may-ohta/iwslt14_prompt

JoeyNMT: iwslt14 de-en-fr multilingual

This is a JoeyNMT model for multilingual MT with language tags, built for a demo purpose. The model is trained on iwslt14 de-en / en-fr parallel data using DDP.

Install JoeyNMT v2.3:

$ pip install git+https://github.com/joeynmt/joeynmt.git

Translation

Torch hub interface:

import torch

iwslt14 = torch.hub.load("joeynmt/joeynmt", "iwslt14_prompt")
translation = iwslt14.translate(
    src=["Hello world!"],  # src sentence
    src_prompt=["<en>"],   # src language code
    trg_prompt=["<de>"],   # trg language code
    beam_size=1,
)
print(translation)  # ["Hallo Welt!"]

(See jupyter notebook for details)

Training

$ python -m joeynmt train iwslt14_prompt/config.yaml --use-ddp --skip-test

(See train.log for details)

Evaluation

$ git clone https://huggingface.co/may-ohta/iwslt14_prompt
$ python -m joeynmt test iwslt14_prompt/config.yaml --output-path iwslt14_prompt/hyp

direction	bleu
en->de	28.88
de->en	35.28
en->fr	38.86
fr->en	40.35

• beam_size: 5
• beam_alpha: 1.0
• sacrebleu signature nrefs:1|case:lc|eff:no|tok:13a|smooth:exp|version:2.4.0

(See test.log for details)

Data Format

We downloaded IWSLT14 de-en and en-fr from https://wit3.fbk.eu/2014-01 and created {train|dev|test}.tsv files in the following format:

src_prompt	src	trg_prompt	trg
<en>	Hello.	<de>	Hallo.
<de>	Vielen Dank!	<en>	Thank you!

(See test.ref.de-en.tsv)