Awesome

JWE

Source codes of our EMNLP2017 paper Joint Embeddings of Chinese Words, Characters, and Fine-grained Subcharacter Components

Preparation

You need to prepare a training corpus and the Chinese subcharacter radicals or components.

• Training corpus. Download Chinese Wikipedia Dump. Preprocess the corpus. First, use wiki extractor or python gensim.corpora to extract plain text from the xml file; Second, remove non Chinese characters by using regular expression; Third, use opencc to convert traditional Chinese to simplified Chinese; Finally, perform Chinese word segmentation by using THULAC package. We provide the corpus after preprocessing at the onlibe baidu box.

• Subcharacter radicals and components. Deploy the scrapy codes in JWE/ChineseCharCrawler on Scrapy Cloud, you can crawl the resource from HTTPCN. We provide a copy of the data in ./subcharacters for reserach convenience. The copyright and all rights therein of the subcharacter data are reserved by the website HTTPCN.

Model Training

• cd JWE/src, compile the code by make all.
• run ./jwe for parameters details.
• run ./run.sh to start the model training, you may modify the parameters in file run.sh.
• Input files format: Corpus wiki.txt contains segmented Chinese words with UTF-8 encoding; Subcharacters comp.txt contains a list of components which are seperated by blank spaces; char2comp.txt, each line consists of a Chinese character and its components in the following format:

侩 亻 人 云
侨 亻 乔
侧 亻 贝 刂
侦 亻 卜 贝

Model Evaluation

Two Chinese word similarity datasets 240.txt and 297.txt and one Chinese analogy dataset analogy.txt in JWE/evaluation folder are provided by (Chen et al., IJCAI, 2015).

cd JWE/src, then

• run python word_sim.py -s <similarity_file> -e <embed_file> for word similarity evaluation, where similarity_file is the word similarity file, e.g., 240.txt or 297.txt, embed_file is the trained word embedding file.
• run python word_analogy.py -a <analogy_file> -e <embed_file> or ./word_analogy <embed_file> <analogy_file> for word analogy evaluation.