Abstract
Repetition improves retrieval from memory; however, under some circumstances, it can also impair performance. Separate literatures have investigated this phenomenon, including studies showing subjective loss of meaning following ‘semantic satiation’, slowed naming and categorisation when semantically related items are repeated and semantic ‘access deficits’ in aphasia. Such effects have been variously explained in terms of habituation of repeatedly accessed representations, increased interference from strongly activated competitors and long-term weight changes reflecting the suppression of non-targets on earlier trials (i.e., retrieval-induced forgetting). While studies of semantic satiation involve massed repetition of individual items, competition and weight changes at the conceptual level should elicit declining comprehension for non-repeated items: this pattern has been demonstrated for picture naming but effects in categorisation are less clear. We developed a paced serial semantic task (PSST), in which participants identified category members among distracters. Performance in healthy young adults deteriorated with ongoing retrieval for non-repeated words belonging to functional categories (e.g.,
Introduction
Repetition and priming largely have beneficial effects: they facilitate processing efficiency (Wagner, Desmond, Demb, Glover, & Gabrieli, 1997) and increase the accessibility of memory representations (Radeau, Besson, Fonteneau, & Castro, 1998). A similar benefit of repetition occurs for semantically related items, where
Patients with semantic access impairment show ‘refractory effects’ or declining comprehension in cyclical word–picture matching tasks (Jefferies et al., 2007; Warrington & McCarthy, 1983). When sets of semantically related items are repeatedly presented, such that the target on one trial becomes a distractor on the next, patients become increasingly unable to select the target (Forde & Humphreys, 1997; Warrington & Cipolotti, 1996; Warrington & Crutch, 2004). This is only observed when the interval between one stimulus and the next is short (Campanella, Mondani, Skrap, & Shallice, 2009; Jefferies et al., 2007) and when the items are highly related in meaning (Cipolotti & Warrington, 1996; Crutch & Warrington, 2008; Forde & Humphreys, 1995, 1997; Jefferies et al., 2007). These effects have largely been documented in verbal comprehension tasks – i.e., word–picture matching (Cipolotti & Warrington, 1996; Jefferies et al., 2007; Schnur et al., 2006; Warrington & Crutch, 2004; Warrington & McCarthy, 1983). However, they have also been demonstrated in non-verbal judgements such as picture–picture matching (Forde & Humphreys, 2007; Gardner et al., 2012; Thompson, Robson, Lambon Ralph, & Jefferies, 2015), suggesting semantic access deficits can occur at a conceptual level. The mechanism that underpins this phenomenon is somewhat unclear (Mirman, 2011): declining performance on cyclical word–picture matching tasks has been linked to difficulty overcoming post-retrieval inhibition of selected representations (Gotts & Plaut, 2002) or to strong competition when several potential responses are activated (Forde & Humphreys, 1997; Jefferies et al., 2007; Schnur et al., 2006). The effect is strongest in patients with damage to prefrontal cortex (Campanella et al., 2009; Gardner et al., 2012; Schnur et al., 2006; Thompson et al., 2015), suggesting that it may reflect damage to control mechanisms that are necessary to maintain performance in the presence of strong competition on later trials and/or to overcome post-retrieval inhibition (i.e., when targets have to be re-selected having being inhibited on previous trials). In line with this view, patients with semantic aphasia (SA) show deficient semantic control across verbal and non-verbal tasks, characterised by difficulty suppressing irrelevant aspects of knowledge and comprehending distant relationships and ambiguous meanings (Corbett, Jefferies, & Lambon Ralph, 2011; Jefferies & Lambon Ralph, 2006; Noonan, Jefferies, Corbett, & Lambon Ralph, 2010). These deficits give rise to declining comprehension in both cyclical word–picture matching and picture–picture matching (Gardner et al., 2012; Thompson et al., 2015).
Healthy participants typically do not show declining performance on cyclical word–picture or picture–picture matching tasks (Damian et al., 2001; Riley, McMahon, & de Zubicaray, 2015, some exceptions discussed below), although they do show increasing latencies in cyclical picture-naming studies when sets of semantically related items are named repeatedly (Kroll & Stewart, 1994; Vitkovitch & Humphreys, 1991). In some previous studies, these effects were explained in terms of competition from activated conceptual representations at the point of lexical selection: items drawn from the same semantic category activate each other via their shared conceptual features, and this might hinder retrieval of a specific object name because other activated concepts act as competitors (Belke, 2008; Belke et al., 2005; Damian et al., 2001). ‘Semantic blocking’ effects in naming are sensitive to semantic variables, such as the strength of the association between the items (Abdel Rahman & Melinger, 2011), supporting the view that they reflect processes at the conceptual level.
A continuous picture-naming paradigm can also elicit declining performance when conceptually related items are presented (i.e., two animals) without the repetition of individual items (Belke, 2013; Belke & Stielow, 2013; Howard et al., 2006; Kleinman, 2013; Navarrete, Mahon, & Caramazza, 2010; Oppenheim et al., 2010; Runnqvist, Strijkers, Alario, & Costa, 2012; Schnur, 2014). For example, in the sequence GOAT, CAR, TOMATO, TRUCK and HORSE, the naming time for HORSE is slower than for GOAT due to their conceptual overlap, and this effect is found even at long ‘lags’ with many unrelated intervening items. Since this effect appears to be relatively long term (cf. Wheeldon & Monsell, 1994), it is argued to reflect cumulative weight changes: each time an item from the animal category is named, the associative link between the word and animal features in the semantic system might be strengthened, increasing competition on future trials (Howard et al., 2006). Oppenheim et al. (2010) proposed that this type of incremental learning might not only reinforce the connections between semantic and lexical representations of targets but also weaken semantic-lexical links for non-targets. Thus, semantic interference effects in naming can arise as a direct consequence of retrieval that renders related items less accessible (see also Anderson, Bjork, Bjork, & Jordan, 2000, for a related phenomenon in memory, termed ‘retrieval-induced forgetting’).
Several recent studies used a continuous paradigm without item repetition to examine categorisation as opposed to picture naming and observed cumulative facilitation rather than inhibition (Belke, 2013; Riley et al., 2015). Nevertheless, under some circumstances, healthy controls can show declining categorisation with repetition (Harvey & Schnur, 2015) and thus resemble patients with semantic access impairment. This pattern was observed in cyclical matching to a deadline when there were repeated presentations of the same target plus minimal delays between trials (Fabio Campanella & Shallice, 2011): these circumstances potentially create competition between the current target and previous targets (which have become distractors), with little time to resolve this competition or to recover from previous processing. In addition, Wei and Schnur (2015) reported semantic interference in a picture matching task, when the same response options were repeatedly used to probe associations with either related or unrelated concepts; in this study, there was initial facilitation (when semantically related items were repeated at short lags, perhaps reflecting faster visual recognition for the probe when immediately following a related item), followed by long-lasting inhibition (when related trials occurred at longer lags, perhaps reflecting response interference when a similar probe had led to a different decision on a previous trial).
Long-lasting declines in comprehension with repetition have also been reported in a third set of studies on ‘semantic satiation’: this research reports that prolonged inspection and repetition of words result in a subjective loss of meaning (Jakobovits & Lambert, 1962; Smith & Klein, 1990). Semantic judgements are slowed under these circumstances, but there is little effect on tasks such as lexical decision, suggesting that this effect again reflects effects at a semantic level (Cohene, Smith, & Klein, 1978; Neely, 1977a; Smith, 1984). Repetition is thought to cause temporary blocking of access to conceptual information (Frenck-Mestre, Besson, & Pynte, 1997; Pynte, 1991), potentially reflecting effects akin to neural fatigue or adaptation (Jakobovits & Lambert, 1962; Lambert & Jakobovits, 1960; Smith & Klein, 1990). Explanations for semantic satiation effects are similar to those above in that they anticipate spreading activation or inhibition to related conceptual representations; conceptual representations become ‘habituated’ via repeated exposure, disrupting subsequent category judgements.
It is clear from the discussion above that semantic similarity across successive targets can produce both behavioural facilitation and inhibition on later trials, yet the situations that elicit these opposing effects remain largely unknown. Picture naming tends to show inhibition, whether or not items are repeated (e.g., Howard et al., 2006), while comprehension tasks tend to show facilitation (Belke, 2013; Riley et al., 2015) – although they sometimes show inhibition with item repetition (Campanella & Shallice, 2011; Wei & Schnur, 2015). There might be multiple mechanisms underpinning these diverse changes in performance, including priming, stimulus adaptation, competitive effects relating to spreading activation and lateral inhibition (thought to occur when target selection gives rise to the suppression of semantically related competitors, which later become targets). The balance between these effects is likely to depend on the specific task that is presented. Some researchers have suggested that reductions in performance are restricted to paradigms requiring lexical selection (e.g., Damian et al., 2001; Riley et al., 2015); however, such effects can have a conceptual locus (e.g., Wei & Schnur, 2015) and therefore other task differences might be critical. A second difference between naming and word–picture matching is that naming involves generating representations from memory, while matching tasks present all the critical information as part of the paradigm. For this reason, retrieval demands are thought to be higher in picture naming. Controlled retrieval processes might give rise to declining performance on picture naming if participants retrieve targets by suppressing semantic neighbours, which later become targets (Oppenheim et al., 2010). These effects might not be prominent in most comprehension tasks, when the need for controlled retrieval is reduced.
Given these considerations, this study used a paced serial semantic task (PSST) in which decisions to inputs were taken on the basis of a previously encoded goal. Participants monitored a stream of inputs and pressed a button every time they detected a target that matched a particular category specified at the start of each block. We considered the ability of participants to sustain semantic processing over time: both within categories – by examining whether comprehension deteriorated over the course of each category as more related targets were presented – and across categories – by quantifying changes in performance across the experiment, as participants became generally fatigued. Neither targets nor distracters were repeated – therefore, the task resembled a continuous naming paradigm but required the comprehension and categorisation of meaningful inputs, rather than the production of speech. Although previous continuous categorisation experiments have shown facilitation (Belke, 2013; Riley et al., 2015) and, to our knowledge, there are no previous demonstrations of declining comprehension in the absence of stimulus repetition, our task might elicit cumulative inhibitory effects since semantic retrieval was guided in a top-down fashion (i.e., participants searched for an overlap between each input and the current goal category). This type of task might be expected to increase the need for controlled retrieval to each stimulus, as participants were required to focus on only currently relevant aspects of knowledge. This aspect of the PSST might encourage the suppression of currently irrelevant yet related concepts, which later become targets. The PSST paradigm allowed us to manipulate factors linked to both semantic representations themselves – such as strength of association, which should influence the spread of activation/inhibition between related concepts – and factors linked to control processes thought to play critical role in focussing retrieval on currently relevant knowledge in the face of strong distractors or weak targets (Badre, Poldrack, Paré-Blagoev, Insler, & Wagner, 2005; Thompson-Schill, D’Esposito, Aguirre, & Farah, 1997; Whitney et al., 2011) – such as conditions of divided versus undivided attention. In this way, the task maps onto contemporary accounts of semantic processing, which envisage conceptual representations that interact with control processes to support context- and task-appropriate semantic retrieval (Lambon Ralph, Jefferies, Patterson, & Rogers, 2016).
In the first part of the study, we manipulated factors that might influence the extent to which comprehension declines in healthy young adults. In Experiment 1, we investigated speed of presentation and semantic relatedness (both factors that produce declining comprehension in patients with semantic access deficits). We replicated the declining performance that we observed for thematic categories using taxonomic and specific feature judgements (Experiments 2 and 3) and found the same pattern of declining semantic performance with ongoing categorisation using a two alternative-forced-choice (2AFC) paradigm (Experiment 4). We also examined whether the effect extended beyond verbal comprehension to a task involving interleaved word and picture stimuli, to establish whether this effect has a conceptual locus (which should transfer between modalities) (cf. Thompson et al., 2015; Wei & Schnur, 2015). Finally, we investigated whether the within-category decline in categorisation was increased by the requirement to divide attention in a dual-task study, as would be expected if executive mechanisms guide retrieval in the face of strengthening competition or inhibition from semantically related items (Experiment 5). Changes in the accessibility of meanings with ongoing categorisation in healthy young adults, in the absence of massed repetition, would have important implications for everyday comprehension.
In the second part of the study, we considered how SA influences performance on the PSST paradigm. These patients have well-documented deficits of semantic control alongside damage to left inferior frontal gyrus, which is implicated in the control of semantic retrieval by neuroimaging (Badre et al., 2005; Noonan, Jefferies, Visser, & Ralph, 2013) and neurostimulation studies (Whitney et al., 2011). Neuropsychology provided a straightforward way of distinguishing between alternative accounts of the within-category decline in categorisation: if these effects reflect a build-up of competition as more and more related items are presented, SA patients should have difficulty resolving this competition and therefore show an increase in the within-category decline in categorisation. In contrast, if retrieval-dependent declines in comprehension are a consequence of successfully dealing with competition on earlier trials through the inhibition of non-target information (cf. Anderson et al., 2000), SA patients might be expected to show attenuated or absent within-category decline, since their retrieval is relatively uncontrolled (e.g., Crutch & Warrington, 2005).
Experiment 1: effect of speed of presentation and strength of association on thematic categorisation in healthy participants
Rationale
This experiment tests whether semantic retrieval declines over time (even in the absence of explicit item repetition) and whether there is a release from this effect when the task switches to a new category. We considered whether this pattern of declining categorisation would be greater at a fast rate of presentation. Previous research has found declining performance with both fast and slower presentation speeds: fast speeds might allow a greater build-up of competition from previously presented semantically related items (Fabio Campanella & Shallice, 2011), yet satiation effects occur when inputs are presented for long durations (Smith, 1984) and performance in continuous paradigms declines even over relatively long lags, reflecting slower cumulative weight changes (Oppenheim et al., 2010). We therefore compared performance at fast (1.1 s) and slower speeds of presentation (2 s) to establish whether this manipulation would alter the extent to which semantic categorisation declined in the PSST paradigm. In addition, we tested whether the effect would vary as a function of the strength of association between the presented words and the target category. Larger effects of within-category decline might occur for weakly associated words if declining comprehension is caused by a build-up in competition, since weak associations should be harder to retrieve in the face of competition from previously presented strong category members. Alternatively, the effect could be larger for strongly associated words, if it reflects weight changes that render related non-target items on previous trials less accessible (Oppenheim et al., 2010). The literature on semantic satiation shows stronger detrimental effects of repetition for strong versus weak associates (Balota & Black, 1997). Similarly, patients with semantic access deficits show declining cyclical word–picture matching for sets of closely related items – but not for sets of repeated but unrelated or distantly related items (Crutch & Warrington, 2003). Warrington and Cipolotti (1996) found a detrimental effect of cycle in these patients even when items in the last cycle were replaced with new items from the same category, suggesting that spreading activation between strongly related concepts is the likely cause of this decline. Therefore, decline in performance was compared for strong and weak members of a thematic category (e.g.,
Method
Participants
A total of 24 undergraduate students (16 females and 8 males) from the University of York participated in the experiment in exchange for course credit or a payment of £5. The mean age of the students was 19 years (range: 18-24 years). All participants were native English speakers. Ethical approval was obtained from the Department of Psychology Ethics Committee at the University of York. All participants gave written informed consent.
Task and design
The PSST required rapid semantic association judgements that linked spoken words to a thematic category, such as
Materials
In total, 20 different category labels were used (such as
Procedure
The experiment was presented using E-Prime 2.0 (Psychology Software Tools, Sharpsburg, PA). Category names were presented as written words that remained visible throughout the block, to reduce demands on working memory. There was an equal distribution of strong and weak targets in the first and second half of each category. The order of categories and items was fully counterbalanced between subjects (there were parallel versions of the experiment utilising two presentation orders; each of these was presented to half of the participants, such that effects at the group level could not reflect effects specific to one order of presentation). These details were repeated across all experiments below.
Participants were asked to press ‘1’ each time they heard a word that was related to the category and not to press for unrelated words. Thus, the task required sustained and rapid attention to semantic information. Each participant was presented with all 20 categories, 10 at one speed (e.g., with a 1.1-s gap between each auditory word) and 10 at another speed (e.g., 2-s inter-stimulus interval [ISI]), with the two speeds counterbalanced using an ABBA or BAAB design.
Results
The main dependent measure in all experiments was response sensitivity (d′), which accounts for response bias (the general tendency to respond yes or no; Stanislaw & Todorov, 1999). A higher d′ score reflects better response sensitivity (i.e., the ability to correctly recognise targets and reject distractors – as opposed to making ‘false alarms’ on non-target items). As there was an equal distribution of strong and weak targets in the first half and second half of each category, within-category changes in performance were examined by computing d′ separately for these two halves of each category. In Experiment 1 and subsequent experiments, generalised linear models (GLMs, using generalised estimating equations) were used to analyse response sensitivity for each category and for each participant, including within-subject fixed effects of within-category position (first vs second half of each category), across-category fatigue (first vs second half of the entire experiment), speed of presentation (1.1 vs 2 s) and semantic relatedness (d′ scores computed separately for targets that were strongly or weakly related to the category), in a fully factorial model that included all interaction terms for these predictor variables. These d′ scores were therefore computed across sets of trials, retaining information about performance per category per participant (i.e., categories were treated in the same way as individual trials in a classic ‘by-items’ analysis). The GLMs allowed for random variation in the intercept across participants. Average response time (RT) was entered for each of the first and second half of each category as a covariate (i.e., the average RT for correct responses per condition and participant) in this and all subsequent GLMs. Given that the task required participants to respond before a deadline (i.e., before the onset of the next item, rather than as quickly as possible), RT was not expected to be sensitive to the effects of interest but by including average RT as a covariate in the analysis, changes in RT were accounted for over the course of a block of trials. This would allow focus on response sensitivity while simultaneously accounting for the possibility of a response accuracy trade-off.
Response sensitivity in each condition is shown in Figure 1. Table 1 shows the results of the GLM analysis (alongside conventional repeated-measures analysis of variance [ANOVA] of response sensitivity, which revealed the same effects of the experimental factors). There was a main effect of speed, indicating better performance at a slower rate of presentation. There was also a main effect of association strength: sensitivity was lower for weakly related items in comparison with stronger associations. There was no significant main effect of within-category position or across-category fatigue (p > 0.1). However, there was a significant interaction of relatedness and within-category decline: participants made more errors in categorisation towards the end of each category especially for the strongly associated targets. Post hoc tests with Bonferroni correction indicated this decline in sensitivity ‘within’ each category affected performance on strongly related targets, t(23) = 2.34, p = 0.028, but not weakly related targets, t(23) = 0.105, p = 0.917. There was also a significant interaction between speed and relatedness: participants found it harder to identify weak items at the faster speed in comparison with the slower speed. Bonferroni-corrected t-tests indicated significant effects of relatedness at both speeds, with a larger effect at the fast speed, t(23) = 7.64, p < 0.001, in comparison with the slow speed, t(23) = 5.37, p < 0.001. All other interaction terms were non-significant. A more detailed breakdown of performance, showing hits, correct rejections, false alarms and misses is reported in Table 1 in Supplemental Material. This shows few false alarms; due to the paced nature of the paradigm, participants tended to fail to respond to targets when performance was poor.
Mean response sensitivity (d’) in Experiment 1 (thematic-matching), for the first and second half of each category (within-category fatigue) and across the testing session (across-category fatigue), split by strong and weak targets, at the two presentation speeds.
Summary of significant results for response sensitivity from GLM and repeated-measures ANOVA analysis, examining effects of speed and relatedness, plus within-category and across-category changes in performance, in Experiment 1: thematic-matching.
GLM: generalised linear model; ANOVA: analysis of variance; RT: response time; df: degrees of freedom.
Two parallel analyses employing (1) mixed-effects modelling (i.e., GLM preserving performance information for each category for each participant and treating participants as a random effect – this allowed RT per category to be included as a covariate of no interest) and (2) ANOVA. Other interaction terms were non-significant (p > 0.1).
Bold indicates significant effects.
Summary of Experiment 1
Even though individual words were not repeated, as in typical semantic satiation or cyclical matching experiments, performance showed a cumulative decline across the targets in each category. Within-category decreases in response sensitivity were greater for targets strongly associated with the category label (e.g., PICNIC – ‘sandwich’), compared with weakly associated targets (e.g., PICNIC – ‘wasp’). Performance on the task was also influenced by speed of presentation: participants showed poorer performance at faster speeds. There was no decline in performance over the course of the experiment (i.e., across-category fatigue effects were not significant). Thus, the PSST revealed several of the hallmarks of ‘semantic access deficits’ in healthy subjects – namely, declining performance with ongoing semantic retrieval, greater effects when there was a deadline to respond (Fabio Campanella & Shallice, 2011), plus a more substantial decline for more strongly related items, suggesting that this effect reflects the spread of activation or inhibition within semantic representations.
Additionally, in line with studies of continuous picture naming (which examine performance as a function of how many trials lie between presentations of semantically related items), we examined accuracy as a function of the number of distracters since the last target (reported in Supplemental Material). We found no evidence that the within-category fatigue effect was reduced in magnitude when the number of intervening items was greater (see Figure 1 in Supplemental Material). This suggests that ongoing neural activation from previous related trials was unlikely to be the cause of the decline in performance, since such an account would predict the greatest decline in performance at zero lag, when there has been no time for decay in activation. These results are more consistent with the view that long-lasting weight changes underpin the effect (cf. Wei & Schnur, 2015).
Experiment 2: taxonomic category matching in healthy participants
Rationale
This experiment provided a replication of the within-category decline effect in Experiment 1 using taxonomic categories as opposed to thematic categories. The target items shared common features (e.g., eyes and fur, for the category
Method
Participants
A total of 24 undergraduate students (16 females and 8 males) were recruited from the University of York and received course credit or a payment of £5 for their participation. The mean age of the students was 21 years (range: 18-30 years). All participants were native English speakers.
Task and procedure
Targets were strong members of each taxonomic category (e.g., ‘apple’, ‘orange’, ‘grapes’ for
Results
Response sensitivity is shown in Figure 2. The effects of two fixed effects were examined in a GLM: (1) ‘across-category fatigue’ (comparison of task performance in the first half of session compared with the second half of session) and (2) ‘within-category decline’ (comparison of task performance in the first compared with the second half of each category) in a fully factorial model, including RT as a covariate of no interest. A parallel analysis using repeated-measures ANOVA revealed the same pattern of results; see Table 2 for Wald χ2, F and p values.
Mean response sensitivity (d’) in Experiment 2 (taxonomic-matching), for the first and second half of each category (within-category fatigue) and across the testing session (across-category fatigue).
Summary of significant results for response sensitivity from GLM and repeated-measures ANOVA analysis, examining effects of across-category and within-category changes in performance, in Experiment 2: taxonomic-matching.
GLM: generalised linear model; ANOVA: analysis of variance; RT: response time; df: degrees of freedom.
Two parallel analyses employing (1) mixed-effects modelling (i.e., GLM preserving performance information for each category for each participant and treating participants as a random effect – this allowed RT per category to be included as a covariate of no interest) and (2) ANOVA.
Bold indicates significant effects.
The GLM analysis found a significant main effect of within-category decline, indicating poorer performance towards the end of each category. There was little evidence that performance changed across the experiment (i.e., no across-category effect). There was also no significant interaction of within-category and across-category decline.
Summary of Experiment 2
Taxonomic categorisation was easier overall than the thematic judgements used in Experiment 1 (as reflected in larger d′ scores). However, a similar pattern of declining performance within each category was observed. Since performance did not decline across categories, this effect did not reflect a general difficulty in sustaining attention to the task.
Experiment 3: specific feature-matching in healthy participants
Rationale
Experiment 3 investigated whether the within-category semantic decline observed in Experiments 1 and 2 would occur when items were categorised on the basis of single feature, such as colour (e.g., ‘post-box’, ‘tomato’ and ‘Santa’ for the category RED). Feature-matching is a demanding semantic task that requires executive resources to focus semantic retrieval on the feature relevant to the task, and away from dominant aspects of knowledge (Badre et al., 2005; Jefferies, 2013): targets in this experiment were not globally related to the category being probed, and shared few (if any) features, except for the feature specified in the instructions (e.g., the targets ‘pancakes’, ‘blackboard’, ‘postcard’ share the feature FLAT but are not globally related). If the requirement to maintain a narrow focus of conceptual retrieval underpins the pattern of deteriorating categorisation, these effects would be expected to be maintained in this experiment. If, in contrast, strong global semantic relationships between target items are necessary for within-category decline in performance, this effect should be reduced in magnitude or even eliminated in this experiment.
Method
Participants
There were 24 participants (18 females and 6 males) recruited from the University of York, in exchange for course credit or payment of £5. Mean age of the students was 19 years (range: 18-24 years). All participants were native English speakers.
Task and procedure
The paradigm was similar to that in Experiment 2, except categorisation was based on a specific feature of the presented items. For example, participants were shown a category, such as
Results
Results are shown in Figure 3. This experiment assessed the effects of two within-subjects factors in a GLM: (1) ‘across-category decline’ (comparison of task performance in the first half of the experiment compared with the second half) and (2) ‘within-category decline’ (comparison of task performance in the first compared with the second half of each category). These fixed effects were entered into a fully factorial model. A parallel analysis using repeated-measures ANOVA was conducted on sensitivity, which yielded similar results (see Table 3 for Wald χ2, F and p values).
Mean response sensitivity (d′) in Experiment 3 (feature-matching), for the first and second half of each category (within-category fatigue) and across the testing session (across-category fatigue).
Summary of significant results for response sensitivity from GLM and repeated-measures ANOVA analysis, examining effects of across-category and within-category changes in performance, in Experiment 3: feature-matching.
GLM: generalised linear model; ANOVA: analysis of variance; RT: response time; df: degrees of freedom.
Two parallel analyses employing (1) mixed-effects modelling (i.e., GLM preserving performance information for each category for each participant and treating participants as a random effect – this allowed RT per category to be included as a covariate of no interest) and (2) ANOVA.
Bold indicates significant effects.
There was a significant main effect of within-category decline: performance was better at the beginning than at the end of each category. There was no main effect of declining performance across categories (p > 0.1). There was a significant across-category by within-category interaction, indicating a greater decline in performance ‘within’ each category towards the end of the experiment (see Figure 3). This was supported by Bonferroni t-tests, which indicated a highly significant within-category decline for the second half of the session: t(23) = 4.40, p < 0.001, but not for the first half of the session (p > 0.1).
Summary of Experiment 3
This experiment provided a second replication of the cumulative decline in categorisation performance within categories, in the absence of item repetition; however, in this case, the pattern was only apparent in the second half of the session. The feature-based classification task used in Experiment 3 was more demanding than the thematic and taxonomic categorisation tasks used in Experiments 1 and 2, and the within-category decrease in performance on this executively demanding semantic task might have been maximal towards the end of testing session, when cognitive control was likely to be lower. Most importantly, this experiment shows that the within-category decline effect extends to situations in which there is not a strong global relationship between the targets. The effect might therefore not emerge from strengthening activation in sets of globally related concepts, but might instead reflect interactions between semantic goal representations (i.e., targets are ‘thin’ or ‘red’ or ‘round’) and the conceptual store.
Experiment 4: within-category decline in categorisation across modalities in healthy participants
Rationale
Experiment 4 considers whether the systematic decline in the meaning of an item occurs in a manner that is independent of a specific modality. Declining performance in the cyclical matching paradigm in patients with aphasia has largely been documented using verbal comprehension tasks – i.e., word–picture matching (Cipolotti & Warrington, 1996; Jefferies et al., 2007; Schnur et al., 2006; Warrington & McCarthy, 1983). It has been suggested that this effect may be exclusive to auditory or verbal materials (Crutch & Warrington, 2008; Warrington & Crutch, 2004). Similarly, in healthy participants, declining performance on cyclical paradigms has been linked to lexical competition during speech production (rather than conceptual retrieval) (e.g., Belke et al., 2005; Harvey & Schnur, 2015; Howard et al., 2006), while in the satiation literature, it has been suggested that declining comprehension comes about due to adaptation of orthographic-to-semantic links (Tian & Huber, 2010) – consequently, these effects of repetition may be restricted to the verbal domain. Nevertheless, SA patients show declining performance across cycles for both word–picture and picture–picture matching tasks (Forde & Humphreys, 2007; Gardner et al., 2012), suggesting that semantic access deficits can occur at a conceptual level, and similar results were obtained recently in healthy participants (Wei & Schnur, 2015). This is consistent with the proposal that semantic cognition draws on heteromodal representations and control processes that operate across modalities. This study characterised the decline in performance for word targets, picture targets and an interleaved condition in which related items were presented as both words and pictures on different trials. If the decline in performance arises at a conceptual level, this effect should not be diminished for the interleaved condition. This pattern would allow ruling out accounts of within-category decline that involve fatigue/adaptation or competition within lexical-level representations.
In this experiment, the PSST paradigm was also modified: participants were asked to make a response on each trial (pressing one of two buttons to indicate whether the item was a member of the category, or not), and there were equal numbers of targets and distractors. This 2AFC design minimised the effects of response bias (relative to the paradigm used above, in which participants only responded when a target was present) and, most importantly, allowed us to characterise within-category changes in performance in terms of not only hits but also correct rejections, to examine whether participants were updating their working definition of the category over the set of trials. If participants showed a similar within-category reduction in accuracy for both targets and non-targets, it would suggest a reduced ability to retrieve the relevant information; in contrast, if they only showed a change for targets, it could suggest a narrower definition of the category is being acquired as the category progresses (i.e., a shift in response criteria).
Method
Participants
A total of 24 participants, native English speakers aged between 18 and 30 years, were recruited from the University of York in exchange for course credit or a payment of £5.
Task and design
The target categories in this experiment were thematic, in line with Experiment 1. Within-category manipulations involved (1) stimulus modality (auditory words, pictures and interleaved auditory words and pictures) and (2) within-category position (first half of each category was compared to the second half). The possibility of declining performance across categories was not examined in this experiment since performance would have been influenced by the order in which the three modality conditions were presented. Strength of association ratings were also used to split the verbal targets into strong and weak. The effect of this factor is reported in a cross-experiment comparison below.
Materials
There were 30 categories, 10 per modality (words, images, interleaved). Each participant saw each category only once (in one of the three modality conditions). There were 40 items in each category (20 related and 20 unrelated to the category). In the visual condition, the images were colour photographs on a white background. In the auditory condition, the word stimuli were audio files (see Appendix D in the Supplemental Material for a complete list of items used).
Procedure
The experiment was presented using E-Prime 2.0 (Psychology Software Tools). Participants completed 15 practice trials (five trials for each modality), before proceeding to the experimental trials. At the start of each category block, the category label was written on screen until the participant pressed the spacebar to continue. Pilot testing indicated that participants were unable to perform the interleaved condition at 1.1 s, so a slightly slower presentation speed of 1.3 s was adopted. Each item was presented for 1.3 s and this was the deadline for responding. Participants pressed one of two buttons to indicate whether the item was related or unrelated to the target category. The order of stimulus modality, categories and items within each category was fully counterbalanced across participants (although in the interleaved condition, a spoken word was followed by a picture).
Results
Results are shown in Figure 4. The GLM included modality (words vs pictures), interleaving (single modality vs interleaved words/pictures) and within-category decline, as fixed within-subjects effects and controlled for RT as a covariate. Parallel analysis using repeated-measures ANOVA on response sensitivity obtained similar results (see Table 4 for Wald χ2, F and p values).
Mean response sensitivity (d′) in Experiment 4 (effect across modalities), shown individually for the pictures and words modality in the interleaved and non-interleaved conditions, in the first and second half of each category (within-category decline).
Summary of significant results for response sensitivity from GLM and repeated-measures ANOVA analysis, examining effects of modality and interleaving, plus within-category changes in performance, in Experiment 4: cross-modality alternative-forced-choice decisions.
GLM: generalised linear model; ANOVA: analysis of variance; RT: response time; df: degrees of freedom.
Two parallel analyses employing (1) mixed-effects modelling (i.e., GLM preserving performance information for each category for each participant and treating participants as a random effect – this allowed RT per category to be included as a covariate of no interest) and (2) ANOVA.
Bold indicates significant effects.
There was a main effect of interleaving: lower performance for interleaved versus non-interleaved trials. There were no other main effects (p > 0.1). There was an interaction between within-category decline and interleaving: the cumulative decrease in performance ‘within’ each category was larger for interleaved than for non-interleaved trials. This was supported by post hoc tests, which indicated a significant decline for the interleaved condition, t(21) = 2.17, p = 0.042, but not for the non-interleaved trials (p > 0.1). There was also a significant interaction between modality and interleaving conditions, reflecting a greater effect of interleaving for pictures than words. Bonferroni-corrected t-tests indicated a highly significant effect of interleaving in the pictures modality, t(21) = 3.20, p = 0.004, but not in the words modality (p > 0.1). Other interactions were not significant (p > 0.1).
Summary of Experiment 4
In an alternative design using 2AFC decisions, a within-category decline was observed for both words and pictures, demonstrating that this effect does not reflect changes within lexical-semantic representations. Instead, within-category decreases in categorisation appear to reflect processes at a conceptual level. The within-category decline effect was also more marked for interleaved blocks, containing both word and picture targets, relative to single-modality blocks, suggesting that the effect can accumulate across these inputs. Interleaved blocks were more difficult, presumably because of the greater need for attentional control and switching: this may explain why the magnitude of within-category decline was strongest in this condition. Whether within-category decline interacts with the availability of control resources is tested in Experiment 5. The current experiment also demonstrated a reduction in performance for both target and non-target trials (i.e., an increase in both misses and false alarms, see Supplemental Material). This result suggests that the within-category decline effect is not a change in participants’ criteria for category membership. Instead, it is more likely to occur because of an increasing inability to identify the targets and reject the non-targets.
Experiment 5: effect of divided attention on within-category decline in healthy participants
Rationale
Research suggests that semantic cognition involves an interaction between conceptual representations and control processes that focus retrieval on currently relevant aspects of knowledge (Jefferies, 2013; Lambon Ralph et al., 2016). This type of control over retrieval may be partly achieved by domain-general executive mechanisms (although there might also be neurocognitive mechanisms that support semantic or memory control specifically; Davey et al., 2016; Noonan et al., 2013). Previous research has already shown that the requirement to perform a secondary task concurrently with semantic retrieval disrupts access to non-dominant knowledge (Almaghyuli, Thompson, Lambon Ralph, & Jefferies, 2012). Thus, if within-category decline reflects an increase in either competition or difficulty retrieving targets following long-term inhibitory weight changes, control mechanisms that can resolve competition or promote weak but currently relevant information should become more important towards the end of each category. Under these circumstances, the requirement to do two tasks at once might have a particularly detrimental effect on comprehension later in the categories.
The neuropsychological literature already points to the importance of control processes since patients with semantic access deficits (who show semantic blocking effects on cyclical word–picture matching tests) tend to have damage to left prefrontal cortex (PFC) and problems focussing retrieval on currently relevant information that correlates with general executive dysfunction (Campanella et al., 2009; Gardner et al., 2012; Jefferies et al., 2007; Jefferies & Lambon Ralph, 2006; Schnur et al., 2006; Schnur et al., 2009; Thompson et al., 2015). There are also several findings from the previous experiments reported here that suggest a role for control processes in within-category decline in categorisation. The interaction of within- and across-category decline seen in Experiment 3 but not the other experiments could reflect the importance of executive resources for the difficult feature-matching task, particularly at the end of each category (due to competition and/or weight changes). Within-category decline was also greater in the interleaved condition in Experiment 4, which required participants to switch between input modalities and may have had higher control demands. To directly test the importance of executive resources, Experiment 5 examined within-category decline in the PSST with and without the requirement to perform a secondary task.
Method
Participants
A total of 24 undergraduate students (20 females and 4 males) from the University of York took part in exchange for course credit or a payment of £5. The mean age of the students was 20 years (range: 18-30 years). All participants were native English speakers.
Task design
This experiment used a thematic category matching task (as for Experiment 1) and manipulated: (1) condition (single or dual) and (2) strength of association (strongly versus weakly associated targets). The effects of within-category and across-category decline were also examined. In the single-task condition, participants were asked to press a button when they detected targets that related to the category (identical to Experiment 1). In the dual-task condition, participants performed the same semantic task, except this time they were also asked to count triangles that appeared on the screen over the course of the category and report this number at the end.
Materials
All stimuli were taken from Experiment 1. A total of 20 thematic category labels were used (CHURCH, AIRPORT, and MUSIC FESTIVAL) with 60 spoken items presented in each category. Twenty items were related to the category, including 10 targets that were strongly related (e.g., CHURCH – ‘priest’) and 10 that were distantly related to the category (e.g., CHURCH – ‘bread’), while the remaining 40 items were unrelated to the category (e.g., CHURCH – ‘oyster’) – these were recycled from other categories (see Appendix A in the Supplemental Material for a complete list of items used). Ten categories were combined with the secondary task, while 10 were presented under single-task conditions. In the dual-task condition, patterned triangles were presented on the screen (with 30 triangles appearing overall, 15 triangles distributed in the first half and 15 triangles in the second half of the session). In order to minimise the difficulty of the dual condition, only two to four triangles were presented per category, and these appeared on randomly selected trials.
Procedure
Each session began with three practice blocks – the first block involved categorising spoken words (i.e., single-task condition); the second block involved presentation of triangles (without any auditory stimuli), participants were asked to count the triangles that appeared from time to time on a blank screen and write down the number they had seen; and the third practice block combined the two tasks. There were three categories with 30 trials (5 strongly related, 5 weakly related and 20 unrelated items) in each of the practice blocks. After the practice blocks, participants were presented with 20 experimental categories, using an ABBA or BAAB design for the single and dual-task conditions. The order of conditions, categories and items was fully counterbalanced between subjects.
Results
The results of Experiment 5 are presented in Figure 5. GLM analysis in this experiment included secondary task condition (single vs dual task), relatedness, across-category and within-category as fixed within-subjects effects and included RT as a covariate. Parallel analysis using repeated-measures ANOVA on the sensitivity data indicated similar results (see Table 5 for Wald χ2, F and p values).
Mean response sensitivity (d′) in Experiment 5 (effect of divided attention), shown individually for the strong/weak targets, in the two conditions (single/dual), and split by first and second half of each category (within-category fatigue) and across the testing session (across-category fatigue).
Summary of significant results for response sensitivity from GLM and repeated-measures ANOVA analysis, examining effects of condition (single/dual), relatedness, plus within-category changes in performance, in Experiment 5: effect of divided attention.
GLM: generalised linear model; ANOVA: analysis of variance; RT: response time; df: degrees of freedom.
Two parallel analyses employing (1) mixed-effects modelling (i.e., GLM preserving performance information for each category for each participant and treating participants as a random effect – this allowed RT per category to be included as a covariate of no interest) and (2) ANOVA. Other interaction terms were non-significant (p > 0.1). Bold indicates significant effects.
The analysis revealed significant main effects of secondary task condition (reduced sensitivity in the dual task compared to the single task), relatedness (lower sensitivity for the weak than strong targets) and a marginal main effect of across-category decline (sensitivity scores declined overall from the first half to the second half the session). There was no main effect of a within-category decline, but there was a significant interaction between relatedness and within-category performance: the decline in categorisation was greater for strong than for weak items. This was supported by post hoc tests with Bonferroni correction that showed highly significant within-category decline for strong items, t(23) = 7.41, p < 0.001, and near-significant decline for weak items, t(23) = 1.99, p = 0.058.
There was a trend-level three-way interaction between dual-task condition, relatedness and within-category decline. This was explored by analysing performance for the strong and weak targets separately. The dual task by within-category interaction was significant for strong items, t(23) = 2.83, p = 0.009, but not for the weak items (p > 0.1). Other interactions were not significant (p > 0.1).
Summary of Experiment 5
The strong items showed a substantial effect of within-category decline, especially under dual-task conditions, i.e., divided attention increased the decline in categorisation seen towards the end of each category. Thus, this experiment provides tentative support for the view that a reduction in executive control in healthy young volunteers increases the effects of within-category decline. Executive resources might allow participants to overcome competition from strongly activated distractors or the effects of retrieval-induced inhibition (i.e., the suppression of related items on previous trials).
Cross-experiment analysis of effects of within-category decline in healthy participants
Declining comprehension within and across categories
Given the similar structure of several of the experiments above, a meta-analysis was conducted examining the magnitude of both within-category decline (i.e., performance in the first and second half of each category) and across-category decline (i.e., performance in the first and second half of each experiment). All experiments employing a 1.1-s presentation speed were included. This included the fast presentation condition of Experiment 1, Experiments 2 and 3, plus the single-task condition from Experiment 5. This analysis collapsed across strong and weak targets in Experiments 1 and 5. GLM analysis included across-category and within-category as within-subjects fixed effects. Experiment was included as a between-subjects factor to establish whether the magnitude of within- or across-category decline varied across these experiments. RT per condition and participant was again used as a covariate (see Table 6 for results). There was a main effect of experiment: the taxonomic categorisation task (Experiment 2) was easier than thematic-matching (Experiment 1), specific feature-matching (Experiment 3) and the single-task condition in Experiment 5. There was a significant main effect of within-category decline across these four experiments (see Figure 6). However, there was no overall change in performance across experiments (i.e., no across-category decline) and no significant interactions (p > 0.1). A decline in performance was revealed towards the end of each category across all of these experiments.
Summary of significant results from GLM and repeated-measures ANOVA analysis, examining across-category and within-category changes in performance across Experiments 1 (thematic-matching), 2 (taxonomic-matching), 3 (feature-matching) and 5 (effect of divided attention, single condition).
GLM: generalised linear model; ANOVA: analysis of variance; RT: response time; df: degrees of freedom.
Two parallel analyses employing (1) mixed-effects modelling (i.e., GLM preserving performance information for each category for each participant and treating participants as a random effect – this allowed RT per category to be included as a covariate of no interest) and (2) ANOVA. Experiment was included as a between-subjects factor. Bold indicates significant effects.
Mean response sensitivity (d′) shown individually for the first and second half of each category (within-category fatigue) and across the testing session (across-category fatigue), for Experiments 1-3 and 5 (single condition), at the presentation speed of 1.1 s.
To further characterise the way in which performance changed across successive items within categories, the average number of hits and false alarms was computed for each individual item position across these experiments. Each category included 60 items (20 targets and 40 non-targets, which were presented in different positions across categories and participants). From these data, we computed the average percentage of hits and false alarms individually for each position (reported in Figure 7). The analysis revealed a largely continuous decline in hits, with no substantial increase in false alarms, consistent with the analyses presented above (see Figure 7). Item position correlated with the percentage of hits across the four experiments; this effect was greatest in Experiment 3 (r = 0.636, p < 0.001), followed by Experiments 1 and 5 (r = 0.559, p < 0.001; r = 452, p < 0.001, respectively) and Experiment 2 (r = 0.291, p = 0.024).
Average (1) percentage of hits and (2) false alarms for items within each category and across participants in Experiments 1 (thematic-matching), 2 (taxonomic-matching), 3 (feature-matching) and 5 (divided attention: single condition).
Interactions with strength of association
In an additional analysis across experiments, the relationship between within-category decline and the strength of association between stimuli and target categories was examined. This analysis included data from Experiment 1 (strong vs weak targets, collapsed across the two speeds – 2 and 1.1 s), Experiment 4 (strong vs weak word targets, collapsed across interleaved and non-interleaved conditions) and Experiment 5 (strong vs weak targets, collapsed across the single and dual conditions). The picture condition from Experiment 4 was omitted since verbal measures of strength of association may not apply to picture-based decisions.
A GLM examining response sensitivity included relatedness and within-category as fixed effects within-subjects, Experiment as a between-subjects factor and RT as a covariate (see Table 7 for results and parallel analysis using repeated-measures ANOVA). The analysis revealed a main effect of relatedness: overall sensitivity scores were lower for weak compared to strong targets (see Figure 8). There was a main effect of within-category decline: sensitivity declined towards the end of each category across the three experiments. Importantly, there was a significant interaction between relatedness and within-category decline. Post hoc tests with Bonferroni correction indicated a significant within-category decline for strong targets, t(69) = 3.83, p < 0.001, but not for weak associations (p > 0.1). Thus, performance declined more substantially for strongly than for weakly related targets across experiments.
Summary of significant results from GLM and repeated-measures ANOVA analysis, examining relatedness and within-category performance across Experiments 1 (thematic-matching), 4 (words modality) and 5 (effect of divided attention).
GLM: generalised linear model; ANOVA: analysis of variance; RT: response time; df: degrees of freedom.
Two parallel analyses employing (1) mixed-effects modelling (i.e., GLM preserving performance information for each category for each participant and treating participants as a random effect – this allowed RT per category to be included as a covariate of no interest) and (2) ANOVA. Experiment was included as a between-subjects factor. Other interaction terms were non-significant (p > 0.1). Bold indicates significant effects.
Mean response sensitivity (d′), shown individually for the first and second half of each category (within-category fatigue) and split by strong and weak trials, for Experiments 1, 4 and 5.
There was also a significant interaction between task and relatedness: there was a stronger effect of relatedness in Experiment 5 (which involved divided attention) and in Experiment 4 (which involved interleaved presentation), in comparison with Experiment 1 (a simpler thematic-matching task). Post hoc tests with Bonferroni correction indicated a significant effect of relatedness in all experiments, which was largest in Experiment 5, t(23) = 21.38, p < 0.001, followed by Experiment 4, t(21) = 8.28, p < 0.001, and smallest in Experiment 1, t(23) = 6.41, p < 0.001. In this way, the effects of strength of association appeared to be greater in more executively demanding paradigms.
Experiment 6: retrieval-induced changes in categorisation in patients with SA
Rationale
The findings so far show that young adults find it progressively harder to rapidly categorise items drawn from a single category. This effect was greater for strong associations, suggesting it resulted from spreading activation or inhibition between related items. It was also increased by divided attention, suggesting that control may be employed to overcome the harmful effects of this spreading activation/inhibition. However, the data do not adequately discriminate between theoretical accounts which propose that declines in categorisation reflect (1) an increasing need for controlled retrieval later in each category – following, for example, the activation of a broad semantic field (giving rise to strengthening competition) and (2) the unintended consequences of controlled retrieval earlier in the list (e.g., if semantically related but irrelevant information is inhibited on earlier trials, and then becomes relevant). By the first account, in a category such as
Using the PSST paradigm, we characterised changes in semantic cognition in terms of (1) the effect of strength of association (strong vs weakly related targets); (2) within-category decline – e.g., contrasting the beginning and end of each category; and (3) across-category decline – e.g., general cognitive fatigue which might produce deteriorating performance on the paradigm over the course of each testing session. The patients were expected to be disproportionately impaired at retrieving weak as opposed to strong associations, in line with previous findings (Jefferies & Lambon Ralph, 2006; Noonan et al., 2010). A secondary research question concerned the extent to which the patients would show cognitive fatigue across the session on a task requiring sustained semantic attention.
Method
Participants
Patients
A total of 12 stroke aphasic patients (eight females, four males) were recruited from stroke clubs and speech and language therapy services in York and Leeds, UK. All patients had chronic impairment after a cardiovascular accident (CVA) at least 1 year prior to testing. Patients were aged between 40 and 78 years, with a mean age of 62 years (standard deviation [SD] = 10.2). The patients were selected to show semantic impairment (Table 8). Seven cases met the definition of SA used by Jefferies and Lambon Ralph (2006) – i.e., multimodal deficits affecting comprehension tasks employing both words and pictures. The other cases were more mildly impaired; nonetheless, all cases were impaired at a demanding synonym judgement test, and they all had difficulty comprehending ambiguous words with multiple meanings, particularly following a miscue that primed the irrelevant interpretation of the word (Noonan et al., 2010). This suggests the patients’ deficit was in controlling semantic retrieval such that the information being accessed was appropriate to the demands of the task: performance was relatively preserved when the task demands aligned with dominant aspects of knowledge, but were more impaired when the task required more unusual aspects of knowledge to be brought to the fore.
Background neuropsychological data for each patient.
NA: not available.
Patients are arranged according to composite semantic severity scores; this is a single factor extracted from WPM (word picture matching), CCT (Camel and Cactus Task) (both from Bozeat, Lambon Ralph, Patterson, Garrard, & Hodges, 2000) and synonym judgement. RCPM: Raven’s Coloured Progressive Matrices (Raven, 1962); BSRA: Brixton Spatial Attainment Task (Burgess & Shallice, 1997); PALPA: Psycholinguistic Assessments of Language Processing in Aphasia (Kay, Lesser, & Coltheart, 1992). Cookie theft description assesses fluency (words-per-minute; Goodglass & Kaplan, 1983).
Impaired performance.
Older controls
A total of 15 older adults, with a mean age of 73 years (SD = 8.1 years; range: 55-84 years) were selected from a participant database at the University of York (10 females, 5 males). They were selected to provide an age-matched control group for our patient sample below. Participants had no prior history of brain injury and showed unimpaired cognitive functioning on the Mini-Mental State Examination (Folstein, Folstein, & McHugh, 1975), see Table 8 for their average performance on background semantic and executive assessments. They left school or college at an average of 18 years (SD = 3.9 years).
Task and design
This experiment used a thematic category matching task (as in Experiments 1 and 5), i.e., categories such as
Materials
Stimuli were taken from Experiment 1. A total of 20 different thematic category labels were used (such as
Procedure
The experiment was presented using E-Prime 2.0 (Psychology Software Tools). Testing was completed across two sessions. Category names were presented as written words that remained visible throughout the block to reduce demands on working memory. Participants were asked to press a button each time they heard a word that was related to the category, and not to press for unrelated words. The session began with 20 min of PSST testing (i.e., 5 categories out of the 20) at a presentation speed of 2 s. This was followed by 15 min of neuropsychological testing, followed by another 20 min of PSST testing (i.e., next five categories). The remaining 10 categories were presented in the second session, which followed the same procedure.
Results
Our main dependent measure was response sensitivity, since we aimed to evaluate the effect of SA on the ability to detect targets belonging to each category, by the deadline imposed by the task. As before, we used (generalised) linear models (GLMs) for all analyses and entered RT as a covariate (i.e., the average speed for correct responses per condition per participant) in the analyses examining response sensitivity.
We examined the effects of three within-subjects factors in a GLM: (1) relatedness (targets with a strong or weak association with the category); (2) set (comparison of task performance in the first half of each session compared with the second half of each session, to assess the possibility of general cognitive fatigue); and (3) within-category position (comparison of task performance in the first half compared with the second half of each category), plus (4) group as a between-subjects factor. These predictors were entered in a fully factorial model, including RT as a covariate. We also modelled category number (i.e., performance for each of the five categories within each set) and session number (i.e., Day 1 or Day 2 of testing) without interaction terms, to capture these aspects of the design. The model elicited a significant main effect of relatedness (see Table 9 for all Wald χ2 and p values) and a significant two-way interaction between group and relatedness (see Figure 9): sensitivity was lower for weak items in comparison with strong items across both groups but separate analyses split by group showed that relatedness had a larger effect on performance for the patients, Wald χ2(1) = 101.43, p < 0.001, than for the controls, Wald χ2(1) = 34.64, p < 0.001. The patients showed poorer performance, relative to controls, for weak associations.
Summary of significant results from the GLM analysis for SA patients and age-matched controls – looking at the effects of group, relatedness, set and within-category performance, for our key-dependent measures – response sensitivity, response accuracy and response times.
GLM: generalised linear model; ANOVA: analysis of variance; RT: response time; df: degrees of freedom.
Analyses employing mixed-effects modelling for response sensitivity (i.e., GLM preserving performance information for each category for each participant and treating participants as a random effect – this allowed RT per category to be included as a covariate of no interest). Bold indicates significant effects.
Mean response sensitivity (d′) for the first and second half of each category (within-category performance), split by strong and weak targets for patients and controls.
There was no main effect of within-category performance, there was however a significant interaction between relatedness and within-category performance: there was greater within-category decline in performance for the weak targets and a subtle improvement in performance for the strong items towards the end of each category (see Figure 9). This interaction was significant in both groups – controls: Wald χ2(1) = 18.98, p < 0.001 and patients: Wald χ2(1) = 8.718, p = 0.003 – and contrasted with the pattern found for healthy young volunteers across multiple experiments above.
Summary of Experiment 6
The results showed that patients with SA (relative to age-matched controls) were less efficient at retrieving weak associations, relative to stronger ones: in this way, the patients showed the expected pattern of poorer performance on trials requiring more controlled retrieval of semantic information. In addition, neither the patients nor the controls showed the within-category decline for strong associations seen repeatedly in healthy young participants. Since participants with deficits of semantic control did not show an increase in within-category decline in categorisation, our findings do not support the hypothesis that this decline in the PSST reflects an increasing need for controlled retrieval towards the end of each category – e.g., following the activation of a broad semantic field which would give rise to strong competition from semantically related items. Instead, within-category decline in comprehension might be minimised in both patients with SA and healthy older volunteers if this effect has its origins in the retrieval-induced suppression of semantically related information on earlier trials. The flexible suppression of semantically linked but non-target information might be an efficient way for younger participants to efficiently categorise targets, but this approach would be unhelpful or impractical for older volunteers and patients with a reduced capacity to control semantic retrieval. We note that, given some subtle changes in the design of this experiment (i.e., categories were tested across two sessions), we were not able to directly assess the effects of age on performance. We cannot exclude the possibility that the experiment lacked the sensitivity to recover differences between controls and patients, but the results do suggest that patients with deficits of controlled retrieval are not highly vulnerable to effects of within-category decline on the PSST. This observation further suggests that declining comprehension in the PSST has different origins from refractory effects in cyclical word–picture matching, since patients with SA have been shown in previous studies to have increasing impairment across cycles in which the same stimuli are presented repeatedly (Gardner et al., 2012; Jefferies et al., 2007; Thompson et al., 2015). Potential reasons for this apparent difference between paradigms are considered below.
Discussion
Retrieval-related declines in categorisation were first assessed in healthy young adults across five experiments. Results demonstrated a cumulative decline in semantic categorisation as a consequence of sustained semantic retrieval: participants’ ability to detect targets belonging to a particular category deteriorated even in the absence of item repetition/satiation. This effect was not equivalent to time-on-task and could not be explained as a general decline in sustained attention as a result of fatigue, since many categories were tested back-to-back over the course of the experiments and there was a release from this phenomenon at category boundaries, when targets were no longer related to recently categorised targets. The effects were observed in two different paradigms: both a vigilance paradigm, in which participants attempted to detect targets that were less frequent than distracters, and also in a 2AFC paradigm, where target and distracter items within and outside of the target category were presented equally often. This confirmed that healthy young participants were less able to categorise accurately towards the end of each category and were not simply changing their response criteria following more experience with each category. The effect was largest for targets strongly related to the category, supporting the suggestion that this phenomenon is linked to spreading activation or inhibition between related concepts. However, it did not require a global semantic relationship between the targets within the category: within-category decline was seen across taxonomic, thematic and individual feature-based classification, suggesting it may be a fairly ubiquitous consequence of sustained semantic retrieval, at least in circumstances such as those created by the PSST paradigm, where the focus of retrieval is pre-defined. The effect was also multimodal (extending to a paradigm in which semantically related pictures and words were interleaved), confirming that it is conceptual and not lexical in origin. Finally, it was increased by conditions of divided attention, suggesting that the capacity to maintain semantic retrieval within a category can be increased through the allocation of executive control.
The second part of this study examined how SA might influence performance on the continuous categorisation task. Patients with SA showed deficits in controlled retrieval (i.e., they maintained close-to-normal performance for strong associations, compared to age-matched controls, but had additional difficulties identifying weakly associated targets). Categorising weakly related items is thought to require greater semantic control, since there are fewer overlapping features for such items, and dominant yet currently irrelevant features potentially have to be suppressed in order to allow weakly instantiated knowledge to come to the fore (Noonan et al., 2010; Whitney, Kirk, O’Sullivan, Lambon Ralph, & Jefferies, 2010). For example, to recognise that
Our results add to a growing body of work showing that conceptual processing can become less efficient following the retrieval of semantically related items. The within-category effect in younger adults resembled the less efficient semantic retrieval that is seen when a sequence of related items are categorised (e.g., Campanella & Shallice, 2011; Harvey & Schnur, 2015; Wei & Schnur, 2015) – however, it occurred without the massed repetition and habituation of individual items common across these previous studies. The paradigm resembled a continuous picture-naming task since items related to a category were interspersed with other items (Belke, 2013; Belke & Stielow, 2013; Howard et al., 2006; Kleinman, 2013; Navarrete et al., 2010; Oppenheim et al., 2010; Runnqvist et al., 2012; Schnur, 2014); yet it involved comprehension and categorisation rather than requiring speech output. Previous work examining categorisation in continuous paradigms has found facilitation (Belke, 2013; Riley et al., 2015). However, we were able to elicit declining performance by adapting this paradigm to require goal-driven categorisation, where the goal changed across successive blocks. We speculate below that this requirement to selectively focus retrieval on specific aspects of knowledge might have been critical to the effects we observed. The within-category decline in categorisation was also stronger when rapid processing was required. This resembles Campanella and Shallice’s (2011) finding of deteriorating comprehension for repeated items in semantically related sets when healthy participants had to respond by a deadline. Sensitivity to the speed of response suggests that a process that takes time (such as the application of control to boost previously suppressed representations) is essential to avoiding the detrimental effects of continuous categorisation.
There has been considerable discussion of (1) whether cumulative interference occurs in comprehension paradigms (or whether these effects are restricted to picture naming); (2) whether such effects arise at the level of lexical or semantic representations; and (3) whether these effects are short-lived (reflecting interference from ongoing activation of semantically related items) or are long-lasting (reflecting weight changes or adaptation within the underlying representations). Our data speak to all of these issues. First, within the literature on semantic access impairment, there has been debate about whether declining comprehension is restricted to auditory-verbal materials (Crutch & Warrington, 2008; Warrington & Crutch, 2004; Warrington & McCarthy, 1983) or whether it extends to non-verbal tasks (Forde & Humphreys, 1997; Gardner et al., 2012; Thompson et al., 2015). Similarly, opposing psycholinguistic studies have argued that semantic interference effects reflect lexical processes (e.g., Damian et al., 2001) or alternatively conceptual processes that extend to picture matching tasks (Wei & Schnur, 2015). We reasoned that if within-category decline effects emerge from within modality-specific representations, this pattern should be weaker in an ‘interleaved’ condition involving both words and pictures, as there would be more time for recovery between successive items and/or fewer related targets presented within a modality to produce a decline in performance. In contrast, if these effects arise at a multimodal conceptual level, they should be strong even when inputs of different modalities are interleaved. We observed a within-category decline in comprehension for interleaved words and pictures, suggesting that our results are unlikely to reflect either effects at a lexical level or effects within the mappings between concepts and specific inputs (e.g., auditory word forms; structural descriptions of objects). We conclude this phenomenon originated within conceptual representations that are not specific to a particular input modality (as envisaged by Patterson, Nestor, & Rogers, 2007; Rogers et al., 2004; see also Wei & Schnur, 2015).
The semantic locus of the effect is also supported by our observation that the within-category decline in comprehension was most marked for strong associates of the target category: this suggests that spreading activation or inhibition between semantically related concepts is a necessary condition for deteriorating categorisation. This finding, revealed across multiple experiments in healthy young volunteers, also speaks to two different mechanisms that have been put forward to explain declining comprehension. First, the literature on cyclical word–picture matching and semantic ‘access’ deficits has associated deteriorating performance with increasing competition within a set of items that are repeatedly presented (Gardner et al., 2012; Jefferies et al., 2007; Schnur, Schwartz, Brecher, & Hodgson, 2006; Thompson, Henshall, & Jefferies, 2016). By this view (and in contrast to our results), we might expect weak category members to show deteriorating performance under increasing levels of competition, since weak targets are less able to withstand strong competition. An alternative perspective is provided by the literature on continuous naming, where it has been suggested that semantically related information is suppressed to facilitate target retrieval, and this effect makes semantically associated items harder to retrieve on subsequent trials (Oppenheim et al., 2010). By this view, strong associates of the category should be suppressed to a greater extent than weak associates: for example, in the category
Another theoretical debate concerns the underlying neurobiological mechanisms that give rise to deteriorating performance. While some studies (particularly those employing cyclical paradigms in which the same set of items are repeatedly presented at fast speeds) have assumed that ongoing activation of a set of competitors gives rise to declining performance (Campanella & Shallice, 2011; Schnur et al., 2006), research examining both continuous picture naming and retrieval-induced forgetting suggests that long-term weight changes underpin these effects. Continuous naming studies have manipulated the lag between semantically linked items and have found relatively long-lasting changes in accessibility that are unlikely to be explained by ongoing activation of conceptual information (Howard et al., 2006; Oppenheim et al., 2010). We conducted a lag analysis and obtained a similar pattern: categorisation was poorer in the second half of the category, particularly when there were lags of between one and three intervening items between successive targets. The experiment was not designed in a way that allowed us to examine longer lags, but targets that immediately followed a previous target without any intervening items did not show an effect of within-category decline, even though ongoing activation would have been at its strongest.
The findings from patients with SA and age-matched controls are also consistent with the view that within-category decline in healthy young adults followed from the inhibition of aspects of meaning that were irrelevant or unhelpful during the classification of specific targets earlier in the list. In the PSST, dominant concepts for the category goal (e.g.,
Our findings in patients with SA further suggest that declining comprehension in the PSST has different origins from refractory effects in cyclical word–picture matching, since these patients have been shown by previous studies to have increasing impairment across cycles in which the same stimuli are presented repeatedly (Gardner et al., 2012; Jefferies et al., 2007; Thompson et al., 2015). These differences might be explained by the specific demands of the two tasks. As noted in the Introduction, the PSST requires participants to decide whether each item fits within a specific category, which repeatedly changes over the course of the experiment – in this sense the task is more goal-driven than either naming or word–picture matching. In contrast, cyclical matching paradigms are stimulus-driven, but their structure might result in strong levels of competition, since they involve selecting items, suppressing them when the target changes and then re-selecting them: deficits of semantic control in patients with SA following damage to left inferior frontal gyrus are argued to produce increased difficultly in target selection in the face of this growing competition. In contrast, when the target for categorisation continuously changes in the PSST, competition from previous targets is minimised, and effects of the suppression of related non-target items on earlier trials might be increased. Patients may show an overall deficit on the PSST because they are unable to focus retrieval on the current category goal, but as a consequence they do not show deteriorating performance. Further research is needed to directly manipulate the extent to which continuous categorisation is goal-driven, to establish whether this factor is critical in creating the conditions for healthy young adults to show declining performance and for patients with semantic access deficits to be relatively immune to such declines in performance.
Evidence across the experiments in healthy young participants also suggests that within-category decline is ameliorated through the application of cognitive control, which can overcome the earlier suppression of items that have now become targets. In young adults, the effect of within-category decline was larger towards the end of more executively demanding paradigms, such as the specific feature-matching judgements in Experiment 3, perhaps because participants were no longer able or willing to constrain semantic retrieval in an effortful way to meet the demands of the task. Similarly, within-category decline was greater in the interleaved condition of Experiment 4, which required greater attentional control to switch between words and pictures. Experiment 5 directly manipulated the availability of executive resources through the use of a secondary task to divide attention, and this appeared to increase the magnitude of the within-category decline in categorisation for highly related items. Thus, it might be that the most substantial effects of within-category decline in goal-driven categorisation occur when (1) new targets are highly related to the specified goal for categorisation and are therefore associated with previous targets; (2) participants are young adults who use retrieval-induced suppression to facilitate efficient target retrieval; and (3) executive resources that could be used to control the negative effect of this previous retrieval are weak. In this way, our findings are readily related to contemporary accounts of semantic processing which envisage that heteromodal conceptual knowledge interacts with control processes to support context- and task-appropriate semantic retrieval (Jefferies, 2013; Lambon Ralph et al., 2016).
Our work shows that controlled retrieval of information pertaining to a specific topic is hard to sustain over time and increasing executive resources are required to overcome the negative effects of controlled retrieval on earlier trials. The phenomenon we have described would tend to promote an evolving pattern of retrieval that does not stay focussed on the same tight category but changes over time. Interestingly, this phenomenon is seen commonly in patterns of semantic retrieval that occur in everyday life, such as during mind-wandering and when conversations spontaneously shift topic (e.g., Humphries, Binder, Medler, & Liebenthal, 2006). In the real world, it might only be a matter of time before one of these alternative avenues becomes the focus of our retrieval.
Footnotes
Acknowledgements
The authors thank Woody Chau for his help with data collection.
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The patient work was supported by a grant from the Stroke Association (TSA/12/02). E.J. was supported by the European Research Council (SEMBIND – 283530). H.E.T. was supported by the Wellcome Trust (ref: 105624) through the Centre for Chronic Diseases and Disorders (C2D2) at the University of York. J.S. was supported by a grant from the European Research Council (Wandering Minds – 303701).
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
