The natural language people use to justify memory decisions predicts recognition accuracy. Under a recollection sensitivity account this occurs because the language reflects the greater tendency of recollective experiences to accompany accurate versus erroneous recognition conclusions. To further test this, two experiments had subjects justify their recognition claims for hits and false alarms. This recognition phase was followed by a surprise free recall test in which they were asked to recall all previously justified recognition items. Both a bag-of-words (BoW) and a BERT embeddings-based classifier were trained to predict whether each justification accompanied a recognition hit or false alarm. In both experiments the language models explained recognition accuracy, and more importantly, their recognition scores reliably predicted the final recall of justified probes. This was predicted under the recollection sensitivity hypothesis because recollective experiences themselves are memorable events. Experiment 2 also incorporated numeric confidence ratings into the recognition decisions and the direct comparison of confidence versus language demonstrated that that language better predicted final recall. These findings indicate that machine learners are sensitive to recollective language in justifications that are important for the prediction of recognition accuracy and the future recall of the recognition testing experience.
Zhang et al. (Sat,) studied this question.