Key moments and topic modeling

When people remember everyday experiences, they tend to focus on a subset of moments rather than everything that happened. Recent work shows that people reliably identify key moments within narratives, and that these moments disproportionately organize event representations in the brain. However, what these key moments represent remains an important unanswered question. Here we test the hypothesis that key moments function as semantic anchors in memory for extended events. Using spoken recall from three naturalistic datasets, we derived latent semantic representations of each narrative using topic modeling. We projected time-resolved movie annotations into this semantic space and systematically deleted narrative segments of the annotation to quantify the contribution of individual segments to the overall semantic structure. Removing segments associated with key moments produced significantly greater degradation of semantic structure than removing other segments, including event boundaries. These findings reveal a mechanism by which narratives are compressed into semantically meaningful representations.