Abstract
Ecologically important attributes in a natural scene arise not only from isolated objects, but also from semantic associations among objects in the scene. How these associations affect scene memory is yet fully understood. Here, we investigate this issue by examining how semantic associations of scene elements affect scene memorability using online crowdsourced methods. Study 1 asked mTurk participants to label objects in 1024 scene images and provide subjective ratings of object relatedness, complexity, valence, and arousal of these scenes (~20 participants per image). We calculated semantic associations among scene elements based on Global Vectors of verbal labels. We found that this measure reliably predicted participants’ subjective ratings of scene object relatedness. In Study 2, participants studied a set of 36 images randomly sampled from the 1024 images (~55 participants per image). During a later recognition test, a test image could repeat a study image (“old”) or have an object added or removed from the image (“changed”). Participants reported whether a test image was “old” or “changed." While certain scenes were consistently recognized across participants in “old” trials, we found that scene memorability in “changed” trials varied by type of change and semantic associations among scene elements. Observers often failed to detect a change when an object related to other elements was added to the scene. In contrast, observers more successfully identified a change when an object related to other elements was removed from the scene. Low-level features (e.g., changed area), scene complexity, valence, or arousal could not explain these observations. These results suggest that pre-existing semantic associations impede retrieval by assimilating additional items into a scene but facilitate associative retrieval of missing scene elements. Collectively, our findings indicate that scene memorability is a property of pre-existing associative long-term memory, separable from contributions of lower-level perceptual features.