RNN-based Derivation Structure Prediction for SMT

Feifei Zhai, Jiajun Zhang, Yu Zhou, Chengqing Zong

Abstract: In this paper, we propose a novel derivation structure prediction (DSP) model for SMT using recursive neural network (RNN). Within the model, two steps are involved: (1) phrase-pair vector representation, to learn vector representations for phrase pairs; (2) derivation structure prediction, to generate a bilingual RNN that aims to distinguish good derivation structures from bad ones. Final experimental results show that our DSP model can significantly improve the translation quality.