| { | |
| "paper_id": "O17-1032", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:59:50.634436Z" | |
| }, | |
| "title": "On the Use of Sequence Labeling and Matching Methods for ASR Error Detection and Correction", | |
| "authors": [ | |
| { | |
| "first": "Chia-Hua", | |
| "middle": [], | |
| "last": "\u5433\u4f73\u6a3a", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Chun-I", | |
| "middle": [], | |
| "last": "Wu", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Hsiao-Tsung", | |
| "middle": [], | |
| "last": "Tsai", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Yu -Chen", | |
| "middle": [], | |
| "last": "Hung", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Berlin", | |
| "middle": [], | |
| "last": "Kao", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan Normal University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "This paper sets out to study several important aspects pertaining to speech recognition errors, especially the out-of-vocabulary (OOV) word problem that is caused by using generic speech recognition systems for a specific application domain. To this end, a two-stage processing method, involving error detection and error correction, is proposed. For error detection, we explore and compare disparate sequence labeling methods to detect possible errors of different types. Further, in the error correction stage, an effective phone-level matching mechanism along with a domain-specific keyword list is exploited to correct errors of different types detected by the previous stage. Extensive experiments conducted on four application domains, including educational issues, industrial technology-related interviews and speech memos and meeting recordings, show that our proposed methods can boot the performance of a given 354", | |
| "pdf_parse": { | |
| "paper_id": "O17-1032", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "This paper sets out to study several important aspects pertaining to speech recognition errors, especially the out-of-vocabulary (OOV) word problem that is caused by using generic speech recognition systems for a specific application domain. To this end, a two-stage processing method, involving error detection and error correction, is proposed. For error detection, we explore and compare disparate sequence labeling methods to detect possible errors of different types. Further, in the error correction stage, an effective phone-level matching mechanism along with a domain-specific keyword list is exploited to correct errors of different types detected by the previous stage. Extensive experiments conducted on four application domains, including educational issues, industrial technology-related interviews and speech memos and meeting recordings, show that our proposed methods can boot the performance of a given 354", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Hinton", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "IEEE Signal Process. Mag", | |
| "volume": "29", | |
| "issue": "6", | |
| "pages": "82--97", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G.Hinton et al., \"Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups,\" IEEE Signal Process. Mag., vol. 29, no. 6, pp. 82-97, 2012.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "An overview of noise-robust automatic speech recognition", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Deng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Gong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Haeb-Umbach", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "IEEE Transactions on Audio, Speech and Language Processing", | |
| "volume": "22", | |
| "issue": "4", | |
| "pages": "745--777", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J.Li, L.Deng, Y.Gong, andR.Haeb-Umbach, \"An overview of noise-robust automatic speech recognition,\" IEEE Transactions on Audio, Speech and Language Processing, vol. 22, no. 4. pp. 745-777, 2014.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Learning Out-of-Vocabulary Words in Automatic Speech Recognition", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Qin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L.Qin, \"Learning Out-of-Vocabulary Words in Automatic Speech Recognition,\" 2013.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Error detection and accuracy estimation in automatic speech recognition using deep bidirectional recurrent neural networks", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Ogawa", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Hori", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Speech Commun", | |
| "volume": "89", | |
| "issue": "", | |
| "pages": "70--83", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A.Ogawa andT.Hori, \"Error detection and accuracy estimation in automatic speech recognition using deep bidirectional recurrent neural networks,\" Speech Commun., vol. 89, pp. 70-83, 2017.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "OOV Detection and Recovery using Hybrid Models with Different Fragments", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Qin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Sun", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Rudnicky", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1913--1916", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L.Qin, M.Sun, andA.Rudnicky, \"OOV Detection and Recovery using Hybrid Models with Different Fragments,\" no. August, pp. 1913-1916, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Automatic pronunciation scoring of specific phone segments for language instruction", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Kim", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Franco", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Neumeyer", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "Proc. of EUROSPEECH", | |
| "volume": "97", | |
| "issue": "", | |
| "pages": "649--652", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y.Kim, H.Franco, andL.Neumeyer, \"Automatic pronunciation scoring of specific phone segments for language instruction,\" in Proc. of EUROSPEECH, 1997, vol. 97, pp. 649- 652.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Automatic correction of ASR outputs by using machine translation", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [ | |
| "F" | |
| ], | |
| "last": "D'haro Andr", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": ".", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Banchs", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Proc. Annu. Conf. Int. Speech Commun. Assoc. INTERSPEECH", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3469--3473", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "L. F.D'Haro andR. E.Banchs, \"Automatic correction of ASR outputs by using machine translation,\" Proc. Annu. Conf. Int. Speech Commun. Assoc. INTERSPEECH, vol. 08- 12-Sept, pp. 3469-3473, 2016.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "CRF-based combination of contextual features to improve a posteriori word-level confidence measures", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Fayolle", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Moreau", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Raymond", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Gravier", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Gros", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "P.Fayolle, J., Moreau, F., Raymond, C., Gravier, G., & Gros, \"CRF-based combination of contextual features to improve a posteriori word-level confidence measures.,\" Elev.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Improving domain-independent cloud-based speech recognition with domain-dependent phonetic post-processing", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Twiefel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Baumann", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Heinrich", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Wermter", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Proceedings of the 28th AAAI Conference on Artificial Intelligence (AAAI-14)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1--7", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J.Twiefel, T.Baumann, S.Heinrich, andS.Wermter, \"Improving domain-independent cloud-based speech recognition with domain-dependent phonetic post-processing,\" in Proceedings of the 28th AAAI Conference on Artificial Intelligence (AAAI-14), 2014, pp. 1-7.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "ASR error segment localization for spoken recovery strategy", | |
| "authors": [ | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Bechet", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Favre", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing -Proceedings", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "6837--6841", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "F.Bechet andB.Favre, \"ASR error segment localization for spoken recovery strategy,\" in ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing -Proceedings, 2013, pp. 6837-6841.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Person name recognition in ASR outputs using continuous context models", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Bigot", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Senay", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Linares", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Fredouille", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Dufour", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "ICASSP, IEEE Int. Conf. Acoust. Speech Signal Process. -Proc", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "8470--8474", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R.Bigot, B., Senay, G., Linares, G., Fredouille, C., & Dufour, \"Person name recognition in ASR outputs using continuous context models.,\" ICASSP, IEEE Int. Conf. Acoust. Speech Signal Process. -Proc., pp. 8470-8474, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Optimizing lexical and n-gram coverage via judicious use of linguistic data", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Rosenfeld", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1995, | |
| "venue": "Fourth European Conference on Speech Communication and Technology", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1763--1766", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R.Rosenfeld, \"Optimizing lexical and n-gram coverage via judicious use of linguistic data,\" in Fourth European Conference on Speech Communication and Technology, 1995, pp. 1763-1766.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Scikit-learn: Machine learning in Python", | |
| "authors": [ | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Pedregosa", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Varoquaux", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "", | |
| "volume": "12", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "F.Pedregosa andG.Varoquaux, Scikit-learn: Machine learning in Python, vol. 12. 2011.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Theano: a CPU and GPU Math Expression Compiler", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Bergstra", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proc. Python Sci. Comput. Conf", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1--7", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J.Bergstra et al., \"Theano: a CPU and GPU Math Expression Compiler,\" Proc. Python Sci. Comput. Conf., pp. 1-7, 2010.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "GitHub", | |
| "authors": [ | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Chollet", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "F.Chollet, \"Keras,\" GitHub, 2015. [Online]. Available: https://github.com/fchollet/keras.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "text": "\u4ee5\u53ca\u97f3\u7d20\u5c0d\u7167\u6cd5[9]\u3002\u901a\u5e38\u662f\u4ee5\u5b57\u8a5e(word)\u3001\u97f3 \u7d20(phone)\u3001\u7b26\u865f(symbol)\u7b49\u4f5c\u70ba\u8f38\u5165\u7279\u5fb5\u3002\u4f8b\u5982\uff1a[10]\u63d0\u51fa\u4e86\u4e00\u500b\u5728\u5c0d\u8a71\u7cfb\u7d71\u4e2d\u7684\u8a9e\u97f3 \u5230\u8a9e\u97f3(speech-to-speech)\u8f49\u63db\u6a5f\u5236\uff0c\u662f\u5229\u7528\u689d\u4ef6\u96a8\u6a5f\u57df\u5075\u6e2c\u932f\u8aa4\u6a19\u8a18\u9054\u5230\u4fee\u6b63\u6587\u5b57\u7684\u76ee \u7684\u3002[11]\u63d0\u51fa\u4e86\u4e00\u7a2e\u57fa\u65bc\u85c9\u7531\u6f5b\u5728\u8a9e\u7fa9\u5206\u6790(latent sematic analysis, LSA)\u63d0\u53d6\u4e0a\u4e0b\u6587\u7684\u5411 \u91cf\u8868\u793a\u65b9\u6cd5\uff0c\u4e26\u5229\u7528\u652f\u6301\u5411\u91cf\u6a5f(support vector machine, SVM)\u5206\u985e\u5668\u4f5c\u4eba\u540d\u8fa8\u8a8d\u3002\u7531\u65bc", | |
| "content": "<table><tr><td>\u8868\u4e00\u3001 \u83ef\u8a9e\u6703\u8b70\u8a9e\u6599\u5167\u5bb9\u4ecb\u7d39 \u8868\u4e8c\u3001 \u83ef\u8a9e\u8a9e\u6599\u5eab\u53ca\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7d50\u679c \u8868\u4e09\u3001 ROC \u5206\u6790\u4e2d\u7684\u56db\u9805\u6307\u6a19\u5728\u8fa8\u8b58\u5075\u6e2c\u4efb\u52d9\u4e2d\u7684\u5b9a\u7fa9</td></tr><tr><td>\u4e00\u3001 \u7dd2\u8ad6 \u7531\u65bc\u6a5f\u5668\u5b78\u7fd2\u53ca\u6df1\u5ea6\u5b78\u7fd2\u7684\u8fc5\u901f\u767c\u5c55[1]\uff0c\u8a31\u591a\u9818\u57df\u7684\u6027\u80fd\u8868\u73fe\u90fd\u6709\u5927\u5e45\u5ea6\u7684\u63d0\u5347\u53ca\u7a81 \u7834\uff0c\u800c\u8a9e\u97f3\u8fa8\u8b58\u4e5f\u4e0d\u4f8b\u5916\u3002\u8a31\u591a\u5927\u578b\u4f01\u696d\u76f8\u7e7c\u6295\u5165\u8a9e\u97f3\u65b9\u9762\u7684\u7814\u7a76\u53ca\u61c9\u7528\u4e0a\uff0c\u4e26\u4e14\u63d0\u4f9b \u4f7f\u7528\u8005\u8a9e\u97f3\u76f8\u95dc\u670d\u52d9\uff0c\u5305\u542b\u96f2\u7aef\u8a08\u7b97\u8207\u7d42\u7aef\u88dd\u7f6e\u7684\u8a9e\u97f3\u8fa8\u8b58\u7684\u61c9\u7528\u7a0b\u5f0f\u4ecb\u9762(API)\u3002\u56e0 \u70ba\u4e0a\u8ff0\u5e73\u53f0\u63d0\u4f9b\u7684\u4fbf\u5229\u6027\uff0c\u4f7f\u5f97\u5927\u91cf\u8a9e\u97f3\u4e92\u52d5\u7684\u667a\u6167\u578b\u88dd\u7f6e\u88ab\u5ee3\u6cdb\u5730\u61c9\u7528\uff0c\u4f8b\u5982\u8eca\u8f09\u96fb \u8166\u7684\u8a9e\u97f3\u5c0d\u8a71\u4ecb\u9762\u548c\u8a9e\u97f3\u5ba2\u670d\u7b49\uff0c\u9019\u985e\u7684\u61c9\u7528\u901a\u5e38\u662f\u4f9d\u9644\u5728\u8a9e\u97f3\u8fa8\u8b58\u5668\u4e4b\u5f8c\u3002\u56e0\u70ba\u8a9e\u6599 \u7684\u6536\u96c6\u4fbf\u5229\u6027\u53ca\u6210\u672c\u7684\u5dee\u7570\uff0c\u4f7f\u5f97\u4e00\u822c\u7684\u4f7f\u7528\u8005\u65e5\u5e38\u5c0d\u8a71\u6216\u71b1\u9580\u8a71\u984c\u90fd\u80fd\u9054\u5230\u826f\u597d\u7684\u8fa8 \u8b58\u6b63\u78ba\u7387\uff0c\u4f46\u5728\u7279\u5b9a\u9818\u57df\uff0c\u4f8b\u5982\u5de5\u696d\u61c9\u7528\uff0c\u5247\u6703\u906d\u9047\u5230\u7279\u6b8a\u8a5e\u5f59\u6216\u8a72\u9818\u57df\u7684\u5c08\u696d\u8853\u8a9e\uff0c \u53ef\u80fd\u9020\u6210\u91cd\u8981\u5b57\u8a5e\u7121\u6cd5\u8fa8\u8a8d\u7684\u554f\u984c\u3002\u6700\u76f4\u63a5\u7684\u89e3\u6c7a\u65b9\u6cd5\u662f\u91cd\u65b0\u6536\u96c6\u8a9e\u6599\uff0c\u5305\u542b\u8a9e\u97f3\u4ee5\u53ca \u6587\u5b57\u5167\u5bb9\uff0c\u518d\u61c9\u7528\u6a21\u578b\u8abf\u9069\u7b49\u65b9\u6cd5\u89e3\u6c7a\u554f\u984c\u3002\u7136\u800c\uff0c\u65b0\u5f0f\u7684\u8a9e\u97f3\u8fa8\u8b58\u6280\u8853\u4f9d\u8cf4\u6df1\u5c64\u985e\u795e \u7d93\u7db2\u8def\u8207\u5927\u6578\u64da\u8cc7\u6599\uff0c\u9700\u8981\u82b1\u8cbb\u66f4\u9577\u7684\u6536\u96c6\u8cc7\u6599\u7684\u6642\u9593\u6210\u672c\uff0c\u624d\u80fd\u9054\u5230\u548c\u4e00\u822c\u60c5\u5883\u76f8\u8fd1 \u7684\u8fa8\u8b58\u7387\uff0c\u800c\u771f\u5be6\u7684\u61c9\u7528\u60c5\u5883\u53ef\u80fd\u7121\u6cd5\u5148\u6536\u96c6\u8a9e\u97f3\u518d\u4f7f\u7528\u670d\u52d9\u3002\u56e0\u6b64\uff0c\u5982\u4f55\u5feb\u901f\u5730\u5c07\u8fa8 \u6539\u5584\u95dc\u9375\u8a5e\u8fa8\u8b58\u932f\u8aa4\u6240\u5c0e\u81f4\u7684\u554f\u984c\u3002\u5047\u8a2d\u9019\u4e9b\u5b57\u8a5e\u82e5\u80fd\u88ab\u6b63\u78ba\u8f49\u5beb\uff0c\u5247\u80fd\u5e6b\u52a9\u8a9e\u97f3\u8fa8\u8b58 \u61c9\u7528\u65bc\u66f4\u591a\u9818\u57df\u53ca\u60c5\u5883\u4e4b\u4e0b\u3002\u901a\u7528\u7684\u8a9e\u97f3\u8fa8\u8b58\u5668\u662f\u7531\u5927\u6578\u64da\u8a13\u7df4\u800c\u5f97\u7684\u8907\u96dc\u8fa8\u8b58\u5668\uff0c\u4e14 \u9700\u8981 GPU \u7b49\u904b\u7b97\u8cc7\u6e90\uff0c\u800c\u6bcf\u500b\u7d42\u7aef\u61c9\u7528\u90fd\u5f9e\u6b64\u8fa8\u8b58\u5668\u5f97\u5230\u7b2c\u4e00\u968e\u6bb5\u7684\u8f49\u5beb\u6587\u5b57\uff0c\u518d\u642d \u914d\u8f15\u91cf\u7684\u6f14\u7b97\u6cd5\uff0c\u9032\u884c\u7b2c\u4e8c\u6b21\u7684\u8f49\u5beb\u5167\u5bb9\u4fee\u6b63\u3002\u6211\u5011\u5617\u8a66\u5169\u968e\u6bb5\u7684\u89e3\u6c7a\u65b9\u6cd5\uff1a\u9996\u5148\u8a2d\u8a08 \u4e00\u500b\u81ea\u52d5\u5206\u6790\u8fa8\u8b58\u5668\u932f\u8aa4\u7684\u5206\u985e\u5668\uff0c\u518d\u6839\u64da\u932f\u8aa4\u985e\u578b\u8cc7\u8a0a\uff0c\u642d\u914d\u5c11\u91cf\u7684\u95dc\u9375\u8a5e\u6e05\u55ae\u4f86\u4fee \u6b63\u5167\u5bb9\u3002 \u5728\u7279\u6b8a\u9818\u57df\u77e5\u8b58\u7684\u8a9e\u97f3\u8fa8\u8b58\u4efb\u52d9\u4e2d\uff0c\u7531\u65bc\u8a9e\u97f3\u5167\u5bb9\u5305\u542b\u5927\u91cf\u7684\u7279\u6b8a\u540d\u8a5e\uff0c\u4f7f\u5f97\u8fad\u5178 \u5916\u7684\u672a\u77e5\u8a5e(out of vocabulary words, OOV words)\u6703\u56b4\u91cd\u5f71\u97ff\u8fa8\u8b58\u6b63\u78ba\u7387\u3002\u7d93\u7531\u8a9e\u97f3\u8fa8\u8b58 \u6d41\u7a0b\u5f8c\u7684\u7d50\u679c\u4ecd\u6709\u8f49\u5beb\u932f\u8aa4\uff0c\u5728\u6b64\u6211\u5011\u6839\u64da\u662f\u5426\u7834\u58de\u5c0d\u8a71\u4efb\u52d9\u7684\u7406\u89e3\uff0c\u5c07\u932f\u8aa4\u5206\u70ba\u5169\u985e\u3002 \u7b2c\u4e00\u7a2e\u8f15\u5fae\u7684\u932f\u8aa4\uff0c\u901a\u5e38\u662f\u4e00\u822c\u5b57\u8a5e\uff0c\u4f46\u767c\u97f3\u4e0d\u6e05\u695a\u5c0e\u81f4\u767c\u751f\u8a9e\u97f3\u8fa8\u8b58\u932f\u8aa4\u3002\u9019\u6a23\u932f\u8aa4 \u5e38\u767c\u751f\u5728\u65bc\u81ea\u7136\u5c0d\u8a71\u4e0a\uff0c\u4f8b\u5982\u4e0d\u6d41\u5229\u7684\u91cd\u8907\u8d05\u8a5e\u6216\u8a9e\u52a9\u8a5e\u7b49\u3002\u7b2c\u4e8c\u7a2e\u6703\u5f71\u97ff\u7406\u89e3\u7684\u932f\u8aa4\uff0c \u5927\u591a\u662f\u7279\u6b8a\u5b57\u8a5e\u4e14\u4e0d\u5b58\u5728\u8a13\u7df4\u8a9e\u6599\u7684\u8fad\u5178\u4e2d\uff0c\u5c0e\u81f4\u8a9e\u97f3\u8fa8\u8b58\u5668\u7121\u6cd5\u53bb\u6b63\u78ba\u8f49\u5beb\u3002\u4f8b\u5982\u5c08 \u6709\u540d\u8a5e\u3001\u4eba\u540d\u3001\u5730\u540d\u3001\u6578\u5b57\u53ca\u4e2d\u82f1\u593e\u96dc\u7684\u5b57\u8a5e\u7b49\u3002\u5728\u5be6\u969b\u61c9\u7528\u4e2d\uff0c\u4e0d\u6d41\u5229\u8a9e\u53e5\u9020\u6210\u7684\u932f \u8aa4\u9069\u5408\u5728\u7b2c\u4e00\u968e\u6bb5\u7684\u901a\u7528\u8a9e\u97f3\u8fa8\u8b58\u5668\u4e2d\u89e3\u6c7a\uff0c\u5728\u6b64\u4e26\u4e0d\u8a0e\u8ad6\u3002\u800c\u5f71\u97ff\u7406\u89e3\u7684\u7279\u6b8a\u8a5e\u5f59\u4e0d \u5bb9\u6613\u6536\u96c6\u5927\u6578\u64da\u4e26\u52a0\u5165\u8a13\u7df4\u8a9e\u6599\u4e2d\uff0c\u9700\u8981\u7368\u7acb\u89e3\u6c7a\u6b64\u554f\u984c\u3002\u91cd\u8981\u8a5e\u5f59\u6e05\u55ae\u8f03\u5bb9\u6613\u4eba\u5de5\u5b9a \u7fa9\uff0c\u800c\u5f71\u97ff\u7406\u89e3\u7684\u8a5e\u53ef\u4ee5\u8996\u70ba\u95dc\u9375\u8a5e\uff0c\u800c\u6211\u5011\u53ef\u4ee5\u5f9e\u95dc\u9375\u8a5e\u7684\u7cbe\u78ba\u7387\u548c\u53ec\u56de\u7387\u8a55\u4f30\u8f49\u5beb \u6587\u5b57\u5426\u80fd\u6eff\u8db3\u5c0d\u8a71\u5167\u5bb9\u7684\u7406\u89e3\u3002 \u672a\u77e5\u8a5e\u662f\u4e00\u500b\u51fa\u73fe\u5728\u6e2c\u8a66\u8a9e\u6599\uff0c\u4f46\u4e26\u4e14\u4e0d\u5b58\u5728\u65bc\u8fa8\u8b58\u8fad\u5178\u4e2d\u7684\u5b57\u8a5e\u3002\u7136\u800c\uff0c\u5927\u591a\u6578\u8a9e\u97f3 \u8fa8\u8b58\u7cfb\u7d71\u90fd\u662f\u5c6c\u65bc\u5c01\u9589\u8a5e\u5f59(closed-vocabulary)\u7684\u8fa8\u8b58\u5668\uff0c\u5373\u53ea\u80fd\u8fa8\u8b58\u56fa\u5b9a\u4e14\u6709\u9650\u7684\u8a5e\u5f59\u3002 \u7576\u9019\u4e9b\u672a\u77e5\u8a5e\u51fa\u73fe\u5728\u6e2c\u8a66\u8a9e\u6599\u4e2d\uff0c\u7cfb\u7d71\u5c07\u7121\u6cd5\u8b58\u5225\uff0c\u5c0e\u81f4\u5b83\u88ab\u8aa4\u8a8d\u6210\u5df2\u77e5\u8a5e\u3002\u6b64\u5916\uff0c\u767c \u751f\u672a\u77e5\u8a5e\u7684\u540c\u6642\uff0c\u66f4\u53ef\u80fd\u9023\u5e36\u5f71\u97ff\u5468\u906d\u5176\u4ed6\u7684\u5df2\u77e5\u8a5e[3]\u3002\u800c\u5e73\u5747\u4f86\u8aaa\uff0c\u4e00\u500b\u672a\u77e5\u8a5e\u53ef\u80fd \u7522\u751f 1.2 \u500b\u5b57\u932f\u8aa4[4]\u3002\u70ba\u4e86\u6539\u5584\u672a\u77e5\u8a5e\u7684\u554f\u984c\uff0c\u8a31\u591a\u7814\u7a76\u63d0\u51fa\u4e86\u4ee5\u6a21\u578b\u8abf\u9069(model adaptation)\u6216\u662f\u958b\u653e\u8a5e\u5f59(open-vocabulary)\u65b9\u6cd5\u4f86\u505a\u6539\u5584\u3002\u4e00\u822c\u800c\u8a00\uff0c\u9700\u8981\u6536\u96c6\u81ea\u7136\u8a9e\u53e5 \u624d\u80fd\u5efa\u7acb\u8a9e\u8a00\u6a21\u578b\u4f9b\u8fa8\u8b58\u5668\u4f7f\u7528\uff0c\u4f46\u4f7f\u7528\u5c08\u6709\u540d\u8a5e\u7684\u8a9e\u53e5\u4e0d\u5bb9\u6613\u6536\u96c6\u3002\u4ee5\u4e0b\u6211\u5011\u91dd\u5c0d\u8a9e \u97f3\u8b58\u5225\u932f\u8aa4\u7684\u6539\u5584\u6240\u4f7f\u7528\u7684\u7279\u5fb5\u53ca\u6a21\u578b\u65b9\u6cd5\u505a\u66f4\u9032\u4e00\u6b65\u7684\u63a2\u8a0e\u3002 \u8fd1\u4e8c\u5341\u5e74\u4f86\uff0c\u5df2\u6709\u8a31\u591a\u7814\u7a76\u5617\u8a66\u6aa2\u6e2c\u548c\u4fee\u5fa9\u8a9e\u97f3\u8fa8\u8b58\u932f\u8aa4\u3002\u6709\u5e7e\u500b\u65b9\u6cd5\u80fd\u5920\u5075\u6e2c\u672a \u77e5\u8a5e\uff1a1)\u4ee5\u6df7\u5408\u8a9e\u8a00\u6a21\u578b(hybrid language model)\u505a\u89e3\u78bc(decoding)\uff0c\u4e26\u4e14\u4ee5\u97f3\u7d20\u3001\u5b50\u8a5e\u7b49 \u4f86\u8868\u793a\u672a\u77e5\u8a5e\uff1b2)\u4ee5\u4fe1\u5fc3\u5206\u6578(confidence score)\u548c\u5176\u4ed6\u8cc7\u8a0a\u4f86\u5c0b\u627e\u53ef\u80fd\u7684\u672a\u77e5\u8a5e\u5340\u57df\uff1b3) \u7d50\u5408\u6df7\u5408\u8a9e\u8a00\u6a21\u578b\u53ca\u4fe1\u5fc3\u5206\u6578\uff0c\u9032\u4e00\u6b65\u63d0\u5347\u6aa2\u7d22\u6027\u80fd[5]\u3002 \u932f\u8aa4\u4fee\u5fa9\u6d41\u7a0b\u5305\u542b\u932f\u8aa4\u5075\u6e2c\u53ca\u932f\u8aa4\u4fee\u6b63\u5169\u968e\u6bb5\u3002\u932f\u8aa4\u5075\u6e2c\u65b9\u6cd5\u53ef\u5206\u70ba\u57fa\u65bc\u8a2d\u5b9a\u9580\u6abb \u503c(threshold-based)\u8207\u5206\u985e\u5668(classification-based)\u70ba\u57fa\u790e\u7684\u5169\u7a2e\u7b56\u7565\u3002\u5169\u8005\u4e4b\u9593\u6709\u4e9b\u8a31\u5dee \u7570\uff0c\u57fa\u65bc\u9580\u6abb\u503c\u7684\u65b9\u6cd5\u662f\u8a2d\u5b9a\u55ae\u4e00\u8a55\u4f30\u6307\u6a19\u6216\u5206\u6578\u4f86\u5224\u5b9a\u662f\u5426\u767c\u751f\u932f\u8aa4\uff1b\u800c\u57fa\u65bc\u5206\u985e\u5668 \u7684\u65b9\u6cd5\u5927\u591a\u662f\u6574\u5408\u591a\u7a2e\u7279\u5fb5\u53bb\u8a13\u7df4\u4e8c\u5143\u5206\u985e\u5668\u3002\u57fa\u65bc\u5236\u5b9a\u9580\u6abb\u503c\u7684\u4f5c\u6cd5\u53ef\u4f9d\u64da\u8072\u5b78\u6a21\u578b \u7684\u767c\u97f3\u5206\u6578[6]\u6216\u8a9e\u8a00\u6a21\u578b\u7684\u6a5f\u7387\u7576\u4f5c\u4fe1\u5fc3\u5206\u6578\u3002\u8072\u5b78\u6a21\u578b\u6240\u64f7\u53d6\u7684\u5c0d\u6578\u4e8b\u5f8c\u6a5f\u7387\u6216\u5c0d \u7de8\u8f2f\u8ddd\u96e2\uff0c\u4e26\u6a19\u8a18\u70ba\u4e09\u7a2e\u6a21\u5f0f\uff1a \uf0d8 \u6b63\u78ba\u5b57\u53ca\u932f\u8aa4\u5b57\u5206\u985e\u6a21\u578b\u6a19\u8a18\uff1a\u6b63\u78ba(C)\u53ca\u932f\u8aa4(C \u0305 )\u4e4b\u5340\u584a \u9304\u88fd\u6a21\u5f0f \u7de8\u865f\u4ee3\u865f \u8a9e\u97f3\u4e3b\u984c Corpus01 Corpus02 Corpus03 Corpus04 \u63cf\u8ff0 0 \u5c07\u7121\u6cd5\u8a08\u7b97 F1 \u5206\u6578\uff0c\u56e0\u6b64\u5728\u5be6\u9a57\u8868\u683c\u4e2d\u5c07\u4ee5 --\u8868\u793a\u3002 \u9996\u5148\uff0c\u5728\u8868\u56db\u6211\u5011\u80fd\u5920\u89c0\u5bdf\u5230\uff0c\u5728\u5be6\u9a57\u5ba4\u9304\u97f3\u8a9e\u6599\u4e2d\uff0c\u8868\u73fe\u8f03\u597d\u5206\u985e\u5668\u70ba Decision \u5b57\u662f\u66f4\u70ba\u91cd\u8981\u7684\u6548\u80fd\u8a55\u4f30\u65b9\u6cd5\uff0c\u800c\u6211\u5011\u4ee5\u522a\u9664\u5b57\u7684\u6548\u80fd\u8868\u73fe\u4f86\u770b\uff0cDecision tree \u8868\u73fe\u76f8 \u8f03\u65bc\u985e\u795e\u7d93\u7db2\u8def\u66f4\u70ba\u7a81\u51fa\uff0c\u91dd\u5c0d\u6b64\u4efb\u52d9 SVM \u53ca BRNN \u7121\u6cd5\u6709\u8f03\u597d\u7684\u6027\u80fd\u8868\u73fe\u6211\u8a8d\u70ba\u53ef \u7a2e\u5b57\u8a5e\uff0c\u800c\u85c9\u6b64\u767c\u73fe\u6211\u5011\u4e5f\u89c0\u5bdf\u5230\uff0c\u4ee5\u6642\u9593\u5e8f\u5217\u4e14\u9577\u8a18\u61b6\u6027\u7684\u795e\u7d93\u7db2\u8def\u5728\u5075\u6e2c\u66ff\u63db\u5b57\u6642 \u80fd\u5920\u6709\u5f88\u4e0d\u932f\u7684\u8868\u73fe\uff0c\u9019\u5c0d\u65bc\u6211\u5011\u5728\u7b2c\u4e8c\u6b65\u9a5f\u7684\u932f\u8aa4\u4fee\u6b63\u662f\u975e\u5e38\u6709\u5229\u7684\u4e00\u7a2e\u73fe\u8c61\u3002 \u9375\u5b57\u8a5e\u3002\u800c\u7531\u6211\u5011\u5728\u57fa\u790e\u5be6\u9a57\u4e2d\u7684\u95dc\u9375\u5b57\u4fee\u6b63\u8868\u73fe\u5c31\u80fd\u9054\u5230\u5e73\u5747\u53ec\u56de\u7387\u7d04 78%\u3001\u7cbe\u78ba\u7387 \u7d04 87%\uff0c\u7136\u800c\u6211\u5011\u66f4\u9032\u4e00\u6b65\u505a\u4fee\u6b63\u6539\u5584\uff0c\u4e26\u4e14\u5448\u73fe\u51fa\u66f4\u597d\u6027\u80fd\u8868\u73fe\u5e73\u5747\u53ec\u56de\u7387\u7d04 78%\u3001 \u6a21\u578b\u3001\u7d71\u8a08\u6a21\u578b\u3001\u6a5f\u5668\u5b78\u7fd2\u3001\u6a5f\u5668\u7ffb\u8b6f[7]\u4e0a\u4e0b\u6587\u8a9e\u610f\u53ca\u4e3b\u984c\u6a21\u578b\u9700\u8981\u5927\u91cf\u8cc7\u6599\u8a13\u7df4\u5b57\u8a5e\u8868\u793a\u6cd5\uff0c\u4e26\u4e14\u4e0d\u9069\u7528\u65bc\u6587\u672c\u7d50\u69cb\u8f03\u5f31\u7684\u6703 Corpus01 \u8ab2\u5802\u8a66\u9a57\u5c0d\u8a71 \u5b57\u6578 3878 2593 1267 1665 \u6b63\u78ba\u63a5\u53d7 \u5be6\u969b\u4e0a\u662f\u6b63\u78ba\u5b57\uff0c\u4e26\u4e14\u88ab\u5206\u985e\u5230\u6b63\u78ba tree \u53ca RNN\uff0c\u800c\u66f4\u70ba\u8907\u96dc\u7684 BRNN \u985e\u795e\u7d93\u7db2\u8def\u67b6\u69cb\u53cd\u800c\u5728\u932f\u8aa4\u5b57\u5075\u6e2c\u4e0a\u4e0d\u5982\u524d\u8005\uff0c\u751a \u80fd\u539f\u56e0\u70ba\uff1a1)\u7531\u65bc\u5927\u91cf\u8cc7\u8a0a\u7686\u70ba\u672a\u522a\u9664\u5b57\u5c6c\u65bc\u5206\u985e\u985e\u5225\u6578\u91cf\u8f03\u70ba\u6975\u7aef\uff0c\u56e0\u6b64 SVM \u8f03\u7121 \u7cbe\u78ba\u7387\u7d04 90%\uff0c\u6709\u6548\u63d0\u5347 3%\u9818\u57df\u8a5e\u7cbe\u78ba\u7387\uff0c\u4e26\u4e14\u6539\u5584\u8a9e\u97f3\u8fa8\u8b58\u6587\u672c\u7684\u932f\u8aa4\u3002 \u8868\u516d\u3001 \u6bd4\u8f03 Corpus01~ Corpus04 \u932f\u8aa4\u578b\u614b\u5075\u6e2c\u6548\u80fd \uf0d8 \u672a\u522a\u9664\u53ca\u522a\u9664\u5b57\u5206\u985e\u6a21\u578b\u6a19\u8a18\uff1a\u672a\u767c\u751f\u522a\u9664\u932f\u8aa4(D \u0305 )\u53ca\u522a\u9664\u932f\u8aa4(D)\u4e4b\u5340\u584a \uf0d8 \u932f\u8aa4\u985e\u578b\u5206\u985e\u6a21\u578b\u6a19\u8a18\uff1a\u6b63\u78ba(C)\u3001\u63d2\u5165\u932f\u8aa4(I)\u53ca\u66ff\u63db\u932f\u8aa4(S)\u4e4b\u5340\u584a 3 \u6a21\u578b \u5be6\u9a57\u5ba4\u9304\u97f3 Corpus02 \u696d\u52d9\u62dc\u8a2a\u5c0d\u8a71 Corpus03 \u53e5\u6578 204 176 323 85 \u8a9e\u8005\u6578 7 8 8 7 (true positives, TP) \u932f\u8aa4\u63a5\u53d7 \u5be6\u969b\u4e0a\u662f\u932f\u8aa4\u5b57\uff0c\u4e26\u4e14\u88ab\u5206\u985e\u5230\u932f\u8aa4 \u81f3 BRNN \u7684\u6b63\u78ba\u5b57\u5075\u6e2c\u4e5f\u76f8\u5c0d\u8868\u73fe\u8f03\u5dee\uff0c\u6211\u8a8d\u70ba\u53ef\u80fd\u539f\u56e0\u5982\u4e0b\uff1a1)\u7531\u65bc\u5be6\u9a57\u5ba4\u9304\u97f3\u662f\u5c6c \u65bc\u9078\u53d6\u91cd\u8981\u8a9e\u53e5\u91cd\u8907\u9304\u97f3\uff0c\u5f8c\u8005\u8f03\u8907\u96dc\u67b6\u69cb\u53ef\u80fd\u76f8\u5c0d\u770b\u4e86\u6bd4\u8f03\u9577\u7684\u524d\u5f8c\u6587\u8cc7\u8a0a\uff0c\u4e26\u975e\u771f \u6cd5\u6709\u6548\u4f5c\u5206\u985e\u30022)\u932f\u8aa4\u522a\u9664\u8cc7\u8a0a\uff0c\u5176\u5be6\u4e0d\u5bb9\u6613\u5f9e\u524d\u5f8c\u6587\u89c0\u5bdf\u5230\uff0c\u6545\u7576\u6211\u5011\u5617\u8a66\u4f7f\u7528\u8f03\u8907 \u96dc\u7684\u985e\u795e\u7d93\u7db2\u8def\u67b6\u69cb\u6642\uff0c\u6975\u53ef\u80fd\u5c0e\u81f4\u5176\u53cd\u6548\u679c\uff0c\u800c\u6c92\u6709\u826f\u597d\u7684\u6548\u80fd\u8868\u73fe\u3002 Type SVM \u8868\u4e03\u3001 \u6bd4\u8f03\u4ee5\u97f3\u7d20\u6bd4\u5c0d\u65b9\u6cd5\u4fee\u6b63\u8fa8\u8b58\u932f\u8aa4\u4e4b\u6548\u80fd Decision DNN RNN LSTM BRNN tree Corpus Name evaluation PM IMP_PM \u8a9e\u97f3\u8a18\u4e8b\u60c5\u5883 \u6e96\u78ba\u7387 93.4% 87.5% 77.7% 75.9% (false negatives, FN) \u7684\u6709\u5229\uff0c\u56e0\u70ba\u5176\u5be6\u6311\u9078\u53e5\u5b50\u91cd\u65b0\u9304\u88fd\u7684\u6587\u672c\u76f8\u5c0d\u65bc\u5be6\u969b\u5c0d\u8a71\u6587\u672c\u7d50\u69cb\u8f03\u5f31\uff0c\u4e0a\u4e0b\u6587\u8a9e\u53e5 \u8868\u4e94\u3001 \u6bd4\u8f03\u5404\u6a21\u578b\u5728 Corpus01~ Corpus04 \u672a\u767c\u751f\u522a\u9664\u53ca\u767c\u751f\u522a\u9664\u932f\u8aa4\u5075\u6e2c\u6548\u80fd C 0.54 0.88 0.96 0.98 0.97 0.97 Precision 75.20% 82.00% \u8b70\u8a9e\u97f3\u8f49\u5beb\u4e2d\uff0c\u6240\u4ee5\u5728\u672c\u6587\uff0c\u6211\u5011\u5c07\u63a1\u7528\u5b57\u4e32\u641c\u5c0b\u6bd4\u5c0d[9]\u4f5c\u70ba\u57fa\u790e\u65b9\u6cd5\u3002 \u5728\u7279\u5b9a\u9818\u57df\u7684\u8a9e\u97f3\u8fa8\u8b58\u4e2d\uff0c\u7f55\u898b\u8a5e\u6216\u672a\u77e5\u8a5e\u7684\u8655\u7406\u90fd\u662f\u6838\u5fc3\u7684\u554f\u984c[2]\u3002\u800c\u672c\u8ad6\u6587\u63a2 \u8a0e\u7684\u60c5\u5883\u662f\u5728\u4e00\u500b\u5177\u6709\u8a9e\u97f3\u5f37\u5065\u6027\u7684\u8fa8\u8b58\u5668\u7684\u60c5\u6cc1\u4e0b\uff0c\u5617\u8a66\u5229\u7528\u8a72\u9818\u57df\u5c11\u91cf\u7684\u8a9e\u6599\u8cc7\u6e90 \u5728\u6a21\u578b\u90e8\u5206\uff0c\u6211\u5011\u63a2\u8a0e\u5728\u4e0d\u540c\u76e3\u7763\u5f0f\u5b78\u7fd2\u65b9\u6cd5\u4e2d\uff0c\u5c0d\u65bc\u5075\u6e2c\u932f\u8aa4\u7684\u6548\u80fd\u3002\u5176\u4e2d\uff0c\u6211\u5011\u4f7f \u7528\u4e86\u5169\u7a2e\u6a5f\u5668\u5b78\u7fd2\u65b9\u6cd5\uff1a \uff1a\u652f\u6301\u5411\u91cf\u6a5f(SVM) \u3001\u6c7a\u7b56\u6a39(DT) \uff0c\u4e26\u4ee5\u9810\u8a13\u7df4\u8a5e\u5411\u91cf(pre-trained word vectors) \u4f5c\u70ba\u8f38\u5165\u3002\u70ba\u4e86\u7b26\u5408\u8a9e\u97f3\u8f49\u5beb\u5bcc\u542b\u6642\u9593\u53ca\u5e8f\u5217\u7279\u6027\uff0c\u6211\u5011\u66f4\u6df1\u5165\u63a2\u8a0e\u4e86\u4ee5 \u56db\u3001 \u5be6\u9a57 Corpus04 \u6280\u8853\u6703\u8b70\u5c0d\u8a71 \u6b63\u78ba\u5b57 94.1% 87.8% 79.2% 83.3% \u66ff\u63db\u932f\u8aa4\u5b57 4.8% 11.1% 19.3% 15.0% \u63d2\u5165\u932f\u8aa4\u5b57 0.7% 0.3% 1.5% 7.4% \u932f\u8aa4\u62d2\u7d55 (false positives, FP) \u5be6\u969b\u4e0a\u662f\u932f\u8aa4\u5b57\uff0c\u4e26\u4e14\u88ab\u5206\u985e\u5230\u6b63\u78ba \u6b63\u78ba\u62d2\u7d55 \u5be6\u969b\u4e0a\u662f\u6b63\u78ba\u5b57\uff0c\u4e26\u4e14\u88ab\u5206\u985e\u5230\u932f\u8aa4 \u7684\u95dc\u806f\u6027\u4e5f\u8f03\u4f4e\uff0c\u5c0e\u81f4\u5176\u5728\u5206\u985e\u4e0a\u53d7\u5230\u5e72\u64fe\uff0c\u6545\u7121\u6cd5\u6709\u6548\u53bb\u505a\u6b63\u78ba\u5206\u985e\u30022)\u4ee5\u6211\u5011\u7684\u8cc7 \u6599\u96c6\u800c\u8a00\uff0c\u985e\u795e\u7d93\u7db2\u8def\u672c\u8eab\u9700\u8981\u5927\u91cf\u8cc7\u6599\u4f86\u8a13\u7df4\u624d\u80fd\u5920\u6709\u8f03\u597d\u7684\u6548\u80fd\uff0c\u76f8\u5c0d\u65bc\u5c0f\u8cc7\u6599\u96c6 \u800c\u8a00\uff0c\u904e\u5ea6\u8907\u96dc\u7684\u985e\u795e\u7d93\u7db2\u8def\u53cd\u800c\u5c0e\u81f4\u5176\u767c\u751f\u904e\u5ea6\u64ec\u5408\u7684\u73fe\u8c61\u3002 Type SVM Decision DNN RNN LSTM CORPUS01 S --0.23 0.38 0.75 0.5 Corpus01 0.64 Recall 94.50% 94.50% BRNN tree CORPUS01 I ------0.33 ----Precision 87.60% 90.00% D \u0305 0.93 0.95 0.98 0.99 0.99 0.98 C 1.0 0.87 0.97 0.97 0.96 Corpus02 0.96 Recall 87.90% 87.90% \u89e3\u6c7a\u7f55\u898b\u8a5e\u8207\u672a\u77e5\u8a5e\u9020\u6210\u7684\u554f\u984c\u3002 \u4e09\u3001 \u8a9e\u97f3\u8fa8\u8b58\u932f\u8aa4\u5075\u6e2c\u548c\u6539\u932f \u5728\u672c\u7bc0\uff0c\u6211\u5011\u63a2\u8a0e\u8fa8\u8b58\u932f\u8aa4\u4fee\u6b63\u7684\u554f\u984c\uff0c\u4e26\u4e14\u63d0\u51fa\u4e86\u4e00\u500b\u5169\u6b65\u9a5f\u932f\u8aa4\u4fee\u6b63\u67b6\u69cb(\u5716\u4e00)\u3002 \u7b2c\u4e00\u6b65\u9a5f\uff0c\u5c0b\u627e\u53ef\u80fd\u6e2c\u8a66\u8a9e\u6599\u4e2d\uff0c\u53ef\u80fd\u767c\u751f\u932f\u8aa4\u7684\u4f4d\u7f6e\u3002\u7b2c\u4e8c\u6b65\u9a5f\uff0c\u4ee5\u97f3\u7d20\u6bd4\u5c0d\u6cd5\u5c0b\u627e \u53ef\u80fd\u767c\u751f\u932f\u8aa4\u7684\u5340\u584a\u3002\u4ee5\u4e0b\u6211\u5011\u5c07\u5be6\u9a57\u67b6\u69cb\u4e2d\u7684\u5169\u5927\u4e3b\u8ef8\uff0c\u932f\u8aa4\u5075\u6e2c\u6a21\u578b\u8207\u8fa8\u8b58\u932f\u8aa4\u4fee \u6b63\uff0c\u505a\u66f4\u9032\u4e00\u6b65\u7684\u6a21\u578b\u53ca\u65b9\u6cd5\u4ecb\u7d39\u3002 (\u4e00)\u3001 \u932f\u8aa4\u5075\u6e2c\u6a21\u578b \u8868\u793a\u6cd5\u3002\u50b3\u7d71\u8868\u793a\u6cd5\u4e2d\uff0c\u4f8b\u5982\u4e00\u5143\u8868\u793a\u6cd5(one-hot representation)\uff0c\u53ef\u80fd\u56e0\u70ba\u8fad\u5178\u592a\u5927\u5c0e \u81f4\u7dad\u5ea6\u8a5b\u5492\u7684\u554f\u984c[12]\u3002\u56e0\u6b64\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u63d0\u51fa\u540c\u6642\u8003\u616e\u8a5e\u8207\u8a5e\u6027\u7684\u65b0\u6a19\u8a18\uff0c\u518d\u8a13 \u7df4\u65b0\u6a19\u8a18\u7684\u8a5e\u5411\u91cf\u3002\u9996\u5148\u5c07\u8fa8\u8b58\u7d50\u679c\u7684\u6bcf\u500b\u8a9e\u53e5\u8a5e\u5e8f\u5217\u505a\u4e2d\u6587\u65b7\u8a5e\u53ca\u6a19\u8a18\u8a5e\u6027\uff0c\u4e26\u4e14\u5c07 \u5716\u4e00\u3001\u4fee\u6b63\u53ca\u5075\u6e2c\u932f\u8aa4\u4e4b\u6d41\u7a0b\u5716 \u4e0b\u5e7e\u7a2e\u65b9\u6cd5\uff0c\u5728\u672c\u4efb\u52d9\u4e0a\u7684\u6548\u80fd\u3002\u5982\uff1a\u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def(DNN)\u3001\u905e\u8ff4\u795e\u7d93\u7db2\u8def(RNN)\u3001 \u9577\u77ed\u671f\u8a18\u61b6\u985e\u795e\u7d93\u7db2\u8def(LSTM)\u3001\u96d9\u5411\u905e\u8ff4\u795e\u7d93\u7db2\u8def(BRNN)\u3002\u4ee5\u8a5e\u5411\u91cf\u4f5c\u70ba\u8f38\u5165\uff0c\u4e26\u4e14 \u8207\u795e\u7d93\u7db2\u8def\u53c3\u6578\u4e00\u540c\u8a13\u7df4\u3002 (\u4e8c)\u3001 \u8fa8\u8b58\u932f\u8aa4\u4fee\u6b63 \u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u4f7f\u7528\u840a\u6587\u65af\u5766\u8ddd\u96e2(Levenshtein distance)[9]\u53bb\u6bd4\u8f03\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u8f38\u51fa \u7684\u97f3\u7d20\u5e8f\u5217\u8207\u5047\u8a2d\u7684\u95dc\u9375\u8a5e\u76f8\u4f3c\u6027\uff0c\u800c\u9019\u6a23\u7684\u65b9\u5f0f\u4e5f\u5e38\u88ab\u4f7f\u7528\u5728\u5b57\u5c64\u7d1a\u7684\u6bd4\u5c0d\u3002\u8a9e\u97f3\u8f49 \u5beb\u7684\u932f\u8aa4\u4e3b\u8981\u5206\u70ba\u4e09\u7a2e\uff0c\u5305\u542b\uff1a\u4ee3\u66ff\u3001\u63d2\u5165\u3001\u522a\u9664\u3002\u7576\u8a9e\u97f3\u8fa8\u8b58\u4e2d\u672a\u77e5\u8a5e\u5c0e\u81f4\u8a9e\u97f3\u932f\u8aa4 \u672c\u7bc0\u4e2d\u4e3b\u8981\u662f\u4ecb\u7d39\u672c\u8ad6\u6587\u4e2d\u5be6\u9a57\u8a9e\u6599\u5eab\u8207\u76f8\u95dc\u5be6\u9a57\u8a2d\u5b9a\uff0c\u7b2c\u4e00\u90e8\u5206\u5c07\u4ecb\u7d39\u672c\u8ad6\u6587\u6240\u4f7f\u7528 \u7684\u5be6\u9a57\u8a9e\u6599\u5eab\u53ca\u8a9e\u6599\u5eab\u5206\u6790\uff1b\u7b2c\u4e8c\u90e8\u5206\u5c07\u8aaa\u660e\u672c\u8ad6\u6587\u6240\u4f7f\u7528\u7684\u76f8\u95dc\u5be6\u9a57\u8a2d\u5b9a\uff1b\u7b2c\u4e09\u90e8\u5206 \u4ecb\u7d39\u5be6\u9a57\u6548\u80fd\u7684\u8a55\u4f30\u65b9\u6cd5\uff1b\u6700\u5f8c\u5c07\u5448\u73fe\u76f8\u95dc\u5be6\u9a57\u7d50\u679c\u53ca\u89c0\u5bdf\u3002 (\u4e00)\u3001 \u8a9e\u6599\u5eab\u4ecb\u7d39 \u672c\u8ad6\u6587\u4f7f\u7528\u83ef\u8a9e\u5c0d\u8a71\u53ca\u6703\u8b70\u8a9e\u6599\u70ba\u53f0\u7063\u5e2b\u7bc4\u5927\u5b78\u8207\u570b\u5167\u4f01\u696d\u7684\u7522\u5b78\u5408\u4f5c\u8a08\u756b\u8a9e\u6599\u5eab\uff0c\u672c \u8a9e\u6599\u90e8\u5206\u8a9e\u97f3\u70ba\u6539\u5584\u8a9e\u97f3\u8fa8\u8b58\u932f\u8aa4\u800c\u91cd\u65b0\u9304\u88fd\u7684\u5be6\u9a57\u5ba4\u9304\u97f3\u8a9e\u6599\u3002\u4e3b\u8981\u7531\u56db\u500b\u4e0d\u540c\u9818\u57df \u4e3b\u984c\u5167\u5bb9\u53ca\u5169\u7a2e\u4e0d\u540c\u7684\u9304\u88fd\u6a21\u5f0f\uff0c\u5176\u4e2d Corpus01~ Corpus04 \u70ba\u5be6\u9a57\u5ba4\u9304\u97f3\uff0c\u5be6\u9a57\u5ba4\u9304\u97f3 \u4e0a\u6703\u8b70\u9304\u97f3\u4e2d\u6703\u8b70\u5ef3\u7684\u9ea5\u514b\u98a8\u6536\u97f3\u6548\u679c\u53ca\u6703\u8b70\u5ba4\u74b0\u5883\u4e0b\u7684\u566a\u97f3\u5e72\u64fe\u7b49\u7b49\uff0c\u5176\u5be6\u9019\u6a23\u7684\u8a9e \u6599\u5eab\u662f\u975e\u5e38\u5177\u6709\u6311\u6230\u6027\uff0c\u56e0\u6b64\u6211\u5011\u4e5f\u7279\u5225\u70ba\u6b64\u5167\u5bb9\uff0c\u91cd\u65b0\u9304\u88fd\u5c0d\u8a71\uff0c\u4e26\u4e14\u5617\u8a66\u53bb\u6539\u5584\u6b64 \u5c0d\u8a71\u5167\u5bb9\u5c0d\u65bc\u8a9e\u97f3\u8fa8\u8b58\u6027\u7684\u56f0\u96e3\u5ea6\uff0c\u4e26\u85c9\u6b64\u9032\u4e00\u6b65\u5206\u6790\u66f4\u6709\u6548\u6539\u5584\u8fa8\u8b58\u7d50\u679c\u7684\u65b9\u6cd5\u3002 (\u4e8c)\u3001 \u5be6\u9a57\u8a2d\u5b9a \u8a9e\u97f3\u932f\u8aa4\u6aa2\u6e2c\u7684\u96e3\u6613\u5ea6\u8207\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u7684\u6027\u80fd\u8868\u73fe\u606f\u606f\u76f8\u95dc\uff0c\u7136\u800c\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u7684\u8868\u73fe \u522a\u9664\u932f\u8aa4\u5b57 1.2% 1.1% 1.5% 1.7% \u95dc\u9375\u5b57 40 52 36 28 \u5c07\u5c0d\u65bc\u6b63\u78ba\u5b57\u53ca\u932f\u8aa4\u5b57\u5075\u6e2c\u7d50\u679c\u505a\u8a55\u4f30\uff0c\u56e0\u6b64\u6211\u5011\u5148\u5b9a\u7fa9\u6b63\u78ba\u5b57\u5075\u6e2c\u7684\u53ec\u56de\u7387( )\u3001 \u7cbe\u6e96\u5ea6( )\u53ca F1 \u5206\u6578( 1 \u2212 )\u7684\u8a08\u7b97\uff0c\u53cd\u4e4b\uff0c\u932f\u8aa4\u8a5e\u4e4b\u8a55\u4f30\u65b9\u6cd5\u4ea6\u53ef\u985e\u63a8\u3002 \u5728\u672c\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u4ee5 F1 \u5206\u6578\u4f5c\u70ba\u672c\u8ad6\u6587\u7814\u7a76\u8a0e\u8ad6\u7684\u8a55\u4f30\u65b9\u6cd5\uff0c\u7531\u65bc F1 \u5206\u6578\u80fd\u5920\u540c \u6642\u8003\u616e\u53ec\u56de\u7387\u8207\u7cbe\u6e96\u5ea6\uff0c\u5c07\u8f03\u65bc\u5206\u985e\u5668\u7684\u6e96\u78ba\u7387(Accuracy)\u8a55\u4f30\u66f4\u80fd\u5920\u770b\u898b\u6b63\u53cd\u5206\u985e\u4e4b (true negatives, TN) \u8868\u56db\u3001 \u6bd4\u8f03\u5404\u6a21\u578b\u5728 Corpus01~ Corpus04 \u6b63\u78ba\u53ca\u932f\u8aa4\u5340\u57df\u5075\u6e2c\u6548\u80fd Type SVM Decision DNN RNN LSTM BRNN D --0.50 0.22 0.44 0.44 0.5 CORPUS02 CORPUS02 S --0.29 0.77 0.79 0.71 0.72 Precision 94.80% 97.00% D \u0305 0.97 0.97 0.99 0.99 0.99 0.99 I ------0.33 --Corpus03 --Recall 91.60% 91.60% tree CORPUS01 C 0.9 0.97 0.96 0.96 0.96 0.94 C \u0305 --0.85 0.62 0.67 0.61 D --0.66 0.66 0.66 0.66 0.66 CORPUS03 C 1.0 0.87 0.97 0.97 0.96 0.96 Precision 93.20% 93.20% D \u0305 0.96 0.98 0.95 0.96 0.94 0.98 CORPUS03 S --0.29 0.77 0.79 0.71 Corpus04 0.72 Recall 39.90% 39.90% 0.55 D --0.66 0.33 0.57 0.44 0.8 I ------0.33 ----\u7cbe\u6e96\u5ea6\u53ca\u5206\u985e\u7684\u7d30\u7bc0\uff0c\u6545\u5728\u672c\u8ad6\u6587\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u5c07\u4ee5 F1 \u5206\u6578\u4f5c\u5be6\u9a57\u7d50\u679c\u8a0e\u8ad6\u7684\u4f9d\u64da\u3002 (\u4e09)\u3001 \u5075\u6e2c\u8fa8\u8b58\u932f\u8aa4\u4e4b\u5be6\u9a57\u7d50\u679c CORPUS02 C 0.9 0.97 0.97 0.94 0.95 0.91 D \u0305 0.95 1.0 0.99 0.99 0.99 \u4e94\u3001 \u7d50\u8ad6\u8207\u672a\u4f86\u5c55\u671b 0.98 C 0.82 0.85 0.91 0.96 0.94 0.94 CORPUS04 C \u0305 --0.88 0.84 0.69 0.73 0.59 D --1.0 0.66 0.72 0.66 --CORPUS04 S ----0.62 0.84 0.76 0.79 \u672c\u8ad6\u6587\u63a2\u8a0e\u4e00\u822c\u7684\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u61c9\u7528\u65bc\u7279\u5b9a\u9818\u57df\u7684\u5c0d\u8a71\u4e2d\u5c0e\u81f4\u7684\u8fa8\u8b58\u932f\u8aa4\uff0c\u4e26\u4e14\u63d0\u51fa\u4e86 \u8b58\u5668\u61c9\u7528\u81f3\u5404\u9805\u9818\u57df\u662f\u500b\u91cd\u8981\u7684\u554f\u984c\u3002 \u8fd1\u5e74\u4f86\uff0c\u5927\u6578\u64da\u53ca\u96fb\u8166\u904b\u7b97\u80fd\u529b\u7684\u5927\u5e45\u63d0\u5347\uff0c\u4ee5\u81f3\u65bc\u8a9e\u97f3\u8fa8\u8b58\u6280\u8853\u5df2\u7d93\u9032\u5c55\u5230\u66f4\u5177 \u6311\u6230\u7684\u61c9\u7528\uff0c\u751a\u81f3\u88ab\u5be6\u8e10\u65bc\u73fe\u5be6\u74b0\u5883\u4e2d[2]\u3002\u800c \u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u4e2d\u7684\u8072\u5b78\u6a21\u578b\u5df2\u7531\u6df1\u5c64\u985e\u795e \u7d93\u7db2\u8def(Deep Neural Network, DNN)\u6280\u8853\u53d6\u4ee3\u50b3\u7d71\u9ad8\u65af\u6df7\u5408\u6a21\u578b(Gaussian Mixture Model, GMM)\uff0c\u4e26\u4e14\u5728\u8a9e\u97f3\u8fa8\u8b58\u4efb\u52d9\u4e0a\u7372\u5f97\u66f4\u597d\u7684\u6548\u80fd[1]\u3002\u800c\u5728\u904e\u53bb\u4e09\u5341\u591a\u5e74\u4f86\uff0c\u5df2\u6709\u6578\u4ee5\u767e \u8a08\u7684\u5f37\u5065\u6027(noise-robust)\u8a9e\u97f3\u8fa8\u8b58\u65b9\u6cd5\u88ab\u63d0\u51fa\uff0c\u4e26\u4e14\u8b49\u660e\u5176\u4e2d\u6709\u8a31\u591a\u65b9\u6cd5\u5728\u7814\u7a76\u53ca\u5546\u696d \u7528\u9014\u4e0a\u5177\u6709\u91cd\u5927\u5f71\u97ff\u53ca\u6548\u7528[2]\u3002\u800c\u672c\u8ad6\u6587\u4e3b\u8981\u8a0e\u8ad6\u5728\u73fe\u5be6\u74b0\u5883\u4e2d\uff0c\u5927\u898f\u6a21\u904b\u7528\u7684\u8a9e\u97f3\u8fa8 \u8b58\u6280\u8853\uff0c\u5c07\u5176\u61c9\u7528\u65bc\u7279\u5b9a\u9818\u57df\u7684\u60c5\u5883\u4e0b\uff0c\u5c0e\u81f4\u8fa8\u8b58\u7387\u5927\u5e45\u964d\u4f4e\uff0c\u91dd\u5c0d\u9019\u6a23\u4e0d\u5339\u914d (mismatch)\u7684\u554f\u984c\u53bb\u63a2\u8a0e\u5176\u4fee\u5fa9\u932f\u8aa4\u7684\u53ef\u80fd\u6027\u3002\u672c\u8ad6\u6587\u6839\u64da\u4eba\u5de5\u6536\u96c6\u95dc\u9375\u8a5e\u6e05\u55ae\uff0c\u7528\u4f86 \u5728\u672c\u8ad6\u6587\u7576\u4e2d\uff0c\u6211\u5011\u5617\u8a66\u89e3\u6c7a\u5728\u7279\u5b9a\u9818\u57df\u4e0a\u8a9e\u97f3\u8fa8\u8b58\u7387\u7684\u4e0d\u8db3\uff0c\u5c07\u91dd\u5c0d\u9019\u500b\u4efb\u52d9\u7684 \u7f3a\u5931\u63d0\u51fa\u5169\u6b65\u9a5f\u7684\u6539\u5584\u6a21\u578b\uff1b\u7b2c\u4e00\u6b65\u9a5f\u5148\u5075\u6e2c\u8a9e\u97f3\u932f\u8aa4\u4e4b\u5340\u584a\uff0c\u7b2c\u4e8c\u6b65\u9a5f\u5247\u4ee5\u95dc\u9375\u8a5e\u56de \u5fa9\u8a9e\u97f3\u932f\u8aa4\uff0c\u4e26\u4e14\u63d0\u5347\u8a9e\u97f3\u8fa8\u8b58\u7387\u53ca\u53ef\u8b80\u6027\u3002\u6211\u5011\u63d0\u51fa\u7684\u65b9\u6cd5\u80fd\u6bd4\u57fa\u672c\u7684\u97f3\u7d20\u5c0d\u7167\u6cd5\u66f4 \u53ef\u9760\uff0c\u4e26\u4e14\u66f4\u6709\u6548\u53bb\u6539\u5584\u6587\u672c\u932f\u8aa4\u3002\u672c\u8ad6\u6587\u5728\u7b2c\u4e8c\u7bc0\u5c07\u4ecb\u7d39\u8a9e\u97f3\u932f\u8aa4\u5075\u6e2c\u53ca\u672a\u77e5\u8a5e\u6539\u932f \u76f8\u95dc\u7814\u7a76\u7684\u767c\u5c55\u8fd1\u6cc1\uff1b\u7b2c\u4e09\u7bc0\u4ecb\u7d39\u76e3\u7763\u5f0f\u5b78\u7fd2\u7684\u8a9e\u97f3\u932f\u8aa4\u5075\u6e2c\u548c\u6539\u932f\u65b9\u6cd5\uff1b\u7b2c \u56db\u7bc0\u5247\u662f \u672c\u6b21\u6539\u932f\u4efb\u52d9\u4e0a\u5be6\u9a57\u7d50\u679c\u53ca\u8a0e\u8ad6\uff1b\u6700\u5f8c\uff0c\u7b2c\u4e94\u7bc0\u63d0\u51fa\u7d50\u8ad6\u53ca\u63a2\u8a0e\u672a\u4f86\u53ef\u4ee5\u5617\u8a66\u7684\u65b9\u5411\u3002 \u4e8c\u3001 \u76f8\u95dc\u7814\u7a76 \u6578\u76f8\u4f3c\u503c\u4f5c\u70ba\u767c\u97f3\u5206\u6578[7]\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5229\u7528\u8a9e\u8a00\u6a21\u578b\u8a08\u7b97\u8a5e\u5e8f\u5217\u6a5f\u7387\u4e5f\u662f\u5e38\u7528\u7684\u65b9\u6cd5\uff0c \u53ef\u4ee5\u4f5c\u70ba\u8fa8\u8b58\u5b57\u8a5e\u7684\u4fe1\u5fc3\u5206\u6578\u3002\u5728\u57fa\u65bc\u5206\u985e\u5668\u7684\u65b9\u6cd5\uff0c\u4e3b\u8981\u662f\u4ee5\u7d71\u8a08\u6a21\u578b\u3001\u6a5f\u5668\u5b78\u7fd2\u6216 \u985e\u795e\u7d93\u7db2\u8def\u7b49\u7684\u9032\u884c\u4e8c\u5143\u5206\u985e\u3002\u4f8b\u5982\u4ee5\u97f3\u9577\u6a21\u578b(duration model)\u3001\u8a9e\u97f3\u8fa8\u8b58\u6a21\u7d44\u4e2d\u7684\u8072 \u5b78\u6a21\u578b\u6a5f\u7387\u53ca\u8fa8\u8b58\u7d50\u679c\u7b49\u4f5c\u70ba\u8f38\u5165\u7279\u5fb5\uff0c\u518d\u642d\u914d\u5408\u9069\u7684\u6a19\u8a18\u65b9\u5f0f\uff0c\u4f8b\u5982\u689d\u4ef6\u96a8\u6a5f\u57df (conditional random field, CRF)[4], [8]\u3001\u985e\u795e\u7d93\u7db2\u8def(neural network, NN)[4]\u7b49\u90fd\u662f\u5e38\u88ab\u63a1 \u7528\u7684\u9078\u9805\u3002 \u5728\u932f\u8aa4\u4fee\u6b63\u65b9\u9762\uff0c\u6f14\u7b97\u6cd5\u53ef\u4ee5\u5206\u6210\u7c21\u6613\u7684\u5b57\u4e32\u641c\u5c0b\u6bd4\u5c0d\uff0c\u548c\u57fa\u65bc\u8a9e\u53e5\u64f7\u53d6\u7279\u5fb5\u518d\u66f4 \u6b63\u6587\u5b57\u7684\u5169\u985e\u3002\u57fa\u65bc\u8a9e\u53e5\u7279\u5fb5\u7684\u65b9\u6cd5\u662f\u85c9\u7531\u4e0a\u4e0b\u6587\u8cc7\u8a0a\u4f86\u5224\u65b7\u4fee\u5fa9\u5167\u5bb9\uff0c\u65b9\u6cd5\u5305\u542b\u6a5f\u7387 \u6211\u5011\u63a2\u8a0e\u4f7f\u7528\u6a5f\u5668\u5b78\u7fd2\u53ca\u985e\u795e\u7d93\u7db2\u8def\u6a21\u578b\u4f86\u6355\u6349\u8fa8\u8b58\u5b57\u7684\u7279\u6027\uff0c\u5728\u6b64\u67b6\u69cb\u4e0b\uff0c\u7db2\u8def\u8f38\u5165 \u5b57\u8a5e\u53ca\u8a5e\u6027\u5b58\u653e\u5728\u8fad\u5178\u4e2d\u3002\u6587\u672c\u4e2d\u7684\u5b57\u8a5e \u8207\u5176\u8a5e\u6027 \u7684\u8a5e\u7d22\u5f15\u503c\u70ba \u0302\uff0c\u53ef\u4ee5\u8868 \u793a\u70ba \u0302= [ ; ] \u3002\u7d93\u7531\u7d50\u5408\u8a5e\u4ee5\u53ca\u5176\u8a5e\u6027\u5f97\u5230\u7684\u65b0\u7d22\u5f15\uff0c\u9810\u671f\u589e\u5f37\u4e2d\u6587\u8a5e\u5f59\u5728\u4e0d\u540c \u6642\uff0c\u53ef\u80fd\u540c\u6642\u767c\u751f\u4ee3\u66ff\u53ca\u522a\u9664\u7684\u9023\u7e8c\u932f\u8aa4\u3002\u56e0\u6b64\uff0c\u70ba\u4e86\u89e3\u6c7a\u9023\u7e8c\u932f\u8aa4\u5c0e\u81f4\u5b57\u8a5e\u908a\u754c\u6a21\u7cca \u7684\u554f\u984c\uff0c\u6211\u5011\u5c07\u4f7f\u7528\u97f3\u7d20\u5c64\u63d0\u5347\u5c0b\u627e\u95dc\u9375\u5b57\u7684\u53ef\u80fd\u6027\u3002\u4e26\u4e14\u7d93\u7531\u6211\u5011\u521d\u6b65\u5be6\u9a57\uff0c\u97f3\u7d20\u5c0d \u4e4b\u5167\u5bb9\u4e3b\u8981\u9078\u53d6\u5c0d\u8a71\u4e2d\u95dc\u9375\u8a5e\u5f59\u7247\u6bb5\u9304\u88fd\uff0c\u4e26\u4e14\u7531\u5c08\u696d\u4eba\u54e1\u8f49\u5beb\u8207\u6a19\u8a18\u3002\u6703\u8b70\u53c3\u8207\u4eba\u6578 \u7d04 7 \u4f4d\u8a9e\u8005\uff0c\u672c\u5be6\u9a57\u5c07\u8a9e\u6599\u5eab\u5206\u6210\u8a13\u7df4\u96c6\u3001\u767c\u5c55\u96c6\u53ca\u6e2c\u8a66\u96c6\uff0c\u4e3b\u8981\u4ee5\u8a9e\u6599\u4e4b\u7e3d\u53e5\u6578\u6bd4\u4f8b \u548c\u8a9e\u8005\u3001\u8a9e\u97f3\u7684\u5167\u5bb9\u53ca\u9304\u88fd\u6a21\u5f0f\u6709\u5f88\u5927\u95dc\u806f\u6027\uff0c\u56e0\u6b64\u5728\u8a9e\u6599\u5eab\u65b9\u9762\uff0c\u6211\u5011\u8a73\u7d30\u5206\u6790\u83ef\u8a9e \u8a9e\u6599\u7684\u9304\u88fd\u5167\u5bb9\u53ca\u7d71\u8a08\u8a9e\u97f3\u8fa8\u8b58\u5f8c\u7684\u7d50\u679c\u3002\u4e26\u91dd\u5c0d\u8fa8\u8b58\u7d50\u679c\u505a\u5206\u985e\uff0c\u671f\u671b\u80fd\u5920\u9810\u6e2c\u8fa8\u7d50 \u672c\u6bb5\u843d\u4e3b\u8981\u5448\u73fe\u7b2c\u4e09\u7bc0\u65b9\u6cd5\u7684\u5be6\u9a57\u7d50\u679c\uff0c\u5075\u6e2c\u932f\u8aa4\u985e\u578b\u5206\u985e\u6a21\u578b\u4e2d\u53c8\u5283\u5206\u51fa\u5169\u500b\u5b50\u6a21\u578b\uff0c C 0.7 0.94 0.98 0.98 0.98 0.98 \u8a0e\u8ad6\u5b8c\u672c\u5be6\u9a57\u4e4b\u5b50\u6a21\u578b\u4e4b\u5f8c\uff0c\u6211\u5011\u5c07\u7531\u5be6\u9a57\u66f4\u6df1\u5165\u8a0e\u8ad6\u8fa8\u8b58\u932f\u8aa4\u76f8\u95dc\u554f\u984c\uff0c\u4e26\u4e14\u7531 I ----0.18 0.33 --\u5169\u6b65\u9a5f\u6539\u5584\u63aa\u65bd\uff0c\u5176\u4e2d\u5305\u542b\u4e86\u8fa8\u8a8d\u932f\u8aa4\u5340\u57df\u548c\u4fee\u88dc\u6bc0\u640d\u5167\u5bb9\u3002\u5728\u7b2c\u4e00\u6b65\u9a5f\u4e2d\uff0c\u6211\u5011\u63a2\u8a0e --CORPUS03 \u5206\u5225\u662f\u6b63\u78ba\u5b57(C ) \u53ca\u932f\u8aa4\u5b57(C \u0305 ) \u5206\u985e\u6a21\u578b\u53ca\u672a\u522a\u9664(D \u0305 ) \u53ca\u522a\u9664\u5b57(D ) \u5206\u985e\u6a21\u578b\uff0c\u800c\u5075\u6e2c\u932f\u8aa4 C \u0305 --0.92 0.97 0.97 0.97 0.97 \u4e0d\u540c\u5206\u985e\u65b9\u6cd5\u53ca\u8a9e\u6599\u53bb\u63a2\u8a0e\u53bb\u5075\u6e2c\u4e4b\u96e3\u6613\u5ea6\uff0c\u800c\u6211\u5011\u7531\u8868\u516d\u53ef\u89c0\u5bdf\u767c\u73fe\u9664\u4e86 Corpus04 \u4e86\u5e8f\u5217\u6a19\u8a18\u7684\u65b9\u6cd5\u61c9\u7528\u65bc\u932f\u8aa4\u6aa2\u6e2c\u7684\u6548\u80fd\uff0c\u5728\u5be6\u9a57\u4e2d\u6211\u5011\u767c\u73fe\u5229\u7528\u6709\u6642\u9593\u5e8f\u5217\u53ca\u8a18\u61b6\u7684 \u70ba\u8fa8\u8b58\u8f49\u5beb\u6587\u4ef6 D\uff0c\u5176\u4e2d n \u500b\u8a5e\u69cb\u6210\u7684\u8a9e\u53e5\u4ee5{ 1 , 2 , 3 \u2026 } \u8868\u793a\u3002\u7db2\u8def\u8f38\u51fa\u70ba\u932f\u8aa4 \u985e\u5225\uff0c\u6211\u5011\u4f7f\u7528 ( | , \u0398) \u4f86\u5b9a\u7fa9\u5b57\u8a5e \u7528\u6cd5\u9593\u7684\u9451\u5225\u6027\uff0c\u518d\u900f\u904e\u9810\u8a13\u7df4\u8a5e\u5411\u91cf\u4f5c\u70ba\u5408\u9069\u7684\u8868\u793a\u6cd5\uff0c\u65b0\u7684\u8a5e\u5411\u91cf\u4ee5 1 , 2 , 3 \u2026 \u7167\u6cd5\u80fd\u6bd4\u6587\u5b57\u5c64\u7d1a\u7684\u6bd4\u5c0d\u627e\u5c0b\u5230\u66f4\u7d30\u90e8\u7684\u5dee\u7570\uff0c\u7531\u65bc\u672c\u8ad6\u6587\u4f7f\u7528\u4e4b\u8a9e\u6599\u5bcc\u542b\u4e4b\u8f03\u591a\u9818\u57df \u70ba 8:1:1\u3002\u6703\u8b70\u8a9e\u8a00\u4e3b\u8981\u70ba\u4e2d\u6587\uff0c\u593e\u96dc\u5c11\u90e8\u5206\u82f1\u6587\u3002 \u679c\u7684\u5b57\u985e\u578b\uff0c\u5728\u8a9e\u6599\u5eab\u65b9\u9762\u6211\u5011\u4e5f\u5c07\u5176\u5206\u6210\u8a13\u7df4\u96c6\u3001\u767c\u5c55\u96c6\u53ca\u6e2c\u8a66\u96c6\u4f5c\u8a13\u7df4\u3002 \u985e\u5225\u4e3b\u8981\u5206\u6210\u6b63\u78ba\u5b57(C ) \u3001\u66ff\u63db\u5b57(S ) \u53ca\u63d2\u5165\u5b57(I ) \uff0c\u4e26\u4e14\u5728\u672c\u5be6\u9a57\u4e2d\uff0c\u5229\u7528\u6a21\u578b\u53ca\u8a9e\u6599\u4e3b C 1.0 0.95 0.96 0.86 0.91 0.86 \u4e4b\u5916\uff0c\u5be6\u9a57\u5ba4\u9304\u97f3\u8a9e\u6599\u5e73\u5747\u63d2\u5165\u7387\u70ba 0.83%\u3002\u5728\u5be6\u9a57\u5ba4\u9304\u97f3\u8a9e\u6599\u4e0a\uff0cRNN \u7684\u5206\u985e\u6548\u679c\u6700 (\u56db)\u3001 \u8fa8\u8b58\u932f\u8aa4\u4fee\u6b63\u4e4b\u5be6\u9a57\u7d50\u679c \u905e\u8ff4\u795e\u7d93\u7db2\u8def\u5c0d\u65bc\u932f\u8aa4\u5075\u6e2c\u662f\u975e\u5e38\u6709\u5e6b\u52a9\u7684\uff1b\u5728\u7b2c\u4e8c\u6b65\u9a5f\u4e2d\uff0c\u6211\u5011\u4ee5\u7b2c\u4e00\u6b65\u9a5f\u7684\u6a19\u8a18\u7d50 \u5c6c\u65bc\u932f\u8aa4\u985e\u5225 \u7684\u4e8b\u5f8c\u6a5f\u7387\uff0c\u5176\u4e2d \u0398 \u8868\u793a \u6a21\u578b\u4e2d\u7684\u53c3\u6578\u3002 1 \u8a5e\u8868\u793a\u6cd5 \u8a5e\u5d4c\u5165(word embedding)\u662f\u4e00\u500b\u5b57\u8a5e\u7684\u5206\u5e03\u8868\u793a(distributed representation)\u3002\u5206\u5e03\u8868\u793a\u9069 \u7528\u5728\u985e\u795e\u7d93\u7db2\u8def\u6a21\u578b\u7684\u8f38\u5165\u503c\uff0c\u4e26\u4e14\u80fd\u8207\u5176\u4e00\u8d77\u8abf\u6574\u53c3\u6578\uff0c\u8a08\u7b97\u51fa\u4e00\u500b\u66f4\u4f73\u7684\u4efb\u52d9\u5b57\u8a5e \u8868\u793a\u3002 2 \u6a19\u8a18 \u8a9e\u97f3\u8fa8\u8b58\u932f\u8aa4\u5075\u6e2c\u4efb\u52d9\u662f\u53bb\u8a55\u4f30\u8fa8\u8b58\u5b57\u548c\u4eba\u5de5\u8f49\u5beb\u5b57\u7684\u6bd4\u5c0d\u7d50\u679c\u3002\u800c\u5728\u672c\u4efb\u52d9\u4e2d\uff0c\u6211\u5011 \u5c07\u8a9e\u97f3\u8fa8\u8b58\u932f\u8aa4\u5075\u6e2c\u4efb\u52d9\u6b78\u985e\u6210\u4e09\u7a2e\u985e\u5225\uff0c\u4e26\u4e14\u5617\u8a66\u4ee5\u6a5f\u5668\u5b78\u7fd2\u53ca\u985e\u795e\u7d93\u7db2\u8def\u7684\u67b6\u69cb\u53bb \u63a2\u8a0e\u5075\u6e2c\u932f\u8aa4\u4e4b\u6548\u679c\u3002\u800c\u5728\u6a21\u578b\u6a19\u8a18\u65b9\u9762\uff0c\u6211\u5011\u4ee5\u8fa8\u8b58\u7d50\u679c\u4e4b\u6587\u672c\u548c\u4eba\u5de5\u8f49\u5beb\u6587\u672c\u8a08\u7b97 \u8a5e\uff0c\u4e26\u4e14\u5167\u5bb9\u901a\u5e38\u4e2d\u82f1\u6df7\u96dc\uff0c\u56e0\u6b64\u5728\u9019\u6a23\u7684\u60c5\u6cc1\u4e0b\uff0c\u4ee5\u5b57\u5c64\u7d1a\u4f86\u505a\u6bd4\u5c0d\u662f\u8f03\u96e3\u7b26\u5408\u6211\u5011 \u83ef\u8a9e\u6703\u8b70\u8a9e\u6599\u4e3b\u8981\u70ba\u5c0d\u8a71\u6216\u6703\u8b70\u4ea4\u8ac7\u5167\u5bb9\uff0c\u8a9e\u97f3\u5167\u5bb9\u662f\u4ee5\u771f\u5be6\u5c0d\u8ac7\u6216\u6703\u8b70\u6a21\u5f0f\u70ba\u4e3b\uff0c \u672c\u5be6\u9a57\u5be6\u4f5c\u5728\u83ef\u8a9e\u5be6\u9a57\u5ba4\u9304\u97f3\u53ca\u6703\u8b70\u8a9e\u6599\u4e0a\uff0c\u4e26\u57fa\u65bc Python \u7a0b\u5f0f\u8a9e\u8a00\u7684\u51fd\u5f0f\u5eab \u984c\u4e4b\u5dee\u7570\u53bb\u63a2\u8a0e\u5075\u6e2c\u8fa8\u8b58\u932f\u8aa4\u4e4b\u8b70\u984c\u3002 CORPUS04 C \u0305 ----0.62 0.70 0.79 \u70ba\u7a81\u51fa\uff0c\u800c\u50b3\u7d71\u7684\u6a5f\u5668\u5b78\u7fd2\u65b9\u6cd5\u90fd\u8868\u73fe\u8f03\u5dee\u4e4b\u5916\uff0c\u672c\u4efb\u52d9\u5728\u63d2\u5165\u932f\u8aa4\u5075\u6e2c\u4e0a\u5e73\u5747\u8868\u73fe\u90fd \u5728\u8868\u4e03\u4e2d\uff0c\u6211\u5011\u505a\u4e86\u932f\u8aa4\u4fee\u6b63\u7684\u57fa\u790e\u5be6\u9a57\u7a31\u70ba\u97f3\u7d20\u6bd4\u5c0d\u6cd5(Phone Match)\u7c21\u7a31\u70ba PM \u4ee5\u53ca \u679c\u4f5c\u70ba\u4f9d\u64da\uff0c\u4e26\u4ee5\u7279\u6b8a\u9818\u57df\u7684\u95dc\u9375\u8a5e\u8868\u8207\u932f\u8aa4\u5b57\u505a\u97f3\u7d20\u6bd4\u5c0d\u3002\u7d93\u7531\u6211\u5011\u7684\u5169\u968e\u6bb5\u6539\u932f\u65b9 0.71 \u7684\u671f\u5f85\u3002 \u840a\u6587\u65af\u5766\u8ddd\u96e2\u80fd\u5920\u7c21\u55ae\u627e\u5230\u4e00\u7d44\u7d66\u5b9a\u53e5\u5b50\u4e2d\u6700\u53ef\u80fd\u7684\u5168\u8c8c\uff0c\u6216\u662f\u7528\u7d66\u5b9a\u8a5e\u5f59\u4e2d\u6700\u76f8 \u4f3c\u7684\u8a5e\u4f86\u66ff\u63db\u8b58\u5225\u7684\u55ae\u8a5e\u3002\u800c\u70ba\u4e86\u6539\u5584\u4e26\u4e14\u5c0b\u627e\u5230\u66f4\u591a\u53ef\u80fd\u9818\u57df\u8a5e\uff0c\u6211\u5011\u5c07\u5728\u672c\u8ad6\u6587\u7b2c \u56db\u7bc0\u4e2d\uff0c\u6211\u5011\u6240\u5be6\u9a57\u7684\u932f\u8aa4\u4fee\u6b63\u662f\u4ee5\u8a2d\u5b9a\u76f8\u4f3c\u5ea6\u9580\u6abb\u503c\u70ba 0.8\u3002(\u5716\u4e00) \u5c0d\u8a71\u5167\u5bb9\u7121\u7d93\u904e\u7279\u6b8a\u8a2d\u8a08\uff0c\u6240\u4ee5\u6b64\u8a9e\u6599\u5167\u5bb9\u76f8\u8f03\u65bc\u5176\u4ed6\u8a9e\u6599\u662f\u8f03\u8cbc\u9032\u4e00\u822c\u9818\u57df\u77e5\u8b58\u5c0d\u8a71 \u6216\u662f\u5be6\u969b\u958b\u6703\u5167\u5bb9\uff0c\u7136\u800c\u9019\u6a23\u7684\u5167\u5bb9\u5c0d\u65bc\u4e00\u822c\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u4e5f\u76f8\u5c0d\u662f\u4e00\u500b\u56f0\u96e3\u7684\u6311\u6230\uff0c \u5176\u4e2d\u672c\u8a9e\u6599\u5167\u5bb9\u4e2d\u53ef\u80fd\u6703\u9762\u81e8\u5230\u4ee5\u4e0b\u5e7e\u500b\u554f\u984c\uff0c\u5982\uff1a\u5c08\u6709\u540d\u8a5e\u3001\u4eba\u540d\u3001\u4e2d\u82f1\u6587\u593e\u96dc\u5167\u5bb9 \u7b49\uff0c\u4e26\u4e14\u6bcf\u4f4d\u8a9e\u8005\u7684\u8aaa\u8a71\u6a21\u5f0f\u53ef\u80fd\u4e5f\u975e\u5e38\u4e0d\u540c\uff0c\u5982\uff1a\u767c\u97f3\u6e96\u78ba\u6027\u3001\u8a9e\u901f\u53ca\u97f3\u91cf\u7b49\uff0c\u518d\u52a0 Scikit-learn[13]\u3001Theano[14]\u53ca Keras[15]\u7b49\u63d0\u4f9b\u6a5f\u5668\u5b78\u7fd2\u53ca\u985e\u795e\u7d93\u7db2\u8def\uff0c\u5728\u7b2c\u4e09\u5c0f\u7bc0\u4e2d\u8a5e \u8868\u793a\u6cd5\u90e8\u5206\u8f38\u51fa\u503c \u4ee5 20 \u7dad\u8868\u793a\u3002\u932f\u8aa4\u4fee\u6b63\u76f8\u4f3c\u5ea6\u9580\u6abb\u503c\u8a2d\u5b9a\u70ba 0.8\u3002 \u5728\u672c\u8ad6\u6587\u7684\u5206\u985e\u554f\u984c\u4e2d\uff0c\u6211\u5011\u5c07\u6839\u64da\u8868\u4e09\u4e2d\u7684\u56db\u9805\u6307\u6a19\u8a08\u7b97\u4e8c\u7a2e\u8a55\u4f30\u65b9\u5f0f\uff1a\u53ec\u56de\u7387 ( )\u548c\u6e96\u78ba\u7387( )\uff0c\u4e26\u4ee5 F1 \u5206\u6578( 1 \u2212 )\u4f5c\u70ba\u672c\u5be6\u9a57\u4e2d\u4e3b\u8981\u8a55\u4f30\u3002\u6211\u5011 \u6211\u5011\u6bd4\u8f03\u4e86\u5206\u985e\u5668\u5728\u5404\u4e3b\u984c\u9818\u57df\u4e4b\u6027\u80fd\u8868\u73fe\uff0c\u4e26\u4e14\u91dd\u5c0d\u5176\u5206\u985e\u7d50\u679c\u8a08\u7b97\u51fa\u8868\u56db~\u8868 \u4e03\u7684 F1 \u5206\u6578\uff0c\u89c0\u5bdf\u6b63\u78ba\u53ca\u932f\u8aa4\u578b\u614b\u5206\u985e\u7684\u60c5\u5f62\uff0c\u4ee5\u4e0b\u6211\u5011\u5c07\u4ee5\u5169\u500b\u50b3\u7d71\u6a5f\u5668\u5b78\u7fd2\u65b9\u6cd5\uff1a SVM\u3001Decision tree\uff0c\u548c\u56db\u500b\u985e\u795e\u7d93\u7db2\u8def\u65b9\u6cd5\uff1aDNN\u3001RNN\u3001LSTM\u3001BRNN\uff0c\u4e26\u66f4\u9032\u4e00 \u6b65\u5206\u6790\u53ca\u63a2\u8a0e\u5176\u6027\u80fd\u8868\u73fe\u3002\u505a\u5be6\u9a57\u4e2d\uff0c\u82e5\u53ec\u56de\u7387( ) \u53ca\u7cbe\u6e96\u5ea6( ) \u4efb\u4e00\u503c\u70ba \u6539\u826f\u65b9\u6cd5\u7c21\u7a31\u70ba IMP_PM\uff0c\u5982\u540c\u7b2c\u4e09\u7bc0\u6240\u63cf\u8ff0\u65b9\u6cd5\uff0c\u6211\u5011\u4f7f\u7528\u97f3\u7d20\u6bd4\u5c0d\u6cd5\u4f86\u53bb\u5c0b\u627e\u8207\u95dc \u6cd5\uff0c\u80fd\u5920\u6709\u6548\u63d0\u9ad8\u95dc\u9375\u5b57\u4fee\u6b63\u7684\u7cbe\u78ba\u7387\uff0c\u4e26\u4e14\u964d\u4f4e\u539f\u672c\u97f3\u7d20\u5c0d\u7167\u6cd5\u9020\u6210\u5047\u8b66\u5831\u6240\u7522\u751f\u7684 \u8f03\u70ba\u666e\u901a\uff0c\u4e3b\u8981\u539f\u56e0\u70ba\uff1a1)\u7531\u65bc\u5be6\u9a57\u5ba4\u9304\u97f3\u54c1\u8cea\u76f8\u5c0d\u8f03\u597d\uff0c\u6240\u4ee5\u5728\u8a9e\u97f3\u8fa8\u8b58\u4e0a\u8f03\u4e0a\u51fa\u73fe \u9664\u4e86\u63a2\u8a0e\u6b63\u78ba\u5b57\u53ca\u932f\u8aa4\u5b57\u5075\u6e2c\u4e4b\u5916\uff0c\u6211\u5011\u66f4\u9032\u4e00\u6b65\u53bb\u8a0e\u8ad6\u5728\u8fa8\u8b58\u5668\u4e2d\u6240\u767c\u751f\u7684\u522a\u9664 \u9375\u8a5e\u76f8\u4f3c\u7684\u4f4d\u7f6e\uff0c\u4f46\u7531\u5be6\u9a57\u4e2d\u89c0\u5bdf\u5230\uff0c\u6b64\u65b9\u6cd5\u5728\u67d0\u4e9b\u8a9e\u6599\u4e0a\u5bb9\u6613\u7522\u751f\u5047\u8b66\u5831(false alarm) \u3002 \u554f\u984c\u3002\u672a\u4f86\u6211\u5011\u5e0c\u671b\u80fd\u5920\u91dd\u5c0d\u8fa8\u8b58\u932f\u8aa4\u53ca\u672a\u77e5\u8a5e\u505a\u66f4\u9032\u4e00\u6b65\u7684\u63a2\u8a0e\u53ca\u5206\u6790\uff0c\u4e26\u4e14\u52a0\u5165\u8a9e \u63d2\u5165\u932f\u8aa4\u7684\u60c5\u5f62\uff0c\u800c\u7531\u8a9e\u6599\u5eab\u63a2\u8a0e\u4e2d\u6211\u5011\u4e5f\u80fd\u89c0\u5bdf\u5230\u5e73\u5747\u63d2\u5165\u7387\u70ba 0.83%\uff0c\u76f8\u8f03\u65bc\u5176\u4ed6 \u932f\u8aa4\uff0c\u4e26\u4e14\u53bb\u63a2\u8a0e\u9810\u6e2c\u522a\u9664\u932f\u8aa4\u5b57\u4e4b\u8b70\u984c\uff0c\u800c\u7531\u8a9e\u6599\u5eab\u5206\u6790\u4e2d\u6211\u5011\u53ef\u4ee5\u89c0\u5bdf\u5230\u7531\u65bc\u8868\u4e94\uff0c \u70ba\u4e86\u6539\u5584\u9019\u500b\u554f\u984c\uff0c\u6211\u5011\u5c07\u5075\u6e2c\u8fa8\u8b58\u932f\u8aa4\u7684\u7d50\u679c\u4f5c\u70ba\u6b64\u90e8\u5206\u7684\u53c3\u8003\u503c\uff0c\u82e5\u6211\u5011\u5075\u6e2c\u6b64\u5340 \u53e5\u53ca\u8a9e\u610f\u8cc7\u8a0a\u5f37\u5316\u5075\u6e2c\u6a21\u578b\uff0c\u8b93\u4fee\u6b63\u932f\u8aa4\u5b57\u80fd\u5920\u6709\u66f4\u7a69\u5b9a\u7684\u6548\u80fd\u8868\u73fe\u3002\u672c\u8ad6\u6587\u671f\u671b\u63d0\u51fa \u932f\u8aa4\u662f\u8f03\u70ba\u5c11\u898b\u7684\u932f\u8aa4\u985e\u578b\u3002\u5728\u66ff\u63db\u932f\u8aa4\u5075\u6e2c\u4e0a\uff0c\u6211\u5011\u4e5f\u89c0\u5bdf\u5230\u4e00\u500b\u6709\u8da3\u7684\u73fe\u8c61\uff0c\u7d93\u5e38 \u672c\u8a9e\u6599\u522a\u9664\u5b57\u5e73\u5747\u767c\u751f\u6a5f\u7387\u7d04 1.4%\uff0c\u6240\u4ee5\u5176\u5be6\u662f\u76f8\u5c0d\u975e\u5e38\u7f55\u898b\u7684\u932f\u8aa4\u578b\u614b\uff0c\u800c\u5728\u672c\u8a9e \u6599\u4e0a\u7684\u522a\u9664\u5b57\u5075\u6e2c\uff0c\u6211\u5011\u4e5f\u767c\u73fe\u7531\u65bc\u5927\u90e8\u5206\u5b57\u90fd\u6b78\u985e\u70ba\u672a\u522a\u9664\uff0c\u6240\u4ee5\u6b64\u4efb\u52d9\u4e0a\u767c\u73fe\u522a\u9664 \u88ab\u8fa8\u8b58\u932f\u8aa4\u4e14\u88ab\u66ff\u63db\u7684\u5b57\u4f3c\u4e4e\u53ef\u4ee5\u5f9e\u4e00\u4e9b\u898f\u5247\u4e2d\u770b\u898b\uff0c\u4f8b\u5982\uff1a\u67d0\u5b57\u8a5e\u5e38\u88ab\u66ff\u63db\u6210\u5176\u4ed6\u5e7e \u57df\u767c\u751f\u8fa8\u8b58\u932f\u8aa4\uff0c\u624d\u4ee5\u95dc\u9375\u5b57\u8a5e\u8868\u505a\u70ba\u66ff\u63db\u7684\u5019\u9078\u8a5e\uff0c\u4e26\u4ee5\u97f3\u7d20\u6bd4\u5c0d\u6cd5\u627e\u51fa\u6700\u76f8\u4f3c\u7684\u95dc \u4e00\u500b\u6539\u5584\u67b6\u69cb\uff0c\u4f86\u89e3\u6c7a\u672a\u77e5\u8a5e\u6240\u5c0e\u81f4\u6587\u672c\u8a9e\u610f\u4e0d\u6e05\u7684\u554f\u984c\u3002</td></tr></table>", | |
| "type_str": "table", | |
| "html": null, | |
| "num": null | |
| } | |
| } | |
| } | |
| } |