COMET-QE and Active Learning for Low-Resource Machine Translation

TL;DR

This paper demonstrates that using COMET-QE as a reference-free evaluation metric in active learning significantly improves sentence selection for low-resource neural machine translation, outperforming RTTL and random methods.

Contribution

It introduces COMET-QE as an effective tool for active learning in low-resource machine translation, showing superior performance over existing methods.

Findings

COMET-QE outperforms RTTL and random selection by up to 5 BLEU points.

Active learning with COMET-QE reduces data requirements for effective translation.

Results are demonstrated on Swahili, Kinyarwanda, and Spanish datasets.

Abstract

Active learning aims to deliver maximum benefit when resources are scarce. We use COMET-QE, a reference-free evaluation metric, to select sentences for low-resource neural machine translation. Using Swahili, Kinyarwanda and Spanish for our experiments, we show that COMET-QE significantly outperforms two variants of Round Trip Translation Likelihood (RTTL) and random sentence selection by up to 5 BLEU points for 20k sentences selected by Active Learning on a 30k baseline. This suggests that COMET-QE is a powerful tool for sentence selection in the very low-resource limit.