Machine Translation, Human Evaluation, Translation Memory, Confidence
Estimation
This paper reports on an initial study that aims to understand whether the acceptability
of translation memory (TM) among translators when contrasted with machine translation (MT)
unacceptability is based on users’ ability to optimise precision in match suggestions. Seven
translators were asked to rate whether 60 English-German translated segments were a usable basis
for a good target translation. 30 segments were from a domain-appropriate TM without a quality
threshold being set, and 30 segments were translated by a general domain statistical MT system.
Participants found the MT output more useful on average, with only TM fuzzy matches of over
90% considered more useful. This result suggests that, were the MT community able to provide an
accurate quality threshold to users, they would consider MT to be the more useful technology.