Flexible Customization of a Single Neural Machine Translation System with Multi-dimensional Metadata Inputs

摘要

Advances in neural machine translation (NMT) technology not only significantly raised machine translation quality for general-purpose, out-of-the-box systems, but also provided a way for additional input signals to an NMT model to effectively influence its output for a given translation unit, so that a single NMT system can serve different customer needs. At the same time, language service providers, media companies, and other businesses started to systematically store metadata associated with their translatable content. In this work, we show how these metadata can be used both for training and at inference time to flexibly customize a given NMT system to produce somewhat different translations of the same translation unit. The metadata as an extra input to NMT can enable such customization across multiple dimensions and at different levels of input granularity: for individual documents or their collections, for a given post-editing session of a professional translator, or even for individual sentences.