Yasmin Moslem

Title: MT system selection and recycling/fixing recycling candidates in a hybrid set-up

Supervision Team: Andy Way, DCU / John Kelleher, TUD

Description: Domain-tuned MT systems outperform general domain MT models, when they are used to translate in-domain data. It may not always be known in advance of translation time which domain is best suited to a particular text or sentence, and even for a known domain like software, some strings may be better translated by a general domain system. This gives rise to a number of research questions, including: Given multiple domain-tuned NMT systems, and translation candidates, how do we analyze an incoming string and determine which system will do the best translation at runtime? How do we best assess which translation candidate is the best choice? What are the best approaches for NMT? Also, if we have access to recycling (in a Translation Memory), when is a recycling match better than an MT candidate? Can NMT help fix high quality TM matches? Can a better translation candidate be found by combining elements of multiple translations, from recycling and MT systems? Can post-editing data be leveraged, e.g. a form of automatic post-editing approach?