Low-Resource NMT: an Empirical Study on the Effect of Rich Morphological Word Segmentation on Inuktitut

Low-Resource NMT: an Empirical Study on the Effect of Rich Morphological Word Segmentation on Inuktitut Ngoc Tan Le le.ngoc tan@courrier.uqam.ca Fatiha Sadat sadat.fatiha@uqam.ca Department of Computer Science, Universite du Quebec a Montreal 201, avenue du President-Kennedy, Montreal, QC, Canada Abstract Nowadays, very little research for Indigenous languages has been studied. Indigenous lan- guages bring significant challenges for Natural Language Processing approaches because of multiple features such as polysynthesis, morphological complexity, dialectal variations with rich morpho-phonemics, spelling with noisy data and low resource scenarios. Particularly, morphological segmentation for polysynthetic languages is challenging because a word may contain many individual morphemes and training data can be extremely scarce. The current research paper focuses on Inuktitut, one of the Indigenous polysynthetic language spoken in Northern Canada. First, a rich word segmentation for Inuktitut is studied using a set of rich features and by leveraging (bi-)character-based and word-based pretrained embeddings from large-scale raw corpora. Second, we incorporated this pre-processing step into our first Neu- ral Machine Translation system. Our evaluations showed promising results and performance improvements in the context of low-resource Inuktitut-English neural machine translation. 1 Introduction In the Americas, there are a wide range of linguistic families about 140 linguistic families in the world. About 900 different Indigenous languages spoken in the Americas approximately are reported in (Mager et al., 2018). In Canada, there is a great diversity of Indigenous languages, grouped into 12 language families, that have been central to the history of First Nations people, M´etis and Inuit in Canada and continue to play a vital role to this day (Rice, 2011). This research paper focuses on Inuktitut, one of the Indigenous polysynthetic languages spoken in Northern Canada and the development of a Neural Machine Translation (NMT) for Inuktitut-English. The main objective and motivation of this project is the revitalization and preservation of Indigenous languages and cultural heritage through major tasks in NLP. However, the development of Indigenous language technology faces many challenges such as polysynthesis with a high rate of morphemes per word, lack of orthographic normalization, dialectal variations and lack of linguistic resources and tool such as corpora (Littell et al., 2018). This first step towards a multilingual NMT framework that will involve several endangered Indigenous languages of Canada, is essential as the only parallel corpus freely available for research is the Nunavut-English Hansard corpus (Joanis et al., 2020). Inspired by the work of Farley (2012), related on the creation of the first Inuktitut morpho- logical analyzer, we build a neural network-based word segmenter for Inuktitut. The main goal Proceedings of the 14th Conference of the Association for Machine Translation in the Americas October 6 - 9, 2020, Volume 1: MT Research Track Page 165 of this research is two folds: (1) to investigate empirically several word segmentation meth- ods on Inuktitut; and (2) to improve low-resource NMT that involves Inuktitut through rich morphological word segmentation. The structure of the paper is described as follows: Section 2 presents the state-of-the-art on MT involving Indigenous languages. In section 3, we …

Content truncated to 500 words