Machine Translation (MT) in Memsource
What is MT?
MT is Machine Translation, which means translation done by a machine – unlike the reverse abbreviation, TM, which is a Translation Memory and stores translations done by humans. Although both originally had some form of human input, but in different ways. MT is done by a reasonably clever machine, naturally depending on the language pair and the engine used. But it’s still quite stupid in some ways. Let’s look at what it can and can’t do.
- Put whole sentences together with the correct grammar (if you’re really lucky!). If this happens to you, congratulations! Some language pairs are more successful at this than others.
- Suggest useful terminology that may be relevant in the context.
- Always understand sentence structure and produce correct word order for all sentence types. You’ll often find that you need to move things like adverbs around to make things make sense in your target language. Some languages may require more drastic post-editing.
- Guarantee the correct terminology. You will still need to check with reference material/TMs/TBs (termbases) to make sure the right term has been used, particularly if you know there are customer-specific terms. Consider the terms produced by MT merely as suggestions. In fact, consider everything produced by MT as a suggestion!
- Learn based on the correct versions you are inputting. (Well, it can but the engines we are currently using don’t do this at present.) This means that any corrections you make will not be stored in the MT engine. They will, however, be stored in the TM for future use.
So how does it work in practice? Let’s look at an example using Memsource.
You’ll see a few interesting things in the third column with the match percentages. Unlike hits from the TM, the MT hits are underlined or are in blue. This tells you that the segment is from the MT engine rather than the TM.
Let’s take a closer look at the match percentages below.
- 100%: Perfect MT output—probably no post-editing required
- 99%: Near-perfect MT output—possibly minor post-editing required, mostly for formatting or punctuation issues
- 75%: Good MT match, but likely to require some human post-editing
- No score (blue): This means MTQE cannot confidently identify the quality (it may be high or low), so the output needs to be checked by a Linguist.
(from the Memsource website)
You can see that it’s basically the same idea as TM matches. And you can also view information about matches for each segment in the CAT pane.
Here, you can clearly see what comes from the MT engine (underlined) and what comes from the TM in the usual way (not underlined) and make your choices accordingly. It may be a good idea to read the CAT Pane page in Memsource Help if you’re not familiar with it (https://help.memsource.com/hc/en-us/articles/115003642952-The-CAT-Pane).
So, you’ve now done your first translation using MT and it’s time for QA. Let’s look at what happens there.
You’ll see in the QA Pane that there are lots of segments of the type “Unedited NT/MT fuzzy match”. These can either be NT (non-translatables – things like numbers and other content that is not usually translated) or MT that is less than a 99% match. You should obviously read through your entire translation before delivering but you should pay particular attention to these segments.
This is most likely to be where you may have missed something! So, this is where the real post-editing work comes in – where you check the terminology with the TM/TB etc. and make sure that the translation is fully accurate and fluent in your target language. It may look good at first glance, but you need to weed out any errors, clumsy wording or inappropriate terminology.
That’s probably enough to get you started. We’ll update this guide as and when new information and any further issues emerge.
Best of luck on your MT journey!
Let us know how it goes email@example.com