Research on Unsupervised Domain Adaptation System for Machine Translation

Abstract

With the intensive research on neural networks and deep learning, neural machine translation proposed in recent years has greatly improved translation quality and gradually replaced traditional statistical-based machine translation. Although neural machine translation can achieve very high translation accuracy with translation models trained on resource-rich outer domains, it tends to perform poorly in other inner domains where resources are scarce. Domain adaptation is one approach to enhance its performance, which uses resource-rich domains to help improve the translation quality of machine translation in resource-scarce domains. In this paper, we try to build a parallel corpus in the inner domain by using the parallel corpus in the outer domain to better train the translation model, which is eventually invoked by the system to realize the translation. The experimental results show that the accuracy of the machine translation system is significantly improved when translating data from the inner domain after applying this method.