Welcome to the world of Information Architecture, where everything can be classified in more than one way. The reason is that anything (or shall I write - any thing) has more than one property, or in the IA jargon - a facet. Even God can be Christian or Muslim, omnipresent or monopresent, religious or scientific.
While it is hard to give a definite answer without a better description of the project; and while in any case your main classification scheme should be based on user research rather than 'documented usability' or educated advice like the one I'm giving you here... I feel you are considering language to be a candidate for an hierarchical (or structural) class, where in fact it is nothing but a facet.
A subject and its publications are clearly linked, language is nothing but what describes a property of a publication. I can easily see year-of-publication becoming a requirement at some stage.
Thus, you should really stick to the subject/publication hierarchy and allow users to filter the list by language using a faceted filter (see example image below).
Having said that, given the actual nature of your project and its users, this may not be the ideal solution.
Shed more details on your project if this answer seems dubious to you.