Home > Positioning > Subjects > Translation theory > Machine translation

Machine Translation

Translation by machine has run alongside the theory for as long as the discipline has existed, and it now performs a large share of the translation done in the world. The theory’s three conflicts do not disappear when the translator does; they return in a new form.

The line begins with Warren Weaver’s memorandum of 1949, which proposed treating a foreign text as English in code. The Georgetown–IBM demonstration of 1954 promised the problem solved within years; the ALPAC report of 1966 found it had not been and cut the funding; rule-based systems continued in niches. Statistical machine translation, from IBM’s Candide in the early 1990s, learned correspondences from aligned corpora and gave up on rules altogether, and neural machine translation, deployed at scale from 2016, replaced the alignments with a learned representation in which sentences of different languages sit near one another. Translation studies has taken the machines up as an object of its own, in work on machine translation and post-editing by Anthony Pym, Minako O’Hagan and Dorothy Kenny among others, and what the systems do can be read as closest to Nida’s dynamic equivalence and Toury’s target-side norms, since they produce whatever the training data made typical, and furthest from Schleiermacher’s, Benjamin’s and Venuti’s, since fluency is what they optimise. Since 2022 large language models, trained on text of every kind rather than on translation pairs, have taken over much of the task: at the WMT24 shared evaluation general-purpose models led the dedicated systems in most language pairs, under the organisers’ own title, “the LLM era is here but MT is not solved yet”. The old questions return in a new form: what such systems preserve, whose language their data privileges, and whether a translation nobody made can be faithful or free.