This is an open source implementation of the Bilingual Distributed Representations without Word Alignments method described in the paper cited below.

http://arxiv.org/abs/1410.2455

BilBOWA: Fast Bilingual Distributed Representations without Word Alignments

Stephan Gouws, Yoshua Bengio, Greg Corrado

We introduce BilBOWA ("Bilingual Bag-of-Words without Alignments"), a simple and computationally-efficient model for learning bilingual distributed representations of words which can scale to large datasets and does not require word-aligned training data. Instead it trains directly on monolingual data and extracts a bilingual signal from a smaller set of raw text sentence-aligned data. This is achieved using a novel sampled bag-of-words cross-lingual objective, which is used to regularize two noise-contrastive language models for efficient cross-lingual feature learning. We show that bilingual embeddings learned using the proposed model outperforms state-of-the-art methods on a cross-lingual document classification task as well as a lexical translation task on the WMT11 data. Our code will be made available as part of an open-source toolkit.

arXiv:1410.2455 Submitted on 9 Oct 2014

Compiling and using Bilbowa

To compile, do the following

cd bilbowa
make all

This creates bilbowa and bidist in the bin directory. Run bilbowa to see the list of command line options. bidist is a modification of the distance module in the word2vec package. It receives 2 binary vectors (let's say English and French) as input, and for any given English word returns the list of the similar French words. Note that similar to the original distance module, it only works with the binary vectors (use -binary 1 switch when calling bilbowa to train the system.)

Generation result

$ ./bilbowa -mono-train1 endata.txt -mono-train2 esdata.txt -par-train1 enes.en -par-train2 enes.es -output1 envec.txt -output2 esvec.txt -size 300 -window 5 -sample 1e-4 -negative 5 -binary 0 -adagrad 1 -xling-lambda 1 -epochs 10 -threads 3 -negative 5 -save-vocab1 en.vocab -save-vocab2 es.vocab -min-count 1
Learning Vocab
pre SortVocab
Vocab size: 3378
Words in train file: 1174846
Done learning vocab
Saving vocab
Saving vocabulary with 3378 entries to en.vocab
Initializing net....done.
Initializing unigram table....done.
Learning Vocab
pre SortVocab
Vocab size: 2985
Words in train file: 1086334
Done learning vocab
Saving vocab
Saving vocabulary with 2985 entries to es.vocab
Initializing net....done.
Initializing unigram table....done.
Starting training.
Alpha: 0.025000  Progress: 99.12%  (epoch 9) Updates (L1: 11.73M, L2: 11.75M) L1L2grad: 0.1533 Words/sec: 2.99K  
Saving model to file: envec.txt

Saving model to file: esvec.txt
TrainModel() took 2136.58 seconds to execute