Separating Grains from the Chaff: Using Data Filtering to Improve Multilingual Translation for Low-Resourced African Languages

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1001 - 1014 December 7-8, 2022 ©2022 Association for Computational Linguistics Separating Grains from the Chaff: Using Data Filtering to Improve Multilingual Translation for Low-Resourced African Languages Idris Abdulmumin1,2∗, Michael Beukman3∗, Jesujoba O. Alabi4∗, Chris Emezue5,6, Everlyn Asiko7,8, Tosin Adewumi9, Shamsuddeen Hassan Muhammad2,10, Mofetoluwa Adeyemi, Oreen Yousuf11, Sahib Singh12, Tajuddeen Rabiu Gwadabe2,13 All: MasakhaneNLP, 1Ahmadu Bello University, Zaria, Nigeria, 2HausaNLP, 3University of the Witwatersrand, South Africa, 4Saarland University, Germany, 5TUM, Germany, 6Mila - Quebec AI Institute, 7University of Cape Town, South Africa, 8African Institute for Mathematical Sciences, 9Luleå University of Technology, Sweden, 10LIAAD-INESC TEC, Porto, Portugal, 11Uppsala University, Sweden, 12Ford Motor Company, 13University of Chinese Academy of Sciences, China iabdulmumin@abu.edu.ng Abstract We participated in the WMT 2022 Large- Scale Machine Translation Evaluation for the African Languages Shared Task. This work de- scribes our approach, which is based on filter- ing the given noisy data using a sentence-pair classifier that was built by fine-tuning a pre- trained language model. To train the classifier, we obtain positive samples (i.e. high-quality parallel sentences) from a gold-standard cu- rated dataset and extract negative samples (i.e. low-quality parallel sentences) from automat- ically aligned parallel data by choosing sen- tences with low alignment scores. Our final machine translation model was then trained on filtered data, instead of the entire noisy dataset. We empirically validate our approach by eval- uating on two common datasets and show that data filtering generally improves overall trans- lation quality, in some cases even significantly. 1 Introduction This paper presents Masakhane NLP’s submission to the WMT 2022 large-scale machine translation evaluation for African languages. We participated in the constrained translation task and chose to focus on a subset of all the language pairs con- sidered for this task due to resource constraints. We specifically explore the language directions {hau, ibo, lug, swa, tsn, yor, zul}↔eng and wol↔fra, and submitted our primary and secondary systems which were competitive with other submissions for this task. Machine translation has received much atten- tion recently, especially for low-resourced lan- guages (Adelani et al., 2022a; Fan et al., 2021; Haddow et al., 2022; Hoang et al., 2018; Nekoto et al., 2020). A promising approach for such se- tups is to fine-tune large pre-trained language mod- els on the available small amount of translation ∗* Equal contribution. data (Neubig and Hu, 2018; Adelani et al., 2021a, 2022a). While most of these language models are trained on predominantly high-resourced language datasets (Conneau et al., 2020; Devlin et al., 2019; Radford et al., 2018), there have been a few mod- els that were pre-trained (Ogueji et al., 2021) or adaptively fine-tuned (Alabi et al., 2022) only on low-resourced languages. Recent works have tried, successfully, to sup- plement the existing small amounts of natural data in low-resource languages with artificially gener- ated parallel data. For instance, in machine trans- lation, Sennrich et al. (2016) and Ueffing (2006) padded the true parallel data with automatic trans- lations of monolingual sentences through back- …

Content truncated to 500 words