Abstract:

In contexts where training resources are extremely limited, neural machine translation models based on deep learning often fail to achieve their desired performance. Current methods in transfer learning that leverage similar languages rely on intuition to select analogous data for rudimentary pre-training and potentially fail to pinpoint the most similar languages or fully exploit the advantages of pre-training. Hence, a framework for low-resource machine translation transfer learning based on linguistic similarity is proposed. This framework selects five low-resource languages translated into English as tasks. Initially, six high-resource languages are chosen, and the model is pre-trained on machine translation datasets from these languages to English. Subsequently, employing linguistic similarity metrics, the translation model that is most similar to the target language pair is selected for transfer, ultimately resulting in enhanced model performance via refined fine-tuning strategies. The experimental findings demonstrate that models trained within this framework exhibit superior performance, compared with baseline models, to thereby offer a viable and versatile approach for low-resource machine translation tasks.

Key words: machine translation, transfer learning, linguistic similarity, low-resource