我爱自然语言处理

I Love Natural Language Processing

自然语言处理 语义相似度 语言模型

中英文维基百科语料上的Word2Vec实验

作者52nlp

3 月 12, 2015 #gensim, #gensim word2vec, #glove, #Mecab, #python gensim, #python glove, #python word2vec, #word2vec, #word2vec实验, #word2vec应用, #word2vec模型, #word2vec相似度, #word2vec相似度计算, #word2vec词语相似度, #中文分词, #中文繁简转换, #中文维基百科语料, #中文编码转换, #文档相似度, #深度学习, #相似度, #维基百科, #维基百科语料, #英文维基百科语料, #词语相似度, #语义相似度, #语言模型

注：目前可以直接在AINLP公众号上体验腾讯词向量，公众号对话直接输入：相似词 词条

最近试了一下Word2Vec, GloVe 以及对应的python版本 gensim word2vec 和 python-glove，就有心在一个更大规模的语料上测试一下，自然而然维基百科的语料进入了视线。维基百科官方提供了一个很好的维基百科数据源：https://dumps.wikimedia.org，可以方便的下载多种语言多种格式的维基百科数据。此前通过gensim的玩过英文的维基百科语料并训练LSI，LDA模型来计算两个文档的相似度，所以想看看gensim有没有提供一种简便的方式来处理维基百科数据，训练word2vec模型，用于计算词语之间的语义相似度。感谢Google，在gensim的google group下，找到了一个很长的讨论帖：training word2vec on full Wikipedia ，这个帖子基本上把如何使用gensim在维基百科语料上训练word2vec模型的问题说清楚了，甚至参与讨论的gensim的作者Radim Řehůřek博士还在新的gensim版本里加了一点修正，而对于我来说，所做的工作就是做一下验证而已。虽然github上有一个wiki2vec的项目也是做得这个事，不过我更喜欢用python gensim的方式解决问题。

关于word2vec，这方面无论中英文的参考资料相当的多，英文方面既可以看官方推荐的论文，也可以看gensim作者Radim Řehůřek博士写得一些文章。而中文方面，推荐 @licstar的《Deep Learning in NLP （一）词向量和语言模型》，有道技术沙龙的《Deep Learning实战之word2vec》，@飞林沙 的《word2vec的学习思路》, falao_beiliu 的《深度学习word2vec笔记之基础篇》和《深度学习word2vec笔记之算法篇》等。

一、英文维基百科的Word2Vec测试

首先测试了英文维基百科的数据，下载的是xml压缩后的最新数据（下载日期是2015年3月1号），大概11G，下载地址：

https://dumps.wikimedia.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2

处理包括两个阶段，首先将xml的wiki数据转换为text格式，通过下面这个脚本(process_wiki.py)实现：
注：因为很多同学留言是在python3.x环境下使用遇到问题，这里修改了一个版本兼容python2.x和python3.x, Ubuntu16.04下测试有效(2017.5.1)

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Author: Pan Yang (panyangnlp@gmail.com)
# Copyrigh 2017
 
from __future__ import print_function
 
import logging
import os.path
import six
import sys
 
from gensim.corpora import WikiCorpus
 
if __name__ == '__main__':
    program = os.path.basename(sys.argv[0])
    logger = logging.getLogger(program)
 
    logging.basicConfig(format='%(asctime)s: %(levelname)s: %(message)s')
    logging.root.setLevel(level=logging.INFO)
    logger.info("running %s" % ' '.join(sys.argv))
 
    # check and process input arguments
    if len(sys.argv) != 3:
        print("Using: python process_wiki.py enwiki.xxx.xml.bz2 wiki.en.text")
        sys.exit(1)
    inp, outp = sys.argv[1:3]
    space = " "
    i = 0
 
    output = open(outp, 'w')
    wiki = WikiCorpus(inp, lemmatize=False, dictionary={})
    for text in wiki.get_texts():
        if six.PY3:
            output.write(b' '.join(text).decode('utf-8') + '\n')
        #   ###another method###
        #    output.write(
        #            space.join(map(lambda x:x.decode("utf-8"), text)) + '\n')
        else:
            output.write(space.join(text) + "\n")
        i = i + 1
        if (i % 10000 == 0):
            logger.info("Saved " + str(i) + " articles")
 
    output.close()
    logger.info("Finished Saved " + str(i) + " articles")

这里利用了gensim里的维基百科处理类WikiCorpus，通过get_texts将维基里的每篇文章转换位1行text文本，并且去掉了标点符号等内容，注意这里“wiki = WikiCorpus(inp, lemmatize=False, dictionary={})”将lemmatize设置为False的主要目的是不使用pattern模块来进行英文单词的词干化处理，无论你的电脑是否已经安装了pattern，因为使用pattern会严重影响这个处理过程，变得很慢。

执行"python process_wiki.py enwiki-latest-pages-articles.xml.bz2 wiki.en.text":

2015-03-07 15:08:39,181: INFO: running process_enwiki.py enwiki-latest-pages-articles.xml.bz2 wiki.en.text
2015-03-07 15:11:12,860: INFO: Saved 10000 articles
2015-03-07 15:13:25,369: INFO: Saved 20000 articles
2015-03-07 15:15:19,771: INFO: Saved 30000 articles
2015-03-07 15:16:58,424: INFO: Saved 40000 articles
2015-03-07 15:18:12,374: INFO: Saved 50000 articles
2015-03-07 15:19:03,213: INFO: Saved 60000 articles
2015-03-07 15:19:47,656: INFO: Saved 70000 articles
2015-03-07 15:20:29,135: INFO: Saved 80000 articles
2015-03-07 15:22:02,365: INFO: Saved 90000 articles
2015-03-07 15:23:40,141: INFO: Saved 100000 articles
.....
2015-03-07 19:33:16,549: INFO: Saved 3700000 articles
2015-03-07 19:33:49,493: INFO: Saved 3710000 articles
2015-03-07 19:34:23,442: INFO: Saved 3720000 articles
2015-03-07 19:34:57,984: INFO: Saved 3730000 articles
2015-03-07 19:35:31,976: INFO: Saved 3740000 articles
2015-03-07 19:36:05,790: INFO: Saved 3750000 articles
2015-03-07 19:36:32,392: INFO: finished iterating over Wikipedia corpus of 3758076 documents with 2018886604 positions (total 15271374 articles, 2075130438 positions before pruning articles shorter than 50 words)
2015-03-07 19:36:32,394: INFO: Finished Saved 3758076 articles

在我的macpro（4核16G机器）大约跑了4个半小时，处理了375万的文章后，我们得到了一个12G的text格式的英文维基百科数据wiki.en.text，格式类似这样的：

anarchism is collection of movements and ideologies that hold the state to be undesirable unnecessary or harmful these movements advocate some form of stateless society instead often based on self governed voluntary institutions or non hierarchical free associations although anti statism is central to anarchism as political philosophy anarchism also entails rejection of and often hierarchical organisation in general as an anti dogmatic philosophy anarchism draws on many currents of thought and strategy anarchism does not offer fixed body of doctrine from single particular world view instead fluxing and flowing as philosophy there are many types and traditions of anarchism not all of which are mutually exclusive anarchist schools of thought can differ fundamentally supporting anything from extreme individualism to complete collectivism strains of anarchism have often been divided into the categories of social and individualist anarchism or similar dual classifications anarchism is usually considered radical left wing ideology and much of anarchist economics and anarchist legal philosophy reflect anti authoritarian interpretations of communism collectivism syndicalism mutualism or participatory economics etymology and terminology the term anarchism is compound word composed from the word anarchy and the suffix ism themselves derived respectively from the greek anarchy from anarchos meaning one without rulers from the privative prefix ἀν an without and archos leader ruler cf archon or arkhē authority sovereignty realm magistracy and the suffix or ismos isma from the verbal infinitive suffix...

有了这个数据后，无论用原始的word2vec binary版本还是gensim中的python word2vec版本，都可以用来训练word2vec模型，不过我们试了一下前者，发现很慢，所以还是采用google group 讨论帖中的gensim word2vec方式的训练脚本，不过做了一点修改，保留了vector text格式的输出，方便debug, 脚本train_word2vec_model.py如下：

#!/usr/bin/env python
# -*- coding: utf-8 -*-
 
import logging
import os
import sys
import multiprocessing
 
from gensim.models import Word2Vec
from gensim.models.word2vec import LineSentence
 
if __name__ == '__main__':
    program = os.path.basename(sys.argv[0])
    logger = logging.getLogger(program)
 
    logging.basicConfig(format='%(asctime)s: %(levelname)s: %(message)s')
    logging.root.setLevel(level=logging.INFO)
    logger.info("running %s" % ' '.join(sys.argv))
 
    # check and process input arguments
    if len(sys.argv) < 4:
        print(globals()['__doc__'] % locals())
        sys.exit(1)
    inp, outp1, outp2 = sys.argv[1:4]
 
    model = Word2Vec(LineSentence(inp), size=400, window=5, min_count=5,
                     workers=multiprocessing.cpu_count())
 
    # trim unneeded model memory = use(much) less RAM
    # model.init_sims(replace=True)
    model.save(outp1)
    model.wv.save_word2vec_format(outp2, binary=False)

执行 "python train_word2vec_model.py wiki.en.text wiki.en.text.model wiki.en.text.vector":

2015-03-09 22:48:29,588: INFO: running train_word2vec_model.py wiki.en.text wiki.en.text.model wiki.en.text.vector
2015-03-09 22:48:29,593: INFO: collecting all words and their counts
2015-03-09 22:48:29,607: INFO: PROGRESS: at sentence #0, processed 0 words and 0 word types
2015-03-09 22:48:50,686: INFO: PROGRESS: at sentence #10000, processed 29353579 words and 430650 word types
2015-03-09 22:49:08,476: INFO: PROGRESS: at sentence #20000, processed 54695775 words and 610833 word types
2015-03-09 22:49:22,985: INFO: PROGRESS: at sentence #30000, processed 75344844 words and 742274 word types
2015-03-09 22:49:35,607: INFO: PROGRESS: at sentence #40000, processed 93430415 words and 859131 word types
2015-03-09 22:49:44,125: INFO: PROGRESS: at sentence #50000, processed 106057188 words and 935606 word types
2015-03-09 22:49:49,185: INFO: PROGRESS: at sentence #60000, processed 114319016 words and 952771 word types
2015-03-09 22:49:53,316: INFO: PROGRESS: at sentence #70000, processed 121263134 words and 969526 word types
2015-03-09 22:49:57,268: INFO: PROGRESS: at sentence #80000, processed 127773799 words and 984130 word types
2015-03-09 22:50:07,593: INFO: PROGRESS: at sentence #90000, processed 142688762 words and 1062932 word types
2015-03-09 22:50:19,162: INFO: PROGRESS: at sentence #100000, processed 159550824 words and 1157644 word 
types
......
2015-03-09 23:11:52,977: INFO: PROGRESS: at sentence #3700000, processed 1999452503 words and 7990138 word types
2015-03-09 23:11:55,367: INFO: PROGRESS: at sentence #3710000, processed 2002777270 words and 8002903 word types
2015-03-09 23:11:57,842: INFO: PROGRESS: at sentence #3720000, processed 2006213923 words and 8019620 word types
2015-03-09 23:12:00,439: INFO: PROGRESS: at sentence #3730000, processed 2009762733 words and 8035408 word types
2015-03-09 23:12:02,793: INFO: PROGRESS: at sentence #3740000, processed 2013066196 words and 8045218 word types
2015-03-09 23:12:05,178: INFO: PROGRESS: at sentence #3750000, processed 2016363087 words and 8057784 word types
2015-03-09 23:12:07,013: INFO: collected 8069236 word types from a corpus of 2018886604 words and 3758076 sentences
2015-03-09 23:12:12,230: INFO: total 1969354 word types after removing those with count<5
2015-03-09 23:12:12,230: INFO: constructing a huffman tree from 1969354 words
2015-03-09 23:14:07,415: INFO: built huffman tree with maximum node depth 29
2015-03-09 23:14:09,790: INFO: resetting layer weights
2015-03-09 23:15:04,506: INFO: training model with 4 workers on 1969354 vocabulary and 400 features, using 'skipgram'=1 'hierarchical softmax'=1 'subsample'=0 and 'negative sampling'=0
2015-03-09 23:15:19,112: INFO: PROGRESS: at 0.01% words, alpha 0.02500, 19098 words/s
2015-03-09 23:15:20,224: INFO: PROGRESS: at 0.03% words, alpha 0.02500, 37671 words/s
2015-03-09 23:15:22,305: INFO: PROGRESS: at 0.07% words, alpha 0.02500, 75393 words/s
2015-03-09 23:15:27,712: INFO: PROGRESS: at 0.08% words, alpha 0.02499, 65618 words/s
2015-03-09 23:15:29,452: INFO: PROGRESS: at 0.09% words, alpha 0.02500, 70966 words/s
2015-03-09 23:15:34,032: INFO: PROGRESS: at 0.11% words, alpha 0.02498, 77369 words/s
2015-03-09 23:15:37,249: INFO: PROGRESS: at 0.12% words, alpha 0.02498, 74935 words/s
2015-03-09 23:15:40,618: INFO: PROGRESS: at 0.14% words, alpha 0.02498, 75399 words/s
2015-03-09 23:15:42,301: INFO: PROGRESS: at 0.16% words, alpha 0.02497, 86029 words/s
2015-03-09 23:15:46,283: INFO: PROGRESS: at 0.17% words, alpha 0.02497, 83033 words/s
2015-03-09 23:15:48,374: INFO: PROGRESS: at 0.18% words, alpha 0.02497, 83370 words/s
2015-03-09 23:15:51,398: INFO: PROGRESS: at 0.19% words, alpha 0.02496, 82794 words/s
2015-03-09 23:15:55,069: INFO: PROGRESS: at 0.21% words, alpha 0.02496, 83753 words/s
2015-03-09 23:15:57,718: INFO: PROGRESS: at 0.23% words, alpha 0.02496, 85031 words/s
2015-03-09 23:16:00,106: INFO: PROGRESS: at 0.24% words, alpha 0.02495, 86567 words/s
2015-03-09 23:16:05,523: INFO: PROGRESS: at 0.26% words, alpha 0.02495, 84850 words/s
2015-03-09 23:16:06,596: INFO: PROGRESS: at 0.27% words, alpha 0.02495, 87926 words/s
2015-03-09 23:16:09,500: INFO: PROGRESS: at 0.29% words, alpha 0.02494, 88618 words/s
2015-03-09 23:16:10,714: INFO: PROGRESS: at 0.30% words, alpha 0.02494, 91023 words/s
2015-03-09 23:16:18,467: INFO: PROGRESS: at 0.32% words, alpha 0.02494, 85960 words/s
2015-03-09 23:16:19,547: INFO: PROGRESS: at 0.33% words, alpha 0.02493, 89140 words/s
2015-03-09 23:16:23,500: INFO: PROGRESS: at 0.36% words, alpha 0.02493, 92026 words/s
2015-03-09 23:16:29,738: INFO: PROGRESS: at 0.37% words, alpha 0.02491, 88180 words/s
2015-03-09 23:16:32,000: INFO: PROGRESS: at 0.40% words, alpha 0.02492, 92734 words/s
2015-03-09 23:16:34,392: INFO: PROGRESS: at 0.42% words, alpha 0.02491, 93300 words/s
2015-03-09 23:16:41,018: INFO: PROGRESS: at 0.43% words, alpha 0.02490, 89727 words/s
.......
2015-03-10 05:03:31,849: INFO: PROGRESS: at 99.20% words, alpha 0.00020, 95350 words/s
2015-03-10 05:03:32,901: INFO: PROGRESS: at 99.21% words, alpha 0.00020, 95350 words/s
2015-03-10 05:03:34,296: INFO: PROGRESS: at 99.21% words, alpha 0.00020, 95350 words/s
2015-03-10 05:03:35,635: INFO: PROGRESS: at 99.22% words, alpha 0.00020, 95349 words/s
2015-03-10 05:03:36,730: INFO: PROGRESS: at 99.22% words, alpha 0.00020, 95350 words/s
2015-03-10 05:03:37,489: INFO: reached the end of input; waiting to finish 8 outstanding jobs
2015-03-10 05:03:37,908: INFO: PROGRESS: at 99.23% words, alpha 0.00019, 95350 words/s
2015-03-10 05:03:39,028: INFO: PROGRESS: at 99.23% words, alpha 0.00019, 95350 words/s
2015-03-10 05:03:40,127: INFO: PROGRESS: at 99.24% words, alpha 0.00019, 95350 words/s
2015-03-10 05:03:40,910: INFO: training on 1994415728 words took 20916.4s, 95352 words/s
2015-03-10 05:03:41,058: INFO: saving Word2Vec object under wiki.en.text.model, separately None
2015-03-10 05:03:41,209: INFO: not storing attribute syn0norm
2015-03-10 05:03:41,209: INFO: storing numpy array 'syn0' to wiki.en.text.model.syn0.npy
2015-03-10 05:04:35,199: INFO: storing numpy array 'syn1' to wiki.en.text.model.syn1.npy
2015-03-10 05:11:25,400: INFO: storing 1969354x400 projection weights into wiki.en.text.vector

大约跑了7个小时，我们得到了一个gensim中默认格式的word2vec model和一个原始c版本word2vec的vector格式的模型: wiki.en.text.vector，格式如下：

1969354 400
the 0.129255 0.015725 0.049174 -0.016438 -0.018912 0.032752 0.079885 0.033669 -0.077722 -0.025709 0.012775 0.044153 0.134307 0.070499 -0.002243 0.105198 -0.016832 -0.028631 -0.124312 -0.123064 -0.116838 0.051181 -0.096058 -0.049734 0.017380 -0.101221 0.058945 0.013669 -0.012755 0.061053 0.061813 0.083655 -0.069382 -0.069868 0.066529 -0.037156 -0.072935 -0.009470 0.037412 -0.004406 0.047011 0.005033 -0.066270 -0.031815 0.023160 -0.080117 0.172918 0.065486 -0.072161 0.062875 0.019939 -0.048380 0.198152 -0.098525 0.023434 0.079439 0.045150 -0.079479 -0.051441 -0.021556 -0.024981 -0.045291 0.040284 -0.082500 0.014618 -0.071998 0.031887 0.043916 0.115783 -0.174898 0.086603 -0.023124 0.007293 -0.066576 -0.164817 -0.081223 0.058412 0.000132 0.064160 0.055848 0.029776 -0.103420 -0.007541 -0.031742 0.082533 -0.061760 -0.038961 0.001754 -0.023977 0.069616 0.095920 0.017136 0.067126 -0.111310 0.053632 0.017633 -0.003875 -0.005236 0.063151 0.039729 -0.039158 0.001415 0.021754 -0.012540 0.015070 -0.062636 -0.013605 -0.031770 0.005296 -0.078119 -0.069303 -0.080634 -0.058377 0.024398 -0.028173 0.026353 0.088662 0.018755 -0.113538 0.055538 -0.086012 -0.027708 -0.028788 0.017759 0.029293 0.047674 -0.106734 -0.134380 0.048605 -0.089583 0.029426 0.030552 0.141916 -0.022653 0.017204 -0.036059 0.061045 -0.000077 -0.076579 0.066747 0.060884 -0.072817...
...

在ipython中，我们通过gensim来加载和测试这个模型，因为这个模型大约有7G，所以加载的时间也稍长一些：

In [2]: import gensim
 
# 注：因为gensim版本更新的问题，如果下面这个load有问题，可以使用新的接口：model = gensim.models.word2vec.Word2Vec.load(MODEL_PATH)
In [3]: model = gensim.models.Word2Vec.load_word2vec_format("wiki.en.text.vector", binary=False)
 
In [4]: model.most_similar("queen")
Out[4]: 
[(u'princess', 0.5760838389396667),
 (u'hyoui', 0.5671186447143555),
 (u'janggyung', 0.5598698854446411),
 (u'king', 0.5556215047836304),
 (u'dollallolla', 0.5540223121643066),
 (u'loranella', 0.5522741079330444),
 (u'ramphaiphanni', 0.5310937166213989),
 (u'jeheon', 0.5298476219177246),
 (u'soheon', 0.5243583917617798),
 (u'coronation', 0.5217245221138)]
 
In [5]: model.most_similar("man")
Out[5]: 
[(u'woman', 0.7120707035064697),
 (u'girl', 0.58659827709198),
 (u'handsome', 0.5637181997299194),
 (u'boy', 0.5425317287445068),
 (u'villager', 0.5084836483001709),
 (u'mustachioed', 0.49287813901901245),
 (u'mcgucket', 0.48355430364608765),
 (u'spider', 0.4804879426956177),
 (u'policeman', 0.4780033826828003),
 (u'stranger', 0.4750771224498749)]
 
In [6]: model.most_similar("woman")
Out[6]: 
[(u'man', 0.7120705842971802),
 (u'girl', 0.6736541986465454),
 (u'prostitute', 0.5765659809112549),
 (u'divorcee', 0.5429972410202026),
 (u'person', 0.5276163816452026),
 (u'schoolgirl', 0.5102938413619995),
 (u'housewife', 0.48748138546943665),
 (u'lover', 0.4858251214027405),
 (u'handsome', 0.4773051142692566),
 (u'boy', 0.47445783019065857)]
 
In [8]: model.similarity("woman", "man")
Out[8]: 0.71207063453821218
 
In [10]: model.doesnt_match("breakfast cereal dinner lunch".split())
Out[10]: 'cereal'
 
In [11]: model.similarity("woman", "girl")
Out[11]: 0.67365416785207421
 
In [13]: model.most_similar("frog")
Out[13]: 
[(u'toad', 0.6868536472320557),
 (u'barycragus', 0.6607867479324341),
 (u'grylio', 0.626731276512146),
 (u'heckscheri', 0.6208407878875732),
 (u'clamitans', 0.6150864362716675),
 (u'coplandi', 0.612680196762085),
 (u'pseudacris', 0.6108512878417969),
 (u'litoria', 0.6084023714065552),
 (u'raniformis', 0.6044802665710449),
 (u'watjulumensis', 0.6043726205825806)]

一切ok，但是当加载gensim默认的基于numpy格式的模型时，却遇到了问题：

In [1]: import gensim 
 
In [2]: model = gensim.models.Word2Vec.load("wiki.en.text.model")
 
In [3]: model.most_similar("man")
... RuntimeWarning: invalid value encountered in divide
  self.syn0norm = (self.syn0 / sqrt((self.syn0 ** 2).sum(-1))[..., newaxis]).astype(REAL)
 
Out[3]: 
[(u'ahsns', nan),
 (u'ny\xedl', nan),
 (u'indradeo', nan),
 (u'jaimovich', nan),
 (u'addlepate', nan),
 (u'jagello', nan),
 (u'festenburg', nan),
 (u'picatic', nan),
 (u'tolosanum', nan),
 (u'mithoo', nan)]

这也是我修改前面这个脚本的原因所在，这个脚本在训练小一些的数据，譬如前10万条text的时候没任何问题，无论原始格式还是gensim格式，但是当跑完这个英文维基百科的时候，却存在这个问题，试了一些方法解决，还没有成功，如果大家有好的建议或解决方案，欢迎提出。

二、中文维基百科的Word2Vec测试

测试完英文维基百科之后，自然想试试中文的维基百科数据，与英文处理过程相似，也分两个步骤，不过这里需要对中文维基百科数据特殊处理一下，包括繁简转换，中文分词，去除非utf-8字符等。中文数据的下载地址是：https://dumps.wikimedia.org/zhwiki/latest/zhwiki-latest-pages-articles.xml.bz2。

中文维基百科的数据比较小，整个xml的压缩文件大约才1G，相对英文数据小了很多。首先用 process_wiki.py处理这个XML压缩文件，执行：python process_wiki.py zhwiki-latest-pages-articles.xml.bz2 wiki.zh.text

2015-03-11 17:39:22,739: INFO: running process_wiki.py zhwiki-latest-pages-articles.xml.bz2 wiki.zh.text
2015-03-11 17:40:08,329: INFO: Saved 10000 articles
2015-03-11 17:40:45,501: INFO: Saved 20000 articles
2015-03-11 17:41:23,659: INFO: Saved 30000 articles
2015-03-11 17:42:01,748: INFO: Saved 40000 articles
2015-03-11 17:42:33,779: INFO: Saved 50000 articles
......
2015-03-11 17:55:23,094: INFO: Saved 200000 articles
2015-03-11 17:56:14,692: INFO: Saved 210000 articles
2015-03-11 17:57:04,614: INFO: Saved 220000 articles
2015-03-11 17:57:57,979: INFO: Saved 230000 articles
2015-03-11 17:58:16,621: INFO: finished iterating over Wikipedia corpus of 232894 documents with 51603419 positions (total 2581444 articles, 62177405 positions before pruning articles shorter than 50 words)
2015-03-11 17:58:16,622: INFO: Finished Saved 232894 articles

得到了大约23万多篇中文语料的text格式的语料:wiki.zh.text，大概750多M。不过查看之后发现，除了加杂一些英文词汇外，还有很多繁体字混迹其中，这里还是参考了 @licstar 《维基百科简体中文语料的获取》中的方法，安装opencc，然后将wiki.zh.text中的繁体字转化位简体字：

opencc -i wiki.zh.text -o wiki.zh.text.jian -c zht2zhs.ini

然后就是分词处理了，这次我用基于MeCab训练的一套中文分词系统来进行中文分词，目前虽还没有达到实用的状态，但是性能和分词结果基本能达到这次的使用要求：

mecab -d ../data/ -O wakati wiki.zh.text.jian -o wiki.zh.text.jian.seg -b 10000000

注意这里data目录下是给mecab训练好的分词模型和词典文件等，详细可参考《用MeCab打造一套实用的中文分词系统》。

有了中文维基百科的分词数据，还以为就可以执行word2vec模型训练了：

python train_word2vec_model.py wiki.zh.text.jian.seg wiki.zh.text.model wiki.zh.text.vector

不过仍然遇到了问题，提示的错误是：

UnicodeDecodeError: 'utf8' codec can't decode bytes in position 5394-5395: invalid continuation byte

google了一下，大致是文件中包含非utf-8字符，又用iconv处理了一下这个问题：

iconv -c -t UTF-8 < wiki.zh.text.jian.seg > wiki.zh.text.jian.seg.utf-8

这样基本上就没问题了，执行：

python train_word2vec_model.py wiki.zh.text.jian.seg.utf-8 wiki.zh.text.model wiki.zh.text.vector

2015-03-11 18:50:02,586: INFO: running train_word2vec_model.py wiki.zh.text.jian.seg.utf-8 wiki.zh.text.model wiki.zh.text.vector
2015-03-11 18:50:02,592: INFO: collecting all words and their counts
2015-03-11 18:50:02,592: INFO: PROGRESS: at sentence #0, processed 0 words and 0 word types
2015-03-11 18:50:12,476: INFO: PROGRESS: at sentence #10000, processed 12914562 words and 254662 word types
2015-03-11 18:50:20,215: INFO: PROGRESS: at sentence #20000, processed 22308801 words and 373573 word types
2015-03-11 18:50:28,448: INFO: PROGRESS: at sentence #30000, processed 30724902 words and 460837 word types
...
2015-03-11 18:52:03,498: INFO: PROGRESS: at sentence #210000, processed 143804601 words and 1483608 word types
2015-03-11 18:52:07,772: INFO: PROGRESS: at sentence #220000, processed 149352283 words and 1521199 word types
2015-03-11 18:52:11,639: INFO: PROGRESS: at sentence #230000, processed 154741839 words and 1563584 word types
2015-03-11 18:52:12,746: INFO: collected 1575172 word types from a corpus of 156430908 words and 232894 sentences
2015-03-11 18:52:13,672: INFO: total 278291 word types after removing those with count<5
2015-03-11 18:52:13,673: INFO: constructing a huffman tree from 278291 words
2015-03-11 18:52:29,323: INFO: built huffman tree with maximum node depth 25
2015-03-11 18:52:29,683: INFO: resetting layer weights
2015-03-11 18:52:38,805: INFO: training model with 4 workers on 278291 vocabulary and 400 features, using 'skipgram'=1 'hierarchical softmax'=1 'subsample'=0 and 'negative sampling'=0
2015-03-11 18:52:49,504: INFO: PROGRESS: at 0.10% words, alpha 0.02500, 15008 words/s
2015-03-11 18:52:51,935: INFO: PROGRESS: at 0.38% words, alpha 0.02500, 44434 words/s
2015-03-11 18:52:54,779: INFO: PROGRESS: at 0.56% words, alpha 0.02500, 53965 words/s
2015-03-11 18:52:57,240: INFO: PROGRESS: at 0.62% words, alpha 0.02491, 52116 words/s
2015-03-11 18:52:58,823: INFO: PROGRESS: at 0.72% words, alpha 0.02494, 55804 words/s
2015-03-11 18:53:03,649: INFO: PROGRESS: at 0.94% words, alpha 0.02486, 58277 words/s
2015-03-11 18:53:07,357: INFO: PROGRESS: at 1.03% words, alpha 0.02479, 56036 words/s
......
2015-03-11 19:22:09,002: INFO: PROGRESS: at 98.38% words, alpha 0.00044, 85936 words/s
2015-03-11 19:22:10,321: INFO: PROGRESS: at 98.50% words, alpha 0.00044, 85971 words/s
2015-03-11 19:22:11,934: INFO: PROGRESS: at 98.55% words, alpha 0.00039, 85940 words/s
2015-03-11 19:22:13,384: INFO: PROGRESS: at 98.65% words, alpha 0.00036, 85960 words/s
2015-03-11 19:22:13,883: INFO: training on 152625573 words took 1775.1s, 85982 words/s
2015-03-11 19:22:13,883: INFO: saving Word2Vec object under wiki.zh.text.model, separately None
2015-03-11 19:22:13,884: INFO: not storing attribute syn0norm
2015-03-11 19:22:13,884: INFO: storing numpy array 'syn0' to wiki.zh.text.model.syn0.npy
2015-03-11 19:22:20,797: INFO: storing numpy array 'syn1' to wiki.zh.text.model.syn1.npy
2015-03-11 19:22:40,667: INFO: storing 278291x400 projection weights into wiki.zh.text.vector

让我们看一下训练好的中文维基百科word2vec模型“wiki.zh.text.vector"的效果：

In [1]: import gensim
 
In [2]: model = gensim.models.Word2Vec.load("wiki.zh.text.model")
 
In [3]: model.most_similar(u"足球")
Out[3]: 
[(u'\u8054\u8d5b', 0.6553816199302673),
 (u'\u7532\u7ea7', 0.6530429720878601),
 (u'\u7bee\u7403', 0.5967546701431274),
 (u'\u4ff1\u4e50\u90e8', 0.5872289538383484),
 (u'\u4e59\u7ea7', 0.5840631723403931),
 (u'\u8db3\u7403\u961f', 0.5560152530670166),
 (u'\u4e9a\u8db3\u8054', 0.5308005809783936),
 (u'allsvenskan', 0.5249762535095215),
 (u'\u4ee3\u8868\u961f', 0.5214947462081909),
 (u'\u7532\u7ec4', 0.5177896022796631)]
 
In [4]: result = model.most_similar(u"足球")
 
In [5]: for e in result:
    print e[0], e[1]
   ....:     
联赛 0.65538161993
甲级 0.653042972088
篮球 0.596754670143
俱乐部 0.587228953838
乙级 0.58406317234
足球队 0.556015253067
亚足联 0.530800580978
allsvenskan 0.52497625351
代表队 0.521494746208
甲组 0.51778960228
 
In [6]: result = model.most_similar(u"男人")
 
In [7]: for e in result:
    print e[0], e[1]
   ....:     
女人 0.77537125349
家伙 0.617369174957
妈妈 0.567102909088
漂亮 0.560832381248
잘했어 0.540875017643
谎言 0.538448691368
爸爸 0.53660941124
傻瓜 0.535608053207
예쁘다 0.535151124001
mc刘 0.529670000076
 
In [8]: result = model.most_similar(u"女人")
 
In [9]: for e in result:
    print e[0], e[1]
   ....:     
男人 0.77537125349
我的某 0.589010596275
妈妈 0.576344847679
잘했어 0.562340974808
美丽 0.555426716805
爸爸 0.543958246708
新娘 0.543640494347
谎言 0.540272831917
妞儿 0.531066179276
老婆 0.528521537781
 
In [10]: result = model.most_similar(u"青蛙")
 
In [11]: for e in result:
    print e[0], e[1]
   ....:     
老鼠 0.559612870216
乌龟 0.489831030369
蜥蜴 0.478990525007
猫 0.46728849411
鳄鱼 0.461885392666
蟾蜍 0.448014199734
猴子 0.436584025621
白雪公主 0.434905380011
蚯蚓 0.433413207531
螃蟹 0.4314712286
 
In [12]: result = model.most_similar(u"姨夫")
 
In [13]: for e in result:
    print e[0], e[1]
   ....:     
堂伯 0.583935439587
祖父 0.574735701084
妃所生 0.569327116013
内弟 0.562012672424
早卒 0.558042645454
曕 0.553856015205
胤祯 0.553288519382
陈潜 0.550716996193
愔之 0.550510883331
叔父 0.550032019615
 
In [14]: result = model.most_similar(u"衣服")
 
In [15]: for e in result:
    print e[0], e[1]
   ....:     
鞋子 0.686688780785
穿着 0.672499775887
衣物 0.67173999548
大衣 0.667605519295
裤子 0.662670075893
内裤 0.662210345268
裙子 0.659705817699
西装 0.648508131504
洋装 0.647238850594
围裙 0.642895817757
 
In [16]: result = model.most_similar(u"公安局")
 
In [17]: for e in result:
    print e[0], e[1]
   ....:     
司法局 0.730189085007
公安厅 0.634275555611
公安 0.612798035145
房管局 0.597343325615
商业局 0.597183346748
军管会 0.59476184845
体育局 0.59283208847
财政局 0.588721752167
戒毒所 0.575558543205
新闻办 0.573395550251
 
In [18]: result = model.most_similar(u"铁道部")
 
In [19]: for e in result:
    print e[0], e[1]
   ....:     
盛光祖 0.565509021282
交通部 0.548688530922
批复 0.546967327595
刘志军 0.541010737419
立项 0.517836689949
报送 0.510296344757
计委 0.508456230164
水利部 0.503531932831
国务院 0.503227233887
经贸委 0.50156635046
 
In [20]: result = model.most_similar(u"清华大学")
 
In [21]: for e in result:
    print e[0], e[1]
   ....:     
北京大学 0.763922810555
化学系 0.724210739136
物理系 0.694550514221
数学系 0.684280991554
中山大学 0.677202701569
复旦 0.657914161682
师范大学 0.656435549259
哲学系 0.654701948166
生物系 0.654403865337
中文系 0.653147578239
 
In [22]: result = model.most_similar(u"卫视")
 
In [23]: for e in result:
    print e[0], e[1]
   ....:     
湖南 0.676812887192
中文台 0.626506924629
収蔵 0.621356606483
黄金档 0.582251906395
cctv 0.536769032478
安徽 0.536752820015
非同凡响 0.534517168999
唱响 0.533438682556
最强音 0.532605051994
金鹰 0.531676828861
 
In [24]: result = model.most_similar(u"习近平")
 
In [25]: for e in result:
    print e[0], e[1]
   ....:     
胡锦涛 0.809472680092
江泽民 0.754633367062
李克强 0.739740967751
贾庆林 0.737033963203
曾庆红 0.732847094536
吴邦国 0.726941585541
总书记 0.719057679176
李瑞环 0.716384887695
温家宝 0.711952567101
王岐山 0.703570842743
 
In [26]: result = model.most_similar(u"林丹")
 
In [27]: for e in result:
    print e[0], e[1]
   ....:     
黄综翰 0.538035452366
蒋燕皎 0.52646958828
刘鑫 0.522252976894
韩晶娜 0.516120731831
王晓理 0.512289524078
王适 0.508560419083
杨影 0.508159279823
陈跃 0.507353425026
龚智超 0.503159761429
李敬元 0.50262516737
 
In [28]: result = model.most_similar(u"语言学")
 
In [29]: for e in result:
    print e[0], e[1]
   ....:     
社会学 0.632598280907
人类学 0.623406708241
历史学 0.618442356586
比较文学 0.604823827744
心理学 0.600066184998
人文科学 0.577783346176
社会心理学 0.575571238995
政治学 0.574541330338
地理学 0.573896467686
哲学 0.573873817921
 
In [30]: result = model.most_similar(u"计算机")
 
In [31]: for e in result:
    print e[0], e[1]
   ....:     
自动化 0.674171924591
应用 0.614087462425
自动化系 0.611132860184
材料科学 0.607891201973
集成电路 0.600370049477
技术 0.597518980503
电子学 0.591316461563
建模 0.577238917351
工程学 0.572855889797
微电子 0.570086717606
 
In [32]: model.similarity(u"计算机", u"自动化")
Out[32]: 0.67417196002404789
 
In [33]: model.similarity(u"女人", u"男人")
Out[33]: 0.77537125129824813
 
In [34]: model.doesnt_match(u"早餐 晚餐 午餐 中心".split())
Out[34]: u'\u4e2d\u5fc3'
 
In [35]: print model.doesnt_match(u"早餐 晚餐 午餐 中心".split())
中心

有好的也有坏的case，甚至bad case可能会更多一些，这和语料库的规模有关，还和分词器的效果有关等等，不过这个实验暂且就到这里了。至于word2vec有什么用，目前除了用来来计算词语相似度外，业界更关注的是word2vec在具体的应用任务中的效果，这个才是更有意思的东东，也欢迎大家一起探讨。

注：原创文章，转载请注明出处“我爱自然语言处理”：www.52nlp.cn

本文链接地址：https://www.52nlp.cn/中英文维基百科语料上的word2vec实验

相关文章：

  1. 微软：Web N-gram Services
  2. 维基百科语料中的词语相似度探索
  3. 玩转腾讯词向量：词语相似度计算和在线查询
  4. 腾讯词向量实战：通过Annoy进行索引和快速查询

作者 52nlp

相关文章

自然语言处理

新浪张俊林：大语言模型的涌现能力——现象与解释

4 月 28, 2023 达观数据
自然语言处理

中科院张家俊：ChatGPT中的提示与指令学习

4 月 11, 2023 达观数据
自然语言处理

“国产类 ChatGPT ”所存在的差距与挑战-专家圆桌

3 月 27, 2023 达观数据
《中英文维基百科语料上的Word2Vec实验》有370条评论

  2. 请问大神在实现word2vec算法生成词向量时，gensim的效果是不是优于tensorflow？

    [回复]

    52nlp 回复:
    25 11 月, 2018 at 19:07

    tensorflow的实现我不清楚；不过gensim的训练速度貌似优于原生c版本

    [回复]

    波特 回复:
    25 11 月, 2018 at 19:26

    谢谢回复，我是百度看到有人做实验测试gensim训练的词向量优于tensorflow,但作者说未查找原因。谢谢您呀！

    [回复]

  3. 对作者的那个bug比较感兴趣，不知道找到原因了没有。我觉得是内存问题。
    估算作者所用“macpro（4核16G机器）”的运算能力在204GFLOPS：
    英文维基百科语料12G（单词数1994415728），词向量训练参数size=400, min_count=5, workers=4, window=5, epoch=1，耗时20916.4s/epoch，处理均速度95350 words/s，总耗时20916.4s，设备macpro 4core 3.2GHz 16 DP FLOPs/cycle，算力204GFLOPS），训练得到模型文件7G（估算为7.3G），预计服务运行时占用内存14.6-18.25G，实际服务运行时占用内存未知（基于numpy格式出现了nan的bug，应该是这种模式下内存不足导致加载时被压缩或覆盖而出现的查询问题）。

    [回复]

    52nlp 回复:
    3 12 月, 2018 at 17:31

    之前没去细究，还有更大的机器可以跑，这个分析有意思，感谢

    [回复]

    娄泽华 回复:
    3 12 月, 2018 at 18:09

    嗯嗯，我是在运行时看到有压缩信息并且加载时间异常长的时候才发现内存确实不够的。现在的gensim的信息会估算一下内存，会好一些，不至于加载半天才发现不能用。
    [2018-12-03 13:59:39 - INFO - word2vec.py:line 1303]: downsampling leaves estimated 44375567 word corpus (96.5% of prior 45962262)
    [2018-12-03 13:59:40 - INFO - base_any2vec.py:line 554]: estimated required memory for 389823 words and 256 dimensions: 993269004 bytes
    [2018-12-03 13:59:40 - INFO - word2vec.py:line 1414]: resetting layer weights
    [2018-12-03 13:59:46 - INFO - base_any2vec.py:line 624]: training model with 4 workers on 389823 vocabulary and 256 features, using sg=0 hs=0 sample=0.001 negative=5 window=5

    [回复]

    michael 回复:
    31 1 月, 2019 at 05:55

    你的这个分析很厉害 看了一遍评论发现全是伸手党 我过一段就能组好32GB的机子 我到时候自己跑一下看看结果

    [回复]

  4. 作者你好，我发现在最后相似度计算中，大部分单词的效果都非常好。但是有一个基础的词汇“男人”，效果反而不太好。1）相似单词有两个韩文词汇，2）相似单词有一个“mc刘”。造成这个的原因是什么呢？
    In [6]: result = model.most_similar(u"男人")
    In [7]: for e in result:
    print e[0], e[1]
    ....:
    女人 0.77537125349
    家伙 0.617369174957
    妈妈 0.567102909088
    漂亮 0.560832381248
    잘했어 0.540875017643
    谎言 0.538448691368
    爸爸 0.53660941124
    傻瓜 0.535608053207
    예쁘다 0.535151124001
    mc刘 0.529670000076

    [回复]

    52nlp 回复:
    25 9 月, 2019 at 14:50

    数据包含噪音，预处理的时候没有清洗干净

    [回复]

  5. 我想问一下，我训练时训练速度平均900words/s，有什么办法可以解决么

    [回复]

    52nlp 回复:
    7 3 月, 2020 at 10:44

    windows下可能会存在这个问题，另外我印象需要cython加速

    [回复]

  6. 对于词组，你们一般会怎么计算相似度呢？比如“勇闯天涯”，“勇闯”、“天涯”在字典里，分别能找到对应的相似词们，但是对于词组的话，不在词库里，是对两个向量平均，然后怎么处理还是？

    [回复]

    52nlp 回复:
    7 6 月, 2020 at 22:56

    建议把这种词组或者短语作为切词自定义词典一起放进去训练，这样效果会好一些

    [回复]

  7. 这个制作诗词软件，好像，不能用了，之前一直好好的，突然的 给它发字词，它不回复了

    [回复]

    52nlp 回复:
    8 8 月, 2020 at 16:16

    你的微信号是哪个？

    [回复]

  8. Windows下运行失败呢？是不是因为压缩包的缩略名？想问下作者大大是不是改成.zip就可以用了

    [回复]

    52nlp 回复:
    14 11 月, 2020 at 21:44

    不清楚

    [回复]

  9. 请问在转化成英文wiki.text的时候，如果我希望保留标点符号，需要在process_wiki.py中修改哪里呢？

    [回复]

    52nlp 回复:
    20 5 月, 2021 at 17:33

    这个你要看一下gensim里 WikiCorpus 的设置，不过时间太久了，我有点忘了，你自己看一下，大致是做了标点符号的正则过滤，看看有没有开关

    [回复]

发表回复

您的电子邮箱地址不会被公开。 必填项已用 * 标注

You missed

Coursera

探索3D打印与增材制造课程：转变创意为现实的艺术之旅

2024年10月16号 课程图谱
Coursera

课程推荐：会计数据分析课程回顾

2024年10月16号 课程图谱
Coursera

如何利用Coursera的《大学成功学术技能》课程提升您的学习能力

2024年10月16号 课程图谱
Coursera

提升写作技能的课程推荐：学术英语写作培训

2024年10月16号 课程图谱