Enhancing Neural Machine Translation with Syntactic Ambiguities

Abstract

Benefiting from the data-driven end-to-end model architecture, neural machine translation has obvious performance advantages over statistical machine translation, but its demand for data is also significantly greater, including monolingual and parallel corpus. Most of the past studies have focused on reducing the demand for parallel corpus or making more effective use of limited parallel corpus. In this work, we have studied a method of using ambiguity of syntactic structure to achieve more effective use of monolingual corpus. Experiments conducted on multiple benchmarks for various languages show that our method has a greater improvement than the method using back-translation only, demonstrating the effectiveness of our proposed method.

Tasks

Reproductions