Skip to content


Repository files navigation

Retrofitting Cross-Lingual Word Embeddings to Dictionaries

This repository reproduces results from our ACL 2020 paper:

Mozhi Zhang*, Yoshinari Fujinuma*, Michael J. Paul, Jordan Boyd-Graber. Why Overfitting Isn't Always Bad: Retrofitting Cross-Lingual Word Embeddings to Dictionaries.

If you find this repository helpful, please cite:

    title = {Why Overfitting Isn't Always Bad: Retrofitting Cross-Lingual Word Embeddings to Dictionaries},
    author = {Zhang, Mozhi and Fujinuma, Yoshinari and Paul, Michael J. and  Boyd-Graber, Jordan},
    booktitle = {Proceedings of the Association for Computational Linguistics},
    doi = "10.18653/v1/2020.acl-main.201",
    year = {2020}

Pre-Train and Preprocess

The first step is to train cross-lingual word embeddings (CLWE) using a projection-based method. In the paper, we start with monolingual Wikipedia FastText embeddings. We normalize the monolingual embeddings with Iterative Normalization and align them with CCA, MUSE, or RCSLS. We preprocess them to lowercase all words and keep only the top 200K words:


Retrofit CLWE to a Dictionary

Once we have pre-trained CLWE, we can retrofit to a dictionary. First, we need to merge the trained CLWE into a single file with For example, if we have aligned 300 dimensional English embeddings at embed/vectors-en.txt and Chinese embeddings at embed/vectors-zh.txt, we merge them with:

python \
    embed/vectors-en.txt \
    embed/vectors-zh.txt \
    embed/vectors-en-zh.txt \
    en zh 300

We then retrofit it to a dictionary. Here we use the English-Chinese training dictionary from MUSE:

python \
    -i embed/vectors-en-zh.txt \
    -l dictionaries/en-zh.0-5000.txt \
    -o embed_retrofit \
    --src_lang en \
    --tgt_lang zh

The retrofitted vectors are at embed_retrofit/vectors-en.txt and embed_retrofit/vectors-zh.txt (specified by -o flag).

Generate Synthetic Dictionary from CLWE

We can use and to build a synthetic dictionary from CLWE. For example, if we want to build an English-Chinese synthetic dictionary, we first translate every English word to Chinese:

python \
    --src embed/vectors-en.txt \
    --tgt embed/vectors-zh.txt \
    --output en-zh.txt

We then translate every Chinese word to English:

python \
    --src embed/vectors-zh.txt \
    --tgt embed/vectors-en.txt \
    --output zh-en.txt

Finally, we build a high-confidence synthetic dictionary using only mutual translations:

python en-zh.txt zh-en.txt en-zh-merged.txt

We can combine the final synthetic dictionary en-zh-merged.txt with the training dictionary. Retrofitting to the combined dictionary often improves downstream task accuracy.

Evaluate on Document Classification

Download RCV2 and split the data with MLDoc. The script trains a CNN with CLWE features and reports test accuracy. For example, we can run a English-Chinese experiment:

python \
    --src embed/vectors-en.txt \
    --tgt embed/vectors-zh.txt \
    --train_lang en \
    --test_lang zh \
    --train mldoc/english.train.5000 \
    --dev mldoc/ \
    --test mldoc/chinese.test

Evaluate on Dependency Parsing

  1. Install AllenNLP with pip install -r requirements.txt.

  2. Download Universal Dependencies (UD) v2.4.

  3. Add a language prefix to all tokens in UD .conllu files with dependency_parse.scripts/ For example, we can add 'en' to all words with the following:

      python dependency_parse/scripts/ \
         --lang en \
         --f_in PATH_TO_UD/en_ewt-ud-train.conllu
  4. Create vocab.txt (and non_padded_namespaces.txt) by running dependency_parse/scripts/ The --emb is a merged embedding file from

     python dependency_parse/scripts/ \
         --emb PATH_TO_EMBEDDING \
         --vocab_path OUTPUT_PATH_OF_VOCAB
  5. Edit the path to vocab.txt file, the path to the pre-trained word vector, and UD training/dev/test path in dependency_parse/allen_configs/parse.jsonnet.

  6. Run sh dependency_parse/scripts/


Why Overfitting Isn't Always Bad: Retrofitting Cross-Lingual Word Embeddings to Dictionaries (ACL 2020)







No releases published


No packages published