The most simple case of a text collection is a set of texts in a single language. Usually, the language in these texts would be described using the Language element. However, in the situation where the text collection is intended to document a language, then it is simultaneously in and about that language. Accordingly, it would be tempting to describe this with both the Language and Subject.language elements. Rather than let the decision about metadata depend on the intent of the creator of the resource (which may not be known), or on the typical usage of the resource (which may change through time), we think it would be simplest to describe such situations using the Language element only. This approach generalizes to the case of a text collection spanning multiple languages (e.g. where each text comes from one of the Formosan languages). Here, collection-level metadata would provide a Language element for each language represented in the collection.
In the case of bilingual aligned texts (bitexts) there is normally a directionality, since one of the texts is primary and the other is a translation. Here, we specify the primary language using Subject.language, and the translation using Language.
A complication arises when the texts or bitexts employ analytical
notations. For a text in language 
transcribed in a notation
(such as the International Phonetic Alphabet) which is
inaccessible to speakers of (assuming any exist), it would
make sense to use Subject.language instead of Language.
The situation becomes more vexing for texts with embedded annotations
(such as the notation of Conversation Analysis), where non-specialists
could discard the annotations to get a conventional text. Here, the
use of both Subject.language and Language seems to be
indicated.