1 INTRODUCTION

Categorization is one of the first things we learn when navigating the environment, constructing knowledge of the world. By categorizing objects and beings according to how similar they are, we form semantic representations and assign them names. Thus, when we speak, we constantly refer to things with different levels of semantic similarity. Semantic similarity denotes the degree of relatedness between two words, for instance, words that share more semantic features under the same taxonomy, such as horse (has legs; is an animal) and sheep (has legs; is an animal), are more semantically similar than words that share fewer features, for example, horse and shark (does not have legs; is an animal). Manipulating semantic similarity offers insights into conceptual preparation and lexical selection during speech production and can reveal the micro-structure of our semantic system.

The present study investigated effects of semantic similarity on different planning stages during language production and aimed to provide a fine-grained time frame of the effects. Until now, to our knowledge, there is no electrophysiological evidence directly associated with semantic similarity effects in the blocked-cyclic naming paradigm. To pursue high temporal resolution information, we employed event-related potentials (ERPs) in a blocked-cyclic naming paradigm. To foreshadow the results, we find evidence for semantic facilitation for the first naming cycle that was predicted by an amplitude modulation of the N1 component, followed by semantic interference for later cycles that was predicted by a posterior negativity in the time range of the P2/N2 component. These effects were present only for close, but not for distant, semantic relations, indicating that only semantically close objects induce observable semantic context effects. Furthermore, the N1 modulation was reduced but persisted in later cycles in which interference dominated, and the posterior negativity was also present in the first cycle in which facilitation dominated, providing first evidence for temporally overlapping conceptual and lexical processing in the blocked-cyclic naming paradigm.

1.1 Investigating semantic similarity in speech production paradigms

Semantic context effects serve as an index of lexical-semantic processing, which has been evidenced to take place around 200 ms after stimulus onset (e.g., Costa et al., 2009; Levelt, 1992; Levelt et al., 1999; Piai et al., 2014; Strijkers et al., 2010). Such effects emerge as facilitation or interference in different language production paradigms. To produce a word (e.g., horse), a speaker activates conceptual representations related to that word (e.g., mammal, fur, hooves), and the related concepts further activate one another's corresponding lexical representations. The most strongly activated target lexical representation (horse) is then selected from all the co-activated lexical candidates (e.g., sheep, camel) for articulation (Howard et al., 2006; Levelt et al., 1999; Roelofs, 1992, 2018). Based on models that assume lexical competition, the general mechanisms behind semantic context effects consist of two parts: conceptual priming and lexical competition. While related concepts facilitate the selection of a target lexical representation, co-activated lexical candidates compete with the target and thus disrupt selection (Abdel Rahman & Melinger, 2009, 2019; Belke et al., 2005; Roelofs, 2018). Since semantic context effects are assumed to require sufficient overlapping semantic features to emerge, enhancing semantic similarity should theoretically amplify the context effects. This hypothesis has been tested by studies applying different naming paradigms.

In the picture-word interference (PWI) paradigm (e.g., Damian & Bowers, 2003; Glaser & Düngelhoff, 1984; Glaser & Glaser, 1989; Hantsch et al., 2005; Hutson & Damian, 2014; Mahon et al., 2007; Piai et al., 2014; Rose et al., 2019; Schriefers et al., 1990; Vigliocco et al., 2004), participants are instructed to name a picture presented together with a superimposed distractor word which they should ignore. When a distractor word is categorically related to the picture, participants typically take more time to respond compared to pictures superimposed with an unrelated distractor word. The role of semantic similarity in lexical-semantic processing has been investigated with the PWI paradigm, but findings are inconsistent. While some studies found stronger interference for semantically close versus distant distractor words (e.g., Aristei & Abdel Rahman, 2013; Rose et al., 2019; Vigliocco et al., 2004), others found the opposite pattern (e.g., Mahon et al., 2007), and yet others found an overall effect of semantic relatedness, but no effect of semantic similarity (Hutson & Damian, 2014). Even when graded effects were observed behaviorally, electrophysiological evidence was only present for the semantically close condition (Rose et al., 2019), revealed as a posterior positivity around 200 ms post-stimulus; no electrophysiological evidence of semantic interference was found for distant distractor words despite the behavioral effects.

In the continuous naming paradigm (e.g., Belke, 2013; Belke & Stielow, 2013; Costa et al., 2009; Howard et al., 2006; Navarrete et al., 2010), participants name pictures in succession. Semantically related picture sets are interleaved with unrelated filler items. A robust effect of ordinal position, that is, a linear increase in naming latency as the number of named exemplars of a given semantic category increases, has been reported for semantically related items, implying increasing difficulty of lexical selection. With regard to the effect of semantic similarity, in the continuous naming paradigm, cumulative interference has been reported for semantically close, but not for semantically distant, items (Rose & Abdel Rahman, 2017). In the same study, the cumulative interference was positively associated with a posterior positivity around 250 ms post-stimulus in the participants' electroencephalogram (EEG) only in the semantically close condition—an electrophysiological signature whose time course and scalp distribution agree with previously reported ERPs for lexical selection (Costa et al., 2009).

In order to make results across paradigms more comparable, our study focuses on another commonly used paradigm, the blocked-cyclic naming task (also referred to as the blocking paradigm; e.g., Belke, 2008; Belke et al., 2005; Damian & Als, 2005; Damian et al., 2001; Kroll & Stewart, 1994; Navarrete et al., 2012, 2014; Schnur et al., 2006; Vigliocco et al., 2002) while using the same set of materials as the two studies mentioned above that investigated the semantic similarity effects (Rose & Abdel Rahman, 2017; Rose et al., 2019). In the blocking paradigm, participants name exemplars from a given stimulus set in small repeated cycles; in homogeneous blocks, exemplars are semantically related, whereas in heterogeneous blocks, exemplars are semantically unrelated. From the second cycle onward, longer naming latencies in the homogeneous compared to the heterogeneous condition have been reliably reported, while either no effect or facilitation has been found in the first cycle (e.g., Abdel Rahman & Melinger, 2011; Crowther & Martin, 2014; Janssen et al., 2015; Navarrete et al., 2012). While findings in the first cycle have been less consistent, a review of studies using the blocked-cyclic paradigm found that facilitation is likely to be observed in the first cycle if the semantic conditions are presented in large blocks instead of in alternating order (Belke, 2017). How semantic similarity influences the process of lexical retrieval has also been investigated using the blocked-cyclic naming paradigm (e.g., Navarrete et al., 2012; Vigliocco et al., 2002). Behavioral studies have reported graded effects, with stronger facilitation in the semantically close versus distant condition in the first cycle (e.g., Navarrete et al., 2012, Experiment 1 & 2), and stronger interference in the semantically close versus distant condition (e.g., Vigliocco et al., 2002, collapsing all repetition cycles). However, there is no electrophysiological evidence directly associated with semantic similarity effects in the blocked-cyclic naming paradigm.

In sum, three different naming paradigms have been applied to test the effect of semantic similarity on lexical retrieval. While close relations usually produce context effects, such effects induced by distant relations are often absent or significantly weaker. One of our goals is thus to provide more evidence with regard to the extent to which semantic similarity modulates context effects. Moreover, even though EEG has been recorded in the PWI and continuous naming paradigms to examine the modulation of semantic similarity, the electrophysiological evidence for such modulation is absent in the blocked-cyclic paradigm.

1.2 Theoretical explanations for semantic blocking effects

The semantic context effects in the blocked-cyclic paradigm, specifically referred to as semantic blocking effects, have been alternatively related to perceptual, conceptual, lexical, or post-lexical planning stages. The view that underpins the current investigation holds that semantic blocking effects, and indeed all semantic context effects, reveal a trade-off between conceptual and lexical processes, which temporally overlap. Specifically, the slower naming times observed in later homogeneous naming cycles result from the competitive nature of the lexical selection process1 (cf. Levelt et al., 1999; Roelofs, 1992, 2018). A Swinging Lexical Network account argues that, in contrast to a heterogeneous blocking context, the repeated naming of a set of semantically related items in the homogeneous blocking condition results in a cohort of strongly co-activated candidates striving to be selected (Abdel Rahman & Melinger, 2009, 2019). The presence of this active cohort slows down lexical selection because all active candidates enter into a competition for selection, much like a tug of war.

Within the Swinging Lexical Network model, the facilitation observed in the first cycle arises due to easier object recognition: Object recognition is most difficult in the first cycle, and participants may profit from the semantic contexts (Abdel Rahman & Melinger, 2007). This gives conceptual priming an upper hand over lexical competition, following the aforementioned assumption that blocking effects reflect a trade-off between conceptual facilitation and lexical competition, two processes that temporally overlap and unfold in parallel (Abdel Rahman & Melinger, 2009, 2019; Rabovsky et al., 2021). Supporting this explanation, a recent study manipulating visual and semantic similarity using a non-repeated semantic priming paradigm found evidence for a perceptually related top-down bias underlying initial facilitation (Scheibel & Indefrey, 2020). In this study, participants named pictures in three conditions: semantic/lexical knowledge (primed by a category word), perceptual/conceptual knowledge (primed by a picture of the category's mean shape, generated by overlaying all exemplars within the given category) and no a priori knowledge (no prime preceding the picture). Naming visually consistent categories (e.g., birds) benefitted from both types of a priori knowledge more than visually variable categories (e.g., buildings). Furthermore, the facilitation based on a priori perceptual/conceptual knowledge was consistently larger than that based on a priori semantic/lexical knowledge. The authors argue that a priori perceptual/conceptual knowledge presented a top-down mechanism that limited the number of target-shape candidates and accelerated the feature matching procedure, thus facilitating the naming response.

1.3 Tracking the functional architecture and time course of lexical selection

The basic assumption about the functional architecture of the just described model is that conceptual processing (typically reflected in facilitatory effects) and lexical selection (typically reflected in effects of interference) proceed in parallel (concurrent activation; for similar arguments of parallel processing, see Abdel Rahman & Sommer, 2003; Abdel Rahman et al., 2003; Feng et al., 2021; Strijkers et al., 2017). Note, this does not necessarily mean that both processes begin and end at the same time. Specifically, we assume continuous spread of activation within and between conceptual and lexical processing levels (cf. Roelofs, 1992, 2018). As a result, conceptual activation is initiated in an initial sweep before lexical activation, but conceptual activation does not stop when lexical activation begins. Rather, there is a period of time when both levels of processing are active and can in fact interact with each other. The period of overlapping conceptual and lexical activation and interaction results in the trade-off between facilitation and interference. The outcome of the trade-off depends on the situation and context in which pictures are named. For instance, in the blocking paradigm, when the semantic context is most helpful in the first cycles, semantic priming (facilitatory effects) dominates, whereas lexical interference dominates if many competitors are fully active at the lexical level, and the influence of conceptual priming should be reduced.

These aspects of the functional architecture of the language production system and the time course of conceptual and lexical processing can be investigated by employing the temporal precision of the EEG, as we will describe below (for EEG evidence in the blocking paradigm, e.g., Aristei et al., 2011; Janssen et al., 2011, 2015; Llorens et al., 2014; Wang et al., 2018; see de Zubicaray & Piai, 2019 for a comprehensive review). Across studies, a common finding of a larger ERP at posterior sites for the related compared to unrelated condition has been associated with the semantic blocking effect, although the polarity of the amplitude seems to differ. Aristei and colleagues (2011) incorporated the PWI task into the blocked-cyclic paradigm and found a larger negativity for homogeneous blocks starting at around 200–250 ms post-stimulus. This ERP was associated with the blocking effect based on the significantly different amplitudes when contrasting homogenous and heterogeneous blocks (see also an MEG study from Maess et al., 2002, which reports similar results), suggesting that the posterior negativity reflects lexical selection. A more recent study by Wang and colleagues (2018) reported a relative positivity at posterior sites for homogeneous blocks from 200 ms after stimulus onset. In sum, the evidence for the polarity of the electrophysiological signature of lexical selection remains inconsistent for the blocking paradigm.

1.4 The present study

The first goal of the present study is to examine whether enhanced semantic similarity leads to stronger blocking effects. We frame our study within the context of competitive models of lexical selection because they provide clear and tractable predictions. Specifically, enhancing semantic similarity should theoretically lead to (a) stronger conceptual priming, thus stronger semantic facilitation in the first cycle, and (b) more intense lexical competition, thus increased semantic interference in later cycles.

In the first presentation cycle, we expect semantic facilitation to emerge. Furthermore, the strength of this effect should be influenced by semantic similarity, that is, stronger blocking effects in the semantically close condition than in the semantically distant condition. The expectation is based on the current large-block design, where initial facilitation has been reliably reported (Belke, 2017). To trace the electrophysiological signature of such facilitation induced by higher semantic similarity, we focused on relatively early ERP modulations within a latency range below 200 ms at posterior sites that should reflect object recognition, including perceptual and conceptual aspects (e.g., Itier & Taylor, 2004; Thorpe et al., 1996; Tokudome & Wang, 2012; Valente et al., 2014; Vogel & Luck, 2000).

For cycles 2–5, we predict that increasing semantic similarity results in stronger interference during lexical selection when naming semantically closely related pictures compared to naming distantly related ones. In addition, we expect a larger posterior negativity at temporal-parietal sites starting in the time range of the P2/N2 component at around 250 ms post-stimulus (cf. Aristei et al., 2011), which has been linked to the process of lexical selection, in the semantically close condition compared to the distant condition. While prior investigations using other naming paradigms reported converging evidence of a positive-going ERP for the semantic context effects (as reviewed in Section 1.1), in the blocked-cyclic naming paradigm, negative-going ERPs have also been reported (e.g., Aristei et al., 2011; Maess et al., 2002). Except for the polarity, both ERP look temporally and topographically similar.

The second goal of this study is to relate behavioral facilitation and interference in the first and later cycles to different planning stages and their relative time courses reflected by ERPs. We assume that conceptual priming and lexical interference start at different points in time relative to picture onset. As proposed above, we expected conceptual facilitation to start first, and be indexed by an earlier ERP modulation. While this process is still ongoing, lexical selection starts and will be indexed by a later ERP modulation (parallel/concurrent activation). Assuming that conceptual and lexical processing occur concurrently with opposite forces, regardless of the behavioral blocking effects, their respective ERP modulations should be present both in the first and in later cycles, modulating naming latency in opposite directions. The two ERPs should interactively be related to naming latency such that an enhanced N1 (related to facilitated object identification) would induce facilitation, while an enhanced P2/N2 (related to concomitant lexical competition) would induce interfering effects. Naming latencies are a result of the relative contributions of both effects. Crucially, these two ERP modulations should interactively influence naming behavior, indicating a joint modulation by both cognitive processes on naming latency across all cycles - in support of the trade-off assumption between conceptual facilitation and lexical competition.

In sum, our basic expectations were graded blocking effects depending on semantic similarity on both naming latencies and ERP amplitudes: Behaviorally, semantic blocking effects emerge as facilitation in the first cycle but turn into interference in later cycles, whereas cycle 1 facilitation should be accompanied by a relatively early ERP modulation and cycles 2–5 interference should be accompanied by a relatively late ERP modulation. In addition, based on the theoretical trade-off assumption, these two ERPs should modulate naming latencies in opposite directions and interact with each other, indicating a joint modulation on naming latencies, but only when semantic similarity is high.

2 METHOD

The experiment was approved by the local ethics committee and was based on ethical principles put forward by the Declaration of Helsinki for research involving human subjects (Version 2013). The data that support the findings of this study are available in OSF at https://osf.io/jkzn9/?view_only=fcd144715c854731904736288dbd48ba.

2.1 Participants

We collected data from 25 healthy, right-handed, native German speakers with normal or corrected-to-normal visual acuity and normal color vision. Participants were compensated with expense allowance or received credit towards their curriculum requirements. All participants gave informed consent prior to participation. The data from one participant had to be excluded due to excessive EEG artifacts, resulting in a total number of 24 participants (18 females, mean age 23.8) for data analyses. The sample size was determined based on previous work investigating semantic interference effects in different naming paradigms (e.g., Rose & Abdel Rahman, 2017; Rose et al., 2019), which are comparable to the blocking effects starting from the second presentation cycle.

2.2 Materials

A total number of 125 colored photographs of objects were selected as picture stimuli. Stimuli and semantic relations were identical to Rose and Abdel Rahman (2017) and Rose and colleagues (2019), which manipulated the degree of semantic similarity in a systematic way. We manipulated semantic similarity by the semantic features that items share with closely related members in a sub-category, or with more distantly related members in an overarching category. For instance, an eagle shares more features with an owl and a parrot, but less with other animals such as a shark or a camel. In short, we varied the semantic feature overlap while keeping the overarching taxonomic category membership constant. Manual classification resulted in five broad categories (animals, clothes, tools, groceries, and furniture), each of which contained five sub-categories (e.g., animals: birds, fish, insects, ungulates, and monkeys; see Appendix A for the full stimulus list). Each sub-category consisted of five exemplars (e.g., birds: eagle, hummingbird, parrot, vulture, and owl). Each item was represented by a unique exemplar (i.e., we only included one image for “eagle”), and the same exemplar (image) was repeatedly named in a block.

The stimulus sets within each sub-category represented the semantically close blocking condition (hereafter the close condition). For the semantically distant blocking condition (hereafter the distant condition), we assigned one exemplar from each sub-category of a given broad category to form the stimulus sets (cf. Rose & Abdel Rahman, 2017; Rose et al., 2019). Finally, for the semantically unrelated blocking condition (hereafter the unrelated condition) we took one exemplar from each broad category. With this, all stimuli appeared in the close, distant, and unrelated stimulus sets. Visually, the selected pictures were typically not confused with other category members and were easy to identify. To avoid higher visual similarities between members in closely related sets influencing the expected effects, we selected our materials following two criteria: (1) pictures of objects are taken from different perspectives, without unnecessary similarities; (2) members in the closely related sets look visually different (e.g., eagle vs. owl). Using a computational similarity measure for images, the Haar wavelet-based perceptual similarity index (HaarPSI; Reisenhofer et al., 2018), we generated perceptual similarity indexes for all possible combinations between stimulus pictures. With the coefficients obtained from a Haar wavelet decomposition, local similarities between two images were assessed, including the relative importance of image areas. The average visual similarity is numerically relatively balanced across conditions (group means for close = .193; distant = .187; unrelated = .180), while an ANOVA test showed a significant difference between condition means, F(2,678) = 6.59, p = .001. However, since the HaarPSI ranges from 0 to 1, differences less than 0.7% have little practical significance on visual similarity. All photographs were edited for a homogenous background color and scaled to the size of 3.5 cm × 3.5 cm. Stimuli were presented on a 4/3 17″ BenQ monitor with a resolution of 1,280 × 1,024 using Presentation® software (Version 18.0, Neurobehavioral Systems, Inc., Berkeley, CA) at a viewing distance of 60 cm, producing an equal stimulus size of 2.7° visual angle for each object stimulus.

2.3 Procedure

Participants were familiarized with the stimuli prior to the main experiment as follows: Color print photographs were presented together with their names in random order on sheets of paper. Participants were asked to study the pictures and the corresponding names carefully. For the main experiment, participants were instructed to name pictures as fast and accurately as possible. On a screen with a light grey background, a fixation cross in the center indicated the start of a trial. After 0.5 s, a picture was presented for maximally 2 s, or disappeared as soon as the voice key was triggered. A blank screen followed and lasted for 1 s until the next trial started.

The experimental session consisted of three sections that corresponded to the three semantic blocking condition (close, distant, and unrelated). The ordering of conditions was counterbalanced across participants. The order of the stimulus sets within each condition was randomized. Each stimulus set was presented five times (five presentation cycles), and an online randomization was performed for each cycle separately. This resulted in 625 trials per semantic blocking condition (125 per presentation cycle) and a total trial number of 1,875.

2.4 Acquisition and analyses 2.4.1 Accuracy

In general, participants showed very high accuracy in naming pictures, and deviations were fairly low (M = 99.88%, SD = 0.12%). We ran a generalized linear mixed-effects model (GLMM) using the function glmer in the lme4 package (Bates, Maechler, et al., 2015, version 1.1-21) in R (R Core Team, 2018) to test the accuracy as a function of block order to control for possible covariate. The random structure consisted of random intercepts of subject and item, and a random slope of block order for subject.

2.4.2 Naming latencies

Naming latencies were measured with a voice key starting from stimulus onset to participants' response. Only those trials were analyzed in which participants named the picture correctly and without speech disfluency. According to these criteria, around 3.8% of the data had to be excluded. Naming latencies shorter than 200 ms (0.87%) were removed. We log-transformed the naming latencies based on the outcome of a Box-Cox Test in order to meet the normality assumption of linear mixed-effect models.

Of all trials, 91.76% entered data analyses. Aside from the pre-defined exclusion criteria, further trials were excluded due to EEG data loss. Since we aimed to predict naming latencies by ERP modulations on a trial-by-trial basis, we analyzed only those trials in which both behavioral and EEG data points survived the exclusion criteria, that is, correct naming within 200 ms and clean EEG signal. This resulted in a total of 8.11% of trials being excluded. While the sum of trials per participant was 1875, the number of trials removed was on average 152 (SD = 132) per participant. One participant's data in a whole condition block (one-third of all trials) was removed due to EEG recording issues.

Linear mixed-effects models (LMMs; Baayen et al., 2008) tested the relationship between log-transformed naming latencies and the predictors using R (R Core Team, 2018) and the lme4 package (Bates, Maechler, et al., 2015, version 1.1-21). Separate analyses were conducted for cycle 1 and cycles 2–5. We entered into the model as fixed effects the critical factor Semantic Blocking, and, for the analyses of presentation cycles 2–5, the control factor Presentation Cycle, and their interaction terms. The predictor Semantic Blocking was contrast coded to compare the semantically close to the unrelated condition (close vs. unrelated) and the semantically distant to the unrelated condition (distant vs. unrelated)2 . The predictor Presentation Cycle was centered and entered as a continuous variable.

To account for random effects, our model included intercepts for participants and items and random slopes for the fixed effect terms. Models were initially run with a maximum random effects structure. Since the maximal model failed to converge, we set all correlation parameters to zero by using the double-bar syntax (cf. Kliegl, 2014). Applying singular value decomposition, this initial random effect structure was simplified by successively removing those random effects whose estimated variance was zero until the maximal informative model was identified (cf. Bates, Kliegl, et al., 2015).

For fixed effects, we report fixed effect estimates, 95% confidence intervals, and t values. Fixed effects are considered significant if |t| ≥ 1.96 (cf. Baayen et al., 2008), but we also computed p-values by Satterthwaite approximation (using the summary function in the lmerTest package, version 3.1-1; Kuznetsova et al., 2017). For random effects, we report estimates of variance as well as the standard deviations. Goodness-of-fit statistics are also reported.

2.4.3 Event-related potentials

The continuous EEG was recorded with 62 Ag/AgCl electrodes, arranged according to the extended 10/20 system, online referenced to an electrode at the left mastoid. The sampling rate was 500 Hz. To register eye movements and blinks, electrodes were placed near the left and right corner of both eyes and above and beneath the left eye. Electrode impedance was kept below 5 kOhm.

Eye movements and blink artifacts in the EEG signals were identified by employing the Multiple Source Eye Correction (MSEC) method implemented in the BESA Research software (Version 6.0, BESA GmbH, Gräfelfing, Germany; Berg & Scherg, 1994). After identifying eye-movements artifacts, the raw EEG data were submitted to BrainVision Analyzer (Version 2.1.2, Brain Products GmbH, Gilching, Germany) for preprocessing. Offline EEG was re-referenced using the common average reference. The identified spatiotemporal patterns reflecting eye-movements artifacts were corrected by a linear derivative. To reduce noise, a low-pass filter was applied (high cutoff = 30 Hz, 24 dB/oct). The data was then segmented based on the reference marker, including all necessary markers. An interval starting from 100 ms before the stimulus onset was used for baseline correction to exclude stimulus-independent activity at the beginning of the segment. Remaining artifacts were eliminated with an automatic artifact rejection procedure, which excluded segments with potentials exceeding 50 μV voltage steps per sampling point and a threshold of 200 μV. The EEG data were then segmented again in epochs of 1,300 ms, starting 100 ms before the onset of the stimulus, to specify conditions for single-trial analysis. The resulting segments arranged according to experimental conditions were exported to MATLAB (Version 2019b, The MathWorks Inc., Natick, Massachusetts) for speech artifact correction.

Speech artifact correction

To tackle the severe artifacts induced by speaking in the EEG signals (e.g., Brooker & Donald, 1980; Grözinger et al., 1975; Wohlert, 1993), we implemented a MATLAB toolbox capable of correcting articulation-related artifacts: residue iteration decomposition (RIDE; Ouyang et al., 2016). This toolbox decomposes ERPs into separate component clusters with different trial-to-trial variabilities (e.g., stimulus-locked, response-locked, and latency-variable component clusters). Articulation artifacts can be identified from the EEG signal based on their large amplitudes and highly variable trial-to-trial latencies. By implementing RIDE, we decomposed the ERPs into the stimulus-locked S-component (search time window 0 to 600 ms after stimulus onset) and the response-locked R-component (i.e., articulation-related artifacts; search time window ±300 ms from response time, see Figure 1). The R-component (per participant and per condition) was then subtracted from the original ERPs for every single trial. The resulting cleaned ERPs were then matched to naming latencies on a trial-by-trial basis and exported to R for statistical analysis.

Component separation for artifact corrections with RIDE. The left plot shows the original ERPs prior to speech artifacts correction. The middle plot shows the S-component free of articulation-related noise, which was submitted to data analyses. The right plot shows the R-component, which was removed from the original ERPs

Analysis procedure

The general parameters and analysis procedure for models predicting ERP amplitudes were the same as the LMMs testing naming latencies, except that the predicted variable was replaced with the averaged EEG activities at the pre-defined ROIs and time windows.

Based on the hypothesis that conceptual priming should be strongest when participants name pictures for the first time in the semantically related contexts, likely restricted to the close context (cf. Abdel Rahman & Melinger, 2007; Scheibel & Indefrey, 2020), for cycle 1, we examined an ERP component reflecting the mechanism of object recognition and identification. Therefore, we analyzed EEG signals from the posterior sites, including electrodes TP9/10, P7/8, PO9/10, O1/2 (ROI for object recognition; Itier & Taylor, 2004) ranging from 140 to 180 ms after stimulus onset (the stage of visual complexity during picture naming; cf. Valente et al., 2014; see also other object recognition studies, e.g., Thorpe et al., 1996; Vogel & Luck, 2000).

For cycles 2–5, we hypothesized that lexical competition should be strongest when participants name pictures in semantically related contexts, particularly pictures with close relations. Here we examined the ERP amplitudes at the ROI and during the time window found in Aristei and colleagues' study (2011), in which brain activities were proposed to reflect changes during lexical retrieval in the blocked-cyclic naming paradigm. We selected two electrodes at the temporal-parietal sites, TP9 and TP10, as the ROI, and analyzed the EEG activities from 250 to 350 ms after stimulus onset.

In addition to our basic hypotheses and predictions that semantic facilitation in cycle 1 should be accompanied by a relatively early ERP modulation, and that lexical interference in cycles 2–5 should be accompanied by a relatively late ERP modulation, we took a step further to examine whether these two ERP modulations are active in parallel. This was based on the theoretical assumption that the concurrent processing of conceptual priming and lexical competition together contribute to the behavioral facilitatory or interfering blocking effects (cf. Swinging Lexical Network, Abdel Rahman & Melinger, 2009, 2019). In support of the trade-off hypothesis proposed in this framework, we should find concurrent traces of the early and late modulation in the selected ROIs across all cycles, with relative strengths. For this purpose, we separately ran a model testing the early modulation in cycles 2–5, as well as a model testing the late modulation in cycle 1.

2.4.4 Relating behavior to brain activities

Assuming that ERPs reflect the underlying cognitive sources, whose changes can be observed behaviorally, the amplitude of ERPs relevant to lexical-semantic processing should serve as a good predictor for naming latencies. Another reason for conducting this brain-behavior analysis was to examine whether participants' naming behavior was really modulated by concurrent processing of conceptual priming and lexical competition across all cycles, and whether their naming behavior was modulated together by these two cognitive processes at the same time. To examine the first question, we entered the mean ERP amplitudes occurring in both early and late time windows as two fixed effects into a single LMM to predict log-transformed naming latencies. To examine the second question, we entered the interaction term between these two ERPs as another fixed effect on the naming latencies. Moreover, we recoded the predictor presentation cycle as a two-level categorical variable (first vs. later) to more precisely capture our research interest, that is, conceptual facilitation in the first cycle versus lexical interference in later cycles.

The random structure was selected following the same procedure as the other models described above. In order to reduce the complexity of the model so that the results could be more interpretable, we split the dataset according to the blocking condition. Since the results of both the naming latency and ERP models indicated that naming latencies in the distant condition did not vary much from the unrelated condition, here we only analyzed data from the close and unrelated conditions.

3 RESULTS 3.1 Accuracy

The overall accuracy was 99.82% in the first block, 99.95% in the second block, and 99.89% in the third block. The GLMM showed a slight trend of significance of the contrast Order 2 versus Order 1 (β = 2.83, SE = 1.68, z = 1.68, p = .09), and a null result for the contrast Order 3 versus Order 2 (β = −1.66, SE = 1.75, z = −0.94, p = .34). The finding indicates that participants' naming accuracy was not influenced by the order in which items were named in the close, distant or unrelated blocks.

3.2 Naming latencies 3.2.1 Cycle 1

In cycle 1, the mean naming latency in the close condition was 39 ms shorter than in the unrelated condition (semantic facilitation), while the naming latency in the distant condition was only 5 ms longer than in the unrelated condition (close: M = 732, SD = 189; distant: M = 777, SD = 179; unrelated: M = 771, SD = 175; see Figure 2). The descriptive statistics were confirmed by the identified LMM testing the semantic blocking effects in cycle 1. The model showed that the hypothesis-based fixed effect contrast close versus unrelated was statistically significant, whereas the contrast distant versus unrelated did not reach significance (see Table 1). Thus, supporting our hypothesis, we found semantic facilitation for semantically close items in cycle 1. Although there was no semantic context effect for semantically distant items, the unexpected finding is still consistent with our hypothesis that semantically distant items induce weaker effects—in this case, too weak to be detected.

Naming latencies plotted against presentation cycles, grouped by semantic blocking condition. The figure shows that in the first cycle, participants named semantically closely related objects faster compared with naming distantly related or unrelated objects (semantic facilitation). From the second cycle onwards, the data pattern was reversed (semantic interference). The error bars refer to the 95% confidence intervals

TABLE 1. Linear mixed-effects model of cycle 1 on log-transformed naming latencies, with the semantic blocking contrasts close versus unrelated and distant versus unrelated as predictors Log-transformed naming latencies (cycle 1) Fixed effects Predictors Estimates 95% CI t p (Intercept) 0.014 −0.02 to 0.05 0.862 .393 Close versus unrelated −0.064 −0.10 to −0.03 −3.376 .002** Distant versus unrelated