Neural Machine Translation for Similar Languages: The Case of Indo-Aryan Languages

Proceedings of the 5th Conference on Machine Translation (WMT), pages 424–429 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 424 Neural Machine Translation for Similar Languages: The Case of Indo-Aryan Languages Santanu Pal1, Marcos Zampieri2 1Wipro AI Lab, India 2Rochester Institute of Technology, USA santanu.pal2@wipro.com Abstract In this paper we present the WIPRO-RIT systems submitted to the Similar Language Translation shared task at WMT 2020. The second edition of this shared task featured parallel data from pairs/groups of similar languages from three different language families: Indo-Aryan languages (Hindi and Marathi), Romance languages (Catalan, Portuguese, and Spanish), and South Slavic Languages (Croatian, Ser- bian, and Slovene). We report the results obtained by our systems in translating from Hindi to Marathi and from Marathi to Hindi. WIPRO-RIT achieved competi- tive performance ranking 1st in Marathi to Hindi and 2nd in Hindi to Marathi trans- lation among 22 systems. 1 Introduction WMT 2020 is the fifth edition of WMT as a conference following a series of well-attended workshops that date back to 2006. WMT be- came a well-established conference due to its blend of research papers and popular shared tasks on different topics such as translation in various domains (e.g. biomedical, news), translation quality estimation, and automatic post-editing. The competitions co-organized with WMT provide important datasets and benchmarks widely used in the MT commu- nity. The vast majority of these tasks so far, however, involved training systems to trans- late to and from English (Bojar et al., 2016, 2017) while only a few of them addressed the problem of translating between pairs of lan- guages with less resources. To address this issue, in 2019, the Simi- lar Language Translation (SLT) shared task was introduced at WMT. SLT’s purpose was to evaluate the performance of state-of-the-art MT systems on translating between pairs of similar languages without English as a pivot language (Barrault et al., 2019). The or- ganizers provided participants with training, development, and testing parallel data from three pairs of languages from three different language families: Spanish - Portuguese (Ro- mance languages), Czech - Polish (Slavic lan- guages), and Hindi - Nepali (Indo-Aryan lan- guages). Systems were evaluated using auto- matic metrics, namely BLEU (Papineni et al., 2002) and TER (Snover et al., 2006). In SLT 2020, the task organizes once again included an Indo-Aryan language track with Hindi and Marathi. Indo-Aryan languages are a sub-family of the Indo-European lan- guage family which includes Bengali, Bo- hjpuri, Hindi, Marathi, and Nepali. These lan- guages are mainly spoken in North and Cen- tral India, and some neighbouring countries such as Nepal, Bangladesh, and Pakistan etc. The script used in most of these languages are derived from the ancient Brahmi script and en- riched with high grapheme to phoneme corre- spondence leading to many orthographic sim- ilarities across these languages. In addition to Hindi and Marathi, SLT 2020 features two other tracks with similar lan- guages from the following language families: Romance languages (Catalan, Portuguese, and Spanish) and South Slavic Languages (Croa- tian, Serbian, …

Content truncated to 500 words