| { |
| "paper_id": "2019", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T14:54:37.003900Z" |
| }, |
| "title": "Deep Learning Based Fake News AI Detection\uff1aEvidence From Taiwan News Report", |
| "authors": [ |
| { |
| "first": "Chih-Chien", |
| "middle": [], |
| "last": "\u6c6a\u5fd7\u5805", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taipei University", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "\u570b\u7acb\u53f0\u5317\u5927\u5b78\u8cc7\u8a0a\u7ba1\u7406\u7814\u7a76\u6240", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taipei University", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "Min-Yuh", |
| "middle": [], |
| "last": "\u6234\u654f\u80b2", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Tamkang University", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Day", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Tamkang University", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "Lin-Lung", |
| "middle": [], |
| "last": "\u80e1\u6797\u8fb3", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taipei University", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "\u570b\u7acb\u53f0\u5317\u5927\u5b78\u8cc7\u8a0a\u7ba1\u7406\u7814\u7a76\u6240", |
| "middle": [], |
| "last": "Hu", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taipei University", |
| "location": {} |
| }, |
| "email": "" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "In recent years, because the Internet acts as a medium, fake news can be quickly spread. Many countries have been seriously affected by fake news. making fake news detection become an important issue. This study collects two Taiwan refute rumors sites. And use the three methods of deep learning for fake news detection., Gated Recurrent Unit (GRU), Long Short Term Memory (LSTM), and Bidirectional Long Short Term Memory (BLSTM). The experimental results show that deep learning can be used in Taiwan's fake news detection, and the BLSTM method works best. Research experiments reduce the proportion of fake news, simulating 25%", |
| "pdf_parse": { |
| "paper_id": "2019", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "In recent years, because the Internet acts as a medium, fake news can be quickly spread. Many countries have been seriously affected by fake news. making fake news detection become an important issue. This study collects two Taiwan refute rumors sites. And use the three methods of deep learning for fake news detection., Gated Recurrent Unit (GRU), Long Short Term Memory (LSTM), and Bidirectional Long Short Term Memory (BLSTM). The experimental results show that deep learning can be used in Taiwan's fake news detection, and the BLSTM method works best. Research experiments reduce the proportion of fake news, simulating 25%", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "and 5% fake news ratios. Let the research sample be closer to the real situation. Finally, this study used a cross-data set test to understand the gap between practice and theory. ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Social media and fake news in the 2016 election", |
| "authors": [ |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Allcott", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Gentzkow", |
| "suffix": "" |
| } |
| ], |
| "year": 2017, |
| "venue": "Journal of economic perspectives", |
| "volume": "31", |
| "issue": "2", |
| "pages": "211--247", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "H. Allcott and M. Gentzkow, \"Social media and fake news in the 2016 election,\" Journal of economic perspectives, vol. 31, no. 2, pp. 211-36, 2017.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Fake news detection on social media: A data mining perspective", |
| "authors": [ |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Shu", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Sliva", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Tang", |
| "suffix": "" |
| }, |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Liu", |
| "suffix": "" |
| } |
| ], |
| "year": 2017, |
| "venue": "ACM SIGKDD Explorations Newsletter", |
| "volume": "19", |
| "issue": "1", |
| "pages": "22--36", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "K. Shu, A. Sliva, S. Wang, J. Tang, and H. Liu, \"Fake news detection on social media: A data mining perspective,\" ACM SIGKDD Explorations Newsletter, vol. 19, no. 1, pp. 22-36, 2017.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "Speech recognition with deep recurrent neural networks", |
| "authors": [ |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Graves", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Mohamed", |
| "suffix": "" |
| }, |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Hinton", |
| "suffix": "" |
| } |
| ], |
| "year": 2013, |
| "venue": "2013 IEEE international conference on acoustics, speech and signal processing", |
| "volume": "", |
| "issue": "", |
| "pages": "6645--6649", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "A. Graves, A.-r. Mohamed, and G. Hinton, \"Speech recognition with deep recurrent neural networks,\" in 2013 IEEE international conference on acoustics, speech and signal processing, 2013: IEEE, pp. 6645-6649.", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Long short-term memory", |
| "authors": [ |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Hochreiter", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Schmidhuber", |
| "suffix": "" |
| } |
| ], |
| "year": 1997, |
| "venue": "Neural computation", |
| "volume": "9", |
| "issue": "8", |
| "pages": "1735--1780", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "S. Hochreiter and J. Schmidhuber, \"Long short-term memory,\" Neural computation, vol. 9, no. 8, pp. 1735-1780, 1997.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "Framewise phoneme classification with bidirectional LSTM and other neural network architectures", |
| "authors": [ |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Graves", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Schmidhuber", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Neural Networks", |
| "volume": "18", |
| "issue": "5-6", |
| "pages": "602--610", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "A. Graves and J. Schmidhuber, \"Framewise phoneme classification with bidirectional LSTM and other neural network architectures,\" Neural Networks, vol. 18, no. 5-6, pp. 602-610, 2005.", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Empirical evaluation of gated recurrent neural networks on sequence modeling", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Chung", |
| "suffix": "" |
| }, |
| { |
| "first": "C", |
| "middle": [], |
| "last": "Gulcehre", |
| "suffix": "" |
| }, |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Cho", |
| "suffix": "" |
| }, |
| { |
| "first": "Y", |
| "middle": [], |
| "last": "Bengio", |
| "suffix": "" |
| } |
| ], |
| "year": 2014, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "arXiv": [ |
| "arXiv:1412.3555" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, \"Empirical evaluation of gated recurrent neural networks on sequence modeling,\" arXiv preprint arXiv:1412.3555, 2014.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "Hyperpartisan Facebook Pages Are Publishing False And Misleading Information At An Alarming Rate", |
| "authors": [ |
| { |
| "first": "C", |
| "middle": [], |
| "last": "Silverman", |
| "suffix": "" |
| }, |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Strapagiel", |
| "suffix": "" |
| }, |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Shaba", |
| "suffix": "" |
| }, |
| { |
| "first": "Ellei", |
| "middle": [], |
| "last": "Hall", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Singer-Vin", |
| "suffix": "" |
| } |
| ], |
| "year": null, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "C. Silverman, L. Strapagiel, H. Shaba, Ellei Hall, and J. Singer-Vin. \"Hyperpartisan Facebook Pages Are Publishing False And Misleading Information At An Alarming Rate.\" buzzfeednews. https://www.buzzfeednews.com/article/craigsilverman/partisan- fb-pages-analysis (accessed.", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "liar, liar pants on fire\": A new benchmark dataset for fake news detection", |
| "authors": [ |
| { |
| "first": "W", |
| "middle": [ |
| "Y" |
| ], |
| "last": "Wang", |
| "suffix": "" |
| } |
| ], |
| "year": 2017, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "arXiv": [ |
| "arXiv:1705.00648" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "W. Y. Wang, \"\" liar, liar pants on fire\": A new benchmark dataset for fake news detection,\" arXiv preprint arXiv:1705.00648, 2017.", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Automatic detection of fake news", |
| "authors": [ |
| { |
| "first": "V", |
| "middle": [], |
| "last": "P\u00e9rez-Rosas", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Kleinberg", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Lefevre", |
| "suffix": "" |
| }, |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Mihalcea", |
| "suffix": "" |
| } |
| ], |
| "year": 2017, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": { |
| "arXiv": [ |
| "arXiv:1708.07104" |
| ] |
| }, |
| "num": null, |
| "urls": [], |
| "raw_text": "V. P\u00e9rez-Rosas, B. Kleinberg, A. Lefevre, and R. Mihalcea, \"Automatic detection of fake news,\" arXiv preprint arXiv:1708.07104, 2017.", |
| "links": null |
| }, |
| "BIBREF10": { |
| "ref_id": "b10", |
| "title": "Detecting rumors from microblogs with recurrent neural networks", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Ma", |
| "suffix": "" |
| } |
| ], |
| "year": 2016, |
| "venue": "Ijcai", |
| "volume": "", |
| "issue": "", |
| "pages": "3818--3824", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "J. Ma et al., \"Detecting rumors from microblogs with recurrent neural networks,\" in Ijcai, 2016, pp. 3818-3824.", |
| "links": null |
| }, |
| "BIBREF11": { |
| "ref_id": "b11", |
| "title": "COMBATING DISINFORMATION: Detecting fake news with linguistic models and classification algorithms", |
| "authors": [ |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Sv\u00e4rd", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [], |
| "last": "Rumman", |
| "suffix": "" |
| } |
| ], |
| "year": 2017, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "M. Sv\u00e4rd and P. Rumman, \"COMBATING DISINFORMATION: Detecting fake news with linguistic models and classification algorithms,\" ed, 2017.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "TABREF0": { |
| "html": null, |
| "type_str": "table", |
| "text": "\u95dc\u9375\u8a5e\uff1a\u5047\u65b0\u805e\u3001\u5047\u65b0\u805e\u5075\u6e2c\u3001\u6df1\u5ea6\u5b78\u7fd2\u3001\u4eba\u5de5\u667a\u6167\u3001\u96d9\u5411\u9577\u77ed\u671f\u8a18\u61b6Keywords: Fake News, Fake News Detection, Deep Learning, Artificial Intelligence, Bidirectional Long Short Term Memory Shared Task \u7af6\u8cfd\u6240\u63d0\u4f9b 185,445 \u7b46\u8cc7\u6599\u3001P\u00e9rez-Rosas, et al.[10]\u900f\u904e GossipCop \u516b\u5366 \u6aa2\u6838\u7db2\u7ad9\u3001Ma, et al.[11].\u5fae\u535a(Weibo)\u7ba1\u7406\u4e2d\u5fc3\u63d0\u4f9b\u5047\u65b0\u805e\u5831\u544a\u8cc7\u6599\u96c6\u3001Sv\u00e4rd and Rumman[12]\u81ea\u884c\u8490\u96c6 201 \u7bc7\u7f8e\u570b\u65b0\u805e\u6587\u7ae0\uff0c\u5176\u4e2d\u6709 120 \u662f\u5047\u65b0\u805e\uff0c81 \u7bc7\u662f\u771f\u65b0\u805e\u3001 Score \u4f86\u5230 0.898\uff0c25%\u7684 F-Score \u4f86\u5230 0.836 \u4ee5\u53ca 5%\u7684 F-Score \u4f86\u5230 0.563\u3002", |
| "num": null, |
| "content": "<table><tr><td>\u4e86\u89e3\u6c7a\u5f8c\u5411\u4fee\u98fe\u524d\u5411\u7684\u554f\u984c\u3002 \u300cCofacts \u771f\u7684\u5047\u7684\u300d\u4ee5\u53ca\u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d \u3002 \u300cCofacts \u771f\u7684\u5047\u7684\u300d\u900f\u904e Line Chatbot \u4f86\u8490 \u672c\u7814\u7a76\u6703\u5c07\u4e09\u500b\u8cc7\u6599\u96c6\u90fd\u4ee5 80-20 \u6cd5\u5247\u5206\u6210 80%\u8a13\u7df4\u96c6\u8207 20%\u7684\u6e2c\u8a66\u96c6\uff0c\u5728\u8b80\u53d6\u8cc7\u6599 \u56db\u3001\u5be6\u9a57\u7d50\u679c \u672c\u7814\u7a76\u8a8d\u70ba\u5047\u65b0\u805e\u4f54\u6a23\u672c 50%\u4e26\u4e0d\u7b26\u5408\u771f\u5be6\u60c5\u6cc1\uff0c\u672c\u7814\u7a76\u53c8\u5be6\u9a57\u4e86\u5047\u65b0\u805e\u6bd4\u4f8b 25%\u4ee5 \u540c\u6bd4\u4f8b\u4e0b\u589e\u52a0\u8a13\u7df4\u8207\u6e2c\u8a66\u7b46\u6578\uff0c\u53ef\u4ee5\u589e\u52a0\u6a21\u578b\u7684\u6548\u7387\u3002 \u96c6\u7576\u4e2d\uff0c\u771f\u65b0\u805e\u7684\u4f86\u6e90\u4e26\u4e0d\u5920\u591a\uff0c\u5c0e\u81f4\u6a21\u578b\u5e38\u5c07\u771f\u65b0\u805e\u932f\u7576\u6210\u5047\u65b0\u805e\u3002\u800c\u5728 5%\u7684\u6a21\u578b</td></tr><tr><td>\u9598\u9580\u5faa\u74b0\u55ae\u5143(Gate Recurrent Unit, GRU)\u8207\u9577\u77ed\u671f\u8a18\u61b6\u985e\u4f3c\u3002\u6839\u64da Chung, et al. [7]\u7814\u7a76 \u96c6\u793e\u7fa4\u4e0a\u6d41\u50b3\u7684\u5047\u65b0\u805e\uff0c\u5728\u85c9\u7531\u7fa4\u773e\u8209\u5831\u7136\u5f8c\u5728\u7fa4\u773e\u6838\u5c0d\uff0c\u8b93\u5927\u5bb6\u4e00\u8d77\u6253\u64ca\u7db2\u8def\u4e0a\u7684\u8b20 \u5f8c\u5efa\u7acb\u51fa Token \u5b57\u5178\uff0c\u5b57\u5178\u4f9d\u64da\u51fa\u73fe\u6b21\u6578\u7684\u591a\u5be1\u5df2\u964d\u51aa\u65b9\u5f0f\u6392\u5e8f\uff0c\u63a5\u4e0b\u4f86\u4f7f\u7528 Token \u628a (\u4e00) \u8a13\u7df4\u8207\u6e2c\u8a66\u7b46\u6578 \u53ca 5%\uff0c\u4e26\u7167\u4e00\u6a23\u7684\u7814\u7a76\u65b9\u6cd5\u4f86\u5be6\u9a57\u3002\u672c\u7814\u7a76\u5728\u5be6\u9a57\u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d\u6642\u767c\u73fe\uff0c\u5728 5%\u5047\u65b0 \u7576\u4e2d Recall \u6709\u5927\u5e45\u5ea6\u4e0b\u964d\u5230 0.4\uff0c\u672c\u7814\u7a76\u8a8d\u70ba\u662f 5%\u6240\u653e\u5165\u5047\u65b0\u805e\u6e1b\u5c11\uff0c\u5c0e\u81f4\u6e96\u78ba\u6293\u51fa</td></tr><tr><td>\u6307\u51fa\u9598\u9580\u5faa\u74b0\u55ae\u5143\u6bd4\u8d77\u9577\u77ed\u671f\u8a18\u61b6\u6703\u4f7f\u7528\u66f4\u5c11\u7684\u8a18\u61b6\u3001\u6e1b\u5c11\u4e2d\u592e\u8655\u7406\u5668(CPU)\u7684\u904b\u7b97\u6642 \u8a00\u4ee5\u53ca\u5047\u65b0\u805e\u3002\u672c\u7814\u7a76\u5c07\u8cc7\u6599\u5b58\u81f3 SQL Server\uff0c\u81f3\u622a\u7a3f\u70ba\u6b62\u8cc7\u6599\u5eab\u771f\u65b0\u805e\u6709 6,107 \u7b46\u8cc7 \u65b0\u805e\u6587\u5b57\u8f49\u6210\u6578\u5b57\u9663\u5217\u3002\u7531\u65bc\u65b0\u805e\u5b57\u6578\u4e26\u4e0d\u4e00\u5b9a\uff0c\u70ba\u4e86\u8b93\u8f49\u63db\u5f8c\u7684\u5b57\u6578\u76f8\u540c\u3002\u672c\u7814\u7a76\u5c07 \u672c\u7814\u7a76\u5c07\u771f\u65b0\u805e\u8207\u5047\u65b0\u805e\u6bd4\u4f8b\u6210 1:1\uff0c\u5047\u65b0\u805e\u7684\u6bd4\u4f8b\u5c07\u6703\u4f54 50%\u3002\u63a5\u8457\u4ee5 80-20 \u6cd5\u5247 \u805e\u6bd4\u4f8b\u4e4b\u4e0b\uff0c\u53ef\u4ee5\u518d\u589e\u52a0\u8cc7\u6599\u7b46\u6578\u3002\u6240\u4ee5\u672c\u7814\u7a76\u53c8\u5be6\u9a57\u4e86\u5728\u76f8\u540c\u5047\u65b0\u805e\u6bd4\u4f8b\u5728 5%\u60c5\u6cc1 \u5047\u65b0\u805e\u4e0b\u964d\u3002</td></tr><tr><td>\u9593\u4ee5\u53ca\u52a0\u5feb\u6536\u6582\u7684\u901f\u5ea6\uff0c\u800c\u9598\u9580\u5faa\u74b0\u55ae\u5143\u5728\u8f03\u5c0f\u7684\u8cc7\u6599\u96c6\u7684\u8868\u73fe\u6bd4\u9577\u77ed\u671f\u8a18\u61b6\u4f86\u7684\u66f4\u52a0 \u6599\uff0c\u5047\u65b0\u805e\u6709 9,141 \u7b46\u8cc7\u6599\u3002 \u63a1\u7528\u622a\u9577\u88dc\u77ed\u7684\u65b9\u5f0f\uff0c\u5229\u7528\u524d\u9762\u6240\u8a08\u7b97\u7684\u8a5e\u6578\uff0c\u8a2d\u5b9a\u4e00\u56fa\u5b9a\u9577\u5ea6\u70ba n\uff0c\u82e5\u65b0\u805e\u6587\u5b57\u9577\u5ea6 \u5206\u6210\u8a13\u7df4\u8207\u6e2c\u8a66\u7b46\u6578\uff0c\u4e0b\u8868\u70ba\u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d\u8207\u300c\u771f\u7684\u5047\u7684\u300d\u8a13\u7df4\u8207\u6e2c\u8a66\u7b46\u6578\uff1a \u4e0b\uff0c\u6703\u6709\u4ec0\u9ebc\u6a23\u7684\u7d50\u679c\u3002\u4ee5\u4e0b\u662f\u672c\u7814\u7a76\u8a13\u7df4\u8207\u6e2c\u8a66\u7b46\u6578\u4ee5\u53ca\u5be6\u9a57\u7d50\u679c\uff1a \u7d9c\u4e0a\u6240\u8ff0\uff0c\u5982\u8981\u5c07\u6a21\u578b\u9032\u4e00\u6b65\u7684\u5be6\u52d9\u5316\uff0c\u9664\u4e86\u5047\u65b0\u805e\u8cc7\u6599\u8490\u96c6\u5916\uff0c\u5fc5\u9808\u8981\u8490\u96c6\u66f4\u591a</td></tr><tr><td>\u512a\u7570\u3002 \u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d\u662f\u4e00\u500b\u95e2\u8b20\u7684\u7db2\u7ad9\uff0c\u4e3b\u8981\u662f\u8b93\u6c11\u773e\u5c0d\u65bc\u53ef\u7591\u65b0\u805e\u4f86\u6b64\u8a62\u554f\uff0c\u7531\u5176\u4ed6\u6c11 \u5927\u65bc n\uff0c\u5247\u622a\u53bb\u5f8c\u9762\u7684\u6578\u5b57;\u53cd\u4e4b\u5982\u679c\u65b0\u805e\u6587\u5b57\u82e5\u5c0f\u65bc n\uff0c\u5247\u6703\u88dc 0 \u76f4\u5230 n \u7684\u9577\u5ea6\u3002\u4e4b\u5f8c \u8868\u4e94\u3001\u5047\u65b0\u805e\u6bd4\u4f8b 25%\u8207 5%\u8a13\u7df4\u8207\u6e2c\u8a66\u7b46\u6578 \u4f86\u6e90\u7684\u771f\u65b0\u805e\uff0c\u8b93\u6a21\u578b\u5728\u9762\u5c0d\u5404\u7a2e\u4e0d\u540c\u771f\u65b0\u805e\u4e4b\u6642\uff0c\u4e0d\u6703\u8aa4\u5c07\u65b0\u805e\u5224\u5225\u6210\u70ba\u5047\u65b0\u805e\u3002 \u8868\u4e8c\u3001\u5047\u65b0\u805e\u6bd4\u4f8b 50%\u8a13\u7df4\u8207\u6e2c\u8a66\u7b46\u6578 \u4e00\u3001\u7dd2\u8ad6 \u5047\u65b0\u805e\u7684\u5b9a\u7fa9\u70ba\u6709\u610f\u64b0\u5beb\u4e14\u8aa4\u5c0e\u65b0\u805e\u8b80\u8005\uff0c\u4f46\u9a57\u8b49\u662f\u865b\u5047\u7684\u65b0\u805e[1]\u3002\u8fd1\u5e74\u4f86\u5404\u570b\u90fd \u53d7\u5230\u5047\u65b0\u805e\u56b4\u91cd\u5f71\u97ff\uff0c\u4f8b\u5982\uff1a2016 \u7f8e\u570b\u7e3d\u7d71\u5927\u9078\u3001\u53f0\u7063\u885b\u751f\u7d19\u4e4b\u4e82\u3001\u53f0\u7063\u95dc\u897f\u6a5f\u5834\u5047\u65b0 \u805e\u4e8b\u4ef6\u7b49\u7b49[2]\u3002\u8fd1\u5e74\u7684\u5047\u65b0\u805e\u90fd\u662f\u900f\u904e\u7db2\u8def\u70ba\u5a92\u4ecb\u3002\u800c\u8981\u6aa2\u6e2c\u7db2\u8def\u4e0a\u7684\u5167\u5bb9\u7684\u771f\u507d\uff0c\u6700 \u539f\u59cb\u7684\u65b9\u5f0f\u662f\u900f\u904e\u300c\u624b\u52d5\u300d\u65b9\u5f0f[3]\uff0c\u4f7f\u5f97 Facebook \u8207 Google \u5169\u5bb6\u7db2\u8def\u5de8\u982d\u516c\u53f8\u90fd\u60f3\u8b93 \u5047\u65b0\u805e\u5075\u6e2c\u81ea\u52d5\u5316\u3002 \u904e\u5f80\u591a\u4f4d\u5b78\u8005\u66fe\u61c9\u7528\u6a5f\u5668\u5b78\u7fd2\u6216\u8005\u6df1\u5ea6\u5b78\u7fd2\u4f86\u9032\u884c\u5047\u65b0\u805e\u5075\u6e2c\uff0c\u5982\uff1aSupport Vector (\u4e8c) \u5047\u65b0\u805e\u8cc7\u6599\u5eab\u8490\u96c6 \u904e\u53bb\u5047\u65b0\u805e\u5075\u6e2c\u7814\u7a76\u4e2d\uff0c\u5927\u591a\u6578\u7814\u7a76\u90fd\u662f\u6839\u64da\u5c08\u5bb6\u6aa2\u6838\u3001\u7814\u7a76\u81ea\u884c\u8490\u96c6\u3001\u7fa4\u773e\u95e2\u8b20\u7db2 \u7ad9\u3001\u793e\u7fa4\u7db2\u7ad9\u63d0\u4f9b\u4ee5\u53ca\u7af6\u8cfd\u63d0\u4f9b\uff0c\u4e94\u7a2e\u65b9\u6cd5\u6240\u5efa\u7acb\u8d77\u8cc7\u6599\u5eab\u3002\u4f8b\u5982\uff1aBuzzFeed \u8a18\u8005 Silverman, et al. [8]\u5c0d\u91dd\u5c0d\u653f\u6cbb\u76f8\u95dc Facebook \u7c89\u7d72\u5c08\u9801\uff0c\u6240\u767c\u7684\u6587\u7ae0\u505a\u4e8b\u5be6\u6aa2\u6838\u3001Wang [9]\u8207\u52a0\u62ff\u5927\u7dad\u591a\u5229\u4e9e\u5927\u5b78\uff0c\u900f\u904e PolitiFact \u4e8b\u5be6\u6aa2\u6838\u7db2\u7ad9\u6240\u8490\u96c6\u7684\u8cc7\u6599\u5eab\u3001The FEVER \u773e\u8490\u96c6\u8cc7\u6599\u5f8c\u95e2\u8b20\uff0c\u4f46\u76ee\u524d\u65bc 2018/10/24 \u505c\u6b62\u670d\u52d9\uff0c\u4f46\u8cc7\u6599\u4f9d\u820a\u6703\u4fdd\u7559\u5728\u7db2\u7ad9\u4e0a\u9762\u3002\u56e0 \u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d\u6c92\u6709\u771f\u65b0\u805e\u7684\u6a19\u8a18\u8cc7\u6599\uff0c\u672c\u7814\u7a76\u8490\u96c6\u4e86\u771f\u65b0\u805e\u8cc7\u6599\uff0c\u4e3b\u8981\u4f86\u6e90\u662f\u860b\u679c\u65e5 \u5831\uff0c\u518d\u900f\u904e\u76f8\u4f3c\u5ea6\u5206\u6790\u5c07\u4ee5\u88ab\u8209\u5831\u5047\u65b0\u805e\u7684\u8cc7\u6599\u7d66\u5254\u9664\u3002\u81f3\u622a\u7a3f\u70ba\u6b62\u771f\u65b0\u805e\u7684\u8cc7\u6599\u4e00\u5171 \u8490\u96c6\u4e86 11,015 \u7b46\uff0c\u5047\u65b0\u805e\u7684\u8cc7\u6599\u4e00\u5171\u6709 5,795 \u7b46\uff0c\u7d93\u8cc7\u6599\u6574\u7406\u5f8c\uff0c\u4f7f\u7528\u4e86 2,014 \u7b46\u3002 (\u4e8c) \u7814\u7a76\u6d41\u7a0b \u4fbf\u52a0\u5165 BLSTM\u3001LSTM \u8207 GRU \u6a21\u578b\uff0c\u5148\u5efa\u7acb\u51fa\u7dda\u6027\u5806\u758a\u6a21\u578b\uff0c\u7136\u5f8c\u518d\u52a0\u5165\u5e73\u5766\u5c64\u3001\u96b1 \u85cf\u5c64\u3001\u8f38\u51fa\u5c64\uff0c\u518d\u4ee5\u4e94\u7a2e\u4e0d\u540c\u5927\u5c0f\u7684 Dropout\u3001\u4e09\u7a2e\u4e0d\u540c\u7684\u640d\u5931\u51fd\u6578\u3001\u4e09\u7a2e\u4e0d\u540c\u7684\u6fc0\u6d3b \u51fd\u6578\u4ee5\u53ca\u4e09\u7a2e\u4e0d\u540c\u7684\u512a\u5316\u5668\uff0c\u7e3d\u5171 135 \u500b\u6392\u5217\u7d44\u5408\uff0c\u627e\u51fa\u6700\u4f73\u7d44\u5408\u7684\u6a21\u578b\u3002\u5728\u8cc7\u6599\u7d93\u904e \u8cc7\u6599\u96c6 \u8a13\u7df4\u7b46\u6578 (\u5047\uff1a\u771f) \u6e2c\u8a66\u7b46\u6578 \u8cc7\u6599\u96c6 \u5047\u65b0\u805e \u6bd4\u7387 \u8a13\u7df4\u7b46\u6578 (\u5047\uff1a\u771f) \u6e2c\u8a66\u7b46\u6578 (\u4e09) \u7814\u7a76\u9650\u5236 (\u5047\uff1a\u771f) (\u5047\uff1a\u771f) \u65b0\u805e\u5c0f\u5e6b\u624b 1611:1611 \u65b0\u805e\u5c0f 25% 1600:4800 400:1200 \u5716\u4e8c\u3001\u4e0d\u540c\u5047\u65b0\u805e\u6bd4\u4f8b\u6548\u7387\u7d71\u8a08\u5716 \u672c\u7814\u7a76\u8a8d\u70ba\u76ee\u524d\u5047\u65b0\u805e\u5075\u6e2c\u6700\u5927\u9650\u5236\u9084\u662f\u5047\u65b0\u805e\u7684\u8cc7\u6599\u4e0d\u8db3\uff0c\u7121\u6cd5\u78ba\u8a8d\u6b64\u8cc7\u6599\u96c6\u662f 403:403 \u771f\u7684\u5047\u7684 4886:4886 \u5e6b\u624b 5% 252:4800 63:1200 1221:1221 5%* 421:8000 \u4e94\u3001\u7d50\u8ad6 \u5426\u80fd\u4ee3\u8868\u6574\u500b\u771f\u5be6\u4e16\u754c\u7684\u72c0\u6cc1\u3002\u8cc7\u6599\u96c6\u7576\u4e2d\u771f\u65b0\u805e\u662f\u5426\u9084\u6709\u5047\u65b0\u805e\u5728\u5176\u4e2d\u6216\u8005\u5047\u65b0\u805e 105:2000 \u4e00\u9023\u4e32\u8fed\u4ee3\u4e4b\u5f8c\uff0c\u4f7f Accuracy\u3001Precision\u3001Recall\u3001F-score \u6aa2\u6e2c\u6a21\u578b\u3002 (\u4e8c) \u6df1\u5ea6\u5b78\u7fd2\u7d50\u679c \u672c\u7814\u7a76\u5c07\u5be6\u9a57\u4e2d\u5f97\u51fa\u7684\u4e09\u7a2e\u65b9\u6cd5\u6700\u4f73\u914d\u7f6e\u5982\u8868\u4e8c\u6240\u793a\u3002\u6700\u5f8c\u4ee5 Accuracy\u3001Precision\u3001 \u771f\u7684\u5047 \u7684 25% 1600:4800 400:1200 5% 252:4800 \u7576\u4e2d\u6709\u771f\u65b0\u805e\u88ab\u8aa4\u5831\u70ba\u5047\u65b0\u805e\uff0c\u4ee5\u53ca\u5047\u65b0\u805e\u6bd4\u4f8b\u5728\u771f\u5be6\u8cc7\u6599\u96c6\u7576\u4e2d\u6bd4\u4f8b\u70ba\u4f55\u3002\u672c\u7814\u7a76 \u672c\u7814\u7a76\u5be6\u9a57\u5169\u500b\u4e2d\u6587\u8cc7\u6599\u96c6\uff0c\u767c\u73fe\u6df1\u5ea6\u5b78\u7fd2\u78ba\u5be6\u53ef\u4ee5\u7528\u65bc\u5075\u6e2c\u5047\u65b0\u805e\u3002\u672c\u7814\u7a76\u4e5f\u6a21 63:1200 \u64ec\u4e86 3 \u7a2e\u5047\u65b0\u805e\u6bd4\u4f8b\u7684\u5be6\u9a57\uff0c50%\u300125%\u4ee5\u53ca 50%\u3002\u5728\u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d\u8cc7\u6599\u96c6\u7576\u4e2d\uff0c \u8a8d\u70ba\u6df1\u5ea6\u5b78\u7fd2\u5df2\u7d93\u76f8\u7576\u6210\u719f\uff0c\u53ef\u4ee5\u904b\u7528\u5728\u5047\u65b0\u805e\u5075\u6e2c\u7576\u4e2d\uff0c\u4f46\u6700\u5927\u9650\u5236\u9084\u662f\u8a9e\u6599\u5eab\u4e0d</td></tr><tr><td>\u7684\u5047\u65b0\u805e\u8cc7\u6599\uff0c\u63a2\u7a76\u4ee5\u6df1\u5ea6\u5b78\u7fd2\u65b9\u5f0f\u5075\u6e2c\u5047\u65b0\u805e\u7684\u53ef\u80fd\u6027\u3002 \u4e8c\u3001\u6587\u737b\u63a2\u8a0e \u6df1\u5ea6\u795e\u7d93\u7db2\u8def(Deep Neural Network)\uff0c\u662f\u591a\u5c64\u7684\u795e\u7d93\u7db2\u8def\uff0c\u53ef\u900f\u904e\u96fb\u8166\u627e\u51fa\u7279\u5fb5\u503c\uff0c 1.0 P\u00e9rez-Rosas, et al. [10]\u81ea\u884c\u8490\u96c6\u5047\u65b0\u805e\uff0c\u4e26\u8acb\u5916\u5305\u516c\u53f8\u6539\u5beb\u3002\u8fd1\u5e74\u4f86\u53f0\u7063\u958b\u59cb\u6709\u4e8b\u5be6\u6aa2 \u6838\u7db2\u7ad9\u4ee5\u53ca\u7fa4\u773e\u95e2\u8b20\u7db2\u7ad9\u7684\u51fa\u73fe\uff0c\u4f46\u4ecd\u5c11\u6709\u76f8\u95dc\u7684\u7814\u7a76\u6210\u679c\u88ab\u63d0\u51fa\u3002 (\u4e09) \u61c9\u7528\u6df1\u5ea6\u5b78\u7fd2\u5047\u65b0\u805e\u5075\u6e2c\u7814\u7a76 \u672c\u7814\u7a76\u5148\u5c07\u5169\u500b\u8cc7\u6599\u96c6\u8490\u96c6\u5f8c\uff0c\u9032\u884c\u6587\u5b57\u7684\u9810\u8655\u7406\u3002\u5c07\u4e2d\u6587\u5b57\u5148\u9032\u884c\u65b7\u8a5e\u5f8c\uff0c\u5728\u5c07\u4e2d \u6587\u5b57\u9032\u884c Tokenization \u4ee5\u53ca\u6587\u5b57\u7684\u8a5e\u6578\u8a08\u7b97\uff0c\u4ee5\u4fbf\u6df1\u5ea6\u5b78\u7fd2\u8655\u7406\u3002\u7136\u5f8c\u5c07\u9810\u8655\u7406\u8cc7\u6599\u4ee5 80-20 \u6cd5\u5247\u5206\u6210\u8a13\u7df4\u96c6\u8207\u6e2c\u8a66\u96c6\u3002\u672c\u7814\u7a76\u5be6\u9a57\u4e09\u7a2e\u6df1\u5ea6\u5b78\u7fd2\u7684\u65b9\u6cd5\uff0cGated Recurrent Unit (GRU)\u3001Long Short Term Memory (LSTM) \u4ee5\u53ca Bidirectional Long Short Term Memory (BLSTM)\u3002\u7136\u5f8c\u4f7f\u7528 Accuracy\u3001Precision\u3001Recall \u53ca F-Score \u5c0d\u6a21\u578b\u9032\u884c\u8a55\u4f30\u3002\u6700\u5f8c\u5c07 \u6b64\u4e0a\u8ff0\u7d93\u904e\u7e6a\u88fd\u6210\u4ee5\u4e0b\u6d41\u7a0b\u5716\uff1a Recall \u4ee5\u53ca F-Score \u5c0d\u5404\u7a2e\u6a21\u578b\u505a\u8a55\u4f30\u5982\u8868\u4e09\u6240\u793a\uff1a \u8868\u4e09\u300150%\u5047\u65b0\u805e\u6bd4\u4f8b\u6700\u4f73\u914d\u7f6e \u8cc7\u6599\u96c6 \u65b9\u6cd5 \u6fc0\u6d3b \u51fd\u6578 \u640d\u5931 \u51fd\u6578 \u512a\u5316\u5668 Dropout \u65b0\u805e \u5c0f\u5e6b\u624b GRU sigmoid MSLE Adam 0.6 LSTM sigmoid binary_crossentropy AdaMax 0.6 BLSTM relu MSE RMSprop GRU sigmoid binary_crossentropy Adam 0.4 5% LSTM tanh binary_crossentropy AdaMax 0.6 \u65b0\u805e\u6bd4\u4f8b\u5b8c\u5168\u7121\u6cd5\u5075\u6e2c\u3002\u672c\u7814\u7a76\u8a8d\u70ba\u300c\u771f\u7684\u5047\u7684\u300d\u6587\u7ae0\u9577\u5ea6\u90fd\u77ed\u65bc\u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d \uff0c\u8cc7 \u771f\u7684\u5047\u7684 25% BLSTM relu binary_crossentropy AdaMax 0.4 0.5 \u8868\u516d\u3001\u5047\u65b0\u805e\u6bd4\u4f8b 25%\u8207 5%\u6700\u4f73\u914d\u7f6e \u8cc7\u6599\u96c6 \u5047\u65b0\u805e\u6bd4 \u4f8b \u6700\u4f73\u65b9\u6cd5 \u6fc0\u6d3b \u51fd\u6578 \u640d\u5931 \u51fd\u6578 \u512a\u5316\u5668 \u8db3\u7684\u554f\u984c\uff0c\u672c\u7814\u7a76\u8a8d\u70ba\u5047\u65b0\u805e\u5075\u6e2c\u9700\u8981\u6709\u66f4\u9032\u4e00\u6b65\u7684\u5206\u985e\uff0c\u4ee5\u53ca\u66f4\u591a\u6a19\u7c64(Label)\u4ee5\u53ca \u66f4\u591a\u7279\u5fb5(Feature)\u53bb\u6a19\u8a18\u3002\u672c\u7814\u7a76\u8a8d\u70ba\u53ef\u4ee5\u518d\u9032\u4e00\u6b65\u5617\u8a66\u8490\u96c6\u66f4\u591a\u771f\u65b0\u805e\uff0c\u5617\u8a66\u66f4\u4f4e 50%\u7684 F-\u672c\u7814\u7a76\u53c8\u5be6\u9a57\u4e86\u5c07 5%\u7684\u8a13\u7df4\u8207\u6e2c\u8a66\u6bd4\u6578\u62c9\u9ad8\uff0c5%\u7684 F-Score \u5f9e\u539f\u672c\u7684 0.563 \u589e\u9577\u5230 Dropout \u65b0\u805e 25% BLSTM sigmoid binary_crossentropy Adam \u7684\u5047\u65b0\u805e\u6bd4\u4f8b\uff0c\u53ef\u4ee5\u8b93\u6a21\u578b\u66f4\u80fd\u904b\u7528\u5728\u5be6\u52d9\u4e0a\u3002\u76f8\u4fe1\u6709\u5929\u5047\u65b0\u805e\u5075\u6e2c\u6709\u5929\u53ef\u4ee5\u904b\u7528\u5728 0.678 \u767c\u73fe\u8a13\u7df4\u8207\u8cc7\u6599\u6e2c\u8a66\u7b46\u6578\u589e\u52a0\uff0c\u53ef\u4ee5\u8b93\u6df1\u5ea6\u5b78\u7fd2\u7684\u6548\u7387\u66f4\u597d\u3002\u5728\u300c\u771f\u7684\u5047\u7684\u300d\u8cc7 0.3 5% BLSTM tanh binary_crossentropy Adam 0.4 \u5404\u7a2e\u88dd\u7f6e\u4e0a\uff0c\u907f\u514d\u5c0e\u81f4\u8a0a\u606f\u63a5\u53d7\u8005\u88ab\u8aa4\u5c0e\uff0c\u9020\u6210\u793e\u6703\u7684\u6050\u614c\u3002 \u6599\u96c6\u7576\u4e2d 50%\u7684 F-Score \u5728\u4f86\u5230 0.728\uff0c25%\u7684 F-Score \u4f86\u5230 0.701\uff0c\u800c 5%\u7684 F-Score \u53ea \u5c0f\u5e6b\u624b 5%* GRU Tanh MSLE Adam 0.6 \u6709 0.396\u3002\u76f8\u8f03\u65bc\u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d\u8cc7\u6599\u96c6\uff0c \u300c\u771f\u7684\u5047\u7684\u300d\u8cc7\u6599\u96c6\u6548\u7387\u8f03\u5dee\uff0c\u751a\u81f3\u5728 5%\u5047 \u53c3\u8003\u6587\u737b</td></tr><tr><td>\u904e\u53bb\u5047\u65b0\u805e\u5075\u6e2c\u4f7f\u7528\u7684\u7279\u5fb5\u5927\u591a\u662f\u4f7f\u7528\u6587\u5b57\u7279\u5fb5\uff0c\u4e26\u4f7f\u7528\u6a5f\u5668\u5b78\u7fd2\u7684\u65b9\u5f0f\u9032\u884c\u5075 \u771f\u7684\u5047\u7684 LSTM sigmoid binary_crossentropy RMSprop 0.4 \u8868\u4e03\u3001\u5047\u65b0\u805e\u6bd4\u4f8b 25%\u8207 5%\u5be6\u9a57\u7d50\u679c \u6599\u96c6\u7b46\u6578\u4e0d\u8db3\u5c0e\u81f4\u6df1\u5ea6\u5b78\u7fd2\u7684\u6548\u7387\u4e0d\u4f73\u3002 \u7d93\u7531\u6df1\u5165\u5b78\u7fd2\u5f8c\uff0c\u8b93\u9810\u6e2c\u7d50\u679c\u66f4\u6709\u6548\uff0c\u76ee\u524d\u4e3b\u8981\u7684\u795e\u7d93\u7db2\u8def\u6a21\u578b\u6709 20 \u5e7e\u7a2e\uff0c\u5176\u4e2d\uff0c\u905e \u8ff4\u795e\u7d93\u7db2\u8def(Recurrent Neural Network, RNN)\u53ef\u7528\u65bc\u89e3\u6c7a\u5e36\u6709\u9806\u5e8f\u6027\u7684\u554f\u984c\uff0c\u4f8b\u5982: \u81ea\u7136 \u6e2c\uff0c\u4f8b\u5982\uff1a\u4f7f\u7528 Support Vector Machines(SVM)\u3001Unsupervised\u3001Naive Bayes\u3002\u800c\u56e0\u6df1 BLSTM relu MSLE AdaMax 0.5 \u8cc7\u6599\u96c6 \u6700\u4f73 \u65b9\u6cd5 \u5047\u65b0\u805e \u6bd4\u4f8b Loss Accuracy Precision Recall F-Score \u672c\u7814\u7a76\u5be6\u9a57\u4e86\u4e09\u7a2e\u65b9\u6cd5\uff0cBLSTM\u3001LSTM \u4ee5\u53ca GRU\uff0c\u767c\u73fe\u5927\u591a\u6578\u7684\u60c5\u6cc1\u4e0b\uff0c</td></tr><tr><td>\u8a9e\u8a00\u8655\u7406\u3001\u8a9e\u97f3\u8fa8\u8b58\u3001\u624b\u5beb\u8fa8\u8b58\u7b49\u7b49[4]\u3002 Hochreiter and Schmidhuber [5]\u63d0\u51fa\u9577\u77ed\u671f\u8a18\u61b6(Long Short-Term Memory, LSTM)\uff0c\u662f \u70ba\u4e86\u89e3\u6c7a RNN \u6240\u7522\u751f\u68af\u5ea6\u6d88\u5931\u53ca\u68af\u5ea6\u7206\u70b8\u800c\u7522\u751f\u3002LSTM \u8207 RNN \u76f8\u6bd4\u591a\u4e86 3 \u500b\u63a7\u5236 \u5668\uff0c\u5206\u5225\u70ba\u8f38\u5165\u9598\u9580(Input Gate)\u3001\u8f38\u51fa\u9598\u9580(Output Gate)\u3001\u5fd8\u8a18\u9598\u9580(Forget Gate)\uff0c\u7576\u6709 \u5ea6\u5b78\u7fd2\u767c\u5c55\u512a\u7570\uff0c\u4e5f\u958b\u59cb\u6709\u7814\u7a76\u5c07\u6df1\u5ea6\u5b78\u7fd2\u61c9\u7528\u5728\u5047\u65b0\u805e\u5075\u6e2c\u4e2d\u3002\u4e0d\u904e\uff0c\u5728\u53f0\u7063\uff0c\u9019 \u65b9\u9762\u7684\u7814\u7a76\u6210\u679c\u4ecd\u8f03\u5c11\u88ab\u63d0\u51fa\u3002\u672c\u7814\u7a76\u6574\u7406\u51fa\u904e\u53bb\u7814\u7a76\u6240\u4f7f\u7528\u6df1\u5ea6\u5b78\u7fd2\u7684\u5047\u65b0\u805e\u5075\u6e2c \u7814\u7a76\uff1a GRU 0.104 0.677 0.635 0.836 0.722 \u4e09\u3001\u7814\u7a76\u65b9\u6cd5 \u8868\u56db\u300150%\u5047\u65b0\u805e\u6bd4\u4f8b\u6a21\u578b\u8a55\u4f30 \u8cc7\u6599\u96c6 \u65b9\u6cd5 Loss Accuracy Precision Recall F-Score GRU 0.053 0.881 0.885 0.876 0.88 BLSTM 25% 0.053 0.922 0.879 0.797 BLSTM \u6548\u679c\u90fd\u662f\u6700\u4f73\u3002\u9664\u4e86\u5728\u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d5%\u5047\u65b0\u805e\u6bd4\u4f8b\u7684 GRU \u6548\u679c\u6700\u597d\u4ee5\u53ca 0.836 \u65b0\u805e BLSTM 5% 0.045 0.964 0.725 0.46 0.563 \u300c\u771f\u7684\u5047\u7684\u300d5%\u5047\u65b0\u805e\u6bd4\u4f8b\u7121\u6cd5\u5075\u6e2c\u4e4b\u5916\uff0c\u5c1a\u9918 5 \u7a2e\u7686\u662f BLSTM \u65b9\u6cd5\u6548\u679c\u6700\u4f73\u3002 \u5c0f\u5e6b\u624b GRU 5%* 0.091 0.972 0.795 0.59 0.678 \u65b0\u805e \u5c0f\u5e6b\u624b LSTM 0.045 0.888 0.881 0.898 0.889 BLSTM 0.091 0.898 0.898 0.898 0.898 BLSTM 25% 0.134 0.845 0.728 0.701 \u800c\u9664\u6b64\u4e4b\u5916\uff0c\u672c\u7814\u7a76\u53c8\u70ba\u4e86\u9032\u4e00\u6b65\u5be6\u9a57\u5c07\u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d\u505a\u4ea4\u53c9\u8cc7\u6599\u96c6\u6e2c\u8a66\uff0c\u767c\u73fe 0.677 \u771f\u7684\u5047\u7684 LSTM 5% 0.616 0.948 0.333 0.396 0.488 \u5982\u8981\u5c07\u6a21\u578b\u5be6\u52d9\u6240\u9700\u7684\u767c\u73fe\u3002</td></tr><tr><td>\u4e86\u63a7\u5236\u5668\u9598\u9580\u7684\u6a5f\u5236 LSTM \u5c31\u80fd\u5920\u5c07\u8a18\u61b6\u9577\u671f\u8a18\u4f4f\u3002 \u96d9\u5411\u9577\u77ed\u671f\u8a18\u61b6(Bidirectional-Long Short-Term Memory, BLSTM)\u5728 2005 \u5e74\u7531 Graves and Schmidhuber [6]\u63d0\u51fa\u4f86\uff0c\u4ed6\u662f\u7531\u524d\u5411\u9577\u77ed\u671f\u8a18\u61b6\u8207\u5f8c\u5411\u9577\u77ed\u671f\u8a18\u61b6\u7d50\u5408\u800c\u6210\uff0c\u662f\u70ba \u5716\u4e00\u3001\u7814\u7a76\u6d41\u7a0b\u5716 \u771f\u7684\u5047\u7684 LSTM 0.108 0.685 0.641 0.844 0.728 \u5728 50%\u300125%\u8207 5%\u6a21\u578b\u6e2c\u8a66\u4e2d\u4e5f\u767c\u73fe\u4e00\u6a23\u7684\u7d50\u679c\uff0cRecall \u4e5f\u504f\u9ad8\u5728 0.93 \u4ee5\u53ca 0.87\uff0c \u6700\u5f8c\u4ee5\u5716\u4e8c\u53ef\u4ee5\u767c\u73fe\uff0c\u96d6\u7136\u964d\u4f4e\u5047\u65b0\u805e\u6bd4\u4f8b\uff0c\u53ef\u4ee5\u8b93\u6a23\u672c\u53ef\u4ee5\u66f4\u8da8\u8fd1\u65bc\u771f\u5be6\u60c5\u6cc1\uff0c\u4f46 (\u4e00) \u8cc7\u6599\u96c6 \u672c\u7814\u7a76\u8490\u96c6\u8cc7\u6599\u96c6\u662f\u53f0\u7063\u7d44\u7e54\u300cg0v.tw \u53f0\u7063\u96f6\u6642\u653f\u5e9c\u300d \uff0c\u6240\u5275\u7acb\u7684\u5169\u500b\u5c08\u6848\uff1a BLSTM 0.218 0.683 0.638 0.846 0.728 \u9019\u4e5f\u767c\u73fe 50%\u300125%\u7684\u6a21\u578b\u4e5f\u80fd\u5920\u6e05\u695a\u4e86\u89e3\u5047\u65b0\u805e\u8108\u7d61\uff0c\u4e26\u4e14\u6210\u529f\u9810\u6e2c\u6b63\u78ba\uff0c\u4f46\u4e00\u6a23 \u53ef\u4ee5\u660e\u986f\u767c\u73fe Precision\u3001Recall \u4ee5\u53ca F-Score \u4e0b\u964d\u3002\u751a\u81f3\u5728\u300c\u771f\u7684\u5047\u7684\u300d\u5728 5%\u5047\u65b0\u805e\u6bd4 (\u4e09) \u6df1\u5ea6\u5b78\u7fd2\u67b6\u69cb (\u4e09) \u4e0d\u540c\u5047\u65b0\u6bd4\u4f8b\u7d50\u679c \u4f8b\u56db\u500b\u6307\u6a19\u7686\u672a\u9054\u5230\u57fa\u6e96\u7dda\u3002\u800c\u5728\u300c\u65b0\u805e\u5c0f\u5e6b\u624b\u300d\u7684\u5047\u65b0\u805e 5%\u5be6\u9a57\u7576\u4e2d\u53ef\u4ee5\u767c\u73fe\uff0c\u76f8 \u5c0d\u591a\u7a2e\u771f\u65b0\u805e\u4f86\u6e90\u8108\u7d61\u5f88\u4e0d\u719f\u6089\uff0c\u5e38\u5e38\u771f\u65b0\u805e\u932f\u7576\u6210\u5047\u65b0\u805e\u3002\u9019\u8aaa\u660e\u4e86\u5728\u8a13\u7df4\u7684\u8cc7\u6599</td></tr></table>" |
| } |
| } |
| } |
| } |
