This software can produce a bilingual lexicon from non-parallel data without any cross-lingual supervision. It does so by learning a transformation between source word embeddings and target ones by adversarial training. The technique is described in the following paper:

Meng Zhang, Yang Liu, Huanbo Luan, and Maosong Sun. Adversarial Training for Unsupervised Bilingual Lexicon Induction. In Proceedings of ACL, 2017.

This software has been tested in the following environment, but should work in a compatible one.

• 64-bit Linux
• Python 2.7 (for adversarial training code in the src folder)
  • Theano
  • Lasagne (bleeding-edge version needed as of April, 2017)
  • scikit-learn
• Python 3.4 (for translation code in the scripts folder)

1. Specify the variables in the config file. For example, if config contains the following lines:

config=zh-en
lang1=zh
lang2=en

then the data should be located in data/zh-en with file extensions zh and en.

2. Prepare the following data in the folder specified in Step 1:

• word2vec.zh/en: Word embeddings, which can be obtained by running word2vec on monolingual data.
• vocab-freq.zh/en: Space-separated word-frequency pairs.

We provide the data for the five language pairs considered in our paper.

3. Train and obtain the bilingual lexicon.

./run.sh

4. The following files will be generated in data/zh-en (the folder specified in config):

• transformed-1.zh: Transformed source embeddings that lie in the same semantic space as target embeddings.
• vocab.zh: Source language vocabulary.
• result.1: Translations of vocab.zh. For each source word, there will be at most 10 translations after the tab character, separated by space and sorted in decreasing order of the cosine similarity.

It is recommended that the bleeding-edge version of Lasagne works with the latest development version of Theano. It has been tested on Lasagne version 0.2.dev1 and Theano version 0.9.0dev4.dev-RELEASE, but NaN may appear on Theano version 0.8.2.