@inproceedings{63493946c01a42daa37847b42252dc1f,
title = "An exploratory study on multilingual quality estimation",
author = "Shuo Sun and Marina Fomicheva and Frederic Blain and Vishrav Chaudhary and Ahmed El-Kishky and Adithya Renduchintala and Francisco Guzman and Lucia Specia",
note = "Predicting the quality of machine translationhas traditionally been addressed withlanguage-specific models, under the assumptionthat the quality label distribution or linguisticfeatures exhibit traits that are notshared across languages. An obvious disadvantageof this approach is the need for labelleddata for each given language pair. We challengethis assumption by exploring differentapproaches to multilingual Quality Estimation(QE), including using scores from translationmodels. We show that these outperform singlelanguagemodels, particularly in less balancedquality label distributions and low-resourcesettings. In the extreme case of zero-shot QE,we show that it is possible to accurately predictquality for any given new language from modelstrained on other languages. Our findingsindicate that state-of-the-art neural QE modelsbased on powerful pre-trained representationsgeneralise well across languages, making themmore applicable in real-world settings.",
year = "2020",
language = "English",
booktitle = "Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing",
publisher = "Association for Computational Linguistics",
}