Improving Transformer based Neural Machine Translation with Source-side Morpho-linguistic Features
Santwana Chimalamarri, Dinkar Sitaram, Rithik Mali, Alex Johnson, K A Adeab · 2020
In this paper, we aim to build and train the extremely popular transformer neural network architectures for carrying out neural machine translation under low resource conditions for a diverse pair of languages. Additionally, we propose to improve the translation predictions in each of these cases by augmenting the source data with a novel linguistically driven morpheme segmentation method as well as additional linguistic features. The experiments were carried out using Kannada as a source language and Telugu, Hindi and English as target languages. These pairs were chosen to reflect diversity in morphological complexity and language similarity, ranging from most similar to the least similar. The evaluations have been done using standard BLEU score metrics.