Revisiting the Estimation of Unigram Surprisal in Computational Psycholinguistics

Data Explorer

Compare unigram surprisal estimators in GAMMs with different predictors, dependent variables (reading time measures with or without log transforms and ERPs), outlier removal, and multiple comparison approaches.

Notes:

• Baseline Comparison: Random Effects for participants and item
• For each predictor in a model, "Spillover" adds another smooth to the GAMM representing the relevant predictor at the prior unit
• Toggling "Remove Outliers" on removes data points more or less than 2.5 sd from the mean.
• p-tests are sign-flip permutation tests between the same model fit with unigram frequency from the chosen two estimators. Because the "surp." model contains only a contextual surprisal smooth, it will always be the same, regardless of the unigram surp. estimators chosen.