The content of language predicts the accuracy of recognition
Updated: Aug 26, 2019
A machine learner can use the language people provide to justify recognition decisions to predict judgment accuracy.
We used a type of regularized regression termed Lasso, to translate the language subjects give when they justify their recognition decisions (memory justifications), into predictions of accuracy. During training the logistic-lasso uses the frequency of individual words in the justification in an attempt to determine whether the recognition decision was a hit (coded 1) or false alarm (coded 0). The plot above shows the words that indicate the justification was during a hit (orange) or a false alarm (blue). We hypothesized that this hit/false alarm (HFA) classifier was sensitive to the presences versus absence of recollection supporting the decision. Supporting this interpretation, the HFA classifier performed as expected under this hypothesis when applied to justifications from a Remember/Know study by Gardiner and colleagues. It 'thought' justifications of remember reports were hits and justifications of know reports were false alarms; just what it should do if it is sensitive to the presence versus absence of recollection signaled by the language.
Dobbins, I. G., & Kantner, J. (2019). The language of accurate recognition memory. Cognition, 192, 103988.