Neural Machine Translation in Low-Resource Setting: a Case Study in English-Marathi Pair

Neural Machine Translation in Low-Resource Setting: a Case Study in English-Marathi Pair Aakash Banerjee abanerjee@cse.iitb.ac.in Aditya Jain adityajainiitb@cse.iitb.ac.in Shivam Mhaskar shivammhaskar@cse.iitb.ac.in Sourabh Deoghare sourabhdeoghare@cse.iitb.ac.in Aman Sehgal aman.sehgal@iiml.org Pushpak Bhattacharyya pb@cse.iitb.ac.in Department of Computer Science and Engineering, IIT Bombay, India. Abstract In this paper, we explore different techniques of overcoming the challenges of low- resource in Neural Machine Translation (NMT), specifically focusing on the case of English-Marathi NMT. NMT systems require a large amount of parallel corpora to obtain good quality translations. We try to mitigate the low-resource problem by augmenting parallel corpora or by using transfer learning. Techniques such as Phrase Table Injection (PTI), back-translation and mixing of language corpora are used for enhancing the parallel data; whereas pivoting and multilingual embeddings are used to leverage transfer learning. For pivoting, Hindi comes in as assisting language for English-Marathi translation. Compared to baseline transformer model, a significant improvement trend in BLEU score is observed across various techniques. We have done extensive manual, automatic and qualitative evaluation of our systems. Since the trend in Machine Translation (MT) today is post-editing and measuring of Human Effort Reduction (HER), we have given our preliminary observations on Translation Edit Rate (TER) vs. BLEU score study, where TER is regarded as a measure of HER. 1 Introduction The aim of this work is to improve the quality of Machine Translation (MT) for the English-Marathi language pair for which less amount of parallel corpora is available. One of the major requirements for good performance of the Neural Machine Translation (NMT) models is the availability of a large parallel corpora. As a result, there is a need to come up with additional resources by augmenting parallel corpora or by using knowledge from other tasks using transfer learning. Kunchukuttan and Bhattacharyya (2020) have shown that the lexical and ortho- graphic similarities among languages can be utilized to improve translation quality be- tween Indic languages when limited parallel corpora is available. English and Marathi does not have common ancestry and hence are not related languages, whereas Hindi and Marathi are related languages. Also, among the various English to Indic language Proceedings of the 18th Biennial Machine Translation Summit Virtual USA, August 16 - 20, 2021, Volume 1: MT Research Track Page 35 corpora, English-Hindi corpus is comparatively larger. In our pivot based transfer learn- ing, combined corpus, and multilingual experiments we try to utilize this high resource English-Hindi language pair in various ways to assist in English-Marathi translation. In our Phrase Table Injection (PTI) experiment, we explore how the phrases generated during Statistical Machine Translation (SMT) model training can be further utilized in NMT. We also explore how the monolingual corpus of the target language can be leveraged to create additional pseudo-parallel sentences using back-translation. We also try to understand the correlation between the BLEU and Translation Edit Rate (TER) scores by fitting a linear regression line on the TER vs BLEU graph, where TER is regarded as a measure of Human Effort Reduction (HER). 2 Related Work Transformer model …

Content truncated to 500 words