Free recall of semantically related words reveals similarity structure

Abstract In free recall, semantic associations between studied items lead to clustering of those items. In prior work, the impact of these associations on recall has been assessed using measures that are independent of participant data. For example, the semantic similarity of lemon and banana can be estimated using a distributional semantic model (e.g., latent semantic analysis) and these estimates can be used to derive semantic clustering scores. We show that instead of using pre-existing estimates of semantic similarity, it is possible to estimate semantic similarity from the data itself. In one experiment, participants study categorized word lists that are either presented randomly or blocked (arranged by category). Using established and novel analyses, we find that temporal and semantic associations interact, but that semantic associations exert a predictable influence on recall order. We use this insight to develop a model for estimating pairwise similarity and a semantic network from free recall data. The estimated networks show high correspondence with both the category structure and a distributional semantic model (word2vec). Compared to word2vec, our model made more accurate predictions of clustering in free recall after controlling for temporal similarity, underscoring that similarity measures from different sources reflect different aspects of semantic information. We further validate the model using a large, pre-existing dataset (PEERS) of uncategorized free recall lists. The work presents a novel methodology that has many potential applications in the study of both episodic and semantic memory.