Extended Parallel Corpus for Amharic-English Machine Translation

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 6644–6653 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 6644 Extended Parallel Corpus for Amharic-English Machine Translation Andargachew Mekonnen Gezmu, Andreas N¨urnberger, Tesfaye Bayu Bati Otto von Guericke Universit¨at Magdeburg Universit¨atsplatz 2, Magdeburg, Germany {andargachew.gezmu, andreas.nuernberger}@ovgu.de tesfayebayu@hu.edu.et Abstract This paper describes the acquisition, preprocessing, segmentation, and alignment of an Amharic-English parallel corpus. It will be helpful for machine translation of a low-resource language, Amharic. We freely released the corpus for research purposes. Furthermore, we developed baseline statistical and neural machine translation systems; we trained statistical and neural machine translation models using the corpus. In the experiments, we also used a large monolingual corpus for the language model of statistical machine translation and back-translation of neural machine translation. In the automatic evaluation, neural machine translation models outperform statistical machine translation models by approximately six to seven Bilingual Evaluation Understudy (BLEU) points. Besides, among the neural machine translation models, the subword models outperform the word-based models by three to four BLEU points. Moreover, two other relevant automatic evaluation metrics, Translation Edit Rate on Character Level and Better Evaluation as Ranking, reflect corresponding differences among the trained models. Keywords: Statistical Machine Translation, Neural Machine Translation, Less-Resourced Language 1. Introduction To automate the intricate task of translation, re- searchers have followed different approaches. The earliest attempt was to use rule-based systems, which are criticized for being tedious and expensive to de- velop. Alternative empirical approaches such as Statis- tical Machine Translation (SMT) and Neural Machine Translation (NMT) came when parallel corpora were more and more available. Such methods take advantage of the authentic translations made by human translators in parallel corpora. They rely on machine learning to build translation models by taking parallel corpora as training data. For empirical machine translation, we need a paral- lel corpus (bitext), a text that has a parallel transla- tion in another language. Machine translation models are trained on a parallel corpus. Some international and governmental institutions provide such texts for public use. For instance, the Canadian Hansard cor- pus (Roukos S. et al., 1995) consists of parallel texts in English and French, drawn from official records of the proceedings of the Canadian Parliament. Simi- larly, the Europarl corpus (Koehn and Monz, 2005), extracted from the proceedings of the European Par- liament, contains parallel corpora for twenty-one Eu- ropean languages. The United Nations (UN) Parallel Corpus (Ziemski et al., 2016) is available in six offi- cial UN languages. The current version of the parallel corpus consists of manually translated UN documents between 1990 and 2014. Other parallel corpora have been made from movie subtitles, like the OpenSubtitles corpus (Lison and Tiedemann, 2016), or from general web text, like the ParaCrawl corpus (Ba˜n´on et al., 2020). The Open Par- allel Corpus (OPUS) (Tiedemann, 2012) collects paral- lel corpora from sources such as open-source software documentations and religious books. Large numbers of parallel corpora are available for dominant international languages such as …

Content truncated to 500 words