Authors: Clare Patterson, Petra B. Schumacher, Bruno Nicenboim, Johannes Hagen and Andrew Kehler
Contact Email(s): clare@reprocessed.org, petra.schumacher@uni-koeln.de, B.Nicenboim@tilburguniversity.edu, johannes.hagen@gmx.net, akehler@ucsd.edu
3 thoughts on “German pronoun use follows Bayesian principles”
Leave a Comment or Question Below
You must be logged in to post a comment.
Thank you for the excellent talk!
Some of the things I was wondering:
(i) It seems that the Bayesian Model is perhaps more accurate (for a lack of a better term) for both demonstratives and pronouns in active-accusative verbs and in experiencer-stimulus verbs. If this is true, could you say more about why that is? Follow-up: Did participants also complete sentences with a free prompt other expressions, like the thief? And if so, did they happen to use non-pronouns more often in the Dative-Experiencer and Stimulus-Experiencer conditions?
(ii) Did you compare the reduced Bayesian Model with the full Bayesian model that also includes p(pronoun)? If so, which fared better? Calculating p(pronoun) would have been a bit more complex, I imagine, if participants also completed free prompt sentences with non-pronominal expressions.
Again, very interesting stuff!
Sorry, looks like I should have posted my comment below as a reply!
Thank you for your interest and the questions! We didn’t statistically compare the performance of the Bayesian Model in accusative versus dative verbs (Exp 1) or ES versus SE (Exp 2). But it does look like the predictions are more accurate in the accusative verbs (compared to datives) and in ES (compared to SE). Looking at participant-level data for Exp 1, there is more variation in responses in dative compared to accusative, so that could have contributed to less accurate predictions. And The SE dieser responses were surprising to us – the longer, thinner violin plot shows that there was less certainty in the predictions there. We think that there is a clash of biases in this condition – a strong bias from the verb to re-mention the stimulus (NP1), and a strong bias from the pronoun (dieser) to refer to NP2, and in a follow up experiment the SE-dieser-NP1 completions were rated as much worse than SE-dieser-NP2 and SE-er-NP1 completions. So that condition is a bit strange.
To your follow up question, we did have some completions with other expressions, making up about 10% and 12% of responses in Exp 1 and Exp 2 respectively. I haven’t currently got that broken down by verb type, but I should definitely look at that, thanks!
(ii). I need to check that I understand your question correctly. Do you man p(pronoun) in the sense of combining er and dieser? (We didn’t look at that). Or do you mean P(pronoun) as the denominator of the formula P(pronoun|referent)P(referent)? The denominator we used was [P(pronoun|NP1)P(NP1) + P(pronoun|NP2)P(NP2)] which has the effect of normalising the probability. I hope that answers your question but please comment below if you want further clarification!