Status Quo of Open Class Written Language Identification

It seems that closed class statistical LID methods, such as n-gram, HMM, etc, have been well established and have rendered good performances so far. However, they all rely on a training set or some known language profiles to work.

I have been researching about open class LID, which, if successful, will rely little or none on pre-determined language profiles, but so far I gathered little relevant info by searching existent literature. Can some people give some suggestion on this? Is this a realm that is worth further researching and developing?

Thank you very much.