This is a word2vec for Chinese douban movie reviews 在豆瓣电影影评上进行word2vec, 一个中文语料word2vec

• Install gcc by sudo apt-get install gcc before install Anaconda and gensim, otherwise gensim will run in slow verison(70x slower).
• Install Anaconda
• Install gensim

• Going to src/ cd src
• Execute process_corpus.py python -i process_corpus.py

  • input: review_douban_movie.tsv is a file in which a douban review lays on one line.

  • output: review_douban_movie.tsv.removeword in which a processed douban review lays on one line. Some dependencies should be installed for those process flow.

  • Tradition2Simple

    • Install opencc sudo apt-get install opencc

  • WordBeark

    • Install jieba pip install jieba

  • RemoveWord

    • TODO: reference chinese-char-rnn

  • Then we maybe cat some files together if we split corpus into several files, using cat file1 file2 > file3 or cat file2 >> file1

• Going to src/ cd src
• Execute train_word2vec.py python -i train_word2vec.py -size 300 -window 5 -sg 1 -min_count 100 -negative 10 -iter 25 -workers 12

• Going to src/ cd src
• Execute test_word2vec.py python -i test_word2vec.py ../data/xxx.model
• questions-words-Zh.txt is a Chinese evaluation file translated from English questions-words.txt using Microsoft Bing Translator API. Some examples:

THIS FUNCTION IS DEPRECATED. It is hard to get word2vec better by Finetune rather than retrain.

• python -i finetune.py -finetune ../model/review_douban_movie.tsv.removewordsize300window5sg1min_count100negative10iter25.model -train ../data/wv_train.data

• Get the vector of a word: model.wv["word"], i.e. model.wv["apple"]
• Get the index of a word: model.wv.vocab["word"].index, i.e. model.wv.vocab["apple"].index
• Get a word by its index: model.wv.index2word[index], i.e. model.wv.index2word[5585]
• Find the most similar word of a given word: model.wv.most_similar("word"), i.e. model.wv.most_similar("apple")
• Evaluate a model on analogies: model.wv.accuracy( 'questions-words.txt')

• 确保安装gcc，否则根据官网的说法word2vec会慢70倍。使用命令sudo apt-get install gcc
• 安装Anaconda
• 安装gensim

• 进入src文件夹
• 执行process_corpus.py, python -i process_corpus.py

  • 其中输入为review_douban_movie.tsv，每一行有一条豆瓣影评，一共65w行

  • 输出为review_douban_movie.tsv.removeword，每一行都被繁体转简体、分词、去除标点。使用之前需要安装依赖项。

  • 繁体转简体

    • 安装opencc sudo apt-get install opencc

  • 分词

    • 安装jieba pip install jieba

  • 去除标点

    • TODO: 可以参考chinese-char-rnn改进

  • 如果多个语料文件分开处理，最后我们可以使用Linux的cat将他们连在一起，如 cat file1 file2 > file3或cat file2 >> file1

• 进入src文件夹
• 执行train_word2vec.py python -i train_word2vec.py -size 300 -window 5 -sg 1 -min_count 100 -negative 10 -iter 25 -workers 12

• 进入src文件夹
• 执行test_word2vec.py python -i test_word2vec.py ../data/xxx.model
• 值得一提的是，在这一步中我贡献了中文word2vec领域第一份评估标准：data/questions-words-Zh.txt，该标准是英文版的analogies任务翻译为中文。是我在Microsoft期间使用Bing Translator API翻译，翻译质量上乘。一些例子：

THIS FUNCTION IS DEPRECATED. 使用Finetune并不会取得很好的效果，还不如retrain

• python -i finetune.py -finetune ../model/review_douban_movie.tsv.removewordsize300window5sg1min_count100negative10iter25.model -train ../data/wv_train.data

• 获取词向量: model.wv["word"], i.e. model.wv["刘德华"]
• 获取词索引index: model.wv.vocab["word"].index, i.e. model.wv.vocab["张艺谋"].index
• 通过index获取词: model.wv.index2word[index], i.e. model.wv.index2word[5585]
• 找一个词最相近的词: model.wv.most_similar("word"), i.e. model.wv.most_similar("周星驰")
• 四词类推评价: model.wv.accuracy( 'questions-words.txt')