Social Metamemory: Metacognition & The Utility of Natural Language Memory Justifications

Observers can forecast the memorability of stimuli during an encoding, a metacognitive skill referred to as a judgment of learning (JOL). In the case of face recognition this is in part thought to reflect the ability to detect the distinctiveness of faces, combined with the correct understanding that unusual faces are more memorable (Sommer et al., 1995). Moreover, JOLs not only anticipate the future recognizability faces but, if they are based on attributes, such as distinctiveness, increasing JOLs should produce increased rates of later recollection or remembering experiences (Dewhurst et al., 2005). However, one critique of introspective memory reports is that they may merely index the general strength of recognition evidence as opposed to signifying anything specific about the nature of conscious experience of prior encounter and its relation to a current memory probe (Donaldson, 1996). Moreover, the remember/know paradigm used in examining conscious retrieval experiences provides subjects with considerable instructions on the putative experiences of remembering, knowing and sometimes, guessing (Gardiner et al., 1998) and this may lead to unwanted demand characteristics. Indeed, minor modifications of these instructions induce changes in the relative rates of remember and know reports (Umanath & Coane, 2020). To address this, researchers have taken a different approach by having subjects justify recognition and rejection decisions in their own words, typically by typing them following the recognition decision. Language classifiers are then trained to distinguish accuracy by translating these reports into a form amenable for statistical analysis (i.e., Natural Language Processing) and train statistical algorithms to distinguish accurate and inaccurate recognition conclusions (Dobbins & Kantner, 2019). Critically, the words or phrases selected by these learners, as well as the cases receiving the highest estimated probabilities of accurate recognition, indicate that conscious recollection, conveyed via natural language, forecasts accuracy. In other words, the reason the machine learners succeed, in part, is because language reliably conveys recollective experiences. Moreover, recent findings also suggest that these language classifiers are also sensitive to various uses of recollection, for example, the use of recollection to reject materials as altered or different from their original form, termed recollection-rejection or recall-to-reject (Wahlheim, Bayley, & Dobbins under review). In these cases the justifications not only contain evidence of recollection but language reflecting the detection of a changed feature or aspect of the probe that evoked recollection of the original. Returning to the nature of metacognition and face recognition, the current study adopts this approach by having observers provide JOLs during face encoding and then later providing natural language justification of three-alternative forced choice (3AFC) decisions during sequent recognition testing. Critically, the report uses the Chicago Face Database (CFD) which is a collection of high quality face photographs that are extensively normed for both subjective impressions and objective physical characteristics (Ma et al., 2015). The goal is to look at the relationship between measures of metacognition (JOLs and retrospective confidence) and the diagnostic value of language in predicting recognition accuracy and JOL resolution.