Analyzing Uncertainty in Neural Machine Translation
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
LICENSE Initial commit Jun 7, 2018
README.md Initial commit Jun 7, 2018
wmt14-en-de.extra_refs.gz Initial commit Jun 7, 2018
wmt14-en-fr.extra_refs.gz Initial commit Jun 7, 2018

README.md

Introduction

This repository contains additional reference translations for the WMT'14 En-De and En-Fr test sets (newstest2014) as described in the paper: Analyzing Uncertainty in Neural Machine Translation.

In particular we provide 10,000 additional reference translations written by professional translators:

  • 10 additional reference translations for each of 500 sentences of the WMT'14 En-De test set (5,000 total)
  • 10 additional reference translations for each of 500 sentences of the WMT'14 En-Fr test set (5,000 total)

Format

Each line in the data file contains a sentence and an identifier, where identifiers represent:

  • S-123 English source for sentence #123 of newstest2014
  • T-123: Original reference translation (German or French)
  • R1-123 through R10-123: Additional reference translations

Sentence numbers correspond to the index of the sentence in the full test set (see below).

Full test sets

The full test sets from which the sentences were sampled can be reconstructed with sacrebleu:

sacrebleu --test-set wmt14 --language-pair en-de --echo src
sacrebleu --test-set wmt14 --language-pair en-de --echo ref
sacrebleu --test-set wmt14/full --language-pair en-fr --echo src
sacrebleu --test-set wmt14/full --language-pair en-fr --echo ref

Citation

If you use the data in your paper, then please cite it as:

@inproceedings{ott2018analyzing,
  author    = {Ott, Myle and Auli, Michael and Grangier, David and Ranzato, Marc'Aurelio},
  title     = {Analyzing Uncertainty in Neural Machine Translation},
  booktitle = {International Conference on Machine Learning},
  year      = 2018,
}

License

The dataset is licenced under CC-by-NC, see the LICENSE file for details.