Effective Cross-lingual Transfer of Neural Machine Translation Models without Shared Vocabularies
This work addresses the challenge of cross-lingual transfer for low-resource machine translation, offering a method that outperforms existing multilingual training approaches.
The paper tackles the problem of transferring pre-trained neural machine translation models to unrelated languages without shared vocabulararies, achieving improvements of up to +5.1% BLEU in low-resource translation tasks.
Transfer learning or multilingual model is essential for low-resource neural machine translation (NMT), but the applicability is limited to cognate languages by sharing their vocabularies. This paper shows effective techniques to transfer a pre-trained NMT model to a new, unrelated language without shared vocabularies. We relieve the vocabulary mismatch by using cross-lingual word embedding, train a more language-agnostic encoder by injecting artificial noises, and generate synthetic data easily from the pre-training data without back-translation. Our methods do not require restructuring the vocabulary or retraining the model. We improve plain NMT transfer by up to +5.1% BLEU in five low-resource translation tasks, outperforming multilingual joint training by a large margin. We also provide extensive ablation studies on pre-trained embedding, synthetic data, vocabulary size, and parameter freezing for a better understanding of NMT transfer.