Improving translation between English, Assamese bilingual pair with monolingual data, length penalty and model averaging

Notwithstanding the success of Neural Machine Translation (NMT), we have seen that translation of resource poor Indic languages and techniques applied to the improvement of translation quality for these languages is still under explored. In this work, we have first tried to build two base NMT systems which are capable of translation in two directions for the languages Assamese and English. To build the system, we have used OpenNMT-py, a neural machine translation framework based on PyTorch. We have achieved BLEU score of 9.01 and 14.71 respectively for English to Assamese and Assamese to English translation direction during our base model training with our custom test set. Next, public domain monolingual data in English and Assamese are translated batchwise and newly translated data is used to train new models. Significant BLEU point improvements for both directions are seen. Two different test sets viz. publicly available FLORES-101 test set and our in-house test set of 500 domain specific sentences are used during our experiments. Use of length penalty and model averaging further improved the BLEU scores in both the direction. In this paper, we have shown that use of phased translation of monolingual data with length penalty and model averaging contributes well in achieving higher BLEU scores for data poor languages.