Extended Study on Using Pretrained Language Models and YiSi-1 for Machine Translation Evaluation

Proceedings of the 5th Conference on Machine Translation (WMT), pages 126–133 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 126 Linguistically Motivated Subwords for English-Tamil Translation: University of Groningen’s Submission to WMT-2020 Prajit Dhar Arianna Bisazza Gertjan van Noord University of Groningen {p.dhar, a.bisazza, g.j.m.van.noord}@rug.nl Abstract This paper describes our submission for the English-Tamil news translation task of WMT- 2020. The various techniques and Neural Machine Translation (NMT) models used by our team are presented and discussed, in- cluding back-translation, fine-tuning and word dropout. Additionally, our experiments show that using a linguistically motivated subword segmentation technique (Ataman et al., 2017) does not consistently outperform the more widely used, non-linguistically motivated Sen- tencePiece algorithm (Kudo and Richardson, 2018), despite the agglutinative nature of Tamil morphology. 1 Introduction In this paper we present the neural machine transla- tion (NMT) systems submitted to the WMT-2020 English-Tamil (EN→TA) news translation task. This task is challenging mainly for two reasons: 1. Differing syntax: English is an Indo- European language which is fusional and SVO (Subject-Verb-Object). On the other hard, Tamil is part of the Dravidian language family and is a SOV language that is agglu- tinative. A good NMT system is expected to discern the various morphological forms on the Tamil target side. 2. Scarcity of training data: Prior to WMT-2020, there existed only a few corpora for paral- lel EN-TA sentences (Ramasamy et al., 2012; Germann, 2001). This left us with the choice of either only utilizing the low amount of par- allel sentences or finding out ways of artifi- cially enlarging the training data. Through our submission we wish to provide solu- tions to the following questions: • Is linguistically motivated subword segmen- tation beneficial for EN-TA translation? • Can the addition of TA monolingual data com- pensate for the small amount of parallel EN- TA sentences despite the domain mismatch? • Can fine-tuning on a corpus of Indian news improve quality on the WMT news transla- tion task? We start our paper with a short description of the Tamil language before delving into the vari- ous techniques adopted by our submitted NMT sys- tems. 2 Tamil Language Tamil is a Dravidian language spoken by around 80 million people. Tamil morphology is agglutina- tive and suffixal, i.e words are formed by suffixing morphemes to a lemma (Annamalai et. al 2014, cited in Sarveswaran et al. (2019)). Tamil suffixes can be either derivational (marking a change in PoS and/or meaning) or inflectional. In particu- lar, nouns in Tamil are inflected for number, gen- der, case and animacy while verbs are inflected for tense, mood, aspect, negation, interrogation, infor- mation about emphasis, speaker perspective, sen- tience or rationality, and conditional and causal re- lations. Table 4 shows examples of the case forms in singular for the noun ��தகம ’book’. All the aforementioned statements substantiate the fact that Tamil morphology is highly complex. In fact, Ramasamy et al. (2012) identified 716 in- flectional rules for nouns and 519 rules for verbs. Furthermore, designing a translation …

Content truncated to 500 words