fairseq-laser

Implementing LASER architecture using fairseq library based on my understanding of the original papers by Artexte and Schwenk (2018, [1] and [2]).

Requirements

• Python version >= 3.6
• PyTorch (tested on version 1.3.1)
• fairseq (tested on version 0.9.0)
• Faiss, for bitext mining

Training

This example shows how to train a LASER model on 4 languages from Europarl v7 (English/French/Spanish/German) with a similar architecture in [1].

# Download and preprocess the data
bash prepare-europarl.sh

# Binarize datasets for each language pair
bpe=europarl_en_de_es_fr/bpe.40k
data_bin=data-bin/europarl.de_en_es_fr.bpe40k
for lang_pair in de-en de-es de-fr en-es en-fr es-fr; do
    src=`echo $lang_pair | cut -d'-' -f1`
    tgt=`echo $lang_pair | cut -d'-' -f2`
    rm $data_bin/dict.$src.txt $data_bin/dict.$tgt.txt
    fairseq-preprocess --source-lang $src --target-lang $tgt \
        --trainpref $bpe/train.$src-$tgt \
        --joined-dictionary --tgtdict $bpe/vocab \
        --destdir $data_bin \
        --workers 20
done

# Train a LASER model. To speed up, we only use 2 target languages
# (English and Spanish) and train for 10 epochs.
checkpoint=checkpoints/laser_lstm
mkdir -p $checkpoint
fairseq-train $data_bin \
  --max-epoch 10 \
  --ddp-backend=no_c10d \
  --task translation_laser --arch laser \
  --lang-pairs de-en,de-es,en-es,es-en,fr-en,fr-es \
  --optimizer adam --adam-betas '(0.9, 0.98)' \
  --lr 0.001 --criterion cross_entropy \
  --dropout 0.1 --save-dir $checkpoint \
  --max-tokens 12000 --fp16 \
  --valid-subset train --disable-validation \
  --no-progress-bar --log-interval 1000 \
  --user-dir laser/

Bitext mining

Here are some results on running the above model on BUCC 2018 shared task data (see bucc.sh). The scores are on training set since the gold standard for the test set is not released (refer to Table 2 of [1] for comparison to a similar model).

Languages	Threshold	Precision	Recall	F1 score
fr-en	1.102786	91.63	91.37	91.50
de-en	1.095823	95.12	94.57	94.84

References

[1] Mikel Artetxe and Holger Schwenk, Margin-based Parallel Corpus Mining with Multilingual Sentence Embeddings arXiv, Nov 3 2018.

[2] Mikel Artetxe and Holger Schwenk, Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond arXiv, 26 Dec 2018.