Despite the introduction of grammatical relations in ATC, most methods still combined these relations into one united representation for distributional similarity computation, which worked analogously to these based on the premise of 'a bag of words'.