VietTayNMT: Directional LoRA for Low-Resource Vietnamese–Tay Neural Machine Translation

Abstract

Low-resource machine translation for minority languages is hindered by limited parallel data, weak lexical coverage, and deployment constraints. This paper studies Vietnamese–Tay translation and introduces VietTayNMT, a lightweight bidirectional framework designed for both translation quality and edge-oriented efficiency. We construct a manually refined Vietnamese–Tay parallel corpus and combine dictionary-guided pretraining with SentencePiece tokenization and Directional LoRA fine-tuning. The proposed adaptation strategy shares encoderside adapters while using direction-specific decoders and cross-attention adapters for Vietnamese-to-Tay and Tay-to-Vietnamese generation. Experiments show that VietTayNMT achieves competitive translation quality, with BLEU scores above 31 in both directions, while remaining fine-tunable on Jetson Orin NX. These results demonstrate the feasibility of VietTayNMT for practical minority-language translation in resource-constrained settings.