Low-resource Neural Machine Translation: Benchmarking State-of-the-art Transformer for Wolof<->French

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 6654–6661 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 6654 Low-resource Neural Machine Translation: Benchmarking State-of-the-art Transformer for Wolof↔French Cheikh M. Bamba Dione, Alla Lo, Elhadji Mamadou Nguer, Sileye Oumar Ba University of Bergen, Gaston-Berger University, UVS, Loreal Research & Innovation Norway, Senegal, Senegal, France dione.bamba@uib.no, lo.alla@ugb.edu.sn, elhadjimamadou.nguer@uvs.edu.sn,sileye.ba@dailymotion.com Abstract In this paper, we propose two neural machine translation (NMT) systems (French-to-Wolof and Wolof-to-French) based on sequence-to-sequence with attention and Transformer architectures. We trained our models on the parallel French-Wolof corpus (Nguer et al., 2020) of about 83k sentence pairs. Because of the low-resource setting, we experimented with advanced methods for handling data sparsity, including subword segmentation, backtranslation and the copied corpus method. We evaluate the models using BLEU score and find that the transformer outperforms the classic sequence-to-sequence model in all settings, in addition to being less sensitive to noise. In general, the best scores are achieved when training the models on subword-level based units. For such models, using backtranslation proves to be slightly beneficial in low-resource Wolof to high-resource French language translation for the transformer-based models. A slight improvement can also be observed when injecting copied monolingual text in the target language. Moreover, combining the copied method data with backtranslation leads to a slight improvement of the translation quality. Keywords: French, neural machine translation, sequence-to-sequence, transformers, Wolof 1. Introduction Neural Machine Translation (NMT) based on the encoder-decoder framework has achieved impressive results in high-resource data conditions (Sutskever et al., 2014; Cho et al., 2014). For low-resource con- ditions, recent studies have shown that NMT perfor- mance can be improved by using subword units (Sen- nrich et al., 2016b), backtranslation (Sennrich et al., 2015), and by adapting NMT systems to low-resource settings (Sennrich and Zhang, 2019). In this research work, we investigate the design and implementation of NMT systems to translate between French (FR)1 and Wolof (WO, ISO 639-3), a low- resource Niger-Congo language mainly spoken in Senegal (Gamble, 1950). We selected two NMT archi- tectures: (i) an attentional sequence-to-sequence model (henceforth S2S) based on gated recurrent units (GRU) (Cho et al., 2014), and (ii) a Transformer (Vaswani et al., 2017; Hassan et al., 2018). According to recent studies, transformers have shown great promise as an approach to NMT for low-resource languages (Abbott and Martinus, 2018). At the same time, however, trans- former models remain difficult to optimize and require careful hyper-parameter tuning to achieve good trans- lation quality (Nguyen and Salazar, 2019). As a main contribution of the paper, we have imple- mented and benchmarked state-of-the-art NMT sys- tems, which are of high interest for French↔Wolof translators, thereby considerably facilitating their work. Our work will provide to large Wolof commu- nities useful language resources they would not other- 1As the official language of Senegal (the country of the most Wolof speakers), it is easier to find parallel data between French and Wolof than between e.g. English and Wolof. wise have access …

Content truncated to 500 words