[关闭]
@hanxiaoyang 2016-11-08T23:15:26.000000Z 字数 9648 阅读 16871

word2vec训练中文模型

word2vec

1.准备数据与预处理

首先需要一份比较大的中文语料数据，可以考虑中文的维基百科（也可以试试搜狗的新闻语料库）。中文维基百科的打包文件地址为
https://dumps.wikimedia.org/zhwiki/latest/zhwiki-latest-pages-articles.xml.bz2

中文维基百科的数据不是太大，xml的压缩文件大约1G左右。首先用 process_wiki_data.py处理这个XML压缩文件，执行：python process_wiki_data.py zhwiki-latest-pages-articles.xml.bz2 wiki.zh.text

  1. #!/usr/bin/env python
  2. # -*- coding: utf-8 -*-
  3. # process_wiki_data.py 用于解析XML，将XML的wiki数据转换为text格式
  5. import logging
  6. import os.path
  7. import sys
  9. from gensim.corpora import WikiCorpus
  11. if __name__ == '__main__':
  12.     program = os.path.basename(sys.argv[0])
  13.     logger = logging.getLogger(program)
  15.     logging.basicConfig(format='%(asctime)s: %(levelname)s: %(message)s')
  16.     logging.root.setLevel(level=logging.INFO)
  17.     logger.info("running %s" % ' '.join(sys.argv))
  19.     # check and process input arguments
  20.     if len(sys.argv) < 3:
  21.         print globals()['__doc__'] % locals()
  22.         sys.exit(1)
  23.     inp, outp = sys.argv[1:3]
  24.     space = " "
  25.     i = 0
  27.     output = open(outp, 'w')
  28.     wiki = WikiCorpus(inp, lemmatize=False, dictionary={})
  29.     for text in wiki.get_texts():
  30.         output.write(space.join(text) + "\n")
  31.         i = i + 1
  32.         if (i % 10000 == 0):
  33.             logger.info("Saved " + str(i) + " articles")
  35.     output.close()
  36.     logger.info("Finished Saved " + str(i) + " articles")

得到信息：

  1. 2016-08-11 20:39:22,739: INFO: running process_wiki.py zhwiki-latest-pages-articles.xml.bz2 wiki.zh.text
  2. 2016-08-11 20:40:08,329: INFO: Saved 10000 articles
  3. 2016-08-11 20:40:45,501: INFO: Saved 20000 articles
  4. 2016-08-11 20:41:23,659: INFO: Saved 30000 articles
  5. 2016-08-11 20:42:01,748: INFO: Saved 40000 articles
  6. 2016-08-11 20:42:33,779: INFO: Saved 50000 articles
  7. ......
  8. 2016-08-11 20:55:23,094: INFO: Saved 200000 articles
  9. 2016-08-11 20:56:14,692: INFO: Saved 210000 articles
  10. 2016-08-11 20:57:04,614: INFO: Saved 220000 articles
  11. 2016-08-11 20:57:57,979: INFO: Saved 230000 articles
  12. 2016-08-11 20:58:16,621: INFO: finished iterating over Wikipedia corpus of 232894 documents with 51603419 positions (total 2581444 articles, 62177405 positions before pruning articles shorter than 50 words)
  13. 2016-08-11 20:58:16,622: INFO: Finished Saved 232894 articles

Python的话可用jieba完成分词，生成分词文件wiki.zh.text.seg
接着用word2vec工具训练：
python train_word2vec_model.py wiki.zh.text.seg wiki.zh.text.model wiki.zh.text.vector

  1. #!/usr/bin/env python
  2. # -*- coding: utf-8 -*-
  3. # train_word2vec_model.py用于训练模型
  5. import logging
  6. import os.path
  7. import sys
  8. import multiprocessing
  10. from gensim.corpora import WikiCorpus
  11. from gensim.models import Word2Vec
  12. from gensim.models.word2vec import LineSentence
  14. if __name__ == '__main__':
  15.     program = os.path.basename(sys.argv[0])
  16.     logger = logging.getLogger(program)
  18.     logging.basicConfig(format='%(asctime)s: %(levelname)s: %(message)s')
  19.     logging.root.setLevel(level=logging.INFO)
  20.     logger.info("running %s" % ' '.join(sys.argv))
  22.     # check and process input arguments
  23.     if len(sys.argv) < 4:
  24.         print globals()['__doc__'] % locals()
  25.         sys.exit(1)
  26.     inp, outp1, outp2 = sys.argv[1:4]
  28.     model = Word2Vec(LineSentence(inp), size=400, window=5, min_count=5,
  29.             workers=multiprocessing.cpu_count())
  31.     # trim unneeded model memory = use(much) less RAM
  32.     #model.init_sims(replace=True)
  33.     model.save(outp1)
  34.     model.save_word2vec_format(outp2, binary=False)

运行信息

  1. 2016-08-12 09:50:02,586: INFO: running python train_word2vec_model.py wiki.zh.text.seg wiki.zh.text.model wiki.zh.text.vector
  2. 2016-08-12 09:50:02,592: INFO: collecting all words and their counts
  3. 2016-08-12 09:50:02,592: INFO: PROGRESS: at sentence #0, processed 0 words and 0 word types
  4. 2016-08-12 09:50:12,476: INFO: PROGRESS: at sentence #10000, processed 12914562 words and 254662 word types
  5. 2016-08-12 09:50:20,215: INFO: PROGRESS: at sentence #20000, processed 22308801 words and 373573 word types
  6. 2016-08-12 09:50:28,448: INFO: PROGRESS: at sentence #30000, processed 30724902 words and 460837 word types
  7. ...
  8. 2016-08-12 09:52:03,498: INFO: PROGRESS: at sentence #210000, processed 143804601 words and 1483608 word types
  9. 2016-08-12 09:52:07,772: INFO: PROGRESS: at sentence #220000, processed 149352283 words and 1521199 word types
  10. 2016-08-12 09:52:11,639: INFO: PROGRESS: at sentence #230000, processed 154741839 words and 1563584 word types
  11. 2016-08-12 09:52:12,746: INFO: collected 1575172 word types from a corpus of 156430908 words and 232894 sentences
  12. 2016-08-12 09:52:13,672: INFO: total 278291 word types after removing those with count<5
  13. 2016-08-12 09:52:13,673: INFO: constructing a huffman tree from 278291 words
  14. 2016-08-12 09:52:29,323: INFO: built huffman tree with maximum node depth 25
  15. 2016-08-12 09:52:29,683: INFO: resetting layer weights
  16. 2016-08-12 09:52:38,805: INFO: training model with 4 workers on 278291 vocabulary and 400 features, using 'skipgram'=1 'hierarchical softmax'=1 'subsample'=0 and 'negative sampling'=0
  17. 2016-08-12 09:52:49,504: INFO: PROGRESS: at 0.10% words, alpha 0.02500, 15008 words/s
  18. 2016-08-12 09:52:51,935: INFO: PROGRESS: at 0.38% words, alpha 0.02500, 44434 words/s
  19. 2016-08-12 09:52:54,779: INFO: PROGRESS: at 0.56% words, alpha 0.02500, 53965 words/s
  20. 2016-08-12 09:52:57,240: INFO: PROGRESS: at 0.62% words, alpha 0.02491, 52116 words/s
  21. 2016-08-12 09:52:58,823: INFO: PROGRESS: at 0.72% words, alpha 0.02494, 55804 words/s
  22. 2016-08-12 09:53:03,649: INFO: PROGRESS: at 0.94% words, alpha 0.02486, 58277 words/s
  23. 2016-08-12 09:53:07,357: INFO: PROGRESS: at 1.03% words, alpha 0.02479, 56036 words/s
  24. ......
  25. 2016-08-12 19:22:09,002: INFO: PROGRESS: at 98.38% words, alpha 0.00044, 85936 words/s
  26. 2016-08-12 19:22:10,321: INFO: PROGRESS: at 98.50% words, alpha 0.00044, 85971 words/s
  27. 2016-08-12 19:22:11,934: INFO: PROGRESS: at 98.55% words, alpha 0.00039, 85940 words/s
  28. 2016-08-12 19:22:13,384: INFO: PROGRESS: at 98.65% words, alpha 0.00036, 85960 words/s
  29. 2016-08-12 19:22:13,883: INFO: training on 152625573 words took 1775.1s, 85982 words/s
  30. 2016-08-12 19:22:13,883: INFO: saving Word2Vec object under wiki.zh.text.model, separately None
  31. 2016-08-12 19:22:13,884: INFO: not storing attribute syn0norm
  32. 2016-08-12 19:22:13,884: INFO: storing numpy array 'syn0' to wiki.zh.text.model.syn0.npy
  33. 2016-08-12 19:22:20,797: INFO: storing numpy array 'syn1' to wiki.zh.text.model.syn1.npy
  34. 2016-08-12 19:22:40,667: INFO: storing 278291x400 projection weights into wiki.zh.text.vector

测试模型效果：

  1. In [1]: import gensim
  3. In [2]: model = gensim.models.Word2Vec.load("wiki.zh.text.model")
  5. In [3]: model.most_similar(u"足球")
  6. Out[3]: 
  7. [(u'\u8054\u8d5b', 0.6553816199302673),
  8.  (u'\u7532\u7ea7', 0.6530429720878601),
  9.  (u'\u7bee\u7403', 0.5967546701431274),
  10.  (u'\u4ff1\u4e50\u90e8', 0.5872289538383484),
  11.  (u'\u4e59\u7ea7', 0.5840631723403931),
  12.  (u'\u8db3\u7403\u961f', 0.5560152530670166),
  13.  (u'\u4e9a\u8db3\u8054', 0.5308005809783936),
  14.  (u'allsvenskan', 0.5249762535095215),
  15.  (u'\u4ee3\u8868\u961f', 0.5214947462081909),
  16.  (u'\u7532\u7ec4', 0.5177896022796631)]
  18. In [4]: result = model.most_similar(u"足球")
  20. In [5]: for e in result:
  21.     print e[0], e[1]
  22.    ....:     
  23. 联赛 0.65538161993
  24. 甲级 0.653042972088
  25. 篮球 0.596754670143
  26. 俱乐部 0.587228953838
  27. 乙级 0.58406317234
  28. 足球队 0.556015253067
  29. 亚足联 0.530800580978
  30. allsvenskan 0.52497625351
  31. 代表队 0.521494746208
  32. 甲组 0.51778960228
  34. In [6]: result = model.most_similar(u"男人")
  36. In [7]: for e in result:
  37.     print e[0], e[1]
  38.    ....:     
  39. 女人 0.77537125349
  40. 家伙 0.617369174957
  41. 妈妈 0.567102909088
  42. 漂亮 0.560832381248
  43. 잘했어 0.540875017643
  44. 谎言 0.538448691368
  45. 爸爸 0.53660941124
  46. 傻瓜 0.535608053207
  47. 예쁘다 0.535151124001
  48. mc刘 0.529670000076
  50. In [8]: result = model.most_similar(u"女人")
  52. In [9]: for e in result:
  53.     print e[0], e[1]
  54.    ....:     
  55. 男人 0.77537125349
  56. 我的某 0.589010596275
  57. 妈妈 0.576344847679
  58. 잘했어 0.562340974808
  59. 美丽 0.555426716805
  60. 爸爸 0.543958246708
  61. 新娘 0.543640494347
  62. 谎言 0.540272831917
  63. 妞儿 0.531066179276
  64. 老婆 0.528521537781
  66. In [10]: result = model.most_similar(u"青蛙")
  68. In [11]: for e in result:
  69.     print e[0], e[1]
  70.    ....:     
  71. 老鼠 0.559612870216
  72. 乌龟 0.489831030369
  73. 蜥蜴 0.478990525007
  74. 猫 0.46728849411
  75. 鳄鱼 0.461885392666
  76. 蟾蜍 0.448014199734
  77. 猴子 0.436584025621
  78. 白雪公主 0.434905380011
  79. 蚯蚓 0.433413207531
  80. 螃蟹 0.4314712286
  82. In [12]: result = model.most_similar(u"姨夫")
  84. In [13]: for e in result:
  85.     print e[0], e[1]
  86.    ....:     
  87. 堂伯 0.583935439587
  88. 祖父 0.574735701084
  89. 妃所生 0.569327116013
  90. 内弟 0.562012672424
  91. 早卒 0.558042645454
  92. 曕 0.553856015205
  93. 胤祯 0.553288519382
  94. 陈潜 0.550716996193
  95. 愔之 0.550510883331
  96. 叔父 0.550032019615
  98. In [14]: result = model.most_similar(u"衣服")
  100. In [15]: for e in result:
  101.     print e[0], e[1]
  102.    ....:     
  103. 鞋子 0.686688780785
  104. 穿着 0.672499775887
  105. 衣物 0.67173999548
  106. 大衣 0.667605519295
  107. 裤子 0.662670075893
  108. 内裤 0.662210345268
  109. 裙子 0.659705817699
  110. 西装 0.648508131504
  111. 洋装 0.647238850594
  112. 围裙 0.642895817757
  114. In [16]: result = model.most_similar(u"公安局")
  116. In [17]: for e in result:
  117.     print e[0], e[1]
  118.    ....:     
  119. 司法局 0.730189085007
  120. 公安厅 0.634275555611
  121. 公安 0.612798035145
  122. 房管局 0.597343325615
  123. 商业局 0.597183346748
  124. 军管会 0.59476184845
  125. 体育局 0.59283208847
  126. 财政局 0.588721752167
  127. 戒毒所 0.575558543205
  128. 新闻办 0.573395550251
  130. In [18]: result = model.most_similar(u"铁道部")
  132. In [19]: for e in result:
  133.     print e[0], e[1]
  134.    ....:     
  135. 盛光祖 0.565509021282
  136. 交通部 0.548688530922
  137. 批复 0.546967327595
  138. 刘志军 0.541010737419
  139. 立项 0.517836689949
  140. 报送 0.510296344757
  141. 计委 0.508456230164
  142. 水利部 0.503531932831
  143. 国务院 0.503227233887
  144. 经贸委 0.50156635046
  146. In [20]: result = model.most_similar(u"清华大学")
  148. In [21]: for e in result:
  149.     print e[0], e[1]
  150.    ....:     
  151. 北京大学 0.763922810555
  152. 化学系 0.724210739136
  153. 物理系 0.694550514221
  154. 数学系 0.684280991554
  155. 中山大学 0.677202701569
  156. 复旦 0.657914161682
  157. 师范大学 0.656435549259
  158. 哲学系 0.654701948166
  159. 生物系 0.654403865337
  160. 中文系 0.653147578239
  162. In [22]: result = model.most_similar(u"卫视")
  164. In [23]: for e in result:
  165.     print e[0], e[1]
  166.    ....:     
  167. 湖南 0.676812887192
  168. 中文台 0.626506924629
  169. 収蔵 0.621356606483
  170. 黄金档 0.582251906395
  171. cctv 0.536769032478
  172. 安徽 0.536752820015
  173. 非同凡响 0.534517168999
  174. 唱响 0.533438682556
  175. 最强音 0.532605051994
  176. 金鹰 0.531676828861
  178. In [24]: result = model.most_similar(u"习近平")
  180. In [25]: for e in result:
  181.     print e[0], e[1]
  182.    ....:     
  183. 胡锦涛 0.809472680092
  184. 江泽民 0.754633367062
  185. 李克强 0.739740967751
  186. 贾庆林 0.737033963203
  187. 曾庆红 0.732847094536
  188. 吴邦国 0.726941585541
  189. 总书记 0.719057679176
  190. 李瑞环 0.716384887695
  191. 温家宝 0.711952567101
  192. 王岐山 0.703570842743
  194. In [26]: result = model.most_similar(u"林丹")
  196. In [27]: for e in result:
  197.     print e[0], e[1]
  198.    ....:     
  199. 黄综翰 0.538035452366
  200. 蒋燕皎 0.52646958828
  201. 刘鑫 0.522252976894
  202. 韩晶娜 0.516120731831
  203. 王晓理 0.512289524078
  204. 王适 0.508560419083
  205. 杨影 0.508159279823
  206. 陈跃 0.507353425026
  207. 龚智超 0.503159761429
  208. 李敬元 0.50262516737
  210. In [28]: result = model.most_similar(u"语言学")
  212. In [29]: for e in result:
  213.     print e[0], e[1]
  214.    ....:     
  215. 社会学 0.632598280907
  216. 人类学 0.623406708241
  217. 历史学 0.618442356586
  218. 比较文学 0.604823827744
  219. 心理学 0.600066184998
  220. 人文科学 0.577783346176
  221. 社会心理学 0.575571238995
  222. 政治学 0.574541330338
  223. 地理学 0.573896467686
  224. 哲学 0.573873817921
  226. In [30]: result = model.most_similar(u"计算机")
  228. In [31]: for e in result:
  229.     print e[0], e[1]
  230.    ....:     
  231. 自动化 0.674171924591
  232. 应用 0.614087462425
  233. 自动化系 0.611132860184
  234. 材料科学 0.607891201973
  235. 集成电路 0.600370049477
  236. 技术 0.597518980503
  237. 电子学 0.591316461563
  238. 建模 0.577238917351
  239. 工程学 0.572855889797
  240. 微电子 0.570086717606
  242. In [32]: model.similarity(u"计算机", u"自动化")
  243. Out[32]: 0.67417196002404789
  245. In [33]: model.similarity(u"女人", u"男人")
  246. Out[33]: 0.77537125129824813
  248. In [34]: model.doesnt_match(u"早餐 晚餐 午餐 中心".split())
  249. Out[34]: u'\u4e2d\u5fc3'
  251. In [35]: print model.doesnt_match(u"早餐 晚餐 午餐 中心".split())
  252. 中心
添加新批注
在作者公开此批注前，只有你和作者可见。
回复批注