Machine in the Loop

Let’s talk turkey.

Businesses and institutions are under pressure to reduce costs. Translation is a time-consuming service performed by skilled professionals. Not surprisingly, it’s expensive.

Neural machine translation has made considerable progress in recent years. Why pay an expensive “human” translator to do something that can be sourced elswhere at zero or near-zero cost?

In a nutshell, because machine translation, despite the name, doesn’t actually translate. Neural machine translation produces a translation by harvesting vast quantities of existing translations, for example from the Internet, then using an algorithm to produce a translation on the basis of statistical probability.

In terms of understanding and reformulation, it literally doesn’t know what it’s doing.

A typical procedure is now for the algorithms to assume the role of the translator, and the resulting text to brought up to scratch by a human being (not necessarily a professional) – an approach termed “human in the loop” or “machine translation post-editing” (MTPE).

Looking at the changes made by a post-editor, it’s easy to conclude that the machine has produced a result accounting for 80% of the final, quality text. But in accordance with the Pareto principle, the remaining 20% is where the heavy lifting, the real work of understanding and reformulation, is to be found.

This is the background to my approach: to consider machine translation merely as a tool in the hands of a professional.

Together with a growing number of colleagues, I use the term “machine in the loop”, or MITL, for this approach.

It should go without saying that since professional translators are already using machine translation in their processes, this use is already priced in.