{
"paper_id": "O08-1013",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T08:02:16.880442Z"
},
"title": "\u4e2d\u6587\u540d\u8a5e\u7d44\u7684\u8fa8\u8b58\uff1a\u76e3\u7763\u5f0f\u8207\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u7684\u5be6\u9a57 Chinese NP Chunking: Experiments with Supervised,and Semisupervised Learning",
"authors": [
{
"first": "Yen",
"middle": [
"Hsi"
],
"last": "\u6797\u664f\u50d6",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "",
"middle": [],
"last": "Lin",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "\u9ad8\u7167\u660e",
"middle": [],
"last": "Zhao",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Ming",
"middle": [],
"last": "Gao",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Yan",
"middle": [],
"last": "Cheng",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "",
"middle": [],
"last": "Kao",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper utilizes Yamcha, a SVM tool designed by Taku Kudo, to train an NP-chunking model for Chinese. In addition to IOB and two words surrounding the focused word, we experimented on new features and exploited unlabeled data from web pages to enhance the previous model. Our experiments with supervised learning indicate that our chosen feature sets outperform those reported in previous studies. In addition, the proposed method of semisupervised learning is proved to be effective in distinguishing a noun phrase from a verb phrase both consisting of V N combination, thus enhancing the overall accuracy. \u95dc\u9375\u8a5e\uff1a\u540d\u8a5e\u7d44\u8fa8\u8b58\u3001YamCha\u3001\u76e3\u7763\u5f0f\u5b78\u7fd2\u3001\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2 Keywords\uff1aNP-chunking\u3001YamCha\u3001supervised learning\u3001semi-supervised learning \u4e00\u3001\u7dd2\u8ad6 \u540d\u8a5e\u7d44\u7684\u8fa8\u8b58\u4e00\u76f4\u4ee5\u4f86\u90fd\u662f\u81ea\u7136\u8a9e\u8a00\u7814\u7a76\u53ca\u5176\u76f8\u95dc\u9818\u57df\uff0c\u5982\u7db2\u8def\u63a2\u52d8(web mining)\u3001\u6587\u4ef6\u5206\u985e(text categorization)\u7b49\u975e\u5e38\u95dc\u9375\u7684\u4e00\u500b\u6b65\u9a5f\u3002\u5728\u81ea\u52d5\u554f\u7b54\u7cfb\u7d71 (question answering)\u4e2d\uff0c\u95dc\u9375\u8a5e\u591a\u534a\u4ee5\u540d\u8a5e\u641c\u5c0b\u70ba\u4e3b\u3002\u5728\u81ea\u7136\u8a9e\u8a00\u554f\u7b54\u7cfb\u7d71\u88e1\uff0c\u540d \u8a5e\u7d44\u7684\u8fa8\u8b58\u4e5f\u662f\u4e0d\u53ef\u6216\u7f3a\u3002\u6211\u5011\u6bcf\u5929\u90fd\u4f7f\u7528\u7684\u641c\u5c0b\u5f15\u64ce\uff0c\u5927\u5bb6\u8f38\u5165\u7684\u95dc\u9375\u8a5e\u4ee5\u53ca\u641c\u5c0b \u5f15\u64ce\u7d71\u8a08\u51fa\u4f86\u7684\u71b1\u9580\u95dc\u9375\u8a5e\u4ea6\u4ee5\u540d\u8a5e\u7d44\u5c45\u591a\uff1b\u5927\u81f3\u641c\u5c0b\u5f15\u64ce\u80cc\u5f8c\u5927\u91cf\u7684\u8cc7\u6599\u5eab\u3001\u5c0f\u81f3 \u666e\u901a\u6587\u672c\u5efa\u6a94\u800c\u6210\u7684\u8cc7\u6599\u5eab\uff0c\u88fd\u4f5c\u7d22\u5f15\u5206\u985e\u6642\uff0c\u540d\u8a5e\u7d44\u7684\u4f7f\u7528\u4e5f\u591a\u65bc\u52d5\u8a5e\u3001\u526f\u8a5e\uff1b\u4e5f \u6709\u8d8a\u4f86\u8d8a\u591a\u7db2\u9801\u90fd\u5728\u91dd\u5c0d\u7db2\u9801\u4e2d\u7684\u540d\u8a5e\u7d44\u505a\u81ea\u52d5\u5075\u6e2c\u4ee5\u53ca\u5916\u90e8\u9023\u7d50 ... \u9664\u4e86\u9019\u4e9b\u4f8b\u5b50\u4e4b \u5916\uff0c\u8a9e\u610f\u89d2\u8272\u6a19\u8a18(semantic role labeling)\u3001\u5c08\u6709\u540d\u8a5e\u8fa8\u8b58(name entity identification)\u3001\u6587\u7ae0\u8655\u7406\u4e2d\u7684\u56de\u6307(coreference)\uff0c\u540d\u8a5e\u7d44\u7684\u8fa8\u8b58\u90fd\u662f\u4e00\u500b\u91cd\u8981\u7684\u6b65 \u9a5f\uff0c\u56e0\u6b64\u6709\u597d\u7684 NP chunker\uff0c\u53ef\u4ee5\u6539\u5584\u8a31\u591a NLP \u7684\u7814\u7a76\u6210\u679c\u4ee5\u53ca\u76f8\u95dc\u61c9\u7528\u3002",
"pdf_parse": {
"paper_id": "O08-1013",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper utilizes Yamcha, a SVM tool designed by Taku Kudo, to train an NP-chunking model for Chinese. In addition to IOB and two words surrounding the focused word, we experimented on new features and exploited unlabeled data from web pages to enhance the previous model. Our experiments with supervised learning indicate that our chosen feature sets outperform those reported in previous studies. In addition, the proposed method of semisupervised learning is proved to be effective in distinguishing a noun phrase from a verb phrase both consisting of V N combination, thus enhancing the overall accuracy. \u95dc\u9375\u8a5e\uff1a\u540d\u8a5e\u7d44\u8fa8\u8b58\u3001YamCha\u3001\u76e3\u7763\u5f0f\u5b78\u7fd2\u3001\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2 Keywords\uff1aNP-chunking\u3001YamCha\u3001supervised learning\u3001semi-supervised learning \u4e00\u3001\u7dd2\u8ad6 \u540d\u8a5e\u7d44\u7684\u8fa8\u8b58\u4e00\u76f4\u4ee5\u4f86\u90fd\u662f\u81ea\u7136\u8a9e\u8a00\u7814\u7a76\u53ca\u5176\u76f8\u95dc\u9818\u57df\uff0c\u5982\u7db2\u8def\u63a2\u52d8(web mining)\u3001\u6587\u4ef6\u5206\u985e(text categorization)\u7b49\u975e\u5e38\u95dc\u9375\u7684\u4e00\u500b\u6b65\u9a5f\u3002\u5728\u81ea\u52d5\u554f\u7b54\u7cfb\u7d71 (question answering)\u4e2d\uff0c\u95dc\u9375\u8a5e\u591a\u534a\u4ee5\u540d\u8a5e\u641c\u5c0b\u70ba\u4e3b\u3002\u5728\u81ea\u7136\u8a9e\u8a00\u554f\u7b54\u7cfb\u7d71\u88e1\uff0c\u540d \u8a5e\u7d44\u7684\u8fa8\u8b58\u4e5f\u662f\u4e0d\u53ef\u6216\u7f3a\u3002\u6211\u5011\u6bcf\u5929\u90fd\u4f7f\u7528\u7684\u641c\u5c0b\u5f15\u64ce\uff0c\u5927\u5bb6\u8f38\u5165\u7684\u95dc\u9375\u8a5e\u4ee5\u53ca\u641c\u5c0b \u5f15\u64ce\u7d71\u8a08\u51fa\u4f86\u7684\u71b1\u9580\u95dc\u9375\u8a5e\u4ea6\u4ee5\u540d\u8a5e\u7d44\u5c45\u591a\uff1b\u5927\u81f3\u641c\u5c0b\u5f15\u64ce\u80cc\u5f8c\u5927\u91cf\u7684\u8cc7\u6599\u5eab\u3001\u5c0f\u81f3 \u666e\u901a\u6587\u672c\u5efa\u6a94\u800c\u6210\u7684\u8cc7\u6599\u5eab\uff0c\u88fd\u4f5c\u7d22\u5f15\u5206\u985e\u6642\uff0c\u540d\u8a5e\u7d44\u7684\u4f7f\u7528\u4e5f\u591a\u65bc\u52d5\u8a5e\u3001\u526f\u8a5e\uff1b\u4e5f \u6709\u8d8a\u4f86\u8d8a\u591a\u7db2\u9801\u90fd\u5728\u91dd\u5c0d\u7db2\u9801\u4e2d\u7684\u540d\u8a5e\u7d44\u505a\u81ea\u52d5\u5075\u6e2c\u4ee5\u53ca\u5916\u90e8\u9023\u7d50 ... \u9664\u4e86\u9019\u4e9b\u4f8b\u5b50\u4e4b \u5916\uff0c\u8a9e\u610f\u89d2\u8272\u6a19\u8a18(semantic role labeling)\u3001\u5c08\u6709\u540d\u8a5e\u8fa8\u8b58(name entity identification)\u3001\u6587\u7ae0\u8655\u7406\u4e2d\u7684\u56de\u6307(coreference)\uff0c\u540d\u8a5e\u7d44\u7684\u8fa8\u8b58\u90fd\u662f\u4e00\u500b\u91cd\u8981\u7684\u6b65 \u9a5f\uff0c\u56e0\u6b64\u6709\u597d\u7684 NP chunker\uff0c\u53ef\u4ee5\u6539\u5584\u8a31\u591a NLP \u7684\u7814\u7a76\u6210\u679c\u4ee5\u53ca\u76f8\u95dc\u61c9\u7528\u3002",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Wi \u662f\u51fa\u73fe\u5728\u7b2c i-th \u500b\u4f4d\u7f6e\u7684\u8a5e, Ti \u662f Wi \u7684\u8a5e\u6027 \u800c Ci \u662f\u7b2c i-th \u500b\u5b57\u7684\u8a5e\u7d44\u985e\u5225\u6a19\u8a18\u3002 \u53e6\u5916\uff0c\u4ed6\u5011\u628a\u7279\u5fb5\u96c6\u4e2d\u7684 (Ci+1, Ci+2) \u63db\u6210 (Ci-1, Ci-2) \u4ee5\u9054\u5230\u53cd\u5411\u5256\u6790\u7684\u6548\u679c\u3002\u7531\u65bc \u5728\u6e2c\u8a66\u6642\uff0c\u8a5e\u7d44\u985e\u5225\u6a19\u8a18\u9019\u500b\u7279\u5fb5 (\u6b63\u5411\u5256\u6790\uff1aCi-1, Ci-2\uff1b\u53cd\u5411\u5256\u6790\uff1aCi+1,Ci+2) \u4e26 \u4e0d\u662f\u4e8b\u5148\u7d66\u5b9a\uff0c\u800c\u662f\u5229\u7528\u7576\u4e0b\u6a21\u578b\u6c7a\u5b9a\u7684\u7d50\u679c\uff0c\u56e0\u6b64\u88ab\u7a31\u70ba\u52d5\u614b\u7279\u5fb5\uff1b\u76f8\u5c0d\u7684 Wi \u548c Ti \u5247\u70ba\u975c\u614b\u7279\u5fb5\u3002\u7d66\u5b9a\u4e00\u500b\u53e5\u5b50\uff0c\u4f8b\u5982\uff1a\u9019/\u662f/\u8a5e\u7d44/\u7bc4\u4f8b/\u6a19\u8a18\uff0c\u8868\u4e00\u662f\u5c0d\u61c9\u7684\u7bc4\u4f8b\u5411 \u91cf\uff0c\u5176\u4e2d B \u548c O \u5206\u5225\u8868\u793a\u8a72\u8a5e\u662f\u540d\u8a5e\u7d44\u7684\u958b\u59cb\u6216\u4e0d\u5728\u540d\u8a5e\u7d44\u5167\u3002 \u8868\u4e00\u3001\u300c\u9019\u662f\u8a5e\u7d44\u7bc4\u4f8b\u6a19\u8a18\u300d\u5728 Kudo \u6f14\u7b97\u6cd5\u4e2d\u5c0d\u61c9\u7684\u7bc4\u4f8b\u5411\u91cf \u95dc\u6ce8\u8a5e \u7684\u985e\u5225 Wi Wi-2 Wi-1 Wi+1 Wi+2 Ti Ti-1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "\u5f35\u5e2d\u7dad\uff0c\u9ad8\u7167\u660e\uff0c\u5289\u662d\u9e9f ( 2005 )\u5229\u7528\u5411\u91cf\u652f\u6490\u6a5f\u8fa8\u8b58\u4e2d\u6587\u57fa\u5e95\u540d\u8a5e\u7d44\u7684\u521d\u6b65\u7814 \u7a76\u3002\u7b2c\u5341\u4e03\u5c46\u81ea\u7136\u8a9e\u8a00\u8207\u8a9e\u97f3\u8655\u7406\u7814\u8a0e\u6703\u3002 pp",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "317--332",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "\u5f35\u5e2d\u7dad\uff0c\u9ad8\u7167\u660e\uff0c\u5289\u662d\u9e9f ( 2005 )\u5229\u7528\u5411\u91cf\u652f\u6490\u6a5f\u8fa8\u8b58\u4e2d\u6587\u57fa\u5e95\u540d\u8a5e\u7d44\u7684\u521d\u6b65\u7814 \u7a76\u3002\u7b2c\u5341\u4e03\u5c46\u81ea\u7136\u8a9e\u8a00\u8207\u8a9e\u97f3\u8655\u7406\u7814\u8a0e\u6703\u3002 pp. 317-332",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Use of Support Vector Learning for Chunk Identification",
"authors": [
{
"first": "Taku",
"middle": [],
"last": "Kudo",
"suffix": ""
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of CoNLL-2000",
"volume": "",
"issue": "",
"pages": "142--144",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kudo, Taku, and Matsumoto, Yuji. (2000). Use of Support Vector Learning for Chunk Identification. In Proceedings of CoNLL-2000, pp. 142-144.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Unsupervised word sense disambiguation rivaling supervised methods",
"authors": [
{
"first": "D",
"middle": [],
"last": "Yarowsky",
"suffix": ""
}
],
"year": 1995,
"venue": "Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "189--196",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yarowsky, D. (1995). Unsupervised word sense disambiguation rivaling supervised methods. Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics (pp. 189-196).",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Chunking with Support Vector Machine",
"authors": [
{
"first": "Taku",
"middle": [],
"last": "Kudo",
"suffix": ""
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of NAACL 2001",
"volume": "",
"issue": "",
"pages": "192--199",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kudo, Taku, and Matsumoto, Yuji. (2001). Chunking with Support Vector Machine. In Proceedings of NAACL 2001, pp. 192-199.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "LIBSVM --A Library for Support Vector Machines",
"authors": [
{
"first": "Chih-Chung And",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Chih-Jen",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chang, Chih-Chung and Lin, Chih-Jen. (2004) LIBSVM --A Library for Support Vector Machines.[On line]. Available. http://www.csie.ntu.edu.tw/~cjlin/libsvm/",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Chinese Chunking Based on Maximum Entropy Markov Models. Computational Linguistics and Chinese Language Processing",
"authors": [
{
"first": "Guang-Lu",
"middle": [],
"last": "Sun",
"suffix": ""
},
{
"first": "Chang-Ning",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "Xiao-Long",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Zhi-Ming",
"middle": [],
"last": "Xu",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "11",
"issue": "",
"pages": "115--136",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Guang-Lu Sun, Chang-Ning Huang, Xiao-Long Wang, and Zhi-Ming Xu .Chinese Chunking Based on Maximum Entropy Markov Models. Computational Linguistics and Chinese Language Processing Vol. 11, No. 2, June 2006, pp. 115-136",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "A high-performance semi-supervised learning method for text chunking",
"authors": [
{
"first": "R",
"middle": [
"K"
],
"last": "Ando",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the Annual Meetings of the Association for Computational Linguistics (ACL)",
"volume": "",
"issue": "",
"pages": "1--9",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. K. Ando and T. Zhang. A high-performance semi-supervised learning method for text chunking. In Proceedings of the Annual Meetings of the Association for Computational Linguistics (ACL), pages 1-9. 2005",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Semi-supervised literature survey",
"authors": [
{
"first": "Xiaojin",
"middle": [],
"last": "Zhu",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Xiaojin Zhu, Semi-supervised literature survey, December 14, 2007",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Machine Learning-based Dependency Analyzer for Chinese",
"authors": [
{
"first": "Yuchang",
"middle": [],
"last": "Cheng",
"suffix": ""
},
{
"first": "Asahara",
"middle": [],
"last": "Masayuki",
"suffix": ""
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": ""
}
],
"year": 2005,
"venue": "Journal of Chinese Language and Computing",
"volume": "15",
"issue": "1",
"pages": "13--24",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yuchang CHENG and Masayuki ASAHARA and Yuji MATSUMOTO,\"Machine Learning-based Dependency Analyzer for Chinese\", Journal of Chinese Language and Computing 15 (1): (13-24) ,2005",
"links": null
}
},
"ref_entries": {
"TABREF0": {
"type_str": "table",
"html": null,
"text": "\u5728\u9019\u7bc7\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u5148\u7528 Taku Kudo \u6240\u63d0\u51fa\u5229\u7528 SVM \u7684\u6f14\u7b97\u6cd5\u7576\u4f5c\u4e00\u958b\u59cb\u7684\u6a21\u578b\uff0c\u9664 \u4e86\u8a31\u591a\u53c3\u8003\u6587\u737b\u4e2d\u5e38\u7528\u7684 IOB \u6a19\u793a\u6cd5\u4ee5\u53ca\u4f4d\u7f6e\uff0c\u6211\u5011\u9084\u5617\u8a66\u4e86\u4ee5\u4e0d\u540c\u6a19\u793a\u6cd5\u4ee5\u53ca\u52a0\u5165 \u4e0d\u540c\u4f4d\u7f6e\u7684\u53e5\u5b50\u90e8\u4efd\u8cc7\u8a0a\u7576\u4f5c\u7279\u5fb5\uff0c\u8b49\u660e\u5c0d\u65bc\u4e2d\u6587\u7684\u8655\u7406\uff0c\u4e0d\u8ad6\u662f\u5c01\u9589\u6216\u958b\u653e\u5be6\u9a57\u4e2d\uff0c IOE \u8868\u793a\u6cd5\u548c\u52a0\u5165\u524d\u5f8c\u5169\u500b\u4f4d\u7f6e\u7684\u8a5e\u53ca\u4e2d\u7814\u9662\u7c21\u5316\u6a19\u8a18\uff0c\u662f\u6211\u5011\u5229\u7528\u4e2d\u7814\u9662\u53e5\u6cd5\u6a39\u5eab Sinica Treebank \u6240\u80fd\u5f97\u5230\u6700\u597d\u7684\u7d50\u679c\u7684\u53c3\u6578\u3002\u63a5\u8457\uff0c\u6211\u5011\u5229\u7528\u4e00\u500b\u6c92\u6709\u53e5\u6cd5\u7d50\u69cb\u8a0a\u606f \u7684\u5927\u578b\u8a9e\u6599\u5eab word sketch engine \u4e2d\u7684\u53e5\u5b50\uff0c\u52a0\u4e0a\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u4e2d\uff0c\u81ea\u6211\u5b78\u7fd2\u7684\u6982\u5ff5\uff0c \u5229\u7528\u7db2\u8def\u4e0a\u5927\u91cf\u672a\u6a19\u8a18\u7684\u7db2\u9801\uff0c\u4f86\u5f4c\u88dc Sinica TreeBank \u88e1\u4e0d\u8db3\u7684\u8a0a\u606f\uff0c\u6539\u5584\u5229\u7528\u76e3\u7763 \u5f0f\u5b78\u7fd2\u6cd5\u5be6\u505a\u51fa\u7684 chunker\u3002 \u5be6\u9a57\u90e8\u4efd\uff0c\u9664\u4e86\u5c01\u9589\u6e2c\u8a66\u5916\uff0c\u7531\u65bc\u4e2d\u7814\u9662\u6a39\u5eab\u5716\u4e2d\u8cc7\u6599\u6709\u9650\uff0c\u6211\u5011\u984d\u5916\u6536\u96c6\u4e86\u4e0d\u540c\u985e \u578b\u7684\u53e5\u5b50\u7576\u4f5c\u958b\u653e\u6e2c\u8a66\u7684\u8a9e\u6599\uff0c\u4ee5\u5206\u5225\u6bd4\u8f03\u5169\u7a2e\u4f5c\u6cd5\u5728\u540d\u8a5e\u7d44\u8fa8\u8b58\u7684\u6548\u679c\u53ca\u9650\u5236\u3002\u5be6 \u9a57\u7d50\u679c\u986f\u793a\uff0c\u6211\u5011\u9078\u7528\u7684\u53c3\u6578\u8f03\u524d\u4eba\u9078\u7528\u7684\u53c3\u6578\u505a\u51fa\u7684\u6a21\u578b\u5728\u7b2c\u4e00\u968e\u6bb5\u958b\u653e\u6e2c\u8a66\u4e2d\u9ad8 \u51fa\u4e86 16 \u500b\u767e\u5206\u6bd4\uff0c\u5728\u7b2c\u4e8c\u500b\u958b\u653e\u6e2c\u8a66\u4e2d\u4e5f\u6709 70\uff05\u7684 f-rate\uff1b\u52a0\u5165 unlabeled data \u9019\u500b\u6b65 \u9a5f\u7684\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\uff0c\u4e5f\u7684\u78ba\u63d0\u6607\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u7684\u6548\u679c\uff0c\u4f7f\u958b\u653e\u6e2c\u8a66\u7684 f-rate \u63d0\u9ad8\u81f3 78.79%\uff0c\u4e0d\u4f46\u4fdd\u5b58\u5206\u985e\u5668\u7684\u512a\u9ede\uff0c\u4e5f\u660e\u986f\u63d0\u6607\u4e2d\u6587\u5728\u96e3\u89e3\u7684\u540d\u7269\u5316\u6b67\u7fa9\u7684\u540d\u8a5e\u8fa8\u8b58\u7d50 \u679c\u3002 \u63a5\u4e0b\u4f86\u7684\u7ae0\u7bc0\u4e2d\uff0c\u7b2c\u4e8c\u7ae0\u70ba\u6587\u737b\u56de\u9867\u5305\u542b Chunking, SVM, \u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u7684\u57fa\u672c\u4ecb\u7d39\uff1b \u7b2c\u4e09\u7ae0\u53ca\u7b2c\u56db\u7ae0\u5206\u5225\u70ba\u5be6\u9a57\u65b9\u6cd5\u8aaa\u660e\u548c\u6578\u64da\u7d50\u679c\u8a0e\u8ad6\u3002\u6700\u5f8c\u70ba\u7d50\u8ad6\u8207\u672a\u4f86\u5c55\u671b\u3002 \u51fa\u4e00\u500b\u9069\u7528\u65bc\u4e0d\u540c\u8a9e\u8a00\u7684 rule-based chunker\u3002\u5373\u4f7f\u7f3a\u4e4f\u5927\u91cf\u7684\u8a13\u7df4\u8a9e\u6599\uff0c\u53ea\u8981\u6709\u4e00\u4e9b \u80fd\u5920\u8fa8\u5225\u7d50\u69cb\u908a\u754c\u7684\u898f\u5247\uff0c\u5c31\u80fd\u4f7f\u7528 Kinyon \u6240\u63d0\u51fa\u7684\u65b9\u6cd5\u3002Igor (2005)\u6bd4\u8f03\u5229\u7528 NLTK \u5de5\u5177\u5be6\u505a\u5b8c\u6210\u7684 rule-based chunker \u4ee5\u53ca\u5229\u7528 TnT(Trigram and tag) \u7d71\u8a08\u65b9\u5f0f\u9019\u5169 \u7a2e\u65b9\u6cd5\u505a\u51fa\u7684 chunker \u5728\u540d\u8a5e\u7d44\u548c\u52d5\u8a5e\u7d44\u4e0a\u7684\u8fa8\u8b58\u6548\u679c\uff0c\u5be6\u9a57\u8b49\u660e\u8fd1\u4f86\u9bae\u5c11\u88ab\u5927\u5bb6\u63a1 \u7528\u7684\u898f\u5247\u65b9\u5f0f\u5be6\u505a\u51fa\u4f86\u7684 chunker \u4e0d\u4f46\u6c92\u6709\u6bd4\u5229\u7528\u7d71\u8a08\u7684 chunker \u905c\u8272\uff0c\u751a\u81f3\u5728\u53ec\u56de\u7387 (recall)\u53ca f-rate \u7684\u8868\u73fe\u4e0a\u8981\u4f86\u7684\u66f4\u597d\u3002\u5728\u4e2d\u6587\u65b9\u9762\uff0c\u96d6\u7136 Zhao \u7b49(1999)\u5c0d\u67d0\u4e9b\u985e\u578b \u7684\u8a5e\u7d44\u6574\u7406\u51fa\u7d50\u69cb\u7684\u898f\u5247\uff0c\u4f46 Zhao \u7b49(1999)\u9084\u662f\u6368\u68c4\u898f\u5247\u5f0f\u7684\u4f5c\u6cd5\uff0c\u4f7f\u7528\u8a18\u61b6\u57fa\u790e\u5b78 \u7fd2(memory-based learning)\u7684\u65b9\u5f0f\u3002\u4ed6\u5011\u7684\u5be6\u9a57\u986f\u793a\u82e5\u4e0d\u52a0\u4e0a\u8a5e\u5f59\u672c\u8eab\u7684\u8a0a\u606f\uff0c\u800c\u53ea \u6709\u5229\u7528\u8a5e\u6027\u7684\u8a0a\u606f\u4e0b\uff0c\u6548\u679c\u6703\u6bd4\u8f03\u5dee\u3002 (\u4e8c)\u3001\u76e3\u7763\u5f0f\u5b78\u7fd2\u53ca\u7d71\u8a08\u65b9\u6cd5",
"content": "
1 \u3001Kudo \u7684\u652f\u6301\u5411\u91cf\u6a5f\u6f14\u7b97\u6cd5 |
Kudo \u7b49(2000) \u7b2c\u4e00\u500b\u5c07 SVM \u6709\u6548\u5229\u7528\u5728\u8a5e\u7d44\u8fa8\u8b58\u4f5c\u696d\u4e0a\u3002\u5b83\u5229\u7528\u5468\u570d\u7684\u8a5e\u3001\u9019\u4e9b\u8a5e |
\u7684\u8a5e\u6027\u4ee5\u53ca\u9810\u6e2c\u7684\u8a5e\u7d44\u985e\u5225\u7576\u4f5c\u8a13\u7df4\u3001\u9810\u6e2c\u904e\u7a0b\u4e2d\u7684\u7279\u5fb5\u96c6\uff0c\u5229\u7528 SVM \u5c0d\u6bcf\u500b\u8a5e\u505a\u6a19 |
\u8a18\u7684\u52d5\u4f5c\u3002\u8981\u8fa8\u8b58\u7b2c i \u500b\u5b57\u7684\u8a5e\u7d44\u985e\u5225 Ci\uff0cKudo \u63a1\u7528\u4e86\u5982\u5716\u4e00\u7684\u7279\u5fb5\uff1a |
\u5716\u4e00\u3001Kudo \u63d0\u51fa\u7684\u6f14\u7b97\u6cd5\u4e2d\u6240\u4f7f\u7528\u7684\u7279\u5fb5[2] |
\u4e8c\u3001 \u6587\u737b\u56de\u9867 |
(\u4e00)\u3001\u898f\u5247\u6cd5 |
Abney(1995)\u5229\u7528\u4e86\u6709\u9650\u72c0\u614b\u6a5f\u505a\u51fa\u7684\u898f\u5247\u5f0f\u5256\u6790\u5668\u3002\u4ed6\u7684\u5be6\u9a57\u5229\u7528\u8a9e\u610f\u7684\u8a0a\u606f\u4f8b\u5982\u5b57 |
\u5f62\u8b8a\u5316\u4f86\u7576\u4f5c\u7279\u5fb5\u3002\u4ed6\u5c0d\u82f1\u6587\u53ca\u5fb7\u6587\u505a\u4e86\u6e2c\u8a66\u90fd\u6210\u529f\u4e26\u4e14\u5feb\u901f\u7684\u53d6\u51fa\u4e3b\u8981\u985e\u578b\uff0c\u5305\u62ec |
\u52d5\u8a5e\u3001\u540d\u8a5e\u3001\u4ecb\u4fc2\u8a5e\u7684\u8a5e\u7d44\u3002\u4e0d\u904e\u4ed6\u9084\u662f\u5f37\u8abf\u9078\u53d6\u7684\u6587\u6cd5\u7684\u91cd\u8981\u6027\u3002Kinyon(2000) \u63d0 |
\u5728\u5927\u898f\u6a21\u8a9e\u6599\u5eab\u5efa\u7acb\u4e4b\u524d\uff0c\u540d\u8a5e\u7d44\u8fa8\u8b58\u5e38\u5229\u7528\u7d44\u6210\u540d\u8a5e\u7d44\u7d50\u69cb\u898f\u5f8b\u900f\u904e\u6709\u9650\u72c0\u614b\u6a5f\u627e |
\u51fa\u7b26\u5408\u7684\u6a21\u5f0f pattern\uff0c\u6216\u5f9e\u6a19\u8a18\u597d\u8a5e\u6027\u7684\u8a9e\u6599\u5eab\u4ee5\u7d71\u8a08\u65b9\u5f0f\u5f97\u5230\uff0c\u6216\u7d50\u5408\u8a9e\u8a00\u898f\u5f8b\u53ca |
\u8a9e\u6599\u5eab\u7d71\u8a08\uff1b\u96a8\u8457\u8cd3\u5dde\u5927\u5b78\u6a39\u5eab\u5716(University of Penn TreeBank)\u958b\u653e\u7d66\u5927\u5bb6\u4f7f\u7528\u4e4b |
\u5f8c\uff0c\u8a5e\u7d44\u8fa8\u8b58\u4e5f\u671d\u5411\u4ee5\u6a5f\u5668\u5b78\u7fd2\u7684\u65b9\u6cd5\u4f86\u89e3\u6c7a\uff1aSkut and Brants(1998)\u3001Koeling (2000) |
and Osborne (2000)\u4f7f\u7528\u6700\u5927\u71b5\u6f14\u7b97\u6cd5\uff1bPark and Zhang \u63a1\u7528\u898f\u5247\u4ee5\u53ca\u8a18\u61b6\u5b78\u7fd2 |
(memory-based learning, MBL)\u7d9c\u5408\u7684\u65b9\u5f0f\uff1bKudo and Matsumoto(2000,2001)\u5229\u7528 8 |
\u500b Support Vector Machine(SVM)\u7cfb\u7d71\u6295\u7968(voting)\u7684\u65b9\u5f0f\u5f97\u51fa chunking \u6a21\u578b\uff0c\u5176 |
\u4ed6\u5229\u7528\u76e3\u7763\u5f0f\u5b78\u7fd2(supervised learning)\u7684\u65b9\u6cd5\u9084\u6709 Hidden Markov Model(HMM)(Li |
(2004))\u3001transform-based learning(Ramshaw and Marcus (1995))\u9019\u5e7e\u7a2e\uff0c\u5927\u90fd\u662f\u5229\u7528\u8a9e |
\u6599\u7684\u7d50\u69cb\u53ca\u524d\u5f8c\u8a9e\u5883\u7684\u7279\u5fb5\u5f97\u5230\u7684\u3002\u9019\u4e9b\u6f14\u7b97\u6cd5\u4e5f\u65e9\u5df2\u88ab\u7528\u5728\u5176\u4ed6\u8ddf\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u6709 |
\u95dc\u7684\u8b70\u984c\u4e0a\u3002 |
",
"num": null
},
"TABREF1": {
"type_str": "table",
"html": null,
"text": "\u7d04\u70ba\u8a13\u7df4\u96c6\u7684 2%)\u7684\u53e5\u5b50\u51fa\u4f86\u6a19\u8a18\u7b54\u6848\uff0c \u4f8b\u5982\uff1a\u53e5\u5b50\u4e2d\u7684 plant \u662f\u690d\u7269\u5247\u6b64\u53e5\u70ba class A\uff0c\u662f\u5de5\u5ee0\u5247\u70ba class B\u3002\u6839\u64da(1)\u7684\u5047\u8a2d\uff0c \u627e\u51fa\u540c\u985e\u5225\u53e5\u5b50\u4e2d\u7684\u642d\u914d\u8a9e\uff0c\u4f8b\u5982\uff1a\u4e00\u958b\u59cb\u9078\u5b9a\u7684 A \u4e2d\u7684\u53e5\u5b50\u7684 plant \u65c1\u908a\u90fd\u6709 life \u9019 \u500b\u5b57\uff0c\u800c B \u7684 plant \u65c1\u908a\u90fd\u6709 manufactoring \u9019\u500b\u5b57\uff0c\u5229\u7528\u9019\u500b\u7279\u5fb5\u5230\u672a\u6a19\u8a18\u904e\u7684\u8cc7\u6599 \u6bcf\u4eba \u4e00 \u672c \u5c0e\u89bd \u624b\u518a\u3002 \u91cf\u8a5e\u7684\u5f8c\u9762\u5e38\u5e38\u6703\u63a5\u540d\u8a5e\u7d44\u3002 \u9019\u500b \u5831\u5c0e \u7d66\u4e88 \u6211 \u5011 \u7121\u9650 \u7684 \u60f3\u50cf \u7a7a\u9593\u3002 \"\u7684\"\u7684\u5f8c\u9762\u5e38\u5e38\u6703\u63a5\u540d\u8a5e\u7d44\u3002 \u5927\u5bb6 \u770b \u4e86 \u5ba3\u5c0e \u77ed\u7247 \u4e4b\u5f8c \u6709 \u4ec0\u9ebc \u611f\u60f3 \u5462\uff1f \u6642\u614b\u8a5e\u7684\u5f8c\u9762\u5e38\u6703\u63a5\u540d\u8a5e\u7d44\u3002 Bi-gram\uff0c\u5f9e word sketch engine \u4e2d\u641c\u5c0b 50 \u53e5\u5305\u542b\u9019\u4e9b bigram \u7684\u53e5\u5b50\uff0c\u5982\u4e0b\u9762\u5e7e\u53e5\u662f\u5305\u542b\u63a1\u8cfc\u4eba\u54e1\u7684\u53e5\u5b50\uff1a ...\u5718\u9ad4 \u4e4b \u63a1\u8cfc\u4eba\u54e1\u3002 ...\u9700\u8981\u4ef0\u8cf4\u7684\u4e0d\u53ea\u6709 \u7e3d\u52d9 \u63a1\u8cfc\u4eba\u54e1 \u81ea\u8eab\u3002 ...\u53c8\u860a\u542b \u4e86 \u63a1\u8cfc\u4eba\u54e1 \u5c0d \u4ed6\u4eba\u548c\u793e\u6703... ...\u5927\u591a\u6578 \u7684 \u63a1\u8cfc\u4eba\u54e1 \u6703 \u856d\u898f\u66f9\u96a8... \u5728\u8a13\u7df4\u8a9e\u6599\u4e2d\uff0c\u7d04\u6709 45000 \u500b\u53e5\u5b50\uff0c\u4f9d\u7167 Church \u7684\u5b9a\u7fa9\uff0c\u6240\u53d6\u51fa\u7684 NP chunk \u5171 \u6709 65009 \u500b\uff0c\u4f46\u662f\u6bcf\u500b chunk \u5e73\u5747\u53ea\u5305\u542b\u4e86 1.57 \u500b\u8a5e\u3002\u53ef\u898b\u5f97\u5728\u4e2d\u7814\u9662\u6a39\u5eab\u5716\u88e1\u6a19\u8a18 \u5f97 np-chunk \u9084\u662f\u4ee5\u55ae\u8a5e\u5c45\u591a\u3002\u5f9e\u4e0a\u4e00\u500b\u53e5\u5b50\u4e2d\uff0c\u4f9d\u7167\u6a39\u5eab\u5716\u4e2d\u7684\u7d50\u69cb\uff0c\u300c\u89c0\u5149\u5c40\u300d \u300c\u5e7e\u8655\u5e02\u90ca\u300d\u300c\u6d3b\u52d5\u300d\u90fd\u6210\u70ba\u4e00\u500b\u55ae\u7368\u7684 chunk\uff0c\u4f46\u5c11\u4e86\u6211\u5011\u8a8d\u5b9a\u7684\u300c\u904a\u89bd\u6d3b\u52d5\u300d\u751a \u81f3\u6709\u53ef\u80fd\u662f\u300c\u5e7e\u8655\u5e02\u90ca\u904a\u89bd\u6d3b\u52d5\u300d\u3002\u5982\u679c\u53ea\u5229\u7528 non-recursive chunk \u7684\u7d50\u69cb\u5f9e Treebank \u7a05\u6350\u8655 \u5de5\u5546 \u7a05\u79d1 \u3001 \u8ca1\u7522\u7a05\u79d1 \u3001 \u7a3d\u5fb5 \u79d1 \u53ca \u7a05\u52d9 \u7ba1\u7406 \u79d1 \u7b49 \u4f9d\u7167 \u6b0a\u8cac \uff0c \u5c07 \u5206\u5225 \u5168\u9762 \u67e5\u7ddd \u9003\u6f0f\u7a05 \u3002 \u5728/P21 \u8ce6\u7a05/Naeb \u65b9\u9762/Nac\uff0c/COMMACATEGORY \u67e5\u7ddd/VC2 \u9003\u6f0f\u7a05/Na \u53ca /Caa \u9032\u884c/VC2 \u6703\u8a08\u5e2b/Nab \u8a55\u9451 \u5728\u8a5e\u6027\u6a19\u8a18\u9019\u90e8\u4efd\u8207\u4e2d\u7814\u9662\u65b7\u8a5e\u7a0b\u5f0f\u4e0d\u540c\u7684\u5730\u65b9\u5728\u65bc\uff0c\u4e2d\u7814\u9662\u65b7\u8a5e\u7a0b\u5f0f\u4f3c\u4e4e\u53c3\u8003\u4e86\u53e5 \u5b50\u7684\u8a9e\u5883\u6a19\u793a\u6bcf\u500b\u8a5e\u7684 POS \u800c WSE \u6c92\u6709\uff0c\u6240\u4ee5\u5728\u4e0d\u5c6c\u65bc\u529f\u80fd\u8a5e(function word)\u7684\u90e8",
"content": "Ti+1 1:SHI 1:NA 1:NA \u6709\u4ee5\u4e0b \u652f\u6301\u5411\u91cf\u6a5f\u4e3b\u8981\u7684\u60f3\u6cd5\u662f\u88fd\u9020\u4e00\u500b\u6700\u4f73\u7684\u5e73\u9762\u53ef\u4ee5\u8b93\u8a13\u7df4\u7bc4\u4f8b\u5411\u91cf\u5206\u6210\u5169\u500b\u985e B 1:\u9019 1:0 1:0 1:\u662f 1:\u8a5e\u7d44 1:NES 1:0 O 1:\u662f 1:0 1:\u9019 1:\u8a5e\u7d44 1:\u7bc4\u4f8b 1:SHI 1:NES B 1:\u8a5e\u7d44 1:\u9019 1:\u662f 1:\u7bc4\u4f8b 1:\u6a19\u8a18 1:NA 1:SHI \u660e\u986f\u7684\u512a\u9ede\uff1a (1)\u5373\u4f7f\u5728\u9ad8\u7dad\u7279\u5fb5\u5411\u91cf\u7a7a\u9593\u4e0b\u9084\u662f\u80fd\u7522\u751f\u597d\u7684\u6548\u80fd\u3002 (2)\u6838\u5fc3\u51fd\u6578\u80fd\u5c07\u8cc7\u6599\u6620\u5c04\u5230\u66f4\u9ad8\u7dad\u7684\u7a7a\u9593\u800c\u6c92\u6709\u589e\u52a0\u8a08\u7b97\u8907\u96dc\u5ea6\u3002 \u5225(positive and negative)\u4e26\u4e14\u628a\u9019\u500b\u5e73\u9762\u7684\u908a\u754c\u6700\u5927\u5316\u3002\u5716\u4e8c\u4e2d\uff0c\u9ed1\u8272\u5be6\u7dda\u5c31\u662f\u5169\u500b \u53ef\u5c07\u8cc7\u6599\u5206\u6210\u5169\u985e\u7684\u5e73\u9762\uff0c\u5169\u689d\u865b\u7dda\u4e2d\u9593\u7684\u8ddd\u96e2\u5c31\u662f\u908a\u754c(margin)\uff0c\u4e5f\u5c31\u662f SVM \u6f14 \u7b97\u6cd5\u8a66\u8457\u6700\u5927\u5316\u7684\u76ee\u6a19\u3002\u5728\u865b\u7dda\u5169\u908a\u7684\u9ede\u7a31\u70ba\u652f\u6301\u5411\u91cf(support vectors)\uff0c\u800c\u4e14\u53ea\u6709 \u5728\u8a13\u7df4\u96c6\u4e2d\u7684\u652f\u6301\u5411\u91cf\u6703\u5f71\u97ff\u6574\u500b\u6a21\u578b\u7684\u7d50\u679c\u3002\u96d6\u7136 SVM \u7684\u5206\u985e\u6e96\u78ba\u5ea6\u5341\u5206\u9a5a\u4eba\uff0c\u8a08 \u7b97\u8907\u96dc\u5ea6\u8ddf\u5176\u4ed6\u6a5f\u5668\u5b78\u7fd2\u65b9\u6cd5\u6bd4\u8d77\u4f86\u4e5f\u76f8\u5c0d\u7684\u9ad8\u4e86\u8a31\u591a\u3002\u5728\u9700\u8981\u9f90\u5927\u7684\u8a13\u7df4\u8a9e\u6599\u96c6\u7684 \u72c0\u6cc1\u4e0b\uff0c\u5229\u7528 SVM \u7684\u8a13\u7df4\u904e\u7a0b\u4e0d\u4f46\u4e0d\u5920\u6709\u6548\u7387\uff0c\u751a\u81f3\u6709\u53ef\u80fd\u56e0\u70ba\u9700\u8981\u7684\u8a13\u7df4\u6642\u9593\u592a\u4e45 \u9019\u7a2e\u56e0\u7d20\uff0c\u5be6\u969b\u60c5\u6cc1\u4e0b\u7121\u6cd5\u770b\u5230\u6210\u679c\u3002 \u5716\u4e8c\u3001\u5169\u7a2e\u53ef\u80fd\u5c07\u8cc7\u6599\u5206\u958b\u7684\u8d85\u5e73\u9762[2] YamCha(Yet Another Multi-purpose Chunking Annatator)\u662f Taku Kudo \u57fa\u65bc Taku \u7b49 (2000)\u4e2d\u7684\u6f14\u7b97\u6cd5\uff0c\u8a2d\u8a08\u5c08\u9580\u7528\u5728\u89e3\u6c7a\u8a5e\u7d44\u8fa8\u8b58\u3001\u8a5e\u6027\u6a19\u8a18\u751a\u81f3\u6587\u4ef6\u5206\u985e\u7b49\u81ea\u7136\u8a9e\u8a00 \u8655\u7406\u61c9\u7528\u7684\u5de5\u5177\u3002\u6574\u500b\u67b6\u69cb\u63a1\u7528\u7684\u5206\u985e\u65b9\u6cd5\u662f SVM\u3002\u8ddf\u55ae\u7d14\u7684 SVM \u5206\u985e\u5668\u4e0d\u540c\u7684\u5730 \u65b9\u662f\uff0cYamcha \u8981\u6c42\u7684\u8f38\u5165\u6a94\u6848\u683c\u5f0f\u6bd4\u8f03\u7b26\u5408\u4eba\u76f4\u89c0\u7684\u60f3\u6cd5\uff0c\u628a\u9700\u8981\u505a\u8a5e\u7d44\u5206\u985e\u7684\u8cc7\u6599 \u5982\u540c\u5716\u4e09\uff0c\u6bcf\u500b\u8a5e\u6703\u5229\u7528\u5230\u7684\u7279\u5fb5\u7c21\u55ae\u7684\u6392\u5217\uff0c\u76f4\u63a5\u4ea4\u7d66 YamCha \u53bb\u57f7\u884c\u5373\u53ef\u3002\u5982\u679c \u4e0d\u505a\u4efb\u4f55\u53c3\u6578\u7684\u8b8a\u52d5\uff0c\u9019\u500b\u5de5\u5177\u5c31\u7528 Taku \u4e2d\u4e00\u6a23\u7684\u9810\u8a2d\u503c\uff0c\u628a (n-2, n-1, n, n+1, n+2) \u4f4d\u7f6e\u4e0a\u7684\u5b57\u548c\u7279\u5fb5\u90fd\u7576\u6210\u95dc\u6ce8\u8a5e(\u7b2c n \u500b\u8a5e)\u7684\u7279\u5fb5\u96c6\u53bb\u8a13\u7df4\u3002\u6240\u4ee5\u8ddf\u50b3\u7d71\u5206\u985e\u5668\u4e0d\u540c\u7684 \u5730\u65b9\uff0c\u53ea\u5728\u65bc YamCha \u5e6b\u4f7f\u7528\u8005\u8655\u7406\u4e86\u8cc7\u6599\u683c\u5f0f\u7684\u554f\u984c\u3002\u53e6\u5916\u503c\u5f97\u6ce8\u610f\u7684\u4e00\u9ede\u662f\uff0c\u96d6 \u7136 SVM \u5206\u985e\u7684\u6548\u679c\u975e\u5e38\u7684\u597d\uff0c\u4f46\u5176\u8017\u8cbb\u7684\u8a08\u7b97\u91cf\u4ee5\u53ca\u6642\u9593\u4e5f\u6bd4\u5176\u4ed6\u5206\u985e\u5668\u4f86\u7684\u5927\uff0c\u800c YamCha \u5728\u9019\u9ede\u4e0a\u505a\u4e86\u6539\u9032\uff0c\u4f7f\u5f97\u8a13\u7df4\u6642\u9593\u4ee5\u53ca\u5206\u985e\u6642\u9593\u90fd\u52a0\u901f\u4e86\u81f3\u5c11\u4e09\u500d\u4ee5\u4e0a\u3002 He reckons the current account deficit PRP VBZ DT JJ NN NN B-NP B-VP B-NP I-NP I-NP I-NP \u5716\u4e09\u3001Yamcha \u8f38\u5165\u8cc7\u6599\u7684\u683c\u5f0f\uff0c\u8207 CoNLL 2000 shared task \u76f8\u540c\u3002\u5176 \u4e2d B, I, O \u5206\u5225\u8868\u793a\u8a72\u8a5e\u662f\u67d0\u7a2e\u8a5e\u7d44\u7684\u958b\u59cb\uff0c\u5167\u90e8\uff0c\u6216\u4e0d\u5728\u8a5e\u7d44\u4e2d\u3002 (\u56db)\u3001\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5 \u5982\u540c\u8868\u9762\u4e0a\u7684\u5b57\u7fa9\u4e00\u6a23\uff0c\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u4ecb\u65bc\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u548c\u975e\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u4e4b\u9593\uff1a \u5229\u7528\u5927\u91cf\u672a\u6a19\u8a18\u904e\u7684\u8cc7\u6599\u7d50\u5408\u4e00\u4e9b\u5df2\u7d93\u6a19\u8a18\u904e\u7684\u8cc7\u6599\u4f86\u505a\u8a13\u7df4\u7684\u6a21\u578b\u4ee5\u89e3\u6c7a\u8cc7\u6599\u91cf\u7a00 \u5c11\u53ca\u5206\u6563\u7684\u554f\u984c\u3002\u800c\u5728\u4e00\u500b\u81ea\u7136\u7684\u8003\u91cf\u4e0b\uff0c\u5982\u679c\u6709\u4e00\u6a23\u6578\u91cf\u6a19\u8a18\u904e\u7684\u8cc7\u6599(labeled data)\uff0c\u6211\u5011\u662f\u4e0d\u662f\u80fd\u5229\u7528\u5927\u91cf\u5bb9\u6613\u53d6\u5f97\u7684\u672a\u6a19\u8a18\u904e\u3001\u672a\u8655\u7406\u904e\u7684\u8cc7\u6599(unlabeled data)\u4f86\u5efa\u9020\u4e00\u500b\u66f4\u7cbe\u78ba\u7684\u5206\u985e\u5668(classifier)\uff1f \u9019\u500b\u554f\u984c\u901a\u5e38\u5c31\u6703\u88ab\u6b78\u985e\u65bc\u534a\u76e3\u7763\u5f0f \u5b78\u7fd2(semi-supervised learning)\u3002\u5728\u771f\u5be6\u751f\u6d3b\u4e2d\uff0c\u6a19\u8a18\u8cc7\u6599\u4e0d\u4f46\u8017\u8cbb\u6642\u9593\u3001\u4eba\u5de5\u3001\u751a \u81f3\u91d1\u9322\uff1b\u76f8\u5c0d\u7684\uff0c\u672a\u7d93\u904e\u6a19\u8a18\u7684\u8cc7\u6599\u91cf\u591a\u800c\u4e14\u96a8\u624b\u53ef\u5f97\u3002\u56e0\u6b64\u5728\u6a5f\u5668\u5b78\u7fd2\u7684\u9818\u57df\u4e0a\uff0c \u5982\u4f55\u5229\u7528\u672a\u6a19\u8a18\u8cc7\u6599\u662f\u4e00\u500b\u91cd\u8981\u7684\u8ab2\u984c\uff0c\u4f8b\u5982\uff1a\u6211\u5011\u53ef\u4ee5\u5229\u7528\u7a0b\u5f0f\u53d6\u5f97\u7db2\u8def\u4e0a\u5927\u91cf\u7684 \u7db2\u9801\u5b58\u6a94\uff0c\u537b\u9700\u8981\u4eba\u5de5\u8655\u7406\u624d\u80fd\u505a\u6b63\u78ba\u7684\u5206\u985e\uff1b\u5728\u8a9e\u97f3\u8fa8\u8b58\u7684\u7814\u7a76\u4e2d\uff0c\u8981\u53d6\u5f97\u5927\u91cf\u7684 \u9304\u97f3\u6a94\u975e\u5e38\u7c21\u55ae\uff0c\u4f46\u662f\u8981\u53bb\u6a19\u8a18\u9019\u4e9b\u9304\u97f3\u6a94\u5247\u9700\u8981\u5927\u91cf\u7684\u6642\u9593\u548c\u4eba\u529b\u53bb\u6253\u9010\u5b57\u7a3f\uff0c\u5728 \u9019\u4e9b\u60c5\u6cc1\u4e0b\uff0c\u5982\u679c unlabeled data \u78ba\u5be6\u80fd\u8b93\u6a21\u578b\u7684\u6548\u80fd\u63d0\u9ad8\uff0c\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u662f\u975e\u5e38\u7684\u6709 \u5e6b\u52a9\u7684\u3002 \u5728\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u4e2d\uff0c\u8cc7\u6599\u578b\u614b\u548c\u4e0d\u540c\u6f14\u7b97\u6cd5\u7684\u6a21\u578b\u7684\u914d\u5c0d\u662f\u5341\u5206\u91cd\u8981\u7684\uff0c\u4e0d\u7136\u4e5f\u6709\u53ef \u80fd\u9020\u6210\u53cd\u6548\u679c\u3002EM with generative mixture model\u3001self-training\u3001co-training\u3001TSVM \u9084 \u6709\u57fa\u65bc\u5716\u5f62\u7684\u6f14\u7b97\u6cd5\u90fd\u662f SSL \u88e1\u5e38\u7528\u7684\u65b9\u6cd5\u4e26\u4e14\u5404\u6709\u512a\u9ede\u3002\u5047\u5982\u6a19\u8a18\u7684\u985e\u5225\u53ef\u4ee5\u628a\u628a \u8cc7\u6599\u5206\u958b\u5f97\u5f88\u660e\u78ba\uff0c\u4f7f\u7528 EM with generative mixture model \u6bd4\u8f03\u597d\uff1b\u5982\u679c\u7279\u5fb5\u7fa4\u5c31\u8db3\u5920 \u96a8\u8457\u8cc7\u6599\u88ab\u5206\u6210\u5169\u534a\uff0c\u90a3\u9ebc co-training \u7684\u65b9\u6cd5\u6703\u6bd4\u8f03\u9069\u5408\uff0c\u56e0\u70ba\u9019\u500b\u6f14\u7b97\u6cd5\u5c31\u662f\u5c0d\u5206 \u958b\u7684\u7279\u5fb5\u505a\u4e00\u4e9b\u5047\u8a2d\uff0c\u4e26\u5728\u4e0d\u540c\u7684\u7279\u5fb5\u96c6\u4e0a\u4f7f\u7528\u4e0d\u540c\u7684\u5b78\u7fd2\u5de5\u5177\u3002\u82e5\u6709\u76f8\u540c\u7684\u7279\u5fb5\u7684 \u9ede\u6703\u88ab\u5206\u5728\u540c\u4e00\u500b\u985e\u5225\uff0c\u800c\u76ee\u524d\u7684\u6a21\u578b\u4e5f\u7121\u6cd5\u88ab\u6539\u9032\u6642\uff0c Mincut, Boltzmann Machine, Tree-based Bayes \u7b49\u57fa\u65bc\u5716\u5f62\u7684\u65b9\u6cd5\u6703\u6bd4\u8f03\u9069\u5408\u3002 \u6700\u65e9\u5f15\u9032\u5229\u7528\u672a\u6a19\u8a18\u904e\u7684\u8cc7\u6599\u9019\u500b\u6982\u5ff5\u61c9\u7528\u5728\u5206\u985e\u4e0a\u7684\u4e5f\u8a31\u5c31\u662f\u81ea\u6211\u5b78\u7fd2\u9019\u7a2e\u6f14\u7b97\u6cd5\u3002 \u4e00\u958b\u59cb\u53ea\u5229\u7528\u5c11\u91cf\u6a19\u8a18\u904e\u7684\u8cc7\u6599\u4f5c\u8a13\u7df4\uff0c\u63a5\u8457\u5229\u7528\u7576\u4e0b\u7684\u6c7a\u7b56\u51fd\u6578(decision function) \u5f9e\u672a\u6a19\u8a18\u7684\u8cc7\u6599\u4e2d\u627e\u51fa\u7b26\u5408\u7684\u9ede\uff0c\u52a0\u9032\u672c\u4f86\u6a19\u8a18\u904e\u7684\u8a13\u7df4\u96c6\uff0c\u91cd\u65b0\u8a13\u7df4\u4e26\u627e \u51fa\u65b0\u7684\u6c7a\u7b56\u51fd\u6578\u3002\u76f4\u5230\u5728\u672a\u6a19\u8a18\u7684\u8cc7\u6599\u4e2d\u7121\u6cd5\u518d\u627e\u5230\u53ef\u6a19\u793a\u7684\u8cc7\u6599\u6216\u662f\u6574\u500b\u60c5\u5f62\u8d85\u904e \u4e00\u4e9b\u95a5\u503c(threshold)\u3002 Yarowsky(1995)\u662f\u5728\u63d0\u5230\u81ea\u6211\u5b78\u7fd2\u6642\u5e38\u88ab\u5f15\u7528\u7684\u6709\u540d\u7684\u4f8b\u5b50\u3002Yarowsky \u5229\u7528\u9019\u500b\u65b9 \u6cd5\u505a\u5206\u8fa8\u8a9e\u610f\u6b67\u7570\u5ea6\uff0c\u8ddf\u53ea\u7528\u6a19\u8a18\u904e\u8cc7\u6599\u548c\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u7684\u6548\u679c\u6bd4\u8d77\u4f86\u6539\u9032\u5f88\u591a\u3002\u5728 \u4e00\u500b\u7bc7\u7ae0\u4e2d\uff0c\u8981\u5982\u4f55\u6c7a\u5b9a plant \u9019\u500b\u5b57\u662f\u690d\u7269\u7684\u610f\u601d\u9084\u662f\u5de5\u5ee0\u7684\u610f\u601d\uff1fYarowsky \u7684\u5047 \u8a2d\u662f\uff1a (1)\u4e00\u500b\u8a5e\u5728\u4e00\u7ae0\u7bc7\u5e45\u4e2d\u53ea\u6703\u6709\u4e00\u500b\u610f\u601d,\uff0c(2)\u4e00\u7d44\u642d\u914d\u8a9e\u4e2d\u4e5f\u53ea\u6703\u6709\u4e00\u500b \u610f\u601d\u3002\u4ed6\u5148\u5f9e\u672a\u6a19\u8a18\u904e\u7684\u8cc7\u6599\u4e2d\uff0c\u53d6\u5c0f\u91cf(\u4e2d\u53bb\u627e\u6709\u540c\u6a23\u7279\u5fb5\u7684\u53e5\u5b50\uff0c\u6536\u9032\u8a13\u7df4\u96c6\u88e1\u3002\u4e5f\u540c\u6642\u5229\u7528\u4e00\u4e9b\u6c7a\u7b56\u51fd\u6578\u5728\u9019\u4e9b\u65b0\u6a19\u8a18\u597d \u7684\u53e5\u5b50\u4e2d\u627e\u65b0\u7684\u642d\u914d\u8a9e\u3002\u63a5\u8457\u5229\u7528(2)\u7684\u5047\u8a2d\uff0c\u5982\u679c\u540c\u4e00\u7bc7\u6587\u7ae0\u4e2d\u7684\u591a\u500b\u53e5\u5b50\u90fd\u5df2\u7d93 \u88ab\u6b78\u5230\u540c\u4e00\u500b\u985e\u5225\uff0c\u5247\u540c\u7bc7\u6587\u7ae0\u4e2d\u5269\u4e0b\u7684\u53e5\u5b50\u90fd\u53ef\u4ee5\u5206\u5230\u540c\u4e00\u500b\u985e\u5225\uff0c\u9019\u500b\u5047\u8a2d\u4e0d\u4f46 \u53ef\u4ee5\u64f4\u5927\u8a13\u7df4\u96c6\uff0c\u9084\u53ef\u4ee5\u4fee\u6b63\u5728\u524d\u9762\u7684\u6b65\u9a5f\u88ab\u6b78\u985e\u932f\u7684\u53e5\u5b50\u3002\u91cd\u8907\u64f4\u5927\u8a13\u7df4\u8cc7\u6599\u3001\u5229 \u7528\u65b0\u7684\u8cc7\u6599\u8a13\u7df4\u6a21\u578b\u7684\u6b65\u9a5f\u76f4\u5230\u672a\u6a19\u8a18\u8cc7\u6599\u7684\u6578\u91cf\u4e0d\u518d\u6709\u592a\u5927\u7684\u8b8a\u5316\u70ba\u6b62\u3002\u5be6\u9a57\u7d50\u679c \u8b49\u660e\uff0c\u9019\u500b\u65b9\u6cd5\u8a13\u7df4\u51fa\u4f86\u7684\u6a21\u578b\u8ddf\u53ea\u5229\u7528\u6a19\u8a18\u904e\u8cc7\u6599\u53ca\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u7684\u6548\u679c\u6bd4\u8d77\u4f86\uff0c \u78ba\u5be6\u5c07\u6548\u80fd\u63d0\u6607\u4e26\u4e14\u5c11\u4e86\u5f88\u591a\u4eba\u5de5\u6a19\u8a18\u7684\u52d5\u4f5c\u3002 \u4e09\u3001\u4f5c\u6cd5\u8aaa\u660e (\u4e00)\u3001\u540d\u8a5e\u7d44\u8868\u793a\u6cd5 Inside/Outside\uff1a Ramshaw and Marcus(1995)\u63d0\u51fa\u4e26\u4f7f\u7528\u4e86\u4e0b\u8ff0\u4e09\u7a2e class(IOB)\u8868 \u793a\u4e00\u500b\u8a5e\u5728\u8a5e\u7d44\u4e2d\u7684\u4f4d\u7f6e\u3002 I\uff1a\u9019\u500b\u8a5e\u5728\u67d0\u500b\u8a5e\u7d44\u4e4b\u4e2d O\uff1a\u9019\u500b\u8a5e\u4e0d\u5c6c\u65bc\u4efb\u4f55\u8a5e\u7d44 B\uff1a\u9019\u500b\u8a5e\u662f\u7dca\u63a5\u8457\u5225\u7684\u8a5e\u7d44\u7684\u8a5e\u7d44\u958b\u982d \u9019\u500b\u8868\u793a\u6cd5\u88ab Tjong Kim Sang \u7a31\u70ba IOB1\uff0c\u53e6\u5916\u4ed6\u9084\u63d0\u51fa\u4e86 IOB2/IOE1/IOE2\uff1a IOB2 \u4e2d B \u662f\u4efb\u4f55\u8a5e\u7d44\u7684\u958b\u982d\uff1b IOE1 \u4e2d E \u662f\u7dca\u9130\u8457\u5225\u7684\u8a5e\u7d44\u7684\u8a5e\u7d44\u7d50\u5c3e\uff1b IOE2 \u4e2d E \u662f\u4efb\u4f55\u4efb\u4f55\u8a5e\u7d44\u7684\u7d50\u5c3e\uff1b \u8868\u4e8c\u662f\u5404\u7a2e\u8868\u793a\u6cd5\u7684\u7bc4\u4f8b\u8aaa\u660e\u3002\u53e6\u5916\u9084\u6709 start/end \u7684\u8868\u793a\u6cd5\uff0c\u7531\u65bc\u5728 Taku(2000)\u4e2d\u5be6 \u9a57\u7d50\u679c\u4e0d\u4f73\uff0c\u56e0\u6b64\u672c\u5be6\u9a57\u4e2d\u4e26\u6c92\u6709\u7528\u5230\uff0c\u4e5f\u4e0d\u52a0\u4ee5\u4ecb\u7d39\u3002 \u8868\u4e8c\u3001\u4ee5\u5404\u7a2e\u8868\u793a\u6cd5\u6a19\u8a18\u300c\u9019\u662f\u8a5e\u7d44\u7bc4\u4f8b\u6a19\u8a18\u8aaa\u660e\u300d IOB1 IOB2 IOE1 IOE2 \u9019 I B I E \u662f O O O O \u8a5e\u7d44 I B I I \u7bc4\u4f8b I I I I I I E E \u80fd\u7684\u5171\u901a\u9ede\uff0c\u5982\u4e0b\u8868\uff1a \u4e2d\u53d6\u51fa\u7684\u6587\u7ae0\u3001\u570b\u5c0f\u8ab2\u672c\u3001\u5149\u83ef\u96dc\u8a8c\u4ee5\u53ca\u4e2d\u7814\u9662\u8a9e\u8a00\u6240\u7684\u8a9e\u97f3\u5e73\u8861\u6a94\u6848\uff0c\u518d\u7d93\u904e\u96fb\u8166 \u7684\u8a5e\u7d44\u5408\u800c\u6210\u7684\uff0c\u5373 NP-chunk \u53ef\u7c21\u55ae\u7684\u60f3\u6210\u4e0d\u5305\u542b\u5225\u7a2e\u8a5e\u7d44\u7684\u540d\u8a5e\u7d44\u3002\u4f46 \u56e0\u6b64\u6211\u5011\u8f49\u5411\u958b\u653e\u6e2c\u8a66\u7684\u7d50\u679c\u3002\u4e0b\u5217\u662f\u4e00\u4e9b\u958b\u653e\u6e2c\u8a66\u7684\u53e5\u5b50\uff1a \u56e0\u6b64\u6211\u5011\u8003\u616e\uff1a \u5ea6\u53ca\u7a69\u5b9a\u5ea6\u3002 supervised II 83.83% 63.06% 69.23% 66% \u6a19\u8a18 \u8aaa\u660e B B I E (\u4e8c)\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\uff1aSupervised-learning \u6211\u5011\u4eff\u7167[1]\u53ca[2]\u4e2d\u7684\u6f14\u7b97\u6cd5(\u5728 2.2.1 \u4e2d\u4e5f\u6709\u63cf\u8ff0)\uff0c\u6211\u5011\u5c07\u524d\u5f8c\u5169\u500b\u8a5e\u53ca\u8a5e\u6027 \u5206\u985e\u7684\u8a0a\u606f\u53ca\u524d\u9762\u5169\u500b\u5df2\u7d93\u5224\u5225\u597d\u7684\u8a5e\u5206\u985e\u7576\u4f5c\u7279\u5fb5\u96c6\u8b93 SVM \u5206\u985e\u5668\u53c3\u8003\uff0c\u5229\u7528 SVM \u5206\u985e\u5668\u5224\u5225\u6bcf\u500b\u8a5e\u662f\u5c6c\u65bc IOB \u4e2d\u7684\u54ea\u4e00\u985e\u3002\u4f8b\u5982\u300c\u9019\u662f\u4e00\u500b\u4f8b\u53e5\u300d\u4e2d\u6709\u54ea\u4e9b\u540d \u8a5e\u7d44\uff1f\u8981\u5224\u5225\u300c\u4e00\u300d\u9019\u500b\u8a5e\u6642\uff0c\u5c07\u524d\u5f8c\u5169\u500b\u8a5e\u53ca\u500b\u5225\u7684\u8a5e\u6027\u6a19\u8a18 \u9019\u3001NEP\u3001\u662f\u3001 SHI\u3001\u500b\u3001NF\u3001\u4f8b\u53e5\u3001NA\u3001\u524d\u9762\u5169\u500b\u5df2\u7d93\u5224\u5225\u51fa\u4f86\u7684\u8a5e\u7d44\u5206\u985e E(\u9019)\u3001O(\u662f) \u4ee5\u53ca\u672c\u8eab\u4e00\u3001NEU \u9019 12 \u500b\u7576\u4f5c\u5206\u985e\u7684\u7279\u5fb5\u503c\u3002\u6700\u5f8c\u9019\u4e94\u500b\u8a5e\u5206\u5225\u7684\u985e\u5225\u53ef\u80fd\u70ba \u9019 E \u662f O \u4e00 I \u500b I \u4f8b\u53e5 E \u5247\u53ef\u4ee5\u5229\u7528 IOE \u5206\u51fa\u8a5e\u7d44\u9593\u7684\u908a\u754c\uff0c\u5224\u65b7\u51fa\u6b64\u53e5\u7684\u540d\u8a5e\u7d44\u70ba\u300c\u9019\u300d\u53ca\u300c\u4e00\u500b\u4f8b \u53e5\u300d\u3002 \u56e0\u6b64\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u53ef\u8aaa\u662f\u5e0c\u671b\u80fd\u627e\u51fa\u6709\u6700\u4f73\u8fa8\u8b58\u6548\u679c\u7684\u7279\u5fb5\u96c6\u3002 (\u4e09)\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5 \u5f9e[1]\u53ca\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u7684\u5be6\u9a57\u53ef\u4ee5\u770b\u51fa\uff0c\u53ea\u5f9e\u6a39\u5716\u8cc7\u6599\u5eab\u7684\u8cc7\u8a0a\u52a0\u4e0a\u67e5\u8a62\u8a5e\u5178\u7684\u8a5e\u7fa9\u5c0d \u6211\u5011\u5224\u65b7\u52d5\u8a5e\u7684\u540d\u7269\u5316\u73fe\u8c61\u4e26\u6c92\u6709\u5e6b\u52a9\uff0c\u65bc\u662f\u6211\u5011\u5e0c\u671b\u80fd\u5229\u7528\u5916\u90e8\u7684\u8cc7\u6e90\u5e6b\u52a9\u6211\u5011\u7372 \u5f97\u4e00\u4e9b\u65b0\u7684\u8a0a\u606f\uff0c\u9748\u611f\u4f86\u81ea\u65bc\u524d\u9762\u63d0\u5230\u7684 Self-training\u3002 \u5f9e\u8a13\u7df4\u8a9e\u6599\u7684\u53e5\u5b50\u88e1\uff0c\u6211\u5011\u5c0d\u52d5\u8a5e\u5f8c\u9762\u63a5\u4e00\u500b\u540d\u8a5e\u7684\u7d44\u5408\u505a\u89c0\u5bdf\u4e4b\u5f8c\u767c\u73fe\u4e00\u4e9b\u898f\u5247\uff0c \u4e0b\u8868\u662f\u4f8b\u53e5\u53ca\u63a8\u6e2c\uff1a \u8868\u4e09\u3001\u8a9e\u6599\u4e2d\u90e8\u4efd\u53e5\u5b50\u53ca\u5176\u7279\u6027 \u4f8b\u53e5 \u63a8\u6e2c \u5c0e\u904a \u767c \u7d66 \u8868\u56db\u3001\u5f9e\u8a9e\u6599\u5eab\u63a8\u6e2c\u51fa\u4f86\u7684\u898f\u5247 \u53ef\u80fd\u51fa\u73fe\u5728\u52d5\u8a5e\u5f8c\u9762\u7684\u8a5e\u985e NEQA\u3001NEP\u3001NEU \u7b49\u91cf\u8a5e NG : \u6642\u9593\u5f8c\u7f6e\u8a5e NC : \u4f4d\u7f6e\u8a5e NH : \u4ee3\u540d\u8a5e \u5148\u5f9e\u8a13\u7df4\u8a9e\u6599\u4e2d\u4efb\u9078\u7576\u4fee\u98fe\u7684\u52d5\u8a5e\u63a5\u540d\u8a5e(\u4f8b\u5982\uff1a\u63a1\u8cfc\u4eba\u54e1\u3001\u904b\u52d5\u7cbe\u795e)\uff0c\u4ee5\u53ca\u52d5\u8a5e\u52a0 \u7684\u53e5\u5b50\u524d\u6216\u5f8c\u6703\u9023\u63a5\u8a31\u591a\u4e0d\u540c\u7684\u8a5e\u985e\uff0c\u6211\u5011\u9084\u662f\u53ef\u4ee5\u5206\u5225\u5c0d\u9019\u5169\u7a2e\u60c5\u5f62\u6b78\u7d0d\u51fa\u4e00\u4e9b\u53ef \u56db\u3001\u5be6\u9a57\u7d50\u679c\u8a0e\u8ad6 (\u4e00)\u3001\u5be6\u9a57\u8a9e\u6599\u4ecb\u7d39 \u672c\u6587\u7684\u5be6\u9a57\u4e2d\uff0c\u76e3\u7763\u5f0f\u4ee5\u53ca\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u6240\u9700\u8981\u7684\u8a13\u7df4\u8a9e\u6599\uff0c\u4f7f\u7528\u7684\u662f\u4e2d\u7814\u9662\u4e2d\u6587 \u53e5\u7d50\u69cb\u6a39\u8cc7\u6599\u5eab Sinica Treebank3.1\u3002\u9019\u500b\u8cc7\u6599\u5eab\u7684\u6587\u7ae0\u4f86\u6e90\u5206\u5225\u6709\u76f4\u63a5\u5f9e\u5e73\u8861\u8a9e\u6599\u5eab \u8868\u516d\u662f\u5c01\u9589\u6e2c\u8a66\u7684\u6578\u64da\uff0c\u5f9e\u6b64\u8868\u4e2d\u53ef\u4ee5\u767c\u73fe\u9019\u56db\u500b\u6a21\u578b\u4e4b\u9593\u4e26\u6c92\u6709\u660e\u986f\u6578\u5b57\u4e0a\u7684\u5dee\u8ddd\uff0c \u653e\u6e2c\u8a66\u88e1\u300c\u6709\u4e00\u9593\u5f88\u6f02\u4eae\u7684\u6559\u5e2b\u4f11\u606f\u5ba4\u300d\u7684\u300c\u4e00\u9593\u300d\u5c31\u6709\u8fa8\u8b58\u932f\u8aa4\u7684\u53ef\u80fd\uff0c \u56e0\u70ba\u5728\u91cf\u8a5e\u5f8c\u9762\u7684\u5169\u500b\u8a5e\u90fd\u9084\u4e0d\u898b\u540d\u8a5e\u7684\u8e64\u5f71\u3002 2. \u5373\u4f7f\u6709\u4e00\u4e9b\u53e5\u5b50\u6709\u8457\u975e\u5e38\u985e\u4f3c\u7684\u5f62\u5f0f(\u4f8b\u5982\uff1a\u5f88\u6f02\u4eae\u7684\u8863\u670d\uff0c\u5f88\u8cb4\u7684\u7968)\uff0c \u4f46\u662f\u6a21\u578b\u8f38\u51fa\u7684\u7d50\u679c\u537b\u4e0d\u76f8\u540c\uff0c\u800c\u6211\u5011\u767c\u73fe\u9019\u6216\u8a31\u662f\u56e0\u70ba \"\u5f88\u6f02\u4eae\" \u5728\u8a13\u7df4 \u96c6\u4e2d\u51fa\u73fe\u904e\u70ba\u540d\u8a5e\u4e00\u90e8\u5206\u7684\u7528\u6cd5\uff0c\u800c\"\u8cb4\"\u5728\u8a13\u7df4\u96c6\u88e1\u53ea\u6709\u7576\u52d5\u8a5e\u7528\u3002 model 6 92.30 84.78 85.34 85.06 model 7 90.55 80.54 80.44 80.49 \u9019\u968e\u6bb5\u958b\u653e\u6e2c\u8a66\u7684\u53e5\u5b50\u4e2d\uff0c\u6709\u5f88\u591a\u8a5e(\u6216\u8a5e\u6027\u6a19\u8a18(pos)\u5e8f\u5217)\u662f\u91cd\u8907\u7684\uff0c\u52a0\u5165\u4e0d \u540c\u7684\u642d\u914d\u8a9e\u6216\u4fee\u98fe\u8a9e\uff0c\u6216\u662f\u4ee5\u4e0d\u540c\u7684\u8a9e\u5e8f\u7d44\u5408\uff0c\u662f\u6e2c\u8a66\u6a21\u578b\u5c0d\u4e0d\u540c\u5f62\u5f0f\u7684\u8a9e\u53e5\u7684\u6e96\u78ba supervised II 91.76% 81.71% 86.05% 83.82% semi-supervised 92.19% 84.85% 86.64% 85.73% \u958b\u653e\u6e2c\u8a66 supervised 89.03% 67.31% 72.92% 70% \u4e0a\u53d7\u8a5e(\u4f8b\u5982\uff1a\u796d\u62dc\u7956\u5148\u3001\u4f86\u81ea\u5bb6\u4eba) \u7684\u9019\u5169\u7a2e\u60c5\u5f62\uff0c\u5404 100 \u7d44 \u8490\u96c6\u5230\u67d0\u500b\u6578\u91cf\u7684\u53e5\u5b50\u4e4b\u5f8c\uff0c\u6211\u5011\u5c0d\u95dc\u6ce8\u7684\u8a5e\u7d44\u524d\u5f8c\u7dca\u9130\u7684\u8a5e\u985e\u505a\u7d71\u8a08\u3002\u5373\u4f7f\u6536\u96c6\u4f86 \u8868\u4e94\u3001\u5728\u4e0d\u540c\u529f\u7528\u7684\u52d5\u8a5e\u524d\u7684\u8a5e\u985e\u6bd4\u8f03 \u985e\u578b \u5256\u6790\u53ca\u4eba\u5de5\u6821\u5c0d\u505a\u6210\u6a39\u5716\u5eab\u3002\u5168\u90e8\u5305\u542b\u4e86\u516d\u500b\u6a94\u6848\uff0c\u5206\u5225\u70ba\u4e0d\u540c\u7684\u80cc\u666f\u3001\u60c5\u5883\uff0c\u5171\u6709 65434 \u500b\u4e2d\u6587\u6a39\u5716\u3001392237 \u500b\u8a5e(\u5e73\u5747\u4e00\u53e5\u5305\u542b\u4e86\u516d\u500b\u8a5e)\u3002\u6211\u5011\u628a\u8cc7\u6599\u5eab\u4e2d\u6bcf\u500b\u6a94 \u56de\u982d\u770b\u8a9e\u6599\u5eab\u4e2d\u6a19\u8a18\u597d\u7684 NP-chunk \u6709\u7d55\u5927\u90e8\u5206\u662f\u5c6c\u65bc\u55ae\u8a5e\uff0c\u53cd\u800c\u4e0d\u7b26\u5408\u6211 \u5011\u4e00\u822c\u7684\u601d\u7dad(4.1 \u7bc0)\uff0c\u56e0\u6b64\u82e5\u5229\u7528\u8a9e\u6599\u5eab\u6a19\u8a18\u597d\u7684 NP-chunk\uff0c\u9664\u4e86\u5f88\u96e3 \u6211\u5011\u8cb7\u4e86\u4e00\u5f35\u5f88\u8cb4\u7684\u7968\u3002 \u6211\u8058\u4e86\u4e00\u500b\u5f88\u512a\u79c0\u7684\u8077\u54e1\u3002 1. \u5f80\u524d\u5f8c\u770b\u4e0d\u540c\u9577\u5ea6\u7684\u8a5e\u53ca\u8a9e\u610f\u7279\u5fb5\u4f5c\u70ba\u7279\u5fb5\u96c6\uff1b 2. \u67d0\u500b\u4f4d\u7f6e\u7684\u8a5e\u53ca\u8a9e\u610f\u7279\u5fb5\u4e0d\u5fc5\u540c\u6642\u5b58\u5728\u3002 \u9664\u4e86\u53d7\u9650\u65bc\u8a13\u7df4\u8a9e\u6599\u9019\u500b\u554f\u984c\u4e4b\u5916\uff0c\u5728\u8a13\u7df4\u8a9e\u6599\u65b9\u9762\u9664\u4e86\u4e4b\u524d\u63d0\u5230\u7684\u7f3a\u9ede\u4ee5\u5916\uff0c\u6b67\u7fa9\u3001 semi-supervised 91.61% 76.47% 81.25% 78.79% \u540d\u7269\u5316\u3001\u672a\u77e5\u8a5e\u9019\u4e9b\u5be6\u969b\u751f\u6d3b\u4e2d\u7684\u73fe\u8c61\uff0c\u5728\u8a9e\u6599\u5eab\u88e1\u662f\u6c92\u6709\u6a19\u8a18\u7684\uff1b\u7531\u65bc\u8a9e\u6599\u5eab\u7684\u7d44 \u53ef\u80fd\u51fa\u73fe\u5728\u524d\u9762\u7684\u8a5e\u985e \u52d5\u8a5e D\uff1a\u526f\u8a5e \u4fee\u98fe\u7528\u7684\u52d5\u8a5e DE\uff1a\u7684\uff0c\u4e4b\u7b49\u3001DI\uff1a\u6642\u614b\u6a19\u8a18\u3001 NEU, NF \u7b49\u91cf\u8a5e \u63a5\u8457\u6211\u5011\u9032\u884c\u521d\u6b65\u5c0f\u898f\u6a21\u7684\u6e2c\u8a66\uff0c\u5f9e\u5c01\u9589\u6e2c\u8a66\u8a9e\u6599\u4e2d\uff0c\u9078\u64c7\u52d5\u8a5e\u5f8c\u9762\u7acb\u5373\u63a5\u4e00\u500b\u540d\u8a5e (V1 N1)\u7684\u7d44\u5408\uff0c\u4e00\u6a23\u7684\u5f9e WSE \u4e2d\u8490\u96c6\u53e5\u5b50\uff0c\u63a5\u8457\u7d71\u8a08\u9019\u4e9b\u53e5\u5b50\u4e2d\u7dca\u9130\u8a5e\u7684\u8a5e\u985e\uff0c\u518d \u5229\u7528\u4e0a\u8868\u5148\u524d\u63a8\u6e2c\u51fa\u7684\u7279\u5fb5\u5224\u65b7(V1 N1)\u662f\u54ea\u4e00\u7a2e\u60c5\u5f62\u3002\u4f8b\u5982\u6211\u5011\u8490\u96c6\u5305\u542b\u300c\u8a2d\u8a08 \u4eba\u54e1\u300d\u7684\u53e5\u5b50(\u5716\u56db)\uff1a \u5716\u56db\u3001WSE \u4e2d\u622a\u53d6\u51fa\u4f86\u7684\u4f8b\u53e5 \u7d71\u8a08\u904e\u5f8c\u7684\u7d50\u679c\uff0c\u524d\u9762\u51fa\u73fe\"DE\"\u3001\"DI\"\u53ca\u91cf\u8a5e\u7684\u6a5f\u6703(\u6b21\u6578)\u6bd4\u51fa\u73fe\u526f\u8a5e\"D\"\u7684\u6a5f\u6703 (\u6b21\u6578)\u9ad8\uff0c\u56e0\u6b64\u5c07\u8a2d\u8a08\u4eba\u54e1\u6b78\u985e\u70ba\u4fee\u98fe\u7528\u7684\u52d5\u8a5e\u52a0\u540d\u8a5e\u7d44\u5408\u800c\u6210\u7684\u8907\u5408\u8a5e(class \u5176\u4e2d Vi = { C, Wi \u7684 pos \u4e0d\u5c6c\u65bc\u52d5\u8a5e\u985e\u5225 A, Wi \u662f\u88ab\u540d\u7269\u5316\u7684\u52d5\u8a5e B, Wi \u662f\u52d5\u8a5e \u6848\u7684 70%\u53d6\u51fa\u6574\u5408\u4f86\u7576\u4f5c\u8a13\u7df4\u8a9e\u6599\u300130%\u7576\u4f5c\u5c01\u9589\u6e2c\u8a66\u7684\u6e2c\u8a66\u8a9e\u6599\u3002 \u6a94\u6848\u4e2d\u7684\u53e5\u5b50\u90fd\u4ee5\u4e0b\u53e5\u7684\u5f62\u5f0f\u8868\u793a\uff0c\u9664\u4e86\u53ef\u5f9e\u4e2d\u5f97\u77e5\u7d50\u69cb\u8a0a\u606f\u4e4b\u5916\u9084\u6709\u4e2d\u6587\u7684\u8a9e\u610f\u89d2 \u8272\u3002\u53e5\u5b50\u4e2d\u7684\u8a5e\u985e\u6a19\u8a18\uff0c\u662f CKIP \u8a5e\u985e\u6a19\u8a18\uff0c\u8207\u4e2d\u7814\u9662\u7684\u5b57\u5178\u6240\u4f7f\u7528\u7684\u8a5e\u6027\u6a19\u8a18\u662f\u540c \u4e00\u7a2e(\u53e6\u5916\u6709\u9084\u6709\u7c21\u5316\u6a19\u8a18\u3001\u7cbe\u7c21\u6a19\u8a18\u5169\u7a2e)\u3002 # S(agent:NP(Head:Nca:\u89c0\u5149\u5c40)|evaluation:Dbb:\u9084|quantity:Daa:\u53e6|Head:VE12:\u5b89 \u6392|aspect:Di:\u4e86|theme:NP(property:NP(quantifier:DM:\u5e7e\u8655|Head:Ncb:\u5e02\u90ca)| property:Nv4:\u904a\u89bd|Head:Nac:\u6d3b\u52d5))# \u3002(PERIODCATEGORY) \u81ea\u52d5\u62bd\u53d6\u7b54\u6848\u51fa\u4f86\uff0c\u6703\u5f97\u5230\u5f88\u4e0d\u7406\u60f3\u7684\u8a13\u7df4\u8a9e\u6599\uff0c\u56e0\u6b64\u5728\u6a19\u8a18\u8a5e\u7d44\u7b54\u6848\u4e0a\uff0c\u6211\u5011\u5c0d\u5e7e \u7a2e\u7d50\u69cb\u505a\u4e86\u4fee\u6539\uff0c\u6a19\u8a18\u51fa\u8f03\u7b26\u5408\u6211\u5011\u601d\u7dad\u7684\u7b54\u6848\uff0c\u5305\u542b (1) \u540d\u8a5e+\u7684+\u540d\u8a5e \uff1b(2) \u5f62\u5bb9\u8a5e+\u7684+\u540d\u8a5e\uff1b(3)\u91cf\u8a5e+\u7684+\u540d\u8a5e (\u4e8c)\u3001\u76f8\u95dc\u8cc7\u6e90\u4ecb\u7d39 \u5728\u4e4b\u5f8c\u7684\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u4f7f\u7528\u5230\u4e00\u4e9b\u5ee3\u70ba\u4eba\u77e5\u7684\u5de5\u5177\u53ca\u8cc7\u6e90\uff1a (1)\u4e2d\u7814\u9662\u65b7\u8a5e\u7a0b\u5f0f\uff1a\u8f38\u51fa\u5305\u542b\u5206\u8a5e\u7d50\u679c\uff0c\u8207\u6bcf\u4e00\u500b\u8a5e\u7684\u8a5e\u6027\u6a19\u8a18\uff0c\u4e26\u53ef\u4ee5\u8655 \u7406\u672a\u77e5\u8a5e\u3002 (2)\u8a5e\u5f59\u7279\u6027\u901f\u63cf\u7cfb\u7d71(Word Sketch Engine) [12]\uff1a\u9019\u662f\u4e00\u500b\u5305\u542b\u4e86\u4e2d\u6587\u3001 \u82f1\u6587\u3001\u6cd5\u6587\u3001\u5fb7\u6587\u7b49\u591a\u7a2e\u8a9e\u8a00\u7684\u5927\u578b\u8a9e\u6599\u5eab\uff0c\u4e26\u4e14\u5df2\u7d93\u5c0d\u9019\u4e9b\u8a9e\u6599\u505a\u4e86\u540c\u7fa9\u8a5e\u3001\u7528\u8a9e \u7d22\u5f15\u3001\u642d\u914d\u8a9e\u5206\u985e\u7684\u6574\u7406\u3002\u9019\u500b\u8a9e\u6599\u5eab\u4e2d\u7684\u53e5\u5b50\uff0c\u662f\u6a19\u793a\u597d\u7684\u8cc7\u6599\uff0c\u9664\u4e86\u5df2\u7d93\u505a\u597d\u65b7 \u8a5e\uff0c\u4e5f\u53ef\u4ee5\u67e5\u770b\u8a5e\u6027\u6a19\u8a18\u3002\u5982\u4e0b\u5217\u5169\u53e5\uff1a \u4efd\uff0c\u4e5f\u5c31\u662f\u4e00\u500b\u8a5e\u53ef\u80fd\u6703\u6709\u591a\u7a2e\u8a5e\u6027\u7684\u60c5\u6cc1\u4e0b\uff0cWSE \u7684\u8a5e\u985e\u6a19\u8a18\u7684\u7cbe\u78ba\u5ea6\u6703\u6bd4\u8f03\u5dee\uff0c \u4e26\u4e14\u6bd4\u8f03\u4e0d\u9069\u5408\u62ff\u4f86\u7576\u53c3\u8003\u3002 (3)Google Soap API\uff1a\u8b93\u4f7f\u7528\u8005\u5408\u6cd5\u505a\u95dc\u9375\u5b57\u641c\u5c0b\uff0c\u6bcf\u5929\u6700\u591a\u53ef\u505a 1000 \u7b46\u7684 \u641c\u5c0b\uff0c\u4e26\u63d0\u4f9b\u56de\u50b3\u7db2\u9801\u7684\u76f8\u95dc\u8cc7\u8a0a\u3002\u3002 (\u4e09)\u6f5b\u5728\u554f\u984c \u5c07\u5168\u90e8\u7684\u7b54\u6848\u6539\u6210\u5982\u540c\u6211\u5011\u6240\u5e0c\u671b\u770b\u5230\u7684\u898f\u5247\uff0c\u9084\u6703\u6709\u4e0b\u9762\u63d0\u5230\u7684\u9577\u8a5e\u7d44\u7684 \u554f\u984c\u3002 2. \u7531\u65bc\u4e2d\u6587\u662f\u4e00\u7a2e\u6c92\u6709\u5c48\u6298\u8a9e\u7d20(inflectional morpheme)\u7684\u8a9e\u8a00\uff0c\u4f8b\u5982\u82f1\u6587\u4e2d \u88ab\u52d5\u5f0f\u7684\u52d5\u8a5e\u6703\u6709\u4e00\u500b\u8b8a\u5316\u578b\uff0c\u8f49\u70ba\u540d\u8a5e\u7684\u7528\u6cd5\u5247\u5b57\u5c3e\u8b8a\u6210 ing\u3001\u52a0\u4e0a tion \u7b49\u7b49\uff0c\u4f46\u662f\u5728\u4e2d\u6587\u88e1\u5247\u662f\u52a0\u500b\u300c\u88ab\u300d\u5b57\u4ee5\u8868\u9054\u88ab\u52d5\u5f0f\uff0c\u5176\u4ed6\u7684\u60c5\u5f62\u5728\u524d\u5f8c\u4e0d \u4e00\u5b9a\u6709\u52a0\u5165\u7684\u95dc\u9375\u8a5e\uff0c\u5fc5\u9808\u7531\u5c0d\u4e2d\u6587\u6709\u4e00\u5b9a\u4e86\u89e3\u7a0b\u5ea6\u7684\u4eba\u81ea\u5df1\u5c0d\u8a9e\u5883\u505a\u63a8\u6e2c \u4f86\u5224\u65b7\u6bcf\u500b\u8a5e\u7684\u8a5e\u6027\u53ca\u529f\u7528\u3002\u4f8b\u5982\uff1a\u5f9e\u4e0b\u9762\u9019\u500b\u53e5\u5b50 The experiment involved the combining of the two chemicals\u3002 \u53ef\u4ee5\u5f88\u6e05\u695a\u7684\u770b\u51fa combining \u662f\u540d\u8a5e\u7684\u7528\u6cd5\uff0c\u4f46\u662f\u5728\u4ee5\u4e0b\u9019\u5169\u500b\u53e5\u5b50\uff0c\u7121\u6cd5\u76f4 \u63a5\u770b\u51fa\u9032\u53e3\u548c\u559c\u611b\u7684\u8a5e\u6027\u3002 \u653f\u5e9c\u7de8\u5b9a\u6c7d\u8eca\u7ba1\u7406\u5236\u5ea6\u4f7f\u9032\u53e3\u6c7d\u8eca\u5f97\u4ee5\u5408\u6cd5\u5316\u3002 \u4ed6\u6df1\u5f97\u5b78\u751f\u7684\u559c\u611b\u3002 \u9019\u4e5f\u662f\u5f35\u5e2d\u7dad\u7b49\u3001Ding \u7b49(2005)\u4e2d\u63d0\u53ca\u7684\u540d\u7269\u5316\u73fe\u8c61\u3002\u7531\u65bc\u5728 Sinica TreeBank \u88e1\u6709\u9019\u7a2e\u73fe\u8c61\u7684\u8a5e\u7d44\u53ea\u6709\u4e0d\u5230 3000 \u7d44\uff0c\u56e0\u6b64\u5f35\u5e2d\u7dad\u7b49\u4e5f\u6839\u64da\u5be6\u9a57\u7d50\u679c\u5f37\u8abf\u5229\u7528\u76e3\u7763\u5f0f \u5b78\u7fd2\u6cd5\u8fa8\u8b58\u4e2d\u6587\u7684\u540d\u8a5e\u7d44\u6642\uff0c\u80fd\u5426\u627e\u51fa\u88ab\u540d\u7269\u5316\u7684\u52d5\u8a5e\u662f\u4e00\u500b\u63d0\u6607\u6b63\u78ba\u7387\u7684\u95dc\u9375\u3002 3. \u5f9e Sinica TreeBank \u4e2d\u53d6\u51fa\u6a19\u8a18\u597d\u7684 chunk \u7684\u5e73\u5747\u9577\u5ea6\u4e0d\u5230\u5169\u500b\u8a5e\uff0cCheng \u7b49 (2005)\u63d0\u5230\u4e2d\u6587\u5be6\u969b\u4e0a\u6709\u975e\u5e38\u591a\u7531\u6578\u500b\u540d\u8a5e\u7d44\u5408\u800c\u6210\u5f97\u540d\u8a5e\u7d44\uff0c\u4f8b\u5982\uff1a\u884c\u653f \u9662/\u570b\u5bb6/\u79d1\u5b78/\u59d4\u54e1\u6703\u3001\u96fb\u8166/\u4eba\u9ad4/\u6a21\u578b...\u7b49\u7b49\uff0c\u9019\u4e9b\u5728\u65e5\u5e38\u751f\u6d3b\u4e2d\u90fd\u4e0d\u662f\u4ee4 \u4eba\u964c\u751f\u7684\u8a5e\u8a9e\u3002\u56e0\u6b64\u4f7f\u7528 Sinica TreeBank \u7576\u505a\u4e00\u7a2e gold standard \u6216\u662f\u8a13\u7df4\u8a9e \u6599\u6642\uff0c\u5f88\u96e3\u89e3\u6c7a\u9577\u8a5e\u7684\u554f\u984c\u3002 (\u56db)\u958b\u653e\u6e2c\u8a66\u96c6 \u7531\u65bc\u4e0a\u9762\u8aaa\u660e\u4e86\u5f88\u591a\u5728\u8a13\u7df4\u53ca\u5c01\u9589\u6e2c\u8a66\u8a9e\u6599\u4e2d\u7121\u6cd5\u89c0\u5bdf\u5230\u7684\u60c5\u5f62\uff0c\u56e0\u6b64\u6211\u5011\u5229 \u7528\u958b\u653e\u6e2c\u8a66\u7684\u7d50\u679c\u4f5c\u70ba\u4e0d\u540c\u65b9\u5f0f\u8a2d\u8a08\u7684\u6a21\u578b\u9593\u6bd4\u8f03\u7684\u6e96\u5247\u3002\u5728\u76e3\u7763\u5f0f\u53ca\u534a\u76e3\u7763\u5f0f\u5b78 \u7fd2\u65b9\u6cd5\u4e5f\u5404\u6709\u4e00\u7d44\u7684\u958b\u653e\u6e2c\u8a66\u8cc7\u6599\u4f5c\u70ba\u8a72\u6b21\u5be6\u9a57\u5167\u7684\u53c3\u6578\u6bd4\u8f03\u3002\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u4e2d\u7684\u958b \u653e\u6e2c\u8a66\u8a9e\u6599\uff0c\u5927\u90e8\u5206\u662f\u5305\u542b\"\u5f62\u5bb9\u8a5e\u63a5\u540d\u8a5e\"\u3001\"\u91cf\u8a5e\u63a5\u540d\u8a5e\"\u3001\"\u91cf\u8a5e\u63a5\u5f62\u5bb9\u8a5e\u63a5\u540d\u8a5e\"\u3001 \u540d\u8a5e\u4e2d\u6709\u6240\u6709\u683c\u7684\u53e5\u5b50\uff0c\u4f8b\u5982\uff1a\u9019\u662f\u6700\u65b0\u7684\u8eca\u6b3e\u3001\u4e8b\u60c5\u767c\u751f\u5728\u53bb\u5e74\u7684\u590f\u5929\u3001\u73ed\u4e0a\u6709\u4e00 \u540d\u5929\u624d\u5b78\u751f..\u7b49\u7b49\uff1b\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u7684\u958b\u653e\u6e2c\u8a66\u8a9e\u6599\u5f37\u8abf\u52d5\u8a5e\u7684\u5224\u5225\uff0c\u56e0\u6b64\u6e2c\u8a66\u8a9e\u6599\u4e2d \u7684\u540d\u8a5e\u7d44\u5305\u542b\u4e00\u4e9b\u5df2\u8f49\u5316\u70ba\u5225\u7684\u4f5c\u7528\u7684\u52d5\u8a5e\uff0c\u4f8b\u5982\uff1a\"\u4ed6\u5728\u62cd\u8ce3\u7db2\u7ad9\u4e0a\u8cb7\u6771\u897f\"\uff0c\"\u8b66\u5bdf \u900f\u904e\u92b7\u8d13\u7ba1\u9053\u6293\u5230\u5c0f\u5077\"\u7b49\u7b49\u3002 \u5be6\u9a57\u7d50\u679c\u6211\u5011\u63a1\u7528\u8207 CoNLL 2000 shared task \u4e00\u6a23\u7684\u8a55\u91cf\u65b9\u6cd5\uff0c\u76f4\u63a5\u5229\u7528\u4ed6\u5011\u63d0\u4f9b\u7684\u8a55 \u91cf\u5de5\u5177\uff0c\u5206\u5225\u7b97\u51fa\u8a5e\u6a19\u8a18\u6b63\u78ba\u7387(tag accuracy)\u4ee5\u53ca\u8a5e\u7d44\u6b63\u78ba(Precision)\u3001\u8a5e\u7d44\u53ec \u56de\u7387(Recall)\u4ee5\u53ca F-rate = 2PR/P+R\uff0c\u800c F-rate \u9084\u662f\u70ba\u4e3b\u8981\u8003\u91cf\u3002 (\u4e94)\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5 Supervised-learning \u963f\u5fe0\u7684\u90a3\u4e00\u9593\u623f\u5b50\u3002 \u8868\u516d\u3001\u6bd4\u8f03 IOB,IOE,CKIP,simplified(\u7c21\u5316\u6a19\u8a18)\u56db\u7a2e\u7279\u5fb5\u5728\u5c01\u9589\u6e2c\u8a66\u6642\u7684\u7d50\u679c feature combination tag accuracy precision recall f-rate W(n-2 .. n+2) , P(n-2 .. n+2) in Simplified tagset , T(n-2 .. n-1), IOB 91.21% 84.85% 86.98% 85.90% W(n-2 .. n+2) , P(n-2 .. n+2) in CKIP tagset, T(n-2 .. n-1), IOB 90.89% 84.44% 86.60% 85.50% W(n-2 .. n+2) , P(n-2 .. n+2) in Simplified tagset, T(n-2 .. n-1), IOE 92.06% 84.65% 86.28% 85.46% W(n-2 .. n+2) , P(n-2 .. n+2) in CKIP tagset, T(n-2 .. n-1), IOE 91.93% 84.34% 86.09% 85.20% W(n-2 .. n+2) , P(n-2 .. n+2) V in CKIP, others in Simplified, T(n-2 .. n-1), IOE 92.11% 84.67% 86.23% 85.44% \u7531\u65bc[1]\u5df2\u7d93\u8aaa\u660e\u76e3\u7763\u5f0f\u5b78\u7fd2\u8a13\u7df4\u51fa\u4f86\u7684\u6a21\u578b\u5c0d\u540d\u7269\u5316\u8a5e\u985e\u6c92\u6709\u597d\u7684\u8fa8\u8b58\u6548\u679c\uff0c\u56e0\u6b64\u5728 \u9019\u90e8\u4efd\u7684\u958b\u653e\u6e2c\u8a66\uff0c\u6211\u5011\u5148\u9078\u64c7\u4e00\u4e9b\u53e5\u5b50\u88e1\u6709\u540d\u8a5e\u7d44\u4e2d\u57fa\u672c\u5f62\u5f0f\uff0c\u50cf\u91cf\u8a5e\u63a5\u540d\u8a5e\u3001\u5f62 \u5bb9\u8a5e\u63a5\u540d\u8a5e\u3001\u91cf\u8a5e\u63a5\u5f62\u5bb9\u8a5e\u52a0\u540d\u8a5e\uff0c\u52a0\u4e0a\u4e00\u4e9b\u5305\u542b\u9577\u8907\u5408\u8a5e\u7d44\u4ee5\u53ca\u6709\u6240\u6709\u683c\u7684\u53e5\u5b50\u3002 \u8868\u516b\u662f\u7279\u5fb5\u96c6\u7684\u7b26\u865f\u4ee5\u53ca\u5176\u4ee3\u8868\u7684\u610f\u7fa9\uff1b \u8868\u516b\u3001\u5be6\u9a57\u4e2d\u63a1\u7528\u7684\u7279\u5fb5\u4ee3\u8868\u7b26\u865f\u4ee5\u53ca\u76f8\u5c0d\u7684\u610f\u7fa9 \u4ee3\u8868\u7b26\u865f \u4ee3\u8868\u610f\u7fa9 \u4ee3\u8868\u7b26\u865f \u4ee3\u8868\u610f\u7fa9 Wn \u7b2c n \u500b\u8a5e IOB \u5229\u7528 IOB \u8868\u793a\u6cd5\u6a19\u8a18\u540d\u8a5e\u7d44 Ln \u7b2c n \u500b\u8a5e\u7684 POS IOE \u5229\u7528 IOE \u8868\u793a\u6cd5\u6a19\u8a18\u540d\u8a5e\u7d44 Tn \u7b2c n \u500b\u8a5e\u7684 tag \u6a19\u8a18 Hn \u7b2c n \u500b\u8a5e\u5728 hownet \u4e2d\u7684\u7fa9\u5143 Simplified \u7c21\u5316\u6a19\u8a18 CKIP CKIP \u6a19\u8a18 F forward parsing B backward parsing \u8868\u4e5d\u662f\u4e0d\u540c\u6a21\u578b\u6240\u4f7f\u7528\u7684\u7279\u5fb5\u7d44\u5408\uff0c\u7531\u65bc\u9019\u4e9b\u7d44\u5408\u5728\u5c01\u9589\u6e2c\u8a66\u7684\u7d50\u679c\u90fd\u5341\u5206\u76f8\u8fd1\uff0c\u6240 \u4ee5\u4e0d\u5c07\u6578\u64da\u4e00\u4e00\u5217\u51fa\uff1b \u8868\u4e5d\u3001\u6a21\u578b\u53ca\u5176\u5229\u7528\u7684\u7279\u5fb5\u5c0d\u7167\u8868 model 1 F, Wi-2.. Wi+2, Pi-2..Pi+2 Ti-2..Ti-1 model 2 F, Wi-2.. Wi+2, Pi-2..Pi+2,Hi-2..Hi+2, Ti-2..Ti-1 model 3 F, Wi-4.. Wi+2, Pi-4..Pi+2 Ti-2..Ti-1 \u6210\u5927\u591a\u662f\u9577\u7bc7\u6587\u7ae0\u5207\u6210\u7684\u53e5\u5b50\uff0c\u67d0\u4e9b\u7528\u6cd5\u6216\u8a9e\u53e5\uff0c\u6703\u56e0\u70ba\u51fa\u81ea\u65bc\u540c\u4e00\u7bc7\u6587\u7ae0\u800c\u91cd\u8907\u51fa \u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u4e2d\uff0c\u6211\u5011\u53ea\u7528\u4e86\u8490\u96c6\u4f86\u7684\u8a9e\u6599\u88e1\u7279\u5b9a\u8a5e\u7d44\u524d\u5f8c\u7684\u8a5e\u6027\u505a\u5224\u5225\uff0c \u73fe\u5f88\u591a\u6b21\uff0c\u5be6\u969b\u7684\u8a13\u7df4\u8a9e\u6599\u4e26\u4e0d\u5982\u7d71\u8a08\u904e\u5f8c\u6578\u5b57\u4e0a\u5f97\u591a\u3002\u96d6\u7136\u5728\u9019\u500b\u8a9e\u6599\u5eab\u88e1\u5171\u6709\u516d \u4f46\u9019\u8207\u5728\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u88e1\u5c07\u524d\u5f8c\u7684\u8a5e\u6027\u7368\u7acb\u5206\u985e(supervised II)\u4f5c\u70ba\u7279\u5fb5\u6709\u4f55\u4e0d\u540c\u5462? \u842c\u591a\u53e5\u7684\u53e5\u5b50\uff0c\u4f46\u662f\u56e0\u70ba\u5206\u53e5\u662f\u4ee5\u6a19\u9ede\u7b26\u865f\u70ba\u539f\u5247\uff0c\u6240\u4ee5\u6709\u5f88\u591a\u53e5\u5b50\u5176\u5be6\u53ea\u5305\u542b\u55ae\u8a5e\uff0c \u7531\u65bc\u5728\u8a13\u7df4\u8a9e\u6599\u4e2d\u5305\u542b\u9019\u4e9b\u8b8a\u5316\u52d5\u8a5e\u7684\u8a5e\u7d44\u7d04\u70ba 5\uff05\uff0c\u4e5f\u5c31\u662f\u8a13\u7df4\u8a9e\u6599\u592a\u5c11\uff1b\u53e6\u5916\u53e5 \u6216\u53ea\u7531\u540d\u8a5e\u7d44\u6210\uff0c\u6216\u6bd4\u6b63\u5e38\u60c5\u6cc1\u66f8\u5beb\u4f86\u7684\u77ed\uff0c\u7121\u6cd5\u5f9e\u4e2d\u770b\u51fa\u7d50\u69cb\u7684\u8a0a\u606f\uff0c\u53cd\u800c\u53ef\u80fd\u8b8a \u5b50\u4e2d\u7684\u524d\u5f8c\u8a5e\u6027\u975e\u5e38\u53ef\u80fd\u53ea\u662f\u6070\u597d\u5728\u7576\u53e5\u4e2d\u51fa\u73fe\u5728\u9694\u58c1\uff0c\u800c\u4e0d\u662f\u771f\u6b63\u5177\u6709\u8fa8\u5225\u4f5c\u7528\u7684 \u6210\u5206\u985e\u5668\u7684\u96dc\u8a0a(noise)\u3002\u4f8b\u5982\uff1a\u300c\u7238\u7238\u8aaa\uff1a\u5c71\u8def\u4e0d\u96e3\u8d70\u300d\u9019\u500b\u53e5\u5b50\u5728\u8a9e\u6599\u5eab\u4e2d\u88ab\u5206 \u6210\u5206\uff0c\u5982\u6211\u5011\u53ef\u80fd\u9700\u8981\u5927\u91cf\u7684 DE \u6216 DI \u985e\u7684\u8a5e\u4f86\u4f50\u8b49\u4e00\u500b VN \u7d44\u5408\u70ba\u540d\u8a5e\u7684\u8907\u5408\u8a5e\u7d44\uff0c \u70ba\u300c\u7238\u7238/\u8aaa\u300d\u548c\u300c\u5c71\u8def/\u4e0d/\u96e3/\u8d70\u300d\u5169\u53e5\uff0c\u4f46\u5be6\u969b\u751f\u6d3b\u4e2d\u591a\u6578\u7684\u5beb\u6cd5\u9084\u662f\u6703\u4ee5\u9577\u53e5\u5b50\u70ba \u4f46\u662f\u5728\u8a9e\u6599\u5eab\u4e2d\u9019\u500b\u8a5e\u7d44\u7684\u524d\u9762\u662f\u500b\u666e\u901a\u540d\u8a5e\uff0c\u5982\u540c\u6211\u5011\u8490\u96c6\u4f86\u7684\u53e5\u5b50\u4e5f\u6709\u975e\u5e38\u591a\u662f \u4e3b\u3002\u4e2d\u7814\u9662\u53e5\u6cd5\u6a39\u5eab\u4e2d\u7684\u8a5e\u53ca\u8a5e\u7d44\u6a19\u8a18\u7531\u65bc\u7d93\u904e\u4eba\u5de5\u6821\u5c0d\uff0c\u6240\u4ee5\u76f8\u7576\u7cbe\u78ba\u3001\u53ef\u4fe1\u5ea6\u9ad8\uff0c \u7121\u6cd5\u5728\u6211\u5011\u7d71\u8a08\u904e\u7a0b\u4e2d\u80fd\u88ab\u62ff\u4f86\u53c3\u8003\u7684\u3002\u56e0\u6b64\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u7684\u6210\u679c\u9084\u662f\u6bd4\u76e3\u7763\u5f0f\u5b78 \u4f46\u662f\u5728\u5be6\u969b\u6e2c\u8a66\u6642\u767c\u73fe\uff0c\u6709\u4e9b\u8a5e\u7684\u8a5e\u6027\u6a19\u8a18\u5229\u7528\u4e2d\u7814\u9662\u65b7\u8a5e\u7a0b\u5f0f\u57f7\u884c\u51fa\u4f86\u7684\u7d50\u679c\uff0c\u7531 \u7fd2\u6cd5\u597d\u3002\u53e6\u5916\u5c0d\u65bc WSE \u548c Google \u9019\u5169\u500b\u8a9e\u6599\u5eab\u7684\u6bd4\u8f03\u4e0a\u5247\u5404\u6709\u512a\u7f3a\u9ede\u3002WSE \u4e2d\u7684\u8cc7 \u65bc\u5206\u8a5e\u7a0b\u5f0f\u5206\u8a5e\u6216\u6a19\u8a18\u932f\u8aa4\u6216\u5176\u5b83\u539f\u56e0\uff0c\u4e26\u4e0d\u6703\u51fa\u73fe\u3002\u4f8b\u5982\uff0c\u5728\u8a9e\u6599\u5eab\u4e2d\u6709\u4e00\u985e NV \u6599\u4e0d\u4f46\u6709\u65b7\u8a5e\u9084\u6709\u8a5e\u985e\u6a19\u8a18\u975e\u5e38\u65b9\u4fbf\u62ff\u4f86\u76f4\u63a5\u4f7f\u7528\uff0c\u800c\u5f9e Google \u5f97\u5230\u7684\u5247\u662f raw \u7684\u8a5e\u5982\u96fb\u8166(NA)/\u6253\u5b57(NV4)/\u53ca(CAA)/\u6392\u7248(NV4)\uff0c\u800c\u4e2d\u7814\u9662\u65b7\u8a5e\u7cfb\u7d71\u4e26\u6c92 data\uff0c\u5fc5\u9808\u518d\u7d93\u904e\u65b7\u8a5e\u7684\u8655\u7406\uff1bWSE \u4e2d\u7684\u53e5\u5b50\u7d93\u904e\u6311\u9078\uff0c\u5f62\u5f0f\u8f03\u55ae\u7d14\uff0c\u800c\u7db2\u8def\u4e0a\u7684\u7db2 \u6709\u8fa6\u6cd5\u5373\u6642\u5224\u65b7\u51fa\u540d\u7269\u5316(NV)\u7684\u73fe\u8c61\uff0c\u56e0\u6b64\u65b7\u8a5e\u5f8c\u7684\u7d50\u679c\u70ba\u300c\u96fb\u8166(Na) \u6253\u5b57(VA) \u9801\u9084\u5305\u542b\u4e86\u95dc\u9375\u5b57\u53ef\u80fd\u5728\u6a19\u984c\u3001\u9023\u7d50\u6216\u641c\u5c0b\u5230\u7684\u7db2\u9801\u7684\u91cd\u8907\u6027\u592a\u9ad8\uff0c\u53ef\u80fd\u6709\u7684\u7279\u5b9a\u7684 \u53ca(Caa) \u6392\u7248(VA)\u300d\uff1b\u5728\u8a13\u7df4\u96c6\u4e2d\uff0c\u6709\u300c\u5169(NEU)/\u8005(NA)/\u540c\u7b49\u91cd\u8981\u300d\u9019\u6a23\u7684\u53e5 \u53e5\u5b50\uff0c\u73fe\u5728\u5f88\u6d41\u884c\u3001\u6216\u88ab\u5f88\u591a\u4eba\u5f15\u7528\u904e\uff0c\u90a3\u9ebc\u51fa\u73fe\u5728\u641c\u5c0b\u9801\u4e0a\u5f88\u591a\u7b46\u90fd\u5728\u63cf\u8ff0\u540c\u4e00\u7b46 \u5b50\uff0c\u4e2d\u7814\u9662\u7684\u7dda\u4e0a\u7a0b\u5f0f\u65b7\u8a5e\u7684\u7d50\u679c\u662f\uff0c\u300c\u5169\u8005(NH)/\u540c\u7b49\u91cd\u8981\u300d\uff0c\u986f\u73fe\u51fa\u958b\u653e\u6e2c\u8a66 \u8cc7\u6599\uff0c\u4e5f\u5c31\u662f\u540c\u4e00\u500b\u53e5\u5b50\uff0c\u800c\u4e14\u5728\u95dc\u6ce8\u8a5e\u524d\u9762\u51fa\u73fe\u7684\u8a5e\uff0c\u662f\u5728\u6211\u5011\u671f\u5f85\u4e4b\u5916 \u548c\u5c01\u9589\u6e2c\u8a66\u7684\u8a9e\u6599\u6709\u4e00\u5b9a\u7684\u5dee\u7570\u3002 \u7684 pattern\u3001\u6216\u7121\u6cd5\u8b93\u6211\u5011\u5229\u7528\u7684 pattern\uff0c\u5206\u5225\u5f62\u6210\u932f\u8aa4\u8cc7\u6599\u592a\u591a\u3001\u80fd\u7528\u7684\u8cc7\u6599\u592a\u5c11 \u7136\u800c\u9019\u500b\u5be6\u9a57\u7684\u6578\u64da\u986f\u793a\uff0c\u5229\u7528\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u8655\u7406\u4e2d\u6587 NP-chunking \u6642\uff0c\u53ea\u6709\u8a13\u7df4\u8a9e \u7684\u60c5\u6cc1\u3002\u56e0\u6b64\u7121\u6cd5\u5728\u6211\u641c\u5c0b\u7684\u7bc4\u570d\u5167\uff0c\u6709\u8db3\u5920\u7b26\u5408\u6211\u9810\u671f\u7684\u6a21\u5f0f\u7684\u8cc7\u6599\uff1b\u5728 WSE \u4e2d \u6599\u4e2d\u7684\u8cc7\u8a0a\u53ef\u4ee5\u88ab\u5229\u7528\u7684\u6642\u5019\uff0cIOE \u540d\u8a5e\u7d44\u8868\u793a\u6cd5\u3001\u7c21\u5316\u8a5e\u985e\u6a19\u8a18\u3001\u95dc\u6ce8\u8a5e\u672c\u8eab\u53ca\u524d \u7121\u6cd5\u627e\u5230\u65b0\u7684\u8a5e\u5f59\u53ca\u4e0d\u9069\u5408\u88ab\u6536\u9304\u5728\u8a9e\u6599\u5eab\u4e2d\u7684\u8a5e\u5f59\uff0c\u5f9e Google \u5f97\u5230\u7684\u8cc7\u6599\u5247\u6c92\u6709\u9019 \u5f8c\u5169\u500b\u4f4d\u7f6e\u7684\u8a5e\uff0c\u9084\u6709\u4ed6\u5011\u7684\u8a5e\u6027(pos)\u662f\u6700\u597d\u7684\u7279\u5fb5\u7d44\u5408\u3002\u4e5f\u56e0\u6b64\u6211\u5011\u56fa\u5b9a\u9019\u5e7e\u500b \u500b\u554f\u984c\u3002 \u7279\u5fb5\uff0c\u7576\u6210\u4e0b\u4e00\u500b\u5be6\u9a57\u7684\u57fa\u672c\u7279\u5fb5\u96c6\u3002 \u8868\u4e03\u662f\u9019\u500b\u5be6\u9a57\u7684\u7d50\u679c\u3002 \u8868\u4e03\u3001IOB IOE \u521d\u6b65\u958b\u653e\u6e2c\u8a66\u6bd4\u8f03\u7d50\u679c feature combination accruracy precision recall f-rate W(n-2 .. n+2) , P(n-2 .. n+2) in Simplified tagset , T(n-2 .. n-1), IOE 92.95% 86.67% 89.66% 88.14% W(n-2 .. n+2) , P(n-2 .. n+2) in Simplified tagset , T(n-2 .. n-1), IOB 88.93% 68.75% 75.86% 72.13% \u5f9e\u8868\u516d\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u770b\u51fa(1)CKIP \u8a5e\u985e\u6a19\u8a18\u7e3d\u5171\u6709\u591a\u9054\u7d04 230 \u500b\u5206\u985e\uff0c\u7c21\u5316\u6a19\u8a18\u7d04 \u6709 45 \u500b\u5206\u985e\uff0c\u56e0\u6b64\u4e0d\u8ad6\u662f\u53ea\u6709\u52d5\u8a5e\u63a1\u7528 CKIP \u7684\u6b21\u5206\u985e\u6216\u662f\u6574\u9ad4\u7684\u8a5e\u985e\u90fd\u5229\u7528\u6b21\u5206\u985e \u4f86\u6a19\u8a18\uff0c\u53ef\u80fd\u7531\u65bc CKIP \u8a5e\u985e\u5206\u9805\u592a\u7d30\u9020\u6210\u5206\u985e\u5668\u4e2d\u8cc7\u6599\u7a00\u758f\u7684\u554f\u984c\uff0c\u4f7f\u5f97\u63a1\u7528 CKIP \u8a5e\u985e\u6a19\u8a18\u7684\u8868\u73fe\u4e26\u6c92\u6709\u4ee5\u7528\u7c21\u5316\u6a19\u8a18\u7684\u8868\u73fe\u597d\uff0c\u4ee5\u53ca(2)\u96d6\u7136\u5728\u5c01\u9589\u6e2c\u8a66\u4e2d\u5169\u7a2e\u8868\u793a \u6cd5\u7cbe\u78ba\u5ea6\u4e0d\u76f8\u4e0a\u4e0b\uff0c\u5f9e\u958b\u653e\u6e2c\u8a66\u7684\u8868\u4e03\u4f86\u770b\uff0c\u4ee5 IOE \u8868\u793a\u6cd5\u7684\u7d50\u679c\u6703\u6bd4 IOB \u8868\u793a\u6cd5\u4f86 \u7684\u7cbe\u78ba\uff0c\u6240\u4ee5\u6211\u5011\u56fa\u5b9a\u9019\u5169\u7a2e\u53c3\u6578\u4f86\u9032\u884c\u4e4b\u5f8c\u7684\u5be6\u9a57(\u5305\u62ec\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5)\u3002\u5229\u7528 \u7c21\u5316\u6a19\u8a18\u7684\u53e6\u4e00\u500b\u597d\u8655\u662f\uff1a\u7576\u6211\u5011\u7528\u4e2d\u7814\u9662\u7684\u65b7\u8a5e\u7a0b\u5f0f\u5c0d\u958b\u653e\u6e2c\u8a66\u7684\u8cc7\u6599\u505a\u524d\u8655\u7406\u6642\uff0c \u5f97\u5230\u7684\u6a19\u8a18\u8207\u8a13\u7df4\u51fa\u4f86\u7684\u6a21\u578b\u4f7f\u7528\u7684\u4e00\u81f4\u3002\u7531\u65bc\u6211\u5011\u5f9e\u9019\u500b\u958b\u653e\u6e2c\u8a66\u7684\u7d50\u679c\u767c\u73fe\uff1a 1. \u521d\u59cb\u6a21\u578b\u5c0d\u9577\u8a5e\u7684\u5075\u6e2c\u4e0d\u592a\u654f\u611f\uff1a\u8a9e\u6599\u4e2d\u6709\"\u4e00\u540d\u9a0e\u6a5f\u8eca\u7684\u5e74\u8f15\u4eba\"\u53ca\"\u4e00\u540d \u9ad8\u7d1a\u5b98\u54e1\"\u3002\u9019\u5169\u53e5\u4e2d\u7684\u300c\u4e00\u540d\u300d\u90fd\u662f\u67d0\u500b NP \u7684\u4e00\u90e8\u5206\uff0c\u4f46\u662f\u524d\u8005\u7684\u8a5e\u7d44\u6a19 \u8a18\u70ba\u4e00 O/ \u540d O \u800c\u5f8c\u8005\u70ba\u4e00 I / \u540d I\u3002\u56e0\u6b64\u91cf\u8a5e\u5f8c\u9762\u88ab tagging \u904e\u7a0b\u8003\u616e\u9032\u4f86\u7684 model 4 F, Wi-1.. Wi+1, Pi-1..Pi+1 Ti-2..Ti-1 model 5 F, Wi, Pi-2..Pi+2 Ti-2..Ti-1 model 6 B, Wi-2.. Wi+2, Pi-2..Pi+2 Ti-2..Ti-1 model 7 F, Pi-2..Pi+2 Ti-2..Ti-1 \u8868\u5341\u662f\u5404\u500b\u6a21\u578b\u958b\u653e\u6e2c\u8a66\u7684\u7d50\u679c\uff0c\u5f9e\u958b\u653e\u6e2c\u8a66\u7684\u8f38\u51fa\u4f86\u770b\u6bcf\u500b\u6a21\u578b\u90fd\u6709\u660e\u986f\u4e0d \u8db3\u7684\u5730\u65b9\uff0c\u300c\u76e3\u5bdf\u4eba\u54e1\u300d\u9019\u7a2e\u578b\u614b\u7684\u8a5e\u66f4\u6c92\u6709\u4e00\u500b\u6a21\u578b\u5224\u65b7\u6b63\u78ba\u3002model 2 \u9664\u4e86\u8a9e\u6599\u5eab \u7684\u5167\u90e8\u8a0a\u606f\u4e4b\u5916\uff0c\u52a0\u5165\u4e86\u6bcf\u500b\u8a5e\u5728 HowNet \u4e2d\u7684\u7fa9\u5143(\u53ef\u4ee5\u8996\u70ba\u8a9e\u610f\u7279\u5fb5\u6216\u985e\u5225)\u7576 \u505a\u4e00\u500b\u7279\u5fb5\uff0c\u96d6\u7136\u6bd4 model 1 \u7684 F-rate \u597d\u4e0a\u7d04 0.8 \u500b\u767e\u5206\u6bd4\uff0c\u8a13\u7df4\u6a21\u578b\u7684\u6642\u9593\u537b\u4e5f\u589e\u52a0 \u70ba\u7d04 1.8 \u500d\uff1bmodel 3 \u53c3\u8003\u524d\u56db\u500b\u8a5e\u53ca\u5f8c\u5169\u500b\u8a5e\uff0c\u96d6\u7136\u5c0d\u4e0a\u9762\u63d0\u5230\u5305\u542b\u91cf\u8a5e\u7684\u9577\u540d\u8a5e\u7d44 \u5408\u6709\u4e9b\u5e6b\u52a9\uff0c\u4f46\u662f\u5c0d\u5176\u4ed6\u554f\u984c\u6c92\u6709\u592a\u5927\u7684\u5f71\u97ff\uff1bbackward parsing(model6) \u5728 closed test \u7684\u90e8\u4efd\uff0c\u96d6\u7136\u6709\u6700\u9ad8\u7684 F-rate, \u4f46\u662f\u5728\u958b\u653e\u6e2c\u8a66\u7684\u90e8\u4efd\u537b\u6c92\u6709\u7279\u5225\u597d\u7684\u8868\u73fe\u3002 \u8868\u5341\u3001\u5404\u6a21\u578b\u958b\u653e\u6e2c\u8a66\u6bd4\u8f03\u7d50\u679c model tag accuracy(%) precision(%) recall(%) F-rate(%) model 1 92.95 86.67 89.66 \u4e94\u3001\u7d50\u8ad6\u53ca\u672a\u4f86\u5c55\u671b (\u516d)\u5be6\u9a57\uff1a\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5 Semi-supervised learning \u6211\u5011\u5229\u7528\u4e0d\u540c\u7684\u7279\u5fb5\u4e26\u5c07\u539f\u672c\u7684\u6a21\u578b\u52a0\u4ee5\u6539\u5584\u904e\u5f8c\uff0c\u5229\u7528\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u5728\u5c0f\u578b\u958b \u6211\u5011\u5c0d\u9019\u500b\u65b9\u6cd5\u5206\u5225\u505a\u4e86\u5c01\u9589\u53ca\u958b\u653e\u6e2c\u8a66\u3002\u7531\u65bc\u6b64\u8655\u8457\u91cd\u5728\u6539\u9032\u8fa8\u8b58\u540d\u7269\u5316\u7684\u52d5\u8a5e\uff0c \u653e\u6e2c\u8a66\u6709 70\uff05\u7684 f-rate\uff0c\u4f46\u5728\u5229\u7528\u672a\u6a19\u8a18\u904e\u7684\u7db2\u9801\u7576\u4f5c\u7279\u5fb5\u4e4b\u5f8c\uff0c\u5c07\u6a21\u578b\u7684 f-rate \u63d0\u6607 \u9019\u90e8\u4efd\u7684\u958b\u653e\u6e2c\u8a66\u8a9e\u6599\u9664\u4e86\u4e0a\u500b\u5be6\u9a57\u7684\u6e2c\u8a66\u8a9e\u6599\uff0c\u6211\u5011\u9084\u52a0\u5165\u5305\u542b\u4e0d\u66fe\u51fa\u73fe\u5728\u8a13\u7df4\u8a9e \u81f3 78.79%\uff0c\u6bd4\u539f\u672c\u9ad8\u51fa\u4e86 8.79\uff05\u3002\u96d6\u7136\u6a21\u578b\u7684 performance \u9084\u662f\u6703\u53d7\u5230\u8a13\u7df4\u8a9e\u6599\u7684\u5f71 \u6599\u4e2d\u7684\u540d\u7269\u5316\u52d5\u8a5e\u8a5e\u7d44\u7684\u53e5\u5b50\uff0c\u4f54\u5168\u90e8\u6e2c\u8a66\u8a9e\u6599\u7684 50%\u3002\u8868\u5341\u4e00\u662f\u9019\u500b\u65b9\u6cd5\u7684\u5be6\u9a57\u7d50 \u97ff\u4f7f\u5f97\u7d50\u679c\u4e0d\u7a69\u5b9a\uff0c\u5076\u723e\u6703\u6709\u8207\u9810\u671f\u4e0d\u7b26\u5408\u7684\u60c5\u5f62\u767c\u751f\uff0c\u6574\u9ad4\u4f86\u8aaa\uff0c\u6211\u5011\u63d0\u51fa\u7684\u5229\u7528 \u679c\uff0csupervised II \u662f\u5c07\u8a5e\u7684\u524d\u4e00\u500b\u8a5e\u6027\u55ae\u7368\u62ff\u51fa\u62ff\u4f5c\u70ba\u4e00\u500b\u7279\u5fb5\u3002\u5728\u5c01\u9589\u6e2c\u8a66\u4e2d\uff0c\u534a\u76e3 \u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u65b9\u6cd5\u5584\u7528\u4e86\u7db2\u8def\u4e0a\u96a8\u624b\u53ef\u5f97\u7684\u8cc7\u6e90\u4e26\u4e14\u7684\u78ba\u589e\u9032\u540d\u8a5e\u8fa8\u8b58\u7684\u6548\u679c\u3002 \u7763\u5f0f\u5b78\u7fd2\u6cd5\u5be6\u9a57\u7d50\u679c\u7684 F-rate \u70ba 85.46\uff05\uff0c\u96d6\u7136\u53ea\u6bd4\u76e3\u7763\u5f0f\u5b78\u7fd2\u4f5c\u6cd5\u7684\u9ad8\u51fa\u4e86 0.2\uff05\u3002 \u672c\u7bc7\u8ad6\u6587\u7684\u8ca2\u737b\u5728\u65bc\uff1a \u4f46\u5728\u958b\u653e\u6e2c\u8a66\u7684\u90e8\u4efd\uff0c\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u660e\u986f\u6bd4\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u7684 F-rate \u9ad8\u51fa\u4e86 8.79\uff05\u4e4b (1)\u96d6\u7136\u8a5e\u7d44\u8fa8\u8b58\u4e00\u76f4\u662f\u8a31\u591a\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u8b70\u984c\u4e2d\u7684\u91cd\u8981\u6b65\u9a5f\uff0c\u4f46\u5728\u4e2d\u6587\u65b9\u9762\u4e26 \u591a\u3002\u7576\u6211\u5011\u53ea\u900f\u904e\u6b64\u5be6\u9a57\u7684\u4f5c\u6cd5\u63cf\u8ff0\u4e2d\u5229\u7528\u672a\u6a19\u8a18\u8cc7\u6599\u5224\u5225\u5c01\u9589\u6e2c\u8a66\u88e1 VN Pair \u4e2d\u7684 \u6c92\u6709\u770b\u5230\u592a\u591a\u76f8\u95dc\u7684\u7814\u7a76\u3002\u7531\u65bc\u4e2d\u6587\u7684\u540d\u8a5e\u7d44\u7d50\u69cb\u76f8\u8f03\u65bc\u5176\u5b83\u8a9e\u8a00\u7684\u540d\u8a5e\u7d44\u7d50\u69cb\u90fd\u8981 \u52d5\u8a5e\u985e\u5225\u6642\u7d04\u6709\u516b\u6210\u7684\u6b63\u78ba\u7387(\u5305\u542b\u932f\u8aa4\u7684\u88ab\u4fee\u6b63\u4ee5\u53ca\u672c\u4f86\u6b63\u78ba\u7684\u7dad\u6301\u6b63\u78ba)\uff0c\u4f46\u662f \u8907\u96dc\u7684\u591a\uff0c\u56e0\u6b64\u672c\u6587\u53ea\u5c08\u6ce8\u65bc\u540d\u8a5e\u7d44\u7684\u8fa8\u8b58\uff0c\u4e26\u4e14\u8b49\u660e\u4e4b\u524d\u7528\u5728\u985e\u4f3c\u4f5c\u6cd5\u4ee5\u53ca\u5728\u5176\u4ed6 \u900f\u904e\u5206\u985e\u5668\u7684\u9810\u6e2c\u7d50\u679c\u4e4b\u5f8c\uff0cVN pairs \u9019\u90e8\u4efd\u537b\u53ea\u6709\u7d04 5\uff05\u7684\u6539\u5584\uff0c\u8981\u5982\u4f55\u6709\u6548\u61c9\u7528\u5206 \u8a9e\u8a00\u4e2d\u88ab\u666e\u904d\u63a1\u7528\u7684\u7279\u5fb5\uff0c\u4e26\u4e0d\u5b8c\u5168\u9069\u5408\u7528\u5728\u4e2d\u6587\u8a9e\u8a00\u4e0a\u3002\u800c\u6211\u5011\u4e5f\u627e\u5230\u4e86\u66f4\u9069\u5408\u7684 \u985e\u5668\u4f86\u7a81\u986f\u51fa\u65b0\u627e\u5230\u7684\u7279\u5fb5\u7684\u91cd\u8981\u6027\u8457\u5be6\u70ba\u4e00\u500b\u8b70\u984c\uff1b\u7531\u65bc\u8f49\u70ba\u540d\u8a5e\u7684\u52d5\u8a5e\u5728\u6a39\u5eab\u5716 \u7279\u5fb5\u3002 \u4e2d\u7684\u8a5e\u985e\u6a19\u8a18\u70ba NV \u6240\u4ee5\u5c01\u9589\u6e2c\u8a66\u8a9e\u6599\u4e2d\u4e26\u4e0d\u9700\u8981\u8003\u616e\u8868\u9762\u4e0a\u70ba\u540d\u8a5e\u63a5\u52d5\u8a5e\u7684\u60c5\u5f62\u3002 (2)\u6211\u5011\u63d0\u51fa\u4e00\u500b\u7c21\u55ae\u7684\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\uff0c\u6539\u5584\u4e86\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u4e2d\u8cc7\u6599\u7a00\u758f (data \u958b\u653e\u6e2c\u8a66\u4e2d\u6709 21 \u500b\u540d\u7269\u5316\u52d5\u8a5e\u7684\u8a5e\u7d44\uff0c\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u4e2d\u6b63\u78ba\u5224\u65b7\u51fa\u516d\u7d44\uff0c\u800c\u534a\u76e3\u7763\u5f0f sparseness) \u53ca\u53ea\u4f9d\u8cf4\u8a13\u7df4\u8a9e\u6599\u6642\u7121\u6cd5\u89e3\u6c7a\u7684\u554f\u984c\u3002\u4e26\u4e14\u8ddf\u539f\u672c\u7684 chunker \u76f8\u6bd4\u4e4b\u4e0b\u63d0\u9ad8 \u5b78\u7fd2\u6cd5\u5224\u65b7\u51fa\u4e5d\u7d44\uff0c\u5269\u9918\u7684 12 \u7d44\u4e2d\u6709 7 \u7d44\u5728\u4f5c\u6cd5\u63cf\u8ff0\u4e2d\u5229\u7528\u672a\u6a19\u8a18\u8cc7\u6599\u5224\u65b7\u52d5\u8a5e\u985e\u5225 \u4e0d\u5c11\u6e96\u78ba\u5ea6\u53ca\u5be6\u7528\u6027\u3002\u5c0d\u65bc\u4e00\u4e9b\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u4e2d\uff0c\u9700\u8981\u5229\u7528\u8f03\u9ad8\u6bd4\u4f8b\u7684\u540d\u8a5e\u7d44\u7684\u61c9\u7528\uff0c \u6642\u7684\u5206\u985e\u662f\u6b63\u78ba\u7684\uff0c\u4f46\u7d93\u904e\u5206\u985e\u5668\u5206\u985e\u4e4b\u5f8c\u8b8a\u6210\u8aa4\u5224\u7684\u60c5\u5f62\u3002 \u4f8b\u5982\uff1a\u53e5\u5b50\u5256\u6790 parsing\u3001\u8a9e\u610f\u89d2\u8272 semantic role labeling\u3001\u6587\u4ef6\u5206\u985e text categorization 88.14 model 2 93.43 87.67 90.63 \u8868\u5341\u4e00\u3001\u76e3\u7763\u5f0f\u53ca\u534a\u76e3\u7763\u5f0f\u5b78\u7fd2\u6cd5\u5be6\u9a57\u7d50\u679c \u7b49\u7b49\uff0c\u90fd\u6709\u5be6\u8cea\u4e0a\u7684\u5e6b\u52a9\u3002 88.97 model 3 91.95 83.37 87.93 85.59 model 4 91.14 86.19 88.26 87.21 Tag accuracy Precision Recall \u672a\u4f86\u6211\u5011\u5e0c\u671b\u80fd\u5920\u627e\u5230\u66f4\u597d\u7684\u7279\u5fb5\uff0c\u52a0\u4e0a\u76e3\u7763\u5f0f\u6a5f\u5668\u5b78\u7fd2\u7684\u65b9\u5f0f\uff0c\u4ee5\u89e3\u6c7a\u66f4\u9577\u7684 F-rate \u5c01\u9589\u6e2c\u8a66 \u8907\u5408\u8a5e\u7d44\u4ee5\u53ca\u5305\u542b\u6709\u5169\u500b\u4ee5\u4e0a\u540d\u7269\u5316\u52d5\u8a5e\u7684\u8907\u5408\u8a5e\u7d44\u3002\u53e6\u5916\uff0c\u7531\u65bc\u7db2\u8def\u4e0a\u5927\u91cf\u672a\u6a19\u8a18 \u904e\u7684\u8cc7\u6599\u96a8\u624b\u53ef\u5f97\uff0c\u56e0\u6b64\u6211\u5011\u4e5f\u5e0c\u671b\u80fd\u63d0\u51fa\u975e\u76e3\u7763\u5f0f(unsupervised)\u7684\u6f14\u7b97\u6cd5\uff0c\u4ee5\u7a81\u7834 } \u8a5e\u6027\u8b8a\u5f97\u5341\u5206\u91cd\u8981\u3002\u7531\u65bc\u6211\u5011\u521d\u59cb\u6a21\u578b\u53ea\u5c07\u524d\u5f8c\u5404\u5169\u500b\u8a5e\u52a0\u5165\u7279\u5fb5\u96c6\uff0c\u5728\u958b model 5 89.95 81.97 81.14 81.55 supervised 92.06% 84.65% 86.28% 85.46% \u53d7\u9650\u65bc\u5c11\u91cf\u4eba\u5de5\u6a19\u8a18\u904e\u8a13\u7df4\u8a9e\u6599\u7684\u9650\u5236\u3002 |
",
"num": null
}
}
}
}