Executive functioning moderates the decline of retrieval fluency in time

Prevailing theoretical accounts consider that automatic and controlled processes are uniformly engaged in memory retrieval across performance of the semantic verbal fluency (SVF) task. We tested this proposal against the alternative, namely that a rapid automatic retrieval, exploiting stable associative structure in the early stages of the performance, is followed by a slower, more executively demanding, retrieval in later stages. Eighty-five healthy adults completed low- and high-demand SVF tasks that were assessed for retrieval rate, response typicality, and inter-response similarity across the performance. Additional measures of executive functioning were collected to estimate individual differences in executive control. We found that decrease in fluency in time was associated with lower typicality and weaker semantic similarity of the responses. Critically, the time-dependent retrieval slowing was steeper in individuals with less efficient interference control, particularly in high-demand SVF tasks. Steeper retrieval slowing was also associated with poorer working-memory capacity. Our findings show that the relative contribution of automatic and controlled processes to semantic retrieval changes with associative sparsity over time and across task demands, and provide implications for the use of SVF tasks in clinical assessment. Supplementary Information The online version contains supplementary material available at 10.1007/s00426-022-01680-0.


Semantic control task stimulus materials
The full list of items is downloadable in English version at the OSF repository (url: osf.io/s63yh).Items belonging to the Global and Feature conditions were divided into separate blocks (four for each condition).These blocks were then organized into four super-ordinate blocks that contained one Global and one Feature condition block.These super-ordinate blocks were presented randomly, however, within them a Global block was always presented first.This ensured that participants always begun with a Global block.Condition blocks were a priori randomly paired into the super-ordinate blocks.Each block from both Global and Feature condition contained equal number of strongly and weakly related, and equal number of congruent and incongruent trials.Furthermore, in Feature blocks the trails were assembled such that each block contained roughly the same number of feature (color, shape, size, texture) types.Trials were displayed randomly within the blocks.Likewise, trials were assigned to individual blocks on random basis (random.org)Regarding the lexical properties of the items, we compared the stimuli in word frequency, length, and semantic relatedness that was operationalized by the logDice measure derived from the national corpus database.
For the Global condition, the cue, target, and alternative words did not differ significantly between strong and weak trials in word frequency and word length, ts(30) < 1.62, ps > .116.However, in strong trials, the target words were significantly more related to the cue words than in the weak trials, t(30) = 30.72,p < .001.Alternative words were not differently related to the cue words in a significant manner, t(30) = -1.00,p = .325.Moreover, for both strong and weak trials, target and alternative words did not differ significantly in word frequency and length, ts(30) < 1.55, ps > .131 In the Feature condition, cue words in congruent trials had statistically higher word frequency compared to incongruent trials, [t(30) = 3.22, p = .003],but did not differ in mean word length, t(30) = -0.26,p = .796.Target and distractor (alternative) words were also not statistically different in word frequency and length, [ts(30) < 1.31, ps > .199] between congruent and incongruent trials.However, the targets were significantly more semantically related to the cue words in the congruent than in incongruent trials, t(30) = 30.03,p < .001.Conversely, the distractor words were significantly more related to the cue words in the incongruent compared to congruent trials, t(30) = -35.75,p < .001.Similarly, in both congruent and incongruent trials, target and alternative words did not differ significantly in word frequency and length, ts(30) < -1.34, ps > .190.The procedure with example trials is depicted in the Figure S1.

Data processing
The SVF task response appropriateness was evaluated by three independent raters (absolute agreement average-measure ICC = .937).Following this procedure, responses judged as inappropriate (5.40%), extremely slow responses (IRI > 35s; 0 .44%),and repetitions (2.01 %) were excluded.The remaining IRT data were winsorized (10% two-sided trimming) across participants and SVF task demands (i.e., individual categories).In the Stroop task data, erroneous trials (1.89%) and trials with RT > 3 SD (1.62%) were excluded before computing individuals' averages.In the SCT data, four test items from the feature-incongruent condition had > 60% error rate and hence were excluded from further analyses.Additionally, we removed erroneous responses (7.77%) and responses with RT > 3 SD (0.64%; across all matching rule × trial types) before computing individual contrasts.
In ACT data, responses that either did not comply with the retrieval rule or took > 20 sec were excluded (8.02%), following which the remaining RTs were winsorized (10% two-sided trimming) across participants, retrieval type (associative, dissociative) and chain type (fixed, alternating) conditions.Note that the final RT contrasts from executive control measures were also winsorized (10% two-sided trimming).

Contrasts from executive control tasks
The data from the Stroop task showed a large and significant interference effect with longer RTs in

Main effect of the SVF task demands on intrusions
Task demands had a moderate and significant main effect on the number of intrusions in a response sequence, F(3, 252) = 14.651, p < .001,η 2 = .149,showing that intrusions occurred more often in high demand than low demand categories.Pairwise comparisons indicated highest intrusion rate in the Liquids category compared to all other categories (pHolm < .007).Tools category produced significantly larger proportion of intrusions than Animals and Vegetables category (pHolm < .049)(Figure S2).

Main effect of the SVF task demands on TYP (pairwise comparisons)
Despite a significant effect of the task demands (without the Vegetables category) on TYP, we did not observe any significant differences between the respective task demands levels in terms of average TYP, t < 1.599, pTukey = .246.

Main effect of the SVF task demands on IRS (pairwise comparisons)
On for the results of category-specific mediation analyses.Note.Lower and upper confidence intervals of estimates were obtained by quasi-Bayesian approximation (5000 Monte Carlo draws).ROretrieval order; TD -SVF task demands; TYPresponse typicality; IRSinter-response similarity; IRT inter-response time.

Table S1 .
Unstandardized parameter estimates for the mediation models forIRT via TYP and IRS  (Animals category)

Table S2 .
Unstandardized parameter estimates for the mediation models forIRT via TYP and IRS  (Vegetables category) Note.Lower and upper confidence intervals of estimates were obtained by quasi-Bayesian approximation (5000 Monte Carlo draws).ROretrieval order; TD -SVF task demands; TYPresponse typicality; IRSinter-response similarity; IRT inter-response time.

Table S3 .
Unstandardized parameter estimates for the mediation models forIRT via TYP and IRS  (Tools category)