Enhancing Neural Machine Translation in Low-Resource Settings through Data Augmentation and Transfer Learning

Abstract

Neural Machine Translation (NMT) has achieved state-of-the-art results in recent years; however, its effectiveness largely depends on the availability of large amounts of parallel data between the target language and a high-resource language. In cases where parallel data is limited, there is often a larger amount of monolingual data that could potentially be exploited to improve translation quality. In this work, we study machine translation in a low-resource scenario where the target language is a distant dialect of a highresource language, which makes it possible to explore transfer learning techniques. We begin by training several neural machine translation architectures from scratch to establish baseline performance under limited data conditions. The models explored include sequence-tosequence architectures based on Long Short-Term Memory (LSTM) networks, attention-based Seq2Seq models, as well as Transformer architectures. We then investigate methods to improve translation performance by applying data augmentation techniques and leveraging transfer learning through the fine-tuning of pretrained models. The models are evaluated using the BLEU score and cosine similarity to assess both translation quality and semantic similarity between the generated translations and the reference translations. The results provide a comparative analysis of different architectures and show that combining modern NMT models with data augmentation and transfer learning enhances translation performance in low-resource settings.