As I understand it, Translation Memory saves users money by reusing translations already completed. For example, the title of an article from English to Spanish.
What I would like to understand is how smart is it, and what can it reuse?
Examples:
This is with the caveat that it's my recent experience; it should be said that I'm sure World Server can discuss any specific requirements / expectations that a client may have and provide the appropriate implementation etc.
So the sentences as we see them are not necessarily the same as the algorithm that parses them for World Server.
From my experience:
Phrases etc. are broken into segments. Segments appear in a particular order. Let's say
When this is translated and sent back to Tridion it's now in the translation memory.
If you resend this you will get what's called an ICE match. That's a 100% of words/terms and in the same order. This is translated directly from memory and (dependant on how you agree your World Server side workflow) sent back immediately.
If you send a component on it's own, (for the sake of argument lets' map our component as a segment) let's say the one that constituted Seg_3
This won't now be returned as an ICE match. It should still be translated directly from memory as the segment as a unit matches but it's context is different (that is, the content before and after it).
In the workflow processes I'm currently dealing with this is checked by a translator to confirm that the context shouldn't result in an update to the translation.
Let's say it got updated by the translator... Now we have two sets of content in the translation memory
and
If you sent a third context for Let me know if you need help the you would see, in the World Server GUI that there is more than one option for the translation - or you may require an explicit one for this context.
Of course, sending
Seg_1: Let me know if you need help
on it's own again would result in the ICE match from our second submission.
............. It goes on then that if you updated a word (and this includes placeholders if you use them for labels in your text) then the whole segment really changes and all the words within it would likely require translation - furthermore it would then change the context of the surrounding segments that would likely require translation and not be seen as ICE matches at this point.
If I have the exact same content in a different component or field (place), is it used?
Yes. And some changes will be applied automatically before the translator see it. For example changed numbers and markup.
If I change one word or add grammar in a sentence, it is used?
Yes. The translator will see the matches and how close they are. If they above a certain threshold (the translator can configure this) they will be presented as a "starting" point for the translation. If they are at 100% they are typically considered to need review only (and both TMS and WorldServer supports skipping this). You typically pay less for matched text (the higher the match the less you pay).
You can for example read how World Server presents the matched levels here: http://producthelp.sdl.com/SDL%20WorldServer%202011/source/topic569.html
TMS will do something similar.
Next to the actual text other factors can also influence the match level. A translation memory can be tagged, and if the memory contain the translation does not have tags matching the current translation a penalty is applied to the match. This can for example apply penalty from a memory used for translating product information if it is being used when translating marketing material. This setup would require an active management of multiple memories and tags for the translation jobs, so not everyone will use this.
TMS has PerfectMatch and WorldServer has ICE match to detect translations that do not even need review (or less review - it is really up to the translation workflow). In World Server this is called ICE (In-context exact), and in TMS this is referred to as Perfect Match. The idea is that if the source text did not change and it is used in the exact same way then it can be used directly.
At what level is content compared for matches word, sentence, paragraph, field?
This is typically done per sentence, though paragraph is seen as well. This is known as segmentation. Notice the translator can override this segmentation if needed.
