candlewill/Chinsese_word_vectors

Chinese word vectors

This project uses Word2vec and GloVe tools to train word vectors for Chinese using data from wikipedia dump.

Steps

1. Download wikipedia dump from: https://dumps.wikimedia.org/zhwiki/latest/zhwiki-latest-pages-articles.xml.bz2

2. Extract Chinese words from the downloaded xml files, using process_wiki.py. This step would take about 11 minutes(from 14:54:17 to 15:05:53) in my computer to finish iterating over Wikipedia corpus of 253646 documents (articles) with 56758668 positions (total 2724783 articles, 68096009 positions before pruning articles shorter than 50 words).

3. We want to get both Traditional version and Simplified version texts. Firstly, Traditional Chinese should be translated to Simplified Chinese, we use opencc package.

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

   Then, simplified Chinese should be translated to traditional Chinese:

   opencc -i wiki.zh.text -o wiki.zh.text.traditional -c zhs2zht.ini
   

4. Tokenization, split the sentence into words. Two tokennization mechanism are used:

   • one character as a words all the time, 因为汉语中单个字往往也能表示许多信息. Namely, we segment the Chinese sentence into character level as follows:

   欧几里得 西元前三世纪的希腊数学家 现在被认为是几何之父
   欧 几 里 得 西 元 前 三 世 纪 的 希 腊 数 学 家 现 在 被 认 为 是 几 何 之 父
   

   • using Jieba packages to separate sentence. Run tokenization.py，考虑到常用分词工具来分词，使用词汇可提供更多信息

5. Train word vectors using train_word2vec_model.py and GloVe model.

   Hyper-parameters:

   • word vector dimension: 300
   • window: 5
   • min_count: 5
   • epochs: 3
   • Training algorithm: skip-gram (sg = 1)
   • Model training: hierarchical sampling (hs = 1)

   Word vectors statistical information:

   word vectors	corpus	words	sentences	dimensions	Time
   word2vec	traditional Chinese	860,835	253,646	300	3,086.8s
   word2vec	simplified Chinese	686,601	253,646	300	3,186.9s
   GloVe	traditional Chinese	860,835	253,646	300	2,160.0s
   GloVe	simplified Chinese	686,601	253,646	300	2,280.0s

   Character vectors ststistical information: (Hyper-parameter is the same as training word-level vectors)

   Character vectors	corpus	character	dimensions	Time
   GloVe	traditional	157,660	300	420s
   GloVe	simplified	157,379	300	410s
   word2vec	traditional	157,660	300	3,991.1s
   word2vec	simplified	157,379	300	4,415.6s

Download

1. word level and character level vectors trained by word2vec:

   Level	Language	syn0	syn1	Text
   word	simplified	Download	Download	Download
   word	traditional	Download	Download	Download
   character	simplified	Download	Download	Download
   character	traditional	Download	Download	Download

   Note: syn0 and syn1 files are produced by saving model using model.save(), which allows to continue training the model when loaded. The Text file is produced by saving model using model.save_word2vec_format(), which allows to view the vectors by a sublime-like software.

2. word level and character level vectors trained by GloVe:

   Level	Language	vocab.txt	vectors.txt	cooccurrence.bin
   word	simplified	Download	Download	Download
   word	traditional	Download	Download	Download
   character	simplified	Download	Download	Download
   character	traditional	Download	Download	Download

   Note: vectors.txt is the vectors produced by GloVe model and vocab.txt is the vocabulary-index mapping file.

3. Original Wikipeida Chinese Corpus: Download

4. The Simplified Chinese version Wikipedia corpus before segmentation: Download, After Segmentation: Download, After character-level segmentation: Download.

5. The Traditional Chinese version Wikipedia corpus before segmentation: Download, After segmentation: Download, After character-level segmentation: Download.

Contact

• [Yunchao He] (https://plus.google.com/+YunchaoHe)
• [email protected]
• YZU at Taiwan
• Weibo
• Facebook
• Twitter