It seems that a new pricing model is being shaped, and it can finally put an end to the longstanding debate on how machine translation post-editing (known as MTPE, but renamed as revision by TAUS lately) can be fairly priced.
The problem is not new: ever since machine translation (MT) appeared (and seduced with the reduction of costs on the buyer side), there has been a lively debate going on over how this relatively new service can be appropriately paid for.
Both sides have understandable viewpoints: clients would like to spare money on their investment in MT, while translators would like to earn as much as with translation (at least timewise). Let us examine the shortcomings of the previous pricing models and the ways forward.
Forget about word-based and time-based pricing
What worked as a unit of measurement for translation and even for revision, unfortunately does not work for MTPE.
The biggest issue with MTPE is that the time spent on a segment can vary greatly depending on the quality of the MT output: reading a badly translated segment plus deleting and retranslating it from scratch obviously takes longer than simply translating it.
However, it can easily happen that the quality is perfect, or close to perfect, and the reviewer can quickly move on to the next segment.
The situation is further complicated by the fact that with the widespread use of neural machine translation (NMT) the occurrence of grammatically correct but otherwise mistranslated sentences is becoming more frequent.
If not thoroughly checked by the reviewer, these segments can be mistakenly skipped in time-sensitive projects. (Other dangers of NMT are detailed in another of our blog articles.)
It is understandable from the above that a price based on the source wordcount is not practical, because it is not representative of the work effort, especially when the text contains both perfect segments and those that need retranslating.
An hourly price (based on the reported number of hours spent) is not ideal either, as clients prefer predictable prices, not to mention the fact that started hours often count as full hours at some places, and this rounded-up pricing can significantly increase the costs.
A slightly more sophisticated method, pricing based on the so-called ‘editing distance’, is on the rise.
This means that the CAT (computer assisted translation) tool logs how much the final improved version of the segment differs from the original machine translanslation.
After that, several pricing categories can be applied that are linked to the percentage of changes: for example, fully retyped sentences are paid at full price, segments modified less than 50% are paid at half price, while unchanged segments are paid at a quarter of the full price or even less.
Minimum charge and maximum charge at the same time?
Pricing based on the editing distance may be the fairest method. However, it needs to be more sophisticated if the ultimate goal is fair payment.
Normally, when using CAT software, not only machine translated content is used, but also matches from the translation memory (TM), which are basically approved, reliable translations. Exact and fuzzy matches are translated from the TM, and only the ‘no match’ segments are pre-translated using the MT engine.
Some of the big language service providers have already found the right recipe: they subtract the editing distance (in percent) from 100%, and the segment is paid based on that percentage as if it were a fuzzy match from the TM. That is, if 20% was rewritten, the linguist receives the same amount of money as for an 80% match in the TM.
By ‘maximum charge’ I mean that the client calculates what would be the highest cost of the project if all the MT segments were fully retranslated and paid at full rate.
At the end of the process, the amount to be paid is modified based on the real editing effort (and not the time spent). If a minimum charge can be added (which is needed in order to make small jobs worth dealing with), the solution might be perfect for both parties.
Even if the final cost is not predictable, the client is aware of the maximum they could pay, while the translator is aware of the minimum.
Of course, there are other questions, for instance: how could we prevent translators from rewriting fully correct sentences?
A translator could introduce some edits into appropriate machine translated segments just to attract a higher price category.
This cannot be completely avoided; however, if we suppose that everybody likes to finish their tasks in a timely manner, and get to the end of a translation quickly, then unnecessary editing will probably not happen very often.
A somewhat related problem is the fact that the review step cannot be skipped if the client needs a reviewed quality, as post-editors are replacements for the translators rather than the reviewers in the process. In other words, machine translation is there for speeding up the translation step, not for replacing it.
If the buyer side of the language industry eventually recognizes what a responsible and accurate job it is to post-edit machine translation, then a mutually beneficial consensus can be reached around the above-mentioned pricing model.