This is pointless overcomplicating. I might agree if the example would be slightly more interesting, but "without stripes" isn't even "absence of <stripes>", it is essentially a colour/pattern and can be easily attributed to a range of things exactly the same way "green" can be. Google translate correctly associates much more dubious and abstract concepts than that, and does it with statistical methods, i.e. associating word combinations with a location in a vector space. The fact all major search engines fail to do it here is just shameful. Especially Amazon, where it is pretty much a primary search function.
Something like: build an representation of the text that's language independent and integrated in other knowledge domains, and predict the receiver's reaction to a verbalization of a representation.
But it isn't necessary to formalize any of it. At the current level of sophistication, our informal common ground of words like "understanding" suffice for a discussion. It's obvious Google Translate doesn't resemble human language processing.
