IITP-MT at WAT2018: Transformer-based Multilingual Indic-English Neural Machine Translation System
Sukanta Kumar Sen, Kamal Kumar Gupta, Asif Ekbal, Pushpak Bhattacharyya · Institutional Repositories DataBase (IRDB) · 2018
This paper describes the systems submitted by the IITP-MT team to WAT 2018 multilingual Indic languages shared task. We submit two multilingual neural machine translation (NMT) systems (Indic-to-English and English-to-Indic) based on Transformer architecture and our approaches are similar to many-to-one and one-to-many approaches of Johnson et al. (2017). We also train separate bilingual models as baselines for all translation directions involving English. We evaluate the models using BLEU score and find that a single multilingual NMT model performs better (up to 14.81 BLEU) than separate bilingual models when the target is English. However, when English is the source language, multi-lingual NMT model improves only for low-resource language pairs (up to 11.60 BLEU) and degrades for relatively high-resource language pairs over separate bilingual models.