MDAT: Model and Dictionary Augmented Translation from Indic Language to English using LLM

The recent advent of Large Language Models (LLM) has paved the way for many downstream Natural Language Processing tasks such as Question Answering, Text Summarization, Chatbot applications etc. Recently, the use of LLMs in Machine Translation is also gaining popularity. Some works have shown significant improvement of translation quality with the use of Large Language Models. But, the research in this direction is still in infancy and it demands more extensive studies and careful experiments to establish the fact that LLMs can actually improve the Machine Translation quality. This motivated us to carry out this present work where an open source LLM with different prompting strategies are used to enhance the translation quality of baseline Multilingual Neural Machine Translation (MNMT) system which is capable of translating in 20 different directions involving English and four other low resource Indic languages. With this work, we propose a novel system MDAT (Model and Dictionary Augmented Translation), where translation output of the baseline MNMT model and dictionary words are injected into the prompt to get the translation from Indic language source sentence to English. At the final stage LASER3 scoring helped to get best translation from competing systems. It is found that, this dictionary aided mixed few-shot approach performed better than all other approach in terms of BLEU, TER and chrF evaluation metrics resulting in increase of approximately +9 to +10 (BLEU, chrF++), decrease of approximately -5 (TER) for IN22-GEN test set while translating from Indic language to English. We have not experimented with the translating in the reverse direction, i.e., from English to Indic language.