RAMEN

A software for transferring pretrained English models to foreign languages. For technical details is provided in the paper From English To Foreign Languages: Transferring Pre-trained Language Models

If you use this code for your research, please cite

@article{trnke2020_ramen,
       author = {{Tran}, Ke},
        title = "{From English To Foreign Languages: Transferring Pre-trained Language Models}",
      journal = {arXiv e-prints},
         year = 2020,
        month = feb,
          eid = {arXiv:2002.07306},
        pages = {arXiv:2002.07306},
archivePrefix = {arXiv},
       eprint = {2002.07306},
 primaryClass = {cs.CL},
}

Requirements

• transformers with can be installed via pip install transformers
• tokenizes via pip install tokenizers

Data Processing

Data pre-processing consists of several steps.

Tokenization

• learning BPE code for target languages
• I used fastBPE in the experiments, but any tokenizers can be used. In demo.sh HuggingFace's tokenizers is used.

$ python src/misc/tokenizer.py --input data/mono/vi.2M --model bert --vocab_size 32000

Tensorize Data

$ python utils/numerize.py --vocab /path/to/vocab --input vi.train --bin_path vi.train.pth

Estimating sub-word translation probabilities

With Parallel data

See demo.sh script

With fastText vectors

See alignment/example_mono.sh

Training RAMEN

See bert_base.sh script

Evaluation

Evaluation using XNLI task can be found in run_xnli.sh script

License

This library is licensed under the MIT-0 License. See the LICENSE file.