Discussion
The TE proved robust for word pairs with both low and high levels of semantic relatedness, and as expected based on prior work, the dual-memory model fitted well for the case of low relatedness. Nevertheless, the efficacy of testing was reduced by about 26% in the high relatedness group (based on the difference between mean TE in that group and the model’s predicted mean), a result that is most compellingly demonstrated by the smaller than predicted TE magnitude at each level of PCR in the cumulative distribution plots. Accordingly, the two alternative scenarios hypothesized at the outset of this study, in which semantic relatedness has either no effect on relative testing efficacy or amplifies it, appear to be ruled out.
Our null proportion correct ANOVA interaction between training task and semantic relatedness mirrors three prior null interaction results on this topic noted in the Introduction. We have demonstrated in the current work, however, that all of those null interactions are theoretically uninterpretable. Further, our modeling and distribution analyses clarify how a null ANOVA interaction can mask a real interaction between relatedness and testing efficacy. Specifically, even when testing is less effective for high relatedness materials, the quadratic relationship between PCR and TE magnitude can obscure this difference in a traditional ANOVA interaction. More generally, the presence or absence of an ANOVA interaction in any TE experiment is likely to be theoretically uninterpretable whenever both the restudy and test condition mean proportions correct change in the same direction across levels of other experimental factors (e.g., semantic relatedness). In that case, alternative bases for theoretical inference are required, such as the application of a process model and the matching strategy in the current work. More generally, the risk of an uninterpretable interaction is not limited to testing effect or psychology experiments. It may be present in the context of any experiment in which an interaction could occur22,23.
Rickard and Pan (2018) suggested that for highly related materials the pre-existing connections in the semantic network may lead to study and test memories that incorporate overlapping information. A plausible consequence of that overlap is positively correlated study and test memory strengths across items for each participant, rather than independent strengths as assumed in the dual-memory model for low relatedness materials. To investigate this possibility, we extended the model through simulation by incorporating a single free parameter, ρ, representing the correlation between study and test memory strengths. The simulation details and predictions of that extended model are described in the Supplementary and a visualization of the correlated memory strengths can be found in Supplementary Fig. 1. It is guaranteed that use of ρ as a free parameter in that model will provide exact fits to the mean TE for the high relatedness group because it acts as a scaling parameter for the mean PCT-predicted. However, ρ does not guarantee good distribution fits. The best fit to the mean data was obtained for ρ = 0.334. For fits of other p values, see Supplementary Fig. 2. And as shown in Fig. 5b, that value of ρ yielded a good fit across the bulk of the distribution for high relatedness items.
The correlated strengths model is a straightforward and plausible extension of the dual-memory model to the high relatedness case, requiring only a single free parameter. One plausible psychological basis of this account is that for highly related words, study and test memories are formed by partially overlapping features drawn from pre-existing semantic associations between the cue and target. This shared representational substrate yields shared feature activations across memories, so items that strongly recruit the cue–target semantic scaffold during encoding and practice will have correlated memory strengths. This view predicts that neural pattern similarity between study and test (e.g., representational similarity analyses) should covary with the model-estimated ρ, and that item-level semantic similarity (e.g., cosine similarity in word embeddings28) should likewise track ρ.
The correlated strengths account raises the possibility that the original dual-memory model, which assumes independence of the two retrieval routes, may be incorrect or incomplete for low relatedness materials. After all, low relatedness does not imply zero relatedness, so some degree of dependence between the two retrieval routes may exist. Thus the question arises: why does the independent routes model fit so well in our current and past work to low relatedness materials? The correlated-strengths account makes clear that when the correlation parameter ρ is small, predicted TE magnitudes are nearly indistinguishable from the independence case (Supplementary Fig. 2). In particular, the difference in TE magnitude across the distribution when ρ = 0 vs when ρ = 0.1 is negligible. It is plausible that for low relatedness pairs ρ is <= 0.1, given for the high relatedness group the estimated value ρ was 0.334 and that the mean semantic relatedness was more than eight times higher in the high than in the low relatedness group. As such, the good fits of the original independent routes model to low relatedness materials is not necessarily problematic for the dual route framework.
Carpenter (2009) hypothesized that elaborative retrieval—by way of mediating concepts—is more likely to occur on test trials than on restudy trials, accounting for the TE. Carpenter also speculated that mediator links are more likely to form for weakly related cue-target pairs than for strongly related pairs. When cues and targets share a strong preexisting association, retrieval can proceed through direct association, leaving little need for elaboration. In contrast, weakly related pairs require the learner to generate additional, non-target associations that can later serve as mediators to the target during final recall. Our results are consistent with that possibility. There are, however, several challenges to the elaborative retrieval model in the recent literature14,16,19,29. Nevertheless, the current findings are at least compatible with a core prediction of that model that the benefits of testing and the formation of mediating links should be strongest when cue–target relatedness is low.
The episodic context model30 proposes that test trials are more likely to evoke recall of episodic context that is active during prior study trials, and that the incorporation of that episodic context into test trial learning provides the basis for the TE. Karpicke et al. also suggested that encoding of episodic context on test trials might occur less frequently for highly related materials, because the answer is more likely to be retrievable without having to search for and activate that context. Hence, the episodic context account is potentially consistent with current results. However, by analogy to the elaborative retrieval account, it is unclear in that model what level of relatedness would be needed for that effect to occur.
For the episodic context model more generally, there is mixed evidence. Two types of evidence can support that model: (1) evidence that a test enhances context memory more than does restudy, and (2) evidence that test-enhanced context memory is the basis of the testing effect31. Evidence of the first type is mixed. A test appears to facilitate both temporal32,33 and spatial34 context memory, but there is no evidence that retrieval strengthens voice (male vs. female) context memory32. Also, Hong et al. (2019) found negative evidence that retrieval enhances color context memory. There appears to be no work in the literature that speaks directly to the second type of evidence. Further, Gupta et al. (2022) showed that repeated study for each item during the initial study phase did not moderate the relative efficacy of cued recall testing, despite the expectation that repeated study would yield stronger or more varied episodic context.
Our finding that the efficacy of testing is reduced for high relatedness materials raises the issue of its efficacy on the other extreme, for very low or zero relatedness materials. For non-linguistic materials with apparent zero relatedness35–37, a complete or near absence of the TE relative to the restudy control has been observed in three studies. The dual-memory model does not anticipate those results, though it could accommodate them if it is assumed that test memory cannot form in the complete absence of cue-response relatedness, in which case the model would predict zero TE. However, that account does not provide insight into why the formation of test memory would require at least some minimal degree cue-target semantic relatedness.
For linguistic materials, in contrast, the cued recall TE appears to hold even for the case of very low relatedness38. Most pertinent is recent work on the TE for language translation pairs. de Lima and Buratto conducted 5 experiments using Swahili – Brazilian Portuguese words pairs and found robust TEs. They also reported good fits of the dual-memory model. Abel and Roediger (2017) observed similar effects for Swahili-English translation pairs, and unpublished results from our laboratory also demonstrated good fits of the model to the data from their Experiment 1 and to the no retroactive interference group of their Experiment 2. Those results suggest that the original dual-memory model may hold without modification for linguistic materials across a range from very low relatedness to low and perhaps moderate relatedness, with the correlated strengths version of that model accommodating highly related materials. As a caveat to that conclusion, the experiments of both de Lima and Buratto and Abel and Roediger involved multiple training phase item repetitions. It is not known whether the same results and model fits will hold for translation pairs involving one training phase exposure to each item, as in the current experiment and much of the cued-recall TE literature.
Our manipulation operationalized semantic relatedness via normative cue-target forward associative strength39. Although we designed the low and high related word pairs to be broadly comparable on basic lexical properties, the present experiment cannot definitively establish that associative strength is the causal factor behind the reduced efficacy of retrieval practice in the high-relatedness condition. It remains possible that forward associative strength covaries with another item property that is the true driver of the effect. However, we are not aware of an accepted alternative candidate in prior work. Future experiments that orthogonalize forward associative strength from other lexical properties would provide a stronger causal test.
The current work constitutes the first strong experimental demonstration that the degree of semantic relatedness for linguistic materials such as word pairs moderates the learning efficacy of retrieval practice. Specifically, the relative efficacy of testing was found to be lower for high relatedness than for low relatedness materials. A modified version of the dual-memory model provided a good fit to the data distribution for the high relatedness group. Our findings may lead to refined recommendations for educational practice by clarifying when retrieval practice is most beneficial, depending on material relatedness.
CC BY 4.0. "Semantic relatedness and the efficacy of retrieval practice" by Gupta MW, Pan SC, Rickard TC, NPJ science of learning (2026). https://doi.org/10.1038/s41539-026-00416-8. Licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Peer-reviewed and open access; full text from PubMed Central (PMC13230953). Headings and layout adapted for reading; figures and tables are in the original.