Lecture

Machine Translation (MT) has been used by many people for some time now as
a productivity tool, with demonstrable success. Recently, a new paradigm --
Neural MT (NMT) -- has emerged as a contender to replace statistical MT as
the new state-of-the-art in the field. While there is no doubt that NMT has
enormous potential, we argue that this has been overhyped, a situation which
if left unchecked will lead to unrealistic expectations of its capabilities
and ultimately to a host of disappointed users. We will demonstrate that this
has been seen before with the advent of new approaches to MT, and each time
this happens it has the potential to undermine the relationship MT developers
need to have with translators, the principal users of the technology. Despite
overblown claims regarding its prospects, we argue – as we have done for many
years now – that the translation community has little to fear from MT, and that
where human input is required, the translator remains the most important link
in the chain. Indeed, MT developers rely on translators in a number of ways,
and if MT is to improve still further, bilateral partnerships need to be formed
between both communities for such advances to be made as fast as possible.