| { |
| "paper_id": "2019", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T14:54:52.508031Z" |
| }, |
| "title": "The Personal Name Modeling in Mandarin ASR System", |
| "authors": [ |
| { |
| "first": "Hong-Bin", |
| "middle": [], |
| "last": "\u6881\u9d3b\u5f6c", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Chiao Tung University", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Liang", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Chiao Tung University", |
| "location": {} |
| }, |
| "email": "hbliang@speech.cm.nctu.edu.tw" |
| }, |
| { |
| "first": "Yih-Ru", |
| "middle": [], |
| "last": "\u738b\u9038\u5982", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Chiao Tung University", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Chiao Tung University", |
| "location": {} |
| }, |
| "email": "yrwang@speech.cm.nctu.edu.tw" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "There are two purposes in the paper, one is training an efficient ASR system, the other is improving the OOV problem caused by the personal name, and we want to recognize it for the purpose of making transcription of different kind of speech data. Name recognition data is also an important training data for the NLP.", |
| "pdf_parse": { |
| "paper_id": "2019", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "There are two purposes in the paper, one is training an efficient ASR system, the other is improving the OOV problem caused by the personal name, and we want to recognize it for the purpose of making transcription of different kind of speech data. Name recognition data is also an important training data for the NLP.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "The paper base on the environment of Kaldi speech recognition toolkit. In the acoustic model part, we use many different kind of neural network such as TDNN to transform the speech information into phone sequence. In the language part, we add Chinese special language information such as variant word combination and name entity decomposition, using n-gram language model and lattice rescoring to transform the phone sequence into word sequence. We also tune the parameters and weights during the decoding process to get the best operation point to obtain a ASR system which is not only good at recognition rate but also efficient at recognition time. Moreover, we focus on the problem of difficulty in personal name recognition. We build a class-based like model to replace the original word-based model of personal name to reach the goal of personal name recognition. ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Mandarin Microphone Speech Corpus-TCC300", |
| "authors": [], |
| "year": null, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "\"Mandarin Microphone Speech Corpus-TCC300,\" [Online]. Available: http://www.aclclp.org.tw/use_mat_c.php#tcc300edu.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Formosa Speech Recognition Challenge", |
| "authors": [], |
| "year": 2018, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "\"Formosa Speech Recognition Challenge 2018,\" [Online]. Available: https://sites.google.com/speech.ntut.edu.tw/fsw/home/challenge#h.p_I_b8U Rx26NXZ.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "AIShell-1: An open-source Mandarin speech corpus and a speech recognition baseline", |
| "authors": [ |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Bu", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Du", |
| "suffix": "" |
| }, |
| { |
| "first": "X", |
| "middle": [], |
| "last": "Na", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Wu", |
| "suffix": "" |
| }, |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Zheng", |
| "suffix": "" |
| } |
| ], |
| "year": 2017, |
| "venue": "Proc. Oriental COCOSDA", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "H. Bu, J. Du, X. Na, B. Wu, and H. Zheng, \"AIShell-1: An open-source Mandarin speech corpus and a speech recognition baseline,\" Proc. Oriental COCOSDA, 2017.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "Semi-OrthogonalLow-RankMatrixFactorizationforDeepNeuralNetworks", |
| "authors": [ |
| { |
| "first": "Daniel", |
| "middle": [], |
| "last": "Povey", |
| "suffix": "" |
| }, |
| { |
| "first": "Gaofeng", |
| "middle": [], |
| "last": "Cheng", |
| "suffix": "" |
| }, |
| { |
| "first": "Yiming", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| }, |
| { |
| "first": "Ke", |
| "middle": [], |
| "last": "Li", |
| "suffix": "" |
| }, |
| { |
| "first": "Hainan", |
| "middle": [], |
| "last": "Xu", |
| "suffix": "" |
| }, |
| { |
| "first": "Mahsa", |
| "middle": [], |
| "last": "Yarmohamadi", |
| "suffix": "" |
| }, |
| { |
| "first": "Sanjeev", |
| "middle": [], |
| "last": "Khudanpur", |
| "suffix": "" |
| } |
| ], |
| "year": 2018, |
| "venue": "in interspeech", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Daniel Povey, Gaofeng Cheng, Yiming Wang, Ke Li, Hainan Xu, Mahsa Yarmohamadi, Sanjeev Khudanpur, \"Semi-OrthogonalLow-RankMatrixFactorizationforDeepNeuralNetworks,\" in interspeech, 2018.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "TABREF0": { |
| "num": null, |
| "text": "The 2019 Conference on Computational Linguistics and Speech Processing ROCLING 2019, pp. 343-357 \u00a9The Association for Computational Linguistics and Chinese Language Processing", |
| "html": null, |
| "content": "<table/>", |
| "type_str": "table" |
| }, |
| "TABREF1": { |
| "num": null, |
| "text": "\u5728\u6240\u6709\u6e2c\u8a66\u8a9e\u6599\u4e4b\u7e3d\u8a5e\u5f59\u6578 387510 \u500b\u3001\u7e3d\u4eba\u540d\u8a5e\u6578\u70ba 1812 \u500b\u4e2d 8 \uff0c\u672c\u5be6\u9a57\u4e4b\u8fa8\u8b58 \u5668\u80fd\u8fa8\u8b58\u4eba\u540d\u7684\u6b63\u78ba\u7387(hit/\u7e3d\u4eba\u540d\u6578)\u7d04\u70ba 81%\uff1b\u8fa8\u8b58\u51fa\u4eba\u540d\u4e14\u97f3\u7bc0\u5b8c\u5168\u6b63\u78ba 9 \u4e4b\u6b63\u78ba\u7387 (hit \u4e14\u97f3\u7bc0\u5b8c\u5168\u6b63\u78ba/\u7e3d\u4eba\u540d\u6578)\u7d04\u70ba 73%\uff0c\u4e5f\u5c31\u662f\u8aaa\uff0c\u5373\u4f7f\u6e2c\u8a66\u8a9e\u6599\u51fa\u73fe\u4e00\u500b\u5f9e\u672a\u51fa\u73fe \u5728\u8a5e\u5178\u4e2d\u4e4b\u4eba\u540d\uff0c\u672c\u8fa8\u8b58\u5668\u4ecd\u6709 73%\u4e4b\u6a5f\u7387\u5c07\u5176\u97f3\u7bc0\u5b8c\u5168\u6b63\u78ba\u7684\u8fa8\u8b58\u51fa\u4f86\u3002 TDNN-F \u67b6\u69cb\uff0c\u8a9e\u8a00\u6a21\u578b\u70ba 4-gram \u8a9e\u8a00\u6a21\u578b\uff0c\u5e73\u6ed1\u5316\u65b9\u6cd5 \u4f7f\u7528 Witten-Bell smoothing\uff0c\u8a13\u7df4\u8a9e\u6599\u70ba 25 \u5104\u8a5e\uff0c\u8fad\u5178\u5927\u5c0f\u70ba 120k\uff0c\u552f\u4e00\u5dee\u5225\u50c5\u5728\u65bc \u8a9e\u8a00\u6a21\u578b\u6709\u6216\u6c92\u6709\u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u505a\u8abf\u6574\u3002 \u5716 \u5341\u4e00\u3001\u52a0\u5165\u8207\u672a\u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u4e4b\u8a5e\u932f\u8aa4\u7387\u6bd4\u8f03 \u516d\u3001 \u7d50\u8ad6 \u53c3\u8003 Kaldi \u4e4b\u4f5c\u8005 Daniel Povey \u5728 2018 \u767c\u8868\u4e4b\u8ad6\u6587\uff0c\u4f7f\u7528 TDNN-F \u4e4b\u67b6\u69cb\u4e0b\u8a13\u7df4", |
| "html": null, |
| "content": "<table><tr><td>Delay Keywords: acoustic model, language model, Mandarin large vocabulary continuous speech Neural Network, TDNN)\u3001\u5c08\u6709\u540d\u8a5e\u8fa8\u8b58(Named Entity Recognition, NER)\u3002 recognition, TDNNs, named entity recognition. \u4e00\u3001 \u7dd2\u8ad6 (\u4e00) \u7814\u7a76\u52d5\u6a5f \u96a8\u8457\u8a13\u7df4\u8a9e\u6599\u7684\u589e\u52a0\uff0c\u96fb\u8166\u904b\u7b97\u901f\u5ea6\u7684\u5927\u5e45\u63d0\u5347\uff0c\u518d\u52a0\u4e0a\u8d8a\u4f86\u8d8a\u591a\u88fd\u4f5c\u8a9e\u97f3\u8fa8\u8b58\u5668 \u7684\u5de5\u5177\u5eab\u7684\u767c\u660e\uff0c\u5982 HTK Speech Recognition Toolkit 1 , Kaldi ASR 2 , et al. \u8a13\u7df4\u4e00\u53ef\u7528\u751a\u81f3\u662f \u5546\u7528\u7d1a\u5225\u7684\u8a9e\u97f3\u8fa8\u8b58\u5df2\u7d93\u662f\u553e\u624b\u53ef\u5f97\u3002 \u4f46\u662f\uff0c\u8fa8\u8b58\u5668\u666e\u904d\u5b58\u5728\u8457\u5c0d\u65bc\u5c08\u6709\u540d\u8a5e\u8fa8\u8b58\u7387\u4e0d\u4f73\u7684\u554f\u984c\uff0c\u7531\u53f0\u5317\u79d1\u6280\u5927\u5b78\u5ed6\u5143\u752b \u8001\u5e2b\u63d0\u4f9b\u4e4b\u6bd4\u8cfd 3 \u7d50\u679c\uff0c\u6bd4\u8f03\u570b\u5167\u591a\u6240\u5b78\u6821\u8207\u4f01\u696d\u6240\u8a13\u7df4\u4e4b\u8fa8\u8b58\u5668\uff0c\u6574\u9ad4\u8fa8\u8b58\u7387\u4e4b\u6b63\u78ba \u7387\u666e\u904d\u53ef\u9ad8\u9054 90%\u505a\u53f3(\u5b57\u932f\u8aa4\u7387 10%\u5de6\u53f3)\uff0c\u4f46\u662f\u4ed4\u7d30\u8a55\u4f30\u8fa8\u8a8d\u932f\u8aa4\u4e4b\u8a5e\u5f59\u5f8c\u767c\u73fe\uff0c\u5404 \u8fa8\u8b58\u5668\u5c0d\u65bc\u5c08\u6709\u540d\u8a5e\u4e4b\u8fa8\u8a8d\u7387\u76f8\u5c0d\u8f03\u5dee\uff0c\u5c24\u5176\u5728\u5c08\u6709\u540d\u8a5e\u4e2d\u6700\u70ba\u91cd\u8981\u4e14\u70ba\u6578\u773e\u591a\u7684\u4eba\u540d \u8fa8\u8b58\u4e4b\u6b63\u78ba\u7387\u50c5\u4e0d\u5230 50%(\u5982\u5716 \u4e00)\uff0c\u4e5f\u5c31\u662f\u8aaa\uff0c\u53ea\u8981\u8fa8\u8b58\u5668\u4e00\u9047\u5230\u4eba\u540d\uff0c\u8fa8\u8b58\u7387\u901a\u5e38\u4e0d \u6703\u592a\u9ad8\u3002\u7136\u800c\uff0c\u4eba\u540d\u5c0d\u8a9e\u8a00\u4f86\u8aaa\u53c8\u662f\u4e00\u500b\u975e\u5e38\u91cd\u8981\u7684\u8cc7\u8a0a\uff0c\u4f8b\u5982\u5728\u65e5\u5e38\u751f\u6d3b\u4e2d\uff0c\u5e38\u7528\u7684 \u53e5\u578b\u88e1\u4e0d\u5916\u4e4e\u7531\u4eba\u4e8b\u6642\u5730\u7269\u6240\u7d44\u6210\uff0c\u800c\u9019\u88e1\u7684\u300c\u4eba\u300d \uff0c\u6307\u7684\u5c31\u662f\u4eba\u540d\uff0c\u518d\u8005\uff0c\u82e5\u8fa8\u8b58\u5668 \u5c07\u4eba\u540d\u5168\u90e8\u7576\u6210 OOV \u8655\u7406\uff0c\u4e0d\u50c5\u4eba\u540d\u9019\u500b\u8a5e\u5f59\u8fa8\u8b58\u4e0d\u51fa\u4f86\u4ee5\u5916\uff0c\u4e5f\u6703\u5f71\u97ff\u5305\u542b\u4eba\u540d\u7684 \u53e5\u5b50\u7684\u8fa8\u8b58\u80fd\u529b\uff0c\u56e0\u6b64\uff0c\u672c\u5be6\u9a57\u5c07\u81f4\u529b\u65bc\u63d0\u5347\u8a9e\u97f3\u8fa8\u8b58\u5668\u5c0d\u4eba\u540d\u4e4b\u8fa8\u8b58\u7387\uff0c\u85c9\u7531\u5c0d\u8a9e\u8a00 \u6a21\u578b\u4e4b\u8a13\u7df4\u8a9e\u6599\u52a0\u5165\u4eba\u540d\u8cc7\u8a0a\uff0c\u4ee5\u671f\u671b\u63d0\u9ad8\u8fa8\u8b58\u5668\u5c0d\u65bc\u4eba\u540d\u4e4b\u8fa8\u8b58\u7387\u3002 \u5716 \u4e00\u3001\u6bd4\u8cfd\u524d\u5341\u540d\u4e4b\u6210\u7e3e(\u5716\u4f8b\u7531\u5de6\u81f3\u53f3\u5206\u5225\u70ba\u5b57\u932f\u8aa4\u7387(CER)\u3001\u8a5e\u932f\u8aa4\u7387(WER)\u8207\u4eba \u540d\u8a5e\u5f59\u932f\u8aa4\u7387(name-WER)) (\u4e8c) \u7814\u7a76\u65b9\u5411 \u672c\u7814\u7a76\u5c07\u81f4\u529b\u65bc\u63d0\u5347\u8a9e\u97f3\u8fa8\u8b58\u5668\u4e4b\u8fa8\u8b58\u7387\u8207\u8fa8\u8b58\u901f\u5ea6\uff0c\u4e26\u85c9\u7531\u8a13\u7df4\u8cc7\u6599\u7684\u6539\u5584\uff0c\u4f7f \u4e4b\u5f97\u4ee5\u8fa8\u8b58\u4ee5\u5f80\u6240\u4e0d\u80fd\u8fa8\u8b58\u4e4b\u4eba\u540d\u3002 \u672c\u7bc7\u8ad6\u6587\u9996\u5148\u900f\u904e\u8abf\u6574\u8072\u5b78\u6a21\u578b\u67b6\u69cb\u3001\u8fad\u5178\u5927\u5c0f\u3001\u8a9e\u8a00\u6a21\u578b\u53c3\u6578\uff0c\u4f7f\u5f97\u8fa8\u8b58\u5668\u5728\u6574 \u500b\u8a9e\u97f3\u8fa8\u8a8d\u904e\u7a0b\u4e0a\u53d6\u5f97\u8fa8\u8b58\u7387\u8207\u8fa8\u8b58\u901f\u5ea6\u7684\u5e73\u8861\u3002\u5728\u5f97\u5230\u4e00\u500b\u8fa8\u8b58\u7387\u8207\u8fa8\u8b58\u901f\u5ea6\u517c\u5177\u7684 \u8fa8\u8b58\u5668\u4e4b\u5f8c\uff0c\u5c07\u91dd\u5c0d\u8fad\u5f59\u4e0d\u5728\u8fad\u5178\u88e1\u4e4b\u554f\u984c(Out of Vocabulary, OOV)\u505a\u6539\u5584\u3002 \u7531\u65bc\u4eba\u540d\u70ba OOV \u7d44\u6210\u4e2d\u8f03\u70ba\u91cd\u8981\u3001\u4e14\u51fa\u73fe\u983b\u7387\u6700\u9ad8\u7684\u8fad\u5f59\uff0c\u672c\u5be6\u9a57\u5c07\u91dd\u5c0d OOV \u88e1 \u7684\u4eba\u540d\u505a\u7279\u5225\u8655\u7406\uff0c\u9996\u5148\uff0c\u5148\u5efa\u7acb\u4e00 word-based Language Model \uff0c\u4e26\u627e\u51fa\u6240\u6709\u4eba\u540d(Word)\uff0c \u5c07\u4eba\u540d\u62c6\u89e3\u6210\u4e8c\u5230\u4e09\u500b\u5c0f\u55ae\u5143(Sub word unit)\uff0c\u56e0\u672c\u5be6\u9a57\u50c5\u8981\u6c42\u8fa8\u8b58\u51fa\u4e4b\u4eba\u540d\u7684\u767c\u97f3\u6b63 \u78ba\uff0c\u90e8\u5206\u672a\u88ab\u9078\u5165\u8fad\u5178\u4e4b\u4eba\u540d\u5c07\u6703\u88ab\u8f49\u6210\u97f3\u7bc0(Syllable)\uff0c\u4ee5\u97f3\u7bc0\u7684\u65b9\u5f0f\u52a0\u5165\u9032\u8fad\u5178\u88e1\uff0c \u4e4b\u5f8c\u518d\u5c0d\u6587\u5b57\u8a9e\u6599\u5f8c\u8655\u7406\uff0c\u5c07\u4eba\u540d\u8cc7\u8a0a\u52a0\u9032\u8a13\u7df4\u8a9e\u6599\uff0c\u5229\u7528\u7d71\u8a08\u7279\u6027\u8a08\u7b97\u5404\u7a2e\u4e0d\u540c\u53ef\u80fd \u4eba\u540d\u4e92\u76f8\u9023\u63a5\u7684\u6a5f\u7387\uff0c\u4ee5\u7522\u751f\u4e0d\u540c\u4eba\u540d\u4e4b\u7d44\u5408\uff0c\u6b64\u65b9\u6cd5\u53ef\u4ee5\u8b93\u8fa8\u8b58\u5668\u5c0d\u4eba\u540d\u7684\u8fa8\u8b58\u4e0d\u518d \u4fb7\u9650\u65bc\u8fad\u5178\u88e1\u51fa\u73fe\u904e\u7684\u4eba\u540d\uff0c\u4f7f\u5f97\u8a9e\u8a00\u6a21\u578b\u7372\u5f97\u8fa8\u8b58\u4eba\u540d\u7684\u80fd\u529b\u3002 \u4e8c\u3001 \u8a9e\u6599\u5eab\u4ecb\u7d39 \u672c\u7bc0\u5c07\u4ecb\u7d39\u7528\u65bc\u672c\u5be6\u9a57\u4e2d\u4e4b\u6240\u6709\u8a9e\u6599\u5eab\u3002\u5176\u4e2d\u7528\u4f86\u7576\u4f5c\u8a13\u7df4\u8072\u5b78\u6a21\u578b\u4e4b\u8a9e\u6599\u7684\u6709 TCC300\u3001NER \u53ca AIShell \u8a9e\u6599\u5eab\uff0c\u800c\u70ba\u4e86\u6e2c\u8a66\u672c\u5be6\u9a57\u4e4b\u8fa8\u8b58\u7cfb\u7d71\u5c0d\u65bc\u4e0d\u540c\u74b0\u5883\u7684\u8fa8\u8b58 \u80fd\u529b\uff0c\u56e0\u6b64\u5728\u6e2c\u8a66\u8a9e\u6599\u7684\u9078\u64c7\u4e0a\uff0c\u4f7f\u7528 TCC300 \u53ca NER \u8a9e\u6599\u5eab\uff0c\u5176\u4e2d NER \u70ba\u5ee3\u64ad\u8a9e\u6599\uff0c \u53c8\u53ef\u7d30\u5206\u70ba\u80cc\u666f\u4e7e\u6de8\u7121\u96dc\u8a0a\u4e4b\u4e7e\u6de8\u8a9e\u6599(Clean)\u4ee5\u53ca\u80cc\u666f\u6709\u4eba\u70ba\u96dc\u8a0a\u6216\u97f3\u6a02\u53c3\u96dc\u5176\u4e2d\u4e4b \u5176\u4ed6\u8a9e\u6599(Other)\u3002 (\u4e00) TCC300 \u8a9e\u6599\u5eab \u672c\u6b21\u5be6\u9a57\u4e2d\u6240\u4f7f\u7528\u7684 TCC300 \u9ea5\u514b\u98a8\u8a9e\u97f3\u8cc7\u6599\u5eab [1]\u662f\u7531\u570b\u7acb\u4ea4\u901a\u5927\u5b78(National Chiao Tung University, NCTU)\u3001\u570b\u7acb\u6210\u529f\u5927\u5b78(National Cheng Kung University, NCKU)\u3001 \u570b\u7acb\u53f0\u7063\u5927\u5b78(National Taiwan University, NTU)\u5171\u540c\u9304\u88fd\u800c\u6210\uff0c\u4e26\u4e14\u7531\u4e2d\u83ef\u6c11\u570b\u8a08\u7b97\u8a9e \u8a00\u5b78\u5b78\u6703(The Association for Computational Linguistics and Chinese Language Processing, ACLCLP)\u767c\u884c\uff0c\u6b64\u8a9e\u6599\u5eab\u5c6c\u65bc\u9ea5\u514b\u98a8\u6717\u8b80\u8a9e\u97f3\uff0c\u4e3b\u8981\u76ee\u7684\u70ba\u63d0\u4f9b\u53f0\u7063\u8154\u4e4b\u4e2d\u6587\u8a9e\u97f3\u8fa8 \u8a8d\u7814\u7a76\u4f7f\u7528\u3002\u5176\u4e2d\u8a13\u7df4\u8a9e\u6599\u7d04\u70ba 24.4 \u5c0f\u6642\uff0c304780 \u500b\u97f3\u7bc0\u6578\uff1b\u6e2c\u8a66\u8a9e\u6599\u7d04\u70ba 2.4 \u5c0f\u6642\uff0c 26357 \u500b\u97f3\u7bc0\u6578\u3002 (\u4e8c) NER \u8a9e\u6599\u5eab NER \u8a9e\u6599\u5eab [2]\uff0c\u5168\u540d\u70ba NER Manual Transcription Vo l1\uff0c\u70ba\u570b\u7acb\u81fa\u5317\u79d1\u6280\u5927\u5b78\u548c \u570b\u5bb6\u6559\u80b2\u5ee3\u64ad\u96fb\u53f0\u5408\u4f5c\u9304\u88fd\u4e4b\u8a9e\u6599\u5eab\uff0c\u4e3b\u8981\u76ee\u7684\u70ba\u5927\u91cf\u8f49\u5beb\u6559\u80b2\u96fb\u53f0\u4e4b\u7bc0\u76ee\uff0c\u7522\u751f\u7bc0\u76ee \u9010\u5b57\u7a3f\uff0c\u4ee5\u5efa\u7f6e\u5927\u898f\u6a21\u53f0\u7063\u8154\u4e4b\u8a9e\u6599\u5eab\uff0c\u5167\u5bb9\u5927\u90e8\u4efd\u70ba\u8ac7\u8a71\u6027\u7bc0\u76ee\uff0c\u591a\u70ba\u81ea\u767c\u6027 (Spontaneous)\u8a9e\u97f3\uff0c\u50c5\u5c11\u90e8\u5206\u70ba\u65b0\u805e\u5831\u5c0e\u4e4b\u6717\u8b80\u5f0f(Reading)\u8a9e\u97f3\u3002\u5176\u4e2d\u8a13\u7df4\u8a9e\u6599\u7d04\u70ba 111.5 \u5c0f\u6642\uff0c1715091 \u500b\u97f3\u7bc0\u6578\uff1b\u6e2c\u8a66\u8a9e\u6599\u53ef\u5206\u70ba\u4e7e\u6de8\u8a9e\u6599(Clean)\u7684 1.9 \u5c0f\u6642\uff0c33660 \u500b \u97f3\u7bc0\u6578\u8207\u5176\u4ed6\u8a9e\u6599(Other)\u7684 9.0 \u500b\u5c0f\u6642\uff0c133746 \u500b\u97f3\u7bc0\u6578\u3002 (\u4e09) AIShell AIShell \u8a9e\u6599\u5eab [3]\uff0c\u662f\u7531\u5317\u4eac\u5e0c\u723e\u8c9d\u6bbc\u79d1\u6280\u6709\u9650\u516c\u53f8\u91cb\u653e\u4e4b\u958b\u6e90\u8a9e\u97f3\u8cc7\u6599\u5eab\uff0c\u9304\u88fd \u5167\u5bb9\u6d89\u53ca\u667a\u80fd\u5c45\u5bb6\u3001\u7121\u4eba\u99d5\u99db\u7b49 11 \u9805\u9818\u57df\uff0c\u9304\u88fd\u904e\u7a0b\u7686\u5728\u5b89\u975c\u7684\u5ba4\u5167\u74b0\u5883\u3002 \u4f7f\u7528\u9ad8\u6548\u80fd\u9ea5\u514b\u98a8\u9304\u88fd\u800c\u6210\uff0c\u53d6\u6a23\u983b\u7387\u70ba 44100 Hz\uff0c\u5f8c\u964d\u4f4e\u53d6\u6a23\u983b\u7387\u81f3 16000 Hz\uff0c \u53d6\u6a23\u4f4d\u5143\u6578\u70ba 16 \u4f4d\u5143\uff0c\u7531 400 \u540d\u4f86\u81ea\u4e2d\u570b\u4e0d\u540c\u53e3\u97f3\u5730\u5340\u7684\u53c3\u8207\u8005\u9304\u88fd\u800c\u6210\uff0c\u6b64\u8a9e\u6599\u5eab \u6587\u672c\u7d93\u4eba\u5de5\u6821\u6b63\u904e\uff0c\u6b63\u78ba\u7387\u70ba 95%\u4ee5\u4e0a\u3002\u5176\u4e2d\u8a13\u7df4\u8a9e\u6599\u7d04\u70ba 162.4 \u5c0f\u6642\uff0c1862171 \u500b \u97f3\u7bc0\u6578\uff1b\u6e2c\u8a66\u8a9e\u6599\uff1a\u7d04\u70ba 16.6 \u5c0f\u6642\uff0c178041 \u500b\u97f3\u7bc0\u6578\u3002 \u4e09\u3001 \u6df1\u5c64\u985e\u795e\u7d93\u7db2\u8def\u6a21\u578b\u914d\u7f6e \u67b6\u69cb\u5982\u5716 \u4e8c\u3002 \u5716 \u4e8c\u3001TDNN-F \u6a21\u578b\u67b6\u69cb\u5716 \u56db\u3001 \u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u4e4b\u8a9e\u97f3\u8fa8\u8b58\u5668 \u5728\u89e3\u78bc\u7684\u904e\u7a0b\u4e2d\uff0c\u82e5\u9047\u5230\u8fad\u5178\u88e1\u6c92\u6709\u51fa\u73fe\u7684\u8a5e(Out of Vocabulary, OOV)\uff0c\u5247\u6b64\u4e00\u8fad \u5f59\u5c07\u6c38\u9060\u4e0d\u53ef\u80fd\u8fa8\u8a8d\u6b63\u78ba\uff0c\u6b64\u6642\uff0c\u8fa8\u8b58\u5668\u6703\u4ee5\u4e00\u76f8\u4f3c\u8b80\u97f3\u7684\u8a5e\u53d6\u4ee3\u4e4b\uff0c\u4f7f\u5f97\u6b64\u4e00 OOV \u9644\u8fd1\u7684\u8fad\u5f59\u5c07\u53d7\u5230\u5f71\u97ff\uff0c\u4e5f\u5c31\u662f\u5f62\u6210\u6240\u8b02\u7684\u6436\u8a5e\uff0c\u9032\u800c\u5c0e\u81f4\u932f\u8aa4\u7387\u7684\u63d0\u9ad8 4 \u3002\u56e0\u6b64\uff0cOOV \u5c0d\u65bc\u8a9e\u97f3\u8fa8\u8b58\u4f86\u8aaa\uff0c\u4e00\u76f4\u662f\u500b\u975e\u5e38\u91cd\u8981\u7684\u554f\u984c\u3002\u5728 OOV \u7d44\u6210\u4e2d\uff0c\u5c24\u4ee5\u5c08\u6709\u540d\u8a5e(Nb)\u7684 \u51fa\u73fe\u70ba\u6700\u5927\u5b97\uff0c\u53c8\u5c08\u6709\u540d\u8a5e\u4e2d\uff0c\u4eba\u540d\u70ba\u8f03\u5177\u610f\u7fa9\u4e14\u51fa\u73fe\u983b\u7387\u6700\u9ad8\u4e4b\u7a2e\u985e\uff0c\u56e0\u6b64\uff0c\u672c\u5be6\u9a57 \u5c07\u91dd\u5c0d\u4eba\u540d\u4e4b\u8fa8\u8b58\u7387\u505a\u6539\u5584\u3002 \u70ba\u4e86\u89e3\u6c7a\u4eba\u540d\u88ab\u7576\u4f5c OOV \u7684\u554f\u984c\uff0c\u672c\u8ad6\u6587\u5c07\u4eba\u540d\u5f9e\u5b57\u8f49\u97f3\u6210\u97f3\u7bc0(G2P)\uff0c\u4f7f\u539f\u8a13 \u7df4\u8a9e\u6599\u4e2d\u4eba\u540d\u7684\u4f4d\u7f6e\u662f\u4ee5\u97f3\u7bc0\u7684\u65b9\u5f0f\u5b58\u5728\uff0c\u4e4b\u5f8c\u518d\u4ee5\u88fd\u9020\u4eba\u540d\u96a8\u6a5f\u586b\u5165\u7684\u65b9\u5f0f\uff0c\u4f7f\u8a9e\u8a00 \u6a21\u578b\u5728\u539f\u4eba\u540d\u4f4d\u7f6e\u80fd\u770b\u5230\u8a31\u591a\u4e0d\u540c\u7a2e\u985e\u4e4b\u4eba\u540d\u3002\u8a73\u7d30\u5167\u5bb9\u5c07\u5728\u63a5\u4e0b\u4f86\u7684\u5c0f\u7bc0\u4e00\u4e00\u89e3\u8aaa\u3002 (\u4e00) \u6587\u5b57\u8a9e\u6599\u7c21\u4ecb \u672c\u7814\u7a76\u7528\u65bc\u8a13\u7df4\u8207\u7814\u6a21\u578b\u4e4b\u6587\u5b57\u8a9e\u6599\u5eab\u5171\u7d04 25 \u5104\u500b\u8a5e\u5f59\uff0c\u5305\u542b\u4ee5\u4e0b\uff1a \uf06c Chinese Gigaword\uff1a\u7531 Linguistic Data Consortium (LDC)\u6574\u5408\u767c\u884c\uff0c\u5167\u5bb9\u5305\u62ec\u53f0 \u7063\u4e2d\u592e\u793e\u3001\u5317\u4eac\u65b0\u83ef\u793e\u7b49\u570b\u969b\u65b0\u805e\u3002 \uf06c \u5176\u4ed6\u7531\u4ea4\u5927\u8a9e\u97f3\u8655\u7406\u5be6\u9a57\u5ba4\u8490\u96c6\u4e4b\u8a9e\u6599 (\u4e8c) \u6587\u5b57\u8a9e\u6599\u5f8c\u8655\u7406\u4e2d\u4e4b\u5f62\u97f3\u7fa9\u5206\u5408\u8a5e(variant word)\u8655\u7406 \u7531\u65bc\u8a13\u7df4\u8a9e\u6599\u591a\u4f86\u81ea\u65bc\u65b0\u805e\u8a9e\u6599\uff0c\u70ba\u4e86\u4f7f\u8a13\u7df4\u8cc7\u6599\u66f4\u63a5\u8fd1\u4e00\u822c\u4eba\u8aaa\u8a71\uff0c\u672c\u5be6\u9a57\u5c07\u65b7 \u8a5e\u5f8c\u7684\u6587\u5b57\u505a\u4e00\u4e9b\u7279\u5225\u7684\u8655\u7406\uff0c\u4f7f\u5176\u80fd\u66f4\u63a5\u8fd1\u53e3\u8a9e\u3002\u4f8b\u5982\u5c07\u6587\u7ae0\u4e2d\u51fa\u73fe\u7684\u79d1\u5b78\u7b26\u865f\u3001\u5ea6 \u91cf\u8861\u7b49\u7b49\u8f49\u70ba\u4e2d\u6587\uff0c\u8ce6\u4e88\u5176\u7d71\u4e00\u7684\u6587\u5b57\u8868\u793a\u65b9\u5f0f\uff0c\u4f7f\u5176\u88ab\u9078\u5165\u8fad\u5178\u5f8c\uff0c\u88ab\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4 \u5230\u3002\u5f8c\u8655\u7406\u4e2d\uff0c\u6700\u91cd\u8981\u4e4b\u6b65\u9a5f\u5982\u540c\u7fa9\u8a5e(variant word)\u4e4b\u4ee3\u63db\u5408\u4f75\u3002\u4e2d\u6587\u4e4b\u540c\u7fa9\u8a5e\u767e\u767e\u7a2e\uff0c \u80fd\u8868\u9054\u540c\u6a23\u610f\u601d\u4e4b\u8a5e\u8a9e\u7684\u9078\u64c7\u898b\u4ec1\u898b\u667a\uff0c\u4f8b\u5982\u540c\u7fa9\u7570\u97f3\u4e4b\u5468\u4e00\u3001\u661f\u671f\u4e00\uff1b\u540c\u7fa9\u540c\u97f3\u4e4b\u5468 \u4e00\u3001\u9031\u4e00\uff0c\u7576\u4eba\u5011\u5728\u4f7f\u7528\u9019\u4e9b\u540c\u7fa9\u8a5e\u69cb\u53e5\u6642\uff0c\u9019\u4e9b\u540c\u7fa9\u8a5e\u4e4b\u524d\u5f8c\u9023\u63a5\u8a5e\u5f59\u5e7e\u4e4e\u4e00\u6a21\u4e00\u6a23\uff0c \u56e0\u6b64\uff0c\u82e5\u672a\u5408\u4f75\u9019\u4e9b\u8a5e\u5f59\uff0c\u7d71\u8a08\u8a9e\u8a00\u6a21\u578b\u5c07\u6703\u8996\u9019\u4e9b\u8a5e\u5f59\u70ba\u500b\u5225\u4e0d\u540c\u7684\u8a5e\u5f59\uff0c\u9019\u5c07\u5f71\u97ff \u8fad\u5178\u6536\u7d0d\u8a5e\u5f59\u4e4b\u80fd\u529b\uff0c\u4e14\u5206\u6563\u8a5e\u53e5\u5728\u7d71\u8a08\u8a9e\u8a00\u6a21\u578b\u6240\u7d71\u8a08\u4e4b\u6a5f\u7387\u3002\u4f9d\u7167\u767c\u97f3\u4e4b\u7570\u540c\uff0c\u5927 \u81f4\u4e0a\u53ef\u4ee5\u5206\u70ba\u5169\u985e\uff0c\u5373\u767c\u97f3\u76f8\u540c\u8207\u767c\u97f3\u76f8\u7570\u4e4b\u8a5e\u985e\uff0c\u7f6e\u63db\u539f\u5247\u662f\u5efa\u7acb\u5728\u5b57\u7fa9\u76f8\u540c\u4e4b\u4e0a\uff0c \u7f6e\u63db\u7684\u76ee\u7684\u4e00\u662f\u70ba\u4e86\u8b93\u6587\u7ae0\u4e2d\u540c\u7fa9\u8a5e\u6b63\u898f\u5316\uff0c\u4ee5\u5229\u9078\u8a5e\u6642\u5bb9\u7d0d\u66f4\u591a\u8a5e\u5f59\uff0c\u4e8c\u662f\u4f7f\u5f97\u5408\u4f75 \u8a13\u7df4\u5f8c\u7684\u8a5e\u53ef\u4ee5\u7372\u5f97\u66f4\u591a\u7684\u8a13\u7df4\u8cc7\u6599\uff0c\u4f46\u662f\u5728\u8a9e\u8a00\u6a21\u578b\u5efa\u7acb\u5f8c\uff0c\u8fa8\u8b58\u7aef\u6703\u7522\u751f\u4e00\u500b\u72c0 \u6cc1\uff1a\u540c\u7fa9\u7570\u97f3\u4e4b\u8a5e\u5f59\u5728\u5408\u4f75\u5f8c\u5c07\u6703\u5c0e\u81f4\u67d0\u4e9b\u767c\u97f3\u5728\u89e3\u78bc\u5716\u8def\u5f91\u4e0a\u6d88\u5931\uff0c\u5c0e\u81f4\u89e3\u78bc\u6642\u7121\u6cd5 \u88ab\u641c\u5c0b\u5230\uff0c\u5982\u7bc4\u4f8b\u4e2d\u7684\u300c\u79ae\u62dc\u4e00\u300d\u88ab\u7f6e\u63db\u6210\u300c\u9031\u4e00\u300d \uff0c\u6545\u5728\u8a9e\u8a00\u6a21\u578b\u4e2d\u7121\u6cd5\u627e\u5230\u300c\u79ae\u62dc \u4e00\u300d\u9019\u500b\u8a5e\u5f59\uff0c\u56e0\u6b64\u6211\u5011\u5728\u8a9e\u8a00\u6a21\u578b\u5efa\u7f6e\u7684\u6700\u5f8c\u4e00\u6b65\uff0c\u9700\u8981\u8655\u7406\u4e0d\u540c\u767c\u97f3\u4e4b\u540c\u7fa9\u7570\u5b57\u8a5e \u7684\u7f6e\u63db\uff0c\u5c07\u300c\u9031\u4e00\u300d\u5c55\u958b\u6210\u300c\u9031\u4e00\u300d \u3001 \u300c\u661f\u671f\u4e00\u300d\u53ca\u300c\u79ae\u62dc\u4e00\u300d \u3002\u5982\u8868 \u4e00\u3002 \u8868 \u4e00\u3001\u540c\u7fa9\u8a5e\u4fee\u6539\u7bc4\u4f8b \u5f62\u97f3\u7fa9\u5206\u5408\u8a5e\u985e\u578b \u7f6e\u63db\u524d\u6587\u5b57 \u7f6e\u63db\u5f8c\u6587\u5b57 \u767c\u97f3\u76f8\u540c \u5468\u4e00 \u9031\u4e00 \u767c\u97f3\u76f8\u7570 \u79ae\u62dc\u4e00\u3001\u661f\u671f\u4e00 \u9031\u4e00 \u672c\u5be6\u9a57\u4f7f\u7528\u5be6\u9a57\u5ba4\u9577\u671f\u7d2f\u7a4d\u8490\u96c6\u4e4b 3160 \u500b\u540c\u7fa9\u8a5e\u4ee3\u63db\uff0c\u5728 7 \u500b\u4e0d\u540c\u4e4b\u8a9e\u6599\u96c6\u4e0a\u505a (\u4e09) \u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u4e4b\u5efa\u7acb \u8a9e\u97f3\u8fa8\u8b58\u5668\u4e4b\u672c\u8cea\u662f\u9054\u6210\u8072\u97f3\u8f49\u6587\u5b57\u4e4b\u76ee\u7684\uff0c\u5c0d\u65bc\u65e5\u5e38\u5c0d\u8a71\u4e4b\u53e5\u5b50\uff0c\u5c07\u5b57\u8a5e\u8fa8\u8a8d\u5b8c \u5168\u6b63\u78ba\u662f\u975e\u5e38\u5408\u7406\u4e14\u5bb9\u6613\u9054\u6210\u7684\uff0c\u4f46\u662f\u7576\u807d\u5230\u4e00\u4eba\u540d\u6642\uff0c\u5373\u4f7f\u662f\u4eba\u5de5\u8f49\u5beb\uff0c\u5c0d\u65bc\u5b57\u8a5e\u4e4b \u9078\u64c7\uff0c\u4ecd\u662f\u7121\u4e00\u500b\u767e\u5206\u4e4b\u767e\u6b63\u78ba\u4e4b\u9078\u64c7\uff0c\u50c5\u80fd\u4f9d\u65e5\u5e38\u751f\u6d3b\u6240\u807d\u5230\u4e4b\u5e38\u898b\u4eba\u540d\u9078\u5b57\u9078\u8a5e\uff0c \u52a0\u4e0a\u4e2d\u6587\u4e4b\u540c\u97f3\u7570\u5b57\u8a5e\u5be6\u5728\u662f\u591a\u5982\u725b\u6bdb\uff0c\u56e0\u6b64\uff0c\u91dd\u5c0d\u8fa8\u8b58\u5668\u5c0d\u65bc\u4eba\u540d\u4e4b\u8fa8\u8b58\u7d50\u679c\uff0c\u50c5\u8981 \u6c42\u5176\u97f3\u7bc0\u6b63\u78ba\uff0c\u4f8b\u5982\u300c\u738b\u5c0f\u660e\u300d\u9019\u500b\u4eba\u540d\u5c07\u4ee5\"wang xiao ming\"\u7684\u5f62\u5f0f\u88ab\u8fa8\u8b58\u5668\u89e3\u78bc\u51fa \u4f86\u3002\u53e6\u5916\uff0c\u672c\u7814\u7a76\u5c07\u4eba\u540d\u5206\u6210\u5169\u985e\uff1a\u5e38\u51fa\u73fe\u4e4b\u4eba\u540d(Somebody)\u8207\u4e0d\u5e38\u51fa\u73fe\u4e4b\u4eba\u540d (Nobody)\uff0c\u6839\u64da\u672c\u5be6\u9a57\u4e4b 25 \u5104\u8a5e\u8a9e\u6599\u5eab\u7d71\u8a08\uff0c\u4eba\u540d\u591a\u70ba\u653f\u6cbb\u4eba\u7269\u6216\u65b0\u805e\u4e0a\u5e38\u5831\u5c0e\u4e4b\u4eba \u7269\uff0c\u91dd\u5c0d\u6b64\u5e38\u51fa\u73fe\u4e4b\u4eba\u540d\uff0c\u53d6\u7d71\u8a08\u6578\u524d 5000 \u540d\u653e\u5165\u8fad\u5178\u4e2d\uff0c\u4f7f\u5176\u80fd\u88ab\u4ee5\u8a5e\u7684\u5f62\u5f0f\u89e3\u78bc \u51fa\u4f86\uff0c\u5269\u4e0b\u4e4b\u4eba\u540d\uff0c\u5c07\u62c6\u6210\u591a\u500b\u97f3\u7bc0\u653e\u5165\u8fad\u5178\u4e2d\uff0c\u4f8b\u5982\u300c\u738b\u5c0f\u660e\u300d\u9019\u4e00\u500b\u4e09\u5b57\u8a5e\u5c07\u88ab\u62c6 \u89e3\u6210\"wang xiao ming\"\u4e09\u500b\u97f3\u7bc0\u653e\u5165\u8fad\u5178\u4e2d\uff0c\u4e5f\u5c31\u662f\u8aaa\uff0c\u8fad\u5178\u4e2d\u6709\u95dc\u4eba\u540d\u7684\u90e8\u5206\u5c07\u7531 5000 \u500b\u4e2d\u6587\u4eba\u540d\u52a0\u4e0a 411 \u500b\u97f3\u7bc0 6 \u6240\u7d44\u6210\u3002\u63a5\u4e0b\u4f86\u5c07\u4ecb\u7d39\u672c\u5be6\u9a57\u5982\u4f55\u5efa\u7acb\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\uff0c\u4f7f \u539f\u672c\u88ab\u8fa8\u8b58\u5668\u7576\u6210 OOV \u4e4b\u4eba\u540d\u80fd\u4ee5\u97f3\u7bc0\u7684\u5f62\u5f0f\u8fa8\u8b58\u51fa\u4f86\u3002 1. \u627e\u51fa\u8a9e\u6599\u4e2d\u6240\u6709\u4eba\u540d\u4f4d\u7f6e \u70ba\u4e86\u627e\u51fa\u6587\u672c\u4e2d\u51fa\u73fe\u4e4b\u4eba\u540d\u4e26\u7f6e\u63db\u70ba#Name\uff0c\u9700\u5148\u5c07\u8a13\u7df4\u8a9e\u6599\u4e2d\u6240\u6709\u4eba\u540d\u4f4d\u7f6e\u627e\u51fa \u4f86\uff0c\u672c\u7814\u7a76\u8a02\u5b9a\u4e86\u4e00\u5957\u641c\u5c0b\u898f\u5247\uff0c\u9996\u5148\uff0c\u5229\u7528\u65b7\u8a5e\u5668\u6240\u6a19\u8a18\u51fa\u4e4b\u8a5e\u6027(Part of Speech, POS)\uff0c\u6311\u51fa\u6240\u6709\u5c08\u6709\u540d\u8a5e(Nb)\uff0c\u63a5\u4e0b\u4f86\uff0c\u67e5\u8a62\u59d3\u6c0f\u6392\u540d\u524d\u4e09\u767e\u540d 7 \u4f5c\u70ba\u5224\u65b7\u4f9d\u64da\uff0c\u5047\u5982 \u4e00\u5c08\u6709\u540d\u8a5e\u7684\u958b\u982d\u51fa\u73fe\u5728\u524d\u4e09\u767e\u59d3\u6c0f\u8868\u88e1\uff0c\u4e14\u5f8c\u7e8c\u9023\u63a5\u5b57\u6578\u70ba\u4e00\u81f3\u4e8c\u5b57\uff0c\u5247\u5224\u65b7\u70ba\u4eba \u540d\uff0c\u4e26\u7528#Name \u53d6\u4ee3\u4e4b(\u5982\u5716 \u56db)\uff0c\u4ee5\u6a19\u8a18\u6b64\u8655\u70ba\u4e00\u4eba\u540d\uff0c\u4f9b\u5f8c\u7e8c\u5c55\u958b\u4eba\u540d\u4e4b\u8655\u7406\u3002 \u4f9d\u6b64\u898f\u5247\u641c\u5c0b\u5728\u672c 25 \u5104\u8a13\u7df4\u8a9e\u6599\u4e2d\uff0c\u7b26\u5408\u4eba\u540d\u7684\u500b\u6578\uff0c\u7e3d\u5171\u7d04\u70ba 4100 \u842c\u500b\uff0c\u7d04\u5360 \u7e3d\u8a13\u7df4\u8a9e\u6599\u4e4b 1.6%\uff0c\u5b57\u8a5e\u4e4b\u6578\u91cf\u50c5\u6b21\u65bc\u300c\u7684\u300d(\u5982\u4e0b\u5716 \u4e94)\u3002\u7531\u6b64\u53ef\u518d\u6b21\u770b\u51fa\u4eba\u540d\u5728\u4e2d \u6587\u8a9e\u53e5\u4e2d\u4e4b\u91cd\u8981\u6027\uff0c\u4e0d\u5bb9\u5ffd\u8996\uff0c\u82e5\u4e0d\u5c0d\u4eba\u540d\u5b57\u8a5e\u505a\u5c08\u9580\u7684\u8655\u7406\uff0c\u50c5\u4ee5\u4e00\u822c\u7684\u8fa8\u8b58\u5668\u8fa8\u8a8d\uff0c \u4e14\u5047\u8a2d\u4eba\u540d\u4e4b\u8fa8\u8b58\u7387\u70ba 0%\uff0c\u5247\u6240\u6709\u4eba\u540d\u5c07\u88ab\u7576\u6210 OOV \u8655\u7406\uff0c\u4e26\u9020\u6210\u932f\u4e00\u500b\u4eba\u540d\u5c0e\u81f4 \u524d\u5f8c\u5b57\u8a5e\u9078\u8a5e\u932f\u8aa4\u4e4b\u9023\u9396\u53cd\u61c9\uff0c\u5982\u8868 \u4e8c\uff0c\u5d14\u84c9\u829d\u70ba\u4eba\u540d\uff0c\u6b64\u8655\u8fa8\u8b58\u5668\u56e0\u70ba\u672a\u5728\u8a9e\u8a00\u6a21 \u578b\u627e\u5230\u300c\u5d14\u84c9\u829d\u300d\u4e4b\u524d\u5f8c\u9023\u63a5\u8a5e\u4e4b\u6a5f\u7387\uff0c\u800c\u5c07\u9019\u4e00\u500b\u4e09\u5b57\u8a5e\u89e3\u78bc\u70ba\u4e00\u500b\u4e00\u5b57\u8a5e\u300c\u63a8\u300d\u52a0 \u4e0a\u4e00\u500b\u4e8c\u5b57\u8a5e\u300c\u878d\u8cc7\u300d \uff0c\u5c0e\u81f4\u7d71\u8a08\u6a21\u578b\u5c07\u8a08\u7b97\u300c\u63a8\u300d\u8207\u300c\u878d\u8cc7\u300d\u4e4b\u524d\u5f8c\u5b57\u8a5e\u9023\u63a5\u6a5f\u7387\uff0c \u9020\u6210\u5b57\u8a5e\u932f\u8aa4\u63a5\u4e8c\u9023\u4e09\u7684\u50b3\u905e\u4e0b\u53bb\uff0c\u4e5f\u5c31\u662f\u5f62\u6210\u6240\u8b02\u7684\u6436\u8a5e\u554f\u984c\u3002\u4ee5\u672a\u7d93\u8655\u7406\u4e4b 12 \u842c \u8a5e\u6240\u5efa\u7acb\u4e4b\u8a9e\u8a00\u6a21\u578b\u4e2d\uff0c\u7d71\u8a08\u5206\u6790\u5176\u8fa8\u8b58\u7b54\u6848\u53ef\u77e5\uff0c\u4e00\u500b\u4e8c\u81f3\u4e09\u5b57\u4eba\u540d\u5e73\u5747\u6703\u9020\u6210\u5de6\u53f3 \u9644\u8fd1 2.03 \u500b\u8a5e\u7684\u932f\u8aa4\uff0c\u4e5f\u5c31\u662f\u8aaa\uff0c\u8a13\u7df4\u8a9e\u6599\u4e2d\u4e4b 1.6%\u4eba\u540d\u6700\u9ad8\u53ef\u9020\u6210\u7d04 3.25%\u4e4b\u8a5e\u932f \u8aa4\u7387\uff0c\u9019\u5c0d\u8fa8\u8b58\u5668\u4f86\u8aaa\uff0c\u5be6\u5728\u662f\u4e00\u4e0d\u5bb9\u5c0f\u89b7\u4e4b\u8a5e\u932f\u8aa4\u7387\u3002 \u8868 \u4e8c\u3001\u4eba\u540d\u89e3\u78bc\u932f\u8aa4\u5c0e\u81f4\u6436\u8a5e\u4e4b\u554f\u984c \u6b63\u78ba\u7b54\u6848 \u2026 \u6b64 \u6848 *** *** \u5d14\u84c9\u829d \u5df2\u7d93 \u52dd\u8a34 \u5728\u671b \u2026 \u8fa8\u8b58\u7b54\u6848 \u2026 \u6b64 \u6848 \u63a8 \u878d\u8cc7 \u4ee5\u53ca \u56e0 \u52dd\u8a34 \u5728\u671b \u2026 \u5716 \u56db\u3001\u627e\u51fa\u4eba\u540d\u4e4b\u4f4d\u7f6e\u4e26\u4ee3\u63db\u70ba#Name \u5716 \u4e94\u3001\u5b57\u8a5e\u6578\u7d71\u8a08\u524d\u5341\u540d 2. class-based \u4eba\u540d\u8a9e\u8a00\u6a21\u578b \u70ba\u4e86\u4f7f\u8a13\u7df4\u8a9e\u6599\u539f\u53e5\u4e2d#Name token \u8655\u4e4b\u4eba\u540d\u7a2e\u985e\u66f4\u591a\u5143\uff0c\u672c\u5be6\u9a57\u8907\u88fd\u597d\u5e7e\u4efd\u8a13\u7df4 \u8a9e\u6599\uff0c\u4e26\u6839\u64da\u5728\u524d\u4e00\u5c0f\u7bc0(\u4e94\u4e4b\u4e09\u4e4b 1)\u6240\u627e\u5230\u7684\u6240\u6709#Name token \u4f4d\u7f6e\uff0c\u5c07\u6240\u6709\u627e\u5230\u7684\u4eba \u540d ( \u7d04 4100 \u842c\u500b) \u4ee5\u96a8\u6a5f\u4e82\u6578\u7684\u65b9\u5f0f\u586b\u5165\uff0c\u82e5\u586b\u5165\u4e4b\u5b57\u8a5e\u70ba\u8a9e\u6599\u4e2d\u5e38\u51fa\u73fe\u4eba\u540d (Somebody)\uff0c\u4ee5\u5b57\u8a5e(word)\u586b\u5165\uff1b\u82e5\u586b\u5165\u4e4b\u5b57\u8a5e\u70ba\u8a9e\u6599\u4e2d\u4e0d\u5e38\u51fa\u73fe\u4e4b\u4eba\u540d(Nobody)\uff0c\u5247 \u4ee5\u97f3\u7bc0\u586b\u5165\uff0c\u4e4b\u5f8c\u518d\u5c07\u9019\u4e9b\u8a9e\u6599\u62ff\u53bb\u505a LM training\u3002\u5982\u5716 \u516d\u3002 \u4eca\u5929_Nd \u738b\u5c0f\u660e_Nb \u53bb_D \u4e0a\u5b78_VA \u4eca\u5929_Nd #Name_Nb \u53bb_D \u4e0a\u5b78_VA Pos tagging & Name extraction \u5716 \u516d\u3001\u88fd\u9020\u4eba\u540d\u586b\u5165#Name \u4e4b\u4f4d\u7f6e \u7d93\u7531\u4e0d\u65b7\u7684\u4ee5\u4e82\u6578\u65b9\u5f0f\u5c07\u4eba\u540d\u586b\u5165\u8a13\u7df4\u8a9e\u6599\u4e2d\uff0c\u5982\u6b64\u4e00\u4f86\uff0c\u539f\u8a9e\u8a00\u6a21\u578b\u4e2d\u7684\u4eba\u540d (#Name)\u4f4d\u7f6e\u5c07\u51fa\u73fe\u8a31\u591a\u4e0d\u540c\u7684\u4eba\u540d\uff0c\u4e5f\u5c31\u662f\u8aaa\uff0c\u8a9e\u8a00\u6a21\u578b\u4e2d\u7684\u4eba\u540d\u4f4d\u7f6e\u5c07\u4e0d\u518d\u50c5\u80fd\u770b \u5230\u539f\u8a13\u7df4\u8a9e\u6599\u4e2d\u6240\u51fa\u73fe\u4e4b\u4eba\u540d\u6a5f\u7387\uff0c\u800c\u662f\u985e\u4f3c\u5c07#Name \u9019\u500b\u6a19\u8a18(Token)\u5c55\u958b\u6210\u4e00\u500b\u64c1 \u6709\u8a31\u591a\u4eba\u540d\u6a5f\u7387\u7684 class(\u5982\u5716 \u4e03)\uff0c\u4f7f\u6211\u5011\u7684\u8a9e\u8a00\u6a21\u578b\u7d44\u5408\u6210 word-based(\u4e00\u822c\u5b57\u8a5e)\u52a0 class-based(\u4eba\u540d\u5b57\u8a5e)\u7684\u8a9e\u8a00\u6a21\u578b\u3002\u4e4b\u5f8c\u7576\u8fa8\u8b58\u5668\u9047\u5230\u4e00\u500b\u672a\u66fe\u51fa\u73fe\u5728\u8fad\u5178\u88e1\u4e4b\u4eba\u540d \u6642\uff0c\u5c07\u6709\u4e00\u5b9a\u6a5f\u7387\u88ab\u89e3\u78bc\u6210\u4eba\u540d\u5b57\u8a5e(\u82e5\u70ba Somebody)\u6216\u4eba\u540d\u97f3\u7bc0(\u82e5\u70ba Nobody)\u3002 \u5716 \u4e03\u3001\u4eba\u540d word \u5c55\u958b\u6210\u4eba\u540d class (\u56db) \u8abf\u6574\u4e00\u822c\u540d\u8a5e\u88ab\u7576\u6210\u4eba\u540d\u4e4b\u6a5f\u7387 \u5728\u4e4b\u524d\u66fe\u7d93\u63d0\u5230\u904e\u4eba\u540d\u51fa\u73fe\u4e4b\u6a5f\u7387\u975e\u5e38\u9ad8(\u5982\u5716 \u4e94)\uff0c\u50c5\u6b21\u65bc\u300c\u7684\u300d \uff0c\u4e5f\u5c31\u662f\u8aaa\uff0c\u8fa8 \u8b58\u5668\u6709\u5f88\u9ad8\u7684\u6a5f\u7387\u6703\u5c07\u4e00\u822c\u540d\u8a5e\u7576\u6210\u4eba\u540d\u4f86\u89e3\u78bc\uff0c\u82e5\u525b\u597d\u6b32\u89e3\u78bc\u4e4b\u8a5e\u5f59\u4e0d\u662f\u4eba\u540d\uff0c\u4e14\u6b64 \u4eba\u540d\u5bb9\u6613\u8207\u4e00\u822c\u540d\u8a5e\u6df7\u6dc6\u6642\uff0c\u6b64\u5c07\u4e0d\u53ea\u5f71\u97ff\u8a72\u4f4d\u7f6e\u8fa8\u8b58\u932f\u8aa4\uff0c\u66f4\u5c07\u5f71\u97ff\u524d\u5f8c\u8a5e\u5f59\u76f8\u9023\u63a5 \u4e4b\u6a5f\u7387\uff0c\u4f8b\u5982\uff1a\u6709\u4e00\u4eba\u540d\u300c\u4f55\u5e73\u300d\u5728\u8a13\u7df4\u8a9e\u6599\u4e2d\u51fa\u73fe\u7684\u6a5f\u7387\u975e\u5e38\u9ad8\uff0c\u7576\u6e2c\u8a66\u53e5\u5b50\u63d0\u5230\u300c\u6211 \u559c\u6b61\u548c\u5e73\u300d\u6642\uff0c\u6b64\u53e5\u4e4b\u610f\u601d\u61c9\u662f\u6307\u559c\u6b61\u548c\u5e73\u9019\u7a2e\u72c0\u614b\uff0c\u7136\u800c\uff0c\u8fa8\u8b58\u5668\u56e0\u4eba\u540d\u4e4b\u51fa\u73fe\u6a5f\u7387 \u4eca\u5929_Nd #Name_Nb \u53bb_D \u4e0a\u5b78_VA \u4eca\u5929_Nd \u738b\u5c0f\u660e_Nb \u53bb_D \u4e0a\u5b78_VA \u4eca\u5929_Nd Liang Hong Bing \u53bb_D \u4e0a\u5b78_VA \u4eca\u5929_Nd Wang Yih Ru \u53bb_D \u4e0a\u5b78_VA LM training \u4eca\u5929_Nd #Name_Nb \u53bb_D \u4e0a\u5b78_VA \u738b\u5c0f\u660e Liang Hong Bing Wang Yih Ru \u2026 \u2026 \u904e\u9ad8\u800c\u5c07\u4e00\u822c\u540d\u8a5e\u300c\u548c\u5e73\u300d\u89e3\u78bc\u70ba\u4eba\u540d\u300c\u4f55\u5e73\u300d \uff0c\u9032\u800c\u89e3\u78bc\u6210\u300c\u6211\u559c\u6b61\u4f55\u5e73\u300d \uff0c\u8868\u793a\u559c\u6b61 \u4e00\u500b\u53eb\u505a\u4f55\u5e73\u7684\u4eba\uff0c\u6b64\u5c07\u5c0e\u81f4\u4e0d\u50c5\u300c\u548c\u5e73\u300d\u9019\u500b\u5b57\u8a5e\u8fa8\u8b58\u932f\u8aa4\uff0c\u8a72\u4f4d\u7f6e\u4e4b\u524d\u5f8c\u8a5e\u5f59\u9023\u63a5 \u6a5f\u7387\u4e5f\u6703\u5927\u4e0d\u76f8\u540c\uff0c\u5c07\u6709\u53ef\u80fd\u6703\u767c\u751f\u6436\u8a5e\u7684\u60c5\u6cc1\u3002 \u56e0\u6b64\uff0c\u70ba\u4e86\u964d\u4f4e\u6b64\u7a2e\u932f\u8aa4\u767c\u751f\uff0c\u672c\u5be6\u9a57\u984d\u5916\u8a13\u7df4\u4e00\u5b8c\u5168\u4e0d\u5305\u542b\u4eba\u540d\u4e4b\u8a9e\u8a00\u6a21\u578b\uff0c\u4e26 \u5b9a\u7fa9\u4e00\u6b0a\u91cd \u03b1\uff0c\u4ee5\u5167\u63d2\u7684\u65b9\u5f0f(interpolation)\u8207\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u5408\u4f75\uff0c\u8a66\u8457\u964d\u4f4e\u8a5e\u5f59\u88ab\u7576\u4eba \u540d\u4e4b\u6a5f\u7387\uff0c\u5982\u5716 \u516b\uff0c\u8a73\u7d30\u4e4b\u5be6\u9a57\u7d50\u679c\u5c07\u5728\u7b2c\u516d\u7ae0\u505a\u66f4\u8a73\u7d30\u7684\u8aaa\u660e\u3002 \u5716 \u516b\u3001\u964d\u4f4e\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u4e4b\u6b0a\u91cd\u5f8c\u4e4b\u8a9e\u8a00\u6a21\u578b\u3002 \u4e94\u3001 \u5be6\u9a57\u7d50\u679c\u5206\u6790\u8207\u8a0e\u8ad6 \u6b64\u5be6\u9a57\u4e4b\u6e2c\u8a66\u8a9e\u6599\u9664\u4e86 tcc300\u3001ner_Clean\u3001ner_Other \u5916\uff0c\u53e6\u52a0\u5165\u540c\u6a23\u70ba\u81ea\u767c\u6027\u8a9e \u97f3\u3001\u65b0\u805e\u5ee3\u64ad\u8a9e\u6599\u4e4b ner2_Clean(\u97f3\u7bc0\u6578 12171 \u500b\uff0c\u7d04 52 \u5206\u9418)\u3001ner2_Other(\u97f3\u7bc0\u6578 223385\uff0c\u7d04 12.6 \u5c0f\u6642)\u3001pilot_Clean(\u97f3\u7bc0\u6578 63695\uff0c\u7d04 4.3 \u5c0f\u6642)\u3001pilot_Other(\u97f3\u7bc0\u6578 107326\uff0c\u7d04 6.8 \u5c0f\u6642)\u3002 \u5728\u8a55\u4f30\u8fa8\u8b58\u7387\u4e4b\u524d\uff0c\u672c\u7ae0\u7bc0\u5c07\u5148\u8aaa\u660e\u56e0\u4eba\u540d\u800c\u5c0e\u81f4\u4e4b\u6436\u8a5e\u60c5\u5f62\u7684\u56b4\u91cd\u6027\uff0c\u4e26\u4ecb\u7d39\u932f \u8aa4\u7387\u4e4b\u8a08\u7b97\u65b9\u5f0f\uff0c\u6700\u5f8c\uff0c\u518d\u4ee5\u8abf\u6574\u8a9e\u8a00\u6a21\u578b\u4e4b\u5167\u63d2\u6b0a\u91cd\u7684\u65b9\u5f0f\uff0c\u63a7\u5236\u5404\u7a2e\u932f\u8aa4\u7684\u767c\u751f\u6a5f \u7387\u3002 \u5728\u524d\u9762\u7684\u7ae0\u7bc0\u66fe\u7d93\u63d0\u5230\u904e\uff0c\u82e5\u653e\u4efb\u4eba\u540d\u5728\u8fa8\u8b58\u5668\u4e2d\u89e3\u78bc\uff0c\u5247\u4eba\u540d\u5c07\u5f88\u6709\u53ef\u80fd\u88ab\u62c6\u89e3 \u6210\u7531\u8a31\u591a\u5c0f\u5b57\u6578\u8a5e\u7d44(Sub-word unit)\u6240\u7d44\u6210\uff0c\u800c\u9019\u4e9b sub-word unit \u53c8\u5c07\u5f71\u97ff\u524d\u5f8c\u5b57\u8a5e\u4e4b \u9023\u63a5\u6a5f\u7387\uff0c\u5c0e\u81f4\u932f\u8aa4\u50b3\u905e\u4e0b\u53bb\uff0c\u4f8b\u5982\u5148\u524d\u63d0\u5230\u7684\u4f8b\u5b50\u300c\u5d14\u84c9\u829d\u300d \uff0c\u4f46\u4e5f\u6709\u53ef\u80fd\u5f88\u5e78\u904b\u5730\uff0c \u8fa8\u8b58\u5668\u4ee5\u5b58\u5728\u65bc\u8fad\u5178\u4e4b\u767c\u97f3\u76f8\u4f3c\u4e4b\u4eba\u540d\u53d6\u4ee3\u4e4b\uff0c\u5728\u6b64\u60c5\u6cc1\u4e0b\uff0c\u932f\u8aa4\u50c5\u767c\u751f\u5728\u4eba\u540d\u672c\u8eab\uff0c \u4e26\u4e0d\u6703\u767c\u751f\u6436\u8a5e\u7684\u60c5\u6cc1\uff0c\u7d9c\u5408\u4e0a\u8ff0\u5169\u7a2e\u60c5\u5f62\uff0c\u5b9a\u7fa9\u4eba\u540d\u4e4b\u932f\u8aa4\u50b3\u905e\u7387(Error Propagation Rate, EPR)\uff0c\u8a08\u7b97\u65b9\u5f0f\u5982\u5f0f(4.1)\u3002\u7531\u65bc\u672c\u5be6\u9a57\u4e4b\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u6703\u5c07\u4eba\u540d\u89e3\u78bc\u70ba\u97f3\u7bc0\u6216\u6587 \u5b57\uff0c\u5728\u8a08\u7b97 EPR \u6642\u4ecd\u6703\u5148\u5c07\u8fa8\u8b58\u6b63\u78ba\u4e4b\u97f3\u7bc0\u8a18\u70ba\u932f\u8aa4\uff0c\u76f4\u5230\u5f8c\u9762\u518d\u505a\u6821\u6b63\u3002\u672c\u6e2c\u8a66\u8a9e \u6599\u4e4b\u7e3d\u8a5e\u6578\u70ba 387510 \u500b\uff0c\u7e3d\u4eba\u540d\u70ba 1812 \u500b\uff0c\u5728\u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u5f8c\uff0c\u6bd4\u8f03\u5404\u6e2c\u8a66\u8a9e\u6599 \u4e4b\u4eba\u540d\u932f\u8aa4\u50b3\u905e\u7387\uff0c\u5373\u6bcf\u55ae\u4f4d\u4eba\u540d\u6240\u9020\u6210\u5de6\u53f3\u8a5e\u5f59\u4e4b\u932f\u8aa4\u91cf\uff0c\u5982\u8868 \u4e09\u3002\u5728\u52a0\u5165\u4eba\u540d\u8a9e \u8a00\u6a21\u578b\u5f8c\uff0c\u50c5\u50c5\u5728\u932f\u8aa4\u50b3\u905e\u7387\u4e0a\uff0c\u5c31\u7531\u539f\u5148\u7684 2.03(\u672a\u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u524d)\u964d\u81f3 1.37\uff0c (4.3) \u91cb\u8fa8\u8b58\u5668\u6574\u9ad4\u4eba\u540d\u8a5e\u5f59\u51fa\u73fe\u4e4b\u6a5f\u7387\u5982\u5f0f(4.5)\u3002 \u5206\u5225\u8abf\u6574 \u03b1 \u70ba 1.0(\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u5b8c\u5168\u672a 8 \u9019\u88e1\u7684 1812 \u500b\u4eba\u540d\u5305\u542b\u5b58\u5728\u65bc\u8a5e\u5178\u4e4b\u4eba\u540d\u8207\u672a\u5b58\u5728\u65bc\u8a5e\u5178\u4e4b OOV \u4eba\u540d\u3002 \u4e5f\u5c31\u662f\u8aaa\uff0c\u539f\u5148\u4eba\u540d OOV \u70ba 1%\u5c07\u9020\u6210 2.03%WER \u7684\u7d50\u679c\uff0c\u5c07\u964d\u70ba 1.37%\u3002 (4.1) \u8868 \u4e09\u3001\u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u524d\u5f8c\u4e4b EPR \u6bd4\u8f03 \u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u524d \u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u5f8c tcc300 1.84 1.13 ner_Clean 2.25 1.52 ner_Other 2.03 1.37 ner2_Clean 1.78 1.23 ner2_Other 2.41 1.66 pilot_Clean 1.73 1.17 pilot_Other 2.38 1.71 Overall 2.03 1.37 \u5728\u8a55\u4f30\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u4e4b\u512a\u52a3\u6642\uff0c\u8207\u4e00\u822c\u8a5e\u5f59\u4e0d\u540c\uff0c\u4e14\u5728\u8a55\u4f30\u6642\uff0c\u672c\u5be6\u9a57\u50c5\u91dd\u5c0d\u4eba\u540d \u89e3\u78bc\u6210\u97f3\u7bc0\u4e4b\u6b63\u78ba\u7387\u505a\u89c0\u5bdf\uff0c\u63db\u53e5\u8a71\u8aaa\uff0c\u90a3\u4e9b\u672c\u4f86\u5c31\u5b58\u5728\u65bc\u8a5e\u5178\u88e1\u7684\u6709\u540d\u4eba\u7269 (Somebody)\u5c07\u81ea\u7136\u800c\u7136\u7684\u89e3\u78bc\u6210\u5b57\u8a5e\uff0c\u4e0d\u662f\u6b64\u5be6\u9a57\u4e4b\u91cd\u9ede\u3002\u5728\u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u5f8c\uff0c\u96d6 \u7136\u7406\u60f3\u4e0a\u53ef\u589e\u52a0\u539f\u672c\u88ab\u7576\u6210 OOV \u7684\u4eba\u540d\u88ab\u8fa8\u8b58\u51fa\u4f86\u7684\u6a5f\u7387\uff0c\u4f46\u540c\u6642\uff0c\u4e5f\u53ef\u80fd\u9020\u6210\u5e7e\u7a2e \u56e0\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u7684\u52a0\u5165\u6240\u5c0e\u81f4\u7684\u932f\u8aa4\uff0c\u56e0\u6b64\uff0c\u672c\u5be6\u9a57\u5b9a\u7fa9\u4e86\u5169\u7a2e\u985e\u578b\u7684\u932f\u8aa4\uff0c\u8207\u4e00\u7a2e\u985e \u578b\u7684\u6b63\u78ba\u65b9\u5f0f\uff0c\u8aaa\u660e\u5982\u4e0b\uff1a 1. False Alarm(type I error): \u4e00\u822c\u5b57\u8a5e\u88ab\u7576\u6210\u4eba\u540d\u89e3\u78bc\u3002\u9019\u88e1\u6307\u7684\u4e00\u822c\u5b57\u8a5e\u4e5f\u5305\u542b OOV\uff0c\u5373\u6240\u6709\u975e\u4eba\u540d\u5b57\u8a5e\u3002\u6b64\u985e \u932f\u8aa4\u662f\u7531\u65bc\u8a9e\u8a00\u6a21\u578b\u4e2d\u4eba\u540d\u8a5e\u5f59\u4e4b\u51fa\u73fe\u6a5f\u7387\u904e\u9ad8\uff0c\u5c0e\u81f4\u8fa8\u8b58\u5668\u9078\u7528\u4eba\u540d\u4f86\u66ff\u63db\u4e00\u822c\u8a5e \u5f59\u3002\u4f8b\uff1a \u300c\u9577\u8a71\u77ed\u8aaa\u300d\u88ab\u89e3\u78bc\u6210\u300c\u5f35\u83ef\u9813 \u8aaa\u300d \u3002 2. Wrong Detection(type II error): \u4eba\u540d\u88ab\u7576\u6210\u4e00\u822c\u5b57\u8a5e\u89e3\u78bc\u3002\u6b64\u985e\u932f\u8aa4\u662f\u7531\u65bc\u4eba\u540d\u8a5e\u5f59\u8207\u4e00\u822c\u5e38\u7528\u8a5e\u5f59\u4e4b\u767c\u97f3\u76f8\u4f3c\uff0c\u5c0e\u81f4 \u8fa8\u8b58\u5668\u89e3\u78bc\u932f\u8aa4\u3002\u4f8b\uff1a \u300c\u4f55\u5e73\u300d\u88ab\u89e3\u78bc\u6210\u300c\u548c\u5e73\u300d \uff1b \u300c\u937e\u570b\u4ec1\u300d\u88ab\u89e3\u78bc\u6210\u300c\u4e2d\u570b\u4eba\u300d \u3002 3. Hit: \u4eba\u540d\u88ab\u7576\u6210\u4eba\u540d\u89e3\u78bc\u3002\u4e0d\u8ad6\u89e3\u78bc\u51fa\u7684\u4eba\u540d\u97f3\u7bc0\u662f\u5426\u5b8c\u5168\u6b63\u78ba\uff0c\u53ea\u8981\u8a72\u8a5e\u5f59\u5728\u8fa8\u8b58\u5668\u4e2d\u662f \u88ab\u7576\u6210\u4eba\u540d\u4f86\u770b\u5f85\uff0c\u5c31\u7b97 Hit\u3002 \u70ba\u4e86\u8a55\u4f30\u8fa8\u8b58\u5668\u5728\u89e3\u78bc\u4eba\u540d\u8a5e\u5f59\u4e4b\u80fd\u529b\uff0c\u5b9a\u7fa9\u4e0b\u5217\u53c3\u6578\uff1a \u5b9a\u7fa9\u8fa8\u8b58\u5668\u6293\u5230\u4e4b\u6240\u6709\u4eba\u540d\u4e14\u70ba Hit \u7684\u6bd4\u7387 Precision \u70ba\uff1a (4.2) \u5b9a\u7fa9\u6e2c\u8a66\u8a9e\u6599\u4e2d\u6240\u6709\u4eba\u540d\u4e14\u70ba Hit \u7684\u6bd4\u7387 Recall \u70ba\uff1a \u4ee5\u8abf\u548c\u5e73\u5747\u6578\u7d9c\u5408\u4ee5\u4e0a\u5169\u500b\u53c3\u6578\uff0c\u8a08\u7b97\u5176 F1-score\uff1a (4.4) \u7576 Precision \u8d8a\u4f4e\uff0c\u8868\u793a False Alarm \u8d8a\u591a\uff0c\u975e\u4eba\u540d\u88ab\u7576\u6210\u4eba\u540d\u7684\u60c5\u6cc1\u8d8a\u56b4\u91cd\uff0c\u53cd\u4e4b\u4ea6\u7136\uff1b \u7576 Recall \u8d8a\u4f4e\uff0c\u8868\u793a Wrong Detection \u8d8a\u591a\uff0c\u4eba\u540d\u6c92\u6709\u88ab\u7576\u6210\u4eba\u540d\u7684\u60c5\u6cc1\u8d8a\u56b4\u91cd\uff0c\u53cd\u4e4b \u4ea6\u7136\u3002 \u5404\u6e2c\u8a66\u8a9e\u6599\u4e0a\u4e4b\u7d50\u679c\u5982\u4e0b\u5716 \u4e5d\u6240\u793a\uff1a \u5716 \u4e5d\u3001\u4eba\u540d\u8fa8\u8b58\u80fd\u529b\u4e4b\u8a55\u4f30 \u7531\u5be6\u9a57\u7d50\u679c\u5f97\u77e5\uff0c\u5404\u6e2c\u8a66\u8a9e\u6599\u4e4b Precision \u666e\u904d\u504f\u4f4e\uff0c\u9019\u662f\u56e0\u70ba\u8fa8\u8b58\u5668\u5bb9\u6613\u5c07 OOV \u8abf\u6574\u524d\uff0c\u5373\u5716 \u4e5d\u4f7f\u7528\u4e4b\u4eba\u540d\u8a9e\u8a00\u6a21\u578b)\u30010.5\u30010.1\uff0c\u7d50\u679c\u5982\u5716 \u5341\u6240\u793a\u3002\u7576\u4eba\u540d\u8a9e\u8a00\u6a21\u578b \u6b0a\u91cd\u8d8a\u4f86\u8d8a\u4f4e\u6642\uff0cPrecision \u56e0 OOV \u8d8a\u4f86\u8d8a\u4e0d\u5bb9\u6613\u88ab\u7576\u6210\u4eba\u540d\u8fa8\u8b58\u800c\u63d0\u9ad8\uff1bRecall \u5c07\u56e0 \u4eba\u540d\u8d8a\u4f86\u8d8a\u4e0d\u5bb9\u6613\u88ab\u7576\u6210\u4eba\u540d\u8fa8\u8b58\u800c\u964d\u4f4e\u3002 (4.5) \u5716 \u5341\u3001\u8abf\u6574\u8a5e\u5f59\u88ab\u7576\u6210\u4eba\u540d\u4e4b\u6a5f\u7387 \u524d\u8ff0\u5be6\u9a57\u5728\u8a08\u7b97 Hit \u7684\u6642\u5019\uff0c\u50c5\u8981\u6c42\u89e3\u78bc\u51fa\u7684\u7b54\u6848\u70ba\u4eba\u540d\uff0c\u800c Hit \u88e1\u53c8\u53ef\u518d\u7d30\u5206\u70ba \u4e09\u7a2e\u985e\u578b\uff1a1.\u97f3\u7bc0\u5b8c\u5168\u8fa8\u8b58\u6b63\u78ba\uff0c\u4f8b\u5982\u300c\u6881\u9d3b\u5f6c\u300d\u89e3\u78bc\u70ba\u300cliang_hong_bin\u300d\u3002 2.\u97f3\u7bc0\u76f8 \u4f3c\u4f46\u4e0d\u5b8c\u5168\u6b63\u78ba\uff0c\u4f8b\u5982\u300c\u90ed\u632f\u8208\u300d\u539f\u672c\u61c9\u89e3\u78bc\u70ba\u300cguo_zheng_xing\u300d\u4f46\u8fa8\u8b58\u5668\u89e3\u78bc\u70ba \u8072\u5b78\u6a21\u578b\u70ba\u4e0a\u8ff0(\u7b2c\u56db\u7ae0)\u7684 \u4e4b\u8072\u5b78\u6a21\u578b\uff0c\u8fa8\u8b58\u901f\u5ea6\u7684\u78ba\u6bd4\u50b3\u7d71\u4e4b LSTM \u5feb\u5f88\u591a\uff0c\u5373\u4f7f\u97f3\u7bc0\u8fa8\u8b58\u7387\u6709\u4e9b\u5dee\u8ddd\uff0c\u90a3\u4e9b \u300cguo_zheng_xin\u300d \uff0c\u4e00\u500b\u662f\u300c\u3112\u3127\u3125\u300d\u7684\u62fc\u97f3\uff0c\u4e00\u500b\u662f\u300c\u3112\u3127\u3123\u300d\u7684\u62fc\u97f3\uff0c\u7531\u65bc\u4eba\u5011\u5728 \u5dee\u8ddd\u5728\u5f8c\u7d1a\u8a9e\u8a00\u6a21\u578b\u89e3\u78bc\u5f8c\u4e5f\u986f\u5f97\u5fae\u4e4e\u5176\u5fae\u3002 \u5e73\u5e38\u8b1b\u8a71\u6642\u300c\u3125\u300d\u8ddf\u300c\u3123\u300d\u672c\u4f86\u5c31\u5206\u8fa8\u4e0d\u51fa\u4f86\uff0c\u6545\u6b64\u7a2e\u97f3\u7bc0\u76f8\u4f3c\u4e4b\u89e3\u78bc\u932f\u8aa4\u5be6\u5728\u662f\u7121\u53ef \u5728\u4eba\u540d\u8fa8\u8b58\u7387\u65b9\u9762\uff0c\u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u53ef\u4ee5\u4f7f\u4ee5\u524d\u88ab\u7576\u6210 OOV \u8655\u7406\u4e4b\u4eba\u540d\u88ab\u89e3\u78bc \u539a\u975e\u30023.\u7531\u4e00\u500b\u767c\u97f3\u76f8\u4f3c\uff0c\u4e14\u5b58\u5728\u65bc\u8a5e\u5178\u88e1\u4e4b\u4eba\u540d\u4ee3\u66ff\u4e4b\uff0c\u4f8b\u5982\u300c\u9ec3\u671d\u8208\u300d\u88ab\u8fa8\u8b58\u5668\u89e3 \u51fa\u4f86\uff0c\u4e0d\u4f46\u6551\u56de\u4e86\u4ee5\u524d\u4e0d\u5728\u8a5e\u5178\u88e1\u4e4b\u4eba\u540d\u4ee5\u5916\uff0c\u4e5f\u6e1b\u5c11\u4e86\u5de6\u53f3\u6436\u8a5e\u7684\u60c5\u5f62\u3002 \u78bc\u70ba\u300c\u9ec3\u662d\u661f\u300d \uff0c\u5176\u4e2d\uff0c \u300c\u9ec3\u662d\u661f\u300d\u70ba\u6709\u88ab\u9078\u5165\u8a5e\u5178\u4e2d\u4e4b\u5e38\u51fa\u73fe\u4eba\u540d\uff0c\u4e5f\u5c31\u662f\u524d\u9762\u7ae0\u7bc0\u6240 \u672c\u5be6\u9a57\u5efa\u69cb\u4e4b\u5927\u8a5e\u5f59\u4e2d\u6587\u8a9e\u97f3\u8fa8\u8b58\u5668\u7d71\u5c0d\u65bc\u4e7e\u6de8\u8a9e\u6599(Clean)\u6709\u7d55\u4f73\u7684\u8fa8\u8b58\u80fd\u529b\uff0c\u4f46 \u5b9a\u7fa9\u7684 Somebody\u3002 \u662f\u5c0d\u65bc\u566a\u97f3\u74b0\u5883(Other)\u8a9e\u6599\u4e4b\u8fa8\u8b58\u7387\u4ecd\u7136\u6709\u6539\u9032\u4e4b\u7a7a\u9593\uff0c\u5c24\u5176\u5728\u8072\u5b78\u6a21\u578b\u65b9\u9762\uff0c\u672c\u5be6\u9a57 \u7576\u6210\u4eba\u540d\u4f86\u8fa8\u8b58\uff0c\u800c\u6b64\u985e\u932f\u8aa4\u53c8\u88ab\u8a08\u5165 Precision \u4e4b\u7de3\u6545\uff0c\u8868 \u56db\u70ba\u5404\u6e2c\u8a66\u8a9e\u6599\uff0cOOV \u4e26\u6c92\u6709\u5c0d\u566a\u97f3\u505a\u7279\u5225\u7684\u8655\u7406\uff0c\u5c0e\u81f4\u5728\u566a\u97f3\u74b0\u5883\u4e0b\u4e4b\u97f3\u7bc0\u932f\u8aa4\u7387\u4ecd\u7136\u504f\u9ad8\uff0c\u9019\u6703\u4f7f\u5f97\u5373\u4f7f \u4f54 False Alarm \u7684\u6bd4\u4f8b\u3002 \u5f8c\u7d1a\u4e4b\u8a9e\u8a00\u6a21\u578b\u518d\u5f37\uff0c\u67d0\u4e9b\u5b57\u8a5e\u9084\u662f\u7121\u6cd5\u633d\u6551\u56de\u4f86\u4e4b\u60c5\u5f62\uff0c\u800c\u8a9e\u8a00\u6a21\u578b\u5982\u679c\u80fd\u514b\u670d\u8a18\u61b6 \u8868 \u56db\u3001False Alarm \u4e2d\u70ba OOV \u7684\u6bd4\u4f8b Tcc300 \u9ad4\u9650\u5236\u4e4b\u554f\u984c\u5f80 5-gram \u8a9e\u8a00\u6a21\u578b\u767c\u5c55\uff0c\u5c07\u6703\u4f7f\u8fa8\u8b58\u7387\u53c8\u518d\u66f4\u70ba\u63d0\u5347\uff0c\u5c24\u5176\u5728\u52a0\u5165\u4eba\u540d Ner_Clean Ner_Other Ner2_Clean Ner2_Other Pilot_Clean Pilot_Other 76.67% 29.63% 53.01% 78.57% 67.19% 79.22% 54.34% \u6b32\u63a7\u5236\u8fa8\u8b58\u5668\u5c07\u8a5e\u5f59\u7576\u6210\u4eba\u540d\u5b57\u8a5e\u8fa8\u8b58\u7684\u6a5f\u7387\uff0c\u4ee5\u8abf\u6574 Flase Alarm \u6216 Wrong Detection \u7684\u9ad8\u4f4e\uff0c\u672c\u5be6\u9a57\u5c07\u8a13\u7df4\u8a9e\u6599\u4e2d\u4e0d\u542b\u4eba\u540d\u4e4b\u53e5\u5b50\u7279\u5225\u5206\u96e2\u51fa\u4f86\u88fd\u505a\u4e00\u500b\u5b8c\u5168\u4e0d \u542b\u4eba\u540d\u4e4b\u8a9e\u8a00\u6a21\u578b\uff0c\u4e26\u7d66\u5b9a\u4e00\u500b\u6b0a\u91cd\u5c07\u6b64\u8a9e\u8a00\u6a21\u578b\u5167\u63d2\u9032\u539f\u672c\u7684\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u4e2d\uff0c\u4ee5\u7a00 \u8a9e\u8a00\u6a21\u578b\u5f8c\uff0c\u8a13\u7df4\u8a9e\u6599\u88e1\u4e4b\u4eba\u540d\u88ab\u5c55\u958b\u81f3 2 \u5230 3 \u500b\u97f3\u7bc0\u6578\uff0c\u4f7f\u7528 5-gram \u8a9e\u8a00\u6a21\u578b\u5c07\u53ef \u6700\u5f8c\uff0c\u70ba\u4e86\u89c0\u5bdf\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u5c0d\u65bc\u6574\u9ad4\u8fa8\u8b58\u7387\u7684\u5f71\u97ff\uff0c\u6211\u5011\u91cd\u65b0\u8a08\u7b97\u52a0\u5165\u4eba\u540d\u8a9e\u8a00 \u4ee5\u770b\u5f97\u66f4\u9060\uff0c\u4ee5\u589e\u52a0\u4eba\u540d\u8a5e\u5f59\u4e4b Hit \u6578\u3002\u6b64\u5916\uff0c\u76ee\u524d\u5be6\u9a57\u5ba4\u8fa8\u8b58\u5668\u5c0d\u65bc\u672a\u51fa\u73fe\u5728\u8a5e\u5178\u88e1 \u6a21\u578b\u5f8c\u8fa8\u8b58\u5668\u4e4b\u8a5e\u932f\u8aa4\u7387\uff0c\u9019\u88e1\u50c5\u5c07 Hit \u4e2d\u97f3\u7bc0\u5b8c\u5168\u89e3\u78bc\u6b63\u78ba\u4e4b\u4eba\u540d\u7576\u4f5c\u89e3\u78bc\u6b63\u78ba\uff0c\u4e26 \u4e4b\u4eba\u540d\u662f\u4ee5\u97f3\u7bc0\u7684\u5f62\u5f0f\u89e3\u78bc\u51fa\u4f86\uff0c\u4e5f\u8a31\u672a\u4f86\u80fd\u4ee5\u6a5f\u7387\u6216\u5176\u4ed6\u641c\u5c0b\u7684\u65b9\u5f0f\uff0c\u9078\u64c7\u5927\u5bb6\u666e\u904d \u8207\u672a\u52a0\u5165\u4eba\u540d\u8a9e\u8a00\u6a21\u578b\u524d\u4e4b\u8fa8\u8b58\u5668\u505a\u6bd4\u8f03\uff0c\u5be6\u9a57\u7d50\u679c\u5982\u4e0b\u5716 \u5341\u4e00\u6240\u793a\u3002\u6b64\u5be6\u9a57\u4f7f\u7528\u4e4b \u80fd\u63a5\u53d7\u7684\u5b57\u8a5e\u5448\u73fe\u3002</td></tr><tr><td>9 \u9019\u88e1\u6307\u50c5\u8a08\u5165 Hit \u4e2d\u97f3\u7bc0\u5b8c\u5168\u6b63\u78ba\u4e4b\u985e\u578b\u3002</td></tr></table>", |
| "type_str": "table" |
| } |
| } |
| } |
| } |
