Assessing autobiographical memory consistency: Machine and human approaches.

Abstract

Memory is far from a stable representation of what we have encountered. Over time, we can forget, modify, and distort the details of our experiences. How autobiographical memory-the memories we have for our personal past-changes has important ramifications in both personal and public contexts. However, methodological challenges have hampered research in this area. Here, we introduce a standardized manual scoring procedure for systematically quantifying the consistency of narrative autobiographical memory recall and review advancements in natural language processing models that might be applied to examine changes in memory narratives. We compare the performance of manual and automated approaches on a large dataset of memories recalled at two time points placed approximately 2 months apart (N(memory pairs) = 1,026). We show that human and automated approaches are moderately correlated (r = .21-.46), though numerically human scorers provide conservative measures of consistency, while machines provide a liberal measure. We conclude by highlighting the strengths and limitations of both manual and automated approaches and recommend that human scoring be employed when the types of mnemonic details that are consistent over time and/or what drives inconsistencies in memory are of interest.