Adam Mickiewicz University’s English-Hausa Submissions to the WMT 2021 News Translation Task

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 167–171 November 10–11, 2021. ©2021 Association for Computational Linguistics 167 Adam Mickiewicz University’s English-Hausa Submissions to the WMT 2021 News Translation Task Artur Nowakowski and Tomasz Dwojak Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Pozna´n, Poland {artur.nowakowski,t.dwojak}@amu.edu.pl Abstract This paper presents the Adam Mickiewicz University’s (AMU) submissions to the WMT 2021 News Translation Task. The submissions focus on the English↔Hausa translation directions, which is a low-resource translation scenario between distant languages. Our approach involves thorough data cleaning, transfer learning using a high-resource lan- guage pair, iterative training, and utilization of monolingual data via back-translation. We experiment with NMT and PB-SMT approaches alike, using the base Transformer architecture for all of the NMT models while utilizing PB-SMT systems as comparable baseline solutions. 1 Introduction We describe the Adam Mickiewicz University’s submissions to the WMT 2021 News Translation Task. We focused on translation between Hausa and English – a low-resource translation scenario between distant languages. Our methods combine data cleaning with OpusFilter (Aulamo et al., 2020) and fastText (Joulin et al., 2016), transfer learning (Aji et al., 2020; Zoph et al., 2016), iterative train- ing, and back-translation (Sennrich et al., 2016a). All NMT models were trained with FAIRSEQ (Ott et al., 2019), while the first it- eration of the back-translation was generated with Moses (Koehn et al., 2007). The results presented in the paper are based on the first released development set ("Dev-1"), which consists of 1000 sentences, the final devel- opment set ("Dev-full"), which adds additional 1000 sentences to the first development set, and the released test set without additional test suites ("Test"). The test set consists of 1000 sentences in English→Hausa direction and 997 sentences in Hausa→English direction. The final submissions significantly outperform the vanilla NMT baselines in terms of BLEU (Pap- ineni et al., 2002) metric results, as implemented in SACREBLEU (Post, 2018) with default settings. All systems were trained in a constrained sce- nario i.e., using the data provided by the organizers of WMT 2021 only. 2 Data preparation The quality of the training data has a great im- pact on the final performance of the NMT mod- els (Rikters, 2018). The data preparation consisted of data cleaning and filtering performed by using OpusFilter (Aulamo et al., 2020) pipelines. We specified separate pipelines for monolingual and parallel data. Data cleaning phase consisted of normalizing punctuation, removing non-printable characters, and decoding HTML entities by using Moses (Koehn et al., 2007) pre-processing scripts. We applied subword segmentation on fil- tered data by using SentencePiece (Kudo and Richardson, 2018) tool with byte-pair-encoding (BPE) (Sennrich et al., 2016b) algorithm. The cor- pora we used for model training, along with the number of sentences before filtering, are specified in Table 1. Number of sentences after filtering is presented in Table 2. Monolingual data filtering For the monolingual data filtering, we defined an OpusFilter pipeline that consists of the following filters: • deduplication filter, • sentence length filter, • word length filter, …

Content truncated to 500 words