Who Pays the Price for Poor MT?

The first time I heard this, it sent a chill down my spine: “It doesn’t matter if the machine translation output is [insert 4-letter expletive starting with an 'S']: the post-editors will clean it up.”

It’s not the MT vendor who suffers when MT is bad, nor is it the end customer. The truth is that it is the post-editors who pay the price for amateurish MT. By the time MT output gets to the end customer, any major flaws will have been fixed. By humans.

For those who rely on post-editors to fix bad MT, what process you use to generate MT doesn’t matter at all. It doesn’t matter which MT engine you use and whether or not it is trained well, because in the end there is always one secret weapon to make things right: the post-editors.

The post-editors, hired to repair machine translation, are the ones who pay the price for faulty processes. They carry the full brunt of a badly-trained MT engine, or of an engine that does not fit the content thrown into it. Post-editors work as hard as they need to so that customers get the right quality, regardless of what shape their material was in when the MT engine spit it out. And they do this at a steep discount over their normal rates.

The problem with this system is that it is not sustainable. For one thing, every post-editor unfairly paid to fix bad MT errors means that there may well be one less post-editor who will say yes to the next project. I believe that poorly-managed MT is the main reason that the pool of those translators who are willing to be post-editors is not growing as it should.

Respecting the time of post-editors is not the only reason, however, that companies should strive for quality MT output. Given that the quality of the raw output determines how quickly a post-editor can progress, ensuring higher quality from the get-go means the kind of productivity gains – and thus cost savings – that make machine translation the most attractive technology today for lowering overall translation spend.

If we go beyond the idea that with MT engines, “One Size Fits All”, and treat MT as an industrial process rather than aims at identifying the highest performing engine in a given context and trains the heck out of it, we can reduce the pain felt by post-editors. And that will herald in a new era of quality MT.