Machine translation (MT) is an important application in our
increasingly connected world. Rapid advances in artificial
intelligence have led to dramatic improvements in data-driven MT
systems (i.e. statistical and neural MT). When an adequate quantity of
training data is available, systems can now produce translations of
sufficient accuracy to be usable in many practical scenarios.
However, the performance of any data-driven system critically depends
on whether the characteristics of its training data match the
characteristics of the data the system encounters in everyday use.

During the summer of 2018, we investigated translation in three
scenarios where domain adaptation is needed: the informal language
common in social media; technical documents such as patents and
scientific articles; and noisy text output by optical character
recognition (OCR) systems. Each of these domains poses important
vocabulary adaptation challenges — scientific documents contain
unknown technical jargon, while social media and OCR results contain
misspellings or variants of known vocabulary. Some of the techniques
for increasing robustness that we explored are use of monolingual
word embeddings, adaptive data filtering, bilingual dictionary mining,
and models of lexical variation. We also compared the performance
of statistical and neural systems on conditions that vary the amount
of available data, the degree of domain mismatch, and approaches to
adaptation.