Improving Translation Quality for Low-Resource Inuktitut with Various Preprocessing Techniques

Proceedings of Recent Advances in Natural Language Processing, pages 475–479 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_053 475 Improving Translation Quality for Low-Resource Inuktitut with Various Preprocessing Techniques Mathias Hans Erik Stenlund, Matilde Nanni, Micaella Bruton, Meriem Beloucif Uppsala University meriem.beloucif@lingfil.uu.se Abstract Neural machine translation has been shown to outperform all other machine translation paradigms when trained in a high-resource set- ting. However, it still performs poorly when dealing with low-resource languages, for which parallel data for training is scarce. This is es- pecially the case for morphologically complex languages such as Turkish, Tamil, Uyghur, etc. In this paper, we investigate various preprocess- ing methods for Inuktitut, a low-resource in- digenous language from North America, with- out a morphological analyzer. On both the orig- inal and romanized scripts, we test various pre- processing techniques such as Byte-Pair Encod- ing, random stemming, and data augmentation using Hungarian for the Inuktitut-to-English translation task. We found that there are ben- efits to retaining the original script as it helps to achieve higher BLEU scores than the roman- ized models. 1 Introduction While state-of-the-art Machine Translation (MT) systems are achieving close to human-like transla- tions on a restricted set of highly researched lan- guages (Luong et al., 2015; Sennrich et al., 2015; Luong and Manning, 2016; Neubig, 2015; Cho et al., 2014; Luong et al., 2017; Vaswani et al., 2017), they fail to obtain equally good results on languages for which there is a lack of resources (Haddow et al., 2022). In fact, these end-to-end neural encoder-decoder MT systems are quite data hungry, requiring parallel datasets in the tens or even hundreds of millions of sentences to outper- form statistical models; datasets which are only available for a few of the spoken languages of the world (Ranathunga et al., 2021). The unavailability of parallel data for most world languages is only the tip of the iceberg because, even when there is data available, the data can be very domain-specific and contain a lot of noise (Haddow et al., 2022). Ranathunga et al. (2021) and Haddow et al. (2022) provide an overview of current research methods tackling low-resource MT, by addressing different aspects and problems. The data and tools scarcity problem in NLP creates the need to simulate low- resource scenarios by taking a small sample of data from a high-resource language so that currently existing tools can be easily tested in low-resource settings (Haddow et al., 2022). The lack of suit- able preprocessing tools hinders research on these languages (Haddow et al., 2022). When available, linguistic tools, such as morphological segmen- tation, are paramount for preprocessing the data and obtaining subword segmentation, to better deal with out-of-vocabulary words; the most common strategies include BPE and SentencePiece (Haddow et al., 2022). In this paper, we tackle the issue of preprocess- ing and its effect on translation quality when deal- ing with a highly agglutinative and morpholog- ically complex low-resource language, Inuktitut. Our goal is to test several preprocessing techniques to determine which yields the best MT results …

Content truncated to 500 words