The stink of machine translation

Listen to Lisa Agostini’s interview of me talking about my presentation. Thanks to MET and Julian Mayers (Yada Yada) per their permission to post the interview here.

Post-editors are asked to do either light post-editing, to get rid of the worst machine translation errors, or full post-editing, to bring the output up to the same standard as human translation.

But is full post-editing in reality a pipe dream?

The speaker has conducted an experiment for two years running with groups of postgraduate university students in which half do an unaided human translation and the other half post-edit machine translation output. Comparison of the texts produced shows that certain turns of phrase, expressions and choices of words occur with greater frequency in the post-edited machine translation output than they do in human translation. This is easily explained by the fact that even neural machine translation systems seem to choose the most statistically frequent solutions even when those solutions occur less frequently than the sum of the frequencies of all the other possible solutions, and post-editors faced with an acceptable solution tend not to edit it. This however implies that post-edited machine translation output, on average, lacks the variety and inventiveness of human translation, and therefore does not in fact reach the same standard. It is evident that the additional post-editing effort required to eliminate what are effectively machine translation markers would nullify most, if not all, of the time and cost-saving advantages of post-edited machine translation. On the other hand, failure to eradicate these markers may eventually lead to lexical and syntactic impoverishment of the target language.

The speaker provides examples of post-editing and translation from English into Italian. However, with the aid of some back-translations, the mechanisms at play should be equally clear to non-Italian speakers, particularly if they are familiar with other Neo-Latin languages.