{ "paper_id": "O15-1015", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:10:04.701606Z" }, "title": "A Tool for Web NER Model Generation Using Search Snippets of Known Entities", "authors": [ { "first": "Ya-Yun", "middle": [], "last": "\u9ec3\u96c5\u7b60", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Central University", "location": {} }, "email": "" }, { "first": "", "middle": [], "last": "Huang", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Central University", "location": {} }, "email": "" }, { "first": "Chia-Hui", "middle": [], "last": "\u5f35\u5609\u60e0", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Central University", "location": {} }, "email": "" }, { "first": "", "middle": [], "last": "Chang", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Central University", "location": {} }, "email": "" }, { "first": "Chia-Hui", "middle": [], "last": "\u5468\u5efa\u9f8d", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Central University", "location": {} }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Named entity recognition (NER) is of vital importance in information extraction and natural language processing. Current NER models are trained mainly on journalistic documents such as news articles. Since they have not been trained to deal with informal documents, the performance drops on Web documents, which may lack sentence structure and contain colloquial expression.", "pdf_parse": { "paper_id": "O15-1015", "_pdf_hash": "", "abstract": [ { "text": "Named entity recognition (NER) is of vital importance in information extraction and natural language processing. Current NER models are trained mainly on journalistic documents such as news articles. Since they have not been trained to deal with informal documents, the performance drops on Web documents, which may lack sentence structure and contain colloquial expression.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "documents. When users want to recognize named entity from Web documents, they certainly have to retrain the new model. Retraining a new model is labor intensive and time consuming. The preparatory work includes preparing a large set of training data, labeling named entity, selecting an appropriate segmentation, symbols unification, normalization, designing feature, preparing dictionary, and so on. Besides, users need to repeat the previous work for different languages or different recognition types. In this research, we propose a NER model generation tool for effective Web entity extraction. We propose a semi-supervised learning approach for NER model training via automatic labeling and tri-training, which makes use of unlabeled data and structured resources containing known named entities. Experiments confirmed that the use of this tool can be applied in different languages for various types of named entities. In the task of Chinese organization name extraction, the generated model can achieve 86.1% F1 score on the 38,692 sentences with 16,241 distinct names, while the performance for Japanese organization name, English organization name, Chinese location name extraction, Chinese address recognition and English address recognition can be reached 80.3%, 83.2%, 84.5%, 97.2% and 94.8% F1-measure, respectively. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "http://nlp.stanford.edu/software/CRF-NER.shtml", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "https://www.iyp.com.tw/ 3 http://itp.ne.jp/?rf=1", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "http://www.yelp.com/ 5 http://data.gov.tw/?q=node/7063", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Nymble: a High-Performance Learning Name-finder", "authors": [ { "first": "D.-M", "middle": [], "last": "Bikel", "suffix": "" }, { "first": "S", "middle": [], "last": "Miller", "suffix": "" }, { "first": "R", "middle": [], "last": "Schwartz", "suffix": "" }, { "first": "R", "middle": [], "last": "Weischedel", "suffix": "" } ], "year": 1997, "venue": "", "volume": "", "issue": "", "pages": "194--201", "other_ids": {}, "num": null, "urls": [], "raw_text": "D.-M. Bikel, S. Miller, R. Schwartz and R. Weischedel, \"Nymble: a High-Performance Learning Name-finder\", Applied natural language processing, pp. 194-201, 1997.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Semi-supervised Sequence Labeling for Named Entity Extraction based on Tri-Training: Case Study on Chinese Person Name Extraction", "authors": [ { "first": "C.-L", "middle": [], "last": "Chou", "suffix": "" }, { "first": "C.-H", "middle": [], "last": "Chang", "suffix": "" }, { "first": "S.-Y", "middle": [], "last": "Wu", "suffix": "" } ], "year": 2014, "venue": "Semantic Web and Information Extraction", "volume": "", "issue": "", "pages": "244--255", "other_ids": {}, "num": null, "urls": [], "raw_text": "C.-L. Chou, C.-H. Chang, S.-Y. Wu, \" Semi-supervised Sequence Labeling for Named Entity Extraction based on Tri-Training: Case Study on Chinese Person Name Extraction,\" Semantic Web and Information Extraction, pp. 244-255, 2014.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "CRF++: Yet Another CRF toolkit", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "CRF++: Yet Another CRF toolkit, http://crfpp.googlecode.com/svn/trunk/doc/index.html 9-1541", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data", "authors": [ { "first": "J", "middle": [], "last": "Lafferty", "suffix": "" }, { "first": "A", "middle": [], "last": "Mccallum", "suffix": "" }, { "first": "F", "middle": [ "C N" ], "last": "Pereira", "suffix": "" } ], "year": 2001, "venue": "ICML Proceedings of the Eighteenth International Conference on Machine Learning", "volume": "", "issue": "", "pages": "282--289", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. Lafferty, A. McCallum and F.C.N. Pereira, \"Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data,\" ICML Proceedings of the Eighteenth International Conference on Machine Learning, pp. 282-289, 2001.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Automatic Recognition of Chinese Personal Name Using Conditional Random Fields and Knowledge Base", "authors": [ { "first": "C", "middle": [], "last": "Gu", "suffix": "" }, { "first": "X.-P", "middle": [], "last": "Tian", "suffix": "" }, { "first": "J.-D", "middle": [], "last": "Yu", "suffix": "" } ], "year": 2015, "venue": "Mathematical Problems in Engineering", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "C. Gu, X.-P. Tian, and J.-D Yu, \"Automatic Recognition of Chinese Personal Name Using Conditional Random Fields and Knowledge Base,\" Mathematical Problems in Engineering, 2015.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Store Name Extraction and Name-Address Matching on the Web", "authors": [ { "first": "Y. -Y", "middle": [], "last": "Lin", "suffix": "" }, { "first": "C.-H", "middle": [], "last": "Chang", "suffix": "" } ], "year": 2014, "venue": "Proceedings of the 26th Conference on Computational Linguistics and Speech Processing", "volume": "", "issue": "", "pages": "91--93", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. -Y. Lin, C.-H. Chang, \"Store Name Extraction and Name-Address Matching on the Web,\" Proceedings of the 26th Conference on Computational Linguistics and Speech Processing, pp. 91-93, 2014.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Chinese Organization Name Recognition Based on Multiple Features", "authors": [ { "first": "Y", "middle": [], "last": "Ling", "suffix": "" }, { "first": "J", "middle": [], "last": "Yang", "suffix": "" }, { "first": "L", "middle": [], "last": "He", "suffix": "" } ], "year": 2012, "venue": "Pacific Asia conference on Intelligence and Security Informatics", "volume": "", "issue": "", "pages": "136--144", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. Ling, J. Yang and L. He, \"Chinese Organization Name Recognition Based on Multiple Features,\" Pacific Asia conference on Intelligence and Security Informatics, pp. 136-144, 2012.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Semi-supervised sequence modeling with syntactic topic models", "authors": [ { "first": "W", "middle": [], "last": "Li", "suffix": "" }, { "first": "A", "middle": [], "last": "Mccallum", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "W. Li, A. McCallum, \"Semi-supervised sequence modeling with syntactic topic models,\"", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "AAAI'05 Proceedings of the 20th national conference on Artificial intelligence", "authors": [], "year": 2005, "venue": "", "volume": "2", "issue": "", "pages": "813--818", "other_ids": {}, "num": null, "urls": [], "raw_text": "AAAI'05 Proceedings of the 20th national conference on Artificial intelligence -Volume 2, pp. 813-818, 2005.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Early Results for Named Entity Recognition with Conditional Random Fields, Feature Induction and Web-Enhanced Lexicons", "authors": [ { "first": "A", "middle": [], "last": "Mccallum", "suffix": "" }, { "first": "W", "middle": [], "last": "Li", "suffix": "" } ], "year": 2003, "venue": "Proceedings of the Seventh Conference on Natural Language Learning HLT-NAACL 2003", "volume": "4", "issue": "", "pages": "188--191", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. McCallum, W. Li, \"Early Results for Named Entity Recognition with Conditional Random Fields, Feature Induction and Web-Enhanced Lexicons,\" Proceedings of the Seventh Conference on Natural Language Learning HLT-NAACL 2003 -Volume 4 (CONLL'03), pp. 188-191, 2003.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Semi-joint labeling for Chinese named entity recognition", "authors": [ { "first": "C.-W", "middle": [], "last": "Wu", "suffix": "" }, { "first": "R", "middle": [ "T" ], "last": "", "suffix": "" }, { "first": ".-H", "middle": [], "last": "Tsai", "suffix": "" }, { "first": "W.-L", "middle": [], "last": "Hsu", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 4th Asia information retrieval conference", "volume": "", "issue": "", "pages": "107--116", "other_ids": {}, "num": null, "urls": [], "raw_text": "C.-W. Wu, R. T.-H. Tsai and W.-L. Hsu, \"Semi-joint labeling for Chinese named entity recognition,\" Proceedings of the 4th Asia information retrieval conference, pp. 107-116, 2008.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "A Method of Chinese Organization Named Entities Recognition Based on Statistical Word Frequency, Part of Speech and Length", "authors": [ { "first": "X", "middle": [], "last": "Yao", "suffix": "" } ], "year": 2011, "venue": "Broadband Network and Multimedia Technology (IC-BNMT)", "volume": "", "issue": "", "pages": "637--641", "other_ids": {}, "num": null, "urls": [], "raw_text": "X. Yao, \"A Method of Chinese Organization Named Entities Recognition Based on Statistical Word Frequency, Part of Speech and Length,\" Broadband Network and Multimedia Technology (IC-BNMT), pp. 637-641, 2011.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Tri-Training: Exploiting Unlabeled Data Using Three Classifiers", "authors": [ { "first": "Z.-H", "middle": [], "last": "Zhou", "suffix": "" }, { "first": "M", "middle": [], "last": "Li", "suffix": "" } ], "year": 2005, "venue": "IEEE Transactions on Knowledge and Data Engineering archive", "volume": "17", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Z.-H. Zhou, M. Li, \"Tri-Training: Exploiting Unlabeled Data Using Three Classifiers\", IEEE Transactions on Knowledge and Data Engineering archive, Volume 17 Issue 11, November 2005, Page 152.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Automatic Recognition of Chinese Organization Name Based on Conditional Random Fields", "authors": [ { "first": "S", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "S", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "X", "middle": [], "last": "Wang", "suffix": "" } ], "year": 2007, "venue": "Natural Language Processing and Knowledge Engineering", "volume": "", "issue": "", "pages": "229--233", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. Zhang, S. Zhang and X. Wang, \"Automatic Recognition of Chinese Organization Name Based on Conditional Random Fields,\" Natural Language Processing and Knowledge Engineering, pp. 229-233, 2007.", "links": null } }, "ref_entries": { "TABREF0": { "num": null, "content": "
\u503c\u6b63\u898f\u5316\uff0c\u4ee5\u53ca\u6e96\u5099\u5177\u6709\u9451\u5225\u5ea6\u7684\u7279\u5fb5\u503c\u3001\u6216\u8a2d\u8a08\u5df2\u77e5\u8fad\u5178\u7b49\u3002\u82e5\u8981\u8fa8\u8b58\u4e0d\u540c\u7684\u8a9e\u8a00\u6216 \u4e0d\u540c\u985e\u578b\u7684\u547d\u540d\u5be6\u9ad4\uff0c\u5c31\u8981\u91cd\u8907\u4ee5\u4e0a\u7684\u52d5\u4f5c\u4f86\u5b8c\u6210\u5de5\u4f5c\uff0c\u9020\u6210\u4e86\u4e0d\u5c11\u4eba\u529b\u8207\u6642\u9593\u7684\u6d6a\u8cbb\uff0c \u52d5\u901a\u8a0a+\u516c\u53f8\u300d \uff0c\u8003\u616e\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u7684\u51fa\u73fe\u983b\u7387\u3001\u8a5e\u6027\u8207\u9577\u5ea6\uff0c\u4e26\u914d\u5408\u81ea\u884c\u8a2d\u8a08\u7684\u7d71\u8a08\u65b9 \u6cd5\u3002\u5be6\u9a57\u4f7f\u7528\u4e86\u4eba\u6c11\u7db2\u7684\u8a9e\u6599\u9032\u884c\u8a13\u7df4\uff0c\u4ee5\u4eba\u6c11\u7db2\u3001\u65b0\u83ef\u7db2\u548c\u5317\u4eac\u90f5\u96fb\u5927\u5b78\u7db2\u7ad9\u9996\u9801\u7684 \u672c\u7cfb\u7d71\u63a1\u7528\u4e0d\u65b7\u8a5e\u7684\u4e2d\u6587\u5b57\u70ba\u57fa\u672c\u8655\u7406\u55ae\u5143 Token\uff0c\u907f\u514d\u6a23\u672c\u56e0\u70ba\u932f\u8aa4\u65b7\u8a5e\u7522\u751f\u547d\u540d\u5be6 \u9ad4\u88ab\u5206\u5272\u6210\u5169\u500b\u8a5e\u7684\u908a\u754c\u932f\u8aa4\u7684\u554f\u984c\uff0c\u6e1b\u5c11\u932f\u8aa4\u7d2f\u7a4d\u3002\u540c\u6642\u5c0d\u65bc\u6bcf\u4e00\u7b46\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u6211 \u7531\u65bc\u5728\u5224\u5b9a\u662f\u5426\u70ba\u6b63\u78ba\u7b54\u6848\u6642\uff0c\u6709\u6642\u6703\u6709\u96e3\u4ee5\u6e96\u78ba\u5b9a\u51fa\u908a\u754c\u7684\u53ef\u80fd\uff0c\u4f8b\u5982\uff1a \u300c7-ELEVEN (\u884c\u5929\u9580\u5e02) \u300d\u4e2d\uff0c \u300c (\u884c\u5929\u9580\u5e02) \u300d\u53ef\u4ee5\u8996\u70ba\u5305\u542b\u5728\u5546\u5bb6\u540d\u7a31\u4e4b\u4e2d\uff0c\u4f46\u82e5\u6c92\u6709\u6a19\u8a18\u51fa\u300c (\u884c \u96c6\u6bcf\u7b46\u67e5\u8a62\u6392\u540d\u524d10\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4ee5\u4eba\u5de5\u7684\u65b9\u5f0f\u6a19\u8a18\u6e2c\u8a66\u8cc7\u6599\u5171809\u500b\u53e5\u5b50\uff0c\u5171\u6a19 \u8a18\u4e0d\u91cd\u8907\u7684\u65e5\u6587\u5546\u5bb6\u7d44\u7e54\u540d\u7a31438\u500b\u3002 \u8868\u4e94\u3001\u4e0d\u540c\u8a9e\u8a00\u8207\u4e0d\u540c\u8fa8\u8b58\u4e3b\u984c\u8cc7\u6599\u96c6 \uf06c \u77ed\u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u6548\u80fd \u8868\u516b\u986f\u793a\u4ee5Google\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u8fa8\u8b58\u4e2d\u6587\u7d44\u7e54\u540d\u70ba\u4f8b\uff0c\u6bd4\u8f03\u7cfb\u7d71\u81ea\u52d5\u7522\u751f\u8a5e\u5eab\u3001\u4eba\u5de5\u7522 \u751f\u95dc\u9375\u8a5e\u5eab\u8207Stanford NER\u6548\u80fd\u7684\u5dee\u7570\u3002\u7e3d\u9ad4\u800c\u8a00\uff0c\u96d6\u7136\u81ea\u52d5\u7522\u751f\u95dc\u9375\u8a5e\u5eab\u6703\u5c0e\u81f4 4.5 \u6bd4\u8f03 ExactMatchLabeling \u53ca AlignmentLabeling \u6a19\u8a18\u6548\u679c \u4e94\u3001 \u7d50\u8ad6
\u4e00\u3001 \u7dd2\u8ad6 \u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u662f\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u7684\u4e00\u9805\u91cd\u8981\u57fa\u790e\u5de5\u4f5c\uff0c\u5176\u8fa8\u8b58\u6b63\u78ba\u7387\u5c0d\u5f8c\u7e8c\u7684\u8a9e\u610f\u5206\u6790 (Semantic Analysis) \u3001\u6a5f\u5668\u7ffb\u8b6f(Machine Translation)\u7b49\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u8b70\u984c\u5177\u91cd\u5927\u7684 \u5f71\u97ff\u3002\u5728\u5927\u91cf\u6587\u5b57\u8cc7\u6599\u4e2d\uff0c\u5e38\u6709\u4eba\u540d\u3001\u5730\u540d\u3001\u7d44\u7e54\u540d\u7b49\u6709\u610f\u7fa9\u7684\u5c08\u6709\u540d\u7a31\u51fa\u73fe\uff0c\u7136\u800c\u56e0 \u61c9\u793e\u6703\u9700\u8981\u53ca\u79d1\u6280\u767c\u5c55\uff0c\u9019\u4e9b\u4e0d\u65b7\u88ab\u5275\u9020\u7684\u8a5e\u5f59\uff0c\u96e3\u4ee5\u88ab\u55ae\u4e00\u8a5e\u5eab\u6240\u6536\u85cf\uff0c\u56e0\u6b64\u9700\u6709\u547d \u540d\u5be6\u9ad4\u8fa8\u8b58\u4ee5\u4fbf\u64f4\u5145\u8a5e\u5eab\u3002\u4e0d\u540c\u985e\u578b\u7684\u547d\u540d\u5be6\u9ad4\u51fa\u73fe\u65bc\u8a9e\u53e5\u4e2d\u7684\u4f4d\u7f6e\u3001\u898f\u5247\u6216\u8a5e\u6027\u7686\u4e0d \u76f8\u540c\uff0c\u56e0\u6b64\u9700\u8981\u7684\u7279\u5fb5\u503c\u4e5f\u90fd\u4e0d\u540c\u3002\u4ee5\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58\u70ba\u4f8b\uff0c\u76ee\u524d\u8a31\u591a\u95dc\u65bc\u7d44\u7e54\u540d\u7a31 \u8fa8\u8a8d\u7684\u7814\u7a76\uff0c\u4e3b\u8981\u662f\u5f9e\u65b0\u805e\u6216\u4e00\u4e9b\u8f03\u6b63\u5f0f\u7684\u6587\u7ae0\u4e2d\u8a13\u7df4\u7d44\u7e54\u540d\u7a31\u64f7\u53d6\u6a21\u578b[7] [11] [13]\uff0c \u4f46\u662f\u7db2\u8def\u4e0a\u5546\u5bb6\u7d44\u7e54\u540d\u7a31\u50be\u5411\u8f03\u4e0d\u6b63\u5f0f\u7684\u547d\u540d\u65b9\u5f0f\uff0c\u4f8b\u5982\uff1a\u5f7c\u5f97\u516c\u96de\u5730\u4e2d\u6d77\u9910\u5ef3\u3001\u9020\u7d19 \u9f8d\u624b\u5275\u9928\u7b49\uff0c\u800c\u65b0\u805e\u7b49\u8f03\u6b63\u5f0f\u7684\u9ad4\u88c1\u5247\u5bb9\u6613\u51fa\u73fe\u516c\u53f8\u884c\u865f\u8207\u6b63\u898f\u7684\u7d44\u7e54\u540d\u7a31\uff0c\u5982\uff1a\u4f0a\u7538 \u57fa\u91d1\u6703\u3001\u570b\u7acb\u4e2d\u592e\u5927\u5b78\u3001\u9ad8\u9435\u516c\u53f8\u7b49\uff0c\u4e14\u7db2\u8def\u4e0a\u767c\u8868\u65bc\u8ad6\u58c7\u6216\u793e\u7fa4\u5a92\u9ad4\u7684\u6587\u7ae0\u8a9e\u53e5\u7d50\u69cb \u8207\u7528\u5b57\u9063\u8a5e\u7686\u8207\u6b63\u5f0f\u6587\u7ae0\u4e0d\u540c\uff0c\u56e0\u6b64\u8fa8\u8b58\u6548\u679c\u4e0d\u4f73\u3002\u5982\u8868\u4e00\u4ee5\u53ca\u8868\u4e8c\u6240\u793a\uff0c\u6211\u5011\u5229\u7528 2,000 \u4e0a\u975e\u6b63\u5f0f\u6587\u7ae0\u7684\u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u6548\u679c\u6709\u9650\uff0c\u4e26\u5c0e\u81f4\u5f8c\u7e8c\u7684\u76f8\u95dc\u7814\u7a76\u6548\u80fd\u6709\u9650\u3002 \u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u53ef\u8996\u70ba\u5e8f\u5217\u6a19\u8a18(Sequence Labeling)\u7684\u554f\u984c\uff0c\u6545\u901a\u5e38\u4f7f\u7528 Conditional Random Field (CRF) \u4f86\u89e3\u6c7a\u6b64\u554f\u984c\uff0cCRF \u70ba\u4e00\u6a5f\uf961\u67b6\u69cb\u7684\u7121\u5411\u5716 (Undirected Graphical) \u6a21\u578b\uff0c\u5e38\u7528\u65bc\u6a19\u6ce8\u5e8f\u5217\u8cc7\u6599\u3002\u6211\u5011\u5229\u7528\u958b\u653e\u7684 CRF++[3]\u7a0b\u5f0f\u9032\u884c\u5be6\u9a57\uff0c\u70ba\u4e86\u4f7f CRF \u6a19 \u8a18\u80fd\u6709\u597d\u7684\u6e96\u78ba\u7387\uff0c\u6211\u5011\u5fc5\u9808\u8655\u7406\u539f\u59cb\u5927\u91cf\u6587\u5b57\u8cc7\u6599\uff0c\u5305\u542b\u4eba\u5de5\u6536\u96c6\u7b54\u6848\u3001\u6a19\u8a18\u7b54\u6848\u7b49\uff0c \u540c\u6642\u70ba\u4e86\u63d0\u5347\u6a21\u7d44\u8fa8\u8b58\u6548\u679c\u4e5f\u5fc5\u9808\u8981\u70ba\u8cc7\u6599\u505a\u9069\u7576\u5207\u5272\u3001\u9078\u64c7\u65b7\u8a5e\u5de5\u5177\u3001\u7d71\u4e00\u7b26\u865f\u3001\u6578 \u56e0\u6b64\u5728\u672c\u7bc7\u8ad6\u6587\u4e2d\u6211\u5011\u5c07\u4ee5\u4e0a\u7684\u52d5\u4f5c\u6a21\u7d44\u5316\uff0c\u4e26\u5c07\u5176\u6574\u5408\u6210\u4e00\u500b\u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u6a21\u578b\u7684\u7522 \u751f\u5de5\u5177\u3002 \u8868\u4e00\u3001\u4ee5 Snippets \u70ba\u6e2c\u8a66\u8cc7\u6599\u5c0d Stanford NER \u6e2c\u8a66\u6548\u80fd Testing Data Chinese Organization Name Chinese Location Name # Queries 2,000 200 # Sentences 38,692 2,638 # Distinct Entities 16,241 600 \u8868\u4e8c\u3001Stanford NER \u5c0d Snippets \u70ba\u8cc7\u6599\u4f86\u6e90\u4e4b\u8fa8\u8b58\u6548\u679c \u4f7f\u7528\u672c\u5de5\u5177\u53ef\u65b9\u4fbf\u7684\u8a13\u7df4\u4e0d\u540c\u8a9e\u8a00\u3001\u985e\u578b\u7684\u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u6a21\u7d44\uff0c\u6211\u5011\u4f7f\u7528\u6b32\u8fa8\u8b58\u7684\u547d\u540d \u5be6\u9ad4\u5217\u8868\u70ba\u672c\u5de5\u5177\u7684\u8f38\u5165\uff0c\u65bc\u7db2\u8def\u6536\u96c6\u5927\u91cf\u7684 Google \u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u900f\u904e\u81ea\u52d5\u6a19\u8a18 (Automatic Labeling)\u8207\u7279\u5fb5\u503c\u7684\u6e96\u5099\uff0c\u7522\u751f\u8a13\u7df4\u8cc7\u6599\u3002\u70ba\u4e86\u6e1b\u5c11\u547d\u540d\u5be6\u9ad4\u6a19\u8a18\u4e0d\u5b8c\u6574 \u7684\u554f\u984c\uff0c\u4ee5\u4e2d\u6587\u7d44\u7e54\u70ba\u4f8b\uff0c\u6211\u5011\u4e0d\u53ea\u5229\u7528\u55ae\u4e00\u7684\u7d44\u7e54\u540d\u7a31\u4f86\u5354\u52a9\u6a19\u8a18(\u7a31\u4e4b\u70ba UniLabeling) \uff0c\u4e5f\u63a1\u7528\u6240\u6709\u5df2\u77e5\u7684\u7d44\u7e54\u540d\u7a31\u4f86\u9032\u884c\u6a19\u8a18(\u7a31\u4e4b\u70ba FullLabeling) \u3002\u56e0\u81ea\u52d5 \u6a19\u8a18\u53ef\u80fd\u9020\u6210\u8a13\u7df4\u8cc7\u6599\u54c1\u8cea\u4e0d\u4f73\uff0c\u56e0\u6b64\u6211\u5011\u63a1\u7528\u81ea\u6211\u6e2c\u8a66(Self-Testing)\u80fd\u9032\u4e00\u6b65\u6539\u5584 \u8cc7\u6599\u54c1\u8cea\uff0c\u518d\u85c9\u7531\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2(Semi-supervised learning)\u65b9\u6cd5\uff0c\u5f15\u5165 Tri-Training \u589e \u52a0\u8a13\u7df4\u8cc7\u6599\u91cf\uff0c\u63d0\u5347\u8fa8\u8b58\u6a21\u578b\u4e4b\u6b63\u78ba\u7387\u3002 \u5be6\u9a57\u986f\u793a\u7cfb\u7d71\u5728\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58\u90e8\u4efd\u4ee5 Tri-Training \u6f14\u7b97\u6cd5\u78ba\u5be6\u4f7f\u5f97 F-Measure \u66f4\u9032 \u4e00\u6b65\u63d0\u5347\u81f3 86.1%\uff0c\u800c\u5728\u65e5\u6587\u7d44\u7e54\u540d\u7a31\u3001\u800c\u5728\u82f1\u6587\u7d44\u7e54\u540d\u7a31\u3001\u4e2d\u6587\u666f\u9ede\u540d\u7a31\u4e5f\u53ef\u9054\u5230 80.3%, 83.2%, 84.5%\u6548\u80fd\uff1b\u53e6\u5916\u5728\u9577\u547d\u540d\u5be6\u9ad4\u4e2d\u6587\u5730\u5740\u4ee5\u53ca\u82f1\u6587\u5730\u5740\u7684\u64f7\u53d6\u4e0a\uff0c F-Measure \u8fa8\u8b58\u6548\u679c\u4e5f\u5206\u5225\u9054\u5230 97.2% \u53ca 94.8%\u3002 \u4e8c\u3001 \u76f8\u95dc\u7814\u7a76 \u547d\u540d\u5be6\u9ad4\u8fa8\u8a8d\u5c6c\u65bc\u8cc7\u8a0a\u64f7\u53d6\u8207\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u7684\u4e00\u500b\u5171\u540c\u5206\u652f\uff0c\u4e5f\u662f\u8a31\u591a\u61c9\u7528\u9818\u57df\u7684\u91cd\u8981 \u57fa\u790e\u5de5\u5177\uff0c\u81ea\u975e\u7d50\u69cb\u5316\u6587\u5b57\u4e2d\u8b58\u5225\u5177\u6709\u7279\u5b9a\u610f\u7fa9\u7684\u547d\u540d\u5be6\u9ad4\uff0c\u5982\u4eba\u540d\u3001\u5730\u540d\u3001\u7d44\u7e54\u540d\u7a31\uff0c \u4ea6\u6216\u547d\u540d\u5be6\u9ad4\u76f8\u95dc\u5c6c\u6027\u5982\u96fb\u5b50\u90f5\u4ef6\u3001\u5730\u5740\u53ca\u5c08\u6709\u540d\u8a5e\u7b49\uff0c\u76ee\u524d\u6709\u8a31\u591a\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u53ca\u4e2d \u6587\u4eba\u540d\u8fa8\u8b58\u7684\u7814\u7a76\uff0c\u5229\u7528\u5e8f\u5217\u6a19\u8a18\u914d\u5408\u6a5f\u7387\u7d71\u8a08\u6a21\u578b\u662f\u4e3b\u8981\u8fa8\u8b58\u65b9\u5f0f\u3002 \uf06c \u8fa8\u8b58\u6b63\u5f0f\u6587\u7ae0\u4e2d\u6587\u547d\u540d\u5be6\u9ad4 Zhang \u7b49\u4eba[13]\u65bc 2007 \u5e74\u5c07\u591a\u500b CRF \u6a21\u578b\u4e32\u9023\u8d77\u4f86\u9032\u884c\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58\uff0c\u63a1\u7528\u7684\u7279\u5fb5\u503c \u5305\u542b\u662f\u5426\u70ba\u524d\u7d1a\u8f38\u51fa\u7684\u5404\u7a2e\u547d\u540d\u5be6\u9ad4\u3001\u5e38\u898b\u7684\u7d44\u7e54\u540d\u7a31\u958b\u982d\u3001\u5167\u5bb9\u8207\u7d50\u5c3e\u3001N-gram\u3002 \u4e26\u4ee5\u4e2d\u6587\u4eba\u6c11\u65e5\u5831\u65b0\u805e\u7a3f\u7576\u4f5c\u8a13\u7df4\u8cc7\u6599\uff0c\u5176\u6700\u7d42\u7684\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58 Recall \u53ef\u4ee5\u9054\u5230 88.78%\uff0cPrecision \u53ef\u9054\u5230 82.35%\u3002 2011 \u5e74 Yao[11]\u5c07\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u5206\u70ba\u4e09\u500b\u90e8\u4efd\u5305\u542b\u524d\u7f6e\u8a5e (Prefix words) \u3001\u4e2d\u9593\u8a5e (Middle words) \u3001\u8a18\u865f\u8a5e(Mark words) \uff0c\u8209\u4f8b\u4f86\u8aaa\uff1a \u300c\u4e2d\u570b\u79fb\u52d5\u901a\u8a0a\u516c\u53f8\u300d\u53ef\u4ee5\u62c6\u6210\u300c\u4e2d\u570b+\u79fb Task Precision Recall F-measure Stanford NER Chinese Organization Name 0.518 0.542 0.530 Chinese Location Name 0.215 0.188 0.201 \u65b0\u805e\u7576\u4f5c\u6e2c\u8a66\u8cc7\u6599\uff0c\u5176\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58 Recall \u53ef\u4ee5\u9054\u5230 87.24%\uff0cPrecision \u53ef\u9054\u5230 95.9%\u3002 2012 \u5e74 Ling \u7b49\u4eba[7]\u5c07\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u8a9e\u6599\u65b7\u8a5e\u5f8c\u62c6\u89e3\u70ba\u591a\u500b\u4fee\u98fe\u8a5e(Modifiers)+\u6838\u5fc3 \u7279\u5fb5\u8a5e(Core Feature Word)\u3002 \u5728\u7d71\u8a08\u8a13\u7df4\u8cc7\u6599\u5f8c\uff0c\u627e\u51fa\u5e38\u7528\u7684\u6838\u5fc3\u7279\u5fb5\u8a5e\uff0c\u5efa\u7acb\u6838\u5fc3 \u7279\u5fb5\u8a5e\u5eab\u7576\u4f5c\u7d44\u7e54\u540d\u7a31\u7684\u7d50\u5c3e\uff0c\u4e26\u4ee5\u7279\u5fb5\u5224\u65b7\u7d44\u7e54\u540d\u7a31\u7684\u8d77\u9ede\u3002\u53d6\u5f97\u5019\u9078\u8005\u4e4b\u5f8c\uff0c\u5229\u7528 \u898f\u5247\u5f0f\u7684\u8fa8\u8a8d\u65b9\u6cd5(Rule-based Named-Entity Recognition)\u9032\u884c\u4fee\u6b63\u3002\u6700\u5f8c\u7684\u5be6\u9a57\u7d50\u679c \u986f\u793a\uff0c F-measure \u6700\u9ad8\u53ef\u9054\u5230 85.7%\u3002 \uf06c \u8fa8\u8b58\u975e\u6b63\u5f0f\u6587\u7ae0\u4e2d\u6587\u547d\u540d\u5be6\u9ad4 \u76ee \u524d \u5df2 \u7d93 \u6709 \u8a31 \u591a \u5982 \u4e0a \u8ff0 \u5728 \u6b63 \u5f0f \u6587 \u7ae0 \u4e2d \u7684 \u4e2d \u6587 \u7d44 \u7e54 \u540d \u7a31 \u8fa8 \u8a8d ( CONER \uff0c Chinese Organization Named Entity Recognition)\u7814\u7a76[7][11][13]\uff0c\u4f46\u7528\u9019\u985e\u8a13\u7df4\u8cc7\u6599\u7522\u751f\u7684\u6a21\u578b \u4ee5\u5f80 CRF \u5e8f\u5217\u6a19\u8a18\u6a21\u578b\u7684\u8a13\u7df4\u8cc7\u6599\u7686\u70ba\u4eba\u5de5\u65b9\u5f0f\u7522\u751f\uff0c\u96d6\u7136\u8cc7\u6599\u7684\u54c1\u8cea\u53ef\u4ee5\u4fe1\u8cf4\uff0c\u4f46 \u9700\u82b1\u8cbb\u5927\u91cf\u7684\u6642\u9593\u8207\u4eba\u529b\u3002\u7531\u65bc\u4eba\u5de5\u5c0d\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u9032\u884c\u7b54\u6848\u6a19\u8a18\u6210\u672c\u904e\u9ad8\uff0c\u70ba\u6b64\uff0c\u672c \u5de5\u5177\u4f7f\u7528\u5df2\u77e5\u7684\u547d\u540d\u5be6\u9ad4\u4f5c\u70ba\u7b54\u6848\uff0c\u5c0d\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u5167\u5bb9\u9032\u884c\u81ea\u52d5\u6a19\u8a18\uff0c\u8a72\u6a19\u8a18\u5373\u70ba\u6b32 \u64f7\u53d6\u7684\u76ee\u6a19\uff0c\u5982\u6b64\u53ef\u4ee5\u7bc0\u7701\u5927\u91cf\u8a13\u7df4\u8cc7\u6599\u6a19\u8a18\u6210\u672c\u3002\u57fa\u65bc Lin \u7b49\u4eba[6]\u7684\u7814\u7a76\u986f\u793a\uff0c\u4f7f\u7528 \u55ae\u4e00\u7684\u5546\u5bb6\u540d\u7a31\u4f86\u81ea\u52d5\u6a19\u8a18(\u7a31\u4e4b\u70ba UniLabeling)\u7684\u8fa8\u8b58\u6548\u679c\u8f03\u63a1\u7528\u6240\u7528\u5546\u5bb6\u540d\u7a31\u4f86 \u9032\u884c\u6a19\u8a18(\u7a31\u4e4b\u70ba FullLabeling)\u8981\u4f86\u7684\u5dee\uff0c\u539f\u56e0\u662f\u5728 UniLabeling \u6a21\u578b\u4e2d\uff0c\u8cc7\u6599\u542b\u6709\u8f03 \u591a\u6a19\u8a18\u4e0d\u5b8c\u5168\u7684\u96dc\u8a0a\uff0c\u4f7f\u5f97\u6548\u80fd\u4e0b\u964d\uff1b\u800c FullLabeling \u6a21\u578b\u4f7f\u7528\u6240\u6709\u7684\u5546\u5bb6\u540d\u7a31\u9032\u884c\u6a19 \u8a18\uff0c\u56e0\u6b64\u96dc\u8a0a\u5927\u5e45\u6e1b\u5c11\u3002\u70ba\u6e1b\u5c11\u96dc\u8a0a\u5f71\u97ff\uff0c\u672c\u7cfb\u7d71\u63a1\u7528 FullLabeling \u7684\u65b9\u5f0f\u9032\u884c\u81ea\u52d5\u6a19 \u8a18\u3002 \uf06c \u6bd4\u5c0d\u6cd5\u6a19\u8a18\u9577\u547d\u540d\u5be6\u9ad4 \u81ea\u52d5\u6a19\u8a18\u7684\u6311\u6230\u5728\u65bc\u5c0d\u65bc\u8f03\u9577\u62fc\u97f3\u6587\u5b57\u7684\u547d\u540d\u5be6\u9ad4\u4f7f\u7528\u5b8c\u5168\u76f8\u914d(Exact Match)\u4e26\u4e0d\u80fd \u6709\u6548\u7684\u6a19\u8a18\u3002\u9019\u662f\u56e0\u70ba\u8f03\u9577\u62fc\u97f3\u6587\u5b57\u7684\u547d\u540d\u5be6\u9ad4\u5728\u4e0d\u6b63\u5f0f\u7db2\u9801\u6587\u7ae0\u4e2d\u7684\u66f8\u5beb\u65b9\u5f0f\u76f8\u8f03\u6b63 \u5f0f\u6587\u7ae0\u5177\u6709\u5f48\u6027\uff0c\u4f8b\u5982\u82f1\u6587\u5730\u5740\u5728\u6b63\u5f0f\u66f8\u5beb\u6642\u6703\u6709\u56fa\u5b9a\u683c\u5f0f\u3001\u62fc\u5beb\u4ee5\u53ca\u7e2e\u5beb\u65b9\u5f0f\u4e00\u81f4\u7b49 \u898f\u5b9a\u3002\u4f46\u6211\u5011\u5229\u7528\"1131 Mountain Rd NW, Albuquerque, NM 87102\"\u5728 Google \u641c\u5c0b\u6642\uff0c \u96d6\u7136\u96d9\u5f15\u865f(\"\")\u80fd\u9650\u5236\u641c\u5c0b\u7d50\u679c\u90fd\u8981\u6709\u5305\u542b\u641c\u5c0b\u8a5e\uff0c\u4f46\u96d9\u5f15\u865f\u4e26\u4e0d\u80fd\u4fdd\u8b49\u641c\u5c0b\u7d50\u679c\u7247 \u6bb5\u5167\u5bb9\u4e2d\u7684\u547d\u540d\u5be6\u9ad4\u8207\u641c\u5c0b\u8a5e\u5b8c\u5168\u4e00\u81f4\uff0c\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u5c31\u7b97\u662f\u5728\u641c\u5c0b\u8a5e\u4e2d\u7a7f\u63d2\u4e0d\u540c\u6a19 \u9ede\u7b26\u865f\uff0c\u6216\u662f\u6c92\u6709\u4efb\u4f55\u6a19\u9ede\u7b26\u865f\u90fd\u6703\u88ab\u641c\u5c0b\u51fa\u4f86\u3002\u5f9e\u5716\u4e8c\u53ef\u4ee5\u770b\u5230\u5728 Google \u641c\u5c0b\u524d 10 \u7b46\u7d50\u679c\u7247\u6bb5\u5c31\u6709 7 \u7a2e\u4e0d\u540c\u5beb\u6cd5\uff0c\u800c\u5b83\u5011\u660e\u986f\u90fd\u662f\u8868\u793a\u6b64\u4e00\u5730\u5740\u3002 \u5716\u4e8c\u3001\"1131 Mountain Rd NW, Albuquerque, NM 87102\"\u5728 Google \u641c\u5c0b\u5f97\u5230\u591a\u7a2e\u5beb\u6cd5 \u5982\u4f7f\u7528\u5b8c\u5168\u76f8\u914d\u65b9\u5f0f\u4f86\u505a\u81ea\u52d5\u6a19\u8a18\uff0c\u9019\u4e9b\u5730\u5740\u5c07\u6c92\u8fa6\u6cd5\u88ab\u6a19\u8a18\u51fa\u4f86\u3002\u70ba\u4e86\u8655\u7406\u8f03\u9577\u547d\u540d \u5be6\u9ad4\u53ef\u80fd\u56e0\u6a19\u9ede\u53ca\u7e2e\u5beb\u7b49\u554f\u984c\u7121\u6cd5\u88ab\u8fa8\u8a8d\u51fa\u4f86\u7684\u60c5\u5f62\uff0c\u6211\u5011\u4f7f\u7528\u6392\u6bd4(Alignment)\u7684 \u65b9\u5f0f\u627e\u51fa\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u5167\u5bb9\u4e2d\u53ef\u80fd\u7684\u540d\u547d\u5be6\u9ad4\u4f4d\u7f6e\u3002\u5728\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u627e\u5c0b\u76ee\u6a19\u547d\u540d\u5be6 \u9ad4\u6642\uff0c\u6211\u5011\u5e0c\u671b\u6a19\u8a18\u76ee\u6a19\u7684\u547d\u540d\u5be6\u9ad4\u5728\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u8207\u67e5\u8a62\u8a5e\u76f8\u5339\u914d\u7684\u5b57\u8d8a\u96c6\u4e2d\u76f8\u9130 \u8d8a\u597d\uff0c\u56e0\u6b64\u6211\u5011\u8a2d\u8a08\u4e86\u6392\u6bd4\u6a19\u8a18\u6cd5(AlignmentLabeling)\u6a19\u8a18\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u518d\u4ee5\u6392\u6bd4 \u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4ee5\u53ca\u641c\u5c0b\u8a5e\u6240\u7522\u751f\u7684\u76f8\u914d Match \u53ca\u9593\u9694 Gap \u5927\u5c0f\uff0c\u505a\u70ba\u6211\u5011\u5224\u65b7\u6b64\u4e00\u6392 \u6bd4\u5f8c\u7684\u7d50\u679c\u662f\u5426\u8a72\u6a19\u8a18\u7684\u4f9d\u64da\u3002\u5982\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u8207\u67e5\u8a62\u8a5e\u7d93\u904e\u6392\u6bd4\u5f8c\u7b26\u5408(1)\u76f8\u914d\u5b57 \u6578\u5927\u65bc\u547d\u540d\u5be6\u9ad4\u9577\u5ea6 Len \u6e1b\u53bb\u9593\u9694\u5927\u5c0f\u7684\u4e00\u534a\uff0c\u4e14(2)\u7b2c\u4e00\u500b\u6392\u6bd4\u914d\u5c0d\u5230\u7684\u5b57 1 \u5230\u6700\u5f8c \u4e00\u500b\u6392\u6bd4\u914d\u5c0d\u5230\u7684\u5b57 \u8207\u547d\u540d\u5be6\u9ad4\u67e5\u8a62\u8a5e\u9577\u5ea6\u5dee\u8ddd\u5c0f\u65bc 3\uff0c\u5247\u7cfb\u7d71\u5c07\u6703\u6a19\u8a18\u70ba\u51fa\u73fe\u7bc4 \u4f8b\u3002 ( \u210e > \u2212 2 ) \u4e14(|( \u2212 1 ) \u2212 Len| < 3) \u7136\u800c\u6392\u6bd4\u6a19\u8a18\u6cd5\u5c0d\u65bc\u975e\u62fc\u97f3\u6587\u5b57\u5982\u4e2d\u6587\u61c9\u7528\u7684\u6548\u679c\u4e0d\u5982\u62fc\u97f3\u6587\u5b57\u3002\u4e2d\u6587\u4e0d\u540c\u65bc\u82f1\u6587\uff0c\u4e2d \u6587\u7684\u7e2e\u5beb\u662f\u5f9e\u9577\u53e5\u5b50\u4e2d\u53d6\u5177\u6709\u4ee3\u8868\u6027\u7684\u5b57\u51fa\u4f86\uff0c\u4e26\u4e14\u4e0d\u6703\u5728\u55ae\u4e00\u547d\u540d\u5be6\u9ad4\u4e2d\u96a8\u610f\u52a0\u5165\u6a19 \u9ede\u7b26\u865f\u3002\u518d\u8005\u672c\u7cfb\u7d71\u5728\u5c0d Google \u641c\u5c0b\u6642\u6703\u4f7f\u7528\u96d9\u5f15\u865f\uff0c\u56e0\u6b64\u80fd\u78ba\u4fdd\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u7684 \u9577 \u547d \u540d \u5be6 \u9ad4 \u6703 \u8207 \u67e5 \u8a62 \u8a5e \u5b8c \u5168 \u4e00 \u81f4 \uff0c \u5982 \u6b64 \u6211 \u5011 \u5c07 \u53ef \u5229 \u7528 \u5b8c \u5168 \u76f8 \u914d \u6a19 \u8a18 \u6cd5 (ExactMatchLabeling)\u6b63\u78ba\u4e14\u6709\u6548\u7387\u7684\u505a\u81ea\u52d5\u6a19\u8a18\u3002 3.2 \u5b57\u4e32\u5207\u5272\u8207\u6a19\u8a18\u6a21\u7d44 \u5728\u8a13\u7df4\u8cc7\u6599\u7684\u6e96\u5099\u4e0a\uff0c\u96d6\u7136\u53ef\u4ee5\u63a1\u7528\u5b8c\u6574 Google \u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u505a\u70ba\u6a23\u672c\u55ae\u5143\u9032\u884c\u8a13\u7df4\uff0c \u4f46\u904e\u9577\u7684\u53e5\u5b50\u6703\u81f4\u4f7f\u8a13\u7df4\u6642\u9593\u62c9\u9577\uff0c\u4e5f\u96e3\u6709\u597d\u7684\u8fa8\u8b58\u6548\u679c\u3002\u4f46\u662f\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u7684\u7db2\u9801 \u6587\u7ae0\u6703\u6709\u6a19\u9ede\u7b26\u865f\u6df7\u7528\u4ee5\u53ca\u683c\u5f0f\u67b6\u69cb\u4e0d\u56b4\u8b39\u7684\u554f\u984c\uff0c\u76f4\u63a5\u5229\u7528\u7d71\u4e00\u7684\u5207\u5272\u65b9\u6cd5\u5c07\u9020\u6210\u8a13 \u7df4\u6a23\u672c\u9577\u5ea6\u76f8\u5dee\u904e\u5927\u4e14\u54c1\u8cea\u4e0d\u826f\u3002\u70ba\u6e96\u5099\u9069\u7576\u9577\u5ea6\u7684\u8a13\u7df4\u53e5\u5b50\uff0c\u6211\u5011\u79fb\u9664\u641c\u5c0b\u7d50\u679c\u7247\u6bb5 \u4e2d\u7684\u7a7a\u767d\u5b57\u5143\uff0c\u5229\u7528\u81ea\u52d5\u6a19\u8a18\u7684\u7b54\u6848\u70ba\u57fa\u6e96\u53d6\u524d\u5f8c W \u5b57\u5143\u70ba\u7a97\u53e3\u5927\u5c0f\uff0c\u5728\u6211\u5011\u5be6\u9a57\u4e2d\uff0c \u4e2d\u6587\u53ca\u65e5\u6587\u8a2d\u5b9a W \u70ba 20\uff0c\u800c\u82f1\u6587\u5247\u8a2d W \u70ba 10\uff0c\u5c07\u6587\u5b57\u5207\u70ba\u8a31\u591a\u5340\u584a\uff0c\u4ee5\u5340\u584a\u70ba\u4e00\u500b \u8a13\u7df4\u6a23\u672c\uff0c\u6700\u5f8c\u53bb\u9664\u91cd\u8907\u7684\u6a23\u672c\uff0c\u5982\u6b64\u53ef\u4f7f\u8a13\u7df4\u6a23\u672c\u6db5\u84cb\u547d\u540d\u5be6\u9ad4\uff0c\u4e5f\u80fd\u6709\u9069\u7576\u7684\u975e\u547d \u540d\u5be6\u9ad4\u7bc4\u4f8b\u3002\u5716\u4e09\u70ba\u8a2d\u5b9a W \u70ba 20 \u7684\u5207\u5272\u7bc4\u4f8b\u3002 \u5716\u4e09\u3001\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58\u641c\u5c0b\u7247\u6bb5\uff0c\u53d6 N=20, \u4ee5\u8a60\u5c55\u5546\u884c\u70ba\u57fa\u6e96 \u5011\u7684\u7cfb\u7d71\u6703\u5148\u5c07\u6240\u6709\u5168\u5f62\u7b26\u865f\u8f49\u63db\u6210\u534a\u5f62\u7b26\u865f\uff0c\u5982\u8868\u4e09\u6240\u793a\u3002 \u8868\u4e09\u3001\u5168\u5f62\u7b26\u865f\u8f49\u63db\u6210\u534a\u5f62\u7b26\u865f\u7bc4\u4f8b \u5713\u5f27\u62ec\u865f \u975e\u5713\u5f27\u62ec\u865f ((\ufe59 ==> ( [ {\u300c \u3014 \uff5b \u3008 \u300e \u300a \u3010\ufe5d\ufe5b ==> [ \u7b54\u6848\u6a19\u8a18\u65b9\u5f0f\u6211\u5011\u9078\u7528 Start/End \u6a19\u8a18\u6cd5\uff0c\u6b64\u7a2e\u6a19\u8a18\u6cd5\u5171\u6709 5 \u500b\u6a19\u8a18 B\u3001I\u3001E\u3001S\u3001O\uff0c \u4f9d\u5e8f\u8868\u793a\u547d\u540d\u5be6\u9ad4\u7684\u958b\u59cb\u3001\u4e2d\u9593\u3001\u7d50\u675f\u3001\u55ae\u4e00\u5e8f\u5217\u55ae\u5143\u4ee5\u53ca\u975e\u547d\u540d\u5be6\u9ad4\u7684\u5e8f\u5217\u55ae\u5143\uff0c\u56e0 \u70ba\u5c0d\u958b\u59cb\u548c\u7d50\u675f\u90fd\u7d66\u4e88\u4e0d\u540c\u7684\u6a19\u8a18\uff0c\u53ef\u4ee5\u63d0\u6607\u908a\u754c\u7684\u5075\u6e2c\u6548\u679c\u3002 3.3 \u7279\u5fb5\u503c\u64f7\u53d6\u6a21\u7d44 \u7279\u5fb5\u503c\u7684\u63d0\u53d6\u662f\u8a13\u7df4\u8cc7\u6599\u6e96\u5099\u4e2d\u975e\u5e38\u91cd\u8981\u7684\u4e00\u6b65\uff0c\u5e38\u898b\u7684\u7279\u5fb5\u662f\u5224\u5b9a\u4e00\u500b\u5b57\u662f\u5426\u70ba\u5177\u6709 \u67d0\u7a2e\u5c6c\u6027\uff0c\u4f8b\u5982\u662f\u5426\u70ba\u6578\u5b57\u6216\u662f\u767e\u5bb6\u59d3\u7b49\uff0c\u56e0\u6b64\u6e96\u5099\u76f8\u95dc\u8a5e\u5eab\u662f\u76f8\u7576\u7e41\u7463\u7684\u4e00\u74b0\u3002\u4e00\u822c \u8aaa\u4f86\uff0c\u5728\u5224\u65b7\u4e00\u6bb5\u6587\u5b57\u662f\u5426\u662f\u7279\u5b9a\u547d\u540d\u5be6\u9ad4\u6642\uff0c\u6703\u4f9d\u9760\u5169\u985e\u7279\u5fb5\uff0c\u7b2c\u4e00\u7a2e\u662f\u5916\u90e8\u7279\u5fb5 (Outside Feature) \uff0c\u9019\u7a2e\u7279\u5fb5\u843d\u5728\u547d\u540d\u5be6\u9ad4\u7684\u5de6\u53f3\uff0c\u7b2c\u4e8c\u7a2e\u5247\u662f\u547d\u540d\u5be6\u9ad4\u7684\u5167\u90e8\u7279\u5fb5 (Inside Feature) \u3002\u7136\u800c\u9019\u4e9b\u7279\u5fb5\u5f80\u5f80\u5fc5\u9808\u8981\u9760\u8457\u719f\u6089\u8a9e\u8a00\u6216\u5c0d\u8a72\u8fa8\u8b58\u9818\u57df\u4e86\u89e3\u7684\u4eba\u4f86 \u9010\u4e00\u7522\u751f\uff0c\u5982\u6211\u5011\u8981\u91dd\u5c0d\u4e2d\u6587\u4ee5\u5916\u7684\u8a9e\u8a00\u9032\u884c\u8fa8\u8b58\uff0c\u95dc\u9375\u8a5e\u5eab\u5c31\u5fc5\u9808\u8981\u7531\u719f\u6089\u8a72\u570b\u8a9e\u8a00 \u4e14\u6709\u8db3\u5920\u80cc\u666f\u77e5\u8b58\u7684\u4eba\u54e1\u4f86\u6e96\u5099\u3002 \u70ba\u4e86\u4f7f\u5f97\u672c\u7cfb\u7d71\u80fd\u5920\u907f\u514d\u9019\u7a2e\u8a9e\u8a00\u80fd\u529b\u53ca\u8fa8\u8b58\u4e3b\u984c\u4e0a\u7684\u9650\u5236\u9054\u5230\u901a\u7528\u7684\u76ee\u7684\uff0c\u6211\u5011\u7684\u505a \u6cd5\u70ba\u7d71\u8a08\u5b57\u8a5e\u51fa\u73fe\u983b\u7387\uff0c\u81ea\u52d5\u7522\u751f\u5e38\u898b\u7684\u95dc\u9375\u8a5e\u5eab\u3002\u5be6\u52d9\u4e0a\uff0c\u6211\u5011\u7d71\u8a08\u547d\u540d\u5be6\u9ad4\u4e2d\u7684\u524d \u4e00\u5b57\u3001\u5169\u5b57\u53ca\u4e09\u5b57\u7684\u983b\u7387\u4ee5\u53ca\u6700\u5f8c\u4e00\u5b57\u3001\u5169\u5b57\u53ca\u4e09\u5b57\u7684\u983b\u7387\uff0c\u5982\u8868\u56db\u4e2d ID 4~9\u3002\u8209\u4f8b\u800c \u8a00\uff0c\u4e2d\u6587\u5546\u5bb6\u540d\u7a31\u6700\u5f8c\u4e00\u5b57\u5e38\u51fa\u73fe\u300c\u5edf\u300d \u3001 \u300c\u838a\u300d \u3001 \u300c\u5e97\u300d\u7b49\u4e00\u5b57\u8a5e\uff0c\u6216\u662f\u300c\u4e8b\u52d9\u300d \u3001 \u300c\u6578\u4f4d\u300d \u7b49\u5169\u5b57\u8a5e\uff0c\u53c8\u6216\u662f\u300c\u57fa\u91d1\u6703\u300d \u3001 \u300c\u96dc\u8ca8\u5e97\u300d\u7b49\u4e09\u5b57\u8a5e\u3002\u6211\u5011\u4e5f\u4ee5\u547d\u540d\u5be6\u9ad4\u51fa\u73fe\u5728\u6a23\u672c\u4e2d\u7684 \u4f4d\u7f6e\u70ba\u57fa\u6e96\uff0c\u7d71\u8a08\u51fa\u73fe\u5728\u5176\u524d\u5f8c\u65b9\u5b57\u3001\u8a5e\u983b\u7387\uff0c\u5982\u8868\u56db\u4e2d ID 10~15 \u5373\u70ba\u5916\u90e8\u7279\u5fb5\u503c\u3002 \u6211\u5011\u5229\u7528\u81ea\u52d5\u9078\u64c7\u524d M \u500b\u5e38\u51fa\u73fe\u7684\u5b57\u6216\u8a5e\u4f86\u7522\u751f\u95dc\u9375\u8a5e\u5eab\uff0c\u5728\u5be6\u9a57\u7ae0\u7bc0\u5c07\u6709\u91dd\u5c0d\u95dc\u9375 \u8a5e\u5eab\u5927\u5c0f\u5c0d\u8fa8\u8b58\u6548\u679c\u7684\u5f71\u97ff\u9032\u884c\u5be6\u9a57\u3002\u9664\u4e0a\u8ff0 12 \u500b\u81ea\u52d5\u7522\u751f\u4e4b\u7279\u5fb5\u503c\u5916\uff0c\u518d\u52a0\u4e0a\u91dd\u5c0d \u8fa8\u8b58\u985e\u5225\u7279\u5225\u6e96\u5099\u7684\u7279\u5fb5\u5982\u7e23\u5e02\u540d\u7a31\u53ca\u5176\u7c21\u7a31\u3001\u8a5e\u6027(POS)tagging\u3001\u662f\u5426\u70ba\u6a19\u9ede\u7b26\u865f \u7b49\u7279\u5fb5\uff0c\u6b64\u5916\u56e0\u70ba\u5728\u7db2\u9801\u4e2d\u547d\u540d\u5be6\u9ad4\u4e5f\u5e38\u6709\u55ae\u7368\u51fa\u73fe\u7684\u60c5\u5f62\uff0c\u56e0\u6b64\u4e00\u6bb5\u6587\u5b57\u7684\u8d77\u9ede\u5c31\u8b8a \u6210\u91cd\u8981\u7279\u5fb5\uff0c\u5982\u679c\u662f\u6a23\u672c\u55ae\u5143\u7684\u8d77\u9ede\u6216\u524d\u4e00\u500b\u5b57\u5143\u5c6c\u65bc\u7b26\u865f\u985e\uff0c\u5c31\u5177\u6709\u958b\u59cb\u7279\u5fb5(Start Feature) \uff0c\u7576\u5b57\u5143\u662f\u6a23\u672c\u55ae\u5143\u7684\u7d50\u5c3e\u6216\u4e0b\u4e00\u500b\u5b57\u5143\u5c6c\u65bc\u7b26\u865f\u985e\uff0c\u5c31\u5177\u6709\u7d50\u5c3e\u7279\u5fb5(End Feature)\uff0c \u5171 6 \u500b\u9810\u8a2d\u7279\u5fb5\u503c\u3002\u5728\u4e0d\u53e6\u5916\u8abf\u6574\u7684\u60c5\u5f62\u672c\u5de5\u5177\u7e3d\u5171 18 \u500b\u7279\u5fb5\u503c\u3002 3.4 \u81ea\u6211\u6e2c\u8a66\u8207\u5354\u540c\u8a13\u7df4 \u6211\u5011\u4f7f\u7528\u958b\u653e\u4e14\u514d\u8cbb\u7684 CRF++[3]\u7a0b\u5f0f\u505a\u70ba\u5e8f\u5217\u6a19\u8a18\u6a21\u578b\u8a13\u7df4\u65b9\u6cd5\u3002\u7531\u65bc\u672c\u7814\u7a76\u63a1\u7528\u81ea \u52d5\u5316\u7684\u6280\u8853\u6536\u96c6\u5927\u91cf\u975e\u7d50\u69cb\u5316\u7684\u8cc7\u6599\u4ee5\u53ca\u81ea\u52d5\u6a19\u8a18\u7522\u751f\u7684\u8a13\u7df4\u8cc7\u6599\uff0c\u9019\u4e9b\u5927\u91cf\u7684\u8a13\u7df4\u8cc7 \u6599\u53ef\u80fd\u5305\u542b\u932f\u8aa4\u6a19\u8a18\uff0c\u70ba\u4e86\u63d0\u5347\u8a13\u7df4\u8cc7\u6599\u7684\u54c1\u8cea\uff0c\u6211\u5011\u7684\u5de5\u5177\u8a2d\u8a08\u5728\u5b78\u7fd2\u904e\u7a0b\u4e2d\u53ef\u9078\u64c7 \u4f7f\u7528 Self-testing \u5c07\u96dc\u8a0a\u79fb\u9664\u63d0\u9ad8\u8a13\u7df4\u8cc7\u6599\u7684\u54c1\u8cea\u3002Self-testing \u7684\u5be6\u4f5c\u65b9\u5f0f\u662f\u4f7f\u7528\u8a13\u7df4\u5b8c \u6210\u7684\u6a21\u578b\u5c0d\u8a13\u7df4\u8cc7\u6599\u505a\u6e2c\u8a66\u4e26\u8f38\u51fa\u6a5f\u7387\uff0c\u82e5\u8a72\u6a5f\u7387\u4f4e\u65bc\u9580\u6abb\u503c\u5247\u8a8d\u5b9a\u8a72\u8a9e\u53e5\u70ba\u96dc\u8a0a\uff0c\u81ea \u8a13\u7df4\u8cc7\u6599\u4e2d\u79fb\u9664\uff0c\u518d\u4ee5\u79fb\u9664\u96dc\u8a0a\u5f8c\u7684\u8cc7\u6599\u91cd\u65b0\u8a13\u7df4\u6a21\u578b\uff0c\u5728\u672c\u7814\u7a76\u7684\u5be6\u9a57\u4e2d\u8a2d\u5b9a\u6a5f\u7387\u70ba 0.7\uff0c\u5176\u503c\u53ef\u4ee5\u8996\u60c5\u6cc1\u8abf\u6574\u3002 \u8868\u793a\uff0c\u524d\u5f8c\u6b21\u758a\u4ee3\u9593\u7684\u932f\u8aa4\u7387\u6bd4 \u4f8b\u516c\u5f0f | | < \u22121 | \u22121 |\u5c07\u7121\u6cd5\u6210\u7acb\uff0c\u6b64\u6642\u5247\u9808\u5c0d \u505a\u53d6\u6a23\u52d5\u4f5c\uff0c\u7531 s= \u2308 \u22121 | \u22121 | \u2212 1\u2309 \u516c\u5f0f\u8a08\u7b97\u53ef\u4ee5\u81ea \u96a8\u6a5f\u6311\u9078 s \u7b46\u8cc7\u6599\u70ba\u65b0\u589e\u7684\u8a13\u7df4\u8cc7\u6599\uff0c\u78ba\u4fdd\u516c\u5f0f | | < \u22121 | \u22121 | Tri-Training \u5728\u5927\u91cf\u8cc7\u6599\u7684\u60c5\u6cc1\u4e0b\uff0c\u50c5\u53ef\u81ea U \u4e2d\u9078\u53d6\u5c11\u91cf\u8cc7\u6599\u4f5c\u70ba\u65b0\u7684\u8a13\u7df4\u8cc7\u6599\uff0c\u5c0d\u7cfb \u7d71\u6548\u80fd\u5e7e\u4e4e\u6c92\u6709\u5f71\u97ff\u7684\u554f\u984c\u3002 \u56db\u3001 \u5be6\u9a57 \u672c\u8ad6\u6587\u76ee\u7684\u5728\u5b8c\u6210\u4e00\u500b\u4e0d\u9650\u8a9e\u8a00\u3001\u4e3b\u984c\u7684 Web NER \u6a21\u578b\u81ea\u52d5\u7522\u751f\u5de5\u5177\uff0c\u6211\u5011\u4e5f\u5c07\u5f9e\u5be6 \u9a57\u4e86\u89e3\u81ea\u52d5\u6a19\u8a18\u7522\u751f\u7684\u8a13\u7df4\u8cc7\u57fa\u672c\u6548\u80fd(Basic) \u3001\u900f\u904e Self-Testing \u8cc7\u6599\u904e\u6ffe\u3001\u4ee5\u53ca Tri-Training \u7b49\u65b9\u6cd5\u5c0d\u65bc\u6548\u80fd\u7684\u5f71\u97ff\u3002\u5c0d\u65bc\u672c\u7cfb\u7d71\u6240\u7522\u751f\u7684\u7279\u5fb5\u64f7\u53d6\u65b9\u6cd5\uff0c\u6211\u5011\u4e5f\u5c07\u61c9 \u7528\u4e2d\u6587\u5546\u5bb6\u540d\u7a31\u8fa8\u8b58\u5be6\u9a57\u6bd4\u8f03\u4eba\u5de5\u6e96\u5099\u95dc\u9375\u8a5e\u5eab\u53ca\u4f7f\u7528\u7d71\u8a08\u51fa\u73fe\u983b\u7387\u7684\u65b9\u5f0f\u81ea\u52d5\u7522\u751f \u95dc\u9375\u8a5e\u5eab\u5c0d\u65bc\u6548\u80fd\u7684\u5f71\u97ff\u3002 \uf06c \u65e5\u6587\u5546\u5bb6\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58 \u6392\u540d\u524d5\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4e26\u5c0d\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u9032\u884c\u5b8c\u5168\u76f8\u914d\u7684FullLabeling\u6a19\u8a18\u7522\u751f\u8a13\u7df4 \u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u5171\u63d0\u53d688,074\u500b\u53e5\u5b50\u3002\u6e2c\u8a66\u8cc7\u6599\u7684\u90e8\u4efd\u5247\u53e6\u5916\u53d6200\u7b46\u5730\u5740\u70ba\u95dc\u9375\u5b57\uff0c\u6536 \u5740257\u500b\u3002 \u5916\u89c0\u5bdf\u5e38\u898b\u65bc\u5546\u5bb6\u540d\u7a31\u524d\u5f8c\u7684\u5b57\u8a5e\u7522\u751f\u66f4\u591a\u8a5e\u5eab\u3002 \u5f8c\u4e4b\u8cc7\u6599\u70ba\u57fa\u790e\uff0c\u9032\u884cTri-Training\u6f14\u7b97\u6cd5\u3002 \u81ea\u52d5\u7522\u751f\u8a5e\u5eab\u6548\u80fd\u7684\u5dee\u7570\u3002\u7279\u5fb5\u503c\u5305\u542b\u4eba\u5de5\u6536\u96c6\u7684\u670d\u52d9\u8a5e\u3001\u7522\u54c1\u8a5e\u4ee5\u53ca\u5730\u6a19\u8a5e\u8a5e\u5eab\uff0c\u53e6 \u5716\u4e94\u3001Basic\u3001Self-Testing \u53ca Tri-Training \u5728\u4e2d\u6587\u7d44\u7e54\u8fa8\u8b58\u4eba\u5de5\u7522\u751f\u95dc\u9375\u8a5e\u5eab\u4e4b\u6548\u80fd \u63a5\u4e0b\u4f86\u7684\u5be6\u9a57\u4e2d\u81ea\u52d5\u7522\u751f\u95dc\u9375\u8a5e\u5eab\u5927\u5c0f\u7686\u8a2d\u70ba100\uff0c\u4e26\u4ee5Self-testing\u4f4e\u65bc0.7\u70ba\u96dc\u8a0a\u53bb\u9664 F-measure 0.607 0.948 0.972 0.490 \u9664\u81ea\u52d5\u7522\u751f\u95dc\u9375\u8a5e\u5eab\u4e4b\u5916\uff0c\u6211\u5011\u4ee5\u4e2d\u6587\u7d44\u7e54\u8fa8\u8b58\u70ba\u4f8b\u63a1\u7528\u4eba\u5de5\u7522\u751f\u95dc\u9375\u8a5e\u5eab\u6bd4\u8f03\u8207\u7cfb\u7d71 \u8cc7\u6599(L)\u3002\u672a\u6a19\u8a18\u8cc7\u6599(U)\u4f7f\u752830,000\u7b46\u5546\u5bb6\u540d\u9032\u884c\u67e5\u8a62\uff0c\u53d6\u6bcf\u7b46\u641c\u5c0b\u6392\u540d\u524d10\u7684\u641c \u4f9b\u63d0\u53d649,851\u500b\u53e5\u5b50\u3002\u6e2c\u8a66\u8cc7\u6599\u53e6\u5916\u53d6200\u7b46\u82f1\u6587\u7d44\u7e54\u540d\u7a31\u70ba\u95dc\u9375\u5b57\uff0c\u6536\u96c6\u6392\u540d\u524d10\u7684 \u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4ee5\u81ea\u52d5\u7684\u65b9\u5f0f\u6a19\u8a18\u6e2c\u8a66\u8cc7\u6599\u5171652\u500b\u53e5\u5b50\uff0c\u5171\u6a19\u8a18\u4e0d\u91cd\u8907\u7684\u81fa\u7063\u5730\u5340\u5730 Recall 0.456 0.958 0.948 0.330 4.3 \u591a\u7a2e\u8a9e\u8a00\u53ca\u8fa8\u8b58\u4e3b\u984c\u4e4b NER \u6548\u80fd 4.3 \u4eba\u5de5\u7522\u751f\u95dc\u9375\u8a5e\u5eab\u4e4b NER \u6548\u80fd Precision 0.911 0.938 0.997 0.951 \u6211\u5011\u900f\u904ei\u30bf\u30a6\u30f3\u30da\u30fc\u30b8 3 \u9019\u500b\u65e5\u672c\u9ec3\u9801\u7db2\u7ad9\u6536\u96c6\u4e8610,000\u7b46\u65e5\u6587\u5546\u5bb6\u540d\u7a31\uff0c\u53d6\u6bcf\u7b46\u641c\u5c0b \u6a19\u8a18\u8cc7\u6599 (U) \u4f7f\u75286,650\u7b46\u5730\u5740\u9032\u884c\u67e5\u8a62\uff0c\u53d6\u6bcf\u7b46\u641c\u5c0b\u6392\u540d\u524d10\u500b\u7d50\u679c\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c F-measure 0.972 Labeled Entity 98,154 68,701 136,366 21,370 0.948 \u7247\u6bb5\u4e2d\u6240\u6709\u53e5\u5b50\u9032\u884cAlignmentLabeling\u6bd4\u5c0d\u4e26\u642d\u914dUniLabeling\u7522\u751f\u8a13\u7df4\u8cc7\u6599(L)\u3002\u672a \u5716\u56db\u3001\u4f7f\u7528\u4e0d\u540c\u5927\u5c0f\u81ea\u52d5\u95dc\u9375\u8a5e\u5eab\u6bd4\u8f03\u6548\u80fd Recall 0.948 0.958 Type Chinese address English address Chinese address English address \u641c\u5c0b\u95dc\u9375\u5b57\uff0c\u6bcf\u6b21\u53d6Google\u641c\u5c0b\u6392\u540d\u524d5\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4e26\u4ee5\u5df2\u77e5\u7684\u5730\u540d\u5c0d\u641c\u5c0b\u7d50\u679c Precision 0.997 0.938 Alignment + UniLabeling ExactMatch + FullLabeling \u70ba\u4e86\u77ad\u89e3\u9577\u5ea6\u8f03\u9577\u7684\u82f1\u6587\u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u6548\u679c\uff0c\u6211\u5011\u900f\u904eYelp\u6536\u96c6\u4e862,400\u7b46\u7f8e\u570b\u5730\u5740\u70ba \u5de5\u6a19\u8a18\u7b54\u6848\u9032\u884cNER\u6548\u80fd\u8a55\u4f30\u3002 Chinese address English address \u8868\u5341\u3001\u9577\u547d\u540d\u5be6\u9ad4\u4f7f\u7528 Alignment + UniLabeling \u53ca ExactMatch + FullLabeling \u4e4b\u6548\u80fd \u6210\u7acb\u3002Chou \u7b49\u4eba[2]\u7684\u6539\u826f\u6f14\u7b97\u6cd5\u4f7f\u5f97 Tri-Training \u53ef\u9069\u7528\u65bc\u8f03\u5927\u7684\u8cc7\u6599\u96c6\uff0c\u907f\u514d\u539f\u59cb ID \u8aaa\u660e \u9577 \u7bc4\u4f8b \u2026 \u2026 \u2026 \u2026 4 POI \u4e2d\u5e38\u898b\u524d\u65b9\u5b57 1 \u4ee3\u3001\u8336 5 POI \u4e2d\u5e38\u898b\u524d\u65b9\u8a5e 2 \u4e8b\u52d9\u3001\u6578\u4f4d 6 POI \u4e2d\u5e38\u898b\u524d\u65b9\u8a5e 3 \u591a\u5a92\u9ad4\u3001\u661f\u5df4\u514b 7 POI \u4e2d\u5e38\u898b\u5012\u6578\u5b57 1 \u5edf\u3001\u838a\u3001\u5e97 8 POI \u4e2d\u5e38\u898b\u5012\u6578\u8a5e 2 \u9580\u5e02\u3001\u516c\u53f8 9 POI \u4e2d\u5e38\u898b\u5012\u6578\u8a5e 3 \u57fa\u91d1\u6703\u3001\u96dc\u8ca8\u5e97 10 \u5e38\u898b\u65bc POI \u524d\u65b9\u7684\u5b57 1 \u5230\u3001\u7684 11 \u5e38\u898b\u65bc POI \u524d\u65b9\u7684\u8a5e 2 \u63a8\u85a6\u3001\u52a0\u76df 12 \u5e38\u898b\u65bc POI \u524d\u65b9\u7684\u8a5e 3 \u540d\u7a31\uff1a\u3001\u5e97\u4ecb\u7d39 13 \u5e38\u898b\u65bc POI \u5f8c\u65b9\u7684\u5b57 1 \u901b\u3001\u662f 14 \u5e38\u898b\u65bc POI \u5f8c\u65b9\u7684\u8a5e 2 \u7d71\u7de8\u3001\u71df\u696d 15 \u5e38\u898b\u65bc POI \u5f8c\u65b9\u7684\u8a5e 3 \u9ad8\u54c1\u8cea\u3001\u71df\u696d\u9805 \u2026 \u2026 \u2026 \u2026 \u5929\u9580\u5e02) \u300d\u53ea\u6709\u300c7-ELEVEN\u300d\u4e5f\u4e0d\u80fd\u7b97\u932f\uff0c\u56e0\u6b64\u5c0d\u65bc\u6bcf\u500b\u8fa8\u8b58\u5230\u7684\u547d\u540d\u5be6\u9ad4 e \u8207\u6b63\u78ba \u7b54\u6848\u7684\u547d\u540d\u5be6\u9ad4 a\uff0c\u6211\u5011\u5b9a\u7fa9 P(e,a) \u3001R(e,a)\u5206\u6578\uff0c\u518d\u53d6\u5e73\u5747\u503c\u5f97\u5230\u6574\u9ad4\u7684 Precision\u3001Recall \u3002 \u5176\u5b9a\u7fa9\u5982\u4e0b\uff1a \uf0d8 ( , ) = | \uf0c7 | | | \uf0d8 ( , ) = | \uf0c7 | | | \uf0d8 = \u2211 ( , ) | | \uf0d8 = \u2211 ( , ) | | \uf0d8 \u2212 = 2 + \u4f9d\u7167\u4e0a\u8ff0\u7684\u8a55\u5206\u516c\u5f0f\uff0c\u5229\u7528\u6a21\u578b\u6a19\u8a18\u51fa\u4f86\u7684\u7b54\u6848(Identified entity)\u8207\u6b63\u78ba\u7b54\u6848(Real entity)\u9593\u91cd\u758a\u7684\u5b57\u6578(Overlap tokens) \uff0c\u5206\u5225\u9664\u4ee5\u6a19\u8a18\u7b54\u6848\u9577\u5ea6\u548c\u6b63\u78ba\u7b54\u6848\u9577\u5ea6\u4f86\u7d66\u4e88 \u90e8\u4efd\u6b63\u78ba\u7684\u6a19\u8a18\u5206\u6578\uff0c\u6b64\u65b9\u6cd5\u53ef\u4ee5\u907f\u514d\u56e0\u70ba\u4e00\u5169\u500b\u5b57\u7684\u8aa4\u5dee\u800c\u5c0e\u81f4\u5b8c\u5168\u6c92\u6709\u5206\u6578\u7684\u72c0 \u6cc1\u3002 4.1 \u5be6\u9a57\u8cc7\u6599\u96c6 \u6211\u5011\u6e2c\u8a66\u4e0d\u540c\u8a9e\u8a00\u4ee5\u53ca\u4e0d\u540c\u8fa8\u8b58\u4e3b\u984c\u7684Web NER\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\uff0c\u5404\u500b\u8cc7\u6599\u96c6\u5982\u8868\u4e94\u3002 \uf06c \u4e2d\u6587\u5546\u5bb6\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58 \u6211\u5011\u900f\u904e\u4e2d\u83ef\u9ec3\u9801 2 \u6536\u96c6\u768411,138\u7b46\u5546\u5bb6\u540d\u7a31\uff0c\u900f\u904eGoogle\u641c\u5c0b\u5f15\u64ce\u9032\u884c\u67e5\u8a62\uff0c\u53d6\u6bcf\u7b46\u641c \u5c0b\u524d5\u500b\u7d50\u679c\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4e26\u4ee5\u5df2\u77e5\u7684\u5546\u5bb6\u540d\u7a31\u5c0d\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u6240\u6709\u53e5\u5b50\u9032\u884c\u5b8c \u5168\u76f8\u914d\u7684FullLabeling\u6a19\u8a18\u7522\u751f\u5df2\u6a19\u8a18\u8a13\u7df4\u8cc7\u6599 (L) \u3002\u672a\u6a19\u8a18\u8a13\u7df4\u8cc7\u6599 (U) \u5247\u4f7f\u752850,000 \u7b46\u5546\u5bb6\u9032\u884c\u67e5\u8a62\uff0c\u53d6\u6bcf\u7b46\u641c\u5c0b\u6392\u540d\u524d10\u500b\u7d50\u679c\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u5171\u63d0\u53d6156,822\u500b\u53e5\u5b50\u3002 \u5de5\u7684\u65b9\u5f0f\u6a19\u8a1838,692\u500b\u53e5\u5b50\uff0c\u6a19\u8a18\u51fa\u4e0d\u91cd\u8907\u7684\u5546\u5bb6\u7d44\u7e54\u540d\u7a31\u517116,241\u500b\uff0c\u6700\u5f8c\u4f7f\u7528\u6b64\u4eba \uf06c \u82f1\u6587\u5730\u5740\u8fa8\u8b58 \u8868\u4e03\u3001\u9577\u547d\u540d\u5be6\u9ad4\u4e4b\u8fa8\u8b58\u6548\u80fd\u63a1\u7528\u81ea\u52d5\u7522\u751f\u4e4b\u95dc\u9375\u8a5e\u5eab \u6e2c\u8a66\u8cc7\u6599\u5247\u4ee5\u53e6\u59162,000\u7b46\u5730\u5740\u70ba\u95dc\u9375\u5b57\uff0c\u6536\u96c6\u6392\u540d\u524d10\u500b\u7d50\u679c\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4ee5\u4eba \uf06c \u82f1\u6587\u5546\u5bb6\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58 \u6211\u5011\u900f\u904eYelp 4 \u6536\u96c6\u768410,000\u7b46\u5546\u5bb6\u540d\u7a31\uff0c\u900f\u904eGoogle\u641c\u5c0b\u5f15\u64ce\u53d6\u5f97\u9032\u884c\u67e5\u8a62\uff0c\u53d6\u6bcf\u7b46\u641c Item Chinese Organization Name Japanese Organization Name English Organization Name Chinese Location Name Chinese address English address Source \u4e2d\u83ef\u9ec4\u9801 i \u30bf\u30a6\u30f3\u30da\u30fc\u30b8 Yelp OpenData \u4e2d\u83ef\u9ec4\u9801 Yelp Precision\u8207F-measure\u964d\u4f4e\uff0c\u4f46\u537b\u80fd\u5920\u7dad\u6301Recall\u6c34\u6e96\u751a\u81f3\u5fae\u5e45\u63d0\u5347\u3002\u800c\u8fa8\u8b58\u6548\u679c\u964d\u4f4e\u4e3b \u5716\u516d\u986f\u793a\u82f1\u6587\u53ca\u4e2d\u6587\u5730\u5740\u7684\u6a19\u8a18\u6548\u679c\u3002\u7576\u4f7f\u7528ExactMatchLabeling\u6bd4\u5c0d\u642d\u914dUniLabeling \u8a13\u7df4\u4e00\u500b\u6a21\u578b\u7684\u6642\u9593\u548c\u4eba\u529b\u6210\u672c\u975e\u5e38\u7684\u9ad8\uff0c\u5305\u542b\u524d\u7f6e\u7684\u5927\u91cf\u8a13\u7df4\u8cc7\u6599\u6e96\u5099\u3001\u4eba\u5de5\u6536\u96c6\u7b54 \u77ed\u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u6548\u80fd\u5982\u8868\u516d\uff0c\u76f8\u8f03\u65bc\u4e2d\u6587\u53ca\u82f1\u6587\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58\u6548\u80fd\uff0c\u65e5\u6587\u7684\u7d44\u7e54\u540d\u7a31\u8fa8 \u8981\u539f\u56e0\u53ef\u80fd\u5728\u65bc\u5546\u5bb6\u7d44\u7e54\u540d\u7a31\u5c6c\u65bc\u8b8a\u7570\u6027\u8f03\u5927\u7684\u4e00\u7a2e\u547d\u540d\u5be6\u9ad4\uff0c\u8cc7\u6599\u80fd\u5426\u76e1\u53ef\u80fd\u7684\u6db5\u84cb \u53caFullLabeling\u7522\u751f\u8a13\u7df4\u8cc7\u6599\u6642\uff0c\u50c5\u53ef\u5f9e86,388\u7b46\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u6a19\u8a18\u51fa21,370\u53ca22,313 \u6848\u3001\u6a19\u8a18\u7b54\u6848\uff0c\u70ba\u4e86\u63d0\u5347\u6a21\u7d44\u8fa8\u8b58\u6548\u679c\u800c\u5fc5\u9808\u8981\u70ba\u8cc7\u6599\u505a\u9069\u7576\u512a\u5316\uff0c\u4ee5\u53ca\u7279\u5fb5\u503c\u7684\u8a2d\u8a08\u3001 \u8b58\u7684F-measure\u7a0d\u4f4e\uff0c\u6211\u5011\u731c\u6e2c\u5176\u539f\u56e0\u53ef\u80fd\u5728\u65bc\u65e5\u6587\u5c6c\u65bc\u97f3\u7bc0\u6587\u5b57(Syllabary)\u662f\u8868\u97f3\u6587 \u5404\u985e\u5546\u5bb6\u7d44\u7e54\u540d\u7a31\u7684\u7279\u6027\u662f\u91cd\u8981\u56e0\u7d20\uff0c\u800c\u81ea\u52d5\u7522\u751f\u7684\u95dc\u9375\u8a5e\u5eab\u76f8\u5c0d\u65bc\u4eba\u5de5\u8a2d\u8a08\u7684\u95dc\u9375\u8a5e \u500b\u82f1\u6587\u5730\u5740\u3002\u4f46\u7576\u4f7f\u7528AlignmentLabeling\u6bd4\u5c0d\u642d\u914dUniLabeling\u7522\u751f\u8a13\u7df4\u8cc7\u6599\u6642\uff0c\u5171\u53ef\u6a19 \u95dc\u9375\u8a5e\u5eab\u6e96\u5099\u7b49\uff0c\u5de5\u4f5c\u975e\u5e38\u7463\u788e\u8907\u96dc\uff0c\u4e14\u5c0d\u65bc\u4e0d\u540c\u8a9e\u8a00\u6216\u4e0d\u540c\u8fa8\u8b58\u4e3b\u984c\u90fd\u8981\u518d\u91cd\u65b0\u8a2d\u8a08 \u5b57\u7684\u4e00\u7a2e\uff0c\u9664\u4e86\u90e8\u5206\u4f7f\u7528\u6f22\u5b57\u5916\u5927\u90e8\u5206\u4f7f\u7528\u5e73\u5047\u540d\u6216\u7247\u5047\u540d\u66f8\u5beb\uff0c\u7576\u5728\u81ea\u52d5\u64f7\u53d6\u5916\u90e8\u8207 \u5eab\u5305\u542b\u8f03\u591a\u7684\u96dc\u8a0a\uff0c\u4e14\u6703\u6709\u5b8c\u5168\u91dd\u5c0d\u8f38\u5165\u7684\u8a13\u7df4\u8cc7\u6599\u8a2d\u8a08\u7b49\u554f\u984c\uff0c\u4f46\u7576\u8a13\u7df4\u8cc7\u6599\u91cf\u5920\u5927 \u8a18\u51fa68,701\u500b\u82f1\u6587\u5730\u5740\u3002\u53e6\u5916\uff0c\u7576\u4f7f\u7528ExactMatchLabeling\u6bd4\u5c0d\u642d\u914dUniLabeling\u7522\u751f\u4e2d \u7279\u5fb5\u503c\u3002\u672c\u7814\u7a76\u671f\u80fd\u8a2d\u8a08\u4e00\u500b\u4f7f\u7528Google\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e4bWeb NER\u8fa8\u8b58\u6a21\u578b\u7684\u7522\u751f\u5de5\u5177\uff0c \u5167\u90e8\u7279\u5fb5\u6642\u5c31\u6703\u9047\u5230\u50c5\u53d6\u5230\u90e8\u4efd\u62fc\u97f3\u800c\u4e0d\u5177\u6709\u610f\u7fa9\u7684\u554f\u984c\u3002 \u4e14\u5546\u5bb6\u985e\u5225\u591a\u6a23\u5316\u6642\u8fa8\u8b58\u61c9\u80fd\u518d\u63d0\u5347\u3002 \u6587\u5730\u5740\u8a13\u7df4\u8cc7\u6599\u6642\uff0c\u5df2\u7d93\u53ef\u5f9e108,435\u7b46\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u6a19\u8a18\u51fa95,558\u500b\u4e2d\u6587\u5730\u5740\uff0c\u82e5\u662f \u4e0d\u50c5\u89e3\u6c7a\u4e0a\u8ff0\u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u904e\u65bc\u8017\u6642\u8cbb\u529b\u7684\u554f\u984c\uff0c\u4e5f\u80fd\u5920\u8f15\u6613\u5730\u61c9\u7528\u5728\u4e0d\u540c\u7684\u8fa8\u8b58\u985e\u578b\u3001 \u5c0b\u524d5\u500b\u7d50\u679c\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4e26\u4ee5\u5df2\u77e5\u7684\u5546\u5bb6\u540d\u7a31\u5c0d\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u6240\u6709\u53e5\u5b50\u9032\u884c\u5b8c \u5168\u76f8\u914d\u7684FullLabeling\u6a19\u8a18\u5373\u70ba\u5df2\u6a19\u8a18\u8a13\u7df4\u8cc7\u6599 (L) \u3002\u672a\u6a19\u8a18\u8a13\u7df4\u8cc7\u6599 (U) \u5247\u4f7f\u752830,000 \u7b46\u5546\u5bb6\u9032\u884c\u67e5\u8a62\uff0c\u53d6\u6bcf\u7b46\u641c\u5c0b\u6392\u540d\u524d10\u500b\u7d50\u679c\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u5171\u63d0\u53d6100,182\u500b\u53e5\u5b50\u3002 \u6e2c\u8a66\u8cc7\u6599\u5247\u4ee5\u53e6\u5916200\u7b46\u5730\u5740\u70ba\u95dc\u9375\u5b57\uff0c\u6536\u96c6\u6392\u540d\u524d10\u500b\u7d50\u679c\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4ee5\u81ea\u52d5 \u7684\u65b9\u5f0f\u6a19\u8a18941\u500b\u53e5\u5b50\uff0c\u6a19\u8a18\u51fa\u4e0d\u91cd\u8907\u7684\u5546\u5bb6\u7d44\u7e54\u540d\u7a31\u5171465\u500b\uff0c\u6700\u5f8c\u4f7f\u7528\u6b64\u81ea\u52d5\u6a19\u8a18\u7b54 \u6848\u9032\u884cNER\u6548\u80fd\u8a55\u4f30\u3002 \uf06c \u4e2d\u6587\u5730\u9ede\u540d\u7a31\u8fa8\u8b58 \u70ba\u4e86\u77ad\u89e3\u672c\u5de5\u5177\u8fa8\u8b58\u4e0d\u540c\u985e\u5225\u7684\u80fd\u529b\uff0c\u6211\u5011\u900f\u904e\u653f\u5e9c\u8cc7\u6599\u958b\u653e\u5e73\u53f0 5 \u6536\u96c6\u4e8610,000\u7b46\u81fa\u7063 \u5730\u5340\u5730\u540d\u8cc7\u6599\uff0c\u6bcf\u7b46\u53d6Google\u641c\u5c0b\u6392\u540d\u524d5\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4e26\u4ee5\u5df2\u77e5\u7684\u5730\u540d\u5c0d\u641c\u5c0b\u7d50 Training: L 11,138 10,000 10,000 10,000 1,800 \u6211\u5011\u4e5f\u6ce8\u610f\u5230\u5728\u4e2d\u6587\u5730\u9ede\u540d\u7a31\u8fa8\u8b58\u90e8\u5206\u6709\u5f88\u9ad8\u7684Precision\uff0c\u4f46Recall\u537b\u660e\u986f\u8f03\u4f4e\u3002\u9020\u6210 \u63a1\u7528FullLabeling\uff0c\u66f4\u53ef\u4ee5\u6a19\u8a18\u51fa136,366\u500b\u4e2d\u6587\u5730\u5740\uff1b\u56e0\u6b64\u4f7f\u7528AlignmentLabeling\u6bd4\u5c0d \u8a9e\u8a00\u4e2d\uff0c\u4e26\u5e0c\u671b\u9054\u5230\u826f\u597d\u7684\u8fa8\u8b58\u6548\u679c\u3002 2,400 #Sentence 87,916 29,999 39,798 53,313 28,739 18,198 Training: U 50,000 30,000 30,000 30,000 10,000 6,650 #Sentence 156,822 88,074 100,182 132,486 78,177 49,851 Testing 2,000 addr 200 addr 200 addr 200 loc 200 organ 200 organ #Sentence 38,692 809 941 2,638 1,519 652 \u9019\u500b\u7d50\u679c\u7684\u539f\u56e0\u662f\u6211\u5011\u5c0d\u65bc\u4e2d\u6587\u5730\u9ede\u540d\u7a31\u6709\u8f03\u5ee3\u6cdb\u7684\u5b9a\u7fa9\uff0c\u4f8b\u5982\uff1a\u300c\u9ad8\u96c4\u5e02\u300d\u3001\u300c\u7d2b\u7af9 \u8868\u516b\u3001\u4ee5\u4e2d\u6587\u7d44\u7e54\u8fa8\u8b58\u70ba\u4f8b\u6bd4\u8f03\u7cfb\u7d71\u81ea\u52d5\u7522\u751f\u8a5e\u5eab\u3001\u4eba\u5de5\u7522\u751f\u95dc\u9375\u8a5e\u5eab\u8207 Stanford NER \u642d\u914dUniLabeling \u6a19 \u8a18 \u51fa 98,154 \u500b \u4e2d \u6587\u5730\u5740\uff0c \u672a \u80fd \u52dd \u904e ExactMatchLabeling \u642d\u914d \u5728\u672c\u7cfb\u7d71\u6211\u5011\u4f7f\u7528\u81ea\u52d5\u6a19\u8a18\u7684\u65b9\u5f0f\u6a19\u8a18\u8a13\u7df4\u8cc7\u6599\u800c\u975e\u4f7f\u7528\u4eba\u5de5\u6a19\u8a18\u7b54\u6848\uff0c\u4e26\u4e14\u70ba\u4e86\u6709\u6548 \u5bfa\u300d\u3001\u300c\u5e73\u6797\u91cc\u300d\u3001\u300c\u72d7\u6bcd\u5c71\u300d\u3001\u300c\u6771\u77f3\u5927\u6a4b\u300d\u3001\u300c\u66f9\u516c\u5733\u300d\u3001\u300c\u53f0\u5317\u706b\u8eca\u7ad9\u300d\u2026\u7b49\u3002 \u6548\u80fd\u7684\u5dee\u7570 FullLabeling\u7684\u6548\u679c\u3002 \u6a19\u8a18\u9577\u7684\u547d\u540d\u5be6\u9ad4\u6211\u5011\u53ef\u4ee5\u4f7f\u7528AlignmentLabeling\u589e\u52a0\u6a19\u8a18\u5230\u7684\u547d\u540d\u5be6\u9ad4\u6578\u91cf\u3002\u96d6\u7136\u81ea \u56e0\u6b64\u6211\u5011\u5728\u6a19\u8a18\u6e2c\u8a66\u8cc7\u6599\u7b54\u6848\u6642\u7684\u7b54\u6848\u5b9a\u4ee5\u4e5f\u8f03\u5ee3\u6cdb\uff0c\u4f46\u5be6\u969b\u6a21\u7d44\u5728\u6a19\u8a18\u6642\u96d6\u7136\u80fd\u6709\u9ad8 \u7684\u6e96\u78ba\u7387\uff0c\u4f46\u537b\u7121\u6cd5\u8fa8\u8b58\u6240\u6709\u985e\u578b\u7684\u4e2d\u6587\u5730\u9ede\u540d\u7a31\u3002 \u8868\u516d\u3001\u77ed\u547d\u540d\u5be6\u9ad4\u4e4b\u8fa8\u8b58\u6548\u80fd\u63a1\u7528\u81ea\u52d5\u7522\u751f\u4e4b\u95dc\u9375\u8a5e\u5eab Chinese organization Japanese organization English organization Manual Dictionary Automatic Dictionary \u52d5\u6a19\u8a18\u53ef\u80fd\u5305\u542b\u96dc\u8a0a\uff0c\u4f46\u6211\u5011\u56e0\u800c\u80fd\u7522\u751f\u5927\u91cf\u7684\u5df2\u6a19\u8a18\u8a13\u7df4\u8cc7\u6599\u3002 Stanford Precision 0.8500 0.8249 \u5916\u90e8\u7279\u5fb5\u662f\u9032\u884c\u547d\u540d\u5be6\u9ad4\u8fa8\u8a8d\u7684\u91cd\u8981\u8f14\u52a9\uff0c\u800c\u5167\u90e8\u7279\u5fb5\u80fd\u63d0\u4f9b\u5f37\u70c8\u7684\u5224\u65b7\u8cc7\u8a0a\uff0c\u6211\u5011\u5229 0.529 Recall 0.8730 0.8753 0.557 \u7528\u983b\u7387\u7d71\u8a08\u7684\u65b9\u5f0f\u80fd\u5920\u81ea\u52d5\u7522\u751f\u4e0a\u8ff0\u5169\u7a2e\u7279\u5fb5\uff0c\u4e26\u5229\u7528\u5b8c\u6574\u6a19\u8a18\u5df2\u77e5\u5927\u91cf\u7684\u547d\u540d\u5be6\u9ad4\u8207 Chinese location F-measure 0.8613 0.8494 Self-Testing\u53caTri-Training\u6f14\u7b97\u6cd5\uff0c\u4f7f\u5f97\u8fa8\u8b58\u6548\u80fd\u66f4\u9032\u4e00\u6b65\u63d0\u5347\uff0c\u89e3\u6c7a\u8a13\u7df4\u8cc7\u6599\u54c1\u8cea\u4e0d\u4f73 0.543 \u7684\u554f\u984c\u3002 #Distinct 16,241 438 465 600 645 257 names names names names 4.4 \u4f7f\u7528 Self-Testing \u53ca Tri-Training \u5f8c\u4e4b NER \u6548\u80fd\u63d0\u5347 \u6211\u5011\u4ee5\u4e2d\u6587\u4e4b\u5546\u5bb6\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58\u505a\u6e2c\u8a66\uff0c\u5be6\u9a57\u986f\u793a\u5728 \u4e2d \u6587 \u7d44 \u7e54\u540d \u7a31 \u8fa8 \u8b58 \u90e8 \u4efd \u4ee5 Entities Precision 0.825 0.845 0.789 0.925 Tri-Training\u6f14\u7b97\u6cd5\u78ba\u5be6\u4f7f\u5f97\u8fa8\u8b58\u6548\u80fd\u66f4\u9032\u4e00\u6b65\u63d0\u5347\uff0cF-Measure\u53ef\u7531DS1\u76840.779\u63d0\u5347\u81f3 \u679c\u7247\u6bb5\u4e2d\u6240\u6709\u53e5\u5b50\u9032\u884c\u5b8c\u5168\u76f8\u914d\u7684FullLabeling\u6a19\u8a18\u7522\u751f\u8a13\u7df4\u8cc7\u6599 (L) \u3002\u672a\u6a19\u8a18\u8cc7\u6599 (U) \u4f7f\u752830,000\u7b46\u5730\u540d\u9032\u884c\u67e5\u8a62\uff0c\u53d6\u6bcf\u7b46\u641c\u5c0b\u6392\u540d\u524d10\u500b\u7d50\u679c\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4f9b\u63d0\u53d6 132,486\u500b\u53e5\u5b50\u3002\u6e2c\u8a66\u8cc7\u6599\u53e6\u5916\u53d6200\u7b46\u5730\u540d\u70ba\u95dc\u9375\u5b57\uff0c\u6536\u96c6\u6392\u540d\u524d10\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c \u4ee5\u4eba\u5de5\u7684\u65b9\u5f0f\u6a19\u8a18\u6e2c\u8a66\u8cc7\u6599\u51712,638\u500b\u53e5\u5b50\uff0c\u5171\u6a19\u8a18\u4e0d\u91cd\u8907\u7684\u81fa\u7063\u5730\u5340\u5730\u540d600\u500b\u3002 \u70ba\u4e86\u77ad\u89e3\u9577\u5ea6\u8f03\u9577\u7684\u4e2d\u6587\u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u6548\u679c\uff0c\u6211\u5011\u900f\u904e\u4e2d\u83ef\u9ec3\u9801\u6536\u96c6\u4e861,800\u7b46\u81fa\u7063\u5730 \u5740\u70ba\u641c\u5c0b\u95dc\u9375\u5b57\uff0c\u6bcf\u6b21\u53d6Google\u641c\u5c0b\u6392\u540d\u524d5\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4e26\u4ee5\u5df2\u77e5\u7684\u5730\u540d\u5c0d\u641c\u5c0b \u7d50\u679c\u7247\u6bb5\u4e2d\u6240\u6709\u53e5\u5b50\u9032\u884c\u5b8c\u5168\u76f8\u914d\u7684FullLabeling\u6a19\u8a18\u7522\u751f\u8a13\u7df4\u8cc7\u6599(L)\u3002\u672a\u6a19\u8a18\u8cc7\u6599 (U)\u4f7f\u752810,000\u7b46\u5730\u5740\u9032\u884c\u67e5\u8a62\uff0c\u53d6\u6bcf\u7b46\u641c\u5c0b\u6392\u540d\u524d10\u500b\u7d50\u679c\u7684\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4f9b\u63d0 \u53d678,177\u500b\u53e5\u5b50\u3002\u6e2c\u8a66\u8cc7\u6599\u53e6\u5916\u53d6200\u7b46\u4e2d\u6587\u5546\u5bb6\u7d44\u7e54\u540d\u7a31\u70ba\u95dc\u9375\u5b57\uff0c\u6536\u96c6\u6392\u540d\u524d10\u7684 \u641c\u5c0b\u7d50\u679c\u7247\u6bb5\uff0c\u4ee5\u81ea\u52d5\u7684\u65b9\u5f0f\u6a19\u8a18\u6e2c\u8a66\u8cc7\u6599\u51711,519\u500b\u53e5\u5b50\uff0c\u5171\u6a19\u8a18\u4e0d\u91cd\u8907\u7684\u81fa\u7063\u5730\u5340 \u5730\u5740645\u500b\u3002 \u5404\u5225\u95dc\u9375\u8a5e\u5eab\u5927\u5c0f\u4e4b\u6548\u80fd\u5982\u5716\u56db\u3001\u4f7f\u7528\u4e0d\u540c\u5927\u5c0f\u81ea\u52d5\u95dc\u9375\u8a5e\u5eab\u6bd4\u8f03\u6548\u80fd\uff0c\u6bd4\u8f03\u5404\u8cc7\u6599\u96c6 \u6211\u5011\u53ef\u4ee5\u767c\u73fe\u7576\u4f7f\u7528\u5927\u91cf\u5b57\u6216\u8a5e\u7684\u95dc\u9375\u8a5e\u5eab\u5c07\u5c0e\u81f4Recall\u5927\u5e45\u964d\u4f4e\u3002 \u5b57\u8207\u5b57\u4e4b\u9593\u9806\u5e8f\u4e00\u81f4\uff0c\u4f46\u4e26\u4e0d\u80fd\u4fdd\u8b49\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u5167\u5bb9\u4e2d\u7684\u547d\u540d\u5be6\u9ad4\u8207\u641c\u5c0b\u8a5e\u5b8c\u5168\u76f8\u540c\uff0c \u5740\uff0c\u6211\u5011\u4f7f\u7528AlignmentLabeling\u6bd4\u5c0d\u4e26\u642d\u914dUniLabeling\u4f86\u6a19\u8a18\u67e5\u8a62\u8a5e\u6240\u5728\u3002\u4f46\u7531\u4e2d\u6587\u5730 \u5740 \u5728\u55ae\u4e00\u547d\u540d\u5be6\u9ad4\u4e2d\u4e0d\u6703\u96a8\u610f\u7684\u63d2\u5165\u6a19\u9ede\u7b26\u865f\u8207\u7e2e\u5beb\uff0c\u56e0\u6b64\u6211\u5011\u53ef\u4f7f\u7528 ExactMatchLabeling\u6bd4\u5c0d\u4e26\u642d\u914dFullLabeling\u6b63\u78ba\u6a19\u8a18\u51fa\u4e2d\u6587\u5730\u5740\u3002\u6211\u5011\u5c07\u6703\u5728\u5f8c\u7e8c\u5be6\u9a57 \u4e2d\u6bd4\u8f03AlignmentLabeling\u8207ExactMatchLabeling\u4e4b\u6a19\u8a18\u6548\u679c\u3002\u5c0d\u65bc\u9577\u547d\u540d\u5be6\u9ad4\u5982\u82f1\u6587\u5730 \u5740\u53ca\u4e2d\u6587\u5730\u5740\u4e4b\u8fa8\u8b58\u6548\u80fd\u898b\u8868\u4e03\u3002 Query 1,000 3,000 4,000 6,000 11,138 50,000 Sentence 6,724 19,437 27,198 45,028 87,916 156,822 \u5728\u8868\u5341\u4e2d\u6211\u5011\u6bd4\u8f03\u4e86 F-measure(0.972)\u3002 ExactMatchLabeling\u6a19\u8a18\u51fa\u9577\u7684\u62fc\u97f3\u6587\u5b57\u547d\u540d\u5be6\u9ad4\u662f\u4e0d\u5bb9\u6613\u7684\u3002\u56e0\u6b64\u70ba\u4e86\u6a19\u8a18\u51fa\u82f1\u6587\u5730 DS1 DS2 DS3 DS4 DS5 Unlabeled \u7576\u4f5c\u662f\u76ee\u6a19\u7d66\u6a19\u8a18\u8d77\u4f86\uff0c\u6b64\u7a2e\u932f\u8aa4\u6703\u5c0e\u81f4\u8f03\u4f4e\u7684\u6e96\u78ba\u7387\u3002 \u56e0\u6b64\u641c \u5c0b \u7d50 \u679c \u7247 \u6bb5 \u4e2d \u7684 \u641c \u5c0b \u8a5e \u4e2d \u53ef \u80fd \u7a7f \u63d2 \u4e0d \u540c \u6a19 \u9ede \u7b26 \u865f \u3002 \u9019\u5c0d\u65bc\u4f7f\u7528 \u5982\uff0c\u300c\u5f70\u5316\u7e23\u9e7f\u6e2f\u5e02\u5834169\u865f\u300d\u4e26\u975e\u662f\u5408\u6cd5\u7684\u53f0\u7063\u5730\u5740\uff0c\u4f46\u5728AlignmentLabeling\u4ecd\u6703\u88ab \u8868\u4e5d\u3001\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58\u4e4b\u5df2\u6a19\u8a18\u8a13\u7df4\u8cc7\u6599(DS1~DS5)\u53ca\u672a\u6a19\u8a18\u8a13\u7df4\u8cc7\u6599(U) AlignmentLabeling\u5bb9\u6613\u6703\u5728\u4e2d\u6587\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u6a19\u8a18\u51fa\u985e\u4f3c\u65bc\u4e2d\u6587\u5730\u5740\u7684\u547d\u540d\u5be6\u9ad4\u3002\u4f8b \u672c\u7cfb\u7d71\u5c0dGoogle\u641c\u5c0b\u6642\u96d6\u4f7f\u7528\u96d9\u5f15\u865f\uff0c\u78ba\u4fdd\u641c\u5c0b\u7d50\u679c\u7247\u6bb5\u4e2d\u7684\u9577\u547d\u540d\u5be6\u9ad4\u6703\u8207\u67e5\u8a62\u8a5e\u4e4b ExactMatchLabeling \u6a19 \u8a18 \u51fa \u5927 \u91cf \u7684 \u9577 \u547d \u540d \u5be6 \u9ad4 \u3002 \u9664 \u6b64 \u4e4b \u5916 \uff0c \u6211 \u5011 \u4e5f \u767c \u73fe \u4f7f \u7528 \uf06c \u4e2d\u6587\u5730\u5740\u8fa8\u8b58 Recall 0.875 0.766 0.881 0.777 \u672c\u5be6\u9a57\u65e8\u5728\u4e86\u89e3\u4f7f\u7528Self-Testing\u4ee5\u53caTri-Training\u7522\u751f\u4e4b\u65b0\u8fa8\u8b58\u6a21\u578b\u5c0dGoogle\u641c\u5c0b\u7d50\u679c DS5\u76840.861\uff0c\u800c\u5728\u65e5\u6587\u7d44\u7e54\u540d\u7a31\u3001\u800c\u5728\u82f1\u6587\u7d44\u7e54\u540d\u7a31\u3001\u4e2d\u6587\u5730\u9ede\u540d\u7a31\u3001\u4e2d\u6587\u5730\u5740\u4ee5\u53ca\u82f1 4.2 \u4f7f\u7528\u4e0d\u540c\u5927\u5c0f\u81ea\u52d5\u95dc\u9375\u8a5e\u5eab\u6bd4\u8f03\u6548\u80fd \u4ee5\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u70ba\u4f8b\uff0c\u6211\u5011\u5206\u5225\u4f7f\u752850\u3001100\u3001150\u3001200\u500b\u5b57\u6216\u8a5e\u7684\u81ea\u52d5\u7522\u751f\u5167\u90e8\u7279\u5fb5 \u4ee5\u53ca\u5916\u90e8\u7279\u5fb5\u5efa\u7acb\u95dc\u9375\u8a5e\u5eab\uff0c\u4f7f\u7528Self-testing\u5c07\u96dc\u8a0a\u79fb\u9664\u63d0\u9ad8\u8a13\u7df4\u8cc7\u6599\u7684\u54c1\u8cea\uff0c\u5be6\u9a57\u4e2d \u5047\u8a2d\u4f4e\u65bc0.7\u70ba\u96dc\u8a0a\u5c07\u5176\u53bb\u9664\uff0c\u4e26\u4ee5Self-testing\u5f8c\u4e4b\u8cc7\u6599\u70ba\u57fa\u790e\u9032\u884cTri-Training\u6f14\u7b97\u6cd5\u3002 F-measure 0.849 0.803 0.832 \u5927\u5c0f\u8fa8\u8b58\u6548\u679c\u7686\u6709\u63d0\u5347\uff0c\u5c0d\u65bcDS5\u63d0\u5347\u5e45\u5ea6\u70ba4.83%\uff0c\u75310.8130\u9054\u52300.8613\u3002 \u4e0d\u540c\u65bc\u82f1\u6587\uff0c\u901a\u5e38\u4e2d\u6587\u4e26\u4e0d\u6703\u5728\u55ae\u4e00\u547d\u540d\u6642\u9ad4\u4e2d\u52a0\u5165\u6a19\u9ede\u7b26\u865f\uff0c\u56e0\u6b64\u6211\u5011\u53ef\u4ee5\u5229\u7528 \uf06c \u9577\u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\u6548\u80fd \u53ef\u4ee5\u770b\u5230\u5229\u7528\u63a1\u7528\u4eba\u5de5\u7522\u751f\u95dc\u9375\u8a5e\u5eab\u65b9\u5f0f\u5728Self-Testing\u4ee5\u53caTri-Training\u7684\u5404\u500b\u8cc7\u6599\u96c6 \u5716\u516d\u3001AlignmentLabeling \u8207 ExactMatchLabeling \u4e4b\u6a19\u8a18\u6548\u80fd \u4e5d\u3002\u5728\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58\u4eba\u5de5\u7522\u751f\u95dc\u9375\u8a5e\u5eab\u7684Self-Testing\u53caTri-Training\u5be6\u9a57\u4e2d\uff0c\u7531\u5716\u4e94 0.845 \u7247\u6bb5NER\u6548\u679c\u7684\u5f71\u97ff\u3002\u6211\u5011\u4ee5\u4e2d\u6587\u7d44\u7e54\u540d\u7a31\u8fa8\u8b58\u70ba\u4f8b\uff0c\u5c07\u8a13\u7df4\u8cc7\u6599\u5206\u70ba\u4e94\u500b\u8cc7\u6599\u96c6\u5982\u8868 \u6587\u5730\u5740\u7684F-
", "type_str": "table", "text": "\u95dc\u9375\u8a5e\uff1a\u547d\u540d\u5be6\u9ad4\u8fa8\u8b58\uff0c\u5354\u540c\u8a13\u7df4\uff0cTri-Training Keywords: Named Entity Recognition, Co-Training, Tri-Training. \u7b46\u5df2\u77e5\u5730\u5740\u70ba\u67e5\u8a62\u95dc\u9375\u5b57\uff0c\u65bc Google \u641c\u5c0b\u7d50\u679c\u7247\u6bb5(Search Snippets)\u4e2d\u5305\u542b\u95dc \u9375\u5b57\u7684\u53e5\u5b50\u70ba\u6e2c\u8a66\u8cc7\u6599\uff0c\u518d\u4f7f\u7528 Stanford NER 1 (Named Entity Recognizer) \u4f86\u505a\u7d44\u7e54\u540d \u7a31\u8fa8\u8b58\u5be6\u9a57\uff0cF1 \u6548\u679c\u53ea\u80fd\u9054\u5230 54.3%\u3002\u53e6\u5916\uff0c\u6211\u5011\u4e5f\u5229\u7528 200 \u7b46\u4e2d\u6587\u5730\u9ede\u540d\u7a31\u70ba\u67e5\u8a62 \u95dc\u9375\u5b57\uff0c\u5229\u7528 Google search snippets \u5305\u542b\u95dc\u9375\u5b57\u7684\u53e5\u5b50\u70ba\u6e2c\u8a66\u8cc7\u6599\uff0c\u540c\u6a23\u5229\u7528 Stanford NER \u4f86\u505a\u5730\u9ede\u540d\u7a31\u8fa8\u8b58\u5be6\u9a57\uff0cF1 \u6548\u679c\u50c5\u9054 20.1%\u3002\u986f\u793a\u73fe\u6709\u7684\u516c\u958b NER \u5de5\u5177\u5c0d\u65bc Web Alignment \u642d\u914d UniLabeling \u4ee5\u53ca Exact Match \u642d\u914d FullLabeling \u5c0d\u65bc\u4e2d\u6587\u5730\u5740\u53ca\u82f1\u6587\u5730\u5740\u7684\u8fa8\u8b58\u5f71\u97ff\u3002\u6211\u5011\u53ef\u4ee5\u5f9e\u8868\u5341\u4e2d\u770b\u51fa\u5c0d\u65bc\u82f1\u6587\u5730\u5740\u8fa8\u8b58\u4f7f\u7528 Alignment \u642d\u914d UniLabeling \u53ef\u4ee5\u5f97\u5230\u8f03\u597d\u7684 Recall \u4ee5\u53ca F-measure(0.948)\uff1b\u7136\u800c\u5728\u4e2d\u6587 \u5730\u5740\u8fa8\u8b58\uff0c\u4f7f\u7528 Exact Match \u642d\u914d FullLabeling \u53ef \u4ee5 \u5f97 \u5230 \u8f03 \u597d \u7684 Recall \u4ee5\u53ca Measure\u8fa8\u8b58\u6548\u679c\u4f9d\u5e8f\u53ef\u905480.3%, 83.2%, 84.5%, 97.2% \u53ca 94.8%\u3002", "html": null } } } }