Abstract
This study investigates machine translation between related languages i.e.,languages within the same family that share similar linguistic traits such asword order and lexical similarity. Machine translation through few-shotprompting leverages a small set of translation pair examples to generatetranslations for test sentences. This requires the model to learn how togenerate translations while simultaneously ensuring that token ordering ismaintained to produce a fluent and accurate translation. We propose that forrelated languages, the task of machine translation can be simplified byleveraging the monotonic alignment characteristic of such languages. Weintroduce a novel approach of few-shot prompting that decomposes thetranslation process into a sequence of word chunk translations. Throughevaluations conducted on multiple related language pairs across variouslanguage families, we demonstrate that our novel approach of decomposedprompting surpasses multiple established few-shot baseline models, therebyverifying its effectiveness. For example, our model outperforms the strongfew-shot prompting BLOOM model with an average improvement of 4.2 chrF++ scoresacross the examined languages.