Tags diminishing machine translation results

Thijs Vissia wrote:

It seems to me that sending the tags along is seriously reducing the quality of the MT results.

Milan Condak wrote:

Translator can remove tags before pretranslation against TMX or using MT, (...)
and put them back after pretranslation.

Milan

Thijs Vissia wrote:
I was wondering about the results I’m getting from Machine Translation through OmegaT. Currently I’m using one service that is available free of charge, MyMemory (Machine). I’m noticing that a lot of idiom is not being recognised and translated accordingly when tags are inbetween several words that make up an idiom.

For example, I was translating the following sentence (in Dutch):
“Ieder instituut gaat beschikken over meer geld en meer gebouwen.”

With some tags/formatting strewn in, this became:
“Ieder <f0>instituut gaat beschikken</f0><f1> </f1><f2>over meer geld</f2> en meer gebouwen.”

In there, there is the widespread idiom “beschikken over” (meaning “have access to”, “have at ones disposal”). I don’t know about the quality of the machine translation at MyMemory, but as a widespread idiom this should be recognised and translated properly.

When the tags were in there, this was returned as:
“Everyone <f0> institute </f0><f1></f1><f2> about more money </f2> and more buildings.”

The connection between “ieder” and “institute” was broken by the tag, so instead of “every institute” it rendered this as “everyone institute”. Similarly, the composite verb “beschikken over”, was also interrupted by a tag, so the MT treated each piece separately, and apparently left out the verb entirely.

However, after creating a new file without tags, it came back as:
“Every institute will have more money and more buildings.”
Which may not be my phrasing of choice but otherwise a fine translation.

So I was wondering why the tags get sent out to the machine translation services in the first place? Is it so that all the formatting doesn’t need to be put back in manually afterwards? Wouldn’t it be almost as easy to strip the strings of the tags before sending the query to the MT service?

Considering that OmegaT already needs to recognise tags as such (to treat them differently in the editor pane), wouldn't it be possible to make sending them to the MT service optional?

It seems to me that sending the tags along is seriously reducing the quality of the MT results.