The Ubiqus English-Inuktitut System for WMT20

Proceedings of the 5th Conference on Machine Translation (WMT), pages 213–217 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 213 The Ubiqus English-Inuktitut System for WMT20 Franc¸ois Hernandez Ubiqus, Paris, France fhernandez@ubiqus.com Vincent Nguyen Ubiqus, Paris, France vnguyen@ubiqus.com Abstract This paper describes Ubiqus’ submission to the WMT20 English-Inuktitut shared news translation task. Our main system, and only submission, is based on a multilingual ap- proach, jointly training a Transformer model on several agglutinative languages. The English-Inuktitut translation task is challeng- ing at every step, from data selection, prepa- ration and tokenization to quality evaluation down the line. Difficulties emerge both be- cause of the peculiarities of the Inuktitut lan- guage as well as the low-resource context. 1 Introduction Ubiqus participated in the English to Inuktitut news translation task of WMT20. We performed a sin- gle submission, based on an unconstrained mul- tilingual setup. The approach consists of jointly training a traditional Transformer (Vaswani et al., 2017) model on several agglutinative languages in order to benefit from them for the low-resource English-Inuktitut task (Aharoni et al., 2019). Though the dataset provided for the task is siz- able, with more than a million segments, it’s quite narrow domain-wise, as it comes from proceedings of the Nunavut Hansard. The task being translation of news, it’s expected to be a much wider domain. For that purpose, we extended the task with datasets of other - linguistically near - languages, as well as in-house datasets introducing more diversity to the domain. All experiments were performed with the OpenNMT (Klein et al., 2017) toolkit, with Tokenizer 1 for data preprocessing and OpenNMT- py 2 for model training and inference. 1https://github.com/OpenNMT/Tokenizer 2https://github.com/OpenNMT/OpenNMT-py 2 Data 2.1 Training corpora Based on prior internal work on English-Inuktitut translation tasks as well as other low-resource tasks, we focused our experiments on multilingual setups. Inuktitut is an agglutinative language, with a lot of particularities. Some Uralic languages like Finnish and Estonian can be considered close to Inuktitut in some linguistic aspects. Most of our experiments are unconstrained with regards to the original WMT task in three ways: • some datasets are taken from previous WMT tasks (English-Finnish, English-Estonian); • some datasets are not in the WMT scope (more recent ParaCrawl3 versions); • some datasets were built in-house at Ubiqus Labs. Some Inuktitut resources can easily be found on the internet, mostly from official government of Nunavut websites and initiatives. We performed two sets of data retrieval: a first one based on paral- lel crawling of multilingual websites, and a second one based on manual retrieval of parallel docu- ments (mostly in PDF format) which then were automatically aligned with a commercial tool. In prior experiments, we also built a set of parallel news articles. Articles were manually retrieved and aligned from both the Inuktitut 4 magazine, which provides parallel versions of all its content in English, French, Inuktitut and Inuinnaqtun, and the Nunatsiaq News5 website, which provides part of its content in both Inuktitut and English. We decided not to include …

Content truncated to 500 words