This article is the "word vector 'with Word2Vec training Chinese word vector (a) - the use of search dogs news data set based on the two corpora merge, and then trained better word vector model.

Download dataset

Click here to download: Wiki Corpus selected date.

The file is downloaded zhwiki-latest-pages-articles.xml.bz2, size 1.75GB (especially slow download speed).
Here Insert Picture Description

Data Processing Set

(A) text extraction

There are two ways to extract text from a compressed file:

Use gensim.corpora of WikiCorpus directly to Wikipedia corpus processing. Save to wikiCorpus / wikiCorpus.txt

from gensim.corpora import WikiCorpus

if __name__ == '__main__':

    print('主程序开始...')

    input_file_name = 'zhwiki-latest-pages-articles.xml.bz2'
    output_file_name = 'wikiCorpus/wikiCorpus.txt'
    print('开始读入wiki数据...')
    input_file = WikiCorpus(input_file_name, lemmatize=False, dictionary={})
    print('wiki数据读入完成！')

    print('处理程序开始...')
    count = 0
    with open(output_file_name, 'wb') as output_file:
        for text in input_file.get_texts():
            output_file.write(' '.join(text).encode("utf-8"))
            output_file.write('\n'.encode("utf-8"))
            count = count + 1
            if count % 10000 == 0:
                print('目前已处理%d条数据' % count)
    print('处理程序结束！')

    output_file.close()
    print('主程序结束！')

Download Wikipedia Extractor extracts text.

Processing the data set is completed, a total of 35w + pieces of data, the following effects. We found that punctuation has been removed, simply do not like too! ! !

(B) simplify

Download opencc , perform character conversion.

The wikiCorpus.txt copied to the folder, the indicator used cd to the folder input opencc -i wikiCorpus.txt -o wikiCorpusSim.txt -c t2s.json
Here Insert Picture Description
to complete the conversion, the converted wikiCorpusSim.txt saved to Python Project , results as shown:

(C) the word document

If done Sogou data set before it, this step should be very familiar.

import jieba

filePath = 'wikiCorpus/wikiCorpusSim.txt'  # 简化后文本
fileSegWordDonePath = 'wikiCorpus/wikiCorpusSegDone.txt'  # 分词处理后文本

# 将每一行文本依次存放到一个列表
fileTrainRead = []
with open(filePath, encoding='utf-8') as fileTrainRaw:
    for line in fileTrainRaw:
        fileTrainRead.append(line)

# 用jieba进行分词
fileTrainSeg = []
file_userDict = 'dict.txt'  # 自定义的词典
jieba.load_userdict(file_userDict)
for i in range(len(fileTrainRead)):
    fileTrainSeg.append([' '.join(jieba.cut(fileTrainRead[i], cut_all=False))])
    if i % 100 == 0:  # 每处理100个就打印一次
        print(i)

# 处理后写入文件
with open(fileSegWordDonePath, 'wb') as fW:
    for i in range(len(fileTrainSeg)):
        fW.write(fileTrainSeg[i][0].encode('utf-8'))
        fW.write('\n'.encode("utf-8"))

Here Insert Picture Description

(D) combining data sets

The good search dog before processing the data sets and merge wiki, generate Fried Chicken invincible Chinese corpus.


filePath1 = 'wikiCorpus/wikiCorpusSegDone.txt'  # wiki分词后语料库
filePath2 = 'sougouCorpus/sougouCorpusSegDone.txt'  # 搜狗语料库
filePath3 = 'corpusFinal.txt'  # 最终语料库


fileFinal = []
countS = 0  # sougou计数
countW = 0  # wiki计数

# 打开搜狗语料库
with open(filePath2, encoding='utf-8') as ff1:
    print("---成功导入搜狗语料库---")
    for line in ff1:
        fileFinal.append(line)
        countS = countS + 1
        if countS % 10000 == 0:
            print("---------已导入%d篇搜狗文章---------" % countS)

# 打开维基语料库
with open(filePath1, encoding='utf-8') as ff2:
    print("---成功导入wiki语料库---")
    for line in ff2:
        fileFinal.append(line)
        countW = countW + 1
        if countW % 10000 == 0:
            print("---------已导入%d篇维基文章---------" % countW)

# 打开最终文档，逐行写入
with open(filePath3, 'wb') as ff:
    for i in range(len(fileFinal)):
        ff.write(fileFinal[i].encode('utf-8'))
        if i % 10000 == 0:
            print("---------已合并%d篇文章---------" % i)

print("---------一共读入%d行---------" % len(fileFinal))
print("---------完成合并---------")

A total of 212w + rows of data (which contribute 71w row + Sogou 141w line)
Here Insert Picture Description
size of 3.35GB

Trainer

Or the old code is used

import logging
import sys
import gensim.models as word2vec
from gensim.models.word2vec import LineSentence, logger
# import smart_open


def train_word2vec(dataset_path, out_vector):
    # 设置输出日志
    logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s', level=logging.INFO)
    logger.info("running %s" % ' '.join(sys.argv))
    # 把语料变成句子集合
    sentences = LineSentence(dataset_path)
    # sentences = LineSentence(smart_open.open(dataset_path, encoding='utf-8'))  # 或者用smart_open打开
    # 训练word2vec模型（size为向量维度，window为词向量上下文最大距离，min_count需要计算词向量的最小词频）
    model = word2vec.Word2Vec(sentences, size=128, sg=1, window=6, min_count=5, workers=4, iter=6)
    # (iter随机梯度下降法中迭代的最大次数，sg为1是Skip-Gram模型)
    # 保存word2vec模型
    model.save("word2vec.model")
    model.wv.save_word2vec_format(out_vector, binary=False)


# 加载模型
def load_word2vec_model(w2v_path):
    model = word2vec.Word2Vec.load(w2v_path)
    return model


# 计算词语最相似的词
def calculate_most_similar(self, word):
    similar_words = self.wv.most_similar(word)
    print(word)
    for term in similar_words:
        print(term[0], term[1])


# 计算两个词相似度
def calculate_words_similar(self, word1, word2):
    print(self.wv.similarity(word1, word2))


# 找出不合群的词
def find_word_dismatch(self, list):
    print(self.wv.doesnt_match(list))


if __name__ == '__main__':
    dataset_path = 'corpusFinal.txt'
    out_vector = 'corpusFinal.vector'
    train_word2vec(dataset_path, out_vector)  # 训练模型

    # model = load_word2vec_model("word2vec.model")  # 加载模型

    # calculate_most_similar(model, "病毒")  # 找相近词

    # calculate_words_similar(model, "法律", "制度")  # 两个词相似度

    # print(model.wv.__getitem__('男人'))  # 词向量

    # list = ["早饭", "吃饭", "恰饭", "嘻哈"]

    # find_word_dismatch(model, list)