license: apache-2.0
language:
- en
- de
- ru
- zh
tags:
- mt-evaluation
- WMT
- MQM
size_categories:
- 100K<n<1M
Dataset Summary
This dataset contains all MQM human annotations from previous WMT Metrics shared tasks and the MQM annotations from Experts, Errors, and Context in a form of error spans. Moreover, it contains some hallucinations used in the training of XCOMET models.
Please note that this is not an official release of the data and the original data can be found here.
The data is organised into 8 columns:
- src: input text
- mt: translation
- ref: reference translation
- annotations: List of error spans (dictionaries with 'start', 'end', 'severity', 'text')
- lp: language pair
While en-ru
was annotated by Unbabel, en-de
and zh-en
was annotated by Google. This means that for en-de and zh-en you will only find minor and major errors while for en-ru you can find a few critical errors.
Python usage:
from datasets import load_dataset
dataset = load_dataset("RicardoRei/wmt-mqm-error-spans", split="train")
There is no standard train/test split for this dataset but you can easily split it according to year, language pair or domain. E.g. :
# split by LP
data = dataset.filter(lambda example: example["lp"] == "en-de")
Citation Information
If you use this data please cite the following works:
- Experts, Errors, and Context: A Large-Scale Study of Human Evaluation for Machine Translation
- Results of the WMT21 Metrics Shared Task: Evaluating Metrics with Expert-based Human Evaluations on TED and News Domain
- Results of WMT22 Metrics Shared Task: Stop Using BLEU – Neural Metrics Are Better and More Robust
- xCOMET: Transparent Machine Translation Evaluation through Fine-grained Error Detection