Abstract:
The present article discusses how to improve translation quality when using limited training data to translate towards morphologically rich languages. The starting point is a neural MT system, used to train translation models, using solely publicly available parallel data. An initial analysis of the translation output has shown that quality is sub-optimal, due mainly to an insufficient amount of training data. To improve translation quality, a hybridized solution is proposed, using an ensemble of relatively simple NMT systems trained with different metrics, combined with an open source module, designed for a low-resource MT system. Experimental results of the proposed hybridized method with multiple independent test sets achieve improvements over (i) both the best individual NMT and (ii) the standard ensemble system provided in the Marian-NMT system. Improvements over Marian-NMT are in many cases statistically significant. Finally, a qualitative analysis of translation results indicates a greater robustness for the hybridized method.
