Acta Psychologica 148 (2014) 96–106

Contents lists available at ScienceDirect

Acta Psychologica journal homepage: www.elsevier.com/ locate/actpsy

Redundancy gain in semantic categorisation☆ Peter Shepherdson ⁎, Jeff Miller University of Otago, PO Box 56, Dunedin 9054, New Zealand

a r t i c l e

i n f o

Article history: Received 27 March 2013 Received in revised form 9 January 2014 Accepted 11 January 2014 Available online 6 February 2014 PsycINFO classification: 2340 2343 2346

a b s t r a c t Redundancy gain refers to the performance enhancements often associated with the presentation of redundant versus single targets (for example, faster, more accurate, or more forceful responses). Though predominantly observed in relatively simple tasks (e.g., stimulus detection), there have been some efforts to investigate similar phenomena in tasks involving higher level processing. We conducted three experiments aimed at determining (a) whether a redundancy gain would be evident in a task unambiguously requiring higher level processing (the semantic categorisation of visually-presented lexical stimuli), and (b) if so, what accounts might be appropriate to explain such findings. We found that redundancy gains are observed in such tasks, and we conclude that both coactivation and race models can account for these gains. © 2014 Elsevier B.V. All rights reserved.

Keywords: Redundancy gain Semantic processing Categorisation Race model Coactivation

1. Introduction In a phenomenon known as redundancy gain, performance is enhanced by the presentation of multiple stimuli prompting the same response (compared to when a single such stimulus is presented). This enhancement can take any number of forms, such as decreased response latency (Grice, Canham, & Boroughs, 1984; Hershenson, 1962; Miller, 1982, 1986; Todd, 1912), decreased error rate (Baird & Burton, 2008; Mohr, Landgrebe, & Schweinberger, 2002; Mohr, Pulvermüller, Mittelst dt, & Rayman, 1996; Mohr, Pulvermüller, & Zaidel, 1994, 2002), and more forceful responses (Giray & Ulrich, 1993; Mordkoff, Miller, & Roch, 1996). However, exactly what mechanisms lead to redundancy gain is a source of some debate (e.g., Miller, 1982; Mordkoff & Yantis, 1991; Raab, 1962; Townsend & Nozawa, 1997). Redundancy gain has predominantly been demonstrated in lower level tasks such as simple sensory detection (e.g., Savazzi & Marzi, 2002; Schwarz & Ischebeck, 1994; Veldhuizen, Shepard, Wang, & Marks, 2010). Nonetheless, there have also been efforts to apply redundant ☆ The experiments reported in this paper were undertaken as part of Peter Shepherdson's doctoral thesis. Peter Shepherdson was supported by a University of Otago Doctoral Scholarship and a University of Otago Publishing Bursary during the preparation of this paper. A preliminary report of Experiments 1 and 3 was presented at the 2012 annual meeting of the Psychonomic Society. We thank Wolf Schwarz for comments on a previous version of the manuscript. ⁎ Corresponding author. E-mail addresses: [email protected] (P. Shepherdson), [email protected] (J. Miller). 0001-6918/$ – see front matter © 2014 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.actpsy.2014.01.011

target paradigms to the investigation of higher level processing. These have included the use of tasks such as fame judgements for faces (Baird & Burton, 2008; Mohr et al., 2002; Schweinberger, Baird, Blümler, Kaufmann, & Mohr, 2003), lexical decision (e.g., Mohr, Pulvermüller, Rayman et al., 1994; Mohr, Pulvermüller, & Zaidel, 1994; Mohr et al., 1996; Mullin & Egeth, 1989), emotion recognition (e.g., Collignon et al., 2008, 2010; Tamietto, Adenzato, Geminiani, & de Gelder, 2007; Tamietto, Latini Corazzini, de Gelder, & Geminiani, 2006), and object recognition (Molholm, Ritter, Javitt, & Foxe, 2004; Suied, Bonneel, & ViaudDelmon, 2009). In each of these cases, redundancy gains have been observed. This seems to indicate that redundancy gain is not limited to simple experimental tasks, but could instead be a more general principle of human information processing. However, whether tasks purporting to demonstrate redundancy gain in higher level processing have actually done so may be questioned. For instance, studies investigating object and emotion recognition, undertaken by Molholm et al. (2004), Collignon et al. (2008, 2010), and Suied et al. (2009), involved small sets of stimuli that were presented repeatedly. This could have allowed participants to build low level S–R associations, which in turn could have obviated the need for any higher level processing in completing the tasks. Thus, the redundancy gain might just have emerged within the processing of low level S–R associations. This was also the case for lexical decision and semantic categorisation experiments undertaken by Mullin and Egeth (1989). Other studies have also involved confounds between redundancy at higher and lower levels of processing. For example, Tamietto et al.'s (Tamietto et al., 2006, 2007) studies of the recognition of emotional facial expressions unavoidably

P. Shepherdson, J. Miller / Acta Psychologica 148 (2014) 96–106

involved perceptual as well as emotional redundancy, since facial expressions are defined by their physical features. As perceptual redundancy has been demonstrated to lead to gains even with visually complex stimuli (e.g., faces; Jiang, Kwon, Shim, & Won, 2010), the effects of redundancy on emotion recognition could have been mediated by the processing of physical features rather than higher level emotional concepts. The results of Mohr and Pulvermüller (e.g., Mohr, Pulvermüller, Rayman et al., 1994; Mohr, Pulvermüller, & Zaidel, 1994; Mohr et al., 1996) using lexical decision tasks (LDTs) appear to provide somewhat stronger evidence of redundancy gain in higher level processing. In these experiments, participants were presented with one or two copies of a word or non-word on each trial, and asked to make manual “word” or “non-word” responses. Findings showed faster and more accurate responding in redundant than single-target trials, but only for words (there were limited or no effects of redundancy on non-words). While individual stimuli were repeated rarely or never, redundant trials involved not only lexical redundancy, but also perceptual redundancy (as multiple copies of the same word or non-word were presented simultaneously). However, had perceptual redundancy been the source of the gains there is no obvious reason they should not have occurred for non-word stimuli as well. Thus, the overall pattern of results provides some support for the idea of redundancy gain in higher level (i.e., word recognition) processes. Mohr et al. (Mohr, Pulvermüller, Rayman et al., 1994; Mohr, Pulvermüller, & Zaidel, 1994; Mohr et al., 1996) explained their results with reference to a neurobiological model of language, based on the concept of Hebbian cell assemblies (e.g., Hebb, 1949). According to this model, words are represented in the brain as collections of cells spread throughout the cortex—that is, cell assemblies. When a word is presented, its cell assembly is activated, leading participants to make a “word” response. When two copies of a word are presented, this provides extra activation to its cell assembly, and the response can be made more rapidly. By contrast, unfamiliar non-words do not possess such representations, and thus are unable to benefit from redundant stimulation. The cell assembly model offers an account of how redundancy gain might occur in a task involving higher level processing (in this case, judgements about “wordness”). Under the theory that similar findings should be evident not just for words but for all complex stimuli possessing existing neural representations, further studies showed redundancy gains in a fame judgement task. In this task photographs of faces were presented singly or redundantly and participants were asked to judge whether or not they belonged to “famous” people (e.g., Baird & Burton, 2008; Mohr et al., 2002; Schweinberger et al., 2003). Analogously to the findings in lexical decision, these gains occurred for “famous” faces, but not “non-famous” faces. Mohr et al. (2002) suggested that this was due to only famous faces being familiar and thus having neural representations whose activation could be enhanced by dual stimulation. However, despite circumventing some of the problems with other experiments showing redundancy gain in tasks which may have involved higher level processing, results from lexical decision and fame judgement tasks also fail to provide unequivocal evidence of a higher level redundancy gain. Though decisions about lexicality and fame can certainly make use of complex information (see, e.g., Balota, Cortese, Sergent-Marshall, Spieler, & Yap, 2004; James, 1975), they also have the potential to be made entirely on the basis of stimulus familiarity (e.g., Balota & Chumbley, 1984). Whether a judgement of familiarity could really be said to require higher level processing is debatable, meaning that the existence of redundancy gain in higher level processing is still an open question. In addition to uncertainties about the level of processing underlying redundancy gains found in the lexical decision and fame judgement tasks, the mechanism behind the gains is also open to debate. Generally, accounts of redundancy gain take one of two forms: coactivation models, and race models. In coactivation models, evidence from redundant stimuli is somehow summed, facilitating responses. As such, the cell assembly model falls under this umbrella. Race models, by contrast, suggest that

97

redundant stimuli are processed in parallel but separately, and that gains result from the statistical facilitation caused by both stimuli “racing” to activate a response. One frequently-used method of determining which explanation is appropriate in a redundant target experiment is through the use of Miller's (1982) race model inequality (RMI). This inequality describes the limit to the possible extent of reaction time enhancement through statistical facilitation. This limit can be modelled using RT data from single-target trials, and if RTs in redundant trials are faster than the modelled RTs—that is, if the RMI is violated—then a race model explanation can be ruled out. Unfortunately, coactivation models do not necessarily produce violations of the RMI, so failures to observe such violations are not strong support for race models. Mohr et al.'s (1996) preference for the cell assembly model was based largely on its biological plausibility. Though they provided other evidence from their results to rule out an alternative race model-based explanation, they only rejected a very specific race model in which the two cerebral hemispheres each processed separate stimuli. There are other race models which fit equally as well with the data from LDTs and fame judgement tasks as does the cell assembly coactivation model. For instance, the finding that redundancy gain occurred only for positive (word/“famous”) stimuli and not for their negative counterparts (non-word/“non-famous”)—used to support Mohr et al.'s argument for the cell assembly coactivation account—could simply be a result of a self-terminating race. In this account, positive responses can be directly elicited by stimuli (and thus benefit from the statistical facilitation associated with redundant targets) but negative responses require a temporal threshold of some sort to be reached (and thus are unable to receive the same benefit irrespective of the number of nontarget stimuli presented). As the RMI was not tested in any of the studies mentioned (e.g., Baird & Burton, 2008; Mohr, Pulvermüller, & Zaidel, 1994, Mohr et al., 1996; Schweinberger et al., 2003), such an alternative model cannot be ruled out. Stronger evidence for redundancy gain in higher level processing comes from a recent study by Fiedler, Schr ter, and Ulrich (2013), who demonstrated redundancy gain in the processing of categorical and physical features associated with the objects denoted by visually presented words. In Fiedler et al.'s experiment, participants completed a go/no-go task where they were asked to respond if a word described an entity either belonging to a specified superordinate category (animals), an entity possessing a certain physical feature (grey in colour), or both. Fiedler et al. found a significant redundancy gain for words designating objects matching both target criteria (e.g., “elephant”), but no RMI violations. They concluded that the redundancy gain could be explained by a race model (i.e., statistical facilitation) in which the processes analysing the categorical and physical features of the indicated object operate in parallel, with the first one to finish activating the response. Thus, this study shows that information about an object's abstract semantic properties (i.e., category) can participate in redundancy gain with information about a more concrete, physical property (i.e., colour) when both are retrieved from memory. Given that Fiedler et al.'s (2013) findings provide strong support for redundancy gain during higher level processing of the semantic representation denoted by a single word (e.g., “elephant”), it seems logical to ask further whether redundancy gain can also arise during the processing of two words with different lexical entries. The present experiments were designed to answer this question within a task where there was a single target attribute defined by semantic category, and the redundant items consisted of different words within that category. In addition, as in Fiedler et al.'s study, we sought to determine whether any observed redundancy gain could best be explained by parallel or serial self-terminating models. 2. Experiment 1 This experiment used a semantic categorisation task with lexical stimuli, based on the LDT studies of Mohr and Pulvermüller (e.g., Mohr,

98

P. Shepherdson, J. Miller / Acta Psychologica 148 (2014) 96–106

Pulvermüller, Rayman et al., 1994, Mohr, Pulvermüller, & Zaidel, 1994, Mohr et al., 1996). Participants were briefly presented with lateralised words and/or non-words, and asked to respond with one hand if either or both of the stimuli were words from a pre-specified target category (animals), and respond with the other hand if either or both of the stimuli were words from outside of that category. Non-word stimuli were used as fillers and were not informative as to the correct response. Short stimulus durations and lateralised presentation were used to mimic the conditions employed in the LDT studies (e.g., Mohr, Pulvermüller, Rayman et al., 1994) insofar as was possible. It should be emphasized that this design controls for a number of potential confounding factors that have limited previous studies. For example, to ensure that any effects of redundancy were not simply a result of varying stimulus number (see, e.g., Mordkoff et al., 1996), non-words were presented as distractors in single-target trials, keeping stimulus number constant. To avoid effects of perceptual and lexical redundancy, redundant trials involved the presentation of two different words from within or outside of the target category. To avoid any confound between category membership and stimulus familiarity, animal and non-animal words were matched for frequency. To avoid the possibility of low level S–R associations, a large set of stimuli was used. In combination, controlling for these factors allowed us to assess redundancy gain between higher level (i.e., categorical) information about two separate lexical entries in a task which was free from numerous other low level forms of redundancy. 2.1. Method 2.1.1. Participants Twenty-four undergraduate psychology students (aged 18–23; 21 females, three males; one left-handed, 22 right-handed, one with no clear handedness) completed Experiment 1 in return for partial course credit, as part of the Department of Psychology's experimental participation programme. 2.1.2. Apparatus and stimuli An IBM-PC Compatible 486 computer was used to run the experimental program, while instructions and stimuli were displayed on a CRT monitor, viewed at a distance of approximately 60 cm. Behavioural responses made using the left and right index fingers were recorded by the use of two force-sensitive keys, located to either side of the computer's keyboard. These keys were linked to strain gauges, by way of which an analogue signal indicating response force was delivered to the experimental program. Force was digitized at the rate of 250 samples per second, with a resolution of approximately 2.8 mN. A response force of 100 cN (approximately the force required to depress a key on a computer keyboard) was set as the threshold for assessing response times. Three types of stimuli were used. First, “target” stimuli were the names of animals, taken from Battig and Montague's (1969) category norms. Only single-word, clear-cut category members were included as stimuli. After pre-testing using the experimental task with a separate participant group, those stimuli classified at a below-chance level were also removed, leaving a final pool of 82 names of animals, ranging from three to ten letters in length. Second, non-animal nouns (taken from no specific category) were used as “non-target” stimuli. These were matched with the target words for length and frequency, determined by way of word frequency lists available online (Wiktionary, 2006). Third, distractor stimuli took the form of pronounceable non-words generated via the ARC non-word database (Rastle, Harrington, & Coltheart, 2002). These were matched with the target and non-target words for length. Stimuli were presented in lowercase, in serif font, with letter size of approximately 0.8° of visual angle in height and width, and with the centre of each word approximately 3.7° to the left or right of fixation. Two stimuli were displayed on each trial, with non-words presented

with equal frequency on target-present and target-absent trials, to control for any possible effect of inter-stimulus contingencies (Mordkoff & Yantis, 1991). 2.1.3. Procedure Each participant completed an experimental session between 45 and 55 min in length. This involved making responses to 240 target-present and 240 target-absent trials over the course of ten experimental blocks (48 trials per block). There were six types of trials defined by a 2 × 3 design, and these six were tested equally often within each block. One factor was target presence, with a word from the target semantic category either present or absent in each display. The other factor was the word location presentation condition: (a) a target or non-target word in left visual field (LVF), with a non-word in the right; (b) a target or non-target word in right visual field (RVF), with a non-word in the left; and (c) two target or non-target words in both visual fields (BVF), with no nonwords. This design controlled for interstimulus and non-target response contingencies (see Mordkoff & Yantis, 1991, Experiment 1). Trial type presentation order was randomised within each block. A practice block of 24 trials preceded the experimental blocks, to allow participants to become accustomed to the task. Each trial commenced with the presentation of a fixation cross for a period of 400 ms, following which stimuli were flashed to the left and right of fixation for 100 ms (a duration chosen to prevent participant saccades). Participants had 1500 ms from stimulus onset to make their response, before the fixation cross disappeared from the screen. If the response was correct, this concluded the trial. If the response was in any way erroneous (e.g., when participants pressed the incorrect button, or when they failed to respond within the 1500 ms window), feedback was presented on-screen. The inter-trial interval was uniformly distributed between 1000 and 2000 ms. Participants were instructed to press one key if an animal word was present on either or both sides of the screen, and the other key if a nonanimal word was present on either or both sides of the screen. Left and right hand response assignment for target-present and target-absent responses was counterbalanced across participants. In addition, participants were informed (a) that non-words could co-occur with both animal and non-animal words, but that animal and non-animal words were never presented together; and (b) that non-word distracters appeared equally frequently with both animal and non-animal words. Finally, they were asked to respond “as quickly as you can without making too many mistakes”. Once any questions they had regarding the experiment had been answered, they commenced the practice block by pressing a foot-switch located underneath the computer desk. At the beginning of each block, written instructions appeared on-screen, and participants were required to press the foot-switch again to proceed to the following block of trials. Once all blocks had been completed, participants filled out a computerised version of the Edinburgh Handedness Inventory (Oldfield, 1971), were debriefed, and were thanked for their participation. 2.1.4. Data analysis After data from practice blocks and response anticipations (RT b150 ms) were discarded, the remaining means of RT, peak force (PF), and accuracy (percent correct) for each participant were entered into separate ANOVA with factors of Target (present versus absent) and Presentation condition (of target/non-target word stimuli: LVF, RVF, or BVF [or redundant/RED]) for each experiment.1 Tests of the race model inequality (Miller, 1982) were also undertaken for RT data using the RMITest program (Ulrich, Miller, & Schröter, 2007), which uses singletarget RTs to determine whether redundant RTs are faster than would be predicted by any race model. Accuracy data were used to perform 1 The ANOVA for accuracy was conducted on both raw and arcsin-transformed data. These analyses did not lead to qualitatively different results; thus the results of the raw data ANOVA are reported.

P. Shepherdson, J. Miller / Acta Psychologica 148 (2014) 96–106

99

signal detection analyses, in which target-present trials for each word presentation condition were considered to be signals (S), and targetabsent trials noise (N), and ANOVA was used to determine whether sensitivity differed across conditions. Criterion α-level for all statistical tests was set at .05, with Greenhouse–Geisser corrections applied where relevant.

(BVF) RTs and RTs from each participant's best single-target condition. Responses to LVF trials (720 ms) were significantly slower than those to RVF (671 ms) trials [t(23) = 2.08, p = .049]. Responses to BVF trials (630 ms) were non-significantly faster than those from each participant's best single-target condition (644 ms) [t(23) = 2.05, p = .051].

2.2. Results

2.2.2. Accuracy data Mean percentages of correct responses are displayed in Fig. 1b. Accuracy in LVF trials (62.7%) was significantly lower than in either RVF (75.8%) or BVF (81.3%) trials, which also differed significantly [F(2, 46) = 21.58, MSE = 202.2, p b .001, followed by Holm– Bonferroni-corrected pairwise comparisons]. There was no significant difference between accuracy in target-present (71.9%) and target-absent (74.6%) trials [F(1, 23) = 0.90, MSE = 300.9, p = .35]; however, there was a significant Target × Presentation interaction [F(1, 46) = 25.62, MSE = 119.0, p b .001]. Separate one-way ANOVA showed significant effects of Presentation condition in target-present trials [F(2, 46) = 30.54, MSE = 240.3, p b .001], but not in target-absent trials [F(2, 46) = 0.91, MSE = 81.0, p = .41]. For target-present trials, accuracy was significantly higher in RVF (75.8%) than LVF (52.8%) trials [t(23) = 3.77, p = .001]. It was also significantly higher in BVF trials (87.1%) than in the average of each participant's best single-target condition (80.9%) [t(23) = 3.16, p = .004]. SDT values are displayed in Table 1. Due to one participant's perfect response rate in RVF trials, Hautus and Lee's (2006)'s adaptive estimator method was used to provide values for that condition. A one-way ANOVA on d′ showed a significant effect [F(2, 46) = 13.98, MSE = 0.64, p b .001]; post-hoc Holm–Bonferroni-corrected pairwise comparisons revealed that sensitivity was significantly lower in LVF trials than in either RVF or BVF trials, which did not significantly differ from one another. An analogous analysis of c also showed a significant effect [F(2, 46) = 22.92, MSE = 0.07, p b .001]; Holm–Bonferroni-corrected pairwise comparisons showed that all three conditions displayed bias significantly different from one another: a bias toward the target-absent response in LVF trials, one toward the target-present response in BVF trials, and a very slight target-absent bias in RVF trials.

2.2.1. Latency data Mean RTs for correct responses (approximately 73% of total responses) are displayed in Fig. 1a. Target-present responses (674 ms) were significantly faster than target-absent responses (766 ms) [F(1, 23) = 93.83, MSE = 3302.4, p b .001]. There was also a significant main effect of Presentation [F(2, 46) = 9.51, MSE = 3756.5, p = .001]; follow-up Holm–Bonferroni-corrected pairwise comparisons revealed significant differences between LVF (748 ms), BVF (693 ms), and RVF RTs (719 ms). Importantly, there was also a significant Target × Presentation interaction [F(2, 46) = 11.96, MSE = 1294.3, p b .001]. Separate one-way ANOVA showed a significant effect of Presentation condition for target-present responses [F(2, 46) = 12.20, MSE = 4026.3, p = .001], but not for target-absent responses [F(2, 46) = 2.04, MSE = 1024.4, p = .14]. We ran two separate planned comparisons to assess the effects of Presentation condition in target-present trials: one comparing RTs from the two single-target conditions, and one using Biederman and Checkosky's (1970) comparison between redundant

a

b

2.2.3. Force data Analysis of peak response force data from correct responses showed that responses were not significantly more forceful in RVF (403 cN) than in LVF (401 cN) trials, but that responses to BVF trials (415 cN) were significantly more forceful than both unilateral conditions [F(2, 46) = 5.53, MSE = 485.6, p = .007, followed by Tukey's tests]. There was no significant main effect of Target [F(1, 23) = 1.15, MSE = 51,806.8, p = .29], nor was the Target × Presentation interaction quite significant [F(2, 46) = 3.05, MSE = 666.03, p = .066].

2.2.4. Test of the race model inequality Every participant provided at least 10 correct responses in each condition, allowing us to test the race model inequality over 10 quantiles (.05,.15,.25…,.95). This showed no significant violations over any of the quantiles tested.

Table 1 Sensitivity and bias for the three presentation conditions in Experiments 1–3. Experiment 1 2 Fig. 1. Latency data for correct responses (panel a) and accuracy data (panel b) in Experiment 1. Error bars show ±1 standard error of a difference between two points, computed by pooling error terms for Target and Presentation factors, and their interaction.

3

d′ c d′ c d′ c

LVF/HI

RVF/LO

BVF/RED

0.74 0.28 1.00 0.42 2.56 0.08

1.49 0.02 1.95 0.14 1.92 0.24

1.94 −0.22 2.26 −0.32 3.58 −0.57

100

P. Shepherdson, J. Miller / Acta Psychologica 148 (2014) 96–106

2.3. Discussion Experiment 1 was designed to determine whether redundancy gain would be found in a semantic categorisation task requiring higher level processing of two separate lexical entries. Results show that it can: both accuracy and latency results showed that performance was enhanced in redundant compared to single-target trials (though non-significantly so for latency). Interestingly, this appears to have been the case only for positive (that is, target-present) stimuli. This mirrors the findings from lexical decision (e.g., Mohr et al., 1996) and fame judgement (e.g., Baird & Burton, 2008) tasks, as described in the Introduction. Response force was also significantly greater in BVF trials compared to LVF and RVF trials, though unlike latency and accuracy gains, this was not specific to target-present responses. It is unclear as to why this would be the case; possibly it implies some form of motor coactivation (e.g., Giray & Ulrich, 1993), but if so it seems strange that this would not also be evident in RT enhancement for redundant non-target trials. We can offer no plausible explanation for this finding. Signal detection analysis revealed varying levels of sensitivity and bias for the three presentation conditions. Tellingly, the bias toward a target-absent response for LVF trials seems to preclude any suggestion that the low accuracy levels in LVF-present trials reflected a chance level of performance (which would have required a far lower percent correct value for that condition and its LVF-absent counterpart). Although the results of this experiment demonstrate redundancy gain, it could be argued that the observed gain was actually a rather uninteresting artefact of participants pursuing a strategy whereby they attend to only one stimulus location on each trial. This “single attended location” account could explain the superior response accuracy in redundant trials—where the attended display location is always inhabited by an informative stimulus—relative to performance in single-target trials—where participants must guess when they attend to a noninformative display location. It could also explain the shorter response latencies in redundant trials, under the plausible additional assumption that participants make rather slow guess responses when an uninformative stimulus is attended, which would happen only in single-word trials. Three pieces of evidence suggest, however, that the single attended location account cannot explain all of the redundancy gain observed in this experiment. First, redundancy gain was larger with target-present stimuli than with target-absent ones. If participants responded accurately when they attended to a word and guessed when they attended to a nonword, then the artefactual redundancy gain should have been present for both stimulus types. Second, accuracy for single-word trials seems to exceed what would be possible if participants were attending to a single location. Consider a simple model in which participants process the stimulus at one location and correctly identify it with probability p. In redundant trials, where the attended location is always informative, the probability of making a correct response is simply p. In single-word trials, the probability of making a correct response is (p + .5) / 2, because half the time a participant attends to the uninformative location and must guess. Table 2 shows the observed percent correct values for single-word and redundant trials, along with the values that would be predicted for singleword trials according to such a model (computed by substituting the

observed redundant-word percent correct values for p). The observed single-word performance is significantly superior to the model's prediction, indicating that this model does not provide a complete account of the difference in accuracy between single-word and redundant trials. Third, a comparison of the RT distributions for the single-word and redundant conditions also provides evidence against the single attended location account. If guesses tend to be rather slow, as must be assumed to explain the slower responses in single-word than redundant trials, then accuracy should be markedly lower for the slowest responses and this effect should be most pronounced for single-word stimuli. However, as is shown in Fig. 2a, the dropoff in response accuracy for the slowest responses is not especially large with single-word stimuli. Instead, response accuracy changes across RT quantiles quite

a

b

c

Table 2 Observed percent correct values for single-word (O-Single) and redundant (O-Redundant) trials in Experiments 1–3, along with single-word percent correct values predicted by a model where participants attend to a single stimulus on each trial (P-Single). The column showing p-values reports the outcomes of within-subjects t-tests comparing O-Single and P-Single in each experiment. Experiment

O-Redundant

O-Single

P-Single

p

1 2 3

81.3 84.6 92.7

69.3 74.7 84.7

65.7 67.3 71.4

.001 b.001 b.001

Fig. 2. Percentage of correct responses across each RT quantile (1 = fast, 8 = slow) in Experiments 1 (a), 2 (b), and 3(c), for single-word and redundant conditions. Error bars show ±1 standard error of a difference between two points, computed by pooling error terms for trial type and quantile factors, and their interaction.

P. Shepherdson, J. Miller / Acta Psychologica 148 (2014) 96–106

similarly for single-word and redundant stimuli, with both conditions appearing to contain some fast guesses and then a gradual accuracy decrease from fast to slow quantiles, consistent with item difficulty effects (i.e., harder items are processed more slowly and less accurately). Given that the single attended location account seems insufficient to account for the redundancy gain observed in this experiment, the question remains as to what mechanism(s) might be responsible for this gain. As discussed earlier, in previous experiments using the lexical decision task both redundancy gain and the asymmetry between positive and negative stimuli were explained via the cell assembly model: only items with existing neural representations (words, and famous faces) should be able to benefit from dual stimulation. An analogous argument might be made to fit our results in with the cell assembly model as follows. Redundant targets came from the same category, for which a neural representation would be likely to exist. However, redundant non-targets were from a variety of disparate categories, and as such no existing neural representation would be expected. Consequently, redundant targets could have coactivated the neural representation of the superordinate category to which they belonged. As this would not have been possible for nontargets, they did not show comparable gains. On the other hand, though this provides a way to reconcile our data with the cell assembly model, the lack of any violations of Miller's (1982) race model inequality means that a race model explanation cannot be ruled out, and can therefore be considered equally plausible.2 To explain our results a race model would also need to contend with the target/non-target asymmetry. This could be accomplished by a model where processing is terminated when a target is identified (e.g., a parallel self-terminating race model, as described in the Introduction). In redundant target trials, a response is enacted once the fastest stimulus is successfully processed. In redundant non-target trials, a response is only enacted once the slowest stimulus has been processed. Consequently, target-present responses benefit from statistical facilitation in redundant trials, whereas target-absent responses do not. It is also worth noting that a serial self-terminating model should show a similar pattern of results to the race model just described. Fiedler et al. (2013) described the difficulties in distinguishing serial from parallel models, and in many ways these difficulties apply to our results also. However, as we used a choice, rather than a go/no-go task, we have the advantage of being able to test RT data from nontarget trials against predictions such a model would make. In particular, if processing in trials with one or two targets is self-terminating, whereas processing in trials with no targets is exhaustive, there should be differences in RT variability between different types of trials. Specifically, RTs should be less variable in non-target trials (where a response is made after processing of both stimuli is complete) and redundant target trials (where a response is made after processing of the first stimulus is complete) than in single-target trials (where a response will be made after the processing of the target, which can occur either first or second).3 This prediction received partial support from our data: A t-test comparing the RT SDs from singletarget trials (170 ms) to the BVF-Present SD (145 ms) showed that the former were more variable [t(23) = 3.49, p = .002], while a second comparison between single-target RT SDs and those from non-target trials (162 ms) showed no significant difference [t(23) = 1.20, p = .24]. As the general trend in both cases was in the predicted direction, we will return to this issue when reporting the experiments to come. Obviously, had we observed RMI violations we would be better able to determine which of these accounts is more suitable. That we did not is not greatly surprising. Due to the high error rates in some conditions, 2 As noted in the Introduction, though violations of the RMI rule out race models, a failure to find violations does not rule out coactivation models. 3 See Townsend and Ashby (1983, p. 193) for an analogous argument relating to visual search.

101

we had limited data available to conduct RMI tests. The following experiment was aimed at rectifying this, by changing task conditions to decrease error rates, thus increasing our chances of observing RMI violations should a race model not be appropriate. Of course, even this does not guarantee RMI violations when coactivation models are correct, because the RMI is statistically conservative and there are even coactivation models that predict no violations of it. 3. Experiment 2 Experiment 1 showed that redundancy gain can occur in a task requiring higher level processing of two superordinate information sources from distinct lexical entries. Unfortunately, the high error rates in some conditions (particularly when targets were presented unilaterally to the left visual field) meant that tests of the RMI—which have the potential to rule out race models—were not very powerful due to the small numbers of correct trials. Consequently, in Experiment 2 we provided participants with a longer window during which to make their responses (3000 ms from stimulus onset, as opposed to 1500 ms in Experiment 1). We also provided a stronger incentive to respond accurately, in the form of a time penalty for incorrect responses (prolonged error feedback). We expected that these manipulations would increase the proportions of correct responses and thus provide a better opportunity to properly test the race model inequality. 3.1. Method 3.1.1. Participants Twenty-four undergraduate psychology students (20 females, 4 males; 19 right-handed, 5 left-handed; aged 18–48 years) took part in this experiment. None had participated in Experiment 1. Participation occurred through the Department of Psychology's experimental participation programme, and participants received partial course credit for their time. 3.1.2. Apparatus and stimuli, procedure, and data analysis Apparatus and stimuli were the same as in Experiment 1. Procedure was altered in only a few respects. First, the response window was increased from 1500 to 3000 ms. Second, error messages appearing after incorrect responses were presented for an extended duration, slowing participants' progress. These two factors were designed to give participants both the opportunity and the incentive to respond as accurately as possible. Participants were also informed of the length of the response window and the time penalty for errors, and instructed to respond as accurately and as quickly as they could. Third, given the longer response window, the number of blocks was reduced to eight (plus a practice block to begin with), so that each session could still be completed in roughly 45 min. Data analysis was identical to Experiment 1, with the addition of comparisons between the two experiments. 3.2. Results 3.2.1. Latency data Mean RTs for correct responses (approximately 78% of total responses) in each condition are displayed in Fig. 3a. ANOVA showed a significant effect of Presentation condition [F(2,46) = 8.98, MSE = 4097.5, p = .002]; follow-up Holm–Bonferroni-corrected pairwise comparisons revealed that responses in BVF trials (839 ms) were significantly faster than those in LVF trials (894 ms), but neither LVF nor BVF responses differed significantly from responses in RVF trials (860 ms). Responses in target-present trials (752 ms) were faster than responses in targetabsent trials (977 ms) [F(1,23) = 55.85, MSE = 32,594.9, p b .001]. A significant target-by-presentation interaction was also apparent [F(2,46) = 13.71, MSE = 4308.0, p b .001]. Separate one-way ANOVA showed a significant effect of presentation condition for target-present

102

P. Shepherdson, J. Miller / Acta Psychologica 148 (2014) 96–106

a

comparisons showed that sensitivity was significantly lower in the LVF condition than the RVF condition, and lower in both of these than in the BVF condition.

3.2.3. Force data Peak force values from correct responses for each trial type are displayed in Fig. 4. There were non-significantly more forceful target-present (540 cN) than target-absent (463 cN) responses [F(1,23) = 4.26, MSE = 50,632.0, p = .051], and a significant main effect of Presentation [F(2,46) = 5.44, MSE = 411.2, p = .008], which Tukey's tests showed was due to significantly more forceful responses in BVF (508 cN) than LVF (494 cN) trials; RVF response force (502 cN) did not differ significantly from either LVF or BVF values. There was also a significant Target × Presentation interaction [F(2,46) = 9.25, MSE = 254.4, p = .001]: target-absent responses were more forceful in LVF (464 cN) than RVF (459 cN) trials, but the reverse was true for target-present responses (LVF: 525 cN; RVF: 545 cN).

b

3.2.4. Test of the race model inequality A test of the RMI showed no significant violations across any of the sixteen quantiles (.031,.094,.156,…,.969) tested.

3.2.5. Between-experiment comparisons 2 (Experiment) × 2 (Target) × 3 (Presentation) ANOVA were conducted on latency, accuracy, and force data. Only those effects involving the Experiment factor, and others of particular interest, are noted here.

Fig. 3. Latency data for correct responses (panel a) and accuracy data (panel b) in Experiment 2. Error bars were computed as in Fig. 1.

trials [F(2,46) = 13.96, MSE = 6769.3, p b .001], but not for targetabsent trials [F(2,46) = 0.83, MSE = 1636.2, p = .44]. A planned comparison between LVF and RVF target-present RTs showed a significant RVF advantage [t(23) = 2.33, p = .029]. A second planned comparison between the redundant target-present RT (mean = 693 ms), and the fastest single target-present RT for each participant (mean = 707 ms), showed that responses to redundant targets were not significantly faster [t(23) = 1.65, p = 0.11]. For RT SDs, single-target (239 ms) RTs were non-significantly more variable than those in the BVF-present (208 ms) condition [t(23) = 2.04, p = .052], whereas a comparison between singletarget and target-absent (287 ms) trials showed SDs in the former to be significantly smaller [t(23) = 4.57, p b .001]. 3.2.2. Accuracy data Percentage correct values for each trial type are displayed in Fig. 3b. Results of the ANOVA on accuracy were generally similar to those from Experiment 1, so here we only report patterns of results that differed qualitatively. First, Holm–Bonferroni-corrected pairwise comparisons showed that RVF (81.2%) and BVF (84.6%) trial accuracy values did not differ significantly. Second, a one-way ANOVA for target-absent trials showed a significant effect of Presentation [F(2,46) = 9.19, MSE = 27.42, p = .001], due to higher accuracy in the RVF condition (84.1%) than the LVF (79.7%) or BVF (77.8%) conditions. As per Experiment 1, a signal detection analysis was also undertaken, the results of which appear in Table 1. A one-way ANOVA showed a significant effect of Presentation for d′ [F(2,46) = 30.46, MSE = 0.34, p b .001]; Holm–Bonferroni-corrected pairwise

3.2.5.1. Latency. RT data showed that correct responses were significantly slower in Experiment 2 (864 ms) than in Experiment 1 (720 ms) [F(1,46) = 20.02, MSE = 74,755.2, p b .001]. This difference was more pronounced for target-absent responses (Δ = 211 ms) than targetpresent responses (Δ = 78 ms) [F(1,46) = 17.50, MSE = 17,948.7, p b .001]. Neither the Experiment × Presentation nor the three-way interaction was significant (p of .934 and .108, respectively). A planned comparison between the redundant target-present RT for the summed data from both experiments (mean = 660 ms), and the fastest single target-present RT for each participant (mean = 677 ms), showed that responses to redundant targets were significantly faster [t(47) = 3.07, p = .004].4

3.2.5.2. Accuracy. For percent correct data, responses were significantly more frequently correct in Experiment 2 (78%) than in Experiment 1 (73.3%) [F(1,46) = 5.96, MSE = 270.4, p = .019]; however, none of the interactions involving the Experiment factor were significant (all p N .4). 3.2.5.3. Force. Neither the main effect of Experiment nor any of the interactions involving that factor were significant (all p N .2), though a significant main effect of Presentation [F(2,92) = 10.29, MSE = 448.4, p b .001, followed by Tukey's tests] showed that BVF responses were more forceful than RVF or LVF responses.

3.2.5.4. Test of the race model inequality. A test of the RMI showed no significant violations across any of the ten quantiles (.05,.15,.25,…,.95) tested using data from all subjects in Experiments 1 and 2. 4 To address possible concerns about using a t-test on a combined sample whose size was not predetermined, we also conducted a Bayesian analysis to assess the extent to which the combined data supported the hypothesis that responses to redundant targets were faster than the fastest single-target response. The value of the JZS Bayes factor—recommended for use with the one-sample t-test by Rouder, Speckman, Sun, Morey, and Iverson (2009)—was 0.13, indicating approximately 8:1 support for that hypothesis.

P. Shepherdson, J. Miller / Acta Psychologica 148 (2014) 96–106

103

than single-target conditions). This indicates that the self-terminating serial model requires some modification if it is to account for our results.

4. Experiment 3

Fig. 4. Mean peak force from correct trials in Experiment 2. Error bars were computed as in Fig. 1.

3.3. Discussion Response accuracy was improved in Experiment 2 over Experiment 1, with the window within which participants could respond doubled from 1500 to 3000 ms. However, perhaps because increases in response accuracy in Experiment 2 were accompanied by increases in response latency, our aim of demonstrating a stronger RT redundancy gain than was found in Experiment 1 was thwarted. Nonetheless, there were sufficient correct responses in Experiment 2 to test the race model inequality over a greater number of quantiles—16, compared to 10 in Experiment 1—and again no violations of the inequality were found. This was also the case when the data from the two experiments were combined. Thus, both race and coactivation models are consistent with the RG found in the target-present trials of this experiment since only race model violations are fully diagnostic. Though combining the data from Experiments 1 and 2 did not lead to any qualitative changes in the results of race model inequality tests, it did lead the previously non-significant differences between the redundant-present RT and fastest single-target RT from each participant to reach statistical significance. This implies that gains in RT were not simply a result of participants having different preferred stimulus locations (Biederman & Checkosky, 1970). It is also worth noting that, as in Experiment 1, the single attended location account seems unable to explain the results found in this experiment. Table 2 shows that single-word accuracy values predicted by this account are significantly lower than the observed values, and Fig. 2b shows that response accuracy dropped off markedly for both redundant and single-word trials in the slowest RT quantile, inconsistent with the idea that slow guesses are responsible for RT differences between conditions. Force data displayed generally the same pattern as in Experiment 1, though here response force in BVF trials was not significantly greater than in LVF or RVF trials (either overall, or for target-present and target-absent trials separately). Combined with the significant effect of Presentation found in the between-experiment analysis, this suggests a consistent, though small effect of redundancy on response force. As mentioned previously, it is difficult to reconcile this finding with the gains for accuracy and latency, which seem specific to target-present trials. Finally, analysis of RT SDs showed a pattern of results inconsistent with the predictions of a self-terminating serial model. As stated in the Discussion for Experiment 1, this model would predict lower variability in RTs for redundant target and all non-target conditions, compared to single-target conditions. However, SDs in Experiment 2 were significantly larger in non-target conditions than single-target conditions (though they were non-significantly smaller for redundant

Experiments 1 and 2 both involved lateralised presentation of stimuli. This mirrored display conditions in most of the LDT experiments undertaken by Mohr and Pulvermüller (e.g., Mohr, Pulvermüller, & Zaidel, 1994, Mohr et al., 1996). Lateralised presentation is designed to ensure that the two cerebral hemispheres each receive initial access to different stimuli, potentially allowing the hemispheres to operate independently, to inhibit one another, or to cooperate. Mohr, Pulvermüller, & Zaidel (1994) suggested that redundancy gain in LDT with lateralised presentation supports a hemispheric cooperation account, such that the two hemispheres work together so that processing is faster and more likely to be accurate in bilateral/redundant trials. To what extent might such interhemispheric cooperation be necessary for the redundancy gains found in semantic categorisation in Experiments 1 and 2? To answer this question, our third experiment was identical to Experiment 1 except that stimuli were presented directly above and below fixation (HI and LO, respectively). If redundancy gain in semantic categorisation relies on the two hemispheres receiving and processing separate stimuli, then this non-lateralised presentation should at least reduce the extent of such gains. On the other hand, if a redundancy gain persists at a similar magnitude with this presentation format, this

a

b

Fig. 5. Latency data for correct responses (panel a), and accuracy data (panel b) for Experiment 3. Error bars were computed as in Fig. 1; where they are not visible, they are smaller than the symbols used to plot the data points.

104

P. Shepherdson, J. Miller / Acta Psychologica 148 (2014) 96–106

would imply that its origin is likely to be at least partially independent of any interaction between cerebral hemispheres. The decision to present stimuli directly above and below fixation was made predominantly because it allowed each stimulus to be located as close to foveal vision as possible, thus avoiding presentation to separate hemispheres. This form of presentation represents a departure from previous redundant target experiments involving upper- and lower-visual field presentation of word stimuli, where stimuli have been presented off-centre to assess the effects of redundant presentation within a single visual field (e.g., Mohr et al., 1996). Here, though, we were more interested in seeing whether a redundancy gain would obtain without the requirement that stimuli be presented separately to the left and right visual fields, than in comparing the contributions of the two separate hemispheres to task performance. 4.1. Method 4.1.1. Participants Twenty-four undergraduate psychology students (16 females, eight males; aged between 18 and 48 years of age; two left-handed, 22 right-handed) from the University of Otago completed this experiment, in return for partial course credit, as part of the Department of Psychology's experimental participation programme. None had participated in Experiments 1 or 2. 4.1.2. Apparatus and stimuli, procedure, and data analysis Apparatus and stimuli were the same as in Experiment 1, with the exception of stimulus location. In this experiment, stimuli were presented with their centres 14 pixels (approximately 1°) above or below fixation. Procedure, instructions, and data analysis were the same as in Experiment 1. 4.2. Results 4.2.1. Latency data Participant mean RTs for correct trials (approximately 87% of total trials) in each condition are displayed in Fig. 5a. Responses to target-present trials (mean RT = 641 ms) were faster than those to target-absent trials (750 ms) [F(1,23) = 173.55, MSE = 2477.8, p b .001], and responses to LO trials (727 ms) were slower than those to HI (685 ms) and redundant (675 ms) trials [F(2,46) = 48.21, MSE = 761.3, p b .001]. There was also a significant Target-by-Presentation interaction [F(2,46) = 64.61, MSE = 584.5, p b .001]. Separate one-way ANOVA showed a significant effect of Presentation for target-present trials [F(2,46) = 70.90, MSE = 1047.1, p b .001], but not for target-absent trials [F(2,46) = 0.77, MSE = 298.7, p = .46]. Again, two planned comparisons were undertaken for the target-present trials. The first showed that responses in HI trials were significantly faster than those in LO trials [t(23) = 7.07, p b .001]. The second showed that responses to redundant target-present trials (mean = 596 ms) were significantly faster [t(23) = 4.41, p b .001] than each participant's faster single target-present condition (mean = 622 ms). RT SDs for single-target (137 ms) conditions were significantly more variable than those in the redundant target-present (118 ms) condition [t(23) = 3.28, p = .003], but non-significantly less variable than those in target-absent (145 ms) conditions [t(23) = 1.88, p = .07]. 4.2.2. Accuracy data Percentage correct means for each trial type are displayed in Fig. 5b. More accurate responses were made in target-absent (88.7% correct) than in target-present (86.0%) trials [F(1,23) = 5.51, MSE = 46.50, p = .028], and responses to LO trials (80.7%) were less accurate than responses for HI (88.6%) and redundant (92.7%) trials [F(2,46) = 32.10, MSE = 55.41, p b .001]. There was also a significant target-by-presentation interaction [F(2,46) = 35.21, MSE =

45.54, p b .001]. Separate one-way ANOVA showed a significant effect of Presentation for target-present trials [F(2,46) = 37.98, MSE = 88.09, p b .001], but no significant effect for target-absent trials [F(2,46) = 2.94, MSE = 12.87, p = .065]. Planned comparisons for target-present trials showed that responses were significantly more frequently correct in HI than LO trials [t(23) = 4.04, p = .001], and that redundant trials (97.3% correct) were significantly more frequently correct than the more accurate single target-present trial type for each participant (mean accuracy = 88.5%) [t(23)=9.18, p b .001]. Results of the SDT analysis appear in Table 1. Six participants had 100% accuracy rates in BVF-present trials; thus, we used the “adaptive estimator” method described by Hautus and Lee (2006) in calculating sensitivity and bias for that condition. ANOVA using d′ showed a significant effect [F(2,46) = 38.94, MSE = 0.43, p b .001]: according to posthoc Tukey's tests, all three conditions had significantly different sensitivity from one another. There was also a significant effect in the ANOVA using c [F(2,46) = 52.94, MSE = 0.08, p b .001]; Tukey's tests showed that the target-present bias in redundant trials was significantly different from the target-absent biases in HI and LO trials, which did not differ significantly from each other. The numerically higher overall proportion of correct responses in this experiment (approximately 87%) compared to Experiments 1 and 2 (approximately 73% and 78%) suggests that participants found the task easier with stimuli presented above and below fixation, rather than to the left and right. This is perhaps not surprising; the offset from fixation to stimulus centre here was approximately 1°, compared to 3.7° of visual angle in Experiments 1 and 2, meaning that stimuli appeared less peripherally in Experiment 3. 4.2.3. Force data Analysis of peak force data showed no significant main effects or interactions (all p N .3). 4.2.4. Test of the race model inequality No significant violations of the RMI occurred at any of the twenty quantiles evenly spaced across the CDF (0.025, 0.075,…,0.975), implying that a race-model explanation could not be ruled out in accounting for the redundancy gain observed in target-present trials. 4.3. Discussion Once again, results of Experiment 3 show that redundancy gain can be obtained in a semantic categorisation task, and that this gain cannot be explained by the single attended location account (see Table 2 and Fig. 2c.) The redundancy gain found here provides evidence that the findings of Experiments 1 and 2 were not simply a result of the lateralised presentation used there, but instead an illustration of a general property of the processes involved in extracting semantic information from written words. In fact, the magnitude of the gain in Experiment 3 (26 ms faster and 8.8% more accurate responses on redundant trials than the average of the fastest and most accurate single-target condition for each participant) was numerically greater than the gains found in either of the preceding experiments. As far as the clearer redundancy gain for latency in this experiment is concerned, this may have a rather simple explanation: as participants made more correct responses (as evident from both higher percent correct values, and higher levels of sensitivity in the SDT analysis), mean RTs for each participant were likely less variable, and thus the data should have been less “noisy”. Importantly, though performance in LO trials appeared worse than in HI or redundant trials, it came nowhere close to approaching the depths of LVF performance in Experiments 1 and 2. As such, participant averages which might have come from a very small number of data points in those experiments were less likely to be found here, making the effects of our experimental manipulations more evident.

P. Shepherdson, J. Miller / Acta Psychologica 148 (2014) 96–106

However, the same phenomenon did not clarify the situation with regard to response force, which showed no effects approaching significance in this experiment. This might indicate that the effects found in Experiments 1 and 2 are somehow specific to lateralised presentation, though without a direct test this suggestion is obviously speculative. Though the extent of redundancy gain for accuracy and latency increased in this experiment, there were still no significant RMI violations in the RT data. As such, a statistical facilitation explanation for redundancy gain cannot be ruled out on this basis. However, once again RT SDs failed to accord with the predictions of the serial self-terminating model, suggesting that this account of our findings is likely inappropriate. 5. General discussion We had two main aims in conducting the experiments reported here. First, we wanted to determine whether a redundancy gain could be obtained in a task requiring higher level processing of superordinate information from two lexical entries, thus providing some indication of the extent to which redundancy gain is a widespread phenomenon in higher level tasks. Second, if such a gain was present, we wanted to investigate what accounts might provide a plausible explanation of that finding. We discuss each of these issues in turn next. As the results of Experiments 1–3 make clear, redundancy gain can occur in a task requiring semantic categorisation of two non-identical word stimuli—a task which unambiguously requires higher level processing of superordinate information about both stimuli. Gains in accuracy were significant in all three experiments, while gains in latency of 14–26 ms were found in all three experiments, although these were only statistically reliable in Experiment 3 (being non-significant for Experiments 1 and 2 considered separately, but highly significant in their combined analysis). Similarly, there were gains in response force of 1–12 cN across the three experiments, though these were only statistically reliable in Experiment 1. Moreover, the observed redundancy gains clearly cannot be explained by the single attended location account. Relative to redundant trials, response accuracy was significantly higher in single-word trials than is predicted by this account (Table 2), and the pattern of accuracy levels across RT quantiles did not vary between single-word and redundant trials as predicted by this account (Fig. 2). In particular, as Fig. 2 shows, the tendency for accuracy to decrease in the slowest quantile occurred across both single-target and redundant-target trials in all three experiments, suggesting that slow guesses do not explain the overall RT differences between the single-word and redundant conditions. Redundancy gain only occurred for positive (that is, target-present) trials; responses in redundant negative trials (that is, those with two non-target words presented) appeared to be neither faster nor more accurate than responses in single-word negative trials. A similar asymmetry between positive and negative stimuli (redundancy gain for the former, but not the latter) was found in LDT (e.g., Mohr et al., 1996) and fame judgement (e.g., Schweinberger et al., 2003) experiments. This might suggest that similar mechanisms underlie redundancy gain in these three domains. Further supporting that idea is the fact that in both tasks, performance was superior in trials where a single target was presented to the right visual field than those where a target was presented to the left (what Mohr et al., 1996, referred to as a right visual field advantage, or RVFA). As mentioned in the Introduction, Mohr et al. (e.g., Mohr, Pulvermüller, Rayman et al., 1994, Mohr, Pulvermüller, & Zaidel, 1994, Mohr et al., 1996) explained redundancy gain in lexical decision by way of an account in which the presentation of redundant targets provides extra stimulation to words' neural representations, leading to faster and more effective responses. This account was then extended to fame judgements for faces (Mohr et al., 2002). According to these authors, asymmetrical redundancy gains in both tasks were due to the fact that only familiar stimuli should possess neural representations which can benefit from the extra stimulation redundant targets provide.

105

Non-words and the faces of unfamiliar people do not have pre-existing representations, and thus they provide no opportunity for a summation of neural activity. Initially, extrapolating this account to the semantic categorisation task appears problematic, because though redundant stimuli in our experiments belonged to the same semantic category, they were always different words. This means that they should have had differing neural representations, thus preventing the summation of activity described by Mohr et al. (1996). However, one possible resolution to this apparent inconsistency would be that here both redundant targets provided stimulation to the neural representation of the superordinate category to which they belonged—that is, the cell assembly for animals. If this were the case, it would also explain the lack of redundancy gain in negative trials, where redundant stimuli were from a diverse range of categories (given the broad “non-animal” nature of non-target stimuli): there was no pre-existing superordinate “non-target” assembly to which stimulation could be redundantly provided. Of course, though a cell assembly coactivation account might thus be consistent with the results of our semantic categorisation experiments, this need not imply it is the only nor even the most appropriate account. As we discussed in the Introduction, the findings from lexical decision and fame judgement tasks also seem wholly consistent with race model explanations. Authors of those studies did not conduct any tests of the race model inequality (Miller, 1982), which tends to be the most common means of ruling out race models. We did perform such tests, however, and found no violations in any of Experiments 1–3. This means that race models cannot be discounted on the basis of faster-than-predicted RTs. In fact, a self-terminating race model of the sort mentioned in the Introduction would seem as suitable an account for results of the semantic categorisation task as it is for results of LDTs and fame judgement tasks, and as suitable as the cell assembly coactivation account. According to this model, target-present responses are made as soon as any stimulus is identified as a target. This means that redundant RTs are faster than single-target RTs due to statistical facilitation (Raab, 1962). By contrast, target-absent responses only occur once some sort of temporal threshold has been reached. This means that trials with two non-target stimuli would evoke responses no faster than trials with one non-target stimulus and one distractor. It is also possible that a similar model with serial rather than parallel processing could account for redundancy gain in semantic categorisation. As stimulus number was held constant from single-target to redundant trials, in a serial processing context redundancy gain could result from the fact that the first stimulus to be processed would always be a target on redundant trials, yet not on single-target trials. If processing is selfterminating (that is, if a response is made once any target is identified), this would make responses to redundant trials faster than those to single-target trials (i.e., it would lead to redundancy gain). However, such a model makes a prediction which was not borne out in our results: there should be less RT variability in target-absent trials than in singletarget trials, because the former always require processing of both display elements whereas the latter may stop after only a single display element is processed (depending on whether a target or a distractor is processed first). Contrary to this prediction, RT standard deviations tended not to significantly differ between these conditions (Experiments 1 and 3), or to differ significantly in the opposite direction (Experiment 2). Note also that this model would not be appropriate for explaining redundancy gain in LDT or fame judgement experiments (e.g., Baird & Burton, 2008; Mohr et al., 1996). This is because there were no distractor stimuli in these experiments, meaning that on both single-target and redundant trials the first stimulus to be processed would always have been a target. As such, given the similarity of the patterns of results in these experiments to our semantic categorisation results, either the parallel self-terminating race or cell assembly coactivation models described above can be considered preferable, as they can account for results from all three tasks (LDT, fame judgement, semantic categorisation).

106

P. Shepherdson, J. Miller / Acta Psychologica 148 (2014) 96–106

Finally, our results differ from those of Mullin and Egeth (1989), who also conducted semantic categorisation experiments with redundant lexical stimuli, and found no gains. As mentioned in the Introduction, Mullin and Egeth's task involved a high level of stimulus repetition (only five target and five non-target words for each category), and this might be partially responsible for the difference. In addition, as mentioned by Fiedler et al. (2013), owing to a confound between stimulus number and target number Mullin and Egeth's experiments also had a higher perceptual processing load for redundant target than single target trials. This could account for the lack of redundancy gain in their data. Nonetheless, it is unclear why such an issue would have prevented redundancy gain in their semantic categorisation experiments, but not in their lexical decision experiments (nor those of other researchers, for that matter; e.g., Mohr et al., 1996). This issue might be worth exploring in future research. 6. Conclusion Results from our experiments provide a new and unambiguous demonstration of redundancy gain in higher level cognition, consistent with the recent findings of Fiedler et al. (2013) showing redundancy gain involving category and colour information about a single lexical entry. Though our results are consistent with the cell assembly coactivation model proposed by Mohr et al. (e.g., Mohr, Pulvermüller, Rayman et al., 1994; Mohr, Pulvermüller, & Zaidel, 1994; Mohr et al., 1996, 2002), the lack of any RMI violations across Experiments 1–3 means that they are also consistent with a race model account. Determining which of these models provides the best explanation for redundancy gain in semantic categorisation—and higher level tasks in general—may provide a productive path for future research. References Baird, L. M., & Burton, A.M. (2008). The bilateral advantage for famous faces: Interhemispheric communication or competition? Neuropsychologia, 46, 1581–1587, http://dx.doi.org/10.1016/j.neuropsychologia.2008.01.001. Balota, D. A., & Chumbley, J. I. (1984). Are lexical decisions a good measure of lexical access? The role of word frequency in the neglected decision stage. Journal of Experimental Psychology: Human Perception and Performance, 10, 340–357. Balota, D. A., Cortese, M. J., Sergent-Marshall, S. D., Spieler, D. H., & Yap, M. (2004). Visual word recognition of single-syllable words. Journal of Experimental Psychology: General, 133, 283–316, http://dx.doi.org/10.1037/0096-3445.133.2.283. Battig, W., & Montague, W. (1969). Category norms for verbal items in 56 categories. Journal of Experimental Psychology, 80, 1–46. Biederman, I., & Checkosky, S. F. (1970). Processing redundant information. Journal of Experimental Psychology, 83, 486–490, http://dx.doi.org/10.1037/h0028841. Collignon, O., Girard, S., Gosselin, F., Roy, S., Saint-Amour, D., Lassonde, M., et al. (2008). Audio-visual integration of emotion expression. Brain Research, 1242, 126–135. Collignon, O., Girard, S., Gosselin, F., Saint-Amour, D., Lepore, F., & Lassonde, M. (2010). Women process multisensory emotion expressions more efficiently than men. Neuropsychologia, 48, 220–225, http://dx.doi.org/10.1016/j. neuropsychologia.2009.09.007. Fiedler, A., Schröter, H., & Ulrich, R. (2013). Redundancy gain for semantic features. Psychonomic Bulletin & Review, 20, 474–480, http://dx.doi.org/10.3758/ s13423-012-0362-3. Giray, M., & Ulrich, R. (1993). Motor coactivation revealed by response force in divided and focused attention. Journal of Experimental Psychology: Human Perception and Performance, 19, 1278–1291. Grice, G. R., Canham, L., & Boroughs, J. (1984). Combination rule for redundant information in reaction time tasks with divided attention. Perception & Psychophysics, 35, 451–463. Hautus, M. J., & Lee, A. (2006). Estimating sensitivity and bias in a yes/no task. The British Journal of Mathematical and Statistical Psychology, 59, 257–273, http://dx.doi.org/10.1348/000711005X65753. Hebb, D. O. (1949). The organization of behavior. New York: John Wiley & Sons.

Hershenson, M. (1962). Reaction time as a measure of intersensory facilitation. Journal of Experimental Psychology, 63, 289–293. James, C. T. (1975). The role of semantic information in lexical decisions. Journal of Experimental Psychology: Human Perception and Performance, 1, 130–136, http://dx.doi.org/10.1037//0096-1523.1.2.130. Jiang, Y. V., Kwon, M., Shim, W. M., & Won, B. -Y. (2010). Redundancy effects in the perception and memory of visual objects. Visual Cognition, 18, 1233–1252, http://dx.doi.org/10.1080/13506281003791074. Miller, J. (1982). Divided attention: Evidence for coactivation with redundant signals. Cognitive Psychology, 14, 247–279. Miller, J. (1986). Timecourse of coactivation in bimodal divided attention. Perception & Psychophysics, 40, 331–343. Mohr, B., Landgrebe, A., & Schweinberger, S. R. (2002). Interhemispheric cooperation for familiar but not unfamiliar face processing. Neuropsychologia, 40, 1841–1848. Mohr, B., Pulvermüller, F., Mittelstädt, K., & Rayman, J. (1996). Multiple simultaneous stimulus presentation facilitates lexical processing. Neuropsychologia, 34, 1003–1013. Mohr, B., Pulvermüller, F., Rayman, J., & Zaidel, E. (1994). Interhemispheric cooperation during lexical processing is mediated by the corpus callosum: Evidence from the split-brain. Neuroscience Letters, 181, 17–21. Mohr, B., Pulvermüller, F., & Zaidel, E. (1994). Lexical decision after left, right and bilateral presentation of function words, content words and non-words: Evidence for interhemispheric interaction. Neuropsychologia, 32, 105–124. Molholm, S., Ritter, W., Javitt, D. C., & Foxe, J. J. (2004). Multisensory visual–auditory object recognition in humans: A high-density electrical mapping study. Cerebral Cortex, 14, 452–465, http://dx.doi.org/10.1093/cercor/bhh007. Mordkoff, J. T., Miller, J., & Roch, A-C. (1996). Absence of coactivation in the motor component: Evidence from psychophysiological measures of target detection. Journal of Experimental Psychology: Human Perception and Performance, 22, 25–41. Mordkoff, J. T., & Yantis, S. (1991). An interactive race model of divided attention. Journal of Experimental Psychology: Human Perception and Performance, 17, 520–538, http://dx.doi.org/10.1037/0096-1523.17.2.520. Mullin, P. A., & Egeth, H. E. (1989). Capacity limitations in visual word processing. Journal of Experimental Psychology: Human Perception and Performance, 15, 111–123. Oldfield, R. C. (1971). The assessment and analysis of handedness: The Edinburgh inventory. Neuropsychologia, 9, 97–113. Raab, D. (1962). Statistical facilitation of simple reaction times. Transactions of the New York Academy of Sciences, 24, 574–590. Rastle, K., Harrington, J., & Coltheart, M. (2002). 358,534 nonwords: The ARC Nonword Database. The Quarterly journal of experimental psychology. A, Human experimental psychology, 55, 1339–1362, http://dx.doi.org/10.1080/02724980244000099. Rouder, J. N., Speckman, P. L., Sun, D., Morey, R. D., & Iverson, G. (2009). Bayesian t tests for accepting and rejecting the null hypothesis. Psychonomic Bulletin & Review, 16, 225–237, http://dx.doi.org/10.3758/PBR.16.2.225. Savazzi, S., & Marzi, C. A. (2002). Speeding up reaction time with invisible stimuli. Current Biology, 12, 403–407. Schwarz, W., & Ischebeck, A. (1994). Coactivation and statistical facilitation in the detection of lines. Perception, 23, 157–168. Schweinberger, S. R., Baird, L. M., Blümler, M., Kaufmann, J. M., & Mohr, B. (2003). Interhemispheric cooperation for face recognition but not for affective facial expressions. Neuropsychologia, 41, 407–414. Suied, C., Bonneel, N., & Viaud-Delmon, I. (2009). Integration of auditory and visual information in the recognition of realistic objects. Experimental Brain Research, 194, 91–102, http://dx.doi.org/10.1007/s00221-008-1672-6. Tamietto, M., Adenzato, M., Geminiani, G., & de Gelder, B. (2007). Fast recognition of social emotions takes the whole brain: Interhemispheric cooperation in the absence of cerebral asymmetry. Neuropsychologia, 45, 836–843, http://dx.doi.org/10.1016/j. neuropsychologia.2006.08.012. Tamietto, M., Latini Corazzini, L., de Gelder, B., & Geminiani, G. (2006). Functional asymmetry and interhemispheric cooperation in the perception of emotions from facial expressions. Experimental Brain Research, 171, 389–404, http://dx.doi.org/10. 1007/s00221-005-0279-4. Todd, J. W. (1912). Reaction to multiple stimuli. Archives of Psychology, 21. Townsend, J. T., & Ashby, F. G. (1983). The stochastic modeling of elementary psychological processes. Cambridge, UK: Cambridge University Press. Townsend, J., & Nozawa, G. (1997). Serial exhaustive models can violate the race model inequality: Implications for architecture and capacity. Psychological Review, 104, 595–602. Ulrich, R., Miller, J., & Schröter, H. (2007). Testing the race model inequality: An algorithm and computer programs. Behavior Research Methods, 39, 291–302. Veldhuizen, M. G., Shepard, T. G., Wang, M. -F., & Marks, L. E. (2010). Coactivation of gustatory and olfactory signals in flavor perception. Chemical Senses, 35, 121–133, http://dx.doi.org/10.1093/chemse/bjp089. Wiktionary (2006). Wiktionary: Frequency lists. http://en.wiktionary.org/wiki/Wiktionary: Frequencylists