Decoding Time Lexical Domain Adaptation for Neural Machine Translation

Machine translation systems are vulnerable to domain mismatch, especially when the task is low-resource. In this setting, out of domain translations are often of poor quality and prone to hallucinations, due to the translation model preferring to predict common words it has seen during training, as opposed to the more uncommon ones from a different domain. We present two simple methods for improving translation quality in this particular setting: First, we use lexical shortlisting in order to restrict the neural network predictions by IBM model computed alignments. Second, we perform $n$-best list reordering by reranking all translations based on the amount they overlap with each other. Our methods are computationally simpler and faster than alternative approaches, and show a moderate success on low-resource settings with explicit out of domain test sets. However, our methods lose their effectiveness when the domain mismatch is too great, or in high resource setting.