Deep transfer network of heterogeneous domain feature in machine translation

Abstract

In order to address the shortcoming of feature representation limitation in machine translation(MT) system, this paper presents a feature transfer method in MT. Meta feature transfer of the decoding process considered not only their own translation system, but also transferred knowledge of another translation system. The domain meta feature and the objective function of domain adaptation are used to better model the domain transfer task. In this paper, extensive experiments and comparisons are made. The experiment results show that the proposed model has a significant improvement in domain transfer task. The first model has better performance than baseline system, which improves 3.06 BLEU score on the news test set, improves 3.27 BLEU score on the education test set, and improves 3.93 BLEU score on the law test set; The second model improves 3.16 BLEU score on the news test set, improves 3.54 BLEU score on the education test set, and improves 4.2 BLEU score on the law test set.