Cross-Lingual Data Augmentation Techniques: Insights from Multilingual Back Translation

Abstract

This paper investigates the effectiveness of utilizing multiple chains of back translation compared to the traditional method of single-chain back translation for enhancing data diversity in natural language processing (NLP). We explore how multiple rounds of translation and back translation across different languages contribute to enriching the training dataset with diverse linguistic variations. We evaluate the effectiveness of multilingual back translation in achieving better data diversity by reporting the BLEU scores of different back translation techniques. Additionally, we investigate the impact of using languages from different language families and the resulting effect on the diversity of data. Our findings highlight the importance of leveraging multiple chains and multiple language families of back translation for augmenting datasets and provide insights for future research and advancement in data augmentation techniques for NLP.