"To be" in semantic role labelling

I’ve been using the standard AllenNLP model for semantic role labeling, and I’ve noticed some striking behavior with respect to the verb “to be”. Even the simplest sentences, such as “The grass is green” give an empty output. This does not appear to be the case with other copular verbs, as in “The grass becomes green”. Is there a reason for this?

My guess is this is due to annotation decisions in Ontonotes, which the model was trained on. I’d try to find a description of the annotation guidelines for Ontonotes SRL to be sure.

Actually, with some more thought, this is probably due to spacy. We use spacy to find verbs in the sentence, and only make predictions when spacy tells us there’s a verb. I think spacy’s treatment of “to be” changed in recent versions, which might be causing what you’re seeing. An easy way to check is to manually tag “is” as a verb in the predictor. If that gives a prediction, then it’s spacy’s POS tagger that’s causing the issue; otherwise, it’s probably an Ontonotes labeling decision.