LEVERAGING CROSS-LINGUAL TRANSFER LEARNING FOR LOW-RESOURCE NATURAL LANGUAGE PROCESSING

Abstract

Abstract—The field of natural language processing (NLP) is growing quickly, yet, many languages are still under-represented because there is a dearth of labelled data. This study investigates the transfer of knowledge from resource-rich to low-resource languages through cross-lingual transfer learning as a way to overcome this difficulty. We test multilingual models like mBERT and XLM-R on tasks including machine translation, named entity recognition, and sentiment analysis. These models are refined utilising task-specific datasets from low-resource languages after being pre-trained on a variety of languages. Significant gains are demonstrated by the results, particularly in tasks with little labelled data and in languages that are closely linked to those used in the pre-training. These results demonstrate how multilingual models can help close the performance disparities between languages.Overall, this study offers useful information and shows how well cross-lingual transfer learning is in low- resource environments. These findings highlight the potential of multilingual models to reduce perfor- mance gaps across languages.Overall, this work demonstrates the effectiveness of cross-lingual transfer learning in low-resource situations and gives practical insights for designing inclusive NLP systems that better reflect global linguistic variety.