Similarity Comes at a Cost: Novel Evidence for Associative Memory Retrieval

Background. A key observation in the study of associative memory is that the greater the number of associations a concept has (its associative fan), the longer the reaction times (RTs) and the lower the accuracy demonstrated by participants during recall. This is called the fan effect ([1]). In classical fan experiments, participants study sentences containing reoccurring concepts in multiple contexts (e.g., ”The doctor is in the desert”, ”The doctor is in the field”). Models of memory in psycholinguistics assume a content-addressable associative memory, in which semantically related items cause interference in retrieval ([2]; [3]; [4]). This assumption is based on the cue-based retrieval framework from the memory module in ACT-R, which in turn was established through the findings from fan experiments. Classic fan experiments, however, did not directly test the role of semantic similarity. To address this gap, we tested whether the fan effect explains how associations occurring in natural language are retrieved by using semantically related concepts based on similarity measures from Dutch vector space models ([5]; [6]).
Method. In Experiment 1 (N = 134, 46 lab, 88 online), we replicated [7] to test the experimental paradigm in Dutch and with an adjusted learning method to make the manipulation of fan size less evident. In this experiment, concepts were repeated in two (fan 2) or four different (fan 4) contexts. Based on this replication, we conducted Experiment 2 (N = 97, online) which used semantically similar instead of repeated concepts. In both experiments, participants learned short sentences (see Table 1). In the testing phase, participants read a sentence and indicated via button press whether it belonged to the set of sentences they studied before (target) or not (foil). For each decision, we recorded the corresponding RT and accuracy.
Results. We fitted Bayesian generalized linear mixed models to analyze RTs and accuracy. In both experiments, participants exhibited longer RTs for fan 4 than for fan 2 conditions (see Figure 1, top), showing that the fan effect persists with the adjusted learning method and semantically similar instead of repeated concepts. Lower accuracy for the fan 4 conditions could only be found in Experiment 1 (see Figure 1, bottom). Accuracy in Experiment 2 was high in all experimental conditions, suggesting a potential ceiling effect.
Discussion. Our results show that the fan effect persists even when we use associated concepts occurring in natural language. These findings indicate that activation in memory spreads along a network of semantic connections as opposed to a network of individual words. This supports the interpretation of the original fan effect as a measure of associative fan, which will enable us to connect similarity measures from vector space models to theories of memory
retrieval.