Lexical selection is not by competition: Evidence from the blocked naming paradigm

A central issue in research on speech production is whether or not the retrieval of words from the mental lexicon is a competitive process. An important experimental paradigm to study the dynamics of lexical retrieval is the blocked naming paradigm, in which participants name pictures of objects that are grouped by semantic category (‘homogenous’ or ‘related’ blocks) or not grouped by semantic category (‘heterogeneous’ or ‘unrelated’ blocks). Typically, pictures are repeated multiple times (or cycles) within both related and unrelated blocks. It is known that participants are slower in related than in unrelated blocks when the data are collapsed over all within-block repetitions. This semantic interference effect, as observed in the blocked naming task, is the strongest empirical evidence for the hypothesis of lexical selection by competition. Here we show, contrary to the accepted view, that the default polarity of semantic context effects in the blocked naming paradigm is facilitation, rather than interference. In a series of experiments we find that interference arises only when items repeat within a block, and only because of that repetition: What looks to be ‘semantic interference’ in the blocked naming paradigm is actually less repetition priming in related compared to unrelated blocks. These data undermine the theory of lexical selection by competition and indicate a model in which the most highly activated word is retrieved, regardless of the activation levels of nontarget words. We conclude that the theory of lexical selection by competition, and by extension the important psycholinguistic models based on that assumption, are no longer viable, and frame a new way to approach the question of how words are retrieved in spoken language production. 2014 Elsevier Inc. All rights reserved. Introduction tradition, based largely on analyses of errors in spontaneAn issue of central importance to all models of speech production concerns the mechanism through which words are retrieved from the mental lexicon. Research on this issue can be divided into two different approaches, according to how they have traditionally emphasized different aspects of the question of how words are retrieved. One ous speech and brain damaged patients, sought to explain how the system determined which representation at the lexical level (of the many possible words) corresponds to the communicative intention (i.e., which word is the ‘target’; Caramazza, 1997; Dell, 1986; Rapp & Goldrick, 2000). That class of models made few if any commitments about the precise timing of when a given target word would be retrieved in different contexts. A second tradition was based on the assumption that the chronometry of word retrieval was a more vital issue to be tackled than explaining which word was produced (Levelt, Roelofs, & 254 E. Navarrete et al. / Journal of Memory and Language 76 (2014) 253–272 Meyer, 1999), and thus, this second tradition stipulated which lexical representation was the target word, and sought to explain the precise timing of when that word was retrieved. The most widely discussed and detailed computational implementation within this second tradition is WEAVER++ (Levelt et al., 1999; Roelofs, 1992, 1993, 2003), although a number of other models have been proposed that involve lexical competition, most notably the models of La Heij and colleagues (Bloem & La Heij, 2003; Bloem, van den Boogaard, & La Heij, 2004; La Heij, 1988; see also Howard, Nickels, Coltheart, & Cole-Virtue, 2006). In addition, a range of authors have argued for lexical competition without proposing a formal model (e.g., Abdel Rahman & Melinger, 2007; Belke, Meyer, & Damian, 2005; Caramazza & Costa, 2000; Costa, Miozzo, & Caramazza, 1999; Cubelli, Lotto, Paolieri, Girelli, & Job, 2005; Damian & Bowers, 2003; Damian & Martin, 1999; Damian, Vigliocco, & Levelt, 2001; Hantsch, Jescheniak, & Schriefers, 2005, 2009; Hantsch & Mädebach, 2013; Mulatti & Coltheart, 2012, 2014; Santesteban, Costa, Pontin, & Navarrete, 2006; Schriefers, Meyer, & Levelt, 1990; Vigliocco, Lauer, Damian, & Levelt, 2002; Vitkovitch & Tyrrell, 1999). According to the hypothesis of lexical selection by competition, the time required to retrieve the target word from the mental lexicon increases proportionally to the summed level of activation of nontarget words. Typically, ‘nontarget’ words refer to words that are semantically related to the target word (but see Abdel Rahman & Melinger, 2009a), as words that are related in meaning to the target are automatically activated in the course of lexical retrieval. Thus if a participant is producing the word ‘dog’, then nontarget words such as ‘cat’, ‘horse’, ‘bark’, ‘leash’ (etc.), will be activated by virtue of automatic spreading of activation within the semantic network, and between semantic representations and their corresponding lexical representations. This leads to the core prediction that is made by the hypothesis of lexical selection: Target picture naming latencies will be slower in contexts in which nontarget words are highly activated compared to contexts in which nontarget words are not highly activated. The project of the current study is to test this prediction in as direct a way as possible. This is important, because the hypothesis of lexical competition is one of the most widely assumed theoretical constructs in research on speech production. Thus, if this hypothesis is not viable, and we are compelled to adopt an alternative model of word retrieval, then a number of phenomena andmodels would need to be reconsidered. We first discuss previous research on this topic that has used the picture–word and continuous naming paradigms, and then turn to the principal focus of the current study: the causes of semantic facilitation and interference in the blocked naming paradigm. The picture–word interference paradigm The picture–word interference paradigm is a variant of the Stroop task (Stroop, 1935), where instead of naming target ink colors as in the Stroop task, participants name pictures of common objects while ignoring embedded distractor words (for early work see Lupker, 1979; Rosinski, 1977). Initially, the principal source of evidence that was marshalled in support of the hypothesis of lexical selection by competition was the so-called semantic interference effect. The semantic interference effect refers to the observation that picture naming latencies (e.g., ‘dog’) are slower when the distractor word is a semantic category coordinate of the target (e.g., cat) compared to when it is unrelated (e.g., car; e.g., Damian & Bowers, 2003; Glaser & Düngelhoff, 1984; Rosinski, 1977; Schriefers et al., 1990). The semantic interference effect seemed to provide strong evidence for the view that lexical selection was a competitive process (e.g., Roelofs, 1992). This is because all models were in agreement that the word representation corresponding to a related distractor would be more highly activated than the word representation corresponding to an unrelated distractor, and thus the slower target naming latencies in the related condition putatively indexed increased competition for target selection. However, that basic logic was challenged by a comprehensive review and analysis that showed the semantic interference effect was the exception rather than the rule. All else equal, the default pattern in the picture–word paradigm is that naming latencies speed up as distractors become more semantically similar to targets (Mahon, Costa, Peterson, Vargas, & Caramazza, 2007). For instance, participants are faster to name target items when the distractors describe actions associated with the targets (picture = dog, distractor = bark), or parts of the target object (picture = dog, distractor = tail), compared to unrelated distractors (Costa, Alario, & Caramazza, 2005; Mahon et al., 2007). Even in the original Stroop task, participants are faster to name the ink color of related noncolor words (name ‘red’ ink color of the word fire) than the same ink color of unrelated distractors (name ‘red’ ink color of the word grass; Dalrymple-Alford, 1972; for discussion see Mahon, Garcea, & Navarrete, 2012; Mahon & Navarrete, 2014; Mulatti & Coltheart, 2014; Roelofs & Piai, 2013). The most direct test of the hypothesis of lexical selection by competition, within the picture–word paradigm, is provided by a manipulation of within-category semantic distance. The hypothesis of lexical selection by competition predicts that participants should be, if anything, slower to name pictures (e.g., dog) with within-category semantically-close distractors (e.g., wolf) compared to within-category semantically-far distractors (e.g., lizard). Contrary to that core prediction, Mahon et al. (2007) found that participants are, if anything, faster to name target pictures as the within-category semantic distance decreases between distractors and pictures. We subsequently argued that the semantic interference effect, as the exception that must be explained, was not a lexical level effect but rather a post-lexical effect at the level of the response buffer (for discussion, see Dhooge, De Baene, & Hartsuiker, 2013; Dhooge & Hartsuiker, 2010, 2011a, 2011b, 2012; Finkbeiner & Caramazza, 2006a; Janssen, Schirm, Mahon, & Caramazza, 2008; Mahon et al., 2007, 2012; Miozzo & Caramazza, 2003; E. Navarrete et al. / Journal of Memory and Language 76 (2014) 253–272 255 but see Mädebach, Oppermann, Hantsch, Curda, & Jescheniak, 2011; Piai, Roelofs, & Schriefers, 2011). In response to that argument, a number of authors have suggested ways in which the scope of competition at the lexical may be restricted or constrained, in one or another way, in order to accommodate observations of semantic facilitation (see Abdel Rahman & Melinger, 2009a; Hantsch, Jescheniak, & Mädebach, 2012; Hantsch & Mädebach, 2013; Hantsch et al., 2005, 2009; Kuipers & La Heij, 2008, 2012; La Heij, Kuipers, & Starreveld, 2006; Mädebach & Hantsch, 2013; Piai, Roelofs, & Schriefers, 2012; for responses see Finkbeiner & Caramazza, 2006b; Mahon & Caramazza, 2009; Mahon & Navarrete, 2014; Mahon et al., 2012; Navarrete & Mahon, 2013; see also Bormann, 2011). Thus, the evidential status of the semantic interference effect with respect to the hypothesis of lexical competition is questionable (for a recent review see Spalek, Damian, & Bölte, 2013). At a minimum, dual-task paradigms such as the picture–word and Stroop paradigms are complex and require many auxiliary assumptions in order to be used to extrapolate inferences about lexical selection. Our own view is that the collective implication of the available data from the picture–word and Stroop paradigms is that lexical selection is not a competitive process: As has been proposed by models based on analyses of error data, the most highly activated word is selected, without regard for the levels of activation of nontarget words (Dell, 1986; Rapp & Goldrick, 2000). According to such a model, related distractor words would only prime the target, but not slow it down, at the lexical level. The cumulative semantic cost An arguably more direct means for evaluating the hypothesis of lexical selection by competition would be to manipulate the context in which a picture is named, while avoiding the complicating influence of a distractor word that must be ignored. One such paradigm that accomplishes this is the continuous naming paradigm. Participants name a series of pictures from multiple semantic categories, in a seemingly random order. Unbeknownst to subjects, the pictures are arranged within the list so that they can be analyzed by ‘ordinal position within category’ across the list. Thus, response times for all pictures that correspond to the first item within category are averaged together (e.g., the first animal, the first vehicle, the first piece of furniture, encountered in the list). Similarly, response times for each subsequent within-category ordinal position encountered in the list are averaged across categories. Brown (1981) and later Howard et al. (2006) found that picture naming latencies increase with each additional within-category item that participants name— a phenomenon referred to as the ‘cumulative semantic cost’ (for related work, see Tree & Hirsh, 2003; Vitkovitch & Cooper-Pye, 2012; Vitkovitch, Cooper-Pye, & Ali, 2010; Vitkovitch, Cooper-Pye, & Leadbetter, 2006; Wheeldon & Monsell, 1994). The cumulative semantic cost is linear at least out to five ordinal positions within category (e.g., linear R = .96 in Navarrete, Mahon, & Caramazza, 2010; see also Alario & Moscoso del Prado Martin, 2010; Belke, 2013; Belke & Stielow, 2013; Costa, Strijkers, Martin, & Thierry, 2009; Howard et al., 2006; Oppenheim, Dell, & Schwartz, 2010; Runnqvist, Strijkers, Alario, & Costa, 2012). The cumulative semantic cost has attracted interest because it is a long lasting effect: interference is propagated across many intervening trials spanning tens of seconds. Howard et al. (2006) argued that lexical competition was a necessary component of an explanation of the cumulative semantic cost, but that lexical competition was not sufficient. The reason that lexical competition was argued to not be sufficient was that, according to all models, including that outlined by Howard et al. (2006), activation at the lexical level decays rapidly, and thus would return to baseline during the intervening time (trials) between within-category items. Howard and colleagues therefore reasoned that a longer lasting mechanism was required and invoked long lasting changes in the weights between semantic and lexical representations. Oppenheim et al. (2010, see also Oppenheim, Dell, & Schwartz, 2007) took that idea further in showing that a model that included only incremental learning, but not lexical competition, was able to explain the cumulative semantic cost (see also Damian & Als, 2005; Navarrete et al., 2010; Vitkovitch & Humphreys, 1991). The basic idea behind the incremental learning account is that when a picture (e.g., dog) is correctly named, the connection between its semantic features and its lexical items will be strengthened, while the connections between those same semantic features and nontarget but related lexical items (e.g., ‘cat’, ‘wolf’, ‘lizard’) will be weakened. This account explains two phenomena: (i) repetition priming, whereby naming latencies decrease when a given item is named a second time (name ‘dog’ a second time), and (ii) the cumulative semantic cost, as the next within category item to be named (e.g., ‘cat’) will have a weaker semantic-to-lexical connection for having previously named ‘dog’. Importantly, Oppenheim et al.’s (2010) computational work demonstrates that the cumulative semantic cost does not distinguish between competitive and noncompetitive models. This is because they showed that lexical competition was not necessary in order to explain the cumulative semantic cost. Thus, while the cumulative semantic cost continues to provide a basis for understanding incremental learning effects on semantic-to-lexical connections it is not an obviously useful tool with which to study whether activation levels of nontarget words influence target selection times. The blocked naming paradigm The final experimental paradigm that has been used to provide insight into the mechanisms of word retrieval is the blocked naming paradigm. In the blocked naming paradigm participants are required to name a series of pictures in two semantic contexts (i.e., two different types of blocks). In homogenous or related contexts, all pictures within a block come from the same superordinate semantic category (e.g., ‘car’, ‘airplane’, ‘truck’, ‘motorbike’, ‘bus’). In heterogeneous or unrelated contexts, all pictures within a block come from different semantic categories (e.g., ‘car’, ‘hammer’, ‘dog’, ‘apple’, ‘chair’; see Kroll & Stewart, 1994). 256 E. Navarrete et al. / Journal of Memory and Language 76 (2014) 253–272 An important aspect of the way that this task has come to be widely implemented, since the work of Damian et al. (2001), is that all pictures are repeated multiple times within a block. Each repetition of the set of pictures within a block is termed a ‘cycle’. There are two principal observations: (i) there is, if anything, semantic facilitation for the first within-block cycle, whereby naming latencies are faster in a semantically related compared to a semantically unrelated context; and (ii) there is semantic interference for the second cycle onward (i.e., all cycles after the first), whereby naming latencies are slower for pictures in a semantically related context compared to an unrelated context. The observation that an initial semantic facilitation effect (first cycle) changes polarity to a semantic interference effect (all subsequent cycles) has been replicated in at least two studies (Abdel Rahman & Melinger, 2007; Navarrete, Del Prato, & Mahon, 2012). Other studies reported an interaction between cycle and semantic context, without specifying the effect at the first cycle (e.g., Abdel Rahman & Melinger, 2011; Belke, 2008; Belke, Brysbaert, Meyer, & Ghyselinck, 2005; Belke, Meyer, et al., 2005; Biegler, Crowther, & Martin, 2008; Damian & Als, 2005; Schnur, Schwartz, Brecher, & Hodgson, 2006; Schnur et al., 2009). Typically, the data from the blocked naming task are not broken down by cycle, so that an overall interference effect is observed when collapsing across cycle. Despite the bivalent polarity of semantic context effects in the blocked naming paradigm, most discussion has focused on the interference effect, and has either ignored or explained away the facilitation effect. Specifically, taking ‘semantic interference’ as the empirical fact that requires explanation, it has been widely argued that due to spreading activation from semantic to lexical levels of processing, the word representations of the other items within a related block will be more highly activated than the word representations of other items within an unrelated block. Accordingly, lexical selection by competition predicts (correctly, it has been argued) slower naming latencies in related blocks than in unrelated blocks (e.g., Abdel Rahman & Melinger, 2007, 2011; Aristei, Melinger, & Abdel Rahman, 2011; Belke, Meyer, et al., 2005; Damian et al., 2001; Santesteban et al., 2006; Schnur et al., 2009). This type of an explanation is the received view in the field—there has been little, if any, critical discussion of this argument. However, if lexical selection is not a competitive process (as we would conclude from the picture–word and Stroop paradigms, e.g., Mahon et al., 2007, 2012), then it follows that the received explanation of what is occurring in the blocked naming task must be incorrect. The project of this article is to test the received view and to unpack, from a new perspective, the inferences about word retrieval that are supported by the full pattern of effects observed in the blocked naming paradigm. Reinterpreting the blocked naming paradigm The fact that semantic facilitation is observed in the first within-block cycle in the blocked naming paradigm has generally been ignored for its potential relevance to understanding the mechanisms involved in word retrieval. However, prior work using related paradigms also found that picture naming is facilitated if the item named on the previous trial is related, compared to when it is unrelated (e.g., Biggs & Marmurek, 1990; Humphreys, Riddoch, & Quinlan, 1988; Huttenlocher & Kubicek, 1983; Lupker, 1988; Sperber, McCauley, Ragain, & Weil, 1979; for more recent work see Hartsuiker, Pickering, & de Jong, 2005; Tydgat, Diependaele, Hartsuiker, & Pickering, 2012). The theory of lexical selection by competition offers no natural explanation of semantic facilitation in the first within-block cycle of the blocked naming task, nor of previous demonstrations of facilitation induced by an immediately-preceding related context. Proponents of lexical competition have sought additional mechanisms to explain the presence of semantic facilitation in the first cycle of the blocked naming paradigm. For instance, Belke, Meyer, et al. (2005) argued that lexical competition requires ‘‘some potentiation, through repetition, to become observable’’ (p. 687), an idea that may resonate with the suggestion of a ‘swinging’ lexical network as proposed by Abdel Rahman and Melinger (2009a; see also Abdel Rahman & Melinger, 2007, 2011; for discussion, see Mahon & Caramazza, 2009; Abdel Rahman & Melinger, 2009b). But, is there any a priori reason why competition should need repetition to be potentiated? In other contexts, repetition was not thought to be important in order for competition to exert an influence over response times. In other words, there is nothing in the theory of lexical competition that indicates repetition should be a relevant factor in determining when lexical competition occurs. The elegance of the theory of lexical competition (e.g., Roelofs, 1992) was that activation levels of words determined competition, and activation levels of words were themselves determined by the structure of lexical semantic representations and the dynamics of activation flow between the semantic and lexical levels. However, setting aside whether there is any motivation for invoking ‘repetition’ as a relevant factor in triggering lexical competition, the account of Belke, Meyer, et al. (2005) explains only the lack of interference in the first cycle, not the presence of facilitation (for discussion on this point see Navarrete et al., 2012). An alternative way to conceptualize the bivalent nature of semantic context effects in the blocked naming paradigm is to consider the facilitation that is observed in the first cycle as what is informative about the mechanism of lexical retrieval. According to a noncompetitive model (e.g., Dell, 1986) the most highly activated word is selected, and the time of that selection event (provided it is a correct trial) is not affected by the levels of activation of nontarget words (for further discussion, see Damian & Als, 2005; Finkbeiner & Caramazza, 2006a; Mahon et al., 2007; Navarrete et al., 2010, 2012; Oppenheim et al., 2010). Such a model naturally explains the semantic facilitation effect that is observed for the first within-block cycle: A related semantic context will prime the word representation of the target, boosting its level of activation, and speeding its retrieval (for a parallel account in word recognition, see e.g., McRae & Boisvert, 1998). This type of semantic priming would occur in related (i.e., homogenous) blocks, 540 560 580 600 620 640 660 680 700 Within-Category Semantically Close Within-Category Semantically Far Unrelated Cycle 1 (Actual Findings) Cycle 2 (Actual Findings) Hypothetical Cycle 2 Data with constant proportion (15%) of repetition priming for all Conditions R es po ns e Ti m e (m s) Fig. 1. Reanalysis of the data from Experiment 2 in Navarrete et al. (2012). The black and gray bars show actual data, indicating that (i) there is semantic facilitation in the first cycle, and that it is, if anything, greater for semantically more close (within-category) than semantically more distant (within-category) blocks, and (ii) there is semantic interference for the second cycle, and it is, if anything, greater for semantically more similar (within-category) than semantically more distant (within-category) blocks. As a demonstration of principle, we then computed hypothetical cycle 2 data by reducing cycle 1 response times by a proportionally equivalent amount (15%). As can be seen, if there were proportionally equivalent repetition priming for all contexts, then semantic interference would not emerge in cycle 2. This suggests that the emergence of semantic interference may derive from a baseline shift in the unrelated condition. In other words, the emergence of semantic interference depends (by hypothesis) on greater repetition priming in the unrelated than the related conditions. E. Navarrete et al. / Journal of Memory and Language 76 (2014) 253–272 257 in which pictures come from the same category, but not in unrelated (i.e., heterogeneous) blocks, in which pictures come from different categories. This explanation also provides a natural account of the observation that if filler items are interposed among the related items, for instance geometric shapes that must be named, the facilitation effect observed in the first cycle is abolished (see Navarrete et al., 2012; see also Damian & Als, 2005). According to a model of lexical access in which lexical selection (i.e., lexical retrieval) is not by competition, the issue then arises as to why there is interference after the first cycle in the blocked naming task. Navarrete et al. (2012) argued that the incremental learning account advanced by Oppenheim, Dell, and Schwartz (2007) and Oppenheim et al. (2010) could explain the emergence of semantic interference, if it were the case that there was reduced repetition priming for items presented in related contexts compared to items presented in unrelated contexts. This would be because while items presented in an unrelated context undergo only strengthening of semantic-to-lexical mappings, the same strengthening that occurs in a related context will be immediately countermanded by incremental weakening from having to name other within category items (see Navarrete et al., 2012, for preliminary evidence). In other words, ‘semantic interference’ in the blocked naming paradigm is not, by hypothesis, an interference effect: it arises because of a baseline shift in the unrelated condition, not because the related condition is slowing down. We believe that an explanation of the emergence of ‘interference’ in the blocked naming paradigm in term of a ‘baseline shift’ in the unrelated condition can already be seen in published data. Fig. 1 replots the findings from an experiment reported in Navarrete et al. (2012) in which within-category semantic distance was manipulated in a blocked naming paradigm. Thus, items within a block could be very close within category (e.g., ‘dog’, ‘fox’, ‘wolf’), relatively far within category (e.g., ‘dog’, ‘lizard’, ‘whale’), or unrelated (e.g., ‘dog’, ‘table’, ‘car’). The graph shows the findings from the first within-block cycle, in which response times are fastest for the within-category semantically-close blocks, intermediate for within-category semantically-far blocks, and slowest for the unrelated blocks. That pattern is entirely in line with what would be predicted by a noncompetitive model of lexical selection, and contrary to what would be predicted by lexical selection by competition: increasing within-block semantic similarity leads to faster response times. What is relevant for present purposes though is what happens in the second within-block cycle: Now there is a flip where the slowest condition is the within-category semanticallyclose condition and the fastest condition is the unrelated condition. A very large repetition priming effect can be seen, comparing the data from the first and second cycles, where there is (averaging across all conditions) a 15% speeding of response time. The critical point here is that the amount of repetition priming is different in different conditions. This can be appreciated if response times from the first cycle across all of the conditions are reduced the same amount, proportionally (i.e., 15%), from values in the first within-block cycle. Those hypothetical data are plotted in Fig. 1, and show that if all that occurred were proportional repetition priming, there would be no ‘semantic interference’ in the second within-block cycle. Thus, the ‘emergence’ of the semantic interference effect in the second cycle is equivocal in that it could be due to interference (i.e., lexical competition) or, it could be due to relatively less repetition priming in the related condition compared to the unrelated condition (see also Navarrete et al., 2012, Fig. 5 for independent evidence consistent with this argument). Because in Navarrete et al. (2012), as in all other studies using the blocked naming paradigm, the same items appear in both the related and unrelated conditions, it is not possible to truly tease apart (putative) semantic interference from differential repetition priming in the unrelated compared to the related condition. Thus, the goal of the present investigation was to test the hypothesis that the emergence of ‘semantic interference’ in the blocked naming paradigm is the result of a baseline shift caused by greater repetition priming in the unrelated condition. If this hypothesis is confirmed, then it would mean that the semantic interference observed in the blocked naming task could not be taken as evidence for the hypothesis of lexical competition. In that case, we would be left with the pattern of semantic facilitation as observed in the first 258 E. Navarrete et al. / Journal of Memory and Language 76 (2014) 253–272 within-block cycle as the data point that requires explanation, which would strongly favor a model of word retrieval that does not include lexical competition. Redesigning the blocked naming task Previous research with the blocked naming paradigm does not permit an evaluation of whether the emergence of semantic interference with within-block repetition is due to relatively faster naming latencies in the unrelated condition (baseline shift) or to relatively slower naming latencies in the related condition (lexical competition). The reason is that previous work confounds ‘repetition within the experiment’ with ‘repetition within a block’. Thus, in a typical implementation of the blocked naming task (e.g., Damian et al., 2001; for recent reviews see Belke & Stielow, 2013; Navarrete et al., 2012), (i) all items appear in both related and unrelated blocks (within subject), and (ii) all items are presented multiple times within each related and unrelated block. In the present research we take a novel approach and separate the factors ‘repetition in the experiment’ from ‘repetition within the block’. To accomplish this, the small but consequential design changes were introduced that: (i) each picture was presented once per block, and blocks were repeated multiple times, and (ii) different pictures were presented in related and unrelated contexts (counterbalanced across participants with a split-half materials/ super-subject design). Fig. 2 illustrates the design used in Experiment 1. All items were first presented twice in unrelated contexts. At the third presentation, half of the items were presented in related blocks, and were presented in related blocks for the remainder of the experiment. The other half of the items continued to be presented in unrelated blocks, and were presented in unrelated blocks for the remainder of the experiment. These design changes allow us to evaluate the polarity of semantic effects as a A Items (half of categories) B Items (half of categories) U