[sqrt17]

It does help to have sketchy knowledge about some language to see how you can figure out things when you cannot assume any knowledge of them.

Contrary to your point, early statistical machine translation only works well for relatively close language pairs, like English-French or English-Spanish. It totally fails for more distant languages such as Chinese, Arabic or even German, which is why you have so many Chinese-speaking people (including English-Chinese bilinguals) in machine translation these days.

Parallel corpora are full of ad-hoc, hard-to-model, human judgements (from people called "translators"). The advantage is that the translators don't come up to you to criticize your translation model; however doing error analysis for an MT system (i.e., the key to actually improving things and not producing garbage) requires at least minimal knowledge of the source language and relatively good knowledge of the target language.