Trying out Artetxe's UNdreaMT on English-Basque

The goal is to see whether Artetxe's method works on a low-resource pair such as English-Basque, in particular if we can do better than the results reported in "English-Basque Statistical and Neural Machine Translation" where they only trained with about 130,000 phrase pairs (small dataset).

• Basque monolingual embeddings taken from https://fasttext.cc/docs/en/pretrained-vectors.html
• English monolingual embeddings taken from https://fasttext.cc/docs/en/english-vectors.html

Results of vecmap on English-Basque on 30 pairs (test_dict.txt)

Method	Coverage	Accuracy - NN	Accuracy - CSLS
Unsupervised	96.67%	0%	0%
Seed = 27 pairs	96.67%	37.93%	31.03%
Seed = 101 pairs	96.67%	41.38%	37.93%

Requirements

• Python 3
• PyTorch (tested with v0.3)
• NLTK
• tqdm
• Also, need to do: import nltk nltk.download('punkt')

Order of pre-processing should be:

• bash prep_eu_wiki.sh
• bash prep_en_data.sh
• bash download_embeddings.sh
• python extract_vocab.py cc.eu.300.vec vocab.eu train_vocab.eu
• python extract_vocab.py wiki-news-300d-1M.vec vocab.en train_vocab.en

No need to do the following two steps as Artetxe's code uses "OOV" if token is not in supplied vocab

• python update_corpus.py sentences.en train_vocab.en train_sentences.en
• python update_corpus.py sentences.eu train_vocab.eu train_sentences.eu

Then to use vecmap we need to git clone it and then something like:

• cd vecmap
• nohup python3 map_embeddings.py --semi_supervised ../undreamt/train_dict.txt ../undreamt/train_vocab.en ../undreamt/train_vocab.eu ../undreamt/train_emb_mapped.en ../undreamt/train_emb_mapped.eu > out.log &

Evaluation data was taken from https://github.com/ijauregiCMCRC/english_basque_MT