everpeace/word2vec-jawiki repository overview

⁠word2vec-jawiki

Tool to build word embeddings with word2vec from japanese wikipedia dump data. For more correct tokenization, mecab-ipdic-neologd⁠ dictionary is used.

⁠How to use

This ships docker image everpeace/word2vec-jawiki⁠ . So, you can do it by hitting simply:

docker run -v <output_directory>:/var/jawiki everpeace/word2vec-jawiki

CAUTION: This takes really long time (maybe 10~12 hours)

This will perform:

• download japanese wikipedia dump data (xml format)
  • latest data will be downloaded in default
• convert downloaded data to plain text
• prepare tokenized plain text by mecab and mecab-ipadic-neologd
• train word embedding with word2vec
  • skip-gram model will be used in default

If all went good, you can see several files generated in your <output_directory>

• vector_jawiki.bin: vector representations for each word in binary format
• vector_jawiki.txt: vector representations for each word in text format
• vector_jawiki.meta: the file contains options used in word2vec and mecab
• vector_jawiki.tgz: the tarball containing above three files

⁠How to configure

All options should be specified via environment variables:

• OUTPUT_DIR: output directory
• JAWIKI_URL: url to download japanese wikipedia dump data
• JAWIKI_FILENAME: local filename for japanese wikipedia dump data
• WP2TXT_OPTIONS: options passed to wp2txt
• MECAB_REPROCESS: switch to reprocess mecab(tokenization) (1: yes, 0: no)
• MECAB_OPTIONS: options passed to mecab
• WORD2VEC_OPTIONS: options passed to word2vec

See Dockerfile⁠ for default values.

⁠Tips

everpeace:word2vec-jawiki docker image contains word2vec tools. So you can try distance or worde-analyze on your word embedding like below once you build them.

$ docker run -it -v <output_directory>:/var/jawiki everpeace/word2vec-jawiki word-analogy /var/jawiki/vector_jawiki.bin

$ docker run -it -v <output_directory>:/var/jawiki everpeace/word2vec-jawiki distance /var/jawiki/vector_jawiki.bin