With the nationwide promotion of Mandarin, regional dialects are gradually fading, especially among the elderly, who often face communication barriers due to limited proficiency in Mandarin. This negatively impacts their quality of life and social participation. This study aims to enable high-quality bidirectional translation between Cantonese and Mandarin, contributing to dialect preservation and the inheritance of intangible cultural heritage. Based on the Transformer architecture, we fine-tuned Metas multilingual translation model, NLLB-200, using a self-constructed Cantonese-Mandarin parallel corpus. Data sources include subtitles from short video platforms, local forums, and community interview transcripts, resulting in a high-quality corpus of 200,000 sentence pairs. Technically, we employed transfer learning and data augmentation strategies to enhance performance in low-resource environments, and evaluated the model using BLEU and chrF metrics. On the test set, the fine-tuned model achieved a 17.3% improvement in BLEU score, with translations showing natural fluency, indicating that NLLB-200 has strong dialect translation capabilities. Additionally, we explored deploying the system on mobile devices to develop a lightweight voice translation application for elderly users, enhancing usability and accessibility. This research not only validates the effectiveness of NLLB-200 in low-resource language translation tasks but also provides a reference path for the promotion and application of multi-dialect translation technologies. By combining technological innovation and social service, it significantly contributes to the protection and revitalization of dialects in the digital era.
Jinyang Wang (Wed,) studied this question.