– I had a throw-back to conferences that accept abstracts and issue certificates of attendance. But I am proud to have implemented a new approach to modelling for that talk: I used GLM with Negative Binomial family for counts of disfluencies in our data, comparing the explanatory and predictive power of corpus measures of complexity vs surprisal from off-the-shelf and domain-adapted GPT2 and MarianMT models. It feels like an achievement.