8b58\u5668\u6240\u7522\u751f\u7684\u8a5e\u5716(\u5167\u542b\u8a31\u591a\u5019\u9078\u8a5e\u5e8f\u5217)\u5f62\u6210\u6240\u8b02\u7684\u5047\u8a2d\u7a7a\u9593\uff1b\u4ee5 Transcription)\uff0c\u540c\u6642\u4e5f\u8003\u616e\u7531\u8a9e\u97f3\u8fa8\u8b58\u5668\u5c0d\u8a9e\u53e5\u9032\u884c\u8fa8\u8b58\u5f8c\u7522\u751f\u7684\u3001\u8207\u6b63\u78ba\u8f49\u5beb\u4e0d\u540c\u7684 \u6642\u4e0b\u6700\u6d41\u884c\u7684\u4e09\u7a2e\u6700\u5177\u4ee3\u8868\u6027\u4e4b\u9451\u5225\u5f0f\u8a13\u7df4\u65b9\u6cd5\u70ba\u4f8b\uff0c\u8aaa\u660e\u5176\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u76ee\u6a19\u51fd\u6578\u7684 \u5019\u9078\u8a5e\u5e8f\u5217\u5047\u8a2d(Candidate Word Sequence Hypotheses)\uff0c\u4ee5\u589e\u9032\u8a13\u7df4\u5f8c\u8072\u5b78\u6a21\u578b\u7684\u9451\u5225 \u6700\u7d42\u76ee\u7684\u90fd\u662f\u5728\u65bc\u4f7f\u7528\u9451\u5225\u5f0f\u51fd\u6578\u63cf\u8ff0\u6b63\u78ba\u8f49\u5beb\u8207\u5176\u5b83\u5728\u8a5e\u5716\u4e0a\u7684\u5019\u9078\u8a5e\u5e8f\u5217\u5728\u5047\u8a2d \u6027\u3002 \u7a7a\u9593\u6240\u5f62\u6210\u7684\u95dc\u4fc2\u3002\u6211\u5011\u53ef\u4ee5\u6b78\u7d0d\u51fa\u9019\u5176\u4e2d\u7684\u5dee\u5225\u4e3b\u8981\u5728\u65bc\u4e0d\u540c\u5c64\u7d1a(\u5982\u8a9e\u53e5\u5c64\u7d1a\u3001\u97f3 \u9577\u4e45\u4ee5\u4f86\uff0c\u9451\u5225\u5f0f\u8a13\u7df4\u65b9\u6cd5\u70ba\u8a9e\u97f3\u8fa8\u8b58\u4e2d\u8072\u5b78\u6a21\u578b\u8fa8\u8b58\u80fd\u529b\u63d0\u6607\u4e2d\u7684\u91cd\u8981\u4e00\u74b0\uff0c\u76f8 \u7d20\u5c64\u7d1a)\u6216\u4e0d\u540c\u7d30\u7dfb\u7a0b\u5ea6\u7684\u8a13\u7df4\u8cc7\u6599\u9078\u53d6\u3002\u7f8e\u570b\u5fae\u8edf\u516c\u53f8(Microsoft)\u7684\u5b78\u8005\u91dd\u5c0d\u9019\u9ede\u5728\u8fd1 \u95dc\u7814\u7a76\u8207\u5ef6\u4f38\u65cf\u7e41\u4e0d\u53ca\u5099\u8f09\uff0c\u4ee5\u4e0b\u5217\u4e09\u7a2e\u65b9\u6cd5\u8f03\u5177\u4ee3\u8868\u6027\uff1a (\u4e00)\u6700\u5c0f\u5316\u5206\u985e\u932f\u8aa4 \u671f\u63d0\u51fa\u4e86\u5b8c\u6574\u7684\u8b49\u660e[13]\uff1b\u540c\u6a23\u5730\uff0c\u65e5\u672c\u96fb\u5831\u96fb\u4fe1\u516c\u53f8(NTT)[14]\u4e5f\u62b1\u6301\u76f8\u540c\u770b\u6cd5\u4e26\u4e14\u5f15 (Minimum Classification Error, MCE)[1]\u8003\u616e\u5230\u8a13\u7df4\u8a9e\u53e5\u7684\u6b63\u78ba\u8f49\u5beb\u8207\u4e0d\u6b63\u78ba\u8f49\u5beb\u5728\u5047 \u5165\u589e\u9032\u6b0a\u91cd\uff0c\u5176\u6240\u96b1\u542b\u7684\u5373\u662f\u908a\u969b\u8cc7\u8a0a\u6982\u5ff5\uff0c\u4e26\u4e14\u8aaa\u660e\u4e86\u57fa\u65bc\u76f8\u540c\u578b\u5f0f\u7684\u4e0d\u540c\u76ee\u6a19\u51fd\u6578 \u8a2d\u7a7a\u9593\u4e0a\u7684\u5206\u96e2\u7a0b\u5ea6\uff1b(\u4e8c)\u6700\u5927\u4ea4\u4e92\u8cc7\u8a0a\u6cd5\u5247(Maximum Mutual Information, MMI)[2]\u4ee5 \u7686\u53ef\u4f7f\u7528\u540c\u6a23\u7684\u6700\u4f73\u5316\u65b9\u6cd5\u4f86\u6c42\u5f97\u6a21\u578b\u53c3\u6578\uff1b\u5176\u5b83\u5982[15][16]\u4e5f\u90fd\u5c0d\u76ee\u6a19\u51fd\u6578\u8207\u6700\u4f73\u5316 \u6700\u5927\u5316\u8a13\u7df4\u8a9e\u53e5\u8207\u5176\u5c0d\u61c9\u8a5e\u6b63\u78ba\u8f49\u5beb\u7684\u4ea4\u4e92\u8cc7\u8a0a\uff1b(\u4e09)\u6700\u5c0f\u5316\u97f3\u7d20\u932f\u8aa4(Minimum Phone Error, MPE)[3]\u76ee\u7684\u70ba\u6700\u5c0f\u5316\u8a9e\u97f3\u8fa8\u8b58\u5668\u8f38\u51fa(\u4ea6\u5373\u6240\u6709\u5019\u9078\u8a5e\u5e8f\u5217)\u7684\u97f3\u7d20\u671f\u671b\u932f\u8aa4\u7387\u3002 \u65b9\u6cd5\u505a\u6df1\u5165\u7684\u89e3\u6790\u4f86\u63a2\u8a0e\u9451\u5225\u5f0f\u8a13\u7df4\u65b9\u6cd5\u7684\u4e00\u81f4\u6027\u3002</td></tr><tr><td>\u9019\u4e09\u7a2e\u4ee5\u4e0d\u540c\u601d\u7dad\u51fa\u767c\u7684\u65b9\u6cd5\uff0c\u5b83\u5011\u80cc\u5f8c\u6db5\u7fa9\u7686\u662f\u63cf\u8ff0\u8a13\u7df4\u8a9e\u53e5\u7684\u6b63\u78ba\u8f49\u5beb\u8207\u5176\u5b83\u5019\u9078 (\u4e00)\u8a9e\u97f3\u89e3\u78bc\u539f\u5247\u8207\u6700\u5927\u5316\u76f8\u4f3c\u5ea6\u6cd5\u5247\uff1a</td></tr><tr><td>\u91cd\u8981\u7684\u89d2\u8272\u3002\u5728\u8a31\u591a\u57fa\u65bc\u4e0d\u540c\u601d\u7dad\u4e14\u80fd\u6709\u6548\u5730\u63d0\u6607\u8fa8\u8b58\u7387\u7684\u9451\u5225\u5f0f\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u65b9\u6cd5\u9678 \u7e8c\u88ab\u63d0\u51fa\u5f8c\uff0c\u5c0d\u65bc\u8a13\u7df4\u65b9\u6cd5\u7684\u76f8\u95dc\u63a8\u5ee3\u8207\u6539\u9032\u4fbf\u5982\u96e8\u5f8c\u6625\u7b4d\u822c\u5730\u8208\u8d77\uff1b\u800c\u9019\u4e9b\u65b9\u6cd5\u5728\u672c \u8cea\u4e0a\uff0c\u7686\u662f\u5728\u63cf\u8ff0\u8a13\u7df4\u8a9e\u53e5\u8207\u8a9e\u97f3\u8fa8\u8b58\u5668\u6240\u7522\u751f\u5c0d\u61c9\u8a5e\u5716(Word Graph)\u4e4b\u9593\u7684\u95dc\u4fc2\u3002\u672c \u8a5e\u5e8f\u5217\u4e4b\u9593\u7684\u95dc\u4fc2\uff0c\u65e5\u5f8c\u8a31\u591a\u88ab\u63d0\u51fa\u7684\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u65b9\u6cd5\u4e5f\u90fd\u662f\u67b6\u69cb\u65bc\u9019\u6a23\u7684\u95dc\u4fc2\u4e0a\u3002 \u5728\u8a9e\u97f3\u8fa8\u8b58\u7684\u89e3\u78bc\u539f\u5247\u4e0a\uff0c\u5927\u591a\u6578\u7684\u4f5c\u6cd5\u901a\u5e38\u662f\u63a1\u7528\u8c9d\u5f0f\u6c7a\u7b56\u5b9a\u7406(Bayesian Decision \u901a\u5e38\u5728\u57f7\u884c\u9451\u5225\u5f0f\u8a13\u7df4\u6642\uff0c\u6bcf\u4e00\u53e5\u8a13\u7df4\u8a9e\u53e5\u6240\u5c0d\u61c9\u7684\u8a5e\u5716(Word Lattice)\u626e\u6f14\u7684\u89d2\u8272 \u4e0d\u50c5\u50c5\u70ba\u6240\u6709\u5019\u9078\u8a5e\u5e8f\u5217(\u53ef\u80fd\u5305\u62ec\u6b63\u78ba\u8f49\u5beb)\u4e4b\u5047\u8a2d\u7a7a\u9593(Hypothesis Space)\uff0c\u66f4\u662f\u63d0\u4f9b Theorem)\uff1a\u5373\u662f\u5728\u7d66\u5b9a\u4e00\u53e5\u8a9e\u53e5\u7684\u8a9e\u97f3\u7279\u5fb5\u5411\u91cf\u5e8f\u5217 \uf07b \uf07d t o o O ,...,</td></tr><tr><td>\u8ad6\u6587\u9996\u5148\u5c07\u7d71\u6574\u8207\u6b78\u7d0d\u8fd1\u5e74\u4f86\u6240\u767c\u5c55\u7684\u591a\u7a2e\u9451\u5225\u5f0f\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u65b9\u6cd5\uff0c\u4e26\u4ee5\u4e09\u7a2e\u6700\u5177\u4ee3 \u8072\u5b78\u6a21\u578b\u53c3\u6578\u4f30\u6e2c\u904e\u7a0b\u4e2d\u9451\u5225\u8cc7\u8a0a\u7684\u91cd\u8981\u4f86\u6e90\u4e4b\u4f9d\u64da\u3002\u5c0d\u65bc\u9451\u5225\u8cc7\u8a0a\u7684\u9700\u6c42\u56e0\u800c\u6709\u4e86\u6240</td></tr><tr><td>\u8868\u6027\u9451\u5225\u5f0f\u8a13\u7df4\u65b9\u6cd5\uff1a\u6700\u5c0f\u5316\u5206\u985e\u932f\u8aa4(Minimum Classification Error, MCE)\u3001\u6700\u5927\u5316\u4ea4 \u8b02\u7684\u8cc7\u6599\u9078\u53d6(Data Selection)\u6982\u5ff5\uff1a\u5728\u6a5f\u5668\u5b78\u7fd2\u4e2d\uff0c\u652f\u6490\u5411\u91cf\u6a5f\u908a\u969b\u8cc7\u8a0a\u6982\u5ff5\u5728\u5206\u985e\u554f</td></tr><tr><td>\u4e92\u8cc7\u8a0a(Maximum Mutual Information, MMI)\u3001\u6700\u5c0f\u5316\u97f3\u7d20\u932f\u8aa4(Minimum Phone Error, \u984c\u4e0a\u6709\u8457\u975e\u5e38\u6210\u529f\u7684\u6210\u6548\uff0c\u908a\u969b\u8cc7\u8a0a\u6240\u95e1\u8ff0\u7684\u7406\u5ff5\u662f\u6c7a\u7b56\u908a\u969b\u8207\u8a13\u7df4\u8cc7\u6599\u5206\u5e03\u5c0d\u5206\u985e\u554f</td></tr><tr><td>MPE)\u70ba\u7bc4\u4f8b\uff0c\u900f\u904e\u6709\u7cfb\u7d71\u5730\u8f49\u63db\u8207\u5316\u89e3\u65b9\u7a0b\u5f0f\uff0c\u5f97\u5230\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u6e96\u5247\u7684\u5171\u901a\u8868\u793a\u51fd \u984c\u6240\u7522\u751f\u5f71\u97ff\uff1b\u76f8\u540c\u7684\u6982\u5ff5\u88ab\u63a8\u5ee3\u5230\u5728\u8a9e\u97f3\u8fa8\u8b58\u7684\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u4e2d\u4f7f\u7528[4][5][6][8]\uff0c\u4e26</td></tr><tr><td>\u6578\u578b\u614b\u3002\u6211\u5011\u53ef\u4ee5\u767c\u73fe\u5230\uff0c\u5c0d\u65bc\u4e0a\u8ff0\u9451\u5225\u5f0f\u8a13\u7df4\u65b9\u6cd5\uff0c\u6b64\u5171\u901a\u8868\u793a\u51fd\u6578\u80cc\u5f8c\u7269\u7406\u610f\u7fa9\u4e4b \u4e14\u8b49\u660e\u4e86\u5176\u6548\u7528[7]\uff0c\u800c\u76f8\u540c\u6709\u6548\u679c\u7684\u8cc7\u6599\u9078\u53d6\u65b9\u6cd5\u5982[9][10][11][12]\u540c\u6a23\u5728\u5c0d\u65bc\u8a9e\u97f3\u8fa8</td></tr><tr><td>\u5dee\u5225\u4e43\u662f\u5728\u65bc\u6b32\u89c0\u5bdf\u8a13\u7df4\u8a9e\u6599\u4e0d\u540c\u5c64\u7d1a\u7684\u9451\u5225\u8cc7\u8a0a\uff0c\u5982\u97f3\u7d20(Phone) \u3001\u8a9e\u53e5(Utterance)\u7b49\uff0c \u8b58\u7387\u7684\u63d0\u6607\u6709\u8457\u6b63\u9762\u5f71\u97ff\u3002\u6b64\u5916\uff0c\u57fa\u65bc\u5c0d\u8a13\u7df4\u8a9e\u6599\u7684\u5206\u5e03\u4ee5\u53ca\u9451\u5225\u5f0f\u8a13\u7df4\u6240\u9700\u7684\u6b0a\u91cd\u7684</td></tr><tr><td>\u4ee5\u53ca\u5171\u901a\u8868\u793a\u51fd\u6578\u4e4b\u53c3\u6578\u8a2d\u5b9a\u3002\u5176\u6b21\uff0c\u672c\u8ad6\u6587\u91dd\u5c0d\u8a9e\u97f3\u8fa8\u8b58\u7d50\u679c\u6240\u5f62\u6210\u7684\u5047\u8a2d\u7a7a\u9593\u4e0a\u6240 \u5206\u6790\uff0c\u543e\u4eba\u7d50\u540840 delta \u8abf\u6574\u8b93\u5169\u985e\u908a\u969b\u8cc7\u8a0a\u65b9\u6cd5\u65bc\u9451\u5225\u5f0f\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u4e0a\uff1a\u4e00\u65b9\u9762\uff0c</td></tr><tr><td>\u89c0\u5bdf\u5230\u932f\u8aa4(\u6216\u6b63\u78ba)\u7387\u7684\u4e0d\u540c\u7d30\u7dfb\u5c64\u5ea6\uff0c\u5728\u6a21\u578b\u8a13\u7df4\u6642\u5f15\u5165\u4e86\u6a5f\u5668\u5b78\u7fd2\u9818\u57df\u4e2d\u7684\u908a\u969b\u6982 \u9664\u4e86\u5728\u8fa8\u8b58\u7387\u4e0a\u53ef\u4ee5\u6709\u6548\u63d0\u6607\u4e4b\u5916\uff1b\u53e6\u4e00\u65b9\u9762\uff0c\u5c0d\u65bc\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u6642\u56e0\u589e\u9032\u6b0a\u91cd\u56e0\u5b50\u6240</td></tr><tr><td>\u5ff5\uff1b\u5176\u80cc\u5f8c\u7684\u7269\u7406\u610f\u7fa9\uff0c\u4e8b\u5be6\u4e0a\u5c31\u662f\u5f9e\u4e0d\u540c\u5c64\u7d1a\u7684\u8a13\u7df4\u8a9e\u6599\u4e2d\u9078\u53d6\u9069\u5408\u7684\u8cc7\u8a0a\u4f9b\u8072\u5b78\u6a21 \u5e36\u4f86\u904e\u5ea6\u8a13\u7df4(Overfitting)\u7684\u554f\u984c\uff0c\u4ea6\u53ef\u4ee5\u6709\u6548\u7684\u7372\u5f97\u89e3\u6c7a\u3002</td></tr><tr><td>\u578b\u8a13\u7df4\u6240\u4f7f\u7528\u3002\u672c\u8ad6\u6587\u7684\u76ee\u7684\u5728\u65bc\u5206\u6790\u8fd1\u4ee3\u5c0d\u65bc\u4ee5\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u70ba\u8072\u5b78\u6a21\u578b\u4e4b\u6a21\u578b \u672c\u8ad6\u6587\u5c0d\u65bc\u8fd1\u5e74\u9451\u5225\u5f0f\u8a13\u7df4\u53ca\u5176\u5ef6\u4f38\u65b9\u6cd5\uff0c\u505a\u51fa\u7d71\u6574\u6b78\u7d0d\u4e4b\u7814\u7a76\uff0c\u4e26\u5f9e\u4ee3\u8868\u6027\u7684\u8cc7</td></tr><tr><td>\u8a13\u7df4\u65b9\u6cd5\u8207\u908a\u969b\u6982\u5ff5\u5728\u6f14\u9032\u4e0a\u7684\u4e00\u81f4\u6027\uff1b\u5f9e\u7433\u746f\u6eff\u76ee\u7684\u8a13\u7df4\u65b9\u6cd5\u4e4b\u4e2d\uff0c\u95e1\u8ff0\u8fd1\u5e74\u4f86\u9451\u5225 \u6599\u9078\u53d6\u65b9\u6cd5\u4e2d\u7d50\u5408\u5176\u512a\u9ede\uff0c\u7372\u5f97\u540c\u6642\u517c\u5177\u5404\u9805\u512a\u9ede\u7684\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u76ee\u6a19\u51fd\u6578\u3002\u543e\u4eba\u5c07\u5404</td></tr><tr><td>\u5f0f\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u767c\u5c55\u6f14\u9032\u4e4b\u4e2d\u5fc3\u601d\u60f3\u3002\u6700\u5f8c\uff0c\u6211\u5011\u5be6\u4f5c\u65bc\u4e2d\u6587\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\uff0c \u985e\u8a13\u7df4\u6e96\u5247\u8207\u672c\u8ad6\u6587\u6240\u63d0\u53ca\u6539\u5584\u65b9\u6cd5\u5be6\u4f5c\u65bc\u4e2d\u6587\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58\u5be6\u9a57\u4e2d\uff0c\u540c\u6642\u5f9e\u5be6</td></tr><tr><td>\u9a57\u8b49\u4e86\u591a\u7a2e\u9451\u5225\u5f0f\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u65b9\u6cd5\u4ee5\u53ca\u6211\u5011\u6240\u63d0\u51fa\u65b9\u6cd5\u4e4b\u6548\u80fd\u3002 \u9a57\u7d50\u679c\u4e2d\u89c0\u5bdf\u8cc7\u6599\u9078\u53d6\u65b9\u6cd5\u7684\u512a\u7f3a\u9ede\u3002\u672c\u8ad6\u6587\u63a5\u4e0b\u7684\u5b89\u6392\u5982\u4e0b\uff1a\u7b2c\u4e8c\u7ae0\u4ee5\u6700\u5177\u4ee3\u8868\u6027\u4e09</td></tr><tr><td>\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\u8b58\u3001\u8072\u5b78\u6a21\u578b\u9451\u5225\u5f0f\u8a13\u7df4\u3001\u908a\u969b\u8cc7\u8a0a\u3001\u8cc7\u6599\u9078\u53d6 \u7a2e\u9451\u5225\u5f0f\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u65b9\u6cd5\u70ba\u4f8b\uff0c\u6574\u5408\u5f9e\u5404\u7a2e\u4e0d\u540c\u89d2\u5ea6\u601d\u8003\u7684\u76ee\u6a19\u51fd\u6578\uff0c\u85c9\u7531\u6578\u5b78\u63a8\u5c0e</td></tr><tr><td>\u8a6e\u91cb\u5176\u76ee\u7684\u7684\u4e00\u81f4\u6027\uff1b\u7b2c\u4e09\u7ae0\u70ba\u6b78\u7d0d\u5404\u7a2e\u908a\u969b\u8cc7\u8a0a\u6982\u5ff5\u7684\u5ef6\u4f38\uff0c\u5206\u6790\u5404\u7a2e\u65b9\u6cd5\u6240\u5c08\u6ce8\u7684</td></tr><tr><td>\u8a13\u7df4\u8cc7\u6599\uff0c\u4e26\u63d0\u51fa\u4ee5\u67d4\u6027\u908a\u969b\u6cd5\u5247\u8207\u589e\u9032\u5f0f\u56e0\u5b50\u70ba\u57fa\u790e\u7684\u76ee\u6a19\u51fd\u6578\uff1b\u7b2c\u56db\u7ae0\u70ba\u5404\u7a2e\u65b9\u6cd5</td></tr></table>", | |
| "type_str": "table" | |
| } | |
| } | |
| } | |
| } |