Addressing Zero-Resource Domains Using Document-Level Context in Neural Machine Translation

Achieving satisfying performance in machine translation on domains for which there is no training data is challenging. Traditional domain adaptation is not suitable for addressing such zero-resource domains because it relies on in-domain parallel data. We show that document-level context can be used to capture domain generalities when in-domain parallel data is not available. We present two document-level Transformer models which are capable of using large context sizes and we compare these models against strong Transformer baselines. We obtain improvements for the two zero-resource domains we study. We additionally present experiments showing the usefulness of large context when modeling multiple domains at once.