Experiment on reiterated translation

Question

I would like to know whether

1. this is feasible or has already been done by someone
2. it's possibile to predict its outcome

The experiment I came up with is the following: a random text written in language A is chosen, alongisde a target language B. Then the following two steps are reiterated N times:

1. The text is translated into language B
2. The translated text gets translated back into A

If done by humans, the translator would have to be a different person each time not to get biased (there will have to be 2*N translators). The meaning of the text is assumed to be preserved at each step.

After this whole process has ended, I'd like to know how the final text in language A differs from the original copy, and what has changed in all the different versions of the text. Possible outcomes I have though of so far:

1. The text converges to a final text for N suffcientely large. (Is the final text simpler, more complex or neither?)
2. The text changes indefinitely at each step.
3. Complex or uncommon words/constructions turn into simpler ones, simpler meaning that this word/construction is more common.
4. All the synonyms appear in the various copies of the text in a ratio similar to that you have in the spoken/written language.
5. Nothing can be said as the choice of A and B influences the outcome too much. (What would be the right parameter to do so then?)

Help with the tags is appreciated, as I'm not a linguist nor a linguistics student.

Answer 1

It has been tried but people seem to have given up on it seeing that it does not help at all. I quote from the paper Europarl: A Parallel Corpus for Statistical Machine Translation

They use a MT system to translate a sentence from English into a foreign language, and then use a reverse MT system to translate the sentence back into English. They then judge the quality of the MT systems by how well the English sentence is preserved.

This method is inspired by an urban legend involving a pair of MT systems between Russian and English. The legend proclaims that once someone fed a English–Russian MT system the bible verse “The spirit is willing, but the flesh is weak.” When back translating the sentence with the Russian– English system, the system returned “The vodka is good but the meat is rotten.“

How well does back translation indicate the translation performance of the MT systems involved? As Table 4 shows, not much.

In conclusion, back translation does not only provide a false sense of the capabilities of MT systems, it is also a lazy and flawed method to compare systems. Back translation unfairly benefits from the ability to reverse errors, which only show up in the foreign language. To drive the point home: a system pair that does nothing, meaning, leaving all English words in place will do perfectly in back translation, while being utterly useless in practise.

Answer 2

This may not be the scientific experiment you wish, but something of the kind has been done at least once.

It is "The Celebrated Jumping Frog of Calaveras County", a short story written by Mark Twain (still Samuel Longhorne Clemens), translated into French, then retranslated by the outraged Twain to Englih. Of course, he then published all three versions together as "The Jumping Frog: in English, then in French, and then Clawed Back into a Civilized Language Once More by Patient, Unremunerated Toil".

Of course, all of it is now available on the web, though the version I found does not seem the best one could hope for.