| { |
| "paper_id": "O18-1005", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T08:09:46.841104Z" |
| }, |
| "title": "\u674e\u5f65\u7487 Yen-Hsuan Lee", |
| "authors": [ |
| { |
| "first": "Yih-Ru", |
| "middle": [], |
| "last": "\u738b\u9038\u5982", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "yrwang@cc.nctu.edu.tw" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "", |
| "pdf_parse": { |
| "paper_id": "O18-1005", |
| "_pdf_hash": "", |
| "abstract": [], |
| "body_text": [ |
| { |
| "text": "Through different famous neural network structure , such as multilayer perceptron and recurrent neural network along with traditional Chinese segmentation, we are able to construct an efficient parser that could transfer a raw traditional Chinese sentences to a dependency tree. We have much better result than similar task that has been proposed at 2012. ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "http://www.aclclp.org.tw/use_conll_c.php", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Automatic Semantic Role Assignment for a Tree Structure", |
| "authors": [ |
| { |
| "first": "Jia-Ming", |
| "middle": [], |
| "last": "You", |
| "suffix": "" |
| }, |
| { |
| "first": "Keh-Jiann", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Proceedings of SIGHAN workshop", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "You, Jia-Ming, Keh-Jiann Chen, 2004, \"Automatic Semantic Role Assignment for a Tree Structure\", Proceedings of SIGHAN workshop.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Efficient Estimation of Word Representations in Vector Space", |
| "authors": [ |
| { |
| "first": "Tomas", |
| "middle": [], |
| "last": "Mikolov", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Kai", |
| "suffix": "" |
| }, |
| { |
| "first": "Greg", |
| "middle": [], |
| "last": "Corrado", |
| "suffix": "" |
| }, |
| { |
| "first": "Jeffrey", |
| "middle": [], |
| "last": "Dean", |
| "suffix": "" |
| } |
| ], |
| "year": null, |
| "venue": "Proceedings of Workshop at ICLR 2013", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Mikolov, Tomas, Chen, Kai, Corrado, Greg, and Dean, Jeffrey. \"Efficient Estimation of Word Representations in Vector Space \", In Proceedings of Workshop at ICLR 2013", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Universal Dependencies v1: A Multilingual Treebank Collection", |
| "authors": [ |
| { |
| "first": "Joakim", |
| "middle": [], |
| "last": "Nivre", |
| "suffix": "" |
| }, |
| { |
| "first": "Marie-Catherine", |
| "middle": [], |
| "last": "De Marneffe", |
| "suffix": "" |
| }, |
| { |
| "first": "Filip", |
| "middle": [], |
| "last": "Ginter", |
| "suffix": "" |
| }, |
| { |
| "first": "Yoav", |
| "middle": [], |
| "last": "Goldberg", |
| "suffix": "" |
| }, |
| { |
| "first": "Jan", |
| "middle": [], |
| "last": "Haji\u010d", |
| "suffix": "" |
| }, |
| { |
| "first": "Christopher", |
| "middle": [ |
| "D" |
| ], |
| "last": "Manning", |
| "suffix": "" |
| }, |
| { |
| "first": "Ryan", |
| "middle": [], |
| "last": "Mcdonald", |
| "suffix": "" |
| }, |
| { |
| "first": "Slav", |
| "middle": [], |
| "last": "Petrov", |
| "suffix": "" |
| }, |
| { |
| "first": "Sampo", |
| "middle": [], |
| "last": "Pyysalo", |
| "suffix": "" |
| } |
| ], |
| "year": 2016, |
| "venue": "Proceedings of LREC", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter, Yoav Goldberg, Jan Haji\u010d, Christopher D. Manning, Ryan McDonald, Slav Petrov, Sampo Pyysalo, Natalia Silveira, Reut Tsarfaty, Daniel Zeman, 2016, \"Universal Dependencies v1: A Multilingual Treebank Collection\". In Proceedings of LREC.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "An Improvement on Chinese Parser", |
| "authors": [ |
| { |
| "first": "Yih-Ru", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Yih-Ru Wang, 2003, \"An Improvement on Chinese Parser\"", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "An Efficient Algorithm for Projective Dependency Parsing", |
| "authors": [], |
| "year": 2003, |
| "venue": "Proceedings of the 8th International Workshop on Parsing Technologies", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Joakim Nivre , 2003, \"An Efficient Algorithm for Projective Dependency Parsing\" In Proceedings of the 8th International Workshop on Parsing Technologies", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "A Fast and Accurate Dependency Parser using Neural Networks", |
| "authors": [ |
| { |
| "first": "Danqi", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| }, |
| { |
| "first": "Christopher", |
| "middle": [ |
| "D" |
| ], |
| "last": "Manning", |
| "suffix": "" |
| } |
| ], |
| "year": 2014, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Danqi Chen, Christopher D. Manning , 2014, \"A Fast and Accurate Dependency Parser using Neural Networks\" in EMNLP", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Transition-Based Dependency Parsing with Stack Long Short-Term Memory", |
| "authors": [ |
| { |
| "first": "Chris", |
| "middle": [], |
| "last": "Dyer", |
| "suffix": "" |
| }, |
| { |
| "first": "Miguel", |
| "middle": [], |
| "last": "Ballesteros", |
| "suffix": "" |
| }, |
| { |
| "first": "Wang", |
| "middle": [], |
| "last": "Ling", |
| "suffix": "" |
| }, |
| { |
| "first": "Austin", |
| "middle": [], |
| "last": "Matthews", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Noah", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Smith", |
| "suffix": "" |
| } |
| ], |
| "year": 2016, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Chris Dyer, Miguel Ballesteros, Wang Ling, Austin Matthews, Noah A. ,Smith, 2016, \"Transition-Based Dependency Parsing with Stack Long Short-Term Memory\"", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "Traditional Chinese Parsing Evaluation at SIGHAN Bake-offs", |
| "authors": [ |
| { |
| "first": "", |
| "middle": [], |
| "last": "Yuen-Hsien", |
| "suffix": "" |
| }, |
| { |
| "first": "Lung-Hao", |
| "middle": [], |
| "last": "Tseng", |
| "suffix": "" |
| }, |
| { |
| "first": "Liang-Chih", |
| "middle": [], |
| "last": "Lee", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Yu", |
| "suffix": "" |
| } |
| ], |
| "year": 2012, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Yuen-Hsien Tseng, Lung-Hao Lee and Liang-Chih Yu, 2012, \"Traditional Chinese Parsing Evaluation at SIGHAN Bake-offs 2012\"", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "TABREF0": { |
| "html": null, |
| "content": "<table><tr><td>Recurrent Neural Network \u4e00\u3001 \u7dd2\u8ad6 \u5728\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u4e2d\uff0c\u8a9e\u610f\u77ad\u89e3\u662f\u5f88\u91cd\u8981\u7684\u4e00\u90e8\u5206\uff0c\u800c\u70ba\u4e86\u505a\u8a9e\u610f\u77ad\u89e3\uff0c\u5c31\u5fc5\u9808\u505a\u53e5\u6cd5 \u5206\u6790\u3002\u65e9\u671f\u5b78\u8005\u5011\u90fd\u4f7f\u7528\u6a39\u72c0\u7d50\u69cb\u53e5\u6cd5\u5206\u6790(tree parsing)\uff0c\u5982\u4e2d\u7814\u9662\u7684 CKIP Chinese Parser[1]\u3002\u6a39\u72c0\u7d50\u69cb\u53c8\u7a31\u77ed\u8a9e\u7d50\u69cb(phrase structure)\uff0c\u4e3b\u8981\u662f\u5c07\u53e5\u5b50\u62c6\u89e3\u6210\u4e0d\u540c\u7684\u7247 \u8a9e\u3002\u8fd1\u5e74\u4f86\uff0c\u5b78\u8005\u5011\u5927\u591a\u4f7f\u7528\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\uff0c\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u4e2d\u7684\u8cc7\u8a0a\u53ef\u7531\u77ed\u8a9e\u7d50\u69cb\u4f9d \u7cfb\u7d71\u3002\u7136\u800c\u4e2d\u7814\u9662\u4e5f\u63a8\u51fa\u4e86\u81ea\u5df1\u7684\u4e2d\u6587\u7d50\u69cb\u6a39\u8cc7\u6599\u5eab 1 \uff0c\u4e26\u4e14\u662f\u4ee5 dependency tree \u7684\u65b9 \u5f0f\u5448\u73fe\uff0c\u65b7\u8a5e\u8207 POS \u6a19\u793a\u90fd\u8207\u4e2d\u7814\u9662\u8a5e\u5eab\u5c0f\u7d44\u4e00\u81f4\uff0c\u4e14\u4e2d\u7814\u9662\u7684\u8a5e\u6027\u6a19\u8a18\u7a2e\u985e\u591a\u9054 47 \u985e\uff0c\u5176\u4e2d P(\u4ecb\u8a5e)\u66f4\u53ef\u7d30\u5206\u6210 66 \u985e\uff0c\u800c\u9019\u4e9b\u80fd\u63d0\u4f9b\u53e5\u6cd5\u5206\u6790\u66f4\u591a\u8cc7\u8a0a\uff0c\u4f7f\u5206\u6790\u6e96 \u78ba\u5927\u70ba\u63d0\u5347\uff0c\u4ee5\u4e0b\u5716\u4e00\u548c\u5716\u4e8c\u300c\u8207\u300d\u70ba\u4f8b\uff0c\u82e5\u6c92\u6709\u6a19\u793a POS \u5728\u53e5\u6cd5\u5224\u65b7\u4e0a\uff0c\u5c31\u7121\u6cd5\u77e5 \u9053\u4ed6\u4ee3\u8868\u7684\u662f\u4f34\u96a8\u9084\u662f\u5c0d\u7b49\u7684\u610f\u601d\uff0c\u5982\u5716\u4e00\u300c\u8207\u300dPOS \u70ba P35 \u4ee3\u8868\u7684\u662f\u4f34\u96a8\uff0c\u524d\u9762\u53ef \u4ee5\u662f\u4eba\u6216\u5176\u4ed6\u4e8b\u7269\uff1b\u800c\u5716\u4e8c\u7684\u300c\u8207\u300d\u70ba Caa\uff0c\u4ee3\u8868\u524d\u5f8c\u5fc5\u9808\u70ba\u5c0d\u7b49\u7684\u540d\u8a5e\u7247\u8a9e\u6216\u662f\u52d5 \u8a5e\uff0c\u7136\u800c\u5728\u65b7\u8a5e\u4e0a\uff0c\u4e26\u7121\u6cd5\u5206\u51fa\u9019\u5169\u8005\u7684\u4e0d\u540c\uff0c\u6545 POS \u5728\u8a13\u7df4\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u5668\u4e2d\u662f\u500b (\u4e00) \u7cfb\u7d71\u4ecb\u7d39 \u672c\u7814\u7a76\u4e2d\u4f7f\u7528\u4e86\u4ea4\u901a\u5927\u5b78\u8a9e\u97f3\u8655\u7406\u5be6\u9a57\u5ba4\u7684\u65b7\u8a5e\u5668\u8207 POS \u6a19\u793a\uff0c\u65b7\u8a5e\u5668\u662f\u7e41\u9ad4\u4e2d \u6587\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u4e0a\u5f88\u91cd\u8981\u7684\u4e00\u500b\u5143\u4ef6\uff0c\u4e26\u4e0d\u50cf\u5176\u4ed6\u8a9e\u7cfb\uff0c\u4e2d\u6587\u53e5\u5b50\u4e2d\u6c92\u6709\u660e\u78ba\u7684\u8a5e\u8a9e \u754c\u7dda\uff0c\u9700\u8981\u6709\u500b\u660e\u78ba\u7684\u6a19\u6e96\u624d\u80fd\u8b93\u5f8c\u7e8c\u7684\u4f7f\u7528\u66f4\u70ba\u9748\u6d3b\uff0c\u800c POS \u6a19\u793a\u4e5f\u5341\u5206\u4f9d\u8cf4\u65b7\u8a5e \u5668\u7684\u7d50\u679c\u3002\u900f\u904e\u9ad8\u6548\u80fd\u7684\u65b7\u8a5e\u5668\u8207 POS \u6a19\u793a\uff0c\u6211\u5011\u5728\u591a\u6b21 share task \u4e2d\u7372\u5f97\u4eae\u773c\u7684\u6210 \u7e3e\uff0c\u56e0\u6b64\u6211\u5011\u6709\u5341\u5206\u5805\u56fa\u7684\u57fa\u790e\u4f86\u5efa\u7acb\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u5668\u3002 Push ROOT into stack Push one word from buffer into stack POP out sencond word on stack Check dependency Check final state Left arc No relation Buffer not empty and ROOT is not only word on stack Parsing finished Right arc POP out first word on stack \u7531 \u56db\u6a13 \u2f9b\u4e0a P06 Nc VC manner agent \u4e94\u6a13 agent Nc \u5728\u5716\u4e5d\u7684\u67b6\u69cb\u4e2d\uff0c\u6211\u5011\u5c07 features \u5206\u6210\u4e09\u90e8\u5206\uff0c\u8a5e\u3001\u8a5e\u6027\u548c Label\uff0c\u5728\u8f38\u5165\u8a5e\u5411\u91cf \u7684\u6642\u5019\uff0c\u53ef\u80fd\u6703\u51fa\u73fe OOV (Out of Vocabulary)\uff0c\u56e0\u6b64\u5c07 POS \u6a19\u8a18\u8207\u8a5e\u5411\u91cf\u5408\u4f75\u505a\u70ba\u65b0 \u7684\u8a5e\u5411\u91cf\u4f86\u4f5c\u8a13\u7df4\uff0c\u800c POS \u7368\u7acb\u51fa\u4f86\u8a13\u7df4\u7684\u76ee\u7684\uff0c\u5247\u662f\u70ba\u4e86\u66f4\u6709\u6548\u7684\u5b78\u7fd2\u5b83\u5728\u4f9d\u5b58\u53e5 \u6cd5\u67b6\u69cb\u7684\u91cd\u8981\u6027\uff0c\u5728\u5f8c\u9762\u7684\u4e00\u4e9b\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u4e5f\u767c\u73fe POS \u5728\u6b64\u6a21\u578b\u4e2d\uff0c\u6709\u8457\u8209\u8db3\u8f15\u91cd \u7684\u5730\u4f4d\u3002 \u5347\u6536\u6582\u901f\u5ea6\u3002\u4e14\u5728\u8cc7\u6599\u91cf\u4e0d\u5920\u7684\u72c0\u614b\u4e0b\uff0cLSTM \u5b78\u7fd2\u6548\u679c\u4e26\u7121\u6cd5\u5c55\u73fe\u51fa\u4f86\u3002\u5728\u5f8c\u9762\u7ae0 \u7bc0\u6703\u76f4\u63a5\u7a31\u6211\u5011\u7684\u6a21\u578b\u70ba stack-GRU\u3002 3. \u5b9a\u5411\u641c\u7d22 Beam search Transition-based \u5728\u524d\u9762\u63d0\u5230\u662f\u500b\u8a08\u7b97\u4e0a\u975e\u5e38\u6709\u6548\u7387\u7684\u6f14\u7b97\u6cd5\uff0c\u53ea\u8981\u4f9d\u64da stack \u8207 buffer \u7684\u72c0\u614b\u63d0\u4f9b\u9078\u64c7\uff0c\u4e26\u4e00\u6b65\u4e00\u6b65\u5256\u6790\u5b8c\u6574\u53e5\u8a71\uff0c\u7136\u800c\u9019\u6a23\u7684\u505a\u6cd5\uff0c\u6709\u500b\u6975\u5927\u7684\u7f3a\u9ede\uff0c\u4e00 \u4f46\u505a\u4e86\u6c7a\u5b9a\uff0c\u4fbf\u7121\u6cd5\u56de\u982d\uff0c\u5373\u4f7f\u5f8c\u4f86\u767c\u751f\u96e3\u4ee5\u5224\u65b7\u7684\u72c0\u614b\u4e5f\u7121\u80fd\u70ba\u529b\u3002\u56e0\u6b64\uff0c\u6211\u5011\u5e0c \u671b\u80fd\u6709\u7cfb\u7d71\u5730\u591a\u770b\u5176\u4ed6\u9078\u9805\uff0c\u4e26\u5f9e\u4e2d\u627e\u51fa\u6700\u9069\u5408\u7684\u53e5\u6cd5\u3002\u800c\u5b9a\u5411\u641c\u7d22\u5247\u4f7f\u7528\u4e86\u6a6b\u5411\u512a \u4e09\u3001 \u7d50\u679c\u8207\u8a0e\u8ad6 (\u4e00) \u5be6\u9a57\u914d\u7f6e \u5728\u8a55\u4f30\u6a19\u6e96\u4e0a\uff0c\u4f9d\u5b58\u53e5\u6cd5\u8f03\u5e38\u4f7f\u7528 UAS( Unlabeled Attachment Score )\u53ca LAS ( Labeled Attachment Score )\uff0c\u4f86\u78ba\u8a8d head-dependent \u662f\u5426\u6b63\u78ba\u914d\u5c0d\uff0cLabeled \u5247\u66f4\u9032\u4e00 \u800c\u9019\u88e1\u7684 precision(\u6e96\u78ba\u7387)\u662f\u4ee5\u6a19\u6e96\u7b54\u6848\u7576\u4f5c\u5206\u6bcd\u4f86\u8a08\u7b97\uff0c\u5982\u4e0b\u5716\u6b63\u78ba\u6307\u5411\u7684\u7bad\u982d \u70ba 4 0.615\u3002 \u8868 \u4e8c\u4e4b\u4e00\u30012012 Traditional Chinese Parsing task \u8cc7\u6599\u96c6 \u8868 \u4e09\u4e4b\u4e8c\u3001\u4e2d\u7814\u9662\u63d0\u4f9b\u7684\u8cc7\u6599\u96c6 \u5fb5\u53ef\u4ee5\u88ab\u64f7\u53d6\u51fa\u4f86\u3002\u5404\u81ea\u96b1\u85cf\u5c64\u7684\u8f38\u51fa\u6703\u8207 stack \u8f38\u51fa\u5f62\u5f0f\u3002 \u6b65\u5730\u8868\u793a head-dependent \u7684\u95dc\u806f\u3002\u5982\u5716\u5341\u4e09\u8207\u5341\u56db\uff0c\u53ea\u770b\u7bad\u982d\u4e0d\u770b\u6a19\u7c64\u53ef\u4ee5\u770b\u51fa 7 \u689d \u5716 \u4e94\u3001\u8868\u4e00\u7684\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u6a39 theme \u5148\u641c\u5c0b\u7684\u7b56\u7565\u4ee5\u968e\u5c64\u5f0f\u7684\u65b9\u6cd5\u7be9\u9078\u4f7f\u6bcf\u4e00\u5c64\u53ea\u9700\u8981\u8655\u7406\u56fa\u5b9a\u5bec\u5ea6(beam width)\u7684\u8cc7 \u95dc\u806f\u4e2d\u53ea\u6709 4 \u689d\u6b63\u78ba\uff0c\u800c\u5e36\u6709\u6a19\u7c64\u7684\u60c5\u6cc1\u5247\u53ea\u6709 3 \u689d\u6b63\u78ba\uff0c\u56e0\u6b64\u9019\u500b\u7bc4\u4f8b\u4e2d\uff0cUAS \u70ba Data Set #Word #Sent Avg. Length Training 337,174 56,957 \u63a5\u8457\u63a2\u8a0e\u52a0\u5165\u5b9a\u5411\u641c\u7d22\u7684\u60c5\u6cc1\uff0c\u5728\u9019\u88e1\u70ba\u4e86\u6e1b\u5c11\u904b\u7b97\u91cf\u6211\u5011\u8a2d\u7f6e\u7684\u5b9a\u5411\u5bec\u5ea6 5.92 (\u4e8c) \u6a21\u578b\u8a2d\u8a08 quantifier \u6599\u3002\u5957\u7528\u5b9a\u5411\u641c\u7d22\u81f3 transition-based \u5256\u6790\u4e0a\uff0c\u6211\u5011\u9700\u8981\u7a0d\u5fae\u4fee\u6539\u6211\u5011\u539f\u672c\u7684\u6f14\u7b97\u6cd5\uff0c 4/7 \u800c LAS \u70ba 3/7\uff0c\u5118\u7ba1\u6211\u5011\u53ef\u4ee5\u770b\u5230\u4fee\u98fe\u300c\u9019\u5834\u300d\u7684\u95dc\u806f\u662f\u300cquantifier\u300d\u4f46\u4fee\u98fe\u5b83\u7684 Test 5,160 690 7.47 (Beam width)\u70ba 10\uff0c\u800c\u6bcf\u6b21\u8f38\u51fa\u5247\u53d6\u524d 4 \u9ad8\uff0c\u4f86\u5206\u6790\u53ef\u80fd\u7684\u53e5\u6cd5\u6a39\uff0c\u56e0\u70ba\u5b9a\u5411\u5bec\u5ea6</td></tr><tr><td>\u7167\u898f\u5247\u6cd5\u8f49\u63db\u800c\u4f86\uff0c\u5118\u7ba1\u9019\u5169\u7a2e\u65b9\u5f0f\u593e\u5e36\u8cc7\u8a0a\u985e\u4f3c\uff0c\u4f46\u5728\u53e5\u6cd5\u5206\u6790\u4e2d\u53ef\u4ee5\u76f4\u63a5\u8868\u793a\u8a5e \u8207\u8a5e\u4e4b\u9593\u7684\u76f8\u4f9d\u95dc\u4fc2\u3002 \u4e0d\u53ef\u7f3a\u5c11\u7684\u7279\u5fb5\u3002\u76f8\u8f03\u5176\u4ed6\u8a9e\u8a00\u4f86\u8aaa\uff0c\u4e2d\u6587\u7684\u8a9e\u6cd5\u67b6\u69cb\u8f03\u70ba\u8907\u96dc\uff0cUD \u8a9e\u6599\u5eab\u4e2d\u70ba\u4e86 \u7d71\u4e00\u5404\u500b\u8a9e\u8a00\u7684 dependency type \u53ea\u4f7f\u7528\u4e86 42 \u500b\uff0c\u76f8\u8f03\u65bc\u4e2d\u7814\u9662\u7684 68 \u985e\uff0cUD \u53ef\u80fd\u7121 \u5716 \u516d\u3001transition-based \u6d41\u7a0b\u5716 \u5c07\u4e0a\u8ff0\u7684\u6d41\u7a0b\u5957\u7528\u5728\u4e00\u500b\u5be6\u969b\u4f8b\u5b50\uff0c\u7531\u65bc\u53e5\u5b50\u7684\u9577\u5ea6\u8207 transition \u70ba\u7dda\u6027\u95dc\u4fc2\uff0c\u6211 \u81ea\u5f9e Mikolov \u63d0\u51fa\u65b0\u7684\u8a5e\u8868\u9054\u65b9\u5f0f\u5f8c\uff0c\u5404\u5f0f\u5404\u6a23\u7684\u795e\u7d93\u7db2\u8def\u8a13\u7df4\u6a21\u578b\uff0c\u53c8\u6210\u70ba\u81ea \u7136\u8a9e\u8a00\u8655\u7406\u754c\u7684\u5bf5\u5152\uff0c\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u5668\u7576\u7136\u4e5f\u4e0d\u4f8b\u5916\uff0c\u900f\u904e\u9810\u5148\u8a13\u7df4\u597d\u7684\u8a5e\u5411\u91cf\uff0c\u795e \u4ed6 \u559d \u4e86 \u5169\u676f \u73cd\u73e0 agent aspect property \u8a5e\u4e0d\u6b63\u78ba\uff0c\u6545\u4e0d\u80fd\u8a08\u7b97 LAS\u3002 \u70ba 10 \u4e14\u53d6 4 \u500b\u53ef\u80fd\u7684\u52d5\u4f5c\uff0c\u6211\u5011\u53ef\u4ee5\u770b 40 \u500b\u4e0d\u540c\u52d5\u4f5c\u9020\u6210\u7684\u4e0d\u540c\u7d50\u679c\u4e26\u5f97\u5230\u5c0d\u61c9\u7684 \u539f\u672c\u50c5\u9700\u9810\u6e2c\u7576\u524d\u7279\u5fb5\u7d44\u614b\u7684\u6700\u9069\u52d5\u4f5c\uff0c\u73fe\u5728\u5247\u6311\u51fa\u6a5f\u7387\u524d\uff2e\u9ad8\u7684\u52d5\u4f5c\uff0c\u6bcf\u500b\u52d5\u4f5c\u90fd \u5976\u8336 Nh VE Di DM Na Na \u6703\u7522\u751f\u65b0\u7684\u7279\u5fb5\u7d44\u614b\uff0c\u9019\u4e9b\u7d44\u614b\u6211\u5011\u6703\u5217\u5165 agenda(\u53ef\u80fd\u52d5\u4f5c\u7684\u96c6\u5408) \uff0c\u53ea\u8981\u6c92\u8d85\u904e \u5716 \u5341\u4e09\u3001\u7cfb\u7d71\u7b54\u6848 \u6211\u5011\u4ee5\u7576\u6642\u5404\u968a\u6bd4\u8cfd\u7684\u7d50\u679c\u7576\u4f5c\u6bd4\u8f03\uff0c\u7531\u65bc\u65b7\u8a5e\u53ef\u80fd\u8207\u6e2c\u8a66\u8a9e\u6599\u4e0d\u4e00\u81f4\uff0c\u5728\u6b64\u4e5f \u6a5f\u7387\uff0c\u518d\u5c07 30 \u500b\u8d85\u904e\u5b9a\u5411\u5bec\u5ea6\u7684\u7bc0\u9ede\u522a\u9664\uff0c\u4ee5\u6b64\u5faa\u74b0\u76f4\u5230\u7d42\u6b62\u72c0\u614b\u3002</td></tr><tr><td>\u7531\u65bc\u8fd1\u5e74\u4f86\u6a5f\u5668\u5b78\u7fd2\u7684\u6280\u8853\u4e00\u65e5\u5343\u91cc\uff0c\u904e\u53bb\u66fe\u7d93\u505a\u904e\u7684\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u65b9\u5f0f\uff0c\u53c8\u88ab\u63d0\u51fa \u4f86\u91cd\u65b0\u6aa2\u8996\uff0c\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u4e5f\u4e0d\u4f8b\u5916\uff0c\u4e26\u7372\u5f97\u76f8\u7576\u597d\u7684\u6210\u679c\u3002\u9996\u5148\u662f 2013 \u5e74 Mikolov \u63d0\u51fa\u65b0\u7684\u8a5e\u5411\u91cf\u8868\u793a\u6cd5(Word Embedding)[2]\uff0c\u4f7f\u53e5\u4e2d\u7684\u5404\u8a5e\u53ef\u4ee5\u5b58\u5728\u4e00\u5b9a\u7684\u95dc\u806f\uff0c\u4e26 \u80fd\u66f4\u597d\u7684\u5229\u7528\u6bcf\u500b\u8f38\u5165\u55ae\u8a5e\uff0c\u63a5\u8457\u662f\u4e0d\u540c\u6a5f\u5668\u5b78\u7fd2\u67b6\u69cb\uff0c\u4f8b\u5982 MLP (Multilayer Perceptron) \u548c LSTM (Long Short Term Memory)\uff0c\u5c07\u8907\u96dc\u7684\u7cfb\u7d71\u7c21\u55ae\u5316\uff0c\u50cf\u662f dependency parser \u539f\u672c\u900f\u904e\u6a5f\u7387\u6a21\u578b\uff0c\u9700\u8981\u8a2d\u8a08\u4e0d\u540c\u898f\u5247\u4f86\u5c0b\u627e\u8a5e\u8207\u8a5e\u4e4b\u9593\u7684\u53ef\u80fd\u95dc \u806f\uff0c\u5728\u5c0b\u627e\u7684\u540c\u6642\u6703\u8017\u8cbb\u8f03\u591a\u7684\u82e6\u5de5\uff0c\u7136\u800c\u4f7f\u7528\u6df1\u5ea6\u5b78\u7fd2\u6280\u8853\uff0c\u5c07\u9078\u5b9a\u597d\u7684\u7279\u5fb5\u8f38 \u6cd5\u6db5\u84cb\u4e2d\u6587\u8a9e\u610f\u66f4\u6df1\u5165\u7684\u5c64\u9762\u3002 \u5716 \u4e00 \u3001\u8207\u7684 POS \u70ba P35 \u5716 \u4e09\u3001End-to-End \u67b6\u69cb\u4e2d\u6240\u5305\u542b\u7684\u5143\u4ef6 \u5728\u672c\u7bc7\u8ad6\u6587\uff0c\u6211\u5011\u8457\u91cd\u65bc\u5716\u4e09\u4e2d\u7684 Dependency Parsing\uff0c\u5373\u70ba\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\uff0c\u5728\u9032 \u5011\u9078\u64c7\u8f03\u77ed\u7684\u53e5\u5b50\u300c\u7531\u56db\u6a13\u8d70\u4e0a\u4e94\u6a13\u300d \uff0c\u65b7\u8a5e\u5b8c\u70ba\u56db\u500b\u8a5e\uff0c\u6bcf\u500b\u8a5e\u7686\u6703\u88ab pop \u51fa Stack \u7d93\u7db2\u8def\u53ef\u4ee5\u66f4\u6709\u6548\u5730\u627e\u51fa\u8a5e\u548c\u8a5e\u4e4b\u9593\u7684\u6a21\u5f0f(pattern) \u3002\u6211\u5011\u900f\u904e\u73fe\u5728\u6700\u5e38\u898b\u7684\u5169\u7a2e\u6a21 \u578b--\u905e\u8ff4\u795e\u7d93\u7db2\u8def\u8207\u591a\u5c64\u611f\u77e5\u5668\u4f86\u5be6\u73fe\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u5668\uff0c\u7531\u65bc transition-based \u662f\u4e00 \u7a2e\u8caa\u5a6a\u6f14\u7b97\u6cd5\uff0c\u5206\u985e\u51fa\u7576\u524d\u7684\u52d5\u4f5c\u5f8c\uff0c\u4e26\u7121\u6cd5\u4f7f\u7528\u56de\u6eaf\u6cd5\u66f4\u6539\u524d\u9762\u5224\u65b7\u7684\u52d5\u4f5c\u3002\u56e0\u6b64 \u5728\u9019\u5169\u7a2e\u6df1\u5ea6\u6a21\u578b\u4e2d\u52a0\u5165\u4e86\u5b9a\u5411\u641c\u7d22(Beam Search) \uff0c\u7522\u751f\u591a\u500b\u53e5\u6cd5\u5206\u6790\u6a39\uff0c\u6700\u5f8c\u518d \u5716 \u516d \u3001\u5256\u6790\u5f8c\u4e4b\u53e5\u6cd5\u6a39 \u4ed6 \u559d \u4e86 Nh VE Di agent aspect \u5169\u676f DM \u73cd\u73e0 \u5976\u8336 Na Na buffer stack \u770b F1 \u505a\u70ba\u53c3\u8003\u4f9d\u64da\uff0c\u5176\u4e2d\u5206\u70ba\u5fae\u5e73\u5747\u8207\u5b8f\u5e73\u5747\uff0c\u800c\u5b8f\u5e73\u5747\u8207\u6211\u5011\u5be6\u9a57\u7684\u7b97\u6cd5\u4e00\u81f4\u3002\u8868 \u6211\u5011\u8a2d\u5b9a\u7684\u5bec\u5ea6(beam width) \uff0c\u5c31\u53ef\u4ee5\u6301\u7e8c\u52a0\u5165\u65b0\u7684\u7d44\u614b\u3002\u7576 agenda \u9054\u5230\u9810\u8a2d\u7684\u5927 \u5c0f\u6642\uff0c\u6211\u5011\u50c5\u6703\u52a0\u5165\u6bd4 agenda \u4e2d\u6700\u5dee\u503c\u6a5f\u7387\u9ad8\u7684\u7d44\u614b\u3002\u70ba\u4e86\u627e\u5230\u6700\u597d\u7684\u53e5\u6cd5\u6a39\uff0c\u4ee5\u8ff4 \u5708\u7684\u65b9\u5f0f\u9032\u884c\u5168\u5c40\u641c\u7d22\u76f4\u5230\u7d42\u6b62\u72c0\u614b\u3002\u5176\u4e2d\u6a5f\u7387\u70ba\u539f\u672c\u6df1\u5ea6\u5b78\u7fd2\u6a21\u578b\u4e0b\uff0c\u5c07\u6700\u7d42\u8f38\u51fa \u6539\u6210\u9078\u64c7\u7684 N \u500b\u6a5f\u7387\uff0c\u800c\u4e0d\u4f7f\u7528\u55ae\u4e00\u7684\u5206\u985e\u7d50\u679c\uff0c\u4e0b\u5716\u70ba\u672c\u7814\u7a76\u4f7f\u7528\u7684\u5b9a\u5411\u641c\u7d22\u793a\u610f \u5716\uff0c\u6bcf\u500b\u7bc0\u9ede\u90fd\u4ee3\u8868\u4e00\u500b\u7279\u5fb5\u7d44\u614b\uff0c\u6bcf\u500b\u7d44\u614b\u6703\u8f38\u5165\u6a21\u578b\u5167\u4e26\u56de\u50b3\u524d\u4e09\u9ad8\u7684\u6a5f\u7387\uff0c\u7576 \u8868 \u4e94 \u3001\u4e0d\u540c\u914d\u7f6e\u7684\u8868\u73fe \u4e09\u4e2d\u662f\u4ee5\u5e36\u8a9e\u610f\u89d2\u8272\u7684\u77ed\u8a9e\u7d50\u69cb\u5256\u6790\u5f8c\u7684\u7d50\u679c\uff0c\u76f8\u7576\u65bc\u5be6\u9a57\u4e2d LAS \u7684\u8a55\u4f30\u65b9\u5f0f\uff0c\u8868\u4e2d \u6700\u597d\u7684\u7d50\u679c\u70ba 0.4287\uff0c\u6545\u4ee5\u6b64\u70ba\u57fa\u6e96\u3002 \u8868 \u56db\u3001\u5e36\u8a9e\u610f\u89d2\u8272\u7684\u77ed\u8a9e\u7d50\u69cb\u5206\u6790\u7d50\u679c UAS LAS UAS f1-score LAS f1-score Stack-GRU 88.5% 83.1% 80.66% 74.17% +beam 89.8% 84.5% 81.05% 74.54% MLP 87.5% 82.7% 79.28% 73.24% \u548c push \u9032 Step Stack Word List Action Relation [root] [\u7531,\u56db\u6a13,\u8d70\u4e0a,\u4e94\u6a13] Shift \u5716 \u5341\u56db\u3001\u6b63\u78ba\u7b54\u6848 \u6bcf\u4e00\u500b\u53ef\u80fd\u7684\u53e5\u6cd5\u3002 \u5716 \u5341\u4e00\u3001\u7cfb\u7d71\u7b54\u6848 \u5716 \u4e03\u3001\u9810\u6e2c\u4e0b\u4e00\u500b\u52d5\u4f5c\u4e4b\u7279\u5fb5\u7d44\u614b \u5716 \u516b\u3001MLP \u67b6\u69cb \u9054\u5230\u7d42\u6b62\u72c0\u614b\u6642\uff0c\u5247\u8a08\u7b97\u6a5f\u7387\u7e3d\u548c\u6700\u9ad8\u4f5c\u70ba\u6700\u7d42\u7b54\u6848\u3002\u9019\u6a23\u4e00\u4f86\u6211\u5011\u4fbf\u80fd\u8f03\u5168\u9762\u7be9\u9078 +beam 88.3% 83.5% 80.45% 74.71%</td></tr><tr><td>\u5165\uff0c\u7d93\u904e\u96b1\u85cf\u5c64\u5f8c\u9762\u7684\u6fc0\u6d3b\u51fd\u6578\uff0c\u53ef\u4ee5\u76f4\u63a5\u5b78\u7fd2\u51fa\u6574\u500b\u53e5\u5b50\u7684\u8a9e\u6cd5\uff0c\u5c07\u898f\u5247\u6bd4\u5c0d\u8207\u8a08 \u7b97\u689d\u4ef6\u6a5f\u7387\u7684\u90e8\u5206\u7c21\u5316\u6210\u5b78\u7fd2\u795e\u7d93\u7db2\u8def\u4e2d\u7684\u6b0a\u91cd\u3002 \u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u6df1\u5ea6\u5b78\u7fd2\u7684\u6a21\u578b\u662f\u4f7f\u7528\u76e3\u7763\u5f0f\u5b78\u7fd2\uff0c\u56e0\u6b64\u6b63\u78ba\u7684\u4eba\u5de5\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u6a19\u793a \u8a9e\u6599\uff0c\u662f\u5efa\u7acb\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u5668\u4e0d\u53ef\u6216\u7f3a\u7684\u8a13\u7df4\u8cc7\u6599\u3002\u9019\u5e7e\u5e74\uff0c\u5b78\u8853\u754c\u7d44\u7e54\u4e86\u6578\u500b\u7d44\u7e54 \u4e14\u8a02\u5b9a\u4e86\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u7684\u4e00\u5957\u6a19\u6e96\uff0c\u4e26\u5efa\u7acb\u5404\u500b\u8a9e\u8a00\u7684\u4f9d\u5b58\u53e5\u6cd5\u6a19\u793a\u8a9e\u6599\u5eab\u3002\u5176\u4e2d\uff0c \u7531 Joakim Nivre \u4e3b\u6301\u7684 Universal Dependencies(UD)[3]\u662f\u4e00\u500b\u591a\u4eba\u53c3\u8207\u4e26\u5efa\u7acb\u8cc7\u6599\u8f03\u5b8c \u6574\u7684\u7d44\u7e54\uff0c\u5305\u542b\u4e86\u5168\u4e16\u754c\u7d55\u5927\u6578\u8a9e\u8a00\uff0c2.1 \u7248\u4e2d\u65b0\u589e\u4e86\u7e41\u9ad4\u8207\u7c21\u9ad4\u4e2d\u6587\u3002\u6211\u5011\u7814\u7a76\u5ba4\u5728 \u904e\u53bb\u5df2\u767c\u5c55\u6709\u7e41\u9ad4\u4e2d\u6587\u7684 word segmenter [4],\u3001POS tagger \u518d\u52a0\u4e0a\u4f9d\u5b58\u53e5\u6cd5\u7684\u8a13\u7df4\u5de5\u5177 \u53ca\u8a13\u7df4\u8a9e\u6599\u61c9\u8a72\u53ef\u4ee5\u9054\u5230\u5f88\u597d\u7684\u8a13\u7df4\u6548\u679c\u3002\u4e8b\u5be6\u4e0a\uff0c\u9084\u6709\u591a\u554f\u984c\u9700\u8981\u89e3\u6c7a\uff0c\u5728 UD \u7e41 \u9ad4\u4e2d\u6587\u8a9e\u6599\u5eab\u4e2d\u7684\u8a13\u7df4\u8a9e\u6599\u50c5\u6709\u5927\u7d04\u5341\u842c\u8a5e\uff0c\u4e14\u5176\u4e2d\u4f7f\u7528\u7684\u8a5e\u6027\u8207\u65b7\u8a5e\u4e26\u6c92\u6709\u4f9d\u7167\u4e2d \u7814\u9662\u7684\u6a19\u6e96\u4f86\u5efa\u7acb\u3002\u5c0d\u65bc\u6211\u5011\u7684\u73fe\u6cc1\u4f86\u8aaa UD \u8a9e\u6599\u5eab\u4e26\u4e0d\u9069\u5408\u5efa\u7acb\u4e00\u5957 End-to-End \u7684 \u5716 \u4e8c \u3001\u8207\u7684 POS \u70ba Caa 2. \u905e\u8ff4\u795e\u7d93\u7db2\u8def Recurrent Neural Network [root,\u7531] [\u56db\u6a13,\u8d70\u4e0a,\u4e94\u6a13] Shift \u5728\u672c\u7814\u7a76\u8a0e\u8ad6\u4e86\u56db\u7a2e\u914d\u7f6e\uff0cStack-GRU \u8207 MLP\uff0c\u4ee5\u53ca\u5404\u5225\u52a0\u4e0a beam search \u5f8c\u7684\u7d50 \u597d\uff0c\u9084\u6c92\u52a0\u5165\u5b9a\u5411\u641c\u7d22\u6642\uff0cLAS \u76f8\u5dee 0.4\uff0c\u4f46\u5728\u52a0\u5165\u5b9a\u5411\u641c\u7d22\u6642\u5247\u5dee\u8ddd 1\uff0c\u53ef\u898b\u905e\u8ff4\u795e dependency relation\uff0c\u8b8a\u6210\u4e00\u500b\u591a\u985e\u5225\u7684\u5206\u985e\u5668\uff0c\u5728\u5224\u65b7\u5de6\u53f3\u7684\u540c\u6642\u4e5f\u80fd\u5224\u65b7\u985e\u5225\u5c31\u80fd [root,\u7531,\u56db\u6a13] [\u8d70\u4e0a,\u4e94\u6a13] RightArc (\u7531\u2192\u56db\u6a13) \u679c\u3002\u5176\u4e2d\u5305\u542b\u4e2d\u7814\u9662\u63d0\u4f9b\u7684\u6e2c\u8a66\u8a9e\u6599\u7d93\u904e\u6a21\u578b\u7684\u6e96\u78ba\u5ea6\uff0c\u8207\u672c\u5be6\u9a57\u4f7f\u7528\u7684\u65b7\u8a5e\u5668\u65b7\u8a5e \u5728\u672c\u5be6\u9a57\u4e2d\uff0c\u591a\u5c64\u611f\u77e5\u5668\u4f7f\u7528\u4e86 250 \u7dad\u7684\u8a5e\u5411\u91cf\uff0c\u8a5e\u6027\u70ba 121 \u985e\u7684\u7368\u71b1\u7de8\u78bc \u7d93\u7db2\u8def\u8f03\u80fd\u627e\u51fa\u904e\u53bb\u7684\u8a5e\u8a9e\u7576\u524d\u8a5e\u7684\u95dc\u806f\u3002\u7136\u800c\u4f7f\u7528\u672c\u7814\u7a76\u4e2d\u7684\u65b7\u8a5e\u5668\u7684\u7d50\u679c\uff0c\u5247\u662f \u5f97\u5230\u5716\u4e03\u5e36\u6709\u6a19\u7c64\u7684\u7d50\u679c\u3002 [root,\u7531] [\u8d70\u4e0a,\u4e94\u6a13] Shift \u5f8c\u7684\u7d50\u679c\uff0c\u800c\u6211\u5011\u4f7f\u7528\u7684\u65b7\u8a5e\u5668\u8207\u4e2d\u7814\u9662\u65b7\u8a5e\u6709\u4e9b\u8a31\u51fa\u5165\uff0c\u56e0\u6b64\u4f7f\u7528 f1-score \u4f5c\u70ba\u5206 (one-hot encoding)\u548c\u4f9d\u5b58\u95dc\u4fc2 69 \u985e\u7684\u7368\u71b1\u7de8\u78bc\u4f5c\u70ba\u8f38\u5165\uff0c\u5404\u81ea\u4e58\u4e0a\u6b0a\u91cd\u77e9\u9663\u5f8c\u518d\u5408 MLP \u52a0\u5165\u5b9a\u5411\u641c\u7d22\u8868\u73fe\u6700\u597d\uff0c\u539f\u56e0\u662f\u7576\u65b7\u8a5e\u932f\u8aa4\u6216\u662f\u8a5e\u6027\u6a19\u793a\u4e0d\u4e00\u81f4\uff0cGRU \u5c0d\u65bc\u904e \u7d9c\u89c0\u4e0a\u9762\u5e7e\u7a2e\u60c5\u5f62\uff0c\u518d\u52a0\u4e0a\u6211\u5011\u5be6\u9a57\u5ba4\u73fe\u6709\u7684\u65b7\u8a5e\u8207 POS \u6a19\u793a\u90fd\u8207\u4e2d\u7814\u9662\u6a19\u6e96\u76f8\u540c\uff0c [root,\u7531,\u8d70\u4e0a] [\u4e94\u6a13] LeftArc (\u7531\u2190\u8d70\u4e0a) \u6790\u7d50\u679c\u3002 \u4f75\uff0c\u4e26\u5728\u5f8c\u9762\u4f7f\u7528\u5169\u5c64 RELU \u96b1\u85cf\u5c64\uff0c\u5982\u5716\u5341\uff0c\u8a5e\u5411\u91cf\u3001\u8a5e\u6027\u548c\u4f9d\u5b58\u95dc\u4fc2\u7684\u96b1\u85cf\u5c64\u795e \u53bb\u7684\u8a5e\u6027\u8f03\u70ba\u654f\u611f\uff0c\u4ee5\u81f3\u65bc\u904e\u53bb\u7684\u932f\u8aa4\u6703\u96a8\u8457\u905e\u8ff4\u795e\u7d93\u5143\u4e2d\u50b3\u64ad\u81f3\u7576\u524d\u7684\u9810\u6e2c\u7d50\u679c\u3002 \u4ee5\u81f4\u65bc\u6211\u5011\u80fd\u66f4\u8457\u91cd\u5728\u8a13\u7df4\u65b9\u6cd5\u4e0a\u3002\u672c\u7814\u7a76\u63d0\u51fa\u4e86\u7aef\u5230\u7aef\u7684\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u5668\uff0c\u4f86\u63d0\u4f9b \u672a\u4f86\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u767c\u5c55\u4e0a\u4e00\u9805\u6709\u5229\u7684\u5de5\u5177\u3002 \u5f8c\u9762\u5c0f\u7bc0\u4e2d\u6703\u63d0\u5230\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u5668\u7684\u505a\u6cd5\uff0c\u4f7f\u7528\u4e86 Joakim Nivre \u63d0\u51fa\u7684 Transition-based algorithm[5]\uff0c\u4e26\u7d50\u5408 Stanford \u63d0\u51fa\u7684\u9ad8\u6548\u7387\u5256\u6790\u5668[6]\u4e2d\u7684 Multilayer perceptron \u67b6\u69cb\uff0c\u9664\u4e86 Feedforward \u5916\uff0c\u4e5f\u4f7f\u7528 Recursive Neural Networks \u4e2d\u7684 Gated Recurrent Unit \u4f86\u5efa\u7acb\u6211\u5011\u7684\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u5668[7]\uff0c\u4f86\u5f4c\u88dc MLP \u4e2d\u7f3a\u5931\u7684\u9806\u5e8f\u554f\u984c\u3002 \u4e8c\u3001 \u65b9\u6cd5 \u5716 \u4e94\u3001Constituency Tree [root,\u8d70\u4e0a] [\u4e94\u6a13] Shift [root,\u8d70\u4e0a,\u4e94\u6a13] [] RightArc (\u8d70\u4e0a\u2192\u4e94\u6a13) [root,\u8d70\u4e0a] [] RightArc (root\u2192\u8d70\u4e0a) [root] [] Done (\u4e8c) \u5be6\u9a57\u7d50\u679c \u76ee\u524d\u7e41\u9ad4\u4e2d\u6587\u4e26\u6c92\u6709\u4efb\u4f55\u5df2\u77e5\u7684\u57fa\u6e96\u63d0\u4f9b\u6211\u5011\u505a\u53c3\u8003\uff0c\u56e0\u6b64\u4ee5 2012 \u5e74 bake-offs \u7684 \u56db\u3001 \u7d50\u8ad6 \u7d93\u5143\u70ba 200 \u80fd\u6536\u6582\uff0c\u56e0\u6b64\u5728\u8a13\u7df4\u4e0a\uff0cMLP \u5341\u5206\u6709\u6548\u7387\u3002 \u5728\u672c\u7814\u7a76\u4e2d\u5be6\u73fe\u4e86\u5b8c\u6574\u7684\u53e5\u6cd5\u5206\u6790\u7cfb\u7d71\uff0c\u7d50\u5408\u4ea4\u901a\u5927\u5b78\u8a9e\u97f3\u5be6\u9a57\u5ba4\u7684\u65b7\u8a5e\u5668\u4ee5\u53ca \u5165\u3002\u4ee5\u4e0a\u4ecb\u7d39\u5f88\u53ef\u6703\u904e\u65bc\u62bd\u8c61\uff0c\u56e0\u6b64\u4ee5\u5716\u516b\u70ba\u4f8b\uff0c\u5256\u6790\u5b8c\u6210\u7684\u4f9d\u5b58\u53e5\u6cd5\u6a39\uff0ctransition-based \u4e2d\u6bcf\u4e00\u500b\u52d5\u4f5c\u90fd\u662f\u900f\u904e\u7279\u5fb5\u7d44\u614b(\u5982\u8868\u4e00\u7684 step)\u4f86\u9810\u6e2c\u4e0b\u4e00\u500b\u52d5\u4f5c\u3002\u4f46\u8f38\u5165\u7684\u7279 \u5fb5\u5c31\u50cf\u4e0a\u9762\u8aaa\u7684\u4e0d\u50c5\u50c5\u53ea\u662f\u62ff\u7576\u524d\u72c0\u614b\u7684 stack \u548c buffer\uff0c\u9084\u5305\u62ec\u4e86\u5df2\u77e5\u7684\u4fee\u98fe\u95dc\u4fc2\u7d44 \u5408\u3002\u5982\u5716\u4e5d\uff0c \u300c\u4ed6\u300d\u548c\u300c\u4e86\u300d\u90fd\u5728\u524d\u9762\u6b65\u9a5f\u88ab\u79fb\u51fa stack\uff0c\u56e0\u70ba\u9019\u5169\u8005\u90fd\u4fee\u98fe\u4e86\u300c\u559d\u300d \u9019\u500b\u8a5e\uff0c\u7136\u800c\u9019\u9805\u95dc\u4fc2\u5728\u4e0b\u4e00\u6b65\u9a5f\u4e26\u6c92\u6709\u56e0\u6b64\u88ab\u5ffd\u7565\uff0c\u9032\u800c\u6210\u70ba\u6211\u5011\u7684\u8f38\u5165\u7279\u5fb5\u3002 \u5716 \u5341\u3001Beam-search \u5716 \u5341\u4e8c\u3001\u6b63\u78ba\u7b54\u6848 \u9664\u4e86\u57fa\u672c\u7684 UAS \u8207 LAS \u5916\uff0c\u4e2d\u6587\u591a\u4e86\u4e00\u9805\u8a55\u4f30\u6a19\u6e96\u4ee5 f1-score \u4f5c\u70ba\u6e2c\u8a66\u7cfb\u7d71\u6548\u80fd\u7684 \u4f9d\u64da\u3002f1-socre \u7684\u516c\u5f0f\u70ba 2 * * Traditional Chinese Parsing task[7]\u7684\u8a55\u4f30\u7d50\u679c\u4f86\u7576\u6211\u5011\u7684\u57fa\u6e96\u3002\u6b64\u7e41\u9ad4\u4e2d\u6587\u5256\u6790\u4f7f\u7528\u7684 \u8a5e\u6027\u6a19\u793a\uff0c\u8ddf\u904e\u53bb\u7684 share task \u6bd4\u5f9e 0.42 \u81f3 0.74 \u6709\u5c07\u8fd1 70%\u7684\u9032\u6b65\u5e45\u5ea6\uff0c\u986f\u793a\u51fa\u5728\u6df1\u5ea6 \u905e\u8ff4\u795e\u7d93\u7db2\u8def\u5728\u67b6\u69cb\u4e0a\u5c31\u8207\u591a\u5c64\u611f\u77e5\u5668\u6709\u76f8\u7576\u5927\u7684\u5dee\u5225\uff0c\u6240\u4ee5\u5728\u53c3\u6578\u7684\u8a2d\u7f6e\u4e5f\u6703 \u662f\u77ed\u8a9e\u7684\u6a39\u72c0\u7d50\u69cb\uff0c\u5982\u4e0a\u9762\u6240\u8aaa\uff0c\u77ed\u8a9e\u53e5\u6cd5\u6a39\u53ef\u4ee5\u8f49\u63db\u6210\u4f9d\u5b58\u53e5\u6cd5\u7d50\u69cb\uff0c\u56e0\u6b64\u4f7f\u7528\u6b64 \u5b78\u7fd2\u6a21\u578b\u5c0d\u65bc\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u7684\u529f\u52de\u4e0d\u53ef\u5c0f\u89b7\u3002\u76ee\u524d\u5728\u6b64\u9818\u57df\u4e2d\u5c1a\u672a\u6709\u4eba\u5b8c\u6574\u7684\u6574\u7406\u51fa \u6709\u6240\u4e0d\u540c\uff0cGRU \u5206\u6210 stack \u8ddf buffer \u5169\u90e8\u4efd\uff0c\u6642\u9593\u6b65\u6578(time step)\u8a2d\u70ba 3\uff0c\u7531\u65bc\u6211\u5011 \u57fa\u6e96\u6211\u5011\u8a8d\u70ba\u662f\u53ef\u4fe1\u7684\u3002\u5982\u8868\u4e8c\u4e4b\u4e00\uff0c2012 \u5e74\u6240\u4f7f\u7528\u7684\u8cc7\u6599\u96c6\uff0c\u8207\u672c\u6b21\u5be6\u9a57\u8a13\u7df4\u8207\u6e2c \u4e00\u5957\u8f38\u5165\u53e5\u5b50(\u672a\u65b7\u8a5e)\u5373\u53ef\u5f97\u5230\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u6a39\u7684\u7cfb\u7d71\uff0c\u4e26\u4e14\u5df2\u5efa\u69cb\u7dda\u4e0a\u7cfb\u7d71\u63d0\u4f9b \u7684\u53e5\u9577\u4e26\u4e0d\u9577\uff0c\u6b65\u6578\u8a2d\u5c11\u4e00\u9ede\u53ef\u4ee5\u4f7f\u6574\u9ad4\u6a21\u578b\u57f7\u884c\u66f4\u5feb\u3002\u4e0d\u540c\u65bc MLP\uff0cGRU \u4e0d\u662f\u66b4 \u8a66\u8a9e\u6599\u6bd4\u7387\u5747\u70ba 1%\u5de6\u53f3\u3002\u8868\u4e8c\u4e4b\u4e00\u7684\u6e2c\u8a66\u8a9e\u6599\u4f54 1.5\uff05\uff0c\u8868\u4e8c\u4e4b\u4e8c\u5247\u4f54 1.2%\u3002 \u5b78\u8853\u4f7f\u7528\u3002\u7531\u65bc\u7e41\u9ad4\u4e2d\u6587\u7684\u81ea\u7136\u8655\u7406\u9818\u57df\uff0c\u624d\u525b\u8d77\u6b65\uff0c\u5c0d\u65bc\u9019\u6a23\u7684\u7cfb\u7d71\u4e0a\u6709\u5341\u5206\u591a\u7684 \u529b\u5730\u5c07\u7279\u5fb5\u5c55\u958b\u518d\u9935\u9032\u6a21\u578b\uff0cstack \u4e2d\uff0c\u5982\u6211\u5011\u5728\u65b9\u6cd5\u4e2d\u6240\u4ecb\u7d39\u7684\uff0c\u6703\u5e36\u6709\u90e8\u5206\u5256\u6790\u6a39 f1 = + \u7684\u8cc7\u8a0a\uff0c\u5206\u5225\u6709 stack \u6700\u4e0a\u9762\u5169\u5c64\u4fee\u98fe\u7684\u8a5e\u8207\u4f9d\u5b58\u95dc\u4fc2\uff0c\u70ba\u4e86\u8b93\u8a5e\u548c\u4f9d\u5b58\u95dc\u4fc2\u53ef\u4ee5\u6709\u66f4 \u6539\u5584\u7a7a\u9593\uff0c\u5982\u4f7f\u7528 graph-based \u65b9\u6cd5\uff0c\u53ef\u7372\u5f97\u66f4\u597d\u7684\u9577\u8ddd\u96e2\u4fee\u98fe\u95dc\u4fc2\uff0c\u5e0c\u671b\u4ee5\u6b64\u8ad6\u6587\u4f5c</td></tr><tr><td>\u5716 \u4e5d\u3001Stack LSTM \u67b6\u69cb \u597d\u7684\u5b78\u7fd2\u6548\u679c\uff0c\u4fbf\u5404\u81ea\u52a0\u5165\u96b1\u85cf\u5c64\u5b78\u7fd2\u7279\u5b9a\u7684\u6b0a\u91cd\u77e9\u9663\uff0c\u5982\u6b64\u4e00\u4f86\u91cd\u8981\u7684\u4f9d\u5b58\u95dc\u4fc2\u7279 \u70ba\u958b\u982d\uff0c\u8b93\u66f4\u591a\u5b78\u8005\u52a0\u5165\u7814\u7a76\u3002</td></tr></table>", |
| "num": null, |
| "type_str": "table", |
| "text": "The 2018 Conference on Computational Linguistics and Speech Processing ROCLING 2018, pp. 61-75 \u00a9The Association for Computational Linguistics and Chinese Language Processing \u7e41\u9ad4\u4e2d\u6587\u4f9d\u5b58\u53e5\u6cd5\u5256\u6790\u5668 Traditional Chinese Dependency Parser Keywords: dependency parser, natural language processing, MultiLayer Perceptron , \u5165\u6a21\u578b\u5efa\u7acb\u7684\u65b9\u6cd5\u524d\uff0c\u5148\u7c21\u55ae\u4ecb\u7d39\u5b83\u7684\u8868\u793a\u65b9\u5f0f\u53ca\u4ee3\u8868\u610f\u7fa9\uff1a \u5728 dependency parser \u8f38\u51fa\u4e2d(\u5982\u5716\u56db\u4e4b\u7bc4\u4f8b)\uff0c\u6703\u4ee5\u5f27 (arc) \u4f86\u8868\u793a\u8a5e\u8207\u8a5e\u9593\u7684\u76f8\u4f9d \u6027\uff0c\u5716\u4e2d arc \u6703\u5f9e\u4e2d\u5fc3\u8a9e(Head)\u6307\u5411\u4fee\u98fe\u8a9e\uff0c\u5982\uff1a\"\u66f8\u67b6\" \u4e00\u8a5e\u6703\u6307\u5411\"\u4e00\u5ea7\u5ea7\"\u3002\u4e26\u4e14\u6bcf \u4e00\u500b arc \u6703\u6a19\u793a dependency relation (\u5982\uff1aquantifier(\u6578\u91cf\u8a5e\u4fee\u98fe\u8a9e))\u3002\u800c\"\u4e00\u5ea7\u5ea7\u66f8\u67b6\"\u5c31 \u7d44\u6210\u4e00\u500b\u8a5e\u7d44\u6216\u7247\u8a9e\u3002\u800c\u518d\u4e0b\u4e00\u5c64\uff0c\"\u4e00\u5ea7\u5ea7\u66f8\u67b6\" \u8a5e\u7d44\u4fee\u98fe \"\u4e0a\" \u3002\u5982\u6b64\uff0c\u4e00\u5c64\u4e00\u5c64\u4e0b \u53bb\uff0c\u6700\u5f8c\uff0c\u53e5\u4e2d\u672a\u4fee\u98fe\u4ed6\u8a5e\u7684\u4e2d\u5fc3\u8a9e(Head)\u5247\u7a31\u70ba root (\u5982\u53e5\u4e2d\"\u653e\"\u4e00\u8a5e)\u3002 \u5982\u6b64\uff0cdependency parser \u4ee5 head \u548c dependent \u4f86\u8868\u793a\u5169\u500b\u8a5e\u4e4b\u9593\u7684\u95dc\u806f\uff0c\u4ee5 head \u4ee3\u8868\u4e00\u500b phrase \u7684\u4e2d\u5fc3\u8a5e(\u540d\u8a5e\u7247\u8a9e\u4e2d\u7684\u540d\u8a5e\uff0c\u52d5\u8a5e\u7247\u8a9e\u4e2d\u7684\u52d5\u8a5e) \uff0c\u5269\u4e0b\u7684\u8a5e\u53ef\u80fd \u8207 head \u6709 direct\u3001indirect \u6216 dependent \u7684\u95dc\u4fc2\u3002\u6211\u5011\u53ef\u4ee5\u900f\u904e\u9019\u6a23\u7684\u95dc\u806f\uff0c\u5c07 headdependent \u7684\u914d\u5c0d\u505a\u66f4\u9032\u4e00\u6b65\u7684\u5206\u985e\uff0c\u7a31\u4f5c dependency relation \u6216 grammatical function\u3002 \u5716 \u56db \u3001\u4f9d\u5b58\u53e5\u6cd5\u7bc4\u4f8b \u800c\u5efa\u69cb\u4f9d\u5b58\u53e5\u6cd5\u5206\u6790\u5668\u5c1a\u9700\u6eff\u8db3\u4e09\u9805\u9650\u5236\uff0c(1) \u5916\u52a0\u7684 root\uff0c\u4e26\u6c92\u6709\u4efb\u4f55 arc \u6307\u5411\u5b83\uff0c (2) root \u4ee5\u5916\u7684 node(\u8a5e\u6216\u6a19\u9ede\u7b26\u865f)\uff0c\u90fd\u6703\u88ab\u4e00\u500b arc \u6307\u5411\uff0c(3) root \u5230\u6240\u6709\u5176\u4ed6 vertex \u53ea\u6709\u4e00\u689d\u552f\u4e00\u8def\u5f91\u3002\u4e0a\u8ff0\u7684 vertex \u6307\u7684\u662f\u5728\u77ed\u8a9e\u7d50\u69cb\u4e2d(\u5982\u5716\u4e94) \uff0c\u8a5e\u7d44\u7684\u4e2d\u5fc3\u8a9e\u3002\u77ed \u8a9e\u7d50\u69cb\u6a39\u4e5f\u53ef\u7a31\u70ba constituency tree\uff0c\u4e26\u7121\u4e0a\u4e0b\u6587\u9806\u5e8f\u95dc\u4fc2\uff0c\u50c5\u4ee5\u7247\u8a9e\u70ba\u5206\u6790\u91cd\u5fc3\u3002 \u4ee5\u4e0a\u9762\u7684\u908f\u8f2f\uff0c\u6211\u5011\u4f7f\u7528 Transition Based Parsing \u7684\u65b9\u5f0f\uff0c\u4f86\u4e00\u4e00\u627e\u51fa\u8a5e\u8207\u8a5e\u7684\u95dc \u806f\uff0c\u9019\u500b\u65b9\u6cd5\u662f\u4f86\u81ea\u65bc shift-reduced parsing\uff0c\u4f7f\u7528\u4e86 stack \u548c\u5373\u5c07\u88ab\u5256\u6790\u7684 word list\u3002 \u5728 transition-based parsing \u4e2d\u6709\u500b\u91cd\u8981\u7684\u7279\u5fb5\u7d44\u614b(configuration)\u6982\u5ff5\uff0c\u5176\u4e2d\u6709\u4e09\u500b\u91cd \u8981\u6210\u5206\uff0cstack\u3001buffer \u548c dependency relations \u7684\u96c6\u5408\u3002\u6700\u521d\u7684\u7279\u5fb5\u7d44\u614b\u4e2d\uff0cstack \u53ea\u5305 \u542b ROOT\uff0cbuffer \u5247\u662f\u53e5\u5b50\u4e2d\u7684\u6240\u6709\u8a5e\uff0c\u800c dependency relations \u4e00\u958b\u59cb\u662f\u7a7a\u7684\u3002\u7d42\u6b62\u72c0 \u614b\u4e2d stack \u4e2d\u50c5\u5269\u4e0b ROOT \u548c buffer \u5fc5\u9808\u662f\u7a7a\u7684\uff0cdependency relations \u7684\u96c6\u5408\u4ee3\u8868\u6700\u7d42 \u7684\u5256\u6790\u7d50\u679c\u3002\u70ba\u4e86\u7522\u751f\u65b0\u7684\u7279\u5fb5\u7d44\u614b\uff0c\u6211\u5011\u6703\u4f7f\u7528\u4e09\u500b operator\uff1a LeftArc : \u5728\u6700\u4e0a\u5c64\u7684\u8a5e\u548c\u4ed6\u4e0b\u4e00\u5c64\u7684\u8a5e\u4e2d\u52a0\u5165\u65b0\u7684 head-dependent relations\uff0c\u4e26 pop \u7b2c\u4e8c\u5c64\u7684\u8a5e\u3002 RightArc : \u5728\u7b2c\u4e8c\u5c64\u7684\u8a5e\u548c\u6700\u4e0a\u5c64\u7684\u8a5e\u4e2d\u52a0\u5165\u65b0\u7684 head-dependent relations\uff0c\u4e26 pop \u6700\u4e0a\u5c64\u7684\u8a5e\u3002 Shift :\u5c07 buffer \u4e2d\u6700\u524d\u9762\u7684\u8a5e push \u9032 stack \u5f8c\uff0c\u5c07\u5176\u79fb\u9664\u3002 \u5728\u5716\u516d\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u5c07 Check dependency \u5206\u6210\u4e0d\u53ea Left-arc \u8ddf Right-arc\uff0c\u52a0\u4e0a Stack \u4e00\u6b21\uff0c\u56e0\u6b64\u53ea\u9700\u8981 4*2 \u6b65 2 \u5c31\u53ef\u4ee5\u5256\u6790\u5b8c\u4e00\u53e5\u8a71\u3002\u5f9e\u8868\u4e00\u6211\u5011\u53ef\u4ee5\u77e5 \u9053\uff0c\u521d\u59cb\u72c0\u614b\u5f9e ROOT \u958b\u59cb\u4e5f\u5230 ROOT \u7d50\u675f\uff0c\u6bcf\u500b\u8a5e\u4e5f\u90fd\u6703\u8f2a\u6d41\u9032\u5165 stack \u6aa2\u8996\uff0c\u76f4 \u5230\u6240\u6709\u95dc\u806f\u88ab\u5efa\u7acb\u3002\u7576\u57f7\u884c\u5b8c\u8868\u4e00\u7684\u6240\u6709\u6b65\u9a5f\uff0c\u5c07\u95dc\u806f\u5716\u50cf\u5316\u5373\u53ef\u5f97\u5230\u5982\u5716\u4e03 \u8868 \u4e00 \u3001Transition based parsing \u7684\u904e\u7a0b \u500b\uff0c\u800c\u6a19\u6e96\u7b54\u6848\u5171\u6709 5 \u500b\uff0c\u56e0\u6b64\u6e96\u78ba\u7387\u70ba 4/5\uff0c\u7136\u800c recall(\u53ec\u56de\u7387)\u662f\u4ee5\u6211\u5011\u7684\u7cfb \u7d71\u8f38\u51fa\u70ba\u5206\u6bcd\u8a08\u7b97\uff0c\u4e0b\u5716\u7cfb\u7d71\u8f38\u51fa\u5171\u6709 8 \u500b\u7bad\u982d\uff0c\u53ec\u56de\u7387\u70ba 4/8\uff0c\u6700\u5f8c\u8a08\u7b97 f1-score \u70ba \u500b\uff0c\u5408\u4f75\u5f8c\u5247\u70ba 600 \u500b\uff0c\u901a\u904e\u6bcf\u4e00\u5c64\u96b1\u85cf\u5c64\u90fd\u9010\u6f38\u6e1b\u5c11 200 \u7dad\uff0c\u6700\u5f8c\u4e00\u5c64 Softmax \u5247\u8f38\u51fa 139 \u985e(transition-based action) \u3002\u5176\u4e2d Batch size \u70ba 100\uff0cEpoch \u70ba 5 \u5c31 \u524d\u5169\u5c64\u7684\u8a5e\u5411\u91cf\u5408\u4f75\uff0c\u518d\u7d93\u904e\u4e00\u5c64\u96b1 \u85cf\u5c64\u5f97\u5230\u8a72 stack \u8a5e\u7684\u65b0\u8a5e\u5411\u91cf\u3002\u9019\u6a23\u9032\u5165 GRU \u7684\u6bcf\u500b\u6642\u9593\u6b65\u6578\u90fd\u70ba\u540c\u6a23\u7dad\u5ea6\u3002\u5728\u6b64 \u6a21\u578b\u4e2d\uff0cGRU \u8a2d\u7f6e\u5169\u5c64\u96b1\u85cf\u66fe\uff0c\u8f38\u51fa\u70ba 200 \u7dad\uff0cstack \u548c buffer \u7684 GRU \u8a2d\u7f6e\u5b8c\u5168\u4e00 \u6a23\uff0c\u6700\u5f8c\u518d\u5c07 stack \u548c buffer \u7684\u8f38\u51fa\u5408\u4f75\u4e00\u8d77\u8f38\u5165 softmax \u5c64\uff0c\u4e26\u5f97\u5230\u8207 MLP \u4e00\u6a23\u7684 \u8868\u56db\u70ba\u500b\u5225\u770b MLP \u8207 GRU \u6a21\u578b\u7684\u7d50\u679c\uff0c\u53ef\u4ee5\u770b\u51fa\u5728\u591a\u6578\u60c5\u6cc1 GRU \u8868\u73fe\u6bd4 MLP" |
| } |
| } |
| } |
| } |
