Resource Sufficient Data Generation Technique for Machine Translation Models Using Named Entity Transliteration and Cross- Lingual Word Embeddings

Abstract

Machine translation offers speed and cost advantages, but human translation remains superior in quality and contextual understanding. For low-resource languages like Hausa, the severe scarcity of parallel data leads to poor model performance, overfitting, and inadequate handling of named entities. While existing synthetic data generation methods such as back-translation and data copying introduce noise, this study proposes a technique that integrates named entity transliteration with cross-lingual word embeddings to generate synthetic parallel sentences. The approach replaces the common practice of entity copying with transliteration to preserve target-language vocabulary integrity, while embeddings translate non-entity words. Evaluated on the FLORES-200 and MAFAND-MT benchmarks, the unsupervised setting achievedBilingual Evaluation Understudy  (BLEU) scores of 1.04 (Hausa→English) and 1.42 (English→Hausa) on FLORES-200, while the semi-supervised setting improved substantially to 15.56 (Hausa→English) and 12.42 (English→Hausa) BLEU on MAFAND-MT outperforming the monolingual data copying baseline by up to 5 BLEU points. These results demonstrate that the proposed technique offers an effective strategy for augmenting scarce parallel data in low-resource MT systems.