Enhancing Neural Machine Translation of Low-Resource Languages: Corpus Development, Human Evaluation and Explainable AI Architectures

Data scarcity is a critical barrier for underrepresented languages: The research demonstrates that low-resource language pairs like English-Irish and English-Marathi face significant challenges in machine translation development due to limited parallel datasets. The creation of gaHealth, the first bilingual health corpus for Irish, shows how domain-specific data collection can dramatically improve translation quality—achieving "very significant improvements in BLEU score" compared to previous benchmarks.

Specialized datasets outperform general ones: Models trained on in-domain health data substantially outperformed general-purpose models, highlighting that for low-resource languages, targeted corpus development in specific sectors (like healthcare) may be more effective than attempting broad language coverage.

Human evaluation reveals important quality differences: Beyond automated metrics, human evaluation using detailed error taxonomies showed that Transformer-based systems significantly reduced both accuracy and fluency errors compared to older RNN-based approaches, demonstrating that the choice of underlying architecture meaningfully impacts real-world translation quality for speakers of these languages.

Accessibility tools are essential for democratizing MT development: The introduction of open-source applications (adaptNMT and adaptMLLM) that simplify model development and deployment makes machine translation technology more accessible to developers and professional translators working with low-resource languages, potentially enabling more communities to build their own translation systems.

Responsible AI requires explainability in translation systems: The paper's emphasis on "explainable AI architectures" suggests that for languages serving smaller communities, understanding why a translation system makes specific choices is important for quality assurance and building trust among users.