NLPRL System for Very Low Resource Supervised Machine Translation

Proceedings of the 5th Conference on Machine Translation (WMT), pages 1075–1078 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 1075 NLPRL System for Very Low Resource Supervised Machine Translation Rupjyoti Baruah, Rajesh Kumar Mundotiya, Amit Kumar, Anil Kumar Singh Department of Computer Science & Engineering Indian Institute of Technology (BHU) Varanasi, India {rupjyotibaruah.rs.cse18, rajeshkm.rs.cse16}@iitbhu.ac.in {amitkumar.rs.cse17, aksingh.cse}@iitbhu.ac.in Abstract This paper describes the results of the sys- tem that we used for the WMT20 very low re- source (VLR) supervised MT shared task. For our experiments, we use a byte-level version of BPE, which requires a base vocabulary of size 256 only. BPE based models are a kind of sub-word models. Such models try to ad- dress the Out of Vocabulary (OOV) word prob- lem by performing word segmentation so that segments correspond to morphological units. They are also reported to work across different languages, especially similar languages due to their sub-word nature. Based on BLEU cased score, our NLPRL systems ranked ninth for HSB to GER and tenth in GER to HSB trans- lation scenario. 1 Introduction We report the results for our system that was used for our participation in the WMT20 shared task (Barrault et al., 2019) on very low resource Machine Translation (MT). The MT systems were built for the language pair Upper Sorbian (HSB) and German (GER) in both translation directions. The Sorbian languages are the West Slavic branch of the Indo-European languages, which have further categorized into two closely related languages, Upper Sorbian and Lower Sorbian. The categories of this language are recognized as a dif- ferent and distinct language in the European Char- ter for Regional or Minority languages (Dolowy- Rybinska, 2011). Upper Sorbian is a minority lan- guage of Germany that is spoken by 10, 000 to 15, 000 speakers (Elle, 2010), although this number is continually declining (Dołowy-Rybi´nska, 2018). To counter this, attempts are being made to increase the number of Sorbian speakers through bilingual educational scenarios and MT1. 1https://minorityrights.org/ minorities/sorbs/ Low resource MT was being attempted even be- fore Neural Machine Translation (NMT) became the state-of-the-art. Several methods are used to improve the accuracy and quality of the low- resource SMT systems by using comparable cor- pora (Irvine and Callison-Burch, 2013; Babych et al., 2007), pivot language (English or non- English) technique (Ahmadnia et al., 2017; Paul et al., 2013), and using related resource-rich lan- guage (Nakov and Ng, 2012). We use a byte-level version of Byte Pair En- coding based model with a Transformer for our experiments. The main motivation was to try out this model for the shared task and see how it works under a shared task setting. 2 Background NMT is an end-to-end learning system (Bahdanau et al., 2015), based on the data-driven approach of machine translation, that requires a massive amount of parallel data for training. To overcome the lack of such data, several tech- niques have been tried out which are based on semi-supervised learning (Zheng et al. (2019)), un- supervised learning (Sun et al. …

Content truncated to 500 words