Dataset Card for DiscEvalMT

Dataset Details

Contrastive test set for English-to-French MT evaluation covering 2 discourse phenomena: anaphora and lexical choice (coherence/cohesion).

Dataset Description

For machine translation to tackle discourse phenomena, models must have access to extra-sentential linguistic context. There has been recent interest in modelling context in neural machine translation, but models have been principally evaluated with standard automatic metrics, poorly adapted to evaluating discourse phenomena. The aim of our article "Evaluating Discourse Phenomena in Neural Machine Translation" was to provide an alternative form of evaluation, specifically targeting discourse phenomena and the need for context beyond the level of the sentence.

• Curated by: Rachel Bawden, Rico Sennrich, Alexandra Birch, Alexandra and Barry Haddow
• Language(s) (NLP): English-to-French
• License: CC-BY-SA-4.0

Dataset Sources [optional]

• Repository: https://github.com/rbawden/discourse-mt-test-sets/tree/master
• Paper [optional]: Bawden et al., 2018. Evaluating Discourse Phenomena in Neural Machine Translation

Uses

The dataset is a contrastive evaluation test used to test the use of previous linguistic context (one previous sentence) when translating from English to French. Each example block is composed of a previous source sentence and two possible translations, accompanied with accompanying context, which can be used to disambiguate the two possible translations. Each instance in the dataset has a correct translation and a contrastive one. MT models should be used to rank each translation given the context provided, and the model is evaluated on its accuracy of ranking the correct translations above the contrastive ones.

from datasets import load_dataset

# for the anaphora split
dataset = load_dataset("rbawden/DiscEvalMT", "anaphora")

# for the lexical choice split
dataset = load_dataset("rbawden/DiscEvalMT", "lexical_choice")

Out-of-Scope Use

The test set is designed for translation from English to French. The test set cannot be used in the opposite direction, as the ambiguity of the source sentence will not hold. The test set is used to be used as a complement to global metrics of evaluation quality and assumes that the model has been designed for translation.

Dataset Structure

Each dataset split (one for anaphora, one for lexical choice) has 200 instances, each of the following form (example shown for anaphora):

{
  "split": "anaphora",
  "ex_num": 1,
  "type": "m.pl",
  "context_src": "The buildings will be finished next week.",
  "current_src": "Soon they will be full of new residents.",
  "context_trg": "Les bâtiments seront terminés la semaine prochaine.",
  "current_trg": "Ils seront bientôt pleins de nouveaux résidents.",
  "contrastive_context_trg": "Les bâtiments seront terminés la semaine prochaine.",
  "contrastive_current_trg": "Elles seront bientôt pleines de nouveaux résidents.",
  "correct_or_semicorrect": "correct"
}

There will be 3 other example of the same ex_num for anaphora, corresponding to the same source sentence, but with different correct/contrastive translations. For lexical choice, there are two instances per example number.

Dataset Creation

Source Data

Handcrafted examples (no specific domain) designed to be spontaneous, natural and varied.

Data Collection and Processing

Examples were manually created by the first author (native British English speaker, bilingual French) and checked by French native speakers. Examples were inspired by similar examples found in subtitle data (in terms of syntactic structures, lexical choices, etc.) in order to encourage diversity and naturalness.

Personal and Sensitive Information

There is no personal or sensitive information in the dataset.

Citation

BibTeX:

@inproceedings{bawden-etal-2018-evaluating,
    title = "Evaluating Discourse Phenomena in Neural Machine Translation",
    author = "Bawden, Rachel and Sennrich, Rico and Birch, Alexandra and Haddow, Barry",
    booktitle = {{Proceedings of the 2018 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers)}},
    month = jun,
    year = "2018",
    address = "New Orleans, Louisiana",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/N18-1118",
    doi = "10.18653/v1/N18-1118",
    pages = "1304--1313"
}

APA:

Bawden, R., Sennrich, R., Birch, A., & Haddow, B. (2018). Evaluating Discourse Phenomena in Neural Machine Translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) (pp. 1304–1313). Association for Computational Linguistics.

Dataset Card Authors

Rachel Bawden: rachel.bawden@inria.fr