The MiMo2 translation theory defines two types of differences in the way in which languages encode meaning: content words, and other syntactic means (word order, function words, morphology). It is based on the idea that content word meaning is difficult to represent in a universal way, and so it uses transfer to deal with this aspect of meaning.
A good example of the reason for having transfer is the difference between the Dutch word `schimmel' and the English translation `white horse'. The logical meaning is the same (in fact, this is not quite true, since a `schimmel' cannot be a black horse painted white - we abstract away here from treating problems like this). But Dutch uses a primitive expression and English a complex one. Now if there is no transfer, there must be one `pivot language' or `interlingua' that serves as the point of communication between the two languages. A question is, whether the pivot language encodes the piece of meaning of this example as a primitive expression (e.g. `schimmel'), or as a complex one (e.g. `white horse'). In the first case, the English grammar must be complicated; in the second case, this applies to the Dutch grammar. Now in the bilingual case, this does not really matter, as the complication has to go somewhere anyway. But in a multilingual situation, each monolingual grammar will be complicated in this way by linguistic peculiarities of all the other languages. Think only of the English-Dutch translation pair `snow-sneeuw' in a multilingual system that has Eskimo as one of its languages.
In the MiMo2 system, we can express the equivalence by a transfer rule like (for the actual notation see below):
In sum, an interlingual approach suffers from potential arbitrariness, and may complicate the overall system; but the issue is relevant only if one wants to keep open the possibility of a multilingual system.