A predictive framework for evaluating models of semantic organization in free recall

• Semantic associations have an important influence on search through episodic memory.
• The mechanisms by which semantic knowledge influences memory search remain unclear.
• We propose a framework for contrasting models of search based on prediction accuracy.
• We contrast item- and context-based models of semantic organization in free recall.
• Semantic influences are short-lived during search, consistent with item-based cuing.

Research in free recall has demonstrated that semantic associations reliably influence the organization of search through episodic memory. However, the specific structure of these associations and the mechanisms by which they influence memory search remain unclear. We introduce a likelihood-based model-comparison technique, which embeds a model of semantic structure within the context maintenance and retrieval (CMR) model of human memory search. Within this framework, model variants are evaluated in terms of their ability to predict the specific sequence in which items are recalled. We compare three models of semantic structure, latent semantic analysis (LSA), global vectors (GloVe), and word association spaces (WAS), and find that models using WAS have the greatest predictive power. Furthermore, we find evidence that semantic and temporal organization is driven by distinct item and context cues, rather than a single context cue. This finding provides important constraint for theories of memory search.