Improving Lexically Constrained Neural Machine Translation with Source-Conditioned Masked Span Prediction

Accurate terminology translation is crucial for ensuring the practicality and\nreliability of neural machine translation (NMT) systems. To address this,\nlexically constrained NMT explores various methods to ensure pre-specified\nwords and phrases appear in the translation output. However, in many cases,\nthose methods are studied on general domain corpora, where the terms are mostly\nuni- and bi-grams (>98%). In this paper, we instead tackle a more challenging\nsetup consisting of domain-specific corpora with much longer n-gram and highly\nspecialized terms. Inspired by the recent success of masked span prediction\nmodels, we propose a simple and effective training strategy that achieves\nconsistent improvements on both terminology and sentence-level translation for\nthree domain-specific corpora in two language pairs.\n