This is awesome! I have found this feature very useful for representing extra information that I required for multiple intent classification. One observation is that it works better(or with fewer training samples) with the DIETClassifier as opposed to CRF.
Great work, thanks for the update! 