Enhancing Zero-Shot Machine Translation through Pre-Trained Multilingual Models and Cross-Lingual Transfer Learning
Multidisciplinary Innovations & Research Analysis · 16.05.2026 · 7 (2) · 22-31
DOI: 10.65923/tsqq8d05 ↗Kurz gefasst
Fachartikel in der Zeitschrift Multidisciplinary Innovations & Research Analysis, online erschienen am 16.05.2026. Verfasst von Hadia Azmat.
Abstract (Originalsprache)
Zero-shot machine translation (ZS-MT) has emerged as an effective solution for translating between language pairs that lack direct parallel training data. Despite recent advances in neural machine translation, conventional models initialized with random parameters often exhibit limited generalization when applied to unseen language pairs. This study investigates the effectiveness of multilingual pre-training and cross-lingual transfer learning in improving zero-shot translation performance. A multilingual Transformer architecture is employed to learn shared semantic representations across multiple languages, enabling knowledge transfer between high-resource and low-resource language pairs. Experimental evaluations are conducted using benchmark multilingual datasets derived from the WMT and IWSLT translation tasks. Translation quality is assessed using BLEU scores and comparative performance analysis. Results demonstrate that multilingual pre-training significantly improves translation accuracy, fluency, and semantic consistency when compared with randomly initialized neural translation systems. Furthermore, cross-lingual transfer learning enables effective utilization of linguistic knowledge acquired from related languages, reducing the dependence on parallel corpora. The findings suggest that multilingual pre-training and transfer learning provide a scalable framework for improving zero-shot translation capabilities and extending machine translation systems to underrepresented languages.
Zitieren
APA
Azmat, H. (2026). Enhancing Zero-Shot Machine Translation through Pre-Trained Multilingual Models and Cross-Lingual Transfer Learning. Multidisciplinary Innovations & Research Analysis, 7(2), 22-31. https://doi.org/10.65923/tsqq8d05
Vancouver
Azmat H. Enhancing Zero-Shot Machine Translation through Pre-Trained Multilingual Models and Cross-Lingual Transfer Learning. Multidisciplinary Innovations & Research Analysis. 2026;7(2):22-31. doi:10.65923/tsqq8d05
BibTeX
@article{Azmat2026,
author = {Azmat, Hadia},
title = {Enhancing Zero-Shot Machine Translation through Pre-Trained Multilingual Models and Cross-Lingual Transfer Learning},
journal = {Multidisciplinary Innovations & Research Analysis},
year = {2026},
volume = {7},
number = {2},
pages = {22--31},
doi = {10.65923/tsqq8d05}
}
EUVIMED