Contact Relatedness can help improve multilingual NMT: Microsoft STCI-MT @ WMT20

Proceedings of the 5th Conference on Machine Translation (WMT), pages 202–206 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 202 Contact Relatedness can help improve multilingual NMT: Microsoft STCI-MT @ WMT20 Vikrant Goyal, Anoop Kunchukuttan, Rahul Kejriwal, Siddharth Jain, Amit Bhagwat STC India, Microsoft, Hyderabad {vikgoyal, ankunchu, rakejriw, sija, amitb}@microsoft.com Abstract We describe our submission for the English→Tamil and Tamil→English news translation shared task. In this submission, we focus on exploring if a low-resource language (Tamil) can benefit from a high-resource language (Hindi) with which it shares contact relatedness. We show utilizing contact related- ness via multilingual NMT can significantly improve translation quality for English-Tamil translation. 1 Introduction In recent years, Neural Machine Translation (Lu- ong et al., 2015; Bahdanau et al., 2015; Johnson et al., 2017; Wu et al., 2018; Vaswani et al., 2017) (NMT) has become the most prominent approach to Machine Translation (MT) due to its simplicity, generality and effectiveness. In NMT, a single neu- ral network often consisting of an encoder and a de- coder is used to directly maximize the conditional probabilities of target sentences given the source sentences in an end-to-end paradigm. NMT mod- els have been shown to surpass the performance of previously dominant statistical machine translation (SMT) (Koehn, 2009) on many well-established translation tasks. However, in order to obtain good translation quality, NMT systems tend to require very large parallel training corpora (Koehn and Knowles, 2017). Such corpora are not yet available for many language pairs. The Indian subcontinent forms a linguistic area where languages from the Dravidian and Indo- Aryan families have been in contact for a long time leading to significant sharing of vocabulary and a convergence of linguistic features (Emeneau, 1956). Tamil is a major language from the Dra- vidian language family spoken in Southern India while Hindi is a widely spoken Indo-Aryan lan- guage. Kunchukuttan and Bhattacharyya (2020) estimate that lexical similarity between Hindi and Tamil to be around 27% in terms of character LCSR (Melamed, 1995), while multiple works have shown that language representations of Tamil and Hindi cluster in the same neighbourhood in a multilingual vector space (Kudugunta et al., 2019; Oncevay et al., 2020). While English-Tamil parallel corpora is limited, more parallel corpora is available for English-Hindi. In this paper, we explore if English-Hindi can im- prove English-Tamil machine translation due to the similarities between Hindi and Tamil on account of contact relatedness. To this end, we train mul- tilingual NMT models for English-Hindi and En- glish Tamil (and vice-versa). Previous work has ex- plored whether high-resource languages can trans- fer knowledge to genetically-related low-resource languages (Nguyen and Chiang, 2017; Dabre et al., 2017). In contrast, we explore if contact relat- edness can benefit low resource languages. We further explore if reducing the divergence between Tamil and Hindi data by representing them in the same script is beneficial. In addition, we explored target agreement models, tagged and noisy back- translation in our submission. 2 Neural Machine Translation Given a bilingual sentence pair (x, y), …

Content truncated to 500 words