Prime time for machine translation postediting

The very reason for this is that Kilgray's founders all used to work on machine translation at MorphoLogic, and there we learned what was missing. And that is the feedback cycle. In the last five years we had expected MT companies to improve on this, but no real improvement was made to date. And we have sympathy for translators - editing the same mistakes again and again is demotivating, to say the least. Another thing that we were completely missing was the translation mindset of MT companies. Most of them thought that MT is only good for translating millions of pages, and never thought about using it as a productivity tool for human translators.

With the advent of memoQ 5, we realized that one tool that we created for the review cycle, the audit trail / version tracking is also the best tool so far for MT post-editing, and it is the only tool that actually allows you to measure the return on investment for machine translation.

Traditionally, machine translation post-editing was paid by the hour, and the problem with hourly rates is that translators tend to have different speed. Even in regular translation there is often a difference of 3-4 times between a slow and a fast translator. For example when I was translating for a living, I only took IT jobs, and I had a lot of experience, so I could offer a relatively modest price (even by Hungarian standards) and make a lot of money. Indeed, there were months when I made more money than today as chief operating officer of a respected technology company. Up to now, post-editing was paid by the hour, so the output of two people were not comparable. memoQ 5 allows you to quantify changes, and this proposes a new business model: pay some checking fee per word or character, and pay some editing fee per words edited. I think no-one will argue that this is unfair.

By knowing what the post-editor has changed, you can also know what should be improved in machine translation, what should be entered into the dictionary, what sort of sentences need to be trained, etc. After many years, one machine translation company has finally decided that they would have a feedback cycle. You can't imagine how happy we are about this decision. This opens a link between MT and TM vendors, and expect a strong partnership between the two companies.

So let's look at some MT tips and tricks now. First of all, how does memoQ support MT post-editing?

Those MT engines that can be integrated with memoQ are usually quite expensive, and we learned that most translation companies that offer MT post-editing services do not use the APIs and server versions. The two-column RTF filter in memoQ does a good job. You can export the imported source text from memoQ into two-column RTF. This gives you the source segments in the second column. The target segments should come into the third column. Perform a machine translation on this file. Now take the memoQ export and the translated file, and copy the second (translated) column of the translated file, and paste it into the third column of the original file. Then import it back into memoQ. Your grid will have the machine translated output.

How else does memoQ support machine translation? First, restructuring sentence order is easy. Use the ctrl+shift+B and ctrl+shift+N shortcuts to highlight a part of the text and move it left and right. The terminology QA check is also of great benefit. While most of the memoQ quality assurance relates to errors committed by human translators, parsing errors might happen in MT and the QA can check for the integrity of terminology. One of the most hilarious translations I have ever read was for this sentence:

"Government members serve fixed terms of five years."

The translation was:

"A kormány, amelyet tagok szolgálnak, megjavította öt év szakkifejezéseit."

Explained in English:

"The government that is served by members has corrected technical terms created in the last five years."

If "government member" is a term, memoQ can automatically point out the shortcoming in this translation. While the translation is perfect grammatically, it hardly makes sense (hello Noam!).

All this functionality is there even in a memoQ translator pro version, so anyone who is a bit innovative can start experimenting with MT post-editing. If the customer knows, you call it post-editing, if the customer does not, and you do a good job, it's a productivity boost. Of course we understand that MT works differently between different languages.