Enhancing distant low-resource neural machine translation with semantic pivot

Prior work has proved that pivot-based method can boost the performance of neural machine translation (NMT). However, in low-resource scenarios, the efficient of pivot-based method is impaired severely due to data sparsity problem. As a typical low-resource language pair, Chinese-Lao NMT suffers the same performance dilemma. In addition, due to the significant linguistic gap between Chinese and Lao, some traditional and effective low-resource translation methods, such as introducing similarity external knowledge, sharing word space, and literal translation, are not suitable for the translation of this language pair. Fortunately, it is highly adaptable to pivot strategy, as there is a pivot language, Thai, which is highly similar to the target language Lao. Here, we propose a novel approach for incorporating similar linguistic features between Thai and Lao into the Chinese-Lao translation model. Firstly, an in-depth linguistic similarity analysis of Thai and Lao is conducted. Secondly, an elaborate pivot-based translation framework with KL adapter is applied. Experiments on the Chinese-Lao translation task show that our approach can help transfer more linguistic knowledges from the Chinese encoder to the Lao decoder via similar linguistic features, achieving substantial improvements compared to the baseline models. • We analysis the linguistic similarities of the Thai-Lao language pair and discuss the practicability of choosing Thai as the pivot language to improve the performance of the Chinese-Lao NMT. • We design a simple but effective data augmentation strategy and apply it sequentially to each stage of pivot translation to overcome the data imbalance in distant low-resource translation scenarios. • We propose a pivot-based NMT model with a new KL-adapter by regrouping the Chinese encoder and Lao decoder exacted from pre-trained Chinese-Thai and Thai-Lao translation models, and then fine-tuning the new model efficiently. • Experiments on the Chinese-Lao translation task show that our approach achieves substantial improvements compared to the strong baseline models and can be a new solution for low-resource NMT.