Paper out in Cogsci @cogscisociety.bsky.social: escholarship.org/content/qt6n...

Predicting substitutions: Furthermore, on the task of predicting word choice, PMI (which models info-processing dependencies between the past and future sequence) surpasses backward predictability in terms of model performance, with no additional gain from including both predictors (8/9)

Predicting substitutions: Our results reveal distinct effects of forward prediction and backward planning on word choice: substitutions are characterized by high forward predicability but appear sub-optimal from the perspective of facilitating production of future context (7/9)

Predicting substitutions: We use both communicative reward (noisy semantic/phonetic distance to repair) and probabilistic measures to predict the identity of the substitution word from a set of 20K alternatives in a naturalistic contexts from the Switchboard corpus. (6/9)

Probabilistic reduction: We also find no asymmetric effects of past and future context predictability on function vs. content word durations contra prior work on reduction in English (5/9)

Probabilistic reduction: We find that PMI qualitatively replicates the inverse effect of backward predictability on duration, but a model with both backward predictability & PMI emerges as the best performing model of word duration, suggesting non-redundant information about planning. (4/9)

Still here? 👀 Okay, deeper dive: We benchmark our PMI-based measure against (decorrelated) backward predictability on (i) probabilistic reduction & (ii) within a generative paradigm that predicts substitution errors in naturalistic productions (3/9)

TL;DR we find that our proposed Pointwise Mutual Information (PMI)-based measure produces results that are comparable to backward predictability in predicting both word form and word choice in context, whilst also being cognitively interpretable as a measure of backward planning in speech (2/9)