A Chinese-Malay Neural Machine Translation Model Based on CA-Transformer and Transfer Learning

Neural machine translation (NMT) has achieved good results in many applications, but large-scale corpora are needed as support, and high-quality corpora are often difficult to obtain, especially for low-resource corpora such as small languages, resulting in poor translation results. Therefore, a transfer learning method is proposed, in which the training parameters of Chinese-English high resource corpus are used to replace the model parameters of Chinese-Malay and English-Malay respectively, so as to solve the problem of insufficient training and scarce corpus. Compared with the basic transformer, this method can effectively reduce the number of parameters, speed up the training, and solve the problem that the pretrained language model is too large and difficult to optimize. Moreover, the performance is basically similar to transformer. Compared with multiple models, the translation model in this paper is faster, with better translation effect and better BLEU.