| { | |
| "paper_id": "O16-3005", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T08:05:05.554391Z" | |
| }, | |
| "title": "Character-Level Linguistic Features Extraction for Text-to-Speech System", | |
| "authors": [ | |
| { | |
| "first": "\u9673\u51a0\u5b8f", | |
| "middle": [ | |
| "\uf02a" | |
| ], | |
| "last": "\u3001\u5ed6\u66f8\u6f22", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\uf02a", | |
| "middle": [], | |
| "last": "\u3001\u5ed6\u5143\u752b", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\uf02a", | |
| "middle": [], | |
| "last": "\u3001\u738b\u9038\u5982", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Kuan-Hung", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Shu-Han", | |
| "middle": [], | |
| "last": "Liao", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Yuan-Fu", | |
| "middle": [], | |
| "last": "Liao", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Yih-Ru", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "High quality linguistic features is the key to the success of speech synthesis. Traditional linguistic feature extraction methods are usually relied on a word-level natural language processing (NLP) parser. Since, a good parser requires a lot of feature engineering to build, it is usually a genral-purpose one and often not specially designed for speech synthesis. To avoid these difficulties, we propose to replace the conventional NLP parser by a character embedding and a chacter-level", | |
| "pdf_parse": { | |
| "paper_id": "O16-3005", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "High quality linguistic features is the key to the success of speech synthesis. Traditional linguistic feature extraction methods are usually relied on a word-level natural language processing (NLP) parser. Since, a good parser requires a lot of feature engineering to build, it is usually a genral-purpose one and often not specially designed for speech synthesis. To avoid these difficulties, we propose to replace the conventional NLP parser by a character embedding and a chacter-level", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "-Of-Words Model)\u548c Skip-gram\u3002\u5716 3 \u70ba word2vec \u7684\u6838\u5fc3 \u67b6\u69cb\u5716\u3002 w(t) w(t) INPUT PROJECTION OUTPUT SUM w(t-2) w(t-1) w(t+1) w(t+2) INPUT PROJECTION OUTPUT w(t-1) w(t+1) w(t+2) w(t-2) \u5716", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u6642\u5e8f\u65b9\u5411 \u7bc4\u4f8b Forward \u5728 \u6c11 \u4e3b \u9ee8 \u7684 \u8868 \u73fe , \u4ed6 \u5011 1 \u81f4 \u8a8d \u70ba , 1 \u5207 \u4e0d \u53ef \u80fd \u6703 \u5c0d \u6211 \u5011 \u7684 \u58d3 \u529b . \u8b66 \u65b9 \u4eca \u5929 \u5728 \u9ad8 \u96c4 \u7e23 \u8b66 \u5bdf \u5c40 \u9577 \u5289 \u677e \u85e9 \u8a2a \u554f \u6642 \u8868 \u793a , \u9019 \u6b21 \u7684 \u6c11 \u9032 \u9ee8 \u7acb \u6cd5 \u9662 \u9ee8 \u5718 \u5c07 \u5728 \u660e \u5929 \u53ec \u958b \u8a18 \u8005 \u6703 , \u6703 \u4e2d \u8868 \u793a , \u6c11 \u9032 \u9ee8 \u6c7a \u5b9a \u5c07 \u5168 \u529b \u652f \u6301 . Backward . \u9047 \u5f85 \u570b \u60e0 \u6700 \u7684 \u570b \u7f8e \u958b \u96e2 \u5b9a \u6c7a \u5df2 , \u5224 \u8ac7 \u7684 \u570b \u7f8e \u8207 \u570b \u5408 \u806f \u5728 \u570b \u7f8e \u517c \u7406 \u7e3d \u526f \u570b \u7f8e , \u793a \u8868 \u7d71 \u7e3d \u674e \u6536 \u4f5c \u9ede 3 8 1 8 \u4ee5 , \u9ede 4 \u5341 7 \u767e 1 \u6f32 \u5927 \u6578 \u6307 \u50f9 \u80a1 \u6b0a \u52a0 \u91cf \u884c \u767c , \u9ede 9 5 . 5 9 1 \u4ee5 \u8ca8 \u671f \u6708 \u671f 1 0 . 9 5 9 \u91d1 \u57fa \u592a \u4e9e \u92b7 \u6c96 \u5929 \u4eca \u5143 \u7f8e \u514c \u5e63 \u53f0", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "(Master's thesis). Retrieved from http://etd.lib.nsysu.edu.tw/ETD-db/ETD-search-c/view_etd?URN=etd-0203110-093833]", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Segmentation", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "A SIMPLE RULE-BASED PART OF SPEECH TAGGER", | |
| "authors": [ | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Brill", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1992, | |
| "venue": "ANLC '92 Proceedings of the third conference on Applied natural language processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "152--155", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Brill, B. (1992). A SIMPLE RULE-BASED PART OF SPEECH TAGGER . In ANLC '92 Proceedings of the third conference on Applied natural language processing, 152-155.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Automatic prosody prediction for Chinese speech synthesis using BLSTM-RNN and embedding features", | |
| "authors": [ | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Ding", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Xie", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Yan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "proceedings of 2015 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "98--102", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ding, C., Xie, L., Yan, J., Zhang, W. & Liu, Y. (2015). Automatic prosody prediction for Chinese speech synthesis using BLSTM-RNN and embedding features. In proceedings of 2015 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU), 98-102.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "LSTM: A Search Space Odyssey. IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Greff", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Srivastava", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Koutn\u00edk", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [ | |
| "R" | |
| ], | |
| "last": "Steunebrink", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Schmidhuber", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Greff, K., Srivastava, R. K., Koutn\u00edk, J., Steunebrink, B. R. & Schmidhuber, J. (in press). LSTM: A Search Space Odyssey. IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, Retrieved from https://arxiv.org/abs/1503.04069", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Deep Learning in NLP (1) Word embedding and Language model", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Licstar", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "\u5e74 7 \u6708 29 \u65e5)\u3002Deep Learning in NLP (\u4e00)\u8a5e\u5411\u91cf\u548c\u8a9e\u8a00\u6a21\u578b\u3002\u3010\u90e8\uf918\u683c\u6587 \u5b57\u8cc7\uf9be\u3011\u3002\u53d6\u81ea", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Licstar, (2013 \u5e74 7 \u6708 29 \u65e5)\u3002Deep Learning in NLP (\u4e00)\u8a5e\u5411\u91cf\u548c\u8a9e\u8a00\u6a21\u578b\u3002\u3010\u90e8\uf918\u683c\u6587 \u5b57\u8cc7\uf9be\u3011\u3002\u53d6\u81ea http://licstar.net/archives/328\u3002[Licstar. (2013, July 29). Deep Learning in NLP (1) Word embedding and Language model [Web blog message]. Retrieved from http://licstar.net/archives/328]", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Context dependent recurrent neural network language model", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Mikolov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Zweig", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "proceedings of 2012 IEEE Spoken Language Technology Workshop (SLT)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/SLT.2012.6424228" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mikolov, T. & Zweig, G. (2012). Context dependent recurrent neural network language model. In proceedings of 2012 IEEE Spoken Language Technology Workshop (SLT). doi: 10.1109/SLT.2012.6424228", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Distributed Representations of Words and Phrases and their Compositionality", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Mikolov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "Sutskever", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Corrado", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Dean", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "proceedings of Advances in neural information processing systems 26 (NIPS 2013", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3111--3119", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mikolov, T., Sutskever, I., Chen,K., Corrado, G. & Dean, J. (2013). Distributed Representations of Words and Phrases and their Compositionality. In proceedings of Advances in neural information processing systems 26 (NIPS 2013), 3111-3119.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Stanford-Parser Version3", | |
| "authors": [], | |
| "year": 2015, | |
| "venue": "The Stanford Natural Language Processing Group", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "The Stanford Natural Language Processing Group. (2015). Stanford-Parser Version3.6.0 Release in 2015/12/09: http://nlp.stanford.edu/software/lex-parser.shtml", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Deep Learning for Chinese Word Segmentation and POS Tagging", | |
| "authors": [ | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Zheng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Xu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings s of the 2013 Conference on EMNLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "647--657", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Zheng, X., Chen, H. & Xu, T. (2013). Deep Learning for Chinese Word Segmentation and POS Tagging. In Proceedings s of the 2013 Conference on EMNLP, 647-657.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "A Study of Chinese Parser(Master's thesis)", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "\u5510 \u5927 \u4efb ; \u3002 \u4e2d \u6587 \u65b7 \u8a5e \u4e4b \u7814 \u7a76 ( \u78a9 \u58eb \u8ad6 \u6587 ) \u3002 \u53d6 \u81ea", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "\u5510 \u5927 \u4efb (2002) \u3002 \u4e2d \u6587 \u65b7 \u8a5e \u4e4b \u7814 \u7a76 ( \u78a9 \u58eb \u8ad6 \u6587 ) \u3002 \u53d6 \u81ea http://140.113.39.130/cgi-bin/gs32/tugsweb.cgi?o=dnctucdr&s=id=%22NT900435069% 22.&searchmode=basic [Tang, D.-R. (2002). A Study of Chinese Parser(Master's thesis). Retrieved from http://140.113.39.130/cgi-bin/gs32/tugsweb.cgi?o=dnctucdr&s=id=%22NT900435069%", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "An Enhanced Conditional Random Field Model for Chinese Word", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "\u9ec3 \u662d \u9298 ; \u3002 \u6539 \u5584 \u689d \u4ef6 \u96a8 \u6a5f \u57df \u6a21 \u578b \u65bc \u4e2d \u6587 \u65b7 \u8a5e ( \u78a9 \u58eb \u8ad6 \u6587 ) \u3002 \u53d6 \u81ea", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "\u9ec3 \u662d \u9298 (2010) \u3002 \u6539 \u5584 \u689d \u4ef6 \u96a8 \u6a5f \u57df \u6a21 \u578b \u65bc \u4e2d \u6587 \u65b7 \u8a5e ( \u78a9 \u58eb \u8ad6 \u6587 ) \u3002 \u53d6 \u81ea http://etd.lib.nsysu.edu.tw/ETD-db/ETD-search-c/view_etd?URN=etd-0203110-093833 [Huang, J.-m. (2010). An Enhanced Conditional Random Field Model for Chinese Word", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "text": "\u5728\u50b3\u7d71\u65b9\u5f0f\u4e0a\uff0c\u6587\u672c\u5206\u6790\u4e3b\u8981\u4f7f\u7528 NLP \u7684 parser(The Stanford Natural Language", | |
| "num": null, | |
| "type_str": "table", | |
| "content": "<table><tr><td>\u57fa\u65bc\u5b57\u5143\u968e\u5c64\u4e4b\u8a9e\u97f3\u5408\u6210\u7528\u6587\u8108\u8a0a\u606f\u64f7\u53d6</td><td>\u9673\u51a0\u5b8f \u7b49 75</td></tr><tr><td colspan=\"2\">Schmidhuber, 2015)\u4f7f\u7528\u4ee5\u5b57\u5143\u70ba\u5c64\u7d1a\u7684\u8a9e\u8a00\u6a21\u578b\uff0c\u5728\u76f8\u540c\u7684\u6027\u80fd\u4e0b\uff0c\u6703\u6bd4 n-gram \u6a21\u578b\u5c0f\u3002 \u8a5e\u6027\u6a19\u8a3b(Part Of Speech Tagging)\u5728 NLP \u4e2d\u4e5f\u662f\u4e00\u5927\u8ab2\u984c\u3002\u4e00\u822c parser \u5c31\u5305\u542b\u4e86\u65b7</td></tr><tr><td colspan=\"2\">\u800c\u4e14\u5c0d\u65bc\u4e2d\u6587\u7684\u8a9e\u97f3\u5408\u6210\u4f86\u8aaa\uff0c\u4ee5\u5b57\u5143\u5c64\u7d1a\u70ba\u8655\u7406\u55ae\u5143\u7684\u65b9\u5f0f\u66f4\u5177\u6709\u610f\u7fa9\uff0c\u56e0\u70ba\u4e2d\u6587\u662f \u8a5e\u8207 POS Tagging \u5169\u90e8\u5206\uff0c\u9806\u5e8f\u4e0a\u662f\u5148\u65b7\u8a5e\u518d\u505a\u8a5e\u6027\u6a19\u8a3b\uff0c\u800c\u8a5e\u6027\u6a19\u8a3b\u5c31\u662f\u900f\u904e\u9069\u7576\u7684</td></tr><tr><td colspan=\"2\">\u6c92\u6709\u7a7a\u683c\u7684\u9023\u7e8c\u5b57\u5143\u4e32\uff0c\u4e5f\u6c92\u6709\u8a5e\u7684\u5206\u9694\u7b26\u865f\uff0c\u6240\u4ee5\u4e2d\u6587\u7684\u8a5e\u5728\u5b9a\u7fa9\u4e0a\u662f\u5f88\u6a21\u7cca\u7684\uff0c\u5728 \u65b9\u5f0f\u5c0d\u7d93\u904e\u65b7\u8a5e\u8655\u7406\u5f8c\u7684\u6bcf\u500b\u8a5e\u7d66\u4e88\u4e00\u500b\u5408\u9069\u7684\u8a5e\u6027\uff0c\u4e5f\u5c31\u662f\u8981\u78ba\u5b9a\u9019\u500b\u8a5e\u662f\u540d\u8a5e\u3001\u52d5</td></tr><tr><td colspan=\"2\">\u9019\u60c5\u6cc1\u4e0b\u8a31\u591a\u4ee5\u5b57\u5143\u5c64\u7d1a\u7684\u8655\u7406\u65b9\u6cd5\u65b9\u6cd5\u88ab\u63d0\u51fa\uff0c\u50cf\u662f(Zheng, Chen & Xu, 2013)\u4e2d\u6587\u65b7 \u8a5e\u6216\u662f\u526f\u8a5e\u7b49\u7b49\uff0c\u95dc\u65bc\u8a5e\u6027\u6a19\u8a3b\u53ef\u4ee5\u53c3\u95b1(Brill, 1992)\u57fa\u65bc POS \u6a19\u8a3b\u7684\u7c21\u55ae\u898f\u5247\u3002\u4f46\u662f\u8a5e</td></tr><tr><td colspan=\"2\">\u8a5e\u8207 POS \u6a19\u8a3b\u7684\u6df1\u5c64\u5b78\u7fd2\u7db2\u8def\u4ee5\u53ca(Ding, Xie, Yan, & Zhang, 2015)\u63a1\u7528\u96d9\u5411\u9577\u77ed\u671f\u8a18\u61b6 \u6027\u6a19\u8a3b\u7684\u56f0\u96e3\u9ede\u5728\u65bc\u8a5e\u6027\u4e0d\u5b9a\u7684\u554f\u984c\u4e0a\uff0c\u9019\u7a2e\u73fe\u8c61\u662f\u81ea\u7136\u8a9e\u8a00\u4e2d\u6709\u5f88\u591a\u8a5e\u8a9e\u672c\u8eab\u5305\u542b\u5f88</td></tr><tr><td colspan=\"2\">\u905e\u8ff4\u985e\u795e\u7d93\u7db2\u8def(BLSTM)\u76f4\u63a5\u5f9e\u4e2d\u6587\u9810\u6e2c\u97fb\u5f8b\u908a\u754c\u7684\u6a19\u8a3b\uff0c\u9019\u4e9b\u7814\u7a76\u90fd\u8b49\u660e\u4e86\u4f7f\u7528 \u591a\u8a5e\u6027\uff0c\u64fa\u5728\u4e0d\u540c\u5730\u65b9\u5c31\u6703\u8b8a\u63db\u8a5e\u6027\u3002\u5c0d\u65bc\u4eba\u5728\u95b1\u8b80\u800c\u8a00\uff0c\u9019\u7a2e\u8a5e\u6027\u6b67\u7fa9\u73fe\u8c61\u6bd4\u8f03\u5bb9\u6613</td></tr><tr><td colspan=\"2\">DNNs \u662f\u80fd\u5920\u5be6\u73fe\u6bd4\u50b3\u7d71 CRF \u6709\u985e\u4f3c\u751a\u81f3\u66f4\u512a\u826f\u7684\u6548\u80fd\u3002 \u6392\u9664\uff0c\u4f46\u662f\u5c0d\u65bc\u6a5f\u5668\u800c\u8a00\u5247\u4e0d\u5bb9\u6613\u5340\u5206\u3002\u50b3\u7d71\u4f7f\u7528\u7684\u8a5e\u6027\u6a19\u8a3b\u898f\u5247\u662f\u7531\u8a9e\u8a00\u5b78\u5bb6\u6839\u64da\u8a9e</td></tr><tr><td colspan=\"2\">\u5728\u9019\u9ebc\u591a\u4f7f\u7528 DNNs \u7372\u5f97\u4e0d\u932f\u6210\u679c\u7684\u7814\u7a76\u4e0b\uff0c\u6211\u5011\u60f3\u8aaa\u770b\u53ef\u4e0d\u53ef\u4ee5\u4f7f\u7528\u5927\u91cf\u6c92\u6709\u7d93 \u8a00\u898f\u5f8b\u9032\u884c\u4eba\u5de5\u6a19\u8a3b\u5b8c\u6210\u7684\uff0c\u6709\u4e86\u6a19\u8a3b\u5b8c\u7684\u8cc7\u6599\u5eab\uff0c\u5728\u4f7f\u7528\u4e0a\u5229\u7528 CRF \u5f9e\u5df2\u77e5\u7684\u5e8f\u5217\u6c42</td></tr><tr><td colspan=\"2\">1. \u7c21\u4ecb \u904e\u6a19\u8a3b\u7684\u8a9e\u6599\u4f86\u8a13\u7df4 DNN\uff0c\u9019\u6a23\u4e0d\u50c5\u53ef\u4ee5\u514d\u53bb\u6a19\u8a3b\u5de5\u4f5c\u4e5f\u7121\u9700\u5c08\u5bb6\u8a2d\u8a08\u7684\u65b7\u8a5e\u6216 POS \u5c0d\u61c9\u5e8f\u5217\u7684\u65b9\u5f0f\uff0c\u4f86\u731c\u6e2c\u8f38\u5165\u8a9e\u6599\u7684\u8a5e\u6027\u3002\u8a73\u7d30\u5167\u5bb9\u53ef\u4ee5\u53c3\u95b1\u76f8\u95dc\u7814\u7a76\uff0c\u4f8b\u5982(\u5510\u5927\u4efb\uff0c</td></tr><tr><td colspan=\"2\">\u7279\u5fb5\u53c3\u6578\uff0c\u800c\u8981\u9054\u6210\u4f7f\u7528\u7121\u6a19\u8a3b\u8a9e\u6599\u7684\u76ee\u7684\uff0c\u6211\u5011\u9700\u8981\u9760\u5b57\u5143\u5c64\u7d1a(character-level)\u7684\u6587 2002)\u4e2d\u6587 parser \u4e4b\u7814\u7a76\uff0c\u5167\u6587\u63a2\u8a0e\u4e86\u8a2d\u8a08 parser \u6642\u5229\u7528\u7684\u65b7\u8a5e\u8207\u69cb\u8a5e\u898f\u5247\uff0c\u548c\u6a19\u8a18\u8a5e\u985e \u5728\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u4e2d\u5206\u70ba\u5169\u5927\u6a21\u7d44(\u5982\u5716 1)\uff0c\u5206\u5225\u662f\u6587\u672c\u5206\u6790(\u81ea\u7136\u8a9e\u8a00\u8655\u7406)\u8207\u8072\u97f3\u5408\u6210(\u8a9e \u97f3\u8a0a\u865f\u8655\u7406)\u3002\u5176\u4e2d\u524d\u7aef\u7684\u6587\u672c\u5206\u6790\u901a\u5e38\u6703\u505a\u6587\u5b57\u6b63\u898f\u5316\u3001\u65b7\u8a5e(word segmentation)\u3001part of speech(POS)\u6a19\u8a3b\u8207\u76f8\u95dc\u6587\u6cd5\u5206\u6790\uff0c\u751a\u81f3\u662f\u85c9\u7531\u97fb\u5f8b\u9810\u6e2c\u5f9e\u6587\u672c\u63d0\u53d6\u6587\u8108\u8a0a\u606f\u7279\u5fb5\u3002\u4f8b \u5b57\u8655\u7406\u7684\u5e6b\u52a9\uff0c\u56e0\u6b64\u6211\u5011\u6368\u68c4\u4ee5\u8a5e\u70ba\u55ae\u5143\uff0c\u5c07\u8f38\u5165\u7c21\u5316\u6210\u4ee5\u5b57\u5143\u70ba\u55ae\u4f4d\uff0c\u4e0d\u7d93\u904e parser \u505a\u65b7\u8a5e\u3001\u6c42\u8a5e\u6027\u7b49\u524d\u8655\u7406\uff0c\u800c\u662f\u4e00\u500b\u5b57\u5143\u4e00\u500b\u5b57\u5143\u9010\u6b21\u8f38\u5165\u7db2\u8def\u4e2d\uff1b\u5e0c\u671b\u80fd\u5c07\u5b57\u5143\u8f49\u5230 \u66f4\u9ad8\u7dad\u7684\u5411\u91cf\u7a7a\u9593\uff0c\u5728\u5411\u91cf\u7a7a\u9593\u9032\u884c\u5206\u6790\uff0c\u80fd\u6709\u6548\u7684\u5f9e\u8a9e\u6599\u4e2d\uff0c\u81ea\u52d5\u5b78\u7fd2\u5b57\u5143\u4e4b\u9593\u96b1\u85cf \u65b9\u5f0f\u7684\u4e00\u4e9b\u554f\u984c\u3002 \u5716 2. \u65b0\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u67b6\u69cb\u5716 \u5728 NLP \u4e2d\u4f7f\u7528 parser \u662f\u975e\u5e38\u666e\u904d\u7684\uff0c\u5c0d\u82f1\u6587\u800c\u8a00\u4f7f\u7528 parser \u80fd\u5f97\u5230\u4e0d\u932f\u7684\u7d50\u679c\uff0c\u4f46 [Figure 2. The proposed character-based TTS apporach] \u662f\u5728\u4e2d\u6587\u4f7f\u7528\u8d77\u4f86\u537b\u4e0d\u76e1\u7406\u60f3\uff0c\u56e0\u70ba\u4e2d\u6587\u7d50\u69cb\u5728\u6b67\u7fa9\u6027\u4e0a\u6709\u592a\u591a\u8b8a\u5316\uff0c\u9700\u8981\u642d\u914d\u4e0a\u4e0b\u6587 \u5982\u5728\u4e2d\u6587\u8a9e\u97f3\u5408\u6210\u4e2d\u7d93\u5e38\u63a1\u7528\u689d\u4ef6\u96a8\u6a5f\u5834(Conditional Random Fields\uff0cCRF)\u505a\u65b7\u8a5e\u548c POS \u6a19\u8a3b\u3002\u53e6\u4e00\u65b9\u9762\u5728\u5f8c\u7aef\u8072\u97f3\u5408\u6210\u4e00\u822c\u6703\u900f\u904e\u6c7a\u7b56\u6a39\u4f9d\u64da\u6587\u8108\u8a0a\u606f\u9078\u64c7\u6700\u9069\u5408\u7684\u8072\u97f3 \u6216\u97fb\u5f8b\u7279\u5fb5\uff0c\u5c07\u9078\u51fa\u4f86\u7684\u8072\u5b78\u53c3\u6578\u7d66\u8a9e\u97f3\u7de8\u78bc\u5668\u5408\u6210\u8a9e\u97f3\u6ce2\u5f62\u7522\u51fa\u8072\u97f3\u3002\u56e0\u6b64\u82e5\u662f\u6211\u5011 \u60f3\u5728\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u4e2d\u7372\u5f97\u81ea\u7136\u3001\u6d41\u66a2\u7684\u5408\u6210\u8072\u97f3\uff0c\u9700\u8981\u80fd\u8403\u53d6\u6709\u6548\u4e14\u6709\u7528\u7684\u6587\u8108\u8a0a\u606f\u3002 \u6574\u7684\u8868\u73fe\u51fa\u6bcf\u6bb5\u8a9e\u53e5\u5167\u7684\u8a5e\u6027\u72c0\u614b\u3002\u4f46\u662f\u8981\u5efa\u7acb\u50b3\u7d71\u65b7\u8a5e\u5f88\u96e3\u505a\uff0c\u9700\u8981\u5927\u91cf\u5c08\u696d\u4eba\u58eb\u6a19 \u6587\u672c\u5206\u6790\u4e2d\u626e\u6f14\u8209\u8db3\u8f15\u91cd\u7684\u89d2\u8272\u3002 \u5728\u8a9e\u8a00\u5b78\u4e0a\uff0c\u8a5e\u662f\u80fd\u5920\u7368\u7acb\u904b\u7528\u800c\u4e14\u542b\u6709\u8a9e\u7fa9\u5167\u5bb9\u7684\u6700\u5c0f\u8a9e\u8a00\u55ae\u4f4d\u3002\u5728\u82f1\u6587\u6587\u672c\u4e2d\uff0c 3.1 \u65b0\u7cfb\u7d71\u67b6\u69cb \u5716 1Processing Group, 2015)\u3002\u4f7f\u7528 parser \u7684\u597d\u8655\u662f\u7d93\u904e\u8a9e\u8a00\u5b78\u5c08\u5bb6\u6240\u8a2d\u8a08\u7684\u65b7\u8a5e\u80fd\u5920\u6bd4\u8f03\u5b8c \u7684\u76f8\u5c0d\u95dc\u4fc2\u3002\u6240\u4ee5\u6211\u5011\u5c07\u900f\u904e\u5efa\u7acb\u8a9e\u6599\u7684\u5b57\u5411\u91cf\u7a7a\u9593\u4e26\u9032\u884c\u5206\u6790\uff0c\u5c07\u6587\u672c\u5b57\u5143\u5206\u985e\uff0c\u7522 \u751f\u5b57\u5143\u8a9e\u610f\u3001\u6587\u6cd5\u8173\u8272\u8cc7\u8a0a\u7b49\u6587\u8108\u8a0a\u606f\uff1b\u53e6\u4e00\u65b9\u9762\u5247\u4f7f\u7528\u905e\u8ff4\u795e\u7d93\u7db2\u8def\u9032\u884c\u5b57\u5143\u4e32\u8a13\u7df4\uff0c \u4f86\u89c0\u5bdf\uff0c\u7121\u6cd5\u55ae\u7d14\u9760\u4eba\u5de5\u6a19\u8a3b\u53ca\u5c08\u5bb6\u8a2d\u8a08\u6240\u6709\u53ef\u80fd\uff0c\u6240\u4ee5\u6211\u5011\u9700\u8981\u4e00\u5957\u65b0\u65b9\u6cd5\u4f86\u514b\u670d\u9019 3.2 \u5b57\u5143\u8a9e\u610f\u8207\u6587\u6cd5\u5c6c\u6027\u4e4b\u6587\u8108\u8a0a\u606f\u64f7\u53d6 \u4e9b\u554f\u984c\u3002 \u6587\u5b57\u63a2\u52d8\u4ee5\u53ca NLP \u5728\u6578\u64da\u5206\u6790\u4e16\u754c\u4e2d\u4e00\u76f4\u662f\u975e\u5e38\u91cd\u8981\u7684\u4e00\u90e8\u5206\u3002\u5176\u4e2d word2vec \u88ab\u983b\u7e41 \u5206\u6790\u76ee\u524d\u8f38\u5165\u5b57\u5143\u5728\u6574\u53e5\u8a71\u4e2d\u7684\u72c0\u614b\uff0c\u4e26\u731c\u6e2c\u4e0b\u4e00\u72c0\u614b\u53ef\u80fd\u70ba\u4f55\uff0c\u6700\u5f8c\u4e26\u64f7\u53d6\u96b1\u85cf\u5c64\u5728 3. \u5b57\u5143\u968e\u5c64\u6587\u8108\u8a0a\u606f\u64f7\u53d6\u65b9\u6cd5\u8207\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71 \u7684\u8a0e\u8ad6\u4ee5\u53ca\u4f7f\u7528\uff0c\u56e0\u70ba\u4f7f\u7528\u5b83\u80fd\u5920\u5c07\u8f38\u5165\u7684\u8a5e\u8f49\u5230\u5411\u91cf\u7a7a\u9593\u4e0a\u4e26\u9032\u884c\u6f14\u7b97\uff0c\u5206\u6790\u5f8c\u53ef\u4ee5 \u5404\u500b\u795e\u7d93\u5143\u7684\u72c0\u614b\u8f38\u51fa\u505a\u70ba\u5b57\u5143\u7684\u6642\u9593\u9806\u5e8f\u8cc7\u8a0a\u7576\u4f5c\u5b57\u5143\u6642\u5e8f\u72c0\u614b\u7684\u6587\u8108\u8a0a\u606f\u3002\u5982\u6b64\u4e00 \u4f86\u6211\u5011\u5c31\u80fd\u5229\u7528\u5927\u91cf\u672a\u6a19\u8a3b\u7684\u6587\u5b57\u8a9e\u6599\uff0c\u81ea\u52d5\u64f7\u53d6\u6587\u8108\u8a0a\u606f\uff0c\u4e26\u63a2\u8a0e\u66f4\u591a\u7a2e\u6587\u8108\u8a0a\u606f\u7684 \u53ef\u80fd\u6027\u3002 2. \u50b3\u7d71\u6587\u8108\u8a0a\u606f\u64f7\u53d6\u65b9\u6cd5 \u4e00\u822c\u6587\u672c\u5206\u6790\u7684\u6587\u8108\u8a0a\u606f\u662f\u8f38\u5165\u6587\u672c\u7d93\u7531 parser \u505a\u65b7\u8a5e\u3001\u6293 POS\u3001\u4f4d\u7f6e\u518d\u52a0\u4e0a\u5207\u5272\u8cc7\u8a0a\u3001 \u8072\u8abf\u8cc7\u8a0a\u7b49\u5408\u8d77\u4f86\u7684\uff0c\u6240\u4ee5\u8981\u5f97\u5230\u597d\u7684\u6587\u8108\u8a0a\u606f\u9019\u4e9b\u53c3\u6578\u9700\u8981\u7cbe\u78ba\u548c\u6709\u7528\uff0c\u800c parser \u5728 \u767c\u73fe\u5728\u5411\u91cf\u7a7a\u9593\u4e2d\uff0c\u76f8\u805a\u5728\u4e00\u8d77\u7684\u8a5e\u5411\u91cf\u8f49\u63db\u56de\u6587\u5b57\u5f8c\u6703\u662f\u76f8\u540c\u5c6c\u6027\u7684\u8a5e\u5f59\u3002\u4e5f\u5c31\u662f\u8aaa \u5728\u524d\u4e00\u7ae0\u4e2d\u6211\u5011\u5df2\u7d93\u77e5\u9053\u50b3\u7d71\u67b6\u69cb\u64f7\u53d6\u8a9e\u8a00\u7279\u5fb5\u6703\u9047\u5230\u7684\u554f\u984c\u9ede\uff0c\u6240\u4ee5\u6211\u5011\u5e0c\u671b\u80fd\u4f7f\u7528 \u5b83\u6709\u80fd\u5c07\u5b57\u8a5e\u8a9e\u610f\u6216\u6587\u6cd5\u89d2\u8272\u505a\u5206\u985e\u7684\u80fd\u529b\uff0c\u800c\u4e14\u5b83\u7121\u9700\u7d66\u5b9a\u6a19\u8a3b\u904e\u7684\u8cc7\u6599\u5eab\u5c31\u80fd\u5c0d\u8a9e \u5927\u91cf\u7121\u6a19\u8a3b\u8cc7\u6599\u5eab\u8207\u975e\u76e3\u7763\u5f0f\u5b78\u7fd2(unsupervised learning)\u4f86\u8a13\u7df4\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\uff0c\u9054\u5230\u7121 \u6599\u76f4\u63a5\u9032\u884c\u8a13\u7df4\uff0c\u9019\u975e\u5e38\u7b26\u5408\u6211\u5011\u60f3\u8981\u907f\u958b\u4eba\u5de5\u6a19\u8a3b\u7684\u521d\u8877\u3002 \u9700\u4eba\u5de5\u4ecb\u5165\u8b93\u6a5f\u5668\u81ea\u52d5\u5b78\u7fd2\u6587\u672c\u53d6\u4ee3 parser\u3002\u5728\u9019\u57fa\u790e\u4e0b\u6211\u5011\u4f7f\u7528\u4ee5\u5b57\u5143\u70ba\u5c64\u7d1a\u7684\u8a13\u7df4 Word2vec \u662f Google \u516c\u53f8\u5728 2013 \u5e74\u958b\u653e\u7684\u4e00\u6b3e\u7528\u65bc\u8a13\u7df4\u8a5e\u5411\u91cf\u7684\u8edf\u9ad4\u5de5\u5177\u3002\u5b83\u6839\u64da \u65b9\u5f0f\u4f86\u9054\u5230\u6211\u5011\u96f6\u6a19\u8a3b\u7684\u76ee\u7684\u3002\u5728\u5b57\u5143\u5c6c\u6027\u6c42\u53d6\u4e0a\u6211\u5011\u63a1\u7528 word2vec \u4f86\u5c0d\u6587\u672c\u9032\u884c\u5b57\u5143 \u7d66\u5b9a\u7684\u8a9e\u6599\u5eab\uff0c\u901a\u904e\u512a\u5316\u5f8c\u7684\u8a13\u7df4\u6a21\u578b\u5feb\u901f\u6709\u6548\u7684\u5c07\u4e00\u500b\u8a5e\u8a9e\u8868\u9054\u6210\u5411\u91cf\u5f62\u5f0f\uff0c\u5176\u6838\u5fc3 \u7684\u8a9e\u610f\u8207\u6587\u6cd5\u89d2\u8272\u5206\u985e\uff0c\u5b57\u5143\u4e4b\u9593\u7684\u6642\u5e8f\u9806\u5e8f\u95dc\u4fc2\u6211\u5011\u5229\u7528\u905e\u8ff4\u5f0f\u795e\u7d93\u7db2\u8def\uff0c\u64f7\u53d6\u5b57\u5143 \u5728\u53e5\u5b50\u4e2d\u7684\u72c0\u614b\u4f86\u7576\u4f5c\u6642\u5e8f\u95dc\u4fc2\uff0c\u6700\u5f8c\u5f97\u5230\u5b57\u5143\u5c6c\u6027\u8207\u5b57\u5143\u6642\u5e8f\u72c0\u614b\u7684\u6587\u8108\u8a0a\u606f\u3002 \u67b6\u69cb\u5305\u62ec CBOW(Continuous Bag</td></tr><tr><td colspan=\"2\">\u8a3b\u7684\u8cc7\u6599\u5eab\u4ee5\u53ca\u5c08\u5bb6\u8a2d\u8a08\u597d\u7684\u65b7\u8a5e\u7279\u5fb5\u53c3\u6578\uff0c\u6240\u4ee5\u6211\u5011\u901a\u5e38\u53ea\u80fd\u4f7f\u7528\u73fe\u6210\u7684\uff0c\u7121\u6cd5\u5c08\u9580 \u6bcf\u500b\u55ae\u5b57(word)\u5373\u662f\u4e00\u500b\u8a5e\uff0c\u5177\u6709\u5b8c\u6574\u610f\u7fa9\uff0c\u800c\u4e14\u6bcf\u500b\u55ae\u5b57\u9593\u90fd\u4ee5\u7a7a\u767d\u505a\u5340\u9694\uff0c\u4f46\u662f\u5728 \u70ba\u4e86\u9054\u6210\u4ee5\u5b57\u5143\u5c64\u7d1a\u70ba\u8f38\u5165\u7684\u76ee\u7684\uff0c\u6211\u5011\u8a2d\u8a08\u53e6\u4e00\u500b\u65b0\u7684\u67b6\u69cb\uff0c\u5c07 parser \u90e8\u5206\u505a\u66ff\u63db\uff0c</td></tr><tr><td colspan=\"2\">\u6839\u64da\u5408\u6210\u7684\u9700\u6c42\u4f86\u8a2d\u8a08\u3002\u518d\u52a0\u4e0a\u6a19\u8a3b\u7684\u904e\u7a0b\u4e2d\u4e0d\u540c\u7684\u6a19\u8a3b\u4eba\u54e1\u53ef\u80fd\u6703\u5c0d\u540c\u4e00\u53e5\u8a71\u7522\u751f\u4e0d \u4e2d\u6587\u6587\u672c\u88e1\uff0c\u8a5e\u8207\u8a5e\u4e4b\u9593\u662f\u4e0d\u6703\u6709\u7a7a\u767d\u505a\u70ba\u5340\u9694\u7684\u3002\u56e0\u6b64\u5728\u4e2d\u6587\u7684 NLP \u4e2d\uff0c\u70ba\u4e86\u8b93\u96fb\u8166 \u4e3b\u8981\u662f\u5c0d\u5b57\u5143\u8a9e\u610f\u3001\u6587\u6cd5\u89d2\u8272\u7b49\u8cc7\u8a0a\u5229\u7528 word2vec \u4f86\u7522\u751f\u3002\u5b57\u5143\u6642\u9593\u524d\u5f8c\u8cc7\u8a0a\u5247\u4f7f\u7528</td></tr><tr><td colspan=\"2\">\u4e00\u81f4\u7684\u6a19\u8a3b\u65b9\u5f0f\uff0c\u9019\u5c0e\u81f4\u6a5f\u5668\u4e0d\u597d\u5b78\u7fd2\u3002\u6240\u4ee5\u70ba\u4e86\u89e3\u6c7a\u9019\u4e9b\u554f\u984c\uff0c\u6211\u5011\u60f3\u8981\u8b93\u6a5f\u5668\u80fd\u5920 \u80fd\u5920\u5206\u8fa8\u6587\u672c\u4e2d\u7684\u8a5e\u7fa9\uff0c\u5c31\u5fc5\u9808\u5148\u6b63\u78ba\u7684\u5c07\u8a5e\u5340\u9694\u958b\u4f86\uff0c\u624d\u80fd\u9032\u4e00\u6b65\u767c\u5c55\u51fa\u76f8\u95dc\u6f14\u7b97\u6cd5\u3002 RNNLM \u64f7\u53d6\uff0c\u6700\u5f8c\u5f62\u6210\u4e00\u500b\u65b0\u7684\u6587\u8108\u8a0a\u606f\u5168\u90e8\u662f\u7531\u6a5f\u5668\u81ea\u884c\u5b78\u7fd2\u672a\u6a19\u8a3b\u8a9e\u6599\u6240\u7522\u751f\u51fa</td></tr><tr><td colspan=\"2\">\u81ea\u52d5\u53bb\u5b78\u7fd2\u6bcf\u4e00\u500b\u5b57\u5143\u7684\u5c6c\u6027\u8207\u5b57\u5143\u4e32\u9593\u7684\u95dc\u4fc2\u3002\u5c31\u50cf\u6211\u5011\u4eba\u5728\u95b1\u8b80\u6587\u7ae0\u6642\u4e0d\u6703\u5148\u5c0d\u6587 \u4f8b\u5982\u6a5f\u5668\u7ffb\u8b6f\u3001\u8cc7\u8a0a\u6aa2\u7d22\u8207\u64f7\u53d6\u3001\u8a9e\u8a00\u5206\u6790\u3001\u8a9e\u97f3\u8fa8\u8b58\u548c\u5408\u6210\uff0c\u70ba\u6b64\u767c\u5c55\u51fa\u65b7\u8a5e\u5668\u4f86\u4f7f \u4f86\u7684\uff0c\u9019\u6a23\u80fd\u9054\u6210\u6211\u5011\u60f3\u4f7f\u7528\u5b57\u5143\u5c64\u7d1a\u7684\u76ee\u7684\u4e5f\u80fd\u907f\u958b\u50b3\u7d71 parser \u7684\u554f\u984c\u3002\u4ee5\u4e0b\u70ba\u65b0\u7cfb</td></tr><tr><td colspan=\"2\">\u7ae0\u505a\u65b7\u8a5e\u53ca POS \u6a19\u8a3b\uff0c\u6211\u5011\u4e5f\u662f\u4e00\u500b\u5b57\u4e00\u500b\u5b57\u8b80\u904e\u53bb\uff0c\u7136\u5f8c\u900f\u904e\u524d\u5f8c\u6587\u7684\u62c6\u89e3\u53bb\u7406\u89e3\u5b57 \u7528\uff0c\u800c\u65b7\u8a5e\u5668\u4e3b\u6d41\u4f7f\u7528 CRF \u8655\u7406\u8f38\u5165\u6587\u672c\uff0c\u900f\u904e\u8a13\u7df4 CRF \u5c0d\u8f38\u5165\u53e5\u5b50\u505a\u731c\u6e2c\u5224\u65b7\u5169\u500b \u7d71\u67b6\u69cb\u4ee5\u53ca\u65b0\u65b9\u6cd5\u7684\u8a73\u7d30\u8aaa\u660e\u3002</td></tr><tr><td colspan=\"2\">\u7fa9\u3001\u8a5e\u6027\u3002\u6240\u4ee5\u6211\u5011\u80fd\u4e0d\u80fd\u60f3\u4e00\u500b\u65b9\u6cd5\u8ddf\u4eba\u5728\u95b1\u8b80\u7684\u6642\u5019\u4e00\u6a23\uff0c\u5229\u7528\u4ee5\u5b57\u70ba\u55ae\u5143\u9019\u7a2e\u6bd4 \u5b57\u4e2d\u9593\u662f\u5426\u70ba\u65b7\u9ede\uff0c\u8a13\u7df4\u65b9\u6cd5\u8981\u4f9d\u7167\u5c08\u5bb6\u8a2d\u8a08\u65b7\u8a5e\u53c3\u6578\u53bb\u5b78\u7fd2\u600e\u9ebc\u9810\u6e2c\u8a5e\u65b7\u9ede\u3002\u4f46\u8a5e\u65b7 \u9996\u5148\uff0c\u5716 2 \u70ba\u6211\u5011\u7684\u65b0\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u67b6\u69cb\uff0c\u5c07\u8a9e\u6599\u7d93\u904e\u6b63\u898f\u5316\uff0c\u6587\u672c\u8f49\u62fc\u97f3\u65b9\u9762\uff0c</td></tr><tr><td colspan=\"2\">\u8f03\u7c21\u55ae\u7684\u65b9\u5f0f\uff0c\u907f\u514d\u4f7f\u7528 parser \u505a\u65b7\u8a5e\u548c POS \u6a19\u8a3b\u7684\u554f\u984c\uff0c\u4ee5\u6539\u5584\u539f\u6709\u7cfb\u7d71\u6c42\u53d6\u8a9e\u8a00\u7279 \u9ede\u8981\u53c3\u7167\u5df2\u7d93\u6a19\u8a3b\u597d\u7684\u8cc7\u6599\u5eab\u4f86\u5b78\u7fd2\uff0c\u6240\u4ee5\u5c0d\u65bc\u672a\u77e5\u8a5e\u7684\u932f\u8aa4\u7387\u7121\u6cd5\u6709\u6548\u964d\u4f4e\u3002\u5982\u4f55\u6539 \u7531\u65bc\u6211\u5011\u7684\u4e2d\u6587\u62fc\u97f3\u5b57\u5178\u4e26\u6c92\u6709\u6b78\u7d0d\u51fa\u54ea\u4e9b\u5b57\u61c9\u8a72\u4f55\u6642\u8b80\u7834\u97f3\u5b57\u7684\u8cc7\u8a0a\uff0c\u5728\u6b64\u9078\u7528\u7b2c\u4e00</td></tr><tr><td colspan=\"2\">\u5fb5\u7684\u7f3a\u9ede\u3002 \u5584\u4e2d\u6587\u65b7\u8a5e\u7684\u689d\u4ef6\u6a5f\u7387\u6a21\u578b\u76f8\u95dc\u8cc7\u6599\u53ef\u4ee5\u53c3\u7167(\u9ec3\u662d\u9298\uff0c2010)\uff0c\u7576\u4e2d\u6709\u63d0\u5230\u4f7f\u7528\u4e00\u500b\u7dda \u7d44\u62fc\u97f3\u4f86\u5c0d\u61c9\uff0c\u518d\u5c07\u62fc\u97f3\u8f49\u97f3\u7d20\u3002\u6211\u5011\u66f4\u52d5\u7684\u5730\u65b9\u662f\u5c07 parser \u62ff\u6389\uff0c\u4ee5 word2vec \u8207</td></tr><tr><td colspan=\"2\">\u8fd1\u5e74\u4f86\uff0c\u985e\u795e\u7d93\u7db2\u8def(DNNs) (Licstar, 2013)(\u5728 NLP \u65b9\u9762\u6709\u8d8a\u4f86\u8d8a\u591a\u7684\u7814\u7a76\uff0c\u63d0\u51fa\u5f88 \u6027\u7684 CRF \u4f86\u9054\u6210\u66f4\u6e96\u78ba\u7684\u4e2d\u6587\u65b7\u8a5e\uff0c\u5982\u679c\u5206\u5272\u51fa\u7684\u8a5e\u8207\u76f8\u5c0d\u61c9\u6a19\u6e96\u8a9e\u6599\u5eab\u7684\u8a5e\u4e0d\u540c\u6642\uff0c RNNLM \u505a\u66ff\u63db\uff0c\u4e3b\u8981\u662f\u7528 word2vec \u8207 RNNLM \u6c42\u53d6\u5b57\u5143\u8a9e\u610f\u3001\u6587\u6cd5\u89d2\u8272\u8cc7\u8a0a\u8207\u5b57\u5143\u6642</td></tr><tr><td colspan=\"2\">\u591a\u5b57\u5143\u5c64\u7d1a\u7684\u8a9e\u8a00\u6a21\u578b\uff0c\u56e0\u70ba\u4f7f\u7528 DNNs \u53ef\u4ee5\u5efa\u7acb\u66f4\u597d\u7684\u6a21\u578b\uff0c\u4e26\u4e14\u5b78\u7fd2\u5927\u91cf\u7121\u9808\u6a19\u8a3b \u900f\u904e\u8a72\u8a5e\u548c\u524d\u5f8c\u65b7\u8a5e\u7684\u91cd\u7d44\uff0c\u53ef\u6c42\u51fa\u66f4\u9069\u7576\u7684\u65b7\u8a5e\u3002 \u9593\u524d\u5f8c\u8cc7\u8a0a\u7576\u4f5c\u6587\u8108\u8a0a\u606f\uff0c\u4ee5\u5408\u6210\u51fa\u65b0\u7684\u8a9e\u97f3\u3002</td></tr><tr><td colspan=\"2\">\u7684\u6587\u672c\u8cc7\u6599\uff0c\u6240\u4ee5\u57fa\u65bc\u4f7f\u7528\u985e\u795e\u7d93\u7db2\u8def\u7684\u65b9\u6cd5\u80fd\u5920\u6709\u6548\u7684\u6e1b\u5c11 POS \u6a19\u8a3b\u9019\u985e\u7684\u8a9e\u8a00\u7279\u5fb5</td></tr><tr><td colspan=\"2\">\u53c3 \u6578 \u8a2d \u8a08 \u5de5 \u7a0b (feature engineering) \uff1b \u4f8b \u5982 (Greff, Srivastava, Koutn\u00edk, Steunebrink, &</td></tr></table>", | |
| "html": null | |
| }, | |
| "TABREF1": { | |
| "text": "", | |
| "num": null, | |
| "type_str": "table", | |
| "content": "<table><tr><td/><td>\u57fa\u65bc\u5b57\u5143\u968e\u5c64\u4e4b\u8a9e\u97f3\u5408\u6210\u7528\u6587\u8108\u8a0a\u606f\u64f7\u53d6</td><td>\u9673\u51a0\u5b8f \u7b49 77 \u9673\u51a0\u5b8f \u7b49</td></tr><tr><td colspan=\"3\">CBOW \u6a21\u578b\u53ef\u4ee5\u7406\u89e3\u70ba\u4ee5\u4e0a\u4e0b\u6587\u6c7a\u5b9a\u7576\u524d\u8a5e\u51fa\u73fe\u7684\u6a5f\u7387\uff0c\u5728 CBOW \u6a21\u578b\u4e2d\uff0c\u4e0a\u4e0b 3.3 \u5b57\u5143\u6642\u5e8f\u72c0\u614b\u4e4b\u6587\u8108\u8a0a\u606f\u64f7\u53d6 \u8868 2</td></tr><tr><td colspan=\"3\">\u6587\u6240\u6709\u7684\u8a5e\u5c0d\u7576\u524d\u8a5e\u51fa\u73fe\u6a5f\u7387\u7684\u5f71\u97ff\u7684\u6b0a\u91cd\u662f\u4e00\u6a23\u7684\uff0c\u56e0\u6b64\u53eb CBOW \u6a21\u578b\u3002\u5982\u5728\u888b\u5b50\u4e2d \u5728\u6211\u5011\u63d0\u51fa\u7684\u5b57\u5143\u6982\u5ff5\u4e0b\uff0c\u6211\u5011\u5e0c\u671b\u80fd\u5f9e\u6587\u672c\u4e2d\u7372\u5f97\u5b57\u5143\u5728\u7576\u524d\u8a9e\u53e5\u4e2d\u7684\u72c0\u614b\uff0c\u8b93\u6211\u5011</td></tr><tr><td colspan=\"3\">\u53d6\u8a5e\uff0c\u53d6\u51fa\u6578\u91cf\u8db3\u5920\u7684\u8a5e\u5c31\u53ef\u4ee5\u4e86\uff0c\u81f3\u65bc\u53d6\u51fa\u7684\u5148\u5f8c\u9806\u5e8f\u662f\u7121\u95dc\u7dca\u8981\u7684\u3002Skip-gram \u6a21\u578b \u5b78\u7fd2\u6587\u7ae0\u7684\u8108\u7d61\uff0c\u53ef\u4ee5\u5f9e\u9019\u53e5\u9810\u6e2c\u4e0b\u53e5\u53ef\u80fd\u70ba\u4f55\u3002\u70ba\u4e86\u5b8c\u6210\u9019\u529f\u80fd\uff0c\u905e\u8ff4\u795e\u7d93\u7db2\u8def(RNN)</td></tr><tr><td colspan=\"3\">\u8207 CBOW \u6a21\u578b\u6b63\u597d\u76f8\u53cd\uff0cSkip-gram \u6a21\u578b\u4f7f\u7528\u7576\u524d\u7684\u8a5e\u5411\u91cf\u4f86\u9810\u6e2c\u8a72\u8a5e\u4e4b\u524d\u548c\u4e4b\u5f8c\u5404 K \u53ef\u80fd\u662f\u500b\u4e0d\u932f\u7684\u9078\u64c7\uff0c\u4f8b\u5982(Mikolov & Zweig, 2012)\u57fa\u65bc RNNLM \u4e0a\u4e0b\u6587\u76f8\u95dc\u6027\u7684\u7814\u7a76\u4e5f</td></tr><tr><td colspan=\"3\">\u500b\u8a5e\u7684\u6a5f\u7387\uff0c\u8ddd\u96e2\u7576\u524d\u8a5e\u8d8a\u9060\u7684\u8a5e\u548c\u7576\u524d\u8a5e\u7684\u76f8\u95dc\u8a5e\u7684\u76f8\u95dc\u6027\u8d8a\u5c0f\uff0c\u6240\u4ee5\u53ef\u4ee5\u7d66\u4e88\u9019\u4e9b \u6307\u51fa\u4f7f\u7528\u905e\u8ff4\u985e\u795e\u7d93\u7db2\u8def\u6a21\u578b\u9032\u884c\u8a13\u7df4\u80fd\u5f9e\u96b1\u85cf\u5c64\u4e2d\u9023\u7e8c\u7684\u8f38\u51fa\u5411\u91cf\u7372\u5f97\u5b57\u8a5e\u5728\u53e5\u5b50\u4e2d</td></tr><tr><td colspan=\"3\">\u8a5e\u4ee5\u8f03\u5c0f\u7684\u6b0a\u91cd\u3002\u8a73\u7d30\u5167\u5bb9\u53ef\u4ee5\u53c3\u95b1(Mikolov, Sutskever, Chen, Corrado & Dean, 2013)\u8a5e \u7684\u72c0\u614b\u3002</td></tr><tr><td colspan=\"3\">\u8a9e\u5206\u4f48\u8868\u793a\u8207\u7d44\u5408\u6027\uff0c\u5167\u6587\u4e2d\u63d0\u5230 word2vec \u63d0\u4f9b\u4e00\u500b\u66f4\u6e96\u78ba\u4ee5\u53ca\u975e\u5e38\u7c21\u55ae\u7684\u8a13\u7df4\u65b9\u5f0f\uff0c \u672c\u6587\u662f\u4ee5 Mikolov \u6539\u826f\u7684 RNNLM \u4f86\u9032\u884c\uff0c\u905e\u8ff4\u5f0f\u985e\u795e\u7d93\u7db2\u8def\u5305\u542b\u8f38\u5165\u5c64(input</td></tr><tr><td colspan=\"3\">\u80fd\u5920\u5927\u5927\u7684\u6539\u5584\u5b78\u7fd2\u55ae\u8a5e\u548c\u77ed\u8a9e\u7684\u76f8\u95dc\u6027\u3002 layer)\u3001\u96b1\u85cf\u5c64(hidden layer)\u3001\u8f38\u51fa\u5c64(output layer)\u548c\u985e\u5225\u5c64(class layer)\u3002\u800c U\u3001V\u3001W</td></tr><tr><td colspan=\"3\">\u56e0\u6b64\u6211\u5011\u5229\u7528 word2vec \u7684\u7279\u6027\uff0c\u5c07\u6240\u6709\u7684\u5b57\u8f49\u5230\u5411\u91cf\u7a7a\u9593\uff0c\u4e26\u9032\u884c\u5206\u985e\uff0c\u5c07\u6027\u8cea\u76f8 \u548c C \u70ba\u5404\u5c64\u7684\u6b0a\u91cd\u3002w(t)\u70ba\u8f38\u5165\uff0ct \u4f9d\u6642\u9593\u6392\u5e8f\u70ba\u70ba 1 \u5230 N\uff0c\u4e5f\u662f RNN \u7684\u6b0a\u91cd\uff0cs(t)\u70ba\u96b1</td></tr><tr><td colspan=\"3\">\u8fd1\u7684\u5b57\u7fa4\u805a\u5728\u4e00\u8d77\u3002\u8868 1 \u70ba\u6211\u5011\u5728\u4e00\u524d\u7f6e\u5be6\u9a57\u4e2d\uff0c\u5229\u7528 word2vec \u5206\u6790 Chinese Gigaword \u85cf\u5c64\u8f38\u51fa\u4e5f\u5c31\u662f\u795e\u7d93\u5143(neurons)\u7684\u503c\u4e5f\u662f\u5b83\u7684 state\uff0cy(t)\u70ba\u8f38\u51fa\u9808\u8207\u8f38\u5165\u540c\u7dad\u5ea6\u3002\u800c c(t)</td></tr><tr><td colspan=\"3\">\u8a9e\u6599\u5eab\u7684\u7d50\u679c\uff0c\u6211\u5011\u53ef\u4ee5\u770b\u898b class1 \u662f\u9b5a\u90e8\u9996\u88ab\u5206\u70ba\u4e00\u985e\uff0c\u88e1\u9762\u90fd\u70ba\u9b5a\u985e\u751f\u7269\u5c45\u591a\uff0cclass2 \u70ba\u985e\u5225\u5c64\uff0cMikolov \u63d0\u51fa\u8f38\u51fa\u5c64\u5206\u89e3\u53ef\u4ee5\u964d\u4f4e\u8a9e\u8a00\u6a21\u578b\u4e2d\u7684\u904b\u7b97\u8907\u96dc\u5ea6\uff0c\u4f7f\u8a13\u7df4\u6548\u7387\u63d0</td></tr><tr><td colspan=\"3\">\u628a\u6728\u90e8\u9996\u7684\u690d\u7269\u985e\u5206\u5728\u4e00\u8d77\uff0cclass3 \u5247\u662f\u628a\u89aa\u621a\u7a31\u8b02\u6b78\u985e\u5728\u4e00\u8d77\uff0cclass5 \u662f\u50f9\u503c\u8861\u91cf\u55ae\u4f4d\uff0c \u9ad8\u3002\u5716 4 \u70ba Mikolov \u6539\u826f\u7684 RNNLM \u67b6\u69cb\u3002</td></tr><tr><td colspan=\"3\">class7 \u662f\u5c07\u8a9e\u52a9\u8a5e\u5206\u5728\u4e00\u8d77\uff0c\u800c class4 \u8207 class6 \u5206\u5225\u662f\u82f1\u8a9e\u548c\u65e5\u8a9e\uff0c\u5c07\u4e0d\u540c\u8a9e\u8a00\u5404\u81ea\u70ba</td></tr><tr><td colspan=\"3\">\u4e00\u985e\uff0c\u7531\u6b64\u53ef\u77e5\u5c07\u5b57\u5143\u8f49\u63db\u5230\u5411\u91cf\u7a7a\u9593\u8655\u7406\u5206\u6790\uff0c\u53ef\u4ee5\u6709\u6548\u7684\u5c07\u5b57\u5143\u8a9e\u610f\u3001\u5b57\u5143\u6587\u6cd5\u89d2</td></tr><tr><td colspan=\"2\">\u8272\u7b49\u5c6c\u6027\u64f7\u53d6\u51fa\u4f86\u3002</td></tr><tr><td>\u8868 1Class</td><td>Characters</td></tr><tr><td/><td colspan=\"2\">\u986a\u3001\u4c81\u3001\u4c97\u3001\u4c7b\u3001\u554d\u3001\u586d\u3001\u72ae\u3001\u86c9\u3001\u877d\u3001\u87fe\u3001\u8e84\u3001\u9b5f\u3001\u9b69\u3001\u9b6e\u3001\u9b81\u3001\u9b8b\u3001</td></tr><tr><td>1</td><td colspan=\"2\">\u9b8e\u3001\u9b9f\u3001\u9ba8\u3001\u9baa\u3001\u9bab\u3001\u9bad\u3001\u9bc9\u3001\u9bca\u3001\u9bd4\u3001\u9bd6\u3001\u9be1\u3001\u9be2\u3001\u9be7\u3001\u9bf0\u3001\u9bf7\u3001\u9c06\u3001</td></tr><tr><td/><td colspan=\"2\">\u9c15\u3001\u9c31\u3001\u9c37\u3001\u9c39\u3001\u9c3e\u3001\u9c47\u3001\u9c48\u3001\u9c52\u3001\u9c56\u3001\u9c58\u3001\u9c77\u3001\u9c78\u3001\u9d37</td></tr><tr><td>2</td><td>\u674f\u3001\u675e\u3001\u6777\u3001\u6787\u3001\u67b3\u3001\u67b8\u3001\u67d1\u3001\u67da\u3001\u67ff\u3001\u682a</td></tr><tr><td/><td colspan=\"2\">\u4e08\u3001\u4e48\u3001\u5144\u3001\u5152\u3001\u53d4\u3001\u592d\u3001\u5973\u3001\u59bb\u3001\u59e8\u3001\u59ea\u3001\u5a36\u3001\u5a66\u3001\u5a7f\u3001\u5ab3\u3001\u5ac1\u3001\u5ac2\u3001</td></tr><tr><td colspan=\"3\">3 4 \u905e\u8ff4\u795e\u7d93\u7db2\u8def\u6700\u5927\u7684\u512a\u52e2\u5728\u65bc\uff0c\u53ef\u4ee5\u771f\u6b63\u5145\u5206\u7684\u5229\u7528\u6240\u6709\u4e0a\u6587\u8a0a\u606f\u4f86\u9810\u6e2c\u4e0b\u4e00\u500b\u8a5e\uff0c\u4e0d \u5b50\u3001\u5b7a\u3001\u5f1f\u3001\u621a\u3001\u6b72\u3001\u6bcd\u3001\u7236\u3001\u7467\u3001\u7525\u3001\u7537\u3001\u8205\u3001\u89aa\u3001\u9f61 \u5716 4. Mikolov \u6539\u826f\u7684 RNNLM \u67b6\u69cb Peanuts\u3001Pentavision\u3001Peranakan\u3001percussion\u3001Perfectv\u3001Persona\u3001phantasy\u3001 [Figure 4. The block diagram of the RNNLM model.] Phantom\u3001 Phoenix\u3001 Pick</td></tr><tr><td colspan=\"3\">5 \u50cf\u5176\u5b83\u795e\u7d93\u7db2\u8def\u53ea\u80fd\u4e00\u6b21\u770b n \u500b\u5b57\uff0c\u53ea\u80fd\u5f9e\u524d n \u500b\u5b57\u4f86\u9810\u6e2c\u4e0b\u4e00\u500b\u8a5e\uff0c\u7c21\u55ae\u4f86\u8aaa RNN \u5104\u3001\u5143\u3001\u5146\u3001\u5343\u3001\u571c\u3001\u5e63\u3001\u65b0\u3001\u767e\u3001\u7f8e\u3001\u80d8\u3001\u842c\u3001\u9214\u3001\u9296\u3001\u938a\u3001\u9918</td></tr><tr><td colspan=\"3\">\u30e0\u30c3\u30af\u3001\u3072\u308d\u307f\u3001\u30d9\u30a3\u30f3\u3001\u30c8\u30e8\u30bf\u3001\u3042\u3084\u304b\u3001\u30bf\u30b1\u30b7\u3001\u30d8\u30ea\u30aa\u30b9\u3001\u30d6\u30eb \u5c31\u662f\u4e00\u500b\u6709\u96b1\u85cf\u5c64\u7684\u81ea\u6211\u76f8\u9023\u7db2\u8def\uff0c\u96b1\u85cf\u5c64\u540c\u6642\u63a5\u6536\u4f86\u81ea t \u6642\u523b\u7684\u8f38\u5165\u548c t-1 \u6642\u523b\u7684\u96b1\u85cf</td></tr><tr><td colspan=\"3\">6 \u5c64\u8f38\u51fa\u505a\u70ba\u8f38\u5165\uff0c\u9019\u4f7f\u5f97 RNN \u5177\u6709\u77ed\u671f\u8a18\u61b6\u80fd\u529b\uff0c\u80fd\u5920\u5b78\u7fd2\u5230\u8f03\u9577\u6642\u9593\u7684\u6587\u7ae0\u8108\u7d61\uff0c \u30fc\u30b9\u3001\u30bf\u30a4\u30c8\u30ed\u30fc\u30d7\u3001\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb</td></tr><tr><td colspan=\"3\">7 \u5efa\u7acb\u8d77\u8a9e\u8a00\u6a21\u578b\u3002 \u54ce\u3001\u54fc\u3001\u5509\u3001\u5514\u3001\u552c\u3001\u5565\u3001\u5582\u3001\u55b2\u3001\u55d0\u3001\u55ef\u3001\u563b\u3001\u563f\u3001\u5662\u3001\u5b24\u3001\u6b38\u3001\u7821</td></tr><tr><td colspan=\"3\">\u8868 2 \u70ba\u6211\u5011\u5728\u4e00\u524d\u7f6e\u5be6\u9a57\u4e2d\uff0c\u5229\u7528 Chinese Gigaword \u8a9e\u6599\u5eab\u8207 RNNLM \u5b78\u7fd2\u5b57\u5143\u6642</td></tr><tr><td colspan=\"3\">\u4e5f\u56e0\u70ba word2vec \u5b83\u6240\u8a13\u7df4\u51fa\u4f86\u7684\u5b57\u5411\u91cf\u7a7a\u9593\u80fd\u5920\u8868\u9054\u5b57\u7684\u5c6c\u6027\uff0c\u4e26\u4e14\u5b83\u80fd\u5c07\u8a13\u7df4\u51fa\u4f86\u7684 \u5e8f\u95dc\u4fc2\u7684\u7d50\u679c\uff0c\u6211\u5011\u53ef\u4ee5\u767c\u73fe RNNLM \u6a21\u578b\u7522\u751f\u51fa\u4f86\u7684\u53e5\u5b50\u4e0d\u7ba1\u5728\u8a9e\u610f\uff0c\u6642\u9593\u9806\u5e8f\uff0c\u6216</td></tr><tr><td colspan=\"3\">\u5b57\u5411\u91cf\u9032\u884c\u6392\u5217\uff0c\u8b93\u5c6c\u6027\u63a5\u8fd1\u7684\u5b57\u5206\u985e\u5728\u4e00\u8d77\uff0c\u6240\u4ee5\u4f7f\u7528\u5b83\u4f86\u4ee3\u66ff\u50b3\u7d71 parser \u53ef\u80fd\u662f\u884c \u662f\u6d41\u66a2\u5ea6\u800c\u8a00\u90fd\u883b\u8cbc\u8fd1\u4e00\u822c\u65b0\u805e\u7684\u6587\u5b57\u3002</td></tr><tr><td>\u5f97\u901a\u7684\u3002</td><td/></tr><tr><td/><td colspan=\"2\">-gram</td></tr><tr><td/><td>word2vec models]</td></tr></table>", | |
| "html": null | |
| }, | |
| "TABREF2": { | |
| "text": "", | |
| "num": null, | |
| "type_str": "table", | |
| "content": "<table><tr><td>80</td><td colspan=\"4\">\u57fa\u65bc\u5b57\u5143\u968e\u5c64\u4e4b\u8a9e\u97f3\u5408\u6210\u7528\u6587\u8108\u8a0a\u606f\u64f7\u53d6 \u57fa\u65bc\u5b57\u5143\u968e\u5c64\u4e4b\u8a9e\u97f3\u5408\u6210\u7528\u6587\u8108\u8a0a\u606f\u64f7\u53d6</td><td/><td>79 \u9673\u51a0\u5b8f \u7b49 81 \u9673\u51a0\u5b8f \u7b49</td></tr><tr><td colspan=\"7\">\u8868 3. \u8a13\u7df4\u8a9e\u6599\u8cc7\u6599\u8868 [Table 3. Statistics of the speech corpus for speech synthesis experiments.] \u8868 6. \u65b0\u820a\u7cfb\u7d71\u7d14\u4e2d\u6587\u8072\u97f3 MOS \u4e3b\u89c0\u5206\u6578\u6bd4\u8f03 4.1.3 \u8a9e\u97f3\u5408\u6210\u8a2d\u5b9a [Table 6. Comparison of the MOS scores of the conventional and the \u672c\u7814\u7a76\u7684\u4e2d\u82f1\u593e\u96dc\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u4f7f\u7528\"NTUT Audiobook Corpus Vol.2\"\u8a9e\u6599\u548c HTS \u5408 proposed approaches on pure Chinese sentence synthesis]</td></tr><tr><td colspan=\"7\">\u4e2d\u6587\u8a9e\u6599 \u6210\uff1b\u9996\u5148\u6211\u5011\u4e2d\u6587\u548c\u82f1\u6587\u8072\u97f3\u7de8\u78bc\u7d71\u4e00\u4f7f\u7528 X-SAMPA \u7de8\u78bc\u70ba\u6a19\u6e96\u3002\u5728\u8a13\u7df4\u8a9e\u97f3\u5408\u6210\u6a21 \u4e2d\u82f1\u593e\u96dc\u8a9e\u6599-CE \u7d14\u82f1\u6587 \u578b\u6642\uff0c\u6240\u6709\u9304\u97f3\u7686\u70ba 48KHz\uff0c\u8072\u5b78\u7279\u6027\u6211\u5011\u53d6 34 \u7dad\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578(MFCCs)\uff0c\u97f3\u8abf(pitch) \u7d14\u4e2d\u6587 \u50b3\u7d71\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71 \u65b0\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71</td></tr><tr><td colspan=\"7\">\u6587\u672c\u5167\u5bb9\u51fa\u8655 \u8f2a\u5ed3\u6bcf 5 \u6beb\u79d2\u81f3 25 \u6beb\u79d2\u70ba\u4e00\u97f3\u6846(frame)\u9577\u5ea6\uff0c\u6700\u5f8c\u6bcf\u500b\u97f3\u7d20(phone)\u6211\u5011\u4f7f\u7528 5 \u500b\u72c0\u614b \u751f\u547d\u79d1\u5b78\u5927\u5e2b\uff1a\u907a\u50b3\u5b78\u4e4b\u7236 \u7dda\u4e0a\u6587\u672c CMU \u81ea\u7136\u5ea6\u8a55\u5206 3.09 3.44</td></tr><tr><td colspan=\"4\">\u5b5f\u5fb7\u723e\u7684\u6545\u4e8b(\u5f35\u6587\u4eae\u8457) (state)\u7684 HMMs \u4f86\u8a13\u7df4\u3002 \u76f8\u4f3c\u5ea6\u8a55\u5206 3.25</td><td>(\u5de5\u7814\u9662\u63d0\u4f9b)</td><td>3.38</td></tr><tr><td colspan=\"2\">\u8a13\u7df4\u8a9e\u6599\u7e3d\u53e5\u6578 4.2 \u8a55\u4f30\u65b9\u6cd5 \u53ef\u7406\u89e3\u5ea6\u8a55\u5206</td><td>\u7d04 4800 \u53e5</td><td>4.27</td><td>\u7d04 3500 \u53e5</td><td>4.27</td><td>\u7d04 990 \u53e5</td></tr><tr><td colspan=\"7\">\u8a13\u7df4\u8a9e\u6599\u6bcf\u53e5\u8a5e\u6578 \u7cfb\u7d71\u504f\u597d\u7684\u8a55\u4f30\u65b9\u5f0f\u662f\u50b3\u7d71\u4f7f\u7528 parser \u7684\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u8207\u6211\u5011\u63d0\u51fa\u4f7f\u7528\u5b57\u5143\u968e\u5c64\u63d0\u53d6\u7279 20-35 \u8a5e 10-30 \u8a5e 5-15 \u55ae\u5b57</td></tr><tr><td colspan=\"7\">\u8a13\u7df4\u8a9e\u6599\u6642\u9593\u9577\u5ea6 \u5fb5\u7684\u65b0\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u4f86\u505a\u6bd4\u8f03\u3002\u6211\u5011\u4ee5\u8072\u97f3\u7684\u76f8\u4f3c\u5ea6\u3001\u81ea\u7136\u5ea6\u548c\u53ef\u7406\u89e3\u5ea6\u7684\u8a55\u4f30\u65b9\u5f0f\uff0c \u7d04 172 \u5206\u9418 \u7d04 201 \u5206\u9418 \u7d04 79 \u5206\u9418</td></tr><tr><td colspan=\"7\">\u5c07\u6e2c\u8a66\u97f3\u6a94\u7d66 10 \u4f4d\u4ee5\u570b\u8a9e\u70ba\u6bcd\u8a9e\u7684\u4eba\u58eb\u9032\u884c\u8a55\u5206\uff0c\u65b0\u820a\u7cfb\u7d71\u504f\u597d\u5ea6\u6e2c\u8a66\u70ba 2 \u9078 1 \u65b9\u5f0f\uff0c</td></tr><tr><td colspan=\"7\">\u65b0 \u56e0\u6b64\u61c9\u7528 RNNLM \u7684\u7279\u6027\u4f86\u8a13\u7df4\u6587\u672c\uff0c\u53ef\u4ee5\u5b78\u7fd2\u524d\u9762\u4ee5\u53ca\u7576\u524d\u770b\u904e\u7684\u53e5\u5b50\u4f86\u731c\u6e2c\u4e0b\u4e00\u7684 \u5b57\u6216\u662f\u4e0b\u4e00\u53e5\u6703\u662f\u751a\u9ebc\uff0c\u4e5f\u56e0\u70ba\u905e\u8ff4\u795e\u7d93\u7db2\u8def\u80fd\u5920\u9019\u6a23\u5206\u6790\u53e5\u5b50\u524d\u5f8c\u6642\u9593\u95dc\u4fc2\uff0c\u6240\u4ee5\u80fd \u5920\u77e5\u9053\u6574\u500b\u8a9e\u53e5\u7684\u8108\u7d61\uff0c\u800c\u9019\u525b\u597d\u8207\u50b3\u7d71\u6587\u672c\u5206\u6790\u7684\u6642\u9593\u9806\u5e8f\u8cc7\u8a0a\u76f8\u4f3c\uff0c\u56e0\u6b64\u6211\u5011\u5c07\u5176 \u7528\u5728\u6211\u5011\u7684\u65b0\u67b6\u69cb\u4e0a\u770b\u770b\u80fd\u5426\u771f\u7684\u6709\u7528\u3002 \u70ba\u6a19\u6e96 A/B/X \u6e2c\u8a66\uff0c\u4e0d\u5b58\u5728\u5169\u8005\u7686\u597d\uff1b\u800c\u65b0\u820a\u7cfb\u7d71\u8a55\u5206\u63a1\u5e73\u5747\u4e3b\u89c0\u503c\u5206\u6578(mean opinion 4.1.2 \u6587\u8108\u8a0a\u606f\u6c42\u53d6\u65b9\u6cd5\u8207\u8a2d\u5b9a \u65b0\u65b9\u6cd5\u8207\u820a\u65b9\u6cd5\u4e2d\u53ea\u6709\u524d\u7d1a\u6587\u672c\u5206\u6790\u4e0d\u540c\uff0c\u5f8c\u7d1a\u7684\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71(HMM-based Speech Synthesis System\uff0cHTS)\u7684\u8a2d\u5b9a\uff0c\u5169\u8005\u5b8c\u5168\u76f8\u540c\uff1b\u820a\u7cfb\u7d71\u6587\u8108\u8a0a\u606f\u4f9d\u7136\u63a1\u7528 parser \u4f86\u5206\u6790 \u6587\u672c\uff0c\u65b0\u7cfb\u7d71\u7684\u6587\u8108\u8a0a\u606f\u5247\u662f\u53bb\u9664 parser \u7522\u751f\u7684\u8cc7\u8a0a\uff0c\u6539\u4f7f\u7528\u5b57\u5143\u5c64\u7d1a\u7684 word2vec\u3001 \u4e2d\u82f1\u593e\u96dc \u50b3\u7d71\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71 \u65b0\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71 score,MOS)\u4f86\u8a55\u4f30\uff0c\u5176\u8a55\u5206\u65b9\u5f0f\u70ba 1~5 \u5206\u3002\u8868 5 \u70ba\u6e2c\u8a66\u97f3\u6a94\u8a2d\u5b9a\u3002 \u8868 5. \u6e2c\u8a66\u97f3\u6a94\u8a2d\u5b9a \u5716 5. \u65b0\u820a\u7cfb\u7d71\u7d14\u4e2d\u6587\u504f\u597d\u6bd4\u8f03 \u81ea\u7136\u5ea6\u8a55\u5206 3.26 3.22 [Figure 5. Experimental results of the A/B/X preference test on [Table 5. Statistics of the synthesized speech database for pure Chinese sentence synthesis] \u76f8\u4f3c\u5ea6\u8a55\u5206 3.26 3.36 all evaluation experiments] RNNLM \u4f86\u5206\u6790\uff0c\u6211\u5011\u7528 word2vec \u5c07\u5b57\u5143\u6b78\u985e\u6210 64 \u985e\uff0cRNNLM \u96b1\u85cf\u5c64\u8a2d 256 \u7dad\uff0c\u4e26 \u70ba\u96b1\u85cf\u5c64\u4e2d\u6bcf\u500b neurons \u7684\u72c0\u614b\u8a2d\u5b9a\u9580\u6abb(threshold)\u628a\u8f38\u51fa\u91cf\u5316\u6210 0 \u6216 1\u3002\u5728\u554f\u984c\u96c6\u4e2d\u6211 NTUT Audiobook Corpus Vol.2 \u53ef\u7406\u89e3\u5ea6\u8a55\u5206 3.24 3.24</td></tr><tr><td colspan=\"7\">4. \u5be6\u9a57\u7d50\u679c\u8207\u5206\u6790 \u672c\u5be6\u9a57\u4e2d\uff0c\u76ee\u7684\u662f\u70ba\u4e86\u4f7f\u7528\u65b0\u8a9e\u8a00\u7279\u5fb5\u53d6\u4ee3\u73fe\u6709 parser \u7684\u65b7\u8a5e\u3001POS \u8cc7\u8a0a\uff0c\u7136\u5f8c\u7522\u751f\u65b0 \u7684\u6587\u8108\u8a0a\u606f\uff0c\u4ee3\u5165\u8a9e\u97f3\u5408\u6210\u7cfb\u7d71\u5408\u6210\u8a9e\u97f3\uff0c\u4e26\u8207\u50b3\u7d71\u65b9\u6cd5 parser \u505a\u6bd4\u8f03\uff1b\u65b0\u7cfb\u7d71\u8207\u820a\u7cfb \u7d71\u7684\u5dee\u5225\u53ea\u6709\u6587\u8108\u8a0a\u606f\u7684\u4e0d\u540c\uff0c\u820a\u7cfb\u7d71\u4f7f\u7528 parser \u505a\u6587\u672c\u7684\u5206\u6790\uff0c\u800c\u65b0\u7cfb\u7d71\u4f7f\u7528\u5b57\u5143\u5c64 \u7d1a\u7684 word2vec \u8207 RNNLM \u4f86\u5206\u6790\u6587\u672c\uff0c\u70ba\u6c42\u5176\u516c\u5e73\u6027\uff0c\u65b0\u820a\u7cfb\u7d71\u7686\u4f7f\u7528\u4e2d\u82f1\u593e\u96dc\u8a9e\u6599\uff0c \u8a13\u7df4\u8a9e\u53e5\u9577\u5ea6\u70ba\u6bb5\u843d\u8a13\u7df4\u3002\u800c\u6211\u5011\u8981\u6bd4\u8f03\u7684\u662f\u65b0\u820a\u7cfb\u7d71\u5408\u6210\u97f3\u6a94\u4e4b\u504f\u597d\uff0c\u9078\u7528\u7684\u5408\u6210\u8a9e \u5011\u53bb\u6389\u6709\u4f7f\u7528\u65b7\u8a5e\u548c POS \u7684\u9805\u76ee\uff0c\u7136\u5f8c\u6dfb\u52a0\u4f7f\u7528 word2vec \u548c RNNLM \u7522\u751f\u7684\u65b0\u9805\u76ee\uff0c \u7136\u5f8c\u5efa\u7acb\u65b0\u7684\u6c7a\u7b56\u6a39\u3002\u8868 4 \u70ba\u820a\u7cfb\u7d71\u8207\u65b0\u67b6\u69cb\u6240\u4f7f\u7528\u7684\u8a9e\u8a00\u7279\u5fb5\uff0c\u5176\u4e2d\u820a\u7cfb\u7d71\u4e2d\u7d05\u8272\u659c \u9ad4\u5b57\u90e8\u5206\u5728\u6211\u5011\u65b0\u67b6\u69cb\u4e2d\u5c07\u88ab\u5ee2\u9664\uff0c\u65b0\u52a0\u5165\u7684\u85cd\u8272\u659c\u9ad4\u5b57\u90e8\u5206\u70ba word2vec \u8207 RNNLM \u7684 \u6e2c\u8a66\u985e\u578b \u4e2d\u6587 \u4e2d\u82f1\u593e\u96dc \u7e3d\u53e5\u6578 60 \u53e5 4.4 \u5be6\u9a57\u8a0e\u8ad6 40 \u53e5 \u97f3\u6a94\u6578 10 \u500b \u5728\u672c\u6b21\u5be6\u9a57\u4e2d\u65b0\u7cfb\u7d71\u5927\u6982\u7372\u5f97 3 \u9ede\u591a\u5206\uff0c\u5728\u5e02\u9762\u4e0a\u7684 TTS \u5408\u6210\u7cfb\u7d71\u591a\u70ba 4 \u5206\u4ee5\u4e0a\uff0c\u96d6\u6bd4 20 \u500b \u4e0d\u4e0a\u5927\u516c\u53f8\u7684\u5408\u6210\u7cfb\u7d71\uff0c\u4e0d\u904e\u9019\u5206\u6578\u9084\u7b97\u662f\u5408\u7406\uff0c\u8b49\u660e\u4f7f\u7528\u5b57\u5143\u5c64\u7d1a\u7684\u6587\u8108\u8a0a\u606f\u5728\u8a9e\u97f3 \u8f38\u51fa\u53c3\u6578\u3002 \u6bcf\u53e5\u5b57\u6578 10-20 \u5b57 10-20 \u5b57 4.3 \u5be6\u9a57\u7d50\u679c \u5408\u6210\u4e0a\u662f\u53ef\u884c\u7684\u3002\u5c0d\u6b64\u6211\u5011\u731c\u6e2c\u65b0\u67b6\u69cb\u6703\u66f4\u597d\u7684\u539f\u56e0\uff0c\u662f\u56e0\u70ba\u50b3\u7d71 parser\u3001POS \u662f\u4eba\u6240 \u8a2d\u8a08\u7684\uff0c\u4f46\u662f\u771f\u6b63\u53e3\u8a9e\u4e0a\u6211\u5011\u4e0d\u4e00\u5b9a\u6703\u9019\u6a23\u5538\uff0c\u800c\u6211\u5011\u65b0\u7cfb\u7d71\u5247\u662f\u4e0d\u9760\u4eba\u5de5\u8a2d\u8a08\uff0c\u8b93\u6a5f \u5668\u81ea\u5df1\u5f9e\u8a9e\u6599\u7576\u4e2d\u5b78\u7fd2\u53e5\u5b50\u95dc\u4fc2\uff0c\u56e0\u6b64\u53ef\u80fd\u5408\u6210\u51fa\u4f86\u7684\u8072\u97f3\u6703\u6bd4\u8f03\u6d41\u66a2\u3002\u4e0d\u904e\u5728\u4e2d\u82f1\u593e \u8868 4. \u50b3\u7d71\u8a9e\u8a00\u7279\u5fb5 \u65b0\u67b6\u69cb\u8a9e\u8a00\u7279\u5fb5 \u5716 5 \u8207\u5716 6 \u5206\u5225\u70ba\u65b0\u820a\u7cfb\u7d71\u5728\u7d14\u4e2d\u6587\u8207\u4e2d\u82f1\u593e\u96dc\u7684\u76f8\u4f3c\u5ea6\u3001\u81ea\u7136\u5ea6\u8207\u53ef\u7406\u89e3\u5ea6\u504f\u597d\u7684\u6bd4 \u96dc\u6e2c\u8a66\u4e0b\u5169\u8005\u5206\u6578\u5dee\u8ddd\u4e0d\u5927\uff0c\u53ef\u80fd\u662f RNNLM \u5728\u4e2d\u82f1\u593e\u96dc\u7684\u82f1\u6587\u53e5\u5b50\u90e8\u5206\u5c11\uff0c\u6240\u4ee5\u7121\u6cd5 \u6599\u662f\u8a13\u7df4\u8a9e\u6599\u4e2d\u6c92\u6709\u88ab\u8a13\u7df4\u5230\u7684\u90e8\u5206\uff0c\u6700\u5f8c\u65b0\u820a\u7cfb\u7d71\u6bd4\u8f03\u7684\u97f3\u6a94\u7686\u70ba\u540c\u4e00\u53e5\u8a71\u4ee5\u793a\u516c\u5e73\uff0c \u800c\u6e2c\u8a66\u8005\u4e0d\u77e5\u9053\u54ea\u500b\u97f3\u6a94\u70ba\u65b0\u7cfb\u7d71\u6240\u5408\u6210\u907f\u514d\u5206\u6578\u704c\u6c34\u3002 4.1 \u5be6\u9a57\u8a2d\u5b9a \u97f3\u7d20(PHONE) \u97f3\u7d20\u5728\u97f3\u7bc0\u4e2d\u7684\u4f4d\u7f6e \u8f03\uff0c\u8868 6 \u8207\u8868 7 \u5247\u5206\u5225\u70ba\u65b0\u820a\u7cfb\u7d71\u5728\u7d14\u4e2d\u6587\u8207\u4e2d\u82f1\u593e\u96dc\u7684\u4e3b\u89c0 MOS \u5206\u6578\u6bd4\u8f03\u3002\u8a55\u6bd4\u504f \u6709\u6548\u5f97\u77e5\u5b57\u5143\u5728\u53e5\u5b50\u72c0\u614b\uff0c\u4e0d\u904e\u6574\u9ad4\u770b\u4f86\u65b0\u67b6\u69cb\u78ba\u5be6\u6709\u6bd4\u8f03\u597d\u7684\u6210\u7e3e\u3002\u4f46\u662f\u672c\u7814\u7a76\u53ea\u662f \u97f3\u7d20\u5728\u97f3\u7bc0\u4e2d\u7684\u4f4d\u7f6e \u97f3\u7bc0(SYLLABLE) \u521d\u6b65\u5be6\u9a57\u65b0\u67b6\u69cb\u65b9\u6cd5\u4ee5\u5be6\u9a57\u6578\u64da\u4f86\u8b49\u660e\u771f\u7684\u6703\u6bd4\u50b3\u7d71\u597d\uff0c\u771f\u6b63\u6bd4\u8f03\u597d\u7684\u8a73\u7d30\u539f\u56e0\u9700\u8981\u672a \u597d\u90e8\u5206\uff0c\u5f9e\u6e2c\u8a66\u7d50\u679c\u4e2d\u6211\u5011\u53ef\u4ee5\u767c\u73fe\uff0c\u5728\u5716 5 \u8207\u5716 6 \u4e2d\u4e0d\u7ba1\u662f\u5728\u76f8\u4f3c\u5ea6\u3001\u81ea\u7136\u5ea6\u8207\u53ef\u7406 \u97f3\u7d20\u6578\u91cf\uff0c\u5728\u8a5e\u4e2d\u7684\u4f4d\u7f6e X \u89e3\u5ea6\u4f86\u770b\uff0c\u5927\u90e8\u5206\u6e2c\u8a66\u8005\u504f\u597d\u65b0\u7cfb\u7d71\u3002\u800c\u7531\u8868 6 \u4f86\u770b\u65b0\u7cfb\u7d71\u7684\u8072\u97f3\u6bd4\u50b3\u7d71\u67b6\u69cb\u6240\u5408\u6210\u51fa \u4f86\u7e7c\u7e8c\u6df1\u5165\u63a2\u8a0e\u3002</td></tr><tr><td colspan=\"7\">4.1.1 \u8a9e\u6599 \u6211\u5011\u4f7f\u7528\u7684\u8a13\u7df4\u8a9e\u6599\u3001\u5408\u6210\u8a9e\u6599\u7686\u70ba\u6211\u5011\u8207\u53f0\u7063\u6578\u4f4d\u6709\u8072\u66f8\u5354\u6703\u5408\u4f5c\u9304\u88fd\u7684\"NTUT Audiobook Corpus Vol.2\"\uff0c\u5728\u6b64\u8a9e\u6599\u5eab\u4e2d\u6211\u5011\u8acb\u5c08\u696d\u9304\u97f3\u54e1\u70ba\u6211\u5011\u9304\u88fd\u7537\u8072\u8a9e\u6599\uff0c\u6240\u6709 \u8a5e(WORD) \u97f3\u7bc0\u6578\u91cf\uff0c\u5728\u77ed\u8a9e\u4e2d\u7684\u4f4d\u7f6e X \u77ed\u8a9e(CLAUSE) \u8a5e\u6578\u91cf\uff0c\u5728\u53e5\u5b50\u4e2d\u7684\u4f4d\u7f6e \u5728\u53e5\u5b50\u4e2d\u7684\u4f4d\u7f6e \u53e5\u5b50(UTTERANCE) \u77ed\u8a9e\u6578\u91cf\uff0c\u5728\u6bb5\u843d\u4e2d\u7684\u4f4d\u7f6e \u77ed\u8a9e\u6578\u91cf\uff0c\u5728\u6bb5\u843d\u4e2d\u7684\u4f4d\u7f6e \u6bb5\u843d(PARAGRAPH) \u53e5\u5b50\u7684\u6578\u91cf \u53e5\u5b50\u7684\u6578\u91cf \u4f86\u7684\u8072\u97f3\u7a0d\u5fae\u81ea\u7136\u8207\u76f8\u4f3c\u539f\u8a9e\u8005\u7684\u8072\u97f3\uff0c\u4f46\u5728\u8868 7 \u4e2d\u82f1\u593e\u96dc\u6e2c\u8a66\u5206\u6578\u770b\u8d77\u4f86\u65b0\u820a\u7cfb\u7d71\u5206 5. \u7d50\u8ad6 \u6578 \u5dee \u5225 \u4e0d \u5927 \uff0c \u6240 \u4ee5 \u53ea \u80fd \u8aaa \u5169 \u8005 \u5927 \u6982 \u76f8 \u7576 \u3002 \u4e0d \u904e \u6982 \u89c0 \u4f86 \u770b \u53ef \u4ee5 \u767c \u73fe \u4f7f \u7528 \u5b57 \u5143 \u5c64 \u7d1a \u5716 6. \u65b0\u820a\u7cfb\u7d71\u4e2d\u82f1\u593e\u96dc\u504f\u597d\u6bd4\u8f03 \u5728\u672c\u7814\u7a76\u4e2d\uff0c\u6211\u5011\u5c07\u4e00\u822c\u8a9e\u97f3\u5408\u6210\u4e2d\u7684\u6587\u672c\u5206\u6790\u505a\u66ff\u63db\uff0c\u5c07\u4ee5\u524d\u4ee5\u8a5e\u70ba\u55ae\u4f4d\u6c42\u53d6\u6587\u8108\u8a0a (character-level)\u7684 word2vec \u8207 RNNLM \u4f86\u53d6\u4ee3\u50b3\u7d71 parser \u9032\u884c\u6587\u8108\u8a0a\u606f\u64f7\u53d6\uff0c\u80fd\u5408\u6210\u51fa \u76f8\u7576\u8cbc\u8fd1\u4eba\u8072\u7684\u8072\u97f3\u3002 [Figure 6. Experimental results of the A/B/X preference test on \u606f\u7684\u65b9\u5f0f\uff0c\u66ff\u63db\u6210\u4ee5\u5b57\u5143\u70ba\u8655\u7406\u55ae\u4f4d\u3002\u7528 word2vec \u6c42\u53d6\u5b57\u5143\u7684\u8a9e\u610f\u5c6c\u6027\u8207\u6587\u6cd5\u89d2\u8272\u5206\u985e mixed Chinese-English sentence synthesis] \u548c\u5229\u7528 RNNLM \u731c\u6e2c\u5b57\u5143\u5728\u53e5\u5b50\u4e2d\u7684\u72c0\u614b\uff0c\u4ee5\u9019\u7a2e\u65b9\u5f0f\u80fd\u5920\u907f\u958b\u65b7\u8a5e\u3001POS \u9700\u8981\u5927\u91cf\u4eba \u97f3\u6a94\u7686\u5728\u5c08\u696d\u9304\u97f3\u5ba4\u9304\u97f3\uff0c\u9304\u88fd\u6642\u662f\u4ee5\u6bb5\u843d\u70ba\u55ae\u4f4d\u5538\u5b8c\uff0c\u4fdd\u7559\u8a9e\u53e5\u4e4b\u9593\u7684\u9023\u7d50\u6027\uff1b\u5408\u6210 \u8a9e\u6599\u5247\u5f9e\u5176\u4e2d\u5404\u5225\u62bd\u53d6\u4e2d\u6587 280 \u53e5\u53ca\u4e2d\u82f1\u593e\u96dc 160 \u53e5\u4f86\u505a\u5408\u6210\uff0c\u62bd\u51fa\u7684\u53e5\u5b50\u7686\u4e0d\u5728\u8a13\u7df4 WORD2VEC \u985e\u5225 X \u5de5\u6a19\u8a3b\u8cc7\u6599\u5eab\u7684\u7f3a\u9ede\uff1b\u800c\u65b0\u820a\u7cfb\u7d71\u5728\u5404\u9805\u8a55\u6bd4\u4e2d\u90fd\u662f\u65b0\u7cfb\u7d71\u5408\u6210\u51fa\u4f86\u7684\u8072\u97f3\u8f03\u70ba\u512a\u826f\uff0c \u5b57\u5143\u662f\u5c6c\u65bc\u54ea\u4e00\u985e \u6240\u4ee5\u6211\u5011\u63d0\u51fa\u7684\u5b57\u5143\u5c64\u7d1a\u7684\u6587\u8108\u8a0a\u606f\u64f7\u53d6\u65b9\u6cd5\u78ba\u5be6\u80fd\u9054\u5230\u76f8\u7576\u751a\u81f3\u8d85\u8d8a\u50b3\u7d71\u65b9\u5f0f\u7684\u6210 \u8a9e\u6599\u7576\u4e2d\u3002\u8868 3 \u70ba\u8a13\u7df4\u8a9e\u6599\u8cc7\u6599\u8868\u3002 RNNLM X \u5b57\u5143\u9593\u524d\u5f8c\u6642\u9593\u9806\u5e8f \u7e3e\u3002</td></tr></table>", | |
| "html": null | |
| } | |
| } | |
| } | |
| } |