Comment by colingauvin

I initially tried annotating everything three times, then judging the annotation with the same model. That failed predictably, so yeah just logprobs right now, judged on a hand-annotated dataset. Mostly correlated - ~80%.