Improving the Performance of Low-Resource SMT Using Neural-Inspired Sentence Generator

Abstract

Machine translation (MT) is an important aspect of natural language processing (NLP) which uses bilingual data set and other language assets to translate text from a source language to text in a target language. The widely used statistical machine translation (SMT) system mainly relies on translation memories and glossaries to learn language pattern and translation rules. However the usage of these SMT systems under low-resource conditions remain a challenge. These SMT systems have a higher output quality when trained using domain specific training data since the texts belonging to same domain follow same pattern or usage of words. The current paper aims to develop and improve the performance of an SMT system under low-resource conditions for translation of health domain specific text from Tamil-to-English and English-to-Tamil. The translation quality of the machine translation system is improved using tuning based on minimum error rate training (MERT) and a novel neural-inspired sentence generator as a post-processor. The quality of translation and its performance analysis is evaluated in terms of the bilingual language understudy (BLEU) score and translation edit rate (TER).