| { | |
| "paper_id": "O15-1002", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T08:09:59.517020Z" | |
| }, | |
| "title": "Exploring Word Embedding and Concept Information for Language Model Adaptation in Mandarin Large Vocabulary Continuous Speech Recognition", | |
| "authors": [ | |
| { | |
| "first": "Ssu-Cheng", | |
| "middle": [], | |
| "last": "\u9673\u601d\u6f84", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Hsiao-Tsung", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "kychen@iis.sinica.edu.tw" | |
| }, | |
| { | |
| "first": "Berlin", | |
| "middle": [], | |
| "last": "Hung", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "berlin@ntnu.edu.tw" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Kuan-Yu", | |
| "middle": [], | |
| "last": "\u9673\u51a0\u5b87", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\u6548\u7528\u3002", | |
| "middle": [], | |
| "last": "\u5be6\u9a57\u7d50\u679c\u986f\u793a\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u7684\u8a9e\u8a00\u6a21\u578b\u8abf\u9069\u65b9\u6cd5\u76f8\u8f03\u65bc\u7576\u4eca\u6700\u597d\u65b9\u6cd5\u6709\u8f03\u4f73\u7684", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\u8b58\u3001\u8a9e\u8a00\u6a21\u578b\u3001\u8a5e\u5411\u91cf\u8868\u793a\u3001\u6982\u5ff5\u6a21\u578b", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "Research on deep learning has experienced a surge of interest in recent years. Alongside the rapid development of deep learning related technologies, various", | |
| "pdf_parse": { | |
| "paper_id": "O15-1002", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "Research on deep learning has experienced a surge of interest in recent years. Alongside the rapid development of deep learning related technologies, various", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "distributed representation methods have been proposed to embed the words of a vocabulary as vectors in a lower-dimensional space. Based on the distributed representations, it is anticipated to discover the semantic relationship between any pair of words via some kind of similarity computation of the associated word vectors. With the above background, this article explores a novel use of distributed representations of words for language modeling (LM) in speech recognition. Firstly, word vectors are employed to represent the words in the search history and the upcoming words during the speech recognition process, so as to dynamically adapt the language model on top of such vector representations. Second, we extend the recently proposed concept language model (CLM) by conduct relevant training data selection in the sentence level instead of the document level. By doing so, the concept classes of CLM can be more accurately estimated while simultaneously eliminating redundant or irrelevant information. On the other hand, since the resulting concept classes need to be dynamically selected and linearly combined to form the CLM model during the speech recognition process, we determine the relatedness of each concept class to the test utterance based the word representations derived with either the continue bag-of-words model (CBOW) or the skip-gram model (Skip-gram). Finally, we also combine the above LM methods for better speech recognition performance. Extensive experiments carried out on the MATBN (Mandarin Across Taiwan Broadcast News) corpus demonstrate the utility of our proposed LM methods in relation to several well-practiced baselines. [3, 4] ", | |
| "cite_spans": [ | |
| { | |
| "start": 1665, | |
| "end": 1668, | |
| "text": "[3,", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 1669, | |
| "end": 1671, | |
| "text": "4]", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u3002N \u9023(N-gram)\u8a9e\u8a00\u6a21\u578b\u70ba\u8a9e\u97f3\u8fa8\u8b58\u4e4b\u4e2d\u6700\u70ba\u5e38\u898b\u7684\u7d71\u8a08\u5f0f\u8a9e\u8a00\u6a21\u578b\uff0c\u7528\u4f86\u4f30\u6e2c\u6bcf\u4e00 \u500b\u5f85\u9810\u6e2c\u8a5e\u5f59\u5728\u5176\u5148\u524d\u7dca\u9130\u7684 N-1 \u500b\u8a5e\u5f59\u5df2\u77e5\u7684\u60c5\u6cc1\u4e0b\u51fa\u73fe\u7684\u689d\u4ef6\u6a5f\u7387\uff1b\u5047\u8a2d\u6bcf \u4e00\u500b\u8a5e\u5f59\u51fa\u73fe\u7684\u6a5f\u7387\u50c5\u8207\u5b83\u7dca\u9130\u7684\u524d N-1 \u500b\u8a5e\u5f59\u76f8\u95dc\uff0c\u53ef\u4ee5\u900f\u904e\u591a\u9805\u5f0f\u5206\u5e03 (Multinomial Distribution)\u4f86\u8868\u793a\u3002\u7136\u800c N \u9023\u8a9e\u8a00\u6a21\u578b\u50c5\u80fd\u64f7\u53d6\u77ed\u8ddd\u96e2\u7684\u8a5e\u5f59\u898f\u5247 \u8cc7\u8a0a\uff0c\u800c\u7121\u6cd5\u8003\u616e\u9577\u8ddd\u96e2\u7684\u8a9e\u53e5\u6216\u7bc7\u7ae0\u8cc7\u8a0a\uff1b\u7576\u8a5e\u5e8f\u5217\u8d8a\u9577\u6642\u53c3\u6578\u91cf\u8d8a\u591a\uff0c\u4f7f\u5f97 N \u9023\u8a9e\u8a00\u6a21\u578b\u6703\u6709\u7dad\u5ea6\u8a5b\u5492\u7684\u554f\u984c\u3002\u53e6\u4e00\u65b9\u9762\uff0cN \u9023\u8a9e\u8a00\u6a21\u578b\u4ea6\u5bb9\u6613\u9762\u81e8\u8a13\u7df4\u8a9e \u6599\u8207\u6e2c\u8a66\u8a9e\u6599\u4e0d\u5339\u914d(Mismatch)\u800c\u9020\u6210\u4f30\u6e2c\u8aa4\u5dee\u3002\u6709\u9451\u65bc\u6b64\uff0c\u8fd1\u5341\u5e7e\u5e74\u4f86\u8a31\u591a\u52d5 \u614b\u8a9e\u8a00\u6a21\u578b\u8abf\u9069\u6280\u8853\u88ab\u63d0\u51fa\uff0c\u7528\u4ee5\u767c\u5c55\u6709\u6548\u7684\u8a9e\u8a00\u6a21\u578b\u8f14\u52a9\u4e26\u5f4c\u88dc\u50b3\u7d71 N \u9023 (N-gram)\u8a9e\u8a00\u6a21\u578b\u4e0d\u8db3\u4e4b\u8655\u3002\u5e38\u898b\u7684\u6709\u5feb\u53d6\u6a21\u578b(Cache Model)[5]\uff0c\u4ee5\u53ca\u5728\u8cc7\u8a0a\u6aa2 \u7d22 \u9818 \u57df \u7684 \u4e3b \u984c \u6a21 \u578b (Topic Model)[6] \u7b49 \u3002 \u5176 \u4e2d \u53c8 \u4ee5 \u6a5f \u7387 \u5f0f \u6f5b \u85cf \u8a9e \u610f \u5206 \u6790 (Probabilistic Latent Semantic", | |
| "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": "\u505a\u70ba\u8a5e\u7684\u8868\u793a\u6cd5\uff0c\u662f\u900f\u904e\u524d\u994b\u5f0f\u985e\u795e\u7d93\u7db2\u8def(Feed-Forward Neural Network)\u8a13\u7df4\u800c \u6210\u3002\u9019\u7a2e\u5411\u91cf\u8868\u793a\u662f\u5c07\u8a5e\u8868\u793a\u6210\u4e00\u500b\u8f03\u4f4e\u7dad\u5ea6\u7684\u5be6\u6578\u5411\u91cf\u3002\u6bcf\u500b\u8a5e\u5f59\u4e4b\u9593\u7684\u95dc\u4fc2 \u53ef\u4ee5\u5229\u7528\u9918\u5f26\u6216\u662f\u6b50\u5f0f\u8ddd\u96e2\u8a08\u7b97\u627e\u51fa\u5169\u500b\u8a5e\u5411\u91cf\u9593\u7684\u8a9e\u610f\u76f8\u4f3c\u5ea6\uff0c\u6211\u5011\u5c07\u9019\u4e9b\u8a5e \u5411\u91cf\u7a31\u70ba\u8a5e\u8868\u793a\u6cd5(Word Representation or Embedding)\u3002 \u6709\u9451\u65bc\u4f7f\u7528\u50b3\u7d71\u985e\u795e\u7d93\u7db2\u8def\u8a9e\u8a00\u6a21\u578b\u4f86\u8a13\u7df4\u8a5e\u5411\u91cf\u6703\u9020\u6210\u8a13\u7df4\u6642\u9593\u904e\u9577\uff0c Tomas Mikolov \u7b49\u4eba[10] \u65bc\u662f\u63d0\u51fa\u6240 \u8b02 \u7684 \u9023 \u7e8c \u578b \u8a5e \u888b \u6a21 \u578b (Continuous Bag-of-Words Model, CBOW)\u8207\u8df3\u8e8d\u5f0f\u6a21\u578b(Skip-Gram Model, SG)\uff0c\u9019\u5169\u7a2e\u6a21\u578b \u4f7f\u7528\u968e\u5c64\u8edf\u5f0f\u6700\u5927\u5316(Hierarchical Soft-max, HS)[10] \u4ee5 \u53ca \u8ca0 \u4f8b \u63a1 \u6a23 (Negative Sampling, NS) [11]\u65b9\u6cd5\u4f86\u63d0\u9ad8\u8a13\u7df4\u7684\u901f\u5ea6\u4e26\u6539\u5584\u8a13\u7df4\u5f8c\u8a5e\u5411\u91cf\u7684\u8868\u793a\u80fd\u529b\u3002 \u5716\u4e00\u3001\u9023\u7e8c\u578b\u8a5e\u888b\u6a21\u578b\u793a\u610f\u5716 (\u4e00)\u3001\u9023\u7e8c\u578b\u6a21\u578b \u9023\u7e8c\u578b\u8a5e\u888b\u6a21\u578b(CBOW)\u8207\u524d\u994b\u5f0f\u985e\u795e\u7d93\u7db2\u8def\u985e\u4f3c\uff0c\u4e0d\u540c\u4e4b\u8655\u5728\u65bc\u9023\u7e8c\u578b\u8a5e\u888b\u6a21 \u578b\u5c07\u975e\u7dda\u6027\u96b1\u85cf\u5c64(Non-Linear Hidden Layer)\u79fb\u9664\uff0c\u4e26\u4e14\u5728\u8f38\u5165\u5c64\u7684\u6240\u6709\u55ae\u8a5e\u7686\u5171 \u4eab\u96b1\u85cf\u5c64\u3002\u5982\u5716\u4e00\u6240\u793a\uff0c\u6b64\u6a21\u578b\u5305\u542b\u4e09\u5c64\uff0c\u5206\u5225\u70ba\u8f38\u5165\u5c64\u3001\u6295\u5f71\u5c64\u3001\u8f38\u51fa\u5c64\u3002\u5df2 \u77e5\u7576\u524d\u8a5e w t \u7684\u4e0a\u4e0b\u6587w t-2 ,w t-1 ,w t+1 ,w t+2 \u7684\u60c5\u6cc1\u4e0b\u9810\u6e2c\u7576\u524d\u8a5ew t \u51fa\u73fe\u7684\u6a5f\u7387\u3002\u5728\u6b64 \u76ee\u6a19\u51fd\u6578\u70ba\u6700\u5927\u5316\u8a13\u7df4\u8a9e\u6599\u5eab\u4e2d\u6240\u6709\u8a5e\u5f59\u5e73\u5747\u7684\u767c\u751f\u6a5f\u7387: 1 T \u2211 log P(w t |w t-k ,\u2026,w t+k ) T-k t=k", | |
| "eq_num": "(1)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u5176\u689d\u4ef6\u6a5f\u7387\u53ef\u4ee5\u900f\u904e Softmax \u51fd\u6578\u8f49\u63db\u70ba: ", | |
| "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": "P(w t |w t-k ,\u2026,w t+k )= e \u2211 e y i i (2) \u5176\u4e2d y ={y 1 ,\u2026, y v }\uff0c\u800c y \u4e2d\u7684\u6bcf\u500b y i \u70ba\u5c0d\u65bc\u6bcf\u4e00\u500b\u8a5e w i \u9084\u672a\u7d93\u904e\u6b63\u898f\u5316\u7684 log \u6a5f \u7387\u503c\uff0c\u8a08\u7b97\u5982\u4e0b\u5f0f: y=b+Uh(w t-k ,\u2026,w t+k ,X) (3) \u5176\u4e2dU\u3001b\u70ba Softmax \u7684\u53c3\u6578\uff0ch \u662f\u5f9e\u77e9\u9663 X \u4e2d\u7684\u8a5e\u5411\u91cf(w t-k \u20d7\u20d7\u20d7\u20d7\u20d7\u20d7 ,\u2026,w t+k \u20d7\u20d7\u20d7\u20d7\u20d7\u20d7\u20d7 )\u52a0\u7e3d\u5e73\u5747\uff0cX\u70ba \u6839\u64da\u6bcf\u500b\u8a5ew i \u7684\u5411\u91cf\u6240\u7d44\u6210\u7684\u77e9\u9663\u3002 (\u4e8c)\u3001\u8df3\u8e8d\u5f0f\u6a21\u578b \u8df3\u8e8d\u5f0f\u6a21\u578b(Skip-gram)\u8207\u9023\u7e8c\u578b\u8a5e\u888b\u6a21\u578b(CBOW)\u76f8\u53cd\uff0c\u4f7f\u7528\u7576\u524d\u7684\u8a5e\u4f86\u9810\u6e2c\u5468 \u570d\u7684\u8a5e\u3002\u5728\u5df2\u77e5\u7576\u524d\u8a5e w t \u7684\u60c5\u6cc1\u4e0b\uff0c\u9810\u6e2c\u5176\u4e0a\u4e0b\u6587 w t-2 ,w t-1 ,w t+1 ,w t+2 \u7684\u6a5f\u7387\u3002\u7d66 \u5b9a\u4e00\u6bb5\u8a5e\u5e8f\u5217 w 1 ,w 2 ,w 3 ,\u2026,w t \uff0c\u5728\u6b64\u6700\u5927\u5316\u76ee\u6a19\u51fd\u6578: 1 T \u2211 \u2211 log P(w t+k |w t ) -c\u2264k\u2264c,k\u22600 T t=1", | |
| "eq_num": "(4)" | |
| } | |
| ], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "( | ) = \u2211 ( | ) \u2208 (5) \u5728\u6b64\u52a0\u5165\u53c3\u6578 \u03b1 j \uff0c\u4e26\u4e14\u5047\u8a2d\u53c3\u6578 1 , 2 , \u2026 , \u52a0\u7e3d\u70ba 1\uff0c\u4f7f\u5f97\u8ddd\u96e2\u8a5e \u8d8a\u8fd1\u7684\u8a5e \u7d66\u4e88\u8f03\u5927\u6b0a\u91cd\uff0c\u4ea6\u5373\u5728\u6b77\u53f2\u8a5e\u5e8f\u5217\u4e2d\u8d8a\u9760\u8fd1\u7576\u524d\u8a5e \u7684\u8a5e\u8d8a\u91cd\u8981\u3002 ( | )\u8868 \u793a\u5728\u7d66\u5b9a\u6b77\u53f2\u8a5e\u5e8f\u5217 H i \u4e2d\u8a5e \u4e0b\u9810\u6e2c\u7576\u524d\u8a5e \u7684\u6a5f\u7387\uff0c\u53ef\u4ee5\u7531(6)\u5f0f\u5f97\u5230: (w i |w m )= e w i \u20d7\u20d7\u20d7 \u2022 w m \u20d7\u20d7\u20d7\u20d7\u20d7\u20d7\u20d7 \u2211 e w i \u20d7\u20d7\u20d7 \u2022W \u20d7\u20d7\u20d7 W\u2208V (6) \u5176\u4e2d w i \u20d7\u20d7\u20d7\u20d7 \u70ba\u7576\u524d\u8a5e w i \u7684\u8a5e\u5411\u91cf\u8868\u793a\uff0c w m \u20d7\u20d7\u20d7\u20d7\u20d7\u20d7 \u70ba\u8a5e\u5716\u4e2d\u7684\u5019\u9078\u8a5e w m \u7684\u8a5e\u5411\u91cf\u8868\u793a\uff0c \u800c W \u70ba\u5c0d\u65bc\u8a5e w i \u7684\u6240\u6709\u5019\u9078\u8a5e\u96c6\u5408\uff0c\u6700\u5f8c\u900f\u904e Softmax", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(8) \uf0e5 \uf0d5 \uf0e5 \uf0d5 \uf0e5 \uf0e5 \uf0ce \uf0a2 \uf03d \uf0a2 \uf0a2 \uf0ce \uf03d \uf0ce \uf0a2 \uf0ce \uf0a2 \uf0a2 \uf03d \uf0a2 \uf0a2 \uf03d C C C C C L l l C L l l i C i C i i i i i i W C P C h P W C P C h P C w P W C P C H P W C P C H w P W H w P 1 1 CLM ) | ( ) | ( ) | ( ) | ( ) | ( ) | ( ) | ( ) | ( ) | , ( ) , | ( \uf0e5 \uf0d5 \uf0a2 \uf0a2 \uf0a2 \uf0e5 \uf0d5 \uf03d \uf0ce \uf0a2 \uf03d \uf0a2 \uf02d \uf0a2 \uf0a2 \uf0ce \uf03d \uf02d C C C L l l l C L l l l L i i i i i W C P C", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Two decades of statistical language modeling: Where do we go from here", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Rosenfeld", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "Proceedings of IEEE", | |
| "volume": "88", | |
| "issue": "", | |
| "pages": "1270--1278", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R. Rosenfeld, \"Two decades of statistical language modeling: Where do we go from here,\" Proceedings of IEEE, vol. 88, no. 8, 2000, pp. 1270-1278, 2000.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Statistical language model adaptation: review and perspectives", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "R" | |
| ], | |
| "last": "Bellegarda", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Speech Communication", | |
| "volume": "42", | |
| "issue": "11", | |
| "pages": "93--108", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J. R. Bellegarda, \"Statistical language model adaptation: review and perspectives,\" Speech Communication, vol. 42, no. 11, pp. 93-108, 2004.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Fundamental technologies in modern speech recognition", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Furui", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Gales", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Ney", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Tokuda", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "IEEE Signal Processing Magazine", | |
| "volume": "29", | |
| "issue": "6", | |
| "pages": "16--17", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Furui, L. Deng, M. Gales, H. Ney and K. Tokuda, \"Fundamental technologies in modern speech recognition,\" IEEE Signal Processing Magazine, vol. 29, no. 6, pp. 16-17, 2012", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Speech information processing: Theory and applications", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Shaughnessy", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of the IEEE", | |
| "volume": "101", | |
| "issue": "5", | |
| "pages": "1034--1037", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. O'Shaughnessy, L. Deng and H. Li, \"Speech information processing: Theory and applications,\" Proceedings of the IEEE, vol. 101, no. 5, pp 1034-1037, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Speech recognition and the frequency of recently used words: A modified Markov model for natural language", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Kuhn", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1988, | |
| "venue": "Proceedings of International Conference on Computational Linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "348--350", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R. Kuhn, \"Speech recognition and the frequency of recently used words: A modified Markov model for natural language,\" in Proceedings of International Conference on Computational Linguistics, pp. 348-350, 1988.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Topic models", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Blei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Lafferty", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Text Mining: Theory and Applications", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Blei and J. Lafferty, \"Topic models,\" in A. Srivastava and M. Sahami, (eds.), Text Mining: Theory and Applications, Taylor and Francis, 2009.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Probabilistic latent semantic indexing", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Hofmann", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1999, | |
| "venue": "Proceeding of the ACM Special Interest Group on Information Retrieval", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "50--57", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T. Hofmann, \"Probabilistic latent semantic indexing,\" in Proceeding of the ACM Special Interest Group on Information Retrieval, pp. 50-57, 1999.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Latent Dirichlet Allocation", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [ | |
| "M" | |
| ], | |
| "last": "Blei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "Y" | |
| ], | |
| "last": "Ng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [ | |
| "I" | |
| ], | |
| "last": "Jordan", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Journal of Machine Learning Research", | |
| "volume": "3", | |
| "issue": "", | |
| "pages": "993--1022", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. M. Blei, A. Y. Ng and M. I. Jordan, \"Latent Dirichlet Allocation,\" Journal of Machine Learning Research, vol. 3, pp. 993-1022, 2003.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Learning distributed representations of concepts", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Hinton", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1986, | |
| "venue": "Proceedings of the Eighth Annual Conference of the Cognitive Science Society", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1--12", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G.E. Hinton, \"Learning distributed representations of concepts,\" in Proceedings of the Eighth Annual Conference of the Cognitive Science Society, pages 1-12, Amherst 1986, 1986. Lawrence Erlbaum, Hillsdale.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Efficient estimation of word representations in vector space", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Mikolov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Corrado", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Dean", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceeding of International Conference on Learning Representations", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T. Mikolov, K. Chen, G. Corrado and J. Dean, \"Efficient estimation of word representations in vector space,\" in Proceeding of International Conference on Learning Representations, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Learning word embeddings efficiently with noise-contrastive estimation", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mnih", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Kavukcuoglu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceeding of Advances in Neural Information Processing Systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "2265--2273", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Mnih and K. Kavukcuoglu, \"Learning word embeddings efficiently with noise-contrastive estimation,\" in Proceeding of Advances in Neural Information Processing Systems, pp. 2265-2273, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Statistical language models for information retrieval: A critical review", | |
| "authors": [ | |
| { | |
| "first": "C", | |
| "middle": [ | |
| "X" | |
| ], | |
| "last": "Zhai", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "137--213", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "C. X. Zhai, \"Statistical language models for information retrieval: A critical review,\" Foundations and Trends in Information Retrieval, nol. 2, no. 3, 137- 213, 2008.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "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 the IEEE International Conference on Acoustics, Speech, 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 the IEEE International Conference on Acoustics, Speech, Signal Processing, 777- 780, 2004.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "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": "1", | |
| "pages": "219--235", | |
| "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. 1, 219-235, 2005.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Training data selection for improving discriminative training of acoustic models", | |
| "authors": [ | |
| { | |
| "first": "S.-H", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F.-H", | |
| "middle": [], | |
| "last": "Chu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S.-H", | |
| "middle": [], | |
| "last": "Lin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H.-S", | |
| "middle": [], | |
| "last": "Lee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Chen", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proceedings of IEEE workshop on Automatic Speech Recognition and Understanding", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "284--289", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S.-H. Liu, F.-H. Chu, S.-H. Lin, H.-S. Lee and Chen, \"Training data selection for improving discriminative training of acoustic models,\" in Proceedings of IEEE workshop on Automatic Speech Recognition and Understanding, 284-289, 2007.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "SRI Language Modeling Toolkit", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Stolcke", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Stolcke, A. (2000). SRI Language Modeling Toolkit. Available at: http://www.speech.sri.com/projects/srilm/.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "A latent semantic analysis framework for large-span language modeling", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "R" | |
| ], | |
| "last": "Bellegarda", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "Proceedings of European Conference on Speech Communication and Technology", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1451--1454", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J. R. Bellegarda, \"A latent semantic analysis framework for large-span language modeling,\" in Proceedings of European Conference on Speech Communication and Technology, pp.1451-1454, 1997.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Relevance language modeling for speech recognition", | |
| "authors": [ | |
| { | |
| "first": "K.-Y", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "5568--5571", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K.-Y. Chen and B. Chen, \"Relevance language modeling for speech recognition,\" in Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, pp. 5568-5571, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Leveraging relevance cues for language modeling in speech recognition", | |
| "authors": [ | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K.-Y.", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Information Processing & Management", | |
| "volume": "49", | |
| "issue": "4", | |
| "pages": "807--816", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "B. Chen and K.-Y. Chen, \"Leveraging relevance cues for language modeling in speech recognition,\" Information Processing & Management, Vol. 49, No 4, pp. 807-816, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Recurrent neural network based language model", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Mikolov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Karafi\u00e1t", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Burget", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "\u010cernock\u00fd", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proceedings of the Annual Conference of the International Speech Communication Association", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1045--1048", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T. Mikolov, M. Karafi\u00e1t, L. Burget, J. \u010cernock\u00fd and S. Khudanpur, \"Recurrent neural network based language model,\" in Proceedings of the Annual Conference of the International Speech Communication Association, 1045-1048, 2010.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF2": { | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td>\u5716\u4e8c\u3001\u8df3\u8e8d\u5f0f\u6a21\u578b\u793a\u610f\u5716</td></tr><tr><td>(\u4e09)\u3001\u5c07\u8a5e\u5411\u91cf\u8868\u793a\u61c9\u7528\u65bc\u8a5e\u5716\u641c\u5c0b</td></tr><tr><td>\u5728\u8a9e\u97f3\u8fa8\u8b58\u7684\u904e\u7a0b\u4e2d\uff0c\u6bcf\u500b\u97f3\u6846\u6703\u8a18\u9304\u8a9e\u8a00\u6a21\u578b\u7684\u6b77\u53f2\u8a5e\u5e8f\u5217\u3001\u5019\u9078\u8a5e\u5c0d\u61c9\u7684\u958b</td></tr><tr><td>\u59cb\u8207\u7d50\u675f\u7684\u97f3\u6846\u3001\u4ee5\u53ca\u641c\u5c0b\u6642\u8072\u5b78\u6a21\u578b\u7684\u89e3\u78bc\u5206\u6578\uff0c\u4f86\u5efa\u7acb\u8a5e\u5716(Word Graph)\uff0c</td></tr><tr><td>\u4e26\u5728\u8a5e\u5716\u4e0a\u4f7f\u7528\u4e09\u9023\u8a5e(Trigram)\u6216\u56db\u9023\u8a5e(Fourgram)\u7b49\u985e\u4f3c\u8a9e\u8a00\u6a21\u578b\uff0c\u5728\u91cd\u65b0\u9032</td></tr><tr><td>\u884c\u4e00\u6b21\u8a5e\u5716\u52d5\u614b\u898f\u5283\u641c\u5c0b(Word Graph Rescoring)\u4e2d\uff0c\u627e\u51fa\u4e00\u689d\u6700\u4f73\u7684\u8fa8\u8b58\u8a5e\u5e8f</td></tr><tr><td>\u5217\uff0c\u5982\u5716\u4e09\u6240\u793a\u3002</td></tr><tr><td>\u5716\u4e09\u3001\u8a5e\u5716\u641c\u5c0b\u793a\u610f\u5716</td></tr><tr><td>\u8a5e\u5716\u662f\u7531\u8a5e\u5f59\u6a39\u8907\u88fd\u641c\u5c0b\u904e\u5f8c\u6240\u5efa\u7acb\u7684\u5716\uff0c\u800c\u8a5e\u5716\u4e2d\u7684\u6bcf\u500b\u5206\u652f(Arc)\u8868\u793a</td></tr><tr><td>\u7d93\u904e\u88c1\u526a\u904e\u5f8c\u6240\u4fdd\u7559\u7684\u8a5e\u6bb5\uff0c\u6bcf\u500b\u8a5e\u6bb5\u6703\u8a18\u9304\u5176\u8072\u5b78\u5206\u6578\u3002\u63a5\u8457\u91dd\u5c0d\u6bcf\u500b\u8a5e\u6bb5\u9032</td></tr><tr><td>\u884c\u7dad\u7279\u6bd4(Viterbi)\u641c\u5c0b\uff0c\u4e26\u8a18\u9304\u8207\u6bcf\u500b\u8a5e\u6bb5\u76f8\u9023\u4e14\u6700\u6709\u53ef\u80fd\u7684\u4e0b\u4e00\u500b\u8a5e\u6bb5(\u4ea6\u5373\u524d</td></tr><tr><td>\u8a5e\u6bb5\u4e4b\u7d50\u675f\u6642\u9593\u8207\u4e0b\u4e00\u8a5e\u6bb5\u7684\u958b\u59cb\u6642\u9593\u76f8\u540c\u4e26\u4e14\u7dad\u7279\u6bd4\u5206\u6578\u70ba\u6700\u9ad8\u8005) \u3002\u7136\u800c\u5f9e</td></tr><tr><td>\u8a5e\u5716\u4e2d\u6240\u4fdd\u7559\u7684\u8a5e\u6bb5\uff0c\u5728\u8072\u5b78\u6a21\u578b\u4e2d\u5927\u591a\u70ba\u540c\u97f3\u7570\u5b57\u6216\u662f\u6df7\u6dc6\u7684\uff0c\u6240\u4ee5\u9700\u8981\u900f\u904e</td></tr><tr><td>\u8a9e\u8a00\u6a21\u578b\u7684\u8f14\u52a9\u3002</td></tr><tr><td>\u5728\u8a5e\u5716\u641c\u5c0b\u6642\uff0c\u7d66\u5b9a\u6b77\u53f2\u8a5e\u5e8f\u5217 \u4e0b\u9810\u6e2c\u7576\u524d\u8a5e w i \u7684\u6a5f\u7387\u53ef\u4ee5\u7531\u4e0b\u5f0f\u8868\u793a:</td></tr></table>", | |
| "text": "\u5176\u4e2d c \u70ba\u8a13\u7df4\u4e0a\u4e0b\u6587\u7684\u7a97\u53e3\u5927\u5c0f\uff0cT \u70ba\u8a13\u7df4\u7684\u6587\u5b57\u8a9e\u6599\u9577\u5ea6\uff0cP(w t+k |w t ) \u8868\u793a\u5728\u7576 \u524d\u8a5e w t \u7684\u689d\u4ef6\u4e0b w t+k \u51fa\u73fe\u7684\u6a5f\u7387\u3002\u8a08\u7b97\u5728\u4e00\u500b\u56fa\u5b9a\u7684\u7a97\u53e3\u5927\u5c0f\u5167\u5169\u5169\u8a5e\u5f59\u4e4b\u9593 \u7684\u6a5f\u7387\uff0c\u53ef\u4ee5\u7528\u4f86\u627e\u51fa\u5728\u4e00\u6bb5\u8a9e\u53e5\u4e2d\u8a5e\u5f59\u5f7c\u6b64\u4e4b\u9593\u7684\u76f8\u4e92\u95dc\u4fc2\u3002\u4e0a\u4e0b\u6587\u7684\u7a97\u53e3\u8d8a \u5927\uff0c\u9810\u6e2c\u7684\u7d50\u679c\u8d8a\u7cbe\u6e96\uff0c\u76f8\u5c0d\u7684\u8a13\u7df4\u6642\u9593\u4ea6\u6703\u96a8\u4e4b\u589e\u52a0\u3002", | |
| "num": null | |
| }, | |
| "TABREF4": { | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td colspan=\"18\">(\u4e8c)\u3001\u57fa\u790e\u5be6\u9a57\u7d50\u679c \u53e6\u5916\u5728\u8a08\u7b97\u6e2c\u8a66\u8a9e\u53e5\u8207\u6982\u5ff5\u7fa4\u805a\u76f8\u4f3c\u5ea6\u90e8\u5206\uff0c\u6211\u5011\u4f7f\u7528\u8a5e\u5411\u91cf\u8868\u793a\u4e26\u900f\u904e\u9918\u5f26\u65b9 \u91cf\u904e\u5c11\u800c\u7121\u6cd5\u63cf\u7e6a\u51fa\u5176\u6982\u5ff5\u7684\u7279\u6027\uff0c\u56e0\u6b64\u7fa4\u805a\u7684\u6578\u76ee\u4ea6\u662f\u6703\u5f71\u97ff\u8fa8\u8b58\u7d50\u679c\u7684\u91cd\u8981 \u8ca2\u737b\u53ef\u4ee5\u5206\u70ba\u5169\u500b\u90e8\u5206: \u7b2c\u4e00\u90e8\u5206\uff0c\u672c\u8ad6\u6587\u5c07\u8a5e\u5411\u91cf\u8868\u793a\u8cc7\u8a0a\u61c9\u7528\u65bc\u8a5e\u5716\u641c\u5c0b\u4e4b</td></tr><tr><td colspan=\"18\">W \u672c\u8ad6\u6587\u5c07\u8a5e\u5411\u91cf\u8868\u793a\u6cd5\u878d\u5165\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b\u4e2d\uff0c\u4e26\u4ee5\u5f0f(8)\u6240\u793a\u7684\u8a5e\u96d9\u9023\u6982\u5ff5\u8a9e\u8a00 C P C h h P C h P C h w P W H w P 2 1 1 BCLM ) | ( ) , | ( ) | ( ) , | ( ) , | ( \u5716\u56db\u3001\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b\u6d41\u7a0b\u5716 (\u4e00)\u3001\u7d50\u5408\u8a5e\u5411\u91cf\u8868\u793a\u8207\u6982\u5ff5\u8cc7\u8a0a\u65bc\u8a9e\u8a00\u6a21\u578b \u6a21\u578b(BCLM)\u70ba\u4f8b\u3002\u9996\u5148\uff0c\u5728\u8abf\u9069\u8a9e\u6599\u6587\u4ef6\u96c6\u5167\u4e4b\u6587\u4ef6\u7531\u4e00\u7d44\u6982\u5ff5\u985e\u5225 C \u4f86\u8868\u793a\uff0c \u4ee5\u7fa4\u805a\u4e4b\u9593\u7684\u76f8\u4f3c\u5ea6\u8fd1\u4f3c\u8a9e\u53e5\u6982\u5ff5\u8868\u9054\u7684\u6db5\u610f\u3002\u5728\u8abf\u9069\u8a9e\u6599\u4e2d\u4ee5\u53e5\u5b50\u7684\u5c64\u6b21\u505a\u6a21 \u578b\u8a13\u7df4\u8cc7\u6599\u9078\u53d6\u4e4b\u4f9d\u64da\uff0c\u5c07\u5177\u6709\u76f8\u4f3c\u8a9e\u610f\u6216\u662f\u76f8\u540c\u6982\u5ff5\u7684\u8a9e\u53e5\u6b78\u70ba\u540c\u4e00\u500b\u985e\u5225 \u4e2d\uff0c\u4f7f\u5f97\u7d93\u7531\u8abf\u9069\u8a9e\u6599\u4e2d\u8a13\u7df4\u51fa\u7684\u6982\u5ff5\u985e\u5225\u66f4\u70ba\u5177\u4ee3\u8868\u6027\u3002\u5176\u4e2d W \u4ee3\u8868\u8a9e\u8005\u6240 \u8b1b\u8a9e\u53e5\u6b32\u8868\u9054\u7684\u8a9e\u8a00\u8cc7\u8a0a\uff0c\u5728\u6b64\u4ee5\u8a9e\u97f3\u8fa8\u8b58\u521d\u6b65\u6240\u7522\u751f\u7684\u8a5e\u5716(Word Graph)\u4f86\u8fd1 \u4f3c\u3002 \u800c P(C|W)\u662f\u900f\u904e\u8a9e\u8a00\u8cc7\u8a0a W \u8207\u6bcf\u4e00\u500b\u6982\u5ff5\u985e\u5225 C \uff0c\u4ee5\u8a5e\u5411\u91cf\u8868\u793a(Word Embedding)\u7684\u65b9\u5f0f\uff0c\u5148\u5c07\u8a5e\u8f49\u63db\u6210\u5411\u91cf\u7684\u5f62\u5f0f\uff0c\u63a5\u8457\u8a08\u7b97\u5176\u9918\u5f26\u76f8\u4f3c\u5ea6\u800c\u5f97\u3002\u5176 \u4e2d\u8a5e\u5411\u91cf\u8868\u793a\u662f\u7531\u9023\u7e8c\u578b\u6a21\u578b(Continue Bag-of-Words Model)\u6216\u662f\u8df3\u8e8d\u5f0f\u6a21\u578b (Skip-gram Model)\u751f\u6210\u3002 ( | )\u8868\u793a\u6982\u5ff5\u985e\u5225 C \u9810\u6e2c\u8a5e\u5f59 \u7684\u55ae\u9023\u8a9e\u8a00\u6a21\u578b\u6a5f \u7387\uff0c\u53ef\u4ee5\u900f\u904e\u6700\u5927\u5316\u76f8\u4f3c\u6a5f\u7387\u4f30\u6e2c\u800c\u5f97\u3002 \u56db\u3001 \u5be6\u9a57\u8a2d\u5b9a\u8207\u7d50\u679c\u8a0e\u8ad6 (\u4e00)\u3001\u5be6\u9a57\u8a9e\u6599 \u672c\u7814\u7a76\u6240\u9032\u884c\u4e4b\u8a9e\u97f3\u8fa8\u8b58\u5be6\u9a57\u662f\u4f7f\u7528\u53f0\u5e2b\u5927\u6240\u81ea\u884c\u7814\u767c\u7684\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58 \u7cfb\u7d71(\u8a5e\u5178\u5927\u5c0f\u7d04\u70ba 7 \u842c 2 \u5343\u8a5e)[14]\u4ee5\u53ca\u516c\u8996\u96fb\u8996\u65b0\u805e\u8a9e\u97f3\u8a9e\uf9be\u5eab(Mandarin Across Taiwan Broadcast News, MATBN)[15]\u3002\u6b64\u65b0\u805e\u8a9e\u97f3\u8a9e\uf9be\u5eab\u662f\u7531\u4e2d\u592e\u7814\u7a76\u9662 \u8cc7\u8a0a\u6240\u53e3\u8a9e\u5c0f\u7d44\u8017\u6642\u4e09\uf98e(2001~2003)\u8207\u516c\u5171\u96fb\u8996\u53f0[PTS]\u5408\u4f5c\uf93f\u88fd\u5b8c\u6210\u3002\u6211\u5011\u521d Model)\uff0c\u6b64\u8a9e\u8a00\u6a21\u578b\u662f\u4f7f\u7528 SRI Language Modeling Toolkit (SRILM)[17]\u8a13\u7df4\u800c \u5f97\uff0c\u63a1\u7528 Good-Turning \u5e73\u6ed1\u5316\u65b9\u6cd5\u4f86\u89e3\u6c7a\u8cc7\u6599\u7a00\u758f\u7684\u554f\u984c\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u6211\u5011\u4ea6 \u8490\u96c6\u540c\u70ba\u516c\u8996\u96fb\u8996\u65b0\u805e\u8a9e\u6599\u5eab\u4e2d\u7684\u540c\u9818\u57df\u6587\u4ef6\u505a\u70ba\u8abf\u9069\u8a9e\u6599\u5eab\uff0c\u7528\u4f86\u4f30\u6e2c\u672c\u8ad6\u6587 \u6240\u63a2\u8a0e\u7684\u5404\u5f0f\u505a\u70ba\u8abf\u9069\u4e4b\u7528\u7684\u8a9e\u8a00\u6a21\u578b\uff0c\u7e3d\u5171\u7d04\u4e09\u5343\u516d\u767e\u56db\u4e09\u53e5\u8a9e\u53e5\u3002\u672c\u8ad6\u6587\u5be6 \u9a57\u6240\u4f7f\u7528\u4e4b\u8a9e\u97f3\u8a9e\u6599\u5eab\u4ee5\u53ca\u6587\u5b57\u8a9e\u6599\u5eab\u7684\u627c\u8981\u7d71\u8a08\u8cc7\u8a0a\u5206\u5225\u5982\u8868\u4e00\u8207\u8868\u4e8c\u6240\u793a\u3002 \u8868\u4e00 \u8a9e\u97f3\u8fa8\u8b58\u5be6\u9a57\u4f7f\u7528\u4e4b\u8a9e\u97f3\u8a9e\u6599\u7d71\u8a08\u8cc7\u8a0a \u8a5e\u5178\u5927\u5c0f \u53e5\u6578 \u9577\u5ea6(\u5c0f\u6642) \u8aaa\u8a71\u901f\u5ea6 \u8a9e\u6599 \u7d04 72000 \u8a5e 292 \u7d04 1.5 8.52 \u5b57/\u79d2 \u8868\u4e8c \u8a9e\u8a00\u6a21\u578b\u4f30\u6e2c\u6240\u4f7f\u7528\u80cc\u666f\u6587\u5b57\u8a9e\u6599\u4ee5\u53ca\u8abf\u9069\u6587\u5b57\u8a9e\u6599\u7d71\u8a08\u8cc7\u8a0a \u80cc\u666f\u8a9e\u6599 \u7d04 80,000,000 2,068,991 \u5728\u57fa\u790e\u5be6\u9a57\u90e8\u5206\uff0c\u9996\u5148\u50c5\u4f7f\u7528\u80cc\u666f\u8a9e\u8a00\u6a21\u578b\u65bc\u4e2d\u6587\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58\uff0c\u89c0\u5bdf\u5176 \u5b57\u8fa8\u8b58\u932f\u8aa4\u7387(Character Error Rate, CER)\uff0c\u6211\u5011\u4ea6\u6bd4\u8f03\u540c\u9818\u57df\u8a9e\u6599\u8a13\u7df4\u7684\u8a9e\u8a00\u6a21 \u578b\u7d50\u5408\u80cc\u666f\u8a9e\u8a00\u6a21\u578b\u7684\u5b57\u932f\u8aa4\u7387\u3002\u53e6\u5916\uff0c\u6211\u5011\u4ee5\u8a5e\u5716\u6700\u4f73\u89e3\u78bc(Oracle)\u4f5c\u70ba\u8a9e\u97f3 \u8fa8\u8b58\u6548\u80fd\u7684\u4e0a\u754c\uff1b\u8a5e\u5716\u4e2d\u6700\u4f73\u89e3\u78bc\u662f\u5229\u7528\u52d5\u614b\u898f\u5283\u65b9\u5f0f\uff0c\u627e\u51fa\u8a5e\u5716\u4e2d\u5b57\u932f\u8aa4\u7387\u6700 \u4f4e\u4e4b\u8def\u5f91\u3002\u57fa\u790e\u5be6\u9a57\u65bc\u6e2c\u8a66\u96c6\u4e4b\u5b57\u8fa8\u8b58\u7387\u7d50\u679c\u5982\u8868\u4e09\u6240\u793a\u3002 \u8868\u4e09\u3001\u8a9e\u97f3\u8fa8\u8b58\u57fa\u790e\u5be6\u9a57\u4e4b\u5b57\u8fa8\u8b58\u7387(%)\u7d50\u679c \u5b57\u932f\u8aa4\u7387(%) \u80cc\u666f\u55ae\u9023\u8a9e\u8a00\u6a21\u578b(UBG) 34.30 \u80cc\u666f\u96d9\u9023\u8a9e\u8a00\u6a21\u578b(BBG) 22.24 \u80cc\u666f\u4e09\u9023\u8a9e\u8a00\u6a21\u578b(TBG) 20.22 \u540c\u9818\u57df\u96d9\u9023\u8a9e\u8a00\u6a21\u578b+TBG 19.12 \u540c\u9818\u57df\u4e09\u9023\u8a9e\u8a00\u6a21\u578b+TBG 19.04 \u8a5e\u5716\u4e2d\u6700\u4f73\u89e3\u78bc(Oracle) 7.72 \u672c\u8ad6\u6587\u5e0c\u671b\u5229\u7528\u8a5e\u5411\u91cf\u8868\u793a\u627e\u5230\u8a5e\u5f59\u9593\u5f7c\u6b64\u7684\u8a9e\u610f\u95dc\u4fc2\uff0c\u5229\u7528\u8a5e\u5411\u91cf\u8868\u793a\u65bc\u8a9e\u97f3 \u8fa8\u8b58\u7684\u8a5e\u5716\u641c\u5c0b\u4e2d\uff0c\u5e0c\u671b\u85c9\u6b64\u80fd\u9054\u5230\u63d0\u5347\u8fa8\u8b58\u7387\u7684\u6548\u679c\u3002\u8868\u56db\u70ba\u6bd4\u8f03\u4e0d\u540c\u7dad\u5ea6\u4ee5 \u53ca\u4e0d\u540c\u8a5e\u5411\u91cf\u8868\u793a(Skip-gram, CBOW)\u65bc\u8a5e\u5716\u641c\u5c0b\u7684\u5b57\u932f\u8aa4\u7387\u7d50\u679c\uff0c\u5728\u6b64\u7dad\u5ea6\u8a2d \u5b9a\u4ee5 10 \u81f3 50 \u4f5c\u70ba\u5be6\u9a57\u4e4b\u6bd4\u8f03\uff0c\u4ee5\u8f03\u5c0f\u7dad\u5ea6\u4e4b\u5dee\u7570\u6bd4\u8f03\uff0c\u6e1b\u5c11\u5176\u8a08\u7b97\u8907\u96dc\u5ea6\u3002 \u8868\u56db\u3001\u61c9\u7528\u8a5e\u5411\u91cf\u8868\u793a\u65bc\u8a5e\u5716\u641c\u5c0b\u4e2d\u4e4b\u5b57\u932f\u8aa4\u7387(%)\u6bd4\u8f03\u8868 \u7dad\u5ea6\u5927\u5c0f \u8df3\u8e8d\u5f0f\u6a21\u578b(Skip-gram) \u9023\u7e8c\u578b\u8a5e\u888b\u6a21\u578b(CBOW) 10 19.85 19.86 20 19.85 19.87 30 19.83 19.84 40 19.85 19.86 50 19.85 19.84 \u7531\u8868\u56db\u4e2d\u53ef\u4ee5\u770b\u51fa\u878d\u5165\u8a5e\u5411\u91cf\u8868\u793a\u7684\u8cc7\u8a0a\u65bc\u8a5e\u5716\u641c\u5c0b\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u5f88\u660e\u986f\u5730\u89c0\u5bdf \u51fa\uff0c\u52a0\u5165\u8a5e\u5411\u91cf\u7684\u8cc7\u8a0a\u5c0d\u65bc\u8a9e\u97f3\u8fa8\u8b58\u6e96\u78ba\u7387\u7684\u63d0\u5347\u6709\u5e6b\u52a9\u3002\u4e0d\u8ad6\u662f\u7528\u8df3\u8e8d\u5f0f\u6a21\u578b (Skip-gram)\u9084\u662f\u4f7f\u7528\u9023\u7e8c\u578b\u8a5e\u888b\u6a21\u578b(CBOW)\u6240\u8a13\u7df4\u5f97\u5230\u7684\u8a5e\u5411\u91cf\u8868\u793a\uff0c\u5c07\u5176\u61c9 \u7528\u65bc\u8a9e\u97f3\u8fa8\u8b58\u7684\u8a5e\u5716\u641c\u5c0b\u4e4b\u4e2d\uff0c\u5b57\u932f\u8aa4\u7387\u5f9e\u539f\u672c\u53ea\u4f7f\u7528\u8a5e\u5716\u641c\u5c0b\u6642\u4e4b\u5b57\u932f\u8aa4\u7387 20.2 \u4e0b\u964d\u81f3 19.83 (\u4f7f\u7528 Skip-gram)\uff0c\u7372\u5f97\u4e0d\u932f\u7684\u6548\u80fd\u63d0\u5347\u3002 (\u56db) \u7d50\u5408\u8a5e\u5411\u91cf\u8868\u793a\u8cc7\u8a0a\u65bc\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b\u4e4b\u5be6\u9a57\u7d50\u679c \u672c\u8ad6\u6587\u5617\u8a66\u5c07\u8a5e\u5411\u91cf\u8cc7\u8a0a\u61c9\u7528\u65bc\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b\u4e4b\u4e2d\uff0c\u5728\u672c\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u5c07\u8abf\u9069\u8a9e \u6599\u4ee5\u53e5\u5b50\u70ba\u55ae\u4f4d\uff0c\u5229\u7528 K-means \u5206\u7fa4\u6cd5\u5c07\u8abf\u9069\u8a9e\u6599\u4e2d\u7684\u8a9e\u53e5\u5206\u70ba\u591a\u500b\u6982\u5ff5\u985e\u5225\u3002 \u95dc\u9375\u3002 \u4e2d\uff0c\u5728\u8a9e\u97f3\u8fa8\u8b58\u7684\u904e\u7a0b\u4e2d\uff0c\u5c0d\u65bc\u52d5\u614b\u7522\u751f\u4e4b\u6b77\u53f2\u8a5e\u5e8f\u5217\u8207\u5019\u9078\u8a5e\u6539\u4ee5\u8a5e\u5411\u91cf\u8868\u793a \u5f0f\u8a08\u7b97\u5176\u76f8\u4f3c\u5ea6\u3002\u672c\u5be6\u9a57\u6bd4\u8f03\u50b3\u7d71\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b(BCLM)\u8207\u7d50\u5408\u8a5e\u5411\u91cf\u8868\u793a\u65bc\u6982 \u5ff5\u8a9e\u8a00\u6a21\u578b(\u7c21\u7a31\u70ba BCLM:WE)\u7686\u4f5c\u7528\u65bc\u4e0d\u540c\u7fa4\u805a\u6578\u76ee\u4e4b\u5b57\u932f\u8aa4\u7387\u7d50\u679c;\u4e0a\u8ff0\u5169 \u7684\u65b9\u5f0f\u4f86\u5efa\u7acb\u5176\u5c0d\u61c9\u7684\u8a9e\u8a00\u6a21\u578b\uff0c\u900f\u904e\u6b64\u7a2e\u8868\u793a\u65b9\u5f0f\u800c\u80fd\u7372\u53d6\u5230\u66f4\u591a\u8a5e\u5f59\u9593\u7684\u8a9e (\u4e94) \u5404\u5f0f\u8a9e\u8a00\u6a21\u578b\u4e4b\u5be6\u9a57\u7d50\u679c\u6bd4\u8f03 \u610f\u8cc7\u8a0a\uff0c\u4ee5\u63d0\u5347\u8fa8\u8b58\u7684\u6e96\u78ba\u5ea6\u3002\u7b2c\u4e8c\u90e8\u5206\uff0c\u6211\u5011\u91dd\u5c0d\u65b0\u8fd1\u88ab\u63d0\u51fa\u7684\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b \u7a2e\u65b9\u6cd5\u7686\u8207\u80cc\u666f\u4e09\u9023\u8a9e\u8a00\u6a21\u578b\u505a\u7dda\u6027\u7d50\u5408\u3002\u672c\u5be6\u9a57\u63a1\u7528\u8df3\u8e8d\u5f0f\u6a21\u578b(Skip-gram) \u4f5c\u70ba\u8a5e\u5411\u91cf\u8a13\u7df4\uff0c\u76f8\u8f03\u65bc\u9023\u7e8c\u578b\u6a21\u578b(CBOW) \u6709\u8f03\u4f73\u5be6\u9a57\u7d50\u679c\u3002 \u5176\u4e2d BCLM:WE(10)\u8868\u793a\u4f7f\u7528\u8df3\u8e8d\u5f0f\u6a21\u578b\u8a13\u7df4\u7dad\u5ea6\u70ba 10 \u4e4b\u8a5e\u5411\u91cf\uff0c\u7d50\u5408\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b \u7684\u5be6\u9a57\u7d50\u679c\u3002\u5be6\u9a57\u7d50\u679c\u5982\u8868\u4e94\u6240\u793a\uff0c\u5716\u4e94\u4ee5\u6298\u7dda\u5716\u65b9\u5f0f\u5448\u73fe\u5176\u5be6\u9a57\u7d50\u679c\u3002 \u8868\u4e94\u3001\u7d50\u5408\u8a5e\u5411\u91cf\u8cc7\u8a0a\u65bc\u6982\u5ff5\u6a21\u578b\u4e4b\u4e0d\u540c\u7fa4\u805a\u6578\u7684\u5b57\u932f\u8aa4\u7387(%)\u6bd4\u8f03\u8868 \u7fa4\u805a\u500b\u6578 2 4 8 16 32 BCLM 19.31 19.14 19.58 19.54 19.59 BCLM:WE(10) 18.89 19.05 19.40 19.39 19.52 BCLM:WE(20) 18.90 19.05 19.40 19.39 19.52 BCLM:WE(30) 18.89 19.04 19.40 19.39 19.52 BCLM:WE(40) 18.88 19.05 19.40 19.39 19.52 BCLM:WE(50) 18.88 19.04 19.40 19.39 19.52 \u5716\u4e94\u3001\u7d50\u5408\u8a5e\u5411\u91cf\u8cc7\u8a0a\u65bc\u6982\u5ff5\u6a21\u578b\u4e4b\u4e0d\u540c\u7fa4\u805a\u6578\u7684\u5b57\u932f\u8aa4\u7387(%)\u6bd4\u8f03\u5716 \u7531\u5716\u4e94\u6211\u5011\u53ef\u4ee5\u770b\u51fa\u7d50\u5408\u8a5e\u5411\u91cf\u8868\u793a\u65bc\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b(BCLM:WE)\u4e2d\u4e4b\u5b57\u932f \u8aa4\u7387\u76f8\u8f03\u65bc\u50b3\u7d71\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b(BCLM)\u7686\u6709\u8f03\u597d\u7684\u8868\u73fe\uff0c\u7576\u7fa4\u805a\u6578\u76ee\u70ba 2 \u6642\uff0c\u4f7f \u7528\u8df3\u8e8d\u5f0f\u6a21\u578b(Skip-gram) \u8a13 \u7df4 \u5f97 \u5230 \u7684 \u8a5e \u5411 \u91cf \u8868 \u793a \u65bc \u6982 \u5ff5 \u8a9e \u8a00 \u6a21 \u578b (BCLM:WE(40))\u7576\u7dad\u5ea6\u70ba 40 \u6642\uff0c\u5b57\u932f\u8aa4\u7387\u53ef\u964d\u4f4e\u81f3 18.88\u3002\u53e6\u5916\uff0c\u4ea6\u53ef\u7531\u5716\u4e94\u4e2d \u8ad6\u6587\u4ee5\u6b64\u70ba\u767c\u60f3\uff0c\u63d0\u51fa\u5c07\u5206\u6563\u5f0f\u8868\u793a\u6cd5\u61c9\u7528\u65bc\u8a9e\u97f3\u8fa8\u8b58\u7684\u8a9e\u8a00\u6a21\u578b\u4e2d\u4f7f\u7528\u3002\u4e3b\u8981 \u770b\u51fa\u7576\u7fa4\u805a\u6578\u76ee\u589e\u52a0\u6642\u6709\u5229\u65bc\u6a21\u578b\u7684\u63cf\u8ff0\uff0c\u4f46\u662f\u7531\u65bc\u5206\u7fa4\u6578\u904e\u591a\u6703\u5c0e\u81f4\u6bcf\u7fa4\u8cc7\u6599 \u5716\u516d\u70ba\u5404\u5f0f\u8a9e\u8a00\u6a21\u578b\u8207\u80cc\u666f\u4e09\u9023\u8a9e\u8a00\u6a21\u578b(TBG)\u7d50\u5408\u5f8c\u4e4b\u5b57\u932f\u8aa4\u7387\u7d50\u679c\u6bd4\u8f03\uff0c\u5176 \u4e2d Baseline \u70ba\u8a5e\u5716\u641c\u5c0b(Word Graph Rescoring)\u50c5\u4f7f\u7528\u80cc\u666f\u4e09\u9023\u6a21\u578b\u7d50\u679c\uff0c\u5176\u5b57\u932f \u8aa4\u7387\u70ba 20.22;\u800c CBOW \u8207 Skip-gram \u70ba\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u5c07\u8a5e\u5411\u91cf\u8868\u793a\u61c9\u7528\u65bc\u8a5e\u5716\u641c \u5c0b\u4e4b\u5be6\u9a57\u7d50\u679c\uff0c\u76f8\u8f03\u65bc\u6c92\u6709\u4f7f\u7528\u8a5e\u5411\u91cf\u8868\u793a\u65bc\u8a5e\u5716\u641c\u5c0b\u7d50\u679c\u6709 0.39 \u7d55\u5c0d\u5b57\u932f\u8aa4 \u7387\u4e0b\u964d\u3002\u63a5\u8457\uff0c\u6211\u5011\u6bd4\u8f03\u6f5b\u85cf\u8a9e\u610f\u5206\u6790 (Latent Semantic Analysis, LSA)[18]\u3001\u6a5f \u7387\u5f0f\u6f5b\u85cf\u8a9e\u610f\u5206\u6790(Probabilistic Latent Semantic Analysis, PLSA)[7]\u3001\u72c4\u5229\u514b\u91cc\u5206 \u914d(Latent Dirichlet Allocation, LDA)[8]\u3001\u95dc\u806f\u6a21\u578b(Relevance Model, RM)[19, 20]\u3001 \u905e\u8ff4\u5f0f\u985e\u795e\u7d93\u7db2\u8def\u8a9e\u8a00\u6a21\u578b(Recurrent Neural Network, RNN)[21]\u3001\u6982\u5ff5\u6a21\u578b (Bigram Concept Language Model, BCLM)[12]\u4ee5\u53ca\u672c\u8ad6\u6587\u63d0\u51fa\u7d50\u5408\u8a5e\u5411\u91cf\u8868\u793a\u65bc \u6982\u5ff5\u8a9e\u8a00\u6a21\u578b(BCLM:WE) \u4e4b \u5be6 \u9a57 \u7d50 \u679c \u3002 \u6700\u5f8c\uff0cBCLM:WE+CBOW \u8207 BCLM:WE+Skip-gram \u70ba\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u7684\u5169\u7a2e\u65b9\u6cd5\u7d50\u5408(\u4ea6\u5373\u7b2c\u4e8c\u7bc0\u4ee5\u53ca\u7b2c\u4e09\u7bc0 \u6240\u63d0\u51fa\u8a9e\u8a00\u6a21\u578b\u8abf\u9069\u65b9\u6cd5\u4e4b\u7d50\u5408)\uff0c\u5be6\u9a57\u679c\u986f\u793a\uff0c\u5169\u8005\u7d50\u5408\u904e\u5f8c\u6548\u679c\u70ba\u6700\u597d\uff0c\u5b57 \u932f\u8aa4\u7387\u53ef\u4e0b\u964d\u81f3 18.70\u3002\u7531\u5716\u516d\u7d50\u679c\u89c0\u5bdf\u5f97\u77e5\uff0c\u672c\u8ad6\u6587\u63d0\u51fa\u5c07\u8a5e\u5411\u91cf\u8868\u793a\u61c9\u7528\u65bc \u8a9e\u8a00\u6a21\u578b\u4e2d\uff0c\u5c0d\u8a9e\u97f3\u8fa8\u8b58\u7684\u63d0\u5347\u78ba\u5be6\u6709\u5e6b\u52a9\u3002 \u5716\u516d\u3001\u5404\u5f0f\u8a9e\u8a00\u6a21\u578b\u4e4b\u5b57\u932f\u8aa4\u7387(%)\u7d50\u679c\u6bd4\u8f03\u5716 \u4e94\u3001 \u7d50\u8ad6\u8207\u672a\u4f86\u5c55\u671b \u8fd1\u5e74\u4f86\u6df1\u5ea6\u5b78\u7fd2(Deep Learning)\u6fc0\u8d77\u4e00\u80a1\u7814\u7a76\u71b1\u6f6e\uff1b\u96a8\u8457\u6df1\u5ea6\u5b78\u7fd2\u7684\u767c\u5c55\u800c\u6709\u5206 \u6563\u5f0f\u8868\u793a\u6cd5(Distributed Representation)\u7684\u7522\u751f\u3002\u6b64\u7a2e\u8868\u793a\u65b9\u5f0f\uff0c\u4e0d\u50c5\u80fd\u4ee5\u8f03\u4f4e\u7dad \u5ea6\u7684\u5411\u91cf\u8868\u793a\u8a5e\u5f59\uff0c\u9084\u80fd\u85c9\u7531\u5411\u91cf\u9593\u7684\u904b\u7b97\uff0c\u627e\u51fa\u4efb\u5169\u8a5e\u5f59\u4e4b\u9593\u7684\u8a9e\u610f\u95dc\u4fc2\u3002\u672c (Concept Language Model)\u52a0\u4ee5\u6539\u9032\uff0c\u5728\u8abf\u9069\u8a9e\u6599\u4e2d\u4ee5\u53e5\u5b50\u7684\u5c64\u6b21\u505a\u6a21\u578b\u8a13\u7df4\u8cc7\u6599 \u9078\u53d6\u4e4b\u4f9d\u64da\uff0c\u53bb\u6389\u591a\u9918\u4e14\u4e0d\u76f8\u95dc\u7684\u8cc7\u8a0a\uff0c\u4f7f\u5f97\u7d93\u7531\u8abf\u9069\u8a9e\u6599\u4e2d\u8a13\u7df4\u51fa\u7684\u6982\u5ff5\u985e\u5225 \u66f4\u70ba\u5177\u4ee3\u8868\u6027\uff0c\u800c\u80fd\u5e6b\u52a9\u52d5\u614b\u8a9e\u8a00\u6a21\u578b\u8abf\u9069\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5728\u8a9e\u97f3\u8fa8\u8b58\u904e\u7a0b\u4e2d\uff0c\u6703 \u9078\u64c7\u76f8\u95dc\u7684\u6982\u5ff5\u985e\u5225\u4f86\u52d5\u614b\u7d44\u6210\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b\uff0c\u800c\u6b64\u662f\u900f\u904e\u8a5e\u5411\u91cf\u8868\u793a\u7684\u65b9\u5f0f\u4f86 \u4f30\u7b97\uff0c\u85c9\u7531\u8a5e\u5411\u91cf\u8868\u793a\u8a18\u9304\u6bcf\u4e00\u500b\u6982\u5ff5\u985e\u5225\u5167\u8a5e\u5f59\u5f7c\u6b64\u9593\u7684\u8a9e\u610f\u95dc\u4fc2\u3002\u6700\u5f8c\uff0c\u6211 \u5011\u5617\u8a66\u5c07\u4e0a\u8ff0\u5169\u7a2e\u8a9e\u8a00\u6a21\u578b\u8abf\u9069\u6280\u8853\u505a\u7d50\u5408\u3002\u6839\u64da\u5be6\u9a57\u7d50\u679c\u986f\u793a\uff0c\u672c\u8ad6\u6587\u63d0\u51fa\u5c07 \u8a5e\u5411\u91cf\u8868\u793a(Word Representation)\u61c9\u7528\u65bc\u8a9e\u8a00\u6a21\u578b\u4e2d\uff0c\u5c0d\u65bc\u8a9e\u97f3\u8fa8\u8b58\u7684\u6e96\u78ba\u7387\u63d0 \u5347\u78ba\u5be6\u6709\u5e6b\u52a9\u3002 \u672a\u4f86\uff0c\u6211\u5011\u5e0c\u671b\u5c07\u8a5e\u5411\u91cf\u8868\u793a\u7684\u8cc7\u8a0a\u61c9\u7528\u65bc\u5176\u4ed6\u7684\u8a9e\u8a00\u6a21\u578b\u4e4b\u4e2d\uff0c\u4f8b\u5982\u61c9\u7528 \u65bc\u95dc\u806f\u6a21\u578b\u3001\u8a5e\u6982\u5ff5\u8a9e\u8a00\u6a21\u578b\u7b49\u3002\u6b64\u5916\uff0c\u6211\u5011\u5e0c\u671b\u4f9d\u64da\u8a5e\u5716\u641c\u5c0b\u7684\u7d50\u679c\u7d50\u5408\u5176\u4ed6 \u8a9e\u8a00\u6a21\u578b\u5f8c\uff0c\u5728\u7b2c\u4e8c\u968e\u6bb5\u7684 N \u689d\u6700\u4f73\u7d50\u679c(N-Best)\u91cd\u65b0\u6392\u540d\u6642\uff0c\u4f7f\u7528\u9577\u77ed\u671f\u8a18\u61b6 \u985e\u795e\u7d93\u7db2\u8def\u6a21\u578b\u3001\u905e\u8ff4\u5f0f\u985e\u795e\u7d93\u7db2\u8def\u7b49\u8a9e\u8a00\u6a21\u578b\u91cd\u65b0\u6392\u5e8f\uff0c\u5e0c\u671b\u85c9\u7531\u6b64\u65b9\u6cd5\u9054\u5230 \u8fa8\u8b58\u6548\u80fd\u7684\u63d0\u5347\u3002 \u6b65\u9078\u64c7\u5916\u5834\u63a1\u8a2a\u8a18\u8005\u8a9e\u6599\u4f5c\u70ba\u5be6\u9a57\u984c\u6750\uff0c\u5c07\u5176\u4e2d\u7d04 25 \u6709\u516b\u5343\u842c\u500b\u8a5e) \u505a \u70ba \u80cc \u666f \u8a9e \u6599 \u5eab \u7528 \u4f86 \u8a13 \u7df4 \u4e09 \u9023 \u8a9e \u8a00 \u6a21 \u578b (Trigram Language \u8a9e\u6599 \u8a5e\u6578 \u53e5\u6578 \u8abf\u9069\u8a9e\u6599 \u7d04 1,000,000 3,643 (\u4e09) \u5c07\u8a5e\u5411\u91cf\u8868\u793a\u61c9\u7528\u65bc\u8a5e\u5716\u641c\u5c0b\u4e4b\u5be6\u9a57\u7d50\u679c \u53c3\u8003\u6587\u737b</td></tr><tr><td>P</td><td>(</td><td>h 1</td><td>|</td><td>C</td><td>)</td><td>2</td><td>P</td><td>(</td><td>h</td><td>|</td><td>h</td><td>1</td><td>,</td><td>)</td><td>(</td><td>|</td><td>)</td></tr></table>", | |
| "text": "\u5c0f\u6642\u6536\u9304\u65bc 2001 \u5e74 11 \u6708 \u81f3 2002 \u5e74 12 \u6708\u671f\u9593\u7684\u8a9e\u6599\u4f5c\u70ba\u6700\u5c0f\u5316\u97f3\u7d20\u932f\u8aa4(Minimum Phone Error, MPE)\u8072 \u5b78\u6a21\u578b\u8a13\u7df4\u7684\u8a9e\u6599\u4f86\u5efa\u7acb\u8072\u5b78\u6a21\u578b(Acoustic Models)[16]\u3002\u672c\u8ad6\u6587\u4ee5 2003 \u5e74\u6240\u8490 \u96c6\u7684\u8a9e\u6599\u4e2d\u6311\u9078\u7d04 1.5 \u500b\u5c0f\u6642\uff0c\u5305\u542b 292 \u53e5\u8a9e\u53e5\u3002 \u5728\u8a9e\u8a00\u6a21\u578b\u7684\u4f30\u6e2c\u4e0a\uff0c\u6211\u5011\u4f7f\u7528\u81ea 2001 \u81f3 2002 \u5e74\u4e2d\u592e\u901a\u8a0a\u793e(Central News Agency, CNA)\u7684\u6587\u5b57\u65b0\u805e\u8a9e\u6599\uff0c\u5167\u542b\u6709\u7d04\u4e00\u5104\u4e94\u5343\u842c\u500b\u4e2d\u6587\u5b57(\u7d93\u7531\u65b7\u8a5e\u4e4b\u5f8c\u7d04", | |
| "num": null | |
| } | |
| } | |
| } | |
| } |