Lion—Tiger—Stripes: Delimiting the Semantic Association Effect on Working Memory With Mediated Association

Abstract

Semantic relatedness of items improves working memory performance. We targeted one type of semantic relatedness, semantic association. The beneficial effect of association is often explained by the spreading-activation process: Encoding/maintaining an item would activate its associated item. Nevertheless, as associated words can be similar to each other (e.g. similarity based on the Latent Semantic Analysis), other processes, rather than spreading activation, may also explain the association effect. To target spreading activation selectively, a novel hypothesis unique to this process was tested. Specifically, we tested the effect of mediated or two-step association, which the spreading-activation theory assumes (e.g. lion → tiger → stripes). To examine the mediated association effect, Experiments 1, 2A, and 2B presented word pairs with indirect association (e.g. “lion” and “stripes”) without mediators (e.g. “tiger”). Only one of the three experiments provided moderate evidence for a beneficial effect of mediated association. Additionally, cross-experiment and item-level analyses did not support the mediated association effect. By contrast, Experiments 3 and 4 presented word pairs with direct association (e.g. “tiger” and “stripes”) and demonstrated extreme evidence for a beneficial effect of direct association. The negligible effect of mediated association would aid in delimiting the scope of association’s influence on working memory.

Introduction

Working memory (WM) refers to a memory system or function that allows us to hold information in mind when we perform a task (Baddeley, 1992; Ishiguro et al., 2024; Jarrold & Towse, 2006; Luck & Vogel, 2013). This system is affected by lexical-semantic factors such as lexicality (Saint-Aubin & Poirier, 2000), frequency (Hulme et al., 1997), imageability/concreteness (Walker & Hulme, 1999), and semantic relatedness/similarity (Saint-Aubin & Poirier, 1999a). Attempts to explore the effect of semantic relatedness can be traced back to classic studies (Baddeley, 1966; Crowder, 1979) and are still ongoing in modern research (Guitard et al., 2025; Kowialiewski et al., 2024). Recent work by Neath et al. (2022) demonstrated that three types of semantic relatedness – semantic association, category membership, and synonymity – consistently improved immediate serial recall.

The present study targets one type of semantic relatedness, namely, semantic association. It is a pre-experimental association between items such as “coffee” and “morning,” which reflects stored knowledge. Semantic association can be quantified by the word association norms (De Deyne et al., 2019; Nelson et al., 2004). In the data collection for the norms, participants see a cue word and freely answer words that come to mind. For instance, in De Deyne et al. (2019), 15 out of 297 participants answered “morning” when they saw the cue “coffee” (i.e. the probability of 0.051). By contrast, unrelated non-associated words were hardly produced; “galaxy” was, for example, not answered to “coffee” at all. The probability of answering B given A is referred to as the associative strength between A and B (e.g. 0.051 for the “coffee-morning” pair; 0.000 for the “coffee-galaxy” pair).

It has been demonstrated that immediate serial recall performance increases when items are associated with each other (Neath et al., 2022; Saint-Aubin et al., 2014; Tse, 2009; Tse et al., 2011). This effect of semantic association is typically explained based on the spreading-activation process: Encoding/maintaining an item activates associated items via links in the semantic network (the spreading-activation account). Encoding “coffee” may pre-activate “morning” prior to the presentation of “morning” (Kowialiewski, Krasnoff, et al., 2022; Kowialiewski & Majerus, 2020; Tse, 2009). In addition, maintaining “coffee” would functionally work as encoding of that item again, and thus it may re-activate “morning” after the presentations of these words (for the refreshing process, see Kowialiewski et al., 2021).

Crucially, an explanation based on spreading activation is recognized in some studies on the semantic relatedness effect, in which items drawn from a category are used as related items (Kowialiewski & Majerus, 2020; Saint-Aubin et al., 2005). This is because items belonging to a category, such as “dog” and “cat”, are assumed to have a semantic association. In fact, “dog” appears to be associated with “cat” because its associative strength, 0.176, is not zero (for the norms, see De Deyne et al., 2019). In addition, dimension-based (or feature-based) similarity seems to be hard to separate from association. Referring to a dimension-based data, a Latent Semantic Analysis (LSA) data (Günther et al., 2015), the cosine similarity between “coffee” and “morning” (0.31) is indeed larger than that between “coffee” and “galaxy” (0.02). These examples highlight the potential confounding of association, category membership, and dimension-based similarity.

To illustrate the relation between all these semantic indices, we took Tse et al. (2011) as an example study on the semantic relatedness effect and reanalyzed the stimuli of their associative, category, and control lists. First, we targeted word pairs from the associative and control lists and checked the distributions of their LSA similarity values (the left plot of Figure 1). Word pairs from the associative lists tend to be more similar to each other in terms of LSA (M = 0.26; Mdn = 0.21) than those from the control lists (M = 0.01; Mdn = −0.00), W = 108,982, p < .001. Second, associative strengths of word pairs from the category and control lists were analyzed (the right plot of Figure 1). Word pairs from the category lists are likely to have larger associative strengths (M = 0.014; Mdn = 0.002) than those from the control lists (M = 0.001; Mdn = 0.000), W = 78,286, p < .001. Therefore, associated words can be similar to each other, and category words can be associated with each other.

Figure 1.

The relations of association to category membership and dimension-based similarity. We analyzed words used in Tse et al. (2011). (a) The distributions of LSA similarity values of the pairs from the associative and control lists. (b) The distribution of associative strengths of the pairs from the category and control lists. Tse et al. (2011) had four sets of 24 six-word lists (the sets of associative, category, control for associative, and control for category). The values of the LSA similarity and associative strength are based on word pairs, and thus we targeted 15 pairs for each list (6C2 = 15), resulting in 360 pairs for each set (24 × 15 = 360). As Tse and colleagues randomized the presentation order of words, we used the mean values of associative strengths in both directions as pairs’ associative strengths (e.g. for the “emerald-diamond” pair, emerald → diamond and diamond → emerald). For details, see our OSF project page (https://osf.io/gh9v6/).

Although Figure 1 clearly illustrates the semantic properties of a typical study on the semantic relatedness effect, a comprehensive overview is also desirable. Thus, we analyzed about 10,000 words that were available in both LSA and association databases (De Deyne et al., 2019; Günther et al., 2015). Figure 2 depicts a scatter plot of three groups: Association, LSA, and Random groups (each contained about 10,000 word pairs). For the Association group, we paired the words with their most associated words based on the associative strength. For the LSA group, words were paired with their most similar words in terms of LSA similarity. To get a baseline, words were paired with randomly selected words for the Random group. As the top boxplot shows, the Association group exhibited higher LSA values (M = 0.22; Mdn = 0.16) than the Random group (M = 0.00; Mdn = 0.00), W = 106,037,432, p < .001. The right boxplots indicate that the LSA group included associated word pairs (M = 0.017; Mdn = 0.000) more than the Random group (M = 0.000; Mdn = 0.000), W = 91,693,747, p < .001, although the median values were equivalent in both groups. The interconnectedness of association and similarity is not confined to a specific study but can be observed in corpora.

Figure 2.

The relation of association to dimension-based similarity in Corpora.

The different types of semantic relatedness have not been clearly separated in the literature (Ishiguro & Saito, 2021, 2024), probably due to their interconnectedness as exemplified by our demonstrations (Figures 1 and 2). Nevertheless, processes brought about by each type of semantic relatedness may not be identical. Category membership would offer an extra cue (e.g. “animal” for animal words), which is different from associations between items (e.g. interitem association between “dog” and “cat”). The retrieval process based on an extra cue or cue-dependent retrieval is thought to play a role for category items (Ishiguro et al., 2025; Saint-Aubin et al., 2005; but see also Tse, 2009). Dimension-based similarity may promote efficient information processing or compression by exploiting redundancy in dimensional representations (Kowialiewski, Lemaire, et al., 2022). Relatedly, summation of elements of dimensional representations would favour similar items (Guitard et al., 2025). Supposing the summation of elements, representations of dissimilar items would work as noise to each other (e.g. sum of orthogonal vectors), whereas those of similar items would preserve themselves. These processes appear to be different from spreading activation.

As for the separation of the processes, a recent study targeting the cue-dependent retrieval process (Ishiguro et al., 2025) provided indirect evidence for spreading activation arising from interitem association. Ishiguro et al. (2025) presented lists of ad hoc category members (e.g. “cat, ball, bone, person, fly, car”) with or without ad hoc category labels (e.g. “things that dogs chase”). The ad hoc category membership was hard to notice (Barsalou, 1983), and thus, the availability of retrieval cues was manipulated by the label presentation. Critically, the interitem association was controlled for because identical lists were used regardless of the label presentation. Their results demonstrated that the presentation of ad hoc category labels led to better item memory, which is empirical evidence for the retrieval-cue process. Nevertheless, their observed effect was smaller than the standard semantic relatedness effect (Kowialiewski & Majerus, 2020). Controlling for the interitem association would lead to a small effect, and thus, association seems to be a factor contributing to the semantic relatedness effect.

To focus on one of the processes relevant to the semantic relatedness effect, it would be desirable to examine an effect unique to the targeted process (for cue-dependent retrieval, see Ishiguro et al., 2025). The current study investigates the mediated association effect as a potential effect unique to spreading activation.

The Spreading Activation Theory and Mediated Association

The spreading-activation theory assumes that activation of an item (or a node in the model) would elevate activations of other associated items (Collins & Loftus, 1975). As the term “spreading” alludes to, “the spread of activation constantly expands, first to all the nodes linked to the first node, then to all the nodes linked to each of these nodes, and so on” (Collins & Loftus, 1975, p. 408). For example, “lion” would activate “tiger,” which would in turn activate “stripes,” even though “lion” is not directly associated with “stripes” (i.e. lion → tiger → stripes; see Figure 3).

Figure 3.

Schematic illustration of two-step spreading activation via mediated association.

Such an influence of mediated or two-step association is posited in the spreading-activation theory (Collins & Loftus, 1975; Nelson et al., 1998). Empirical evidence of the mediated association’s influence has been provided by the mediated priming effect, in which the processing of an item (e.g. “lion”) facilitates the processing of an indirectly associated item (e.g. “stripes”) in priming tasks (Balota & Lorch, 1986; Jones, 2010, 2012; McNamara & Altarriba, 1988; Sass et al., 2009; Shelton & Martin, 1992). For instance, Jones (2012) systematically manipulated Stimulus Onset Asynchronies (SOAs) between the prime and target (i.e. 200-, 1,000-, and 1,800-ms SOAs) and demonstrated priming facilitations in lexical decision with all SOAs (the effect on reaction time ranged from 14 ms to 46 ms).

The mediated association effect appears to be unique to the process responsible for the semantic association effect (i.e. spreading activation), and thus is worth examining. Moreover, the relationship between the first and third items, such as “lion” and “stripes,” may be free from other semantic relations, given that (a) these items do not necessarily belong to the same category (i.e. category membership) and (b) they are not always similar to each other (i.e. dimension-based similarity).

Overview of the Current Study

It has been established that semantic association boosts WM performance with studies using direct association between items (Neath et al., 2022; Saint-Aubin et al., 2014; Tse, 2009; Tse et al., 2011). Nonetheless, direct association is likely to be suboptimal for the test of spreading activation, given that directly associated words would be related in terms of other semantic properties. For instance, as evidenced in our demonstrations (Figures 1 and 2), associated words tend to be similar in terms of LSA, which would facilitate similarity-based processes. The current study aimed at extending the findings of direct association to indirect mediated association, thereby testing the spreading-activation account for the association effect.

We report the results of five experiments, item-level analysis, and cross-experiment comparison. The to-be-remembered items were based on triplets of associated words (e.g. “heavy, light, feather”). To examine the mediated association effect, Experiments 1, 2A, and 2B contrasted six-word lists consisting of the first and third words of the triplets (e.g. “heavy” and “feather”) and six-word lists consisting of an unrelated word with the third word of the triplet (e.g. “war” and “feather”) (see Figure 4). The second (mediating) words of the triplets were never presented in Experiments 1, 2A, and 2B. By contrast, Experiments 3 and 4 presented the second words of the triplets as well as the third words to replicate the standard direct association effect (e.g. “light” and “feather”). To preview the results, Experiments 1–4 indicated a substantial effect of direct association and an elusive or negligible effect of mediated association.

Figure 4.

Schematic illustration of triplets and lists in the current study.

The five experiments were preregistered separately on the Open Science Framework (OSF): Experiment 1 at https://osf.io/w7658, Experiment 2A at https://osf.io/zfx79, Experiment 2B at https://osf.io/7ngb4, Experiment 3 at https://osf.io/hpjqk, and Experiment 4 at https://osf.io/zn6yk. The materials and analysis scripts of all experiments are also available on the current project’s OSF page https://osf.io/gh9v6/.

Experiment 1

To test the mediated association effect on WM, Experiment 1 presented words with indirect mediated association (e.g. “lion, stripes”) in the absence of mediating words (e.g. “tiger”) in the experimental condition. Supposing the spreading-activation process based on mediated association, lists of words with indirect association would lead to better memory performance compared to the control lists.

Method

Hypothesis, Measures, and Data Collection Plan

It was hypothesized that item memory should be higher for the experimental lists with indirect association than for the control lists without the indirect association in a WM task. Item memory was targeted based on the distinction between item versus order memory (i.e. memory of item identity versus item position) (Saint-Aubin & Poirier, 1999b). Specifically, we reasoned that, if mediated association has an effect, its locus would be in item memory. This is because the direct association effect on item memory is robust while its effect on order memory is mixed and awaits further investigation (Neath et al., 2022; Saint-Aubin et al., 2014; Tse, 2009; Tse et al., 2011). Order memory was subject to an exploratory analysis.

As a measure of item memory, we used item correct (also known as the free recall criterion), in which correctly recalled items were scored regardless of their recalled positions. As a measure of order memory in the exploratory analysis, we used conditionalized order recall, which refers to the proportion of correctly recalled items at correct positions to correctly recalled items regardless of recalled positions. Additionally, correct-in-position (also known as the strict serial recall criterion) was used in the exploratory analysis. This scores correctly recalled items at correct positions and is thought to reflect both item and order memory (Saint-Aubin & Poirier, 1999b).

To test the hypothesis, we used a sequential testing approach with Bayesian statistical tests (Schönbrodt et al., 2017). First, 60 participants were recruited. We planned to stop data collection after testing this initial sample if the Bayes Factor (BF) was greater than 6 for null or alternative. Otherwise, we were to recruit an additional sample of 60 participants and stop data collection (i.e. the maximum sample size was 120).

BFs were calculated with the R package, BayesFactor (version 0.9.12-4.7; Morey et al., 2025). The prior of its default setting, the Jeffreys-Zellner-Siow (JZS) prior, was used. To conduct Bayesian ANOVA, we used the anovaBF function with a top-down approach (i.e. the comparison of the full model and the full model dropping a target term). To conduct Bayesian t-tests, we used the ttestBF function.

BFs were interpreted according to a classification scheme (Schönbrodt & Wagenmakers, 2018): 1 to 3 (anecdotal evidence); 3 to 10 (moderate evidence); 10 to 30 (strong evidence); 30 to 100 (very strong evidence); and >100 (extreme evidence). This classification scheme was used in all experiments (Experiments 1–4) and the cross-experiment comparison. We used two notations: BF₁₀ indicates the BF for the alternative hypothesis, while BF₀₁ (1/BF₁₀) indicates the BF for the null hypothesis.

Participants

In total, 60 participants were recruited online via Prolific. Recruitment criteria were as follow: (a) They had an approval rating of at least 90% on prior submissions at Prolific; (b) their first language was English; (c) they had no language-related disorders; (d) they were English monolinguals; (e) they had normal or corrected-to-normal vision; (f) they had no cognitive impairment or dementia; (g) they were born in UK, US, or Canada; (h) their ages were between 18 and 30 years; and (i) their nationality was UK, US, or Canada.

One participant did not follow the instructions and answered nonwords as their responses in most of the trials (their overall item correct was 0.181). We excluded that participant’s data and targeted the remaining 59 participants’ data (age: M = 25.88 years, SD = 3.14; gender: 39 women, 19 men, 1 “prefer not to say”).

Participants were compensated £1.50 for their participation. The experiment was completed within approximately 5 min (M: 6.02 min; Mdn: 5 min; range 3–14 min).

Materials

Semantic Databases

We checked the semantic properties of word pairs based on existing databases. LSA values and associative strengths were retrieved from https://sites.google.com/site/fritzgntr/software-resources/semantic_spaces (the Touchstone Applied Science Associates [TASA] corpus) and https://smallworldofwords.org/en/project (see also De Deyne et al., 2019; Günther et al., 2015).

Mediated Association Values

Mediated association values were achieved by multiplying direct association values, that is, the associative strength between the first and second words (e.g. for “lion-tiger,” 0.113) and that between the second and third words (e.g. for “tiger-stripes,” 0.137) (e.g. 0.113 × 0.137 = 0.015). There can be more than 1 s word between the first and third words. For instance, “lion” has mediated association with “stripes” via another second word, “zebra” (0.003 × 0.187 = 0.001). Therefore, the sum of all mediated association values was calculated for each word pair. The mediated association value for “lion-stripes” was, for example, calculated by summing the mediated association values via “tiger” and “zebra” (0.015 + 0.001).

List Construction

We selected 36 triplets of words (e.g. “lion-tiger-stripes”) from the 48 triplets used in a priming study (Balota & Lorch, 1986). Balota and Lorch demonstrated the mediated priming effect by using pairs of the first and third words of the triplets (e.g. “lion-stripes”). Note that Balota and Lorch (1986) showed mediated priming in naming, but not in lexical decision. They pointed out that their lists included pairs of directly associated words (e.g. “tiger-stripes”), and which may have made participants develop a checking strategy (i.e. when an obvious association between the prime and target was detected, the target was judged as a word). To address this issue, McNamara and Altarriba (1988) used Balota and Lorch’s materials but excluded directly associated word pairs. Their results indicated mediated priming in lexical decision. Therefore, mediated priming was observed in naming (Balota and Lorch, 1986) and lexical decision (McNamara & Altarriba, 1988) with the indirectly associated word pairs of Balota and Lorch.

Some word pairs in Balota and Lorch (1986) had relatively large LSA values, such as “bat-bounce” (0.56) and “pants-collar” (0.41). To minimize the effect of LSA similarity, we targeted 36 pairs with the 36 lowest LSA values. Then, the 36 pairs were split into two sets (Experimental Sets A and B). The pairs were first sorted based on their mediated association values, and the odd-numbered ones and even-numbered ones were assigned to the Experimental Set A and Experimental Set B, respectively. In the experiment, we used three pairs as an experimental list. As illustrated in Figure 4, an experimental list was “heavy, feather, gas, slick, school, stop,” which is based on three triplets: “heavy-light-feather,” “gas-oil-slick,” and “school-bus-stop.” The allocation of the three pairs to a list was fixed across participants, but the presentation order of the pairs in a list was randomized by each participant (e.g. “gas, slick, school, stop, heavy, feather”) (see also the “Procedure” section).

Word pairs for control lists were created by recombining the first words of original pairs with the first words of other original pairs. Re-combination of words was run by each Experimental Set, resulting in the Control Set A (based on the words of the Experimental Set A) and the Control Set B (based on the words of the Experimental Set B). We conducted 20,000 runs of re-combinations and selected word pairs for the Control Sets. The selected word pairs met two criteria: (a) they should have LSA values equivalent to the original word pairs of the Experimental Sets, and (b) they should have low mediated association values.

Semantic Properties

We assessed the semantic properties of our materials with Bayesian t-tests. ¹ As expected, the pairs of Experimental Set A had larger mediated association values (M = 0.007, SD = 0.011) than those of Control Set A (M = 0.000, SD = 0.000), BF₁₀ = 5.003. By contrast, two types of pairs hardly differed in the direct association value (Experimental Set A: M = 0.000, SD = 0.001; Control Set A: M = 0.000, SD = 0.000), BF₀₁ = 2.658, or in the LSA value (Experimental Set A: M = 0.02, SD = 0.06; Control Set A: M = 0.02, SD = 0.06), BF₀₁ = 3.095. As for the comparison of the Experimental and Control Sets B, they indeed differed in the mediated association value (Experimental Set B: M = 0.006, SD = 0.007; Control Set B: M = 0.000, SD = 0.000), BF₁₀ = 16.362. However, they barely differed in the direct association value (Experimental Set A: M = 0.000, SD = 0.001; Control Set A: M = 0.000, SD = 0.001), BF₀₁ = 2.690, or in the LSA value (Experimental Set A: M = 0.04, SD = 0.04; Control Set A: M = 0.03, SD = 0.06), BF₀₁ = 3.050. Therefore, we guaranteed the difference in mediated association values while controlling for direct association and LSA values. Note that the direct association values of the pairs were zero or close to zero in all sets, suggesting that direct association was minimized.

Task and Procedure

We used a Web-based immediate serial recall task. It had six experimental and six control lists. The presentation order of these 12 test lists was randomized for each participant (i.e. mixed presentation).

Each trial presented a six-word list. It began with a prompt, “PRESS THE SPACE BAR TO START THE TRIAL.” Upon the key press, the presentation of six words started. The words were sequentially presented one per second without an interstimulus interval. Immediately after the offset of the last word, serial recall was asked about. Participants recalled and typed each word in the presentation order of the words.

The experimental lists were based on an Experimental Set, and the control lists were based on a Control Set. For counterbalancing, participants were allocated to either a combination of the Experimental Set A and Control Set B, or of the Experimental Set B and Control Set A. The combination of the three pairs for a list was fixed across participants. The presentation order of pairs in a list was randomized by each participant, while the order of words in a pair was fixed, such that “feather” was preceded by “heavy” in the experimental list or preceded by “war” in the control list.

Analysis Plan

Experiment 1 had a 2 × 6 factorial design with the List Type factor (experimental vs. control lists; within-participants) and the Position factor (1–6 presentation positions; within-participants). We used Bayesian ANOVA.

Results

Item Correct

The main effect of Position was evident, BF₁₀ = 2.127 × 10⁵⁵ (Figure 5). However, the main effect of List Type or an interaction was not confirmed, BF₀₁ = 12.271 and BF₀₁ = 111.998, respectively. BF₀₁ (12.271) is taken as strong evidence against the mediated association effect. In fact, accuracy on the experimental lists (M = 0.763, SD = 0.160) was nearly identical to that on the control lists (M = 0.762, SD = 0.161).

Figure 5.

Results of Experiment 1.

Correct-in-Position and Conditionalized Order Recall

As an exploratory analysis, we also report the results with correct-in-position and conditionalized order recall scores. The patterns of correct-in-position were similar to those of item correct: The presence of the main effect of Position, BF₁₀ = 1.065 × 10⁸⁶ and absence of the main effect of List Type, BF₀₁ = 10.940 and of an interaction, BF₀₁ = 222.772. Correct-in-position scores were equivalent for the experimental lists (M = 0.667, SD = 0.196) and the control lists (M = 0.661, SD = 0.207). As for conditionalized order recall scores, again, the experimental lists (M = 0.849, SD = 0.128) and the control lists (M = 0.847, SD = 0.144) showed equivalent scores, BF₀₁ = 6.986.

Discussion

Contrary to our hypothesis, and to previous findings with psycholinguistic tasks (Balota & Lorch, 1986; McNamara & Altarriba, 1988), Experiment 1 did not show the mediated association effect on item correct. The Bayesian analyses provided strong evidence against the mediated association effect. The exploratory analysis with two other measures (correct-in-position and conditionalized order recall) also showed evidence against an effect of mediated associations. Nevertheless, Experiment 1 may have used a weak manipulation in terms of the mediated association value (M = 0.007 and 0.006 for the two Experimental Sets) and the number of lists (six experimental lists for each participant). These issues were addressed in Experiment 2.

Experiment 2

Experiment 2 was a follow-up of Experiment 1. The general procedure was identical to Experiment 1 except for the used lists. We created new materials rather than the materials of Balota and Lorch (1986). We report the results of two experiments. Experiments 2A and 2B presented 11 experimental lists for each participant. The new materials showed large mediated association values compared to those of Experiment 1 (Experiment 2A: M = 0.036 and 0.035 for the two Experimental Sets; Experiment 2B: M = 0.027 for both Experimental Sets).

Method

Hypothesis, Measures, and Data Collection Plan

The hypothesis of Experiment 1 was tested again. The item was correct, relevant to the hypothesis, while correct-in-position and conditionalized order recall were used in the exploratory analysis. The previous data collection plan was used, with a modification on the upper limit of sampling in Experiment 2B. Experiment 2A set N = 120 as the upper limit same as in Experiment 1. Experiment 2B, however, increased it to 180. This is because mediated association values were smaller in Experiment 2B than in Experiment 2A.

Participants

We recruited 120 participants in Experiment 2A (age: M = 25.25 years, SD = 3.16; gender: 66 women, 50 men, 4 other) and 60 participants in Experiment 2B (age: M = 26.15 years, SD = 3.28; gender: 35 women, 25 men). The recruitment criteria of Experiment 1 as well as a new criterion of “not having participated in the previous experiment” were applied. Participants were compensated £2.25. The experiment was completed within approximately 10 min (Experiment 2A: M = 13.47 min, Mdn = 11 min, range = 5–39 min; Experiment 2B: M = 11.65 min, Mdn = 10 min, range = 6–34 min).

Materials

For the Experimental Sets, we first prepared word pairs with zero direct associations and low LSA similarity values (<0.02) as candidates, by using the association and LSA datasets (De Deyne et al., 2019; Günther et al., 2015). Then, we manually excluded associatively and/or taxonomically related pairs (e.g. “gulp–bird” or “leopard–dog”) and selected the top 132 pairs with the largest mediated association values. We allocated the top 66 pairs to Experiment 2A and the remaining 66 pairs to Experiment 2B. For the Control Sets, pairs were created by recombining pairs from the Experiment Sets. In both experiments, all pairs of Experimental and Control Sets had zero direct associations.

Semantic Properties

In both experiments, the Experimental Sets had stronger mediated associations than the Control Sets while controlling for LSA similarity. In Experiment 2A, the Experimental Set A had larger mediated association values (M = 0.036, SD = 0.009) than the Control Set A (M = 0.000, SD = 0.000), BF₁₀ = 8.288 × 10²⁷. The Experimental Set B also showed larger mediated associations (M = 0.035, SD = 0.009) relative to the Control Set B (M = 0.000, SD = 0.000), BF₁₀ = 7.148 × 10²⁸. By contrast, they were well matched on LSA similarity: The Experimental Set A (M = −0.026, SD = 0.029) and the Control Set A (M = −0.023, SD = 0.043), BF₀₁ = 3.801; the Experimental Set B (M = −0.029, SD = 0.040) and the Control Set B (M = −0.019, SD = 0.053), BF₀₁ = 2.905.

Experiment 2B showed similar patterns. The mediated association values were higher for the Experimental Set A (M = 0.027, SD = 0.007) than for the Control Set A (M = 0.000, SD = 0.001), BF₁₀ = 2.136 × 10²⁹. They were again higher for the Experimental Set B (M = 0.027, SD = 0.006) than for the Control Set B (M = 0.000, SD = 0.000), BF₁₀ = 5.063 × 10³⁰. In contrast, the LSA similarity was equivalent between the Experimental Set A (M = −0.022, SD = 0.036) and the Control Set A (M = −0.021, SD = 0.042), BF₀₁ = 3.940 and between the Experimental Set B (M = −0.034, SD = 0.034) and the Control Set B (M = −0.024, SD = 0.038), BF₀₁ = 2.232.

Task, Procedure, and Analysis Plan

The task and procedure were identical to those of Experiment 1 except for the lists. Following the list construction of Experiment 1, lists of three pairs were prepared for each set of 33 pairs. Therefore, in total, 22 experimental and 22 control lists were prepared in each experiment. For counterbalancing, 11 experimental and 11 control lists were presented for each participant. We planned to use Bayesian ANOVA with the List Type factor (experimental vs. control lists; within-participants) and the Position factor (1–6 presentation positions; within-participants) same as in Experiment 1.

In Experiment 2A, due to technical errors, the system was unable to record the last character of the long word “experienced.” We corrected participants’ responses “experience” to “experienced” and conducted the analyses. This correction did not alter the statistical pattern.

Results

Item Correct

For the main effect of List Type, item correct scores were numerically higher for the experimental lists (M = 0.667, SD = 0.167) than for the control lists (M = 0.643, SD = 0.157) in Experiment 2A (Figure 6). Nevertheless, BF remains anecdotal evidence, BF₁₀ = 2.767. The main effect of Position was confirmed, BF₁₀ = 1.985 × 10¹²⁵ while it did not interact with List Type, BF₀₁ = 431.592. In Experiment 2B, item correct was equivalent for the experimental lists (M = 0.663, SD = 0.155) and the control lists (M = 0.654, SD = 0.173), BF₀₁ = 9.014. Position had an effect, BF₁₀ = 5.090 × 10⁵⁵ but it did not interact with List Type, BF₀₁ = 220.335.

Figure 6.

Results of Experiment 2.

Correct-in-Position and Conditionalized Order Recall

For correct-in-position, in Experiment 2A, the main effects of List Type, BF₁₀ = 8,845.450, and Position, BF₁₀ = 4.500 × 10²¹⁵, were confirmed. The experimental lists exhibited better correct-in-position performance (M = 0.572, SD = 0.214) relative to the control lists (M = 0.532, SD = 0.200). Their interaction was not present, BF₀₁ = 322.429. In Experiment 2B, the main effect of Position, BF₁₀ = 1.435 × 10⁹⁰, was evident while neither the main effect of List Type, BF₀₁ = 5.197, or an interaction, BF₀₁ = 143.072, was supported. The experimental lists (M = 0.569, SD = 0.199) and the control lists (M = 0.554, SD = 0.217) yielded similar performance.

Experiment 2A provided anecdotal evidence for better conditionalized order recall of the experimental lists (M = 0.818, SD = 0.160) over the control lists (M = 0.792, SD = 0.176), BF₁₀ = 1.283. Experiment 2B, by contrast, indicates equivalent performance for the experimental lists (M = 0.816, SD = 0.145) and the control lists (M = 0.809, SD = 0.178), BF₀₁ = 6.387.

Additional Analysis

Experiment 2A suggested an advantage of the experimental lists in item correct or the mediated association effect. Nonetheless, evidence was anecdotal (BF₁₀ = 2.767). Experiment 2B, in fact, demonstrated moderate evidence against the effect (BF₀₁ = 9.014). To obtain clear evidence, we increased the sample sizes by recruiting additional samples of 60 participants for Experiment 2A (age: M = 25.43 years, SD = 3.41; gender: 32 women, 27 men, 1 other) and Experiment 2B (age: M = 25.43 years, SD = 3.14; gender: 23 women, 36 men, 1 other). Note that adding data is not problematic in Bayesian inference (cf. p-hacking). In fact, “the Bayes factor is not affected by the sampling plan” (Wagenmakers et al., 2018, p. 47). We, therefore, added data in a post-hoc way. With 180 participants’ data, Experiment 2A provided moderate evidence for the mediated association effect on item correct (BF₁₀ = 3.312). With 120 participants’ data, Experiment 2B showed moderate evidence against the effect (BF₀₁ = 4.301).

We also conducted a post-hoc analysis with subsets of the item correct data. During the experiment, participants may have noticed the mediated associations, which would have affected the results. Thus, we divided the data into a subset of the first five experimental and control lists and another subset of the last five experimental and control lists. Experiment 2A (N = 180) provided merely anecdotal evidence for the effect in the first part: for the first subset, BF₁₀ = 1.085; for the last subset, BF₀₁ = 4.678. Experiment 2B (N = 120), again, provided at best anecdotal evidence for the effect: for the first subset, BF₁₀ = 1.077, for the last subset, BF₀₁ = 14.939.

Therefore, Experiment 2A among the three experiments, only when adding the data, provided moderate evidence for the effect. We further examined the data at an item level as an additional post-hoc analysis.

Item Level Analysis: Experiments 1, 2A, and 2B

We examined whether the strength of the mediated association of an item was related to the recall of that item. Specifically, we calculated item correct scores of the third words of triplets (averaged across participants) and examined whether these scores were explained by mediated association values.

Method

In total, 168 unique pairs were used in the three experiments: 36 pairs, 66 pairs, and 66 pairs for Experiments 1, 2A, and 2B, respectively. Item correct was calculated for each word in each experiment using available data (Experiment 1, N = 59; Experiment 2A, N = 180; Experiment 2B, N = 120). Although some words were used across experiments, there were no overlaps of pairs among the three experiments. For example, “pie” was used in all experiments, but it followed different words (“birthday–pie,” “fragment–pie,” “component–pie”). Mediated association values were specified depending on the preceding word. As lexical-semantic properties of a target word (e.g. “pie”) would affect its recall, we also entered these values as well as a random effect of word.

A linear mixed-effects model was used. The dependent variable was the mean item correct score of each word. The independent variables were the mediated association, number of letters, number of syllables, word frequency, and concreteness (Balota et al., 2007). Random effects were random intercepts of word, list, and position. The random effects of list and position were added to incorporate the influences of other words and presented position of target words (e.g. “pie” was presented with the same other five words, and it was presented at position 2, 4, or 6).

Results and Discussion

Word frequency, t(119.381) = 2.270, p = .025, and concreteness, t(112.856) = 2.764, p = .007, affected item correct scores. By contrast, the number of letters, t(116.168) = −0.974, p = .332, or the number of syllables, t(126.558) = 0.090, p = .929, did not affect the scores. Dropping either the number of letters or the number of syllables (word length in the orthographic or phonological form) did not change the pattern. More importantly, mediated association did not affect item correct scores, t(108.442) = −0.708, p = .481.

To sum, only one of the three experiments demonstrated the mediated association effect in the comparison between the experimental and control lists. A further attempt to find the effect, an item-level analysis, failed to support the mediated association effect. These results suggest that the effect is elusive. That said, the results might reflect an experimental failure. To address the possibility of an experimental failure, Experiments 3 and 4 aimed at replicating the standard direct association effect.

Experiment 3

Experiment 3 presented the second (mediating) words along with the first and third words of the triplets for the experimental lists (see Figure 4). Given that there were direct associations between the first and second words and between the second and third words, a beneficial effect of association was expected in line with previous studies using direct association (Saint-Aubin et al., 2014).

Method

Hypothesis, Measures, and Data Collection Plan

It was hypothesized that item memory should be higher for the experimental lists (with mediating words) than for the control lists (without mediating words). Item correct was used for hypothesis testing, while correct-in-position and conditionalized order recall were used in the exploratory analysis same as in Experiment 1. The data collection plan followed a two-step sampling procedure of Experiment 1. As BF was greater than the criterion (BF >6) with the sample of 60 participants, we stopped the data collection after the initial recruitment in the current experiment.

Participants

We recruited 60 participants via Prolific (age: M = 25.22 years, SD = 3.36; gender: 37 women, 22 men, 1 “prefer not to say”). The recruitment criteria of Experiment 2 was applied. Participants were compensated £1.80 for their participation. The experiment was completed within approximately 5 min (M: 7.43 min; Mdn: 7 min; range 4–17 min).

Materials

List Construction

The 36 triplets used in Experiment 1 were targeted. Although Experiment 1 did not use the second words of the triplets, Experiment 3 used these words in the experimental lists. The Experiment Sets A and B of Experiment 1 were again used in Experiment 3, but with the second words of the triplets. For each set, two triplets out of 18 triplets were used as a six-word list: Each set had nine experimental lists. An example of the experimental lists is “heavy, light, feather, gas, oil, slick” based on the two triplets, “heavy-light-feature” and “gas-oil-slick” (see Figure 4).

We also created two sets of control lists (Control Sets A and B) by re-combining words for each of the Experimental Sets. The Control Set A, for example, included the words of the Experimental Set A but re-combined the words as new triplets. Specifically, the second word of a triplet was replaced with the second word of another triplet. The new triplets met the criteria that the associative strength between the first and second words and that between the second and third words were zero. In other words, the second words in the Control Sets were not directly associated with the first or third words.

Semantic Properties

As the associative strengths between the first and second words and those between the second and third words were all zeros in the Control Sets, we checked if the associative strengths were indeed greater than zero in the Experimental Sets with one-sample Bayesian t-tests. As for LSA similarity, we used Bayesian two-sample t-tests to compare the Experimental and Control Sets.

For simplicity, associative strengths and LSA similarity values between the nth and mth are indicated as AS and LSA with subscripts n and m (e.g. AS₁₂ = the associative strength or direct association between the first and second words). AS₁₂ in the Experimental Set A (M = 0.087, SD = 0.072) was greater than zero, BF₁₀ = 329.310. AS₂₃ in the Experimental Set A (M = 0.065, SD = 0.060) was, again, greater than zero, BF₁₀ = 119.124. In addition, LSA₁₂ was larger in the Experimental Set A (M = 0.35, SD = 0.18) than in the Control Set A (M = −0.00, SD = 0.05), BF₁₀ = 2.790 × 10⁶. LSA₂₃ was also larger in the Experimental Set A (M = 0.20, SD = 0.17) than in the Control Set A (M = −0.01, SD = 0.03), BF₁₀ = 2,015.893. As for the Experimental Set B, AS₁₂ (M = 0.101, SD = 0.066) was greater than zero, BF₁₀ = 3,800.156, and AS₂₃ (M = 0.050, SD = 0.046) was also greater than zero, BF₁₀ = 116.093. LSA₁₂ was larger in the Experimental Set B (M = 0.39, SD = 0.26) than in the Control Set B (M = 0.00, SD = 0.05), BF₁₀ = 2.818 × 10⁴. LSA₂₃ was also larger in the Experimental Set B (M = 0.12, SD = 0.15) than in the Control Set B (M = −0.01, SD = 0.07), BF₁₀ = 17.851. Therefore, for the Experiment Sets, there were indeed direct associations between the first and second words of triplets and between the second and third words of triplets. These words were, however, more similar to each other in terms of LSA similarity, compared to the words of the Control Sets. As directly associated words are likely to be similar, the direct association effect would be affected by similarity-based processes. We note that Experiment 3 was designed to examine the direct association effect, complementary to Experiment 1, but not to test the spreading-activation process. This was true for the next experiment (Experiment 4).

Task, Procedure, and Analysis Plan

The Web-based immediate serial recall task was used same as in Experiment 1. The procedure was identical to Experiment 1 expect that the new nine experimental and nine control lists were presented. Experiment 3 had a 2 × 6 factorial design with the List Type factor (experimental vs. control lists; within-participants) and the Position factor (1–6 presentation positions; within-participants).

Results

Item Correct

The main effects of Position, BF₁₀ = 6.937 × 10³⁹, and List Type, BF₁₀ = 3.274 × 10²⁰, and their interaction, BF₁₀ = 9,931.105, were confirmed (Figure 7). These BFs indicate extreme evidence. The performance for the experimental lists (M = 0.891, SD = 0.095) was better than for the control lists (M = 0.785, SD = 0.139).

Figure 7.

Results of Experiment 3.

Given the interaction, we compared performance for the experimental and control lists on each position with separate Bayesian t-tests. The effect of List Type on the first position was not supported, BF₀₁ = 7.080. By contrast, its effects on the other positions were confirmed: BF₁₀ = 1.464 × 10⁴ (position 2), BF₁₀ = 39.876 (position 3), BF₁₀ = 327.491 (position 4), BF₁₀ = 1.982 × 10⁵ (position 5), and BF₁₀ = 1.157 × 10⁴ (position 6).

Correct-in-Position and Conditionalized Order Recall

The patterns of correct-in-position scores resembled those of item correct. The main effects of Position, BF₁₀ = 1.556 × 10⁷⁶, and List Type, BF₁₀ = 3.704 × 10²³, and their interaction, BF₁₀ = 1.627 × 10⁶, were supported. The experimental list (M = 0.808, SD = 0.152), again, outperformed the control list (M = 0.684, SD = 0.185). As for the interaction, separate Bayesian t-tests revealed that the List Type’s effect was evident on positions 2 to 6 (position 2: BF₁₀ = 5.966 × 10⁴; position 3: BF₁₀ = 425.437; position 4: BF₁₀ = 459.868; position 5: BF₁₀ = 2.174 × 10⁹; position 6: BF₁₀ = 5.929 × 10⁴) but not on position 1, BF₀₁ = 6.902. Furthermore, conditionalized order recall was better for the experimental lists (M = 0.891, SD = 0.111) than for the control lists (M = 0.850, SD = 0.131), BF₁₀ = 9.013.

Discussion

Consistent with the hypothesis, item correct was better in the experimental lists than in the control lists. This provided evidence for the direct association effect on item memory. The follow-up analysis on the interaction between List Type and Position also supported the direct association effect. Note that the words presented at positions 3 and 6 correspond to the third words of the triplets (e.g. heavy, light, feather, gas, oil, slick). As shown in Figure 7, item correct scores for these words were better when they were preceded by directly associated words (in the experimental list) than by unrelated words (in the control list: heavy, fantasy, feather, gas, tiger, slick). As for the other measures, correct-in-position and conditionalized order recall were also better in the experimental lists than in the control lists. The results of Experiments 1 and 3 would suggest the presence of the direct association effect and absence of the mediated association effect in the current study’s experimental settings using Balota and Lorch’s materials.

We point out the procedural differences in the two experiments: Experiment 3 explicitly presented triplets (e.g. “heavy, light, feather”) while Experiment 1 presented pairs based on triplets (e.g. “heavy, feather”). To provide further evidence for the direct association effect, the next experiment presented pairs of direct associations (e.g. “light, feather”).

Experiment 4

Experiment 4 used pairs of words with direct association, the second and third words of the triplets, for the experimental lists (see Figure 4). These pairs were preceded by unrelated words rather than the first words. Based on the previous study using pairs of associated words (Saint-Aubin et al., 2014), we predicted that the association should lead to a memory advantage.

Method

Hypothesis, Measures, and Data Collection Plan

It was hypothesized that item memory should be higher for the experimental lists with direct association than for the control lists without direct association. This hypothesis was tested with item correct scores, whereas correct-in-position and conditionalized order recall were additionally used in an exploratory analysis.

The data collection plan was identical to the previous experiments’ plan. We recruited an initial sample of 60 participants and checked BF. As it met the criterion, we stopped data collection after the first recruitment.

Participants

With the recruitment criteria of Experiment 2, 60 participants who had not taken part in previous experiments were recruited via Prolific (age: M = 25.88 years, SD = 3.43; gender: 30 women, 29 men, 1 Other). They were compensated £1.80 for their participation. They completed the experiment within approximately 5 min (M: 7.62 min; Mdn: 6.5 min; range 5–15 min).

Materials

List Construction

The experimental and control lists were created by replacing the first words of the triplets in the experimental and control lists of Experiment 3 with unrelated words. The unrelated words were selected with constraints that (1) they had no direct associations to the second words of the triplets in the experimental lists and (2) they had no direct associations to the unrelated words of the control lists. As for the examples in Figure 4, “tooth” was not associated with either “light” or “fantasy” and “sky” was not associated with either “oil” or “tiger.”

Semantic Properties

The second and third words of the triplets of Experiment 3 were re-used in Experiment 4. Thus, AS₂₃ and LSA₂₃ were as reported in Experiment 3. With the above-mentioned constraints, AS₁₂ values were all zeros for all sets. As for LSA₁₂, the pairs of the Experimental Set A (M = 0.01, SD = 0.05) were hardly different from those of the Control Set A (M = −0.00, SD = 0.04), BF₀₁ = 2.519. The pairs of the Experimental Set B (M = 0.02, SD = 0.07) were, again, hardly different from those of the Control Set B (M = −0.00, SD = 0.06), BF₀₁ = 2.335. Therefore, the Experimental and Control Sets were matched in terms of AS₁₂ and LSA₁₂.

Task, Procedure, and Analysis Plan

The procedure was identical to that of Experiment 3 except that we used the new nine experimental and nine control lists. Following the previous experiments, it had a 2 × 6 factorial design with the List Type factor (experimental vs. control lists; within-participants) and the Position factor (1–6 presentation positions; within-participants).

Results

Item Correct

We found extreme evidence for the main effects of Position, BF₁₀ = 4.881 × 10⁵¹, List Type, BF₁₀ = 1.784 × 10⁷, and their interaction, BF₁₀ = 115.062 (Figure 8). On average, the experimental list (M = 0.841, SD = 0.099) indeed exhibited better performance than the control list (M = 0.773, SD = 0.122).

Figure 8.

Results of Experiment 4.

We examined the interaction by comparing the experimental and control lists on each position with Bayesian t-tests. On positions 1, 2, and 4, moderate evidence against the effect of List Type was provided: BF₀₁ = 4.748 (position 1), BF₀₁ = 3.353 (position 2), and BF₀₁ = 5.841 (position 4). On positions 3, 5, and 6, however, strong or extreme evidence for its effect was obtained: BF₁₀ = 26.820 (position 3), BF₁₀ = 3,305.427 (position 5), and BF₁₀ = 7.745 × 10⁴ (position 6).

Correct-in-Position and Conditionalized Order Recall

With correct-in-position scores, the main effects of Position, BF₁₀ = 5.569 × 10⁸⁸, and of List Type, BF₁₀ = 2.248 × 10⁵ were revealed. The experimental list (M = 0.739, SD = 0.143) outperformed the control list (M = 0.673, SD = 0.180). The interaction between Position and List Type was not confirmed, BF₀₁ = 11.780. As for conditionalized order recall, performance was equivalent in the experimental list (M = 0.865, SD = 0.112) and the control list (M = 0.852, SD = 0.142), BF₀₁ = 4.624.

Discussion

Consistent with the hypothesis, on average, item correct scores were higher for the experimental list than for the control list. Furthermore, as for the interaction, the List Type’s effects were confirmed at positions 3 and 6. This indicated that item correct was facilitated when items (at positions 3 and 6) were preceded by directly associated words compared to unrelated words (at positions 2 and 5).

The results of the three experiments using Balota and Lorch’s materials are straightforward: The mediated association hypothesis was not supported (Experiment 1), while the direct association effect was demonstrated (Experiments 3 and 4). Nevertheless, with our rich dataset, independent of Experiment 1, a further investigation of the mediated association hypothesis was possible. This test was implemented by conducting a cross-experiment comparison with the data of Experiments 3 and 4.

Cross-Experiment Comparison: Experiments 3 and 4

In this post-hoc analysis, we compared the performance of the triplets’ third words on the control lists of Experiments 3 and 4 (i.e. the third and sixth words in the lists) (see Figure 4). In Experiment 3, both first and third words of the triplets (with mediated association) were presented but separated by an unrelated word (i.e. first, unrelated, third: heavy, fantasy, feather). In Experiment 4, the first words of the triplets were not presented, but two unrelated words preceded the third words of the triplets (i.e. unrelated, unrelated, third: tooth, fantasy, feather). If mediated associations have an effect, the third and sixth words in the control condition of Experiment 3 should be better recalled than those of Experiment 4. This possibility was suggested by previous findings with direct associations. Saint-Aubin et al. (2014) showed that directly associated words separated by an unrelated word (i.e. associated, unrelated, associated) were better recalled than control non-associated words. Although this effect is smaller than the standard association effect, in which associated words are presented contiguously, it is still detectable. Therefore, we suggest that the mediated association effect would be potentially detected in our cross-experiment comparison.

Method

We focused on the control lists of Experiments 3 and 4. The third word of the triplets was identical in both experiments. That is to say, the third and sixth words in the lists were the same. These words were indirectly associated with the words at two positions ahead in Experiment 3 (e.g. “heavy, fantasy, feather”) while they were not in Experiment 4 (e.g. “tooth, fantasy, feather”). The averaged accuracy of the third and sixth words in the lists was calculated for each experiment, and these values were compared with Bayesian t-tests.

Results and Discussion

Item correct scores of the third and sixth words in the control lists were close to each other in Experiment 3 (M = 0.714, SD = 0.185) and Experiment 4 (M = 0.693, SD = 0.175), BF₀₁ = 4.252. ² Correct-in-position scores were also equivalent in Experiment 3 (M = 0.606, SD = 0.234) and Experiment 4 (M = 0.582, SD = 0.228), BF₀₁ = 4.440. Therefore, indirectly associated words separated by an unrelated word did not lead to a memory advantage, which provides further evidence against the mediated association effect.

General Discussion

Semantic relatedness has been extensively investigated to assess the lexical-semantic influences on WM (Neath et al., 2022; Poirier & Saint-Aubin, 1995; Saint-Aubin & Poirier, 1999a). Among the types of semantic relatedness is association, which was targeted in the current study. We proposed and tested a novel hypothesis that items with mediated association should lead to a memory advantage, building on the spreading-activation theory. Experiment 1 tested this hypothesis by presenting word pairs of indirectly associated words such as “heavy” and “feather” (i.e. the first and third words of the triplets, see Figure 4). If mediated association plays a role, “heavy” would activate non-presented “light,” which would in turn activate “feather.” Counter to our hypothesis, the pairs of indirectly associated words did not lead to a memory advantage compared to the control pairs. Experiment 2 was a follow-up of Experiment 1. Experiment 2A provided evidence for the mediated association effect, while Experiment 2B did not. Therefore, only one of the three experiments demonstrated the mediated association effect, while the other two experiments did not. Note that moderate evidence for the effect was present only when data were added. Furthermore, an item-level analysis using the data from Experiments 1, 2A, and 2B did not show that the recall performance of an item was related to its mediated association value (to the preceding item). A cross-experiment comparison again did not support the mediated association effect. In Experiments 3 and 4, the direct association rather than the mediated association was examined. The results of these experiments clearly showed the facilitative effect of direct association.

Taken together, the current study indicated a substantial effect of direct association and an elusive or negligible effect of mediated association. This does not mean that moderate evidence for the mediated association effect in Experiment 2A itself can be neglected. As Experiment 2A had the largest mediated association values among Experiments 1, 2A, and 2B, the presence of the effect may correspond to the strength of mediated association. We, however, suggest that it is negligible in relation to the previous and current findings. With Balota and Lorch’s materials, the mediated association effect has been demonstrated in lexical decision (Balota & Lorch, 1986) and naming (McNamara & Altarriba, 1988). Experiment 1 used their materials but failed to show the mediated association effect on WM. Experiments 3 and 4 again used Balota and Lorch’s materials and demonstrated the direct association effect. Therefore, the mediated association effect on WM seems to be less reliable compared to the mediated association effect on priming and the direct association effect on WM.

Priming and WM tasks differ in terms of required processing, timescale, and target measure. In a lexical decision task (priming), participants decide whether an item is a word. Typically, reaction time is targeted and measured in milliseconds. In an immediate serial recall task (WM), participants remember items. Even if recall is prompted immediately after the presentation of the last item, they must maintain the items for over seconds. This is because the earlier items must be maintained during encoding of the later items, and the later ones must be maintained during retrieval of the earlier ones. Despite these differences, the commonality between these tasks can also be recognized. Direct association facilitates performance in both priming (Fischler, 1977; Perea & Rosa, 2002; Shelton & Martin, 1992; for semantic priming, see also Buchanan et al., 2025) and WM tasks (Saint-Aubin et al., 2014; Experiments 3 and 4 in the current study). Therefore, it is possible that mediated association would facilitate performance in both priming and WM tasks, but it was not the case. It is an open question why mediated association affects priming while it does not reliably affect WM. To answer this, future research should conduct a detailed investigation into required processing, timescales, and measures of priming and WM tasks.

Theoretical Implications

The spreading-activation process has been incorporated into WM models of semantic information (Kowialiewski, Krasnoff, et al., 2022; Kowialiewski et al., 2021; Kowialiewski & Majerus, 2020; Saint-Aubin et al., 2005; Tse, 2009). The direct association effect corroborates the spreading-activation process, but it may also support other processes of dimension-based similarity, such as compression (Kowialiewski, Lemaire, et al., 2022) or summation of representations’ elements (Guitard et al., 2025). This is because items with direct associations can be, at the same time, similar to each other (see Figures 1 and 2). To address the spreading-activation process exclusively, the current study focuses on mediated association, which does not rely on the direct interitem relationship, such as dimension-based similarity.

Word pairs with indirect association can be viewed as dissimilar word pairs. In fact, we successfully matched these pairs with the control pairs in terms of LSA in Experiments 1 and 2, and thus dimension-based similarity was controlled for. They were, however, expected to facilitate memory performance based on a strong assumption of the spreading-activation theory, namely, the mediated association process. Theoretically, the spreading-activation process is assumed to work not only on direct association but also on mediated association (Collins & Loftus, 1975; Nelson et al., 1998), which is at least supported by findings of priming studies (Balota & Lorch, 1986; Jones, 2010, 2012; McNamara & Altarriba, 1988). Thus, it can be naturally inferred that mediated association as well as direct association would bring about a facilitative effect.

Nonetheless, our findings suggest that the mediated association effect on WM is not reliable. We suggest that the negligible effect of mediated association would set an important constraint in modelling the association effect or the semantic relatedness effect in general. Even though the spreading-activation theory assumes the propagation of activations more than one step (e.g. two-step activation), such a process may not be substantial for WM. Given the direct association effect, we do not reject the existence of a spreading-activation process in the maintenance of associated words. Our demonstration would rather delimit the scope of the association’s influence in one-step association.

The delimitation of association’s influence is critical for modelling the cue-dependent retrieval process as well as the spreading-activation process. Category membership is thought to offer an extra cue such as “animal” for the items “cat” and “dog” (Ishiguro et al., 2025). According to the cue-dependent retrieval account, the extra cue serves as a retrieval cue (Ishiguro & Saito, 2024). Nevertheless, if mediated association played a substantial role (e.g. “cat” → “animal” → “dog”), cue-dependent retrieval might be redundant and unnecessary. Based on the negligible effects of mediated association, we suggest that direct association between the cue and item (e.g. “animal” → “cat” and “animal” → “dog”) would be a critical factor for the cue-dependent retrieval process.

Conclusions

Building on the spreading-activation theory, we tested a novel hypothesis that indirectly associated words should lead to a memory advantage (e.g. for indirectly associated words “lion” and “stripes,” lion → tiger → stripes). This hypothesis was not clearly supported. Although our results are not diagnostic of the spreading-activation process, they would delimit the scope of spreading activation in WM. In a broader context, the conceptualization of “semantic relatedness” is recognized as key to interpreting the semantic relatedness effect (Hart et al., 2025; Ishiguro & Saito, 2024). The contrasting effects of mediated and direct associations in the current study highlight the importance of direct relatedness between items. We suggest that the current study’s approach – deriving a prediction from a theory (e.g. the spreading-activation theory) and introducing a concept from another research line (e.g. priming research) – would be effective in conceptualizing “semantic relatedness” in future research.