Curriculum Learning Strategies for Low-Resource Neural Machine Translation

Neural machine translation (NMT) systems for low-resource language pairs remain severely limited by data scarcity, leading to poor generalization and hallucination artifacts. We systematically investigate curriculum learning strategies—ranging from difficulty scoring based on sentence length and vocabulary coverage to competence-based approaches—for improving NMT performance in seven low-resource African language pairs. Our analysis reveals that curriculum order matters substantially: training from easy to hard improves BLEU scores by 4.2–8.9 points across all language pairs, with particularly pronounced gains for morphologically rich languages. We further propose a multi-criteria difficulty scorer that combines linguistic complexity with corpus statistics, achieving an additional 1.8–3.4 BLEU improvement over single-criterion approaches.