The advances in the field of neural machine translation (NMT) have
both led to an exciting leap forward in translation quality and
motivated scholars to re-examine the models of the human
translation process. Even though NMT systems are good at
simulating rule-like language behaviour and making generalisations,
they tend to overgeneralise and they tend to forget infrequent
translation patterns while generating translations. Recent work
showed that the quality of NMT systems can substantially be
improved by integrating explicit translation exemplars into the NMT
architecture. These improvements in translation quality brings up
new and exciting research questions about the role of exemplars in
the field of MT, such as the importance of string-based, semantic and
syntactic similarity measures in finding useful exemplars, which can
be transferred to the human translation process and language
behaviour.
This research proposal uses techniques from computational
linguistics to (i) study the impact of adapting NMT systems through
exemplars; (ii) determine the role of different similarity levels, e.g.
string-based, semantic and syntactic, in retrieving useful translation
exemplars; examine (iii) whether the improvements in translation
quality can be observed in different domains, language pairs; and (iv)
whether translation exemplars can successfully be used to adapt
general-domain NMT systems towards specific domains.