Enriching the Transfer Learning with Pre-Trained Lexicon Embedding for Low-Resource Neural Machine Translation

Most State-Of-The-Art (SOTA) Neural Machine Translation (NMT) systems today achieve outstanding results based only on large parallel corpora. The large-scale parallel corpora for high-resource languages is easily obtainable. However, the translation quality of NMT for morphologically rich languages is still unsatisfactory, mainly because of the data sparsity problem encountered in Low-Resource Languages (LRLs). In the low-resource NMT paradigm, Transfer Learning (TL) has been developed into one of the most efficient methods. It is difficult to train the model on high-resource languages to include the information in both parent and child models, as well as the initially trained model that only contains the lexicon features and word embeddings of the parent model instead of the child languages feature. In this work, we aim to address this issue by proposing the language-independent Hybrid Transfer Learning (HTL) method for LRLs by sharing lexicon embedding between parent and child languages without leveraging back translation or manually injecting noises. First, we train the High-Resource Languages (HRLs) as the parent model with its vocabularies. Then, we combine the parent and child language pairs using the oversampling method to train the hybrid model initialized by the previously parent model. Finally, we fine-tune the morphologically rich child model using a hybrid model. Besides, we explore some exciting discoveries on the original TL approach. Experimental results show that our model consistently outperforms five SOTA methods in two languages Azerbaijani (Az) and Uzbek (Uz). Meanwhile, our approach is practical and significantly better, achieving improvements of up to $\mathrm{4.94}$ and $\mathrm{4.84}$ BLEU points for low-resource child languages Az $\mathrm{\to }$ Zh and Uz $\mathrm{\to }$ Zh, respectively.