- Email: [email protected]
Visible homonyms are ambiguous, subliminal homonyms are not: A close look at priming
Consciousness and Cognition 20 (2011) 1327–1343
Contents lists available at ScienceDirect
Consciousness and Cognition journal homepage: www.elsevier.com/locate/concog
Visible homonyms are ambiguous, subliminal homonyms are not: A close look at priming Doris Eckstein a,⇑, Matthias Kubat a,b, Walter J. Perrig a a b
Department of Psychology, University of Bern, Switzerland Distance Learning University Switzerland, Brig, Switzerland
a r t i c l e
i n f o
Article history: Received 9 February 2010 Available online 12 June 2011 Keywords: Homonym Polysemy Ambiguity Priming Awareness Consciousness Masked priming
a b s t r a c t Homonyms, i.e. ambiguous words like ‘score’, have different meanings in different contexts. Previous research indicates that all potential meanings of a homonym are first accessed in parallel before one of the meanings is selected in a competitive race. If these processes are automatic, these processes of selection should even be observed when homonyms are shown subliminally. This study measured the time course of subliminal and supraliminal priming by homonyms with a frequent (dominant) and a rare (subordinate) meaning in a neutral context, using a lexical decision task. In the subliminal condition, priming across prime-target asynchronies ranging from 100 ms to 1.5 s indicated that the dominant meaning of homonyms was facilitated and the subordinate meaning was inhibited. This indicates that selection of meaning was much faster with subliminal presentation than with supraliminal presentation. Awareness of a prime might decelerate an otherwise rapid selection process. Ó 2011 Elsevier Inc. All rights reserved.
1. Introduction One of the characteristics of intelligent thinking is flexibility, which is needed when ambiguous information is processed. For example, owing to our ability to resolve ambiguity, we intuitively assign different meanings to the word ‘score’ when talking about exams compared to when talking about music. Ambiguous words like ‘score’ or ‘paper’ are called homonyms, i.e. words that have at least two distinct meanings which cannot be distinguished by either the word’s writing or pronunciation. Due to their ambiguity, homonyms are ideal for investigating presumed stages of word recognition, including access to meaning, selection of meaning and the role of context in this process. Although most studies on homonyms have focused on the influence of context on meaning selection (e.g., Binder & Morris, 1995; Chen & Boland, 2008; Gottlob, Goldinger, Stone, & Van Orden, 1999; Marslen-Wilson, 1987; Moritz, Mersmann, Quast, & Andresen, 2001; Onifer & Swinney, 1981; Paul, Kellas, Martin, & Clark, 1992; Schvaneveldt, Meyer, & Becker, 1976; Simpson, 1981; Simpson & Krueger, 1991; Tabossi & Zardon, 1993; Tanenhaus, Leiman, & Seidenberg, 1979; Vu, Kellas, & Paul, 1998), one strand of research has concentrated on measuring meaning selection when no context or a neutral context is given. In single-word priming, neutral trials are usually formed by pairing homonym primes with unrelated target words (e.g., ‘calf’–‘wine’) and experimental trials are formed by pairing homonym primes with targets related to one of their meanings (e.g., ‘calf’–‘cow’). The difference in reaction times between experimental and biased trials indicates the amount of priming related to the prime-target relationship. A couple of single-word priming time course studies using polar homonyms, i.e. homonyms with a dominant (more frequent) (e.g., ‘calf’–‘cow’) and a subordinate (less
⇑ Corresponding author. Address: Institut für Psychologie, Muesmattstr. 45, 3000 Bern 9, Switzerland. Fax: +41 316318212. E-mail address: [email protected] (D. Eckstein). 1053-8100/$ - see front matter Ó 2011 Elsevier Inc. All rights reserved. doi:10.1016/j.concog.2011.05.010
1328
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343 Table 1 Examples of possible prime-target pairs within a test. In Experiment 1, control primes were neutral words. In Experiment 2, homonym primes and their subordinate/dominant related words were recombined to form unrelated pairs in the control condition.
Experiment 1 Subordinate Dominant Control Experiment 2 Experimental – Subordinate Experimental – Dominant Control – Subordinate Control – Dominant
Prime
Target
PORT BANK POPPY
WINE ACCOUNT SCENE
PORT BANK BANK PORT
WINE ACCOUNT WINE ACCOUNT
Remark: English examples are given for illustration. See the Appendices for the word set and stimulus rotations.
frequent) meaning (e.g., ‘calf’–‘leg’) (e.g., Marcel, 1980; Simpson & Burgess, 1985), reported positive priming for both meanings of a homonym at short SOAs (i.e., SOAs between 0 and 300 ms), indicating that all meanings of a word are more accessible first, whereas later on, the dominant meaning has a clear advantage over other meanings (cf. Duffy, Morris, & Rayner, 1988; Simpson, 1981; Simpson & Burgess, 1985). A similar time course was found with sentence studies and single word studies with normal words (e.g., Tanenhaus et al., 1979; Till, Mross, & Kintsch, 1988). These findings of short-SOA priming were taken as evidence for an early parallel access to meanings (cf. Simpson, 1984; Simpson & Burgess, 1985), occurring within 300 ms after word presentation, followed by selection of one meaning and suppression of nonselected meanings. Whereas researchers seem to agree on the conclusion that parallel access is automatic in nature (given that priming at short SOAs is likely to be automatic, cf. Neely, 1977; Neely, 1991), it is unclear whether the late selection/suppression pattern of priming is reflecting automatic competitive selection (interactive activation model, McClelland & Rumelhart, 1981; TRACE model, McClelland & Rumelhart, 1986; Contextual-Feature-Sensitive Model, Vu et al., 1998), or strategic and expectationrelated controlled selection (e.g., Neely, 1977; Neely, 1991; Simpson, 1984). Modular selection accounts (e.g. Forster, 1976) identify this stage with context integration, during which one is in a state of openness for multiple interpretations of a word, or a readiness for revision of the first choice. In this study, we used subliminal priming to determine the extent to which processes of meaning selection are indeed automatic. Subliminal presentation of primes was used to reduce the contribution of strategic, controlled processes in priming.1 By using a design with several prime-target SOAs (100, 300, 600, 900 and 1500 ms), we sought to determine whether there was a time point at which priming for different meanings of the homonyms diverged. More specifically, the time point of automatic selection of a meaning would be indicated by the inhibition of one meaning (the subordinate meaning) compared to the other meaning of the homonym (the dominant meaning). For comparison, the same test was also used with visible primes. The time course of subliminal and supraliminal homonym priming in a neutral context was measured in three experiments. Polar homonyms like ‘port’ were used, which have one dominant meaning (harbour) which is statistically chosen more frequently than the subordinate meaning (beverage) in neutral contexts. In Experiment 1, prime-target SOA was varied between 100 ms and 900 ms SOAs. Experiment 2 varied the prime-target SOA across a range of 100–1500 ms. 2. Experiment 1 In this experiment, priming of subordinate and dominant meanings of polar homonyms was measured at four prime-target SOAs (100 ms, 300 ms, 600 ms, 900 ms) using a lexical decision task. In the subliminal priming test, primes were sandwich masked using individually titrated presentation timing in order to ensure subliminal presentation. In the visible priming test, primes were shown without masks and were therefore clearly visible. In each test, two experimental conditions and one control condition was measured, whereby each prime and target word was shown not more than once across both tests. In the experimental conditions, homonym primes were paired with a target word related to the subordinate or the dominant meaning of the homonym. In the control condition, neutral primes were paired with unrelated targets, which were related to the subordinate or the dominant meaning of the homonyms used in the experimental condition (cf. Table 1). Priming was computed as the difference in lexical decision times to target words preceded by a related homonym and the same words preceded by an unrelated word.
1 When using subliminal perception to reduce controlled processing, one has to be careful in controlling influences of attentional and strategic processes, which have been shown to still have influence on subliminal priming (e.g. Dehaene, Changeux, Naccache, Sackur, & Claire Sergent, 2006; Eckstein & Perrig, 2007; Van den Bussche, Van Den Noortgate, & Reynvoet, 2009). In this study, attentional requirements were identical in the masked and the visible priming tests and strategic influences were minimized by reducing across-trial repetitions of words.
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1329
2.1. Method 2.1.1. Participants Subjects were 85 students (27 men, 58 women) aged 23.6 years (SD = 4.5) of the Institute of Psychology of the University of Bern, who received credit for an introductory psychology course. All of them were naive as to the hypothesis of the experiment. 2.1.2. Design Two 4 3 mixed-models ANOVA with the factor SOA (100 ms, 300 ms, 600 ms and 900 ms) varied between subjects and the factor ‘Prime-Target Relation’ (homonym prime – subordinate target, homonym prime – dominant target, neutral prime – unrelated target) varied within subjects were implemented separately for the masked and nonmasked tests. Examples of prime-target pairs are given in Table 1. 2.1.3. Material and apparatus For the experimental set, 80 German polar homonym words and their associates were selected from our own standardisation of homonyms and a list of homonyms published in Moritz et al. (2001), in which the naming frequency of different meanings of homonyms was measured, i.e. the number of times a specific meaning comes to mind first when asking a large sample of participants. All homonym words were true homonyms, i.e. their subordinate and dominant meaning represented not merely different senses of the same word but described different meanings. The minimum naming frequency of dominant associates was 52% (M = 77%), the maximum naming frequency of subordinate associates was 39% (M = 20%) and the difference between the two ranged between 23% and 93% (M = 57%). Then, 80 additional neutral words were selected and assigned to the homonyms such that they were unrelated to any of the homonym’s meanings. The 80 homonym-subordinate-dominant-neutral word quadruples (see Appendix 2) were divided into four groups equated in word length, syllable length, naming frequency, log word frequency (using the ‘‘Tagesanzeiger 1996–2000’’ Cosmas corpus of text words containing 75 million words, COSMAS-I Corpora, 2002) and orthographic similarity of prime and target. Subordinate words and dominant words did not differ in average in all the controlled variables, and prime-target pairs had very low average orthographic similarity (M = 65 as computed according to Weber, 1970). The four groups of homonyms and their associates were used to counterbalance word pairs over conditions, tests and participants. Finally, 80 additional words with linguistic properties similar to the homonyms were selected to be used as primes in the nonword target condition, and 30 additional words were selected for the threshold-setting conditions (see Appendices 1 and 2 for the word set and its characteristics). A total of 210 pronounceable nonwords were constructed by recombining consonant/vowel- and vowel/consonant-pairs occurring in each position in the words of the word set. Nonword length and number of syllables was equated to the word set. Of these, 120 nonwords were used as targets in the masked and nonmasked test, and the remaining 90 nonwords were used in the threshold-setting test. A standard IBM-compatible PC with CRT monitor (refresh frequency at 70 Hz) was used for display of the stimuli and an external button box was used for response. 2.1.4. Procedure SOA was determined for each participant before test, and remained the same throughout the whole test session. The experiment started with a prime visibility threshold-setting procedure, which was followed by the masked priming test and then the nonmasked priming test. Each of the three parts of the experiments was preceded by a short practice with a separate set of words and nonwords in which response feedback was given. 2.1.4.1. Trial sequence. All stimuli were written in black on a light grey background. The timeline is illustrated on the left in Fig. 1. Appearance of the fixation point indicated that participants could start the trial, which they did by simultaneously pressing both response keys. Display of premask, prime, postmask and target was synchronized with the screen refresh signal. The prime and postmask durations were individually set in the threshold-setting task. Participants had to give their response within a 1.5 s response window and the target was cleared when a response was given or after 1.5 s, whichever shorter. The intertrial interval lasted 500 ms. 2.1.4.2. Threshold-setting procedure. Participants were informed as to the trial structure and also informed on the sequence of the threshold-setting procedure. Their task was to decide whether the masked letter string was a word or not. Trial sequence and display was the same as in the test trials except that (a) words and nonwords were taken from another stimulus set and (b) all targets were nonwords. Prime duration was set to 57 ms at the start. Blocks of 10 trials each were presented, whereby each time when the mean accuracy over a block was higher than 60%, prime duration was lowered by 14.3 ms and postmask duration was increased by 14.3 ms, so that prime-target SOA remained constant. Accuracy feedback was given to participants at the end of each block. If accuracy was lower or equal to 60% in two succeeding blocks or if 10 blocks had been done, the threshold-setting procedure was concluded. After threshold setting, prime and postmask durations were held constant over the masked priming and nonmasked priming tests. Presentation of prime words and nonwords was randomized within the constraint that 50% of primes within a block should be words. Presentation of target nonwords was random.
1330
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
Fig. 1. Trial timing in Experiment 1 (left) and Experiment 2 (right). Prime and postmask presentation times were individually set in a threshold-setting test. In the visible conditions, pre- and postmasks were replaced by empty screens.
2.1.4.3. Masked priming test. Participants were instructed to concentrate on the target from now on and to decide whether the target was a word or not. A total of 120 trials were shown separated by a short break; 50% of the trials presented nonword targets and 50% presented word targets. Of the word trials, 20 trials presented a homonym-subordinate prime-target pair, 20 trials presented a homonym-dominant prime-target pair and 20 trials presented a neutral-subordinate or a neutraldominant prime-target pair. In order to counterbalance the prime-target pairs across conditions and the masked and visible tests such that none of the prime and target words was shown more than once during the masked and visible priming tests, four lists of homonym-associate pairs were rotated across participants such that (a) each homonym occurred only once in a list, (b) each homonym occurred once with its dominant associate and once with its subordinate associate across the four lists, (c) each neutral prime occurred once in a list (half in the control condition, half in the nonword condition), (d) each homonym’s subordinate and dominant associates were paired once with a neutral word across the four lists and (e) none of the targets occurred more than once in each list (see Appendix 3 for the counterbalancing procedure). The sequence of trial conditions was randomised. 2.1.4.4. Visible priming test. Participants were informed that although they would now see the primes they should carry on doing the lexical decision on targets. Here, 60 nonword target and 60 word target trials were shown with randomisation similar to the masked test. Each homonym prime was shown once across the masked and visible priming tests. None of the word targets shown in the masked test was shown in the visible test. Assignment of word pairs to conditions and tests was counterbalanced across participants as described in the masked priming test paragraph (cf. Appendix 3). 2.1.5. Analysis For each subject, session, test and condition, trials with accurate word responses in trials that were within 2.5 standard deviations from the mean were used for RT analysis. Nonword trials were not analysed. Mean RTs by condition and subject were used as dependent variables in the RT analyses, and arcsine transformed accuracies were used in the accuracy analyses. Greenhouse–Geisser corrections were used when sphericity of data was violated, and a 5% type I error criterion was used unless stated otherwise. 2.2. Results Data of 11 participants (3, 2, 2 and 4 in each of the 100–900 ms SOA conditions) had to be excluded from analysis due to low diligence (less than 80% correct responses), leading to valid data from 23, 22, 20 and 20 participants in the 100 ms, 300 ms, 600 ms and 900 ms SOA conditions, in that order. Distribution of prime presentation durations across participants was comparable in each SOA condition: Prime duration was set to 57.1 ms, 42.9 ms, 28.6 ms and 14.3 ms for 77%, 19%, 2% and 2% of the participants, respectively. Accuracy was high with 94.4% correct responses in the masked priming test and 94.8% correct responses in the visible priming test. 2.2.1. Analyses across the masked and visible priming tests 2.2.1.1. Reaction times and accuracies. The hypothesis that related primes would lead to different classification times compared to unrelated primes was tested with two separate omnibus ANOVAs on RTs and arcsine transformed accuracies, using
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1331
the factors Prime-Target Relation, Test and SOA (see Fig. 2, Tables 2 and 3 for descriptives and Table 4 for the outcome of statistical analyses). In both RTs and accuracies, a highly significant main effect of Prime-Target Relation indicated that prime-target relationship and masking affected RTs. Furthermore, Prime-Target Relation and Test interacted with each other, suggesting that priming was different in the masked compared with the visible priming test. Neither the main effect of SOA, nor any interaction with SOA reached significance. The masked and visible tests were then analysed separately. 2.2.2. Masked primes 2.2.2.1. Reaction times. Descriptives of the results are given in Table 2 (see also Fig. 2). Outlier RTs and errors led to removal of 7.2% of trials from analysis. The hypothesis that related primes would lead to different classification times compared to unrelated primes was tested with an omnibus ANOVA on RTs across Prime-Target Relation and SOA (see Table 4). A highly significant main effect of Prime-Target Relation indicated that the prime-target relationship affected RTs. Separate ANOVAs comparing the experimental conditions with the control condition indicated that dominant priming was significantly greater than zero and that subordinate priming was significantly negative. Neither the main effect of SOA nor the interaction of Prime-Target Relation with SOA reached significance. Descriptives and outcomes of t-tests for comparisons at single SOAs are given in Table 2: only in the 300 ms SOA condition was subordinate priming significantly negative. 2.2.2.2. Power analyses. Given the clear effect of priming across SOAs, post hoc power analyses were done on RT results in order to determine the achieved power in testing for priming of experimental vs. control trials at single SOAs (using the GPower 3 program by Faul, Erdfelder, Lang, & Buchner, 2007). The priming effects of about 10 ms corresponded to a small effect size (d = 0.3), where average power was at 70% with 75 participants, and at 30% with 23 participants. Hence, tests at single SOAs lacked the power to detect 10 ms-priming. In contrast, the 20 ms difference between homonym-subordinate and homonym-dominant RTs represented a medium effect (d = 0.5), which corresponded to an achieved power of over 90% for 75 participants, and 70% for 23 participants. These analyses indicate that the statistical tests used to test for simple priming effects at single SOAs lacked power. 2.2.2.3. Accuracies. Mean errors by condition are given in Table 3. In the omnibus ANOVA on arcsine transformed accuracies across Prime-Target Relation and SOA, a significant main effect of Prime-Target Relation indicated that the prime-target relationship affected accuracies (cf. Table 4). Neither the main effect of SOA nor the interaction of Prime-Target Relation with
Fig. 2. RT priming in Experiments 1 and 2. Grey lines indicate priming for the subordinate meaning, i.e. RTs to target words related to the less frequent meanings of homonyms when preceded by the homonym itself, compared to an unrelated word. Dark lines indicate priming for the dominant meaning, i.e. RTs to target words related to the more frequent meaning of homonyms when preceded by the homonym itself, compared to unrelated word. ⁄⁄p < .01, ⁄ p < .05. Bars indicate standard errors.
1332
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
Table 2 Reaction times (ms) and priming in Experiments 1 and 2. SOA
RT neutral
RT subordinate
RT dominant
Subordinate priming
Dominant priming
M
SE
M
SE
M
SE
M
SE
Experiment 1 – Masked primes 100 ms 610 15 300 ms 621 15 600 ms 627 14 900 ms 635 16 All SOAs 623 8
621 638 638 634 632
17 17 18 16 9
608 613 616 613 612
14 12 12 13 6
11 17* 11 1 10*
8 7 11 12 5
1 8 11 22 10*
8 7 10 12 5
Experiment 1 – Visible primes 100 ms 621 17 300 ms 607 18 600 ms 595 14 900 ms 605 16 All SOAs 608 8
614 588 593 604 600
16 16 11 15 7
589 578 582 584 583
14 18 13 14 7
7 19* 2 1 8*
8 10 10 7 5
33** 29** 13 22 24**
8 11 9 14 5
Experiment 2 – Masked primes 100 ms 634 15 300 ms 554 11 900 ms 542 13 1500 ms 564 12 All SOAs 571 7
657 566 549 572 583
16 12 13 13 8
636 545 535 551 564
16 11 14 11 7
23** 12** 7 8a 12**
6 4 5 5 3
2 9** 7 13** 7**
5 3 5 3 2
Experiment 2 – Visible primes 100 ms 604 13 300 ms 534 12 900 ms 538 11 1500 ms 552 11 All SOAs 555 6
615 528 537 547 554
16 12 13 11 7
599 532 522 538 546
14 12 10 11 6
11 7 1 6 1
7 4 3 4 2
5 2 17** 14** 10**
5 5 4 4 2
M
SE
Remark: One-tailed t-tests were used in the visible conditions. a p < .10. * p < .05. ** p < .01.
Table 3 Errors (in%) and error priming in Experiments 1 and 2. SOA
RT neutral
RT subordinate
RT dominant
Subordinate priming
Dominant priming
M
SE
M
SE
M
SE
M
SE
Experiment 1 – Masked primes 100 ms 6.1 1.0 300 ms 3.9 0.4 600 ms 4.0 1.0 900 ms 5.3 0.8 All SOAs 4.8 0.4
7.6 3.4 5.5 6.0 5.6
0.8 0.5 0.9 0.9 0.4
5.4 3.4 4.3 2.8 4.0
1.1 0.7 1.1 1.0 0.6
1.5 0.5 1.5 0.7 0.8
1.4 0.7 1.6 1.4 0.5
0.7 0.5 0.3 2.5 0.8
1.9 1.0 2.0 1.6 0.8
Experiment 1 – Visible primes 100 ms 10.0 1.0 300 ms 4.8 0.5 600 ms 7.5 0.9 900 ms 7.3 0.9 All SOAs 7.4 0.5
7.0 6.1 6.0 6.3 6.4
0.8 0.9 1.0 0.9 0.4
3.5 4.5 3.8 3.3 3.8
0.8 0.8 1.3 0.7 0.4
3.0 1.4 1.5 1.0 1.1
1.6 0.7 1.6 1.4 0.8
Experiment 2 – Masked primes 100 ms 5.4 0.8 300 ms 2.3 0.4 900 ms 3.3 0.5 1500 ms 2.4 0.4 All SOAs 3.2 0.3
5.1 3.1 4.3 2.7 3.7
0.9 0.7 0.8 0.6 0.4
4.1 2.9 2.8 2.7 3.1
0.6 0.6 0.6 0.6 0.3
0.3 0.8 1.1 0.3 0.5
0.8 0.6 0.7 0.7 0.4
Experiment 2 – Visible primes 100 ms 5.3 1.1 300 ms 3.5 0.5 900 ms 3.1 0.6 1500 ms 2.9 0.5 All SOAs 3.6 0.3
5.5 4.9 3.2 3.1 4.1
1.0 0.8 0.6 0.6 0.4
3.6 1.6 2.8 1.6 2.3
0.8 0.4 0.5 0.4 0.3
0.2 1.4 0.1 0.2 0.5
1.0 0.8 0.5 0.7 0.4
M
SE
Remark: Analyses were done on arcsine transformed accuracies. * p < .05. p < .01.
**
6.5** 0.2 3.8** 4.0** 3.6** 1.3 0.6 0.5 0.3 0.2* 1.7* 1.9** 0.3 1.3** 1.3**
1.6 1.0 2.0 1.6 0.8 0.9 0.6 0.5 0.7 0.3 0.9 0.6 0.5 0.6 0.3
1333
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343 Table 4 ANOVA results of Experiment 1. Analysis/effect
df1
df2
Reaction times
Accuracies
F
MSE
p
F
MSE
p
11.49 6.99 3.54
0.041 0.032 0.034
.010 .035
3.41 1.72 1.78
0.040 0.096 0.061
.036 .19 .19
13.86 30.40 2.27
0.032 0.055 0.062
⁄⁄
Omnibus ANOVA Prime-Target Relation Test Prime-Target Relation Test
2 1 2
162 81 162
18.40 27.81 3.53
973 3074 922
⁄⁄
Masked test Prime-Target Relation Dominant vs. Control Subordinate vs. Control
2 1 1
162 81 81
8.77 5.12 16.59
955 1856 2016
⁄⁄
Visible test Prime-Target Relation Dominant vs. Control Subordinate vs. Control
2 1 1
162 81 81
13.59 20.67 2.52
940 2352 1766
⁄⁄
⁄⁄
.031
.026 ⁄⁄
⁄⁄
.11
⁄⁄
⁄⁄
.14
Remark: Significant p-values are printed in bold. ⁄⁄p < .001, ns: p > .20. The main effect of SOA and interactions with SOA did not reach significance in any of the analyses, F’s < 1.0, p > .20 (except for the interaction of Test SOA, F (3, 81) = 2.20, p = .09).
SOA reached significance. In contrast to RT analyses, priming effects were not significant in follow-up contrasts and at single SOAs. Significant priming effects were thus only observed in reaction times. 2.2.3. Visible primes 2.2.3.1. Reaction times. The overall ANOVA with the factors Prime-Target Relation and SOA yielded a highly significant main effect of Prime-Target Relation. Neither the main effect of SOA nor the interaction of Prime-Target Relation with SOA was significant. Only in the dominant condition were RTs significantly faster than in the control condition, whereas the difference between the subordinate and the control condition only approached significance (corresponding to a marginal p = .051 when using a one-sided t-test). As can be seen in the right panel in Fig. 2, separate t-tests at single SOAs indicated that at 100 ms and 300 ms SOAs, RTs to dominant targets were significantly faster than RTs in the control condition, whereas RTs to subordinate targets were faster than RTs in the control condition at 300 ms SOA (see also Table 2). 2.2.3.2. Accuracies. The overall ANOVA with the factors Prime-Target Relation and SOA yielded a highly significant main effect of Prime-Target Relation. Neither the main effect of SOA nor the interaction of Prime-Target Relation with SOA was significant. Only in the dominant condition were accuracies significantly higher than in the control condition, whereas the difference between the subordinate and the control condition was not significant. Separate t-tests at single SOAs indicated that responses to dominant targets were significantly better than in the neutral condition at SOAs of 100 ms, 600 ms and 900 ms, whereas responses to subordinate targets where not more accurate than in the neutral condition at any SOA (cf. Tables 3 and 4). 2.3. Discussion In visible priming, the following pattern was observed: Dominant priming was significant across SOAs, whereby priming was significant in RTs or accuracies at all SOAs. Subordinate priming was only marginally significant across SOAs, whereby there was an apparent maximum of priming at 300 ms SOA. This pattern is consistent with a model of parallel access of meanings, as suggested by previous research. Masked homonym RT priming, in contrast, indicated facilitation of dominant meanings and inhibition of subordinate meanings across SOAs. There was no apparent change of priming with time (i.e., no interaction with SOA). This pattern of results, and the finding that subordinate priming was significantly negative at 300 ms SOA, supported an automatic selection account: an early selection of the dominant meaning and simultaneous suppression of the subordinate meaning. However, there is an intriguing peculiarity in the data, which is that subordinate meanings of homonyms were actually inhibited and not just neutral. What could be the reason for the inhibition observed in subordinate priming? One known cause of negative priming in the priming literature has been investigated with negative priming paradigms (e.g., Neill, 1997; Tipper, 2001), in which a slowdown of responses to items is observed if these items had to be actively ignored earlier on. This might also be an issue in our study, as negative priming has been reported with masked primes in within-trial priming (e.g., Frings & Eder, 2009). Negative priming is thought to be due either to (a) attentional inhibition, whereby withdrawing attention from a to-be-ignored prime slows the processing of succeeding related stimuli (Tipper, 2001), or to (b) a conflict between episodic retrieval of prime (non) response and response to a target related to it (Neill, 1997). In the present study, both causes are conceivable: Firstly, because the mask-prime-mask sequence appears as a series of flashes, participants might withdraw attention from this sequence altogether, which would also inhibit any attentional processing of the prime word; secondly, the requirement
1334
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
that one is not to respond to an (unconsciously) perceived word might lead to retrieval of a no-response trigger in lexical decision to the target. In both cases, one would expect such negative priming by a masked prime to be stronger the more effective a prime is as a distractor, which is the case if it is related to the target word. Before we expand on this possibility of negative priming and other explanations, we present Experiment 2, which aimed to replicate the subordinate meaning inhibition effect found in this experiment.
3. Experiment 2 Experiment 2 consisted of two experiments, Experiment 2a and 2b, which were mainly designed to validate the results of Experiment 1. However, given that power was low for most between-SOA tests in Experiment 1, a couple of changes were made in the design used in Experiment 1: (a) The word set was reduced to include only homonyms whose dominant and subordinate meaning were more clearly separated in naming frequency, (b) the number of trials per condition was increased by using all targets of the word set in each test, (c) each homonym prime was shown once in the control condition and once in the experimental condition, (d) two control conditions were used in which homonym primes were paired with (subordinate or dominant) targets related to other homonyms of the experimental set, (e) the range of SOAs was extended in order to detect potentially slow decay of priming (SOAs were 100 ms, 300 ms, 900 ms and 1500 ms) and (f) a prime visibility test was added in order to measure whether prime visibility increased over the masked test. Experiment 2a included SOAs of 300, 900 and 1500 ms that were varied within participants, using three separate sessions. In Experiment 2b, priming at the 100 ms SOA was measured using the same procedure for the masked and visible priming tests, using an extended threshold-setting procedure and a longer discrimination test, and an additional forced-choice discrimination test at the end of the experiment. The reason for this was that it is quite uncommon to find significant priming effects with truly invisible primes, especially when using large stimulus sets as in the reported experiments. One likely explanation for the effects found in the previous experiments could be partial prime perception: when a participant sees a target that is related to the preceding prime, this might help him/her to recognise the otherwise illegible prime word, which directly leads to faster lexical decision. Because we used a relatively benign threshold-setting procedure and the resulting prime presentation times were relatively long, compared to other studies of this type, partial prime visibility might be the main driver for the priming effects. We addressed this hypothesis in Experiment 2b (a) by using a more restrictive threshold-setting procedure and testing whether this led to shorter prime presentations, (b) by doubling the trials in the prime visibility test in order to obtain a more sensitive measure of prime recognition and (c) by adding a further prime visibility test assessing partial prime visibility. Priming analyses are reported on the data collapsed across Experiments 2a and 2b; correlation analyses of priming and visibility measures are reported separately for Experiments 2a and 2b. Hypotheses were not changed, although we set out to validate results of Experiment 1, which were not in line with any of our a priori hypotheses, in particular (a) slower classification in homonym-subordinate trials than in unrelated trials in the masked priming test, and (b) no interaction of priming with SOA. The visible condition was expected to show the pattern of RTs found in Experiment 1. 3.1. Method 3.1.1. Participants Participants in Experiment 2a were 30 students (24 women, 6 men) aged 25 years on average (SD = 8.8). Participants of Experiment 2b were 25 students (22 women, 3 men) aged 22.7 years on average (SD = 3.7). All participants were students of the Institute of Psychology of the University of Bern who received credit for an introductory psychology course. All of them were naive as to the hypothesis of the experiment. 3.1.2. Design In Experiment 2a, the factors SOA (300 ms, 900 ms and 1500 ms) and Prime-Target Relation (homonym prime – subordinate target, homonym prime – dominant target, homonym prime – unrelated target collapsed across target types) were varied within subjects for both the masked and nonmasked tests. Experiment 2b implemented the same design except that SOA was fixed at 100 ms. Analyses on data collapsed across Experiments 2a and 2b thus used a mixed-model design with SOA varied between participants (four levels: 100 ms, 300 ms, 900 ms and 1500 ms) and Prime-Target Relation varied within participants. 3.1.3. Material and apparatus For the experimental set, 60 homonyms and their associates were selected from the word set of Experiment 1 such that the probability of naming the dominant associate when presented with the homonym was higher than 70%, and the probability of naming the subordinate associate when presented with the homonym was lower than 30%, according to the word norms mentioned in Experiment 1 (cf. Appendix 1). Thirty pairs of homonyms were formed such that their associates were not related to the other homonym of a pair. The 30 pairs were divided into two lists of 15 pairs equated in frequency, naming probability and orthographic similarity. These lists of homonym pairs and their associates were used to counterbalance word
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1335
pairs over tests, sessions and participants (see procedure and Appendix 3). Sixty words with similar linguistic properties as the homonyms were selected as primes in the nonword target condition. Fifty additional neutral words were selected for the threshold-setting conditions. PCs with same configuration as in Experiment 1 were used with a standard keyboard as input device. Two-hundred and seventy nonwords used in the threshold-setting and priming test were constructed in the same way as described in Experiment 1. 3.1.4. Procedure of Experiment 2a Each participant did three sessions that were 1 week apart from each other. In each session, one SOA condition was tested. The order of SOAs over sessions was counterbalanced across participants. Each session started with a prime visibility threshold-setting procedure, which was followed by the masked priming test. After this, prime visibility was measured, followed by the visible priming test. The threshold-setting, masked priming and visible priming test were each preceded by a short practice with a separate set of words and nonwords. Trial timing was the identical to Experiment 1, with the exception that the postmask was shown for 100 ms and then replaced by an empty screen (cf. Fig. 1 on the right). The threshold procedure was identical to Experiment 1, with the exception that prime duration was set to 71 ms at the start of the procedure. Trial sequence and display was the same as in the test trials, except that different words and nonwords were used for primes and targets and that all targets were nonwords. After threshold setting, the prime and mask presentation times were held constant over the masked priming, visibility and visible priming tests within a session. The masked priming test consisted of 240 trials separated by three short breaks. Instruction, trial timing and masks were otherwise identical to Experiment 1. Across every 120 trials, 60 nonword target trials, 15 homonym-subordinate, 15 homonym-dominant, 15 unrelated-subordinate and 15 unrelated-dominant trials were shown in a randomised sequence. In order to avoid repetition effects, homonyms were never shown with the same target within and across the masked and visible tests of one session. Each homonym was shown once in 120 trials, and over 240 trials, each homonym was shown once in an experimental condition and once in a control condition, and each target was shown once in each test (see Appendix 3 for the counterbalancing procedure). The visibility test consisted of a random sequence of 15 nonword prime and 15 word prime trials with identical trial timing and material as in the threshold-setting test. Participants were asked to decide whether the prime was a word or not. In the visible priming test, instruction and trial timing were identical to Experiment 1. A total of 120 nonword target and 120 word target trials were shown, whereby prime and target pairing was controlled as described in the masked test paragraph. The order of related and unrelated trials for a given homonym and dominance was counterbalanced across participants and sessions. 3.1.5. Procedure of Experiment 2b Participants did one session only with the SOA set at 100 ms, using the same procedure as in Experiment 2a, with the following adaptations: (a) an attempt to increase the sensitivity of the threshold procedure was made by increasing the threshold block length to 20 instead of 10 trials, and by setting the threshold exit criterion to smaller or equal to 55%, (b) the prime visibility test using lexical decision was extended to 30 nonword prime and 30 word prime trials and (c) at the end of the test, an additional forced choice test assessing partial awareness of primes was conducted. 3.1.5.1. Partial prime visibility test. In this final test, participants were again asked to try to recognise the prime in each trial, whereby the usual prime-target word trial sequence was extended by a two-stage decision task, i.e. a visibility rating and a forced-choice word selection. After each prime-target sequence, participants were asked to do a visibility rating, i.e. to choose the best fitting response of three alternatives describing what they had seen: (a) they had not seen the prime, (b) they had seen a part of the prime, as for instance single letters or letter combinations or (c) they had seen the prime word. In the second, forced choice selection task, they were asked to indicate which of two choice words matched the previously presented prime, whereby they were encouraged to guess the word if they were unsure. Stimulus presentation and input was controlled by the PC: After the mask-prime-mask sequence (cf. Fig. 1), the target was shown as usual, but was not erased after 1.5 s. Instead, a selection scale appeared at the bottom screen, indicating the three response alternatives of the visibility rating, corresponding to the outer left, middle and outer right keys of the button box. They were given 9.5 s for this decision. After their response, a second screen was presented for the forced choice response, with two choice words displayed on the left and on the right of the screen centre. The two choice words were the two homonyms that had been combined with the presented target word in the priming tests (i.e., the two homonyms of a homonym pair, one related and one unrelated with the target, see the Material section); the prime was thus one of the choice words. Participants could then give in their second response, using the outer left and right keys. Response window was set to 10 s. A total of 64 target word trials from the 120 word–word trials from the priming test were shown in a random sequence, whereby the following variables were balanced: Test half within which the prime-target pair occurred, prime-target relatedness, target dominance and display side of the prime word in the selection pair. 3.1.6. Analysis The main analysis was identical to Experiment 1. A priori power analyses showed that with a sample size of 35 participants, single-sided tests of priming effects of 10 ms (d = 0.3) at single SOAs would yield a power of 70% if alpha was increased to 0.10. P-values below .10 are therefore mentioned separately in the analyses.
1336
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
Responses in the visibility rating of the partial prime visibility test in Experiment 2b were used to classify responses in the forced choice test. Responses in the forced choice test in Experiment 2b were registered as correct or incorrect (i.e., correct if the selected word matched the prime). In order to test whether priming was driven by greater partial prime visibility for related than for unrelated prime-target pairs, a discrimination measure for related vs. unrelated prime-target pairs was computed. The hit rate was defined as the proportion of correct prime word choices in related prime-target trials and the false alarm rate was defined as the proportion of incorrect word choices in unrelated prime-target trials. 3.2. Results 3.2.1. Visibility and accuracy in Experiment 2a Four participants who responded to the prime instead to the target in the visible priming test and eight participants with average high prime discrimination (d0 P .80) in the final test for prime visibility at single SOAs were replaced. Average d0 was not significantly different from zero in any condition, M = 0.05 (SE = 0.08), M = 0.10 (SE = 0.09) and M = 0.09 (SE = 0.06) for the 300, 900 and 1500 ms SOAs, respectively. Prime duration was set to 71.4 ms, 57.1 ms, 42.9 ms and 28.6 ms in 62%, 26%, 10% and 2% of the sessions, respectively (distribution of prime durations was similar at all SOAs). Accuracy was high with 96.5% correct responses in the masked priming test and 96.3% correct responses in the visible priming test. 3.2.2. Visibility and accuracy in Experiment 2b No participant had to be replaced. Prime duration was set to 71.4 ms, 57.1 ms, 42.9 ms and 28.6 ms in 64%, 16%, 8% and 12% of the sessions, respectively. Because the threshold-setting procedure was stricter in this experiment, the distribution of prime durations across subjects was compared with that of the previous experiments. The distribution of prime durations was not significantly different from the one measured in Experiment 2a or in Experiment 1, v2’s < 1.0. Accuracy was high with 94.4% correct responses in the masked priming test and 94.5% correct responses in the visible priming test. Average d0 in the lexical decision prime visibility test was not significantly different from zero, M = 0.09 (SE = 0.07), p = .20. In contrast, average d0 in the partial prime visibility test was significant both when the prime was followed by a subordinate target, M = 0.38 (SE = 0.18), p = .045 and when the prime was followed by a dominant target, M = 0.53 (SE = 0.19), p = .009. 3.2.3. Analyses across the masked and visible priming tests The following priming analyses were done on the data collapsed across Experiments 2a and 2b. 3.2.3.1. Reaction times. The hypothesis that related primes led to different classification times compared to unrelated primes was tested with an omnibus ANOVA on RTs across Prime-Target Relation, Test and SOA (see Fig. 2 and Table 2 for descriptives and Table 5 for inference tests). The main effects of Prime-Target Relation and Test were highly significant, which indicated that prime-target relationship and masking affected RTs. Furthermore, Prime-Target Relation and Test interacted, suggesting that priming was different in the masked compared with the visible priming test. There was also a significant main effect of SOA, which was qualified by Prime-Target Relation and, marginally, by Test. No other interaction reached significance.2 3.2.3.2. Accuracies. In the omnibus ANOVA on arcsine transformed accuracies across Prime-Target Relation, Test and SOA, the main effect of Prime-Target Relation was highly significant, as was the effect of SOA (cf. Tables 3 and 5). There was also a marginal interaction of Prime-Target Relation and Test, suggesting that there was a tendency for priming to be different in the masked compared with the visible priming test. No other main effect or interaction reached significance. 3.2.4. Masked primes 3.2.4.1. Reaction times. The main hypotheses were tested with ANOVAs on average RTs. In the omnibus ANOVA with the within factor Prime-Target Relation (neutral, subordinate, dominant) and the factor SOA (100, 300, 900 and 1500 ms) varied between subjects, the main effects of Prime-Target Relation and SOA were significant. The interaction of SOA and Prime-Target Relation however did not reach significance, indicating that there was little variation in priming across SOAs. Follow-up contrasts between experimental and the neutral condition indicated that dominant priming was significantly positive and subordinate priming was significantly negative (see Fig. 2 and Table 5). Priming for the subordinate meaning was significant at SOAs of 100 ms and 300 ms and priming for the dominant meaning was significant at SOAs of 300 ms and 1500 ms (cf. Table 2). In order to determine whether prime visibility influenced priming (across Experiments 2a and 2b), we (a) analyzed the correlation of subordinate and dominant priming with d0 from the prime visibility task and (b) determined the intercept and slopes of priming vs. d0 using linear regression. The correlation of priming with d0 was neither significant for subordinate priming, r = .15, p = .11, nor for dominant priming, r = .18, p = .058. The scatterplots are given in Fig. 3. Because correlation of dominant priming approached significance, we also estimated the intercept and slope of the d0 -priming relation using linear regression: The intercept of subordinate priming was negative, b1 = 12.3 ms, SE = 2.6 ms, t(113) = 4.80, p < .001 and 2 Because SOAs were varied within participants in Experiment 2a, a first analysis was done using a repeated measures ANOVA across SOA conditions 300 ms, 900 ms and 1500 ms only. Results were comparable to the one reported here, except that the main effect of SOA was not significant in any of the RT and accuracy analyses.
1337
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343 Table 5 ANOVA results of Experiment 2. Analysis/effect
df1
df2
Reaction times
Accuracies
F
MSE
p
F
MSE
p
13.69
0.018
⁄⁄
4.60 2.87
0.019 0.020
1.77 4.10
0.022 0.034
4.42
0.012
14.74 24.50
0.017 0.028
2.71 1.29
0.014 0.017
Omnibus ANOVA Prime-Target Relation Test SOA Prime-Target Relation Test Prime-Target Relation SOA Test SOA 3-way interaction
1.84 1 3 1.81 6 3 6
204 111 111 200 222 111 222
37.08 37.01 10.91 7.03 3.13 2.38 1.28
328 2154 7804 439 301 2154 396
⁄⁄
Masked test Prime-Target Relation Dominant vs. Control Subordinate vs. Control SOA Prime-Target Relation SOA
1.67 1 1 3 6
187 111 111 111 222
29.30 11.05 24.14 11.74 1.28
440 460 740 4600 370
⁄⁄
Visible test Prime-Target Relation Dominant vs. Control Subordinate vs. Control SOA Prime-Target Relation SOA Dominant-Control SOA Subordinate-Control SOA
2 1 1 3 6 1 1
222 111 111 111 222 111 111
9.52 16.52 0.10 8.37 2.98 2.33 3.09
327 305 282 3922 327 305 282
⁄⁄
⁄⁄ ⁄⁄
.002 .006 .074 ns
.001 ⁄⁄ ⁄⁄
ns
⁄⁄
ns ⁄⁄
.008 .08 .030
ns .005 .059 ns ns ns .18 .045 ns .006 ns ⁄⁄ ⁄⁄
ns .049 ns
Remark: Significant p-values are printed in bold. SOA was treated as a between-factor. ⁄⁄p < .001, ns: p > .20. F-values and MSE’s are not listed for F’s smaller than 1.0.
Fig. 3. Scatterplots of subordinate (left) and dominant (right) masked priming in function of lexical decision prime visibility for the data collapsed across Experiments 2a and 2b. Grey lines are regression lines (correlations were not significant).
the slope was not significantly different from zero, b2 = 10.38 ms, SE = 6.4 ms, t(113) = 1.62, p = .11. The intercept of dominant priming was positive, b1 = 6.8 ms, SE = 2.0 ms, t(113) = 3.40, p = .001 and the slope was not significantly different from zero, b2 = 9.7 ms, SE = 5.1 ms, t(113) = 1.92, p = .058.3 This indicates that subordinate priming was negative and dominant priming was positive even when conscious prime perception was completely suppressed. 3.2.4.2. Accuracies. The main effect of Prime-Target Relation in the omnibus ANOVA was not significant but SOA had a significant effect on accuracies. The interaction of SOA and Prime-Target Relation did not reach significance, however. Follow-up contrasts between experimental and neutral conditions indicated that dominant priming was significantly positive and subordinate priming was not significant. Priming was not significant when tested at single SOAs (see Table 3). 3.2.5. Visible primes 3.2.5.1. Reaction times. In the omnibus ANOVA, the main effects of Prime-Target Relation and SOA were significant. More important, the interaction of SOA and Prime-Target Relation was also significant, indicating that priming varied with SOA. Two follow-up ANOVAs on the single experimental and the neutral condition indicated that the variation in dominant 3 It has been suggested that an adapted regression method be used in order to correct for the measurement error of the predictor variable d0 (Klauer, Draine, & Greenwald, 1998). This method leads to slightly greater intercepts when the slope is negative, which is the case with the present data.
1338
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
priming across the whole range of SOAs was small, given that the interaction of SOA and Prime-Target Relation did not reach significance. In contrast, subordinate priming was not significant across the whole range of SOAs but changed with SOA, as indicated by a significant interaction of Prime-Target Relation and SOA (see Fig. 2 and Table 5). When tested at single SOAs, priming for the subordinate meaning was not significant, whereas priming for the dominant meaning was significant at the 900 ms and1500 ms SOAs (see Table 2). 3.2.5.2. Accuracies. The main effects of Prime-Target Relation and SOA were significant. However, the interaction of SOA and Prime-Target Relation was not significant, indicating that variation of priming across SOAs was small. Follow-up contrasts on the single experimental and neutral conditions indicated that dominant priming was significantly positive across the whole range of SOAs, whereas subordinate priming was not significant. Priming for the dominant meaning was significant at the 100 ms, 300 ms and1500 ms SOAs (see Table 3). 3.2.6. Correlation of prime visibility and priming in Experiment 2b In order to determine whether prime visibility influenced priming in Experiment 2b, we analysed the correlation of subordinate and dominant priming with d0 from the two prime visibility tasks (see Fig. 4 for scatterplots). Correlation of RT priming with lexical decision d0 was not significant either for subordinate priming, r = .02, p > .20 or for dominant priming, r = .05, p > .20. Similarly, subordinate priming did not correlate with partial visibility d0 for homonyms when subordinate targets were shown, r = .13, p > .20 and dominant priming did not correlate with partial visibility d0 when dominant targets were shown, r = .26, p > .20. This suggests that in this experiment, prime visibility did not have a measurable influence on masked priming. The correlation of accuracy priming with lexical decision d0 was not significant for subordinate priming, r = .08, p > .20, but significantly negative for dominant priming, r = .49, p = .013. Correlations of accuracy priming with partial visibility d0 were not significant, r’s < .18, p’s > .20. 3.2.7. Partial prime visibility test in Experiment 2b The additional partial prime visibility test in Experiment 2b provided more detailed data on whether and how often participants were able to recognise parts of the primes, and whether the target following the prime facilitated prime perception. In the prime visibility task, all participants declared they had not perceived anything of the prime in some of the 64 trials (average number of trials: M = 40.5, SE = 3.8), while 16 participants reported partial perception of the prime in some trials (M = 22.5, SE = 2.9) and eight participants reported seeing the prime word in some trials (M = 24.6, SE = 4.9). Participants’ subjective reports were in good agreement with the objective measure of discrimination: The accuracy (ac) in selecting the prime word in the partial visibility task was indeed at chance in the ‘‘not seen’’ category, acunrelated = 0.52, acrelated = 0.52,
Fig. 4. Scatterplots of subordinate (left) and dominant (right) masked priming in function of prime visibility measures for Experiment 2b. Lexical decision d0 was assessed with a task asking for lexical decisions with respect to the prime; partial prime visibility d0 was assessed with a task wherein participants had to decide which of two choice words was the prime. Grey lines are regression lines (correlations were not significant).
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1339
d0 = 0.13, SE = 0.12, t(24) = 1.01, p > .20, but better than chance in the ‘‘partially seen’’ category, acunrelated = 0.65, acrelated = 0.71, d0 = 1.5, SE = 0.38, t(15) = 3.91, p = .001 and in the ‘‘seen a word’’ category, acunrelated = 0.78, acrelated = 0.84, d0 = 2.44, SE = 0.66, t(7) = 3.70, p = .008.4 Hence, when participants indicated perceiving part of the prime or the whole prime, they indeed more often selected the correct word in the forced choice task than what would be expected by chance. Prime discrimination was not better for dominant than for subordinate meanings across response categories, Dd0 = 0.15, SE = 0.15, t(24) = 1.0, p > .20 and within response categories, t’s < 1.4, p > .20 for every category. Furthermore, discrimination in the partial prime visibility task was not significantly correlated with discrimination in the lexical decision task, r = .16, p > .20. In summary, when participants directed their full attention to the primes, they were better than chance at identifying which of two words was the prime, whereby there were large differences between participants. Eight participants were able to read whole words in about a third of the trials; across all participants, primes were not perceived at all in two thirds of the trials. It is thus conceivable that partial prime perception occurred in the priming test; however, it appears unlikely that partial prime visibility was the main driver of priming, given that correlations of priming with d0 were nonsignificant (or negative). 3.3. Discussion In Experiment 2, the basic masked priming pattern seen in Experiment 1 was replicated. When lexical decisions in trials with related homonym prime-target word pairs were compared to a control condition, responses were slower for subordinate meanings and faster for dominant meanings. Differences between SOAs did not reach significance, which indicates that this pattern of priming remained fairly stable across SOAs. As in Experiment 1, the pattern suggests some fast selection of the (dominant) word meaning, whereby the subordinate meaning is simultaneously inhibited. In the visible priming test, dominant priming was positive across all SOAs, whereby the priming effect at single SOAs was observed in RTs or errors. In contrast, there was no significant subordinate priming across SOAs, but subordinate priming varied with SOA. Visible subordinate priming was thus weaker in Experiment 2 than in Experiment 1 and did not replicate the qualitative time course found in Experiment 1. One possible explanation for this difference comes from the partial within-variation of SOA: participants in Experiment 2a did three sessions using the identical word set (but different prime-target SOAs), wherein primes and targets were combined using two variants. In the visible test of the first session, participants usually noticed that the words were ambiguous; it is possible that this led to a bias in the second and third sessions which favoured the dominant compared to the subordinate meanings of homonyms. In Experiment 2b, which assessed priming at 100 ms SOA, masked subordinate priming was again negative, but masked dominant priming was only positive in accuracies. In the visible test, the finding that subordinate priming was not significant indicates that RT effects were smaller than in Experiment 1; the finding that dominant priming was significant in errors only supports this contention. The more sensitive threshold-setting procedure of Experiment 2b, compared with Experiments 1 and 2a, did not produce shorter prime presentation times. Prime visibility was very low (mean d0 = 0.09) and nonsignificant when assessed with lexical decision, but higher (mean d0 = 0.5) and significant when assessed with a partial prime visibility test in which participants had to identify the prime out of a list of two words in each trial. Given that it is difficult to focus essentially on the prime in the priming test, it is likely that prime visibility during the priming test was even lower. The results thus indicate that participants probably perceived some letters of the primes in a few trials during test. However, we think that it is unlikely that such perception led to the observed priming effects, because masked priming did not correlate with prime visibility measures. 4. General discussion The characteristic time course of supraliminal homonym priming indicates early parallel access of all meanings of a word, and later selection of one of the meanings. This study endeavoured to measure subliminal priming by polar homonyms in order to gain more insight into the automatic processes of access and selection of word meaning. If selection of meaning were automatic, a difference in priming for different meanings of homonyms should be seen at some point when measuring priming in a range of prime-target SOAs, even when the primes are shown subliminally. Results in the visible prime conditions indicated that meaning selection occurred after 300 ms, which confirmed previous research with visible primes (e.g., Simpson, 1981; Simpson & Burgess, 1985). Conversely, in the subliminal condition, the dominant meaning was facilitated and the subordinate meaning was inhibited not only with respect to the dominant meaning, but also with respect to a neutral condition at prime-target SOAs ranging from 100 ms to 1500 ms. In both experiments, subliminal and dominant priming did not increase or decrease significantly across SOAs. The reliable difference between priming for the dominant meaning and priming for the subordinate meaning indicates that word meanings of masked homonyms were not only accessed, but one of the meanings was selected early on. The results therefore support an automatic early selection account. However, the finding that access to
4 Participants did not report seeing the prime more often when the prime and target were related, compared with when the prime and target were unrelated, as the average difference in the number of trials between the two conditions was not significant, Ndiff = 0.6, SE = 0.4, t(24) = 1.28, p > .20, Ndiff = 0.1, SE = 0.6, t(15) = 0.20, p > .20 and Ndiff = 1.4, SE = 0.7, t(7) = 1.95, p = .09 for the ‘‘not seen’’, ‘‘partially seen’’ and ‘‘seen a word’’ response categories in this order.
1340
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
the subordinate meaning of homonyms was actually inhibited and not just neutral deserves special consideration. In the following we therefore discuss the possible sources of such an inhibition. 4.1. Negative priming and inhibition In both masked priming experiments, priming by subordinate meanings was negative across SOAs, which indicates selective inhibition of the subordinate meaning. In principle, such type of inhibition is present either in automatic selection or in strategic checking. Automatic selective inhibition is found in most models that provide mechanisms of selection, such as inhibition between lexical nodes in spreading activation models (McClelland & Rumelhart, 1981; McClelland & Rumelhart, 1986), inhibition by lexical competition (Norris, 1994) or inhibition between competing attractors for lexical entries (Rodd, Gaskell, & Marslen-Wilson, 2002). There is also one theory of priming suggesting that inhibition is needed to enhance discrimination of weak traces (Dagenbach & Carr, 1994). Although the various models locate inhibition at different levels of representation, the main principle is similar for all of them: Inhibition is a means of counterweighting activation or influence by other, competing concepts during a selection phase. Mutual inhibition of competing concepts will lead to amplification of small differences between the concepts, which ultimately allows for selection of one of the candidates. Therefore, one would expect to find inhibition between competing meanings of a homonym at some point in the selection process. Therefore, inhibition might be a sign of selection or interrupted selection. The positive masked priming of dominant primes found in both experiments might indeed support this thesis. There are, however, other sources of automatic selective inhibition which are due to intertrial inhibition. Simpson and Kang (1994) for instance found evidence for long-term inhibition of unselected meanings at lags of 0, 1, 4 or 12 trials after presentation of a homonym. Because the inhibition was obviously linked to earlier conscious access to one of the meanings, the authors suggested that ‘‘an inhibitory effect is diagnostic of previous activation’’ of a homonym (p. 377). Following this line of reasoning, one could speculate that even at lags of hours, days or weeks, the memory trace associated with a homonym might be preferentially linked with the context associated with its last encounter, and therefore inhibition might be inherent in any word retrieval from memory. Another explanation for inhibition due to inter- and intratrial repetition might be deduced from the negative priming literature. As we mentioned in the discussion of Experiment 1, if the prime is ignored but relevant for a target decision, it can function as a distractor, provoking attentional inhibition or response interferences. Frings and Eder (2009) for instance observed that negative priming by a masked distractor that was repeated as a visible distractor in the probe display produced negative priming of about 10 ms in a probe naming task. This priming effect decayed over time: It was significant at 138 ms and 538 ms SOA, but not at 1038 ms SOA. Although primes were not repeated as targets in our study, the mere fact that primes resembled the targets semantically might be a cause of negative priming, because response to primes must be inhibited. Interestingly, the inhibition observed with subordinate meanings in our study had a (nonsignificant) tendency to decrease across SOAs, similar to the decrease reported in the Frings and Eder study. A data pattern similar to the one reported here was also reported by Wentura and Frings (2005) in four experiments on far and near category concepts, which Wentura and colleagues termed ‘‘dominant’’ and ‘‘nondominant’’ words. When they presented subliminal category words followed by words denominating elements of a category, they observed negative nondominant priming and (less reliably) positive dominant priming. Strategic sources of inhibition might have been an influence in Experiment 2a, because our participants probably realized that primes were homonyms in the visible test, which could have influenced their behaviour in sessions 2 and 3. According to Kahan (2000), participants will use their knowledge on the possible prime-target relationship to guess the prime’s content. Therefore, if participants realize that primes are sometimes related to targets, they might start to expect such a relationship between primes and targets. This expectation will lead to positive priming if perception of the prime is above a certain threshold. However, it might lead to negative priming if perception of the prime is below that threshold, because the weak match between prime and target creates interference in the decision process. In Experiment 2a, stimuli were repeated across sessions. That is, in sessions 2 and 3, participants probably guessed that related prime-target pairs were shown in the masked conditions, too. Therefore, any relationship between prime and target could have led to negative priming if prime visibility was sub-threshold, which was most likely the case in Experiment 2 (our detection rate was much lower than the ‘poor performer’ detection rate in Kahan’s study). However, according to this logic, RTs to dominant and subordinate targets would be expected to be inhibited due to a related prime, which is not exactly what we observed. Nevertheless, this is an interesting explanation which should be followed up. 4.2. Generalizability of results The above suggested interpretation of results rests on the assumption that word identification times were not systematically biased by task demands. In order to exclude systematic biases, care was taken to show equal proportions of trials in all relevant conditions of the reported experiments. Thus, the number of related and unrelated prime-target trials, the number of dominant and subordinate prime-target trials and the number of dominant and subordinate neutral trials were held equal in all parts of the experiments. The rationale for this approach was that unequal proportions of trials would contribute to the formation of biases within an experiment, such that lexical decision RTs for dominant vs. subordinate target words, for instance, would be shorter if dominant words were more frequent. There might be other sources for biases, e.g. biases could result from different salience of groups of word pairs, as for instance homonym-dominant target pairs, that are by definition
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1341
more readily associated than homonym-subordinate target pairs. Furthermore, it is conceivable that partial prime perception had a similar effect in the masked priming tests. Although we have no evidence that our pattern of results could be produced or strengthened by pre-existing or experiment-induced bias for dominant meanings, and by pre-existing or experiment-induced bias for related word pairs, future experiments might aim to directly influence such biases, in order to test the robustness of the observed pattern of results. 4.3. Were primes subliminal or not? There were clear qualitative differences between masked and visible priming. In the masked priming test, we aimed to obtain a measure of prime perception when the prime is not perceived, in order to exclude controlled processes that occur when participants have insight into the aim of the experiment and the type of prime-target relationships occurring during the tests. In order to reduce prime awareness to a minimum, a threshold-setting procedure was used which adapted presentation times to individual perception thresholds for prime words in each session. Were we successful, then, in reducing prime awareness in the masked test? As a general rule, prime visibility tests should use conditions that are as similar as possible to the ones used in the actual priming test. In Experiment 2, participants did a lexical decision prime visibility test in which they decided whether the prime was a word or not. Prime discrimination was not higher than zero and did not correlate with priming. However, the prime words used in this test were not the same as the ones used in the priming test and targets were nonwords (the latter change was made in order to facilitate the prime detection task). Therefore, one could argue that the lexical decision prime visibility test used here did not measure the relevant features of prime visibility that influence priming. As an additional measure of prime visibility, a partial prime visibility test was included in Experiment 2b, which assessed after each priming trial whether participants could identify the prime word in a list of two words, whereby primes and choice words were identical to primes and targets in the masked priming test. This test indicated (a) that partial perception of the primes was occasionally possible and (b) that between-subject differences in prime visibility were large. Some of the subjects, for instance, reported being able to read the prime words in some trials. It is therefore important to note that partial prime perception did not correlate with masked priming, which leads us to two important conclusions: Firstly, partial prime perception was not the main driver of the masked priming effects found here and, secondly, partial prime perception did not produce a pattern resembling the visible priming pattern (i.e., more positive priming for both meanings of homonyms). Given that subjective reports supported the validity and reliability of the partial priming test, it therefore appears that partial prime visibility was not informative enough to produce priming at the level of word meaning.5 From these analyses, we conclude (a) that prime awareness was not totally suppressed in the masked priming tests and (b) that the masked priming results presented are most probably not a consequence of spurious conscious prime perception. It is, however, important to bear in mind that our ‘subliminal’ primes were probably ‘periliminal’. 4.4. Awareness, lexical access and priming The differences between priming by subliminal words and priming by visible words give us insight into processing differences associated with perception with and without awareness. Priming by visible primes indicated that both meanings of homonyms were accessed (at least in Experiment 1), and the dominant meaning was selected between 300 ms and 600 ms after prime onset. Priming by invisible primes indicated that the dominant meaning was more easily accessed and the subordinate meaning less easily accessed than a control, even at short SOAs. The finding that positive priming for multiple meanings is only observed with visible primes, and only at short SOAs, indicates that the stage of meaning selection assumed to occur between 0 and 300 ms depends on awareness of both prime and target words. Indeed, it has been shown that it is difficult to mentally separate two stimuli under such conditions (e.g., Shapiro, Driver, Ward, & Sorensen, 1997; Treisman & Gelade, 1980), indicating that this pre-selection stage also reflects binding of prime and target properties. Accordingly, visible priming at short SOAs might be more sensitive to mental combination of prime and target, rather than being a measure of passive influence by the perception of a prime. In consequence, it is conceivable that the supraliminal homonym priming pattern reflects a post-lexical stage of context integration which is characterised by malleability of prime meaning due to awareness of both the prime and the target. This would imply that subliminal perception is selective and rigid and supraliminal perception is integrative and flexible (for a similar conclusion with respect to different levels of awareness, see Davis et al., 2007). The observed dissociation between subliminal and supraliminal priming might also have even more profound implications for visible priming effects in general. This is because the flexible choice of meaning when both prime and target are visible could have been partly or completely driven by strategic influences: When primes are visible, participants may notice that targets sometimes reflect the dominant and sometimes the subordinate meaning of the prime words, which could induce an inclination to consider the different meanings of the words throughout the test.6 If this were the case, visible priming 5 Strong influences of partial priming on masked ‘‘semantic’’ priming have been reported by Kouider and Dupoux (2004) with colour words. Compared with the Kouider and Dupoux study, the present study used a set of stimuli that was much larger and less easy to predict. Therefore, partial prime visibility was less useful here for second-guessing the prime. 6 We thank an anonymous reviewer for mentioning this possibility.
1342
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
effects assessed with a transparent design such as the one used here would overestimate the ‘real’ priming effect. Such an influence is less likely if primes are presented subliminally (e.g., Van den Bussche & Reynvoet, 2008; Van den Bussche, Segers, & Reynvoet, 2008; see also Bodner & Masson, 2003). In conclusion, we suggest that when the prime is shown subliminally, the initial choice of meaning cannot be revised and the resulting priming is based on within-lexicon interactions of prime and target representations. Because the dominant meaning of a homonym is its most frequently used meaning, the combination of a homonym prime with its subordinate meaning as a target leads to interference within the lexicon. 4.5. Conclusion Compared to other words, homonyms have a more variable meaning, which renders them more dependent on context information. Priming experiments with visible primes indicate that in the first moments during which we become aware of a homonym, we are open to different interpretations of it. The rise and decline of priming, which indicates access to and selection of meanings, is slow. In contrast, our results indicate that when we are not aware of seeing a (subliminal) homonym, we behave selectively. Access to meanings and selection of a meaning is fast and favours the most frequent meaning. Further research is needed to determine whether the observed dissociation between visible and masked priming is due to strategies or, rather, to qualitative differences in perception with and without consciousness. Acknowledgments This manuscript was partly funded by Swiss National Foundation Fellowship No. PA001 – 113106/1 granted to the first author and Swiss National Foundation Grant No. 1114-067180 granted to the third author. We thank Bettina Gerber, Melanie Hegi, Matthias Minder, Ajshe Ponxha and Helen Wyler for conducting parts of Experiment 1, and the University of Bern and the MRC Cognition and Brain Sciences Unit for granting use of infrastructure. Many thanks also to Matt Davis and Rik Henson, who supported the writing process with valuable comments, and to four anonymous reviewers for their helpful comments. Appendix A. Supplementary material Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.concog.2011.05.010. References Binder, K., & Morris, R. K. (1995). Eye movements and lexical ambiguity resolution: Effects of prior encounter and discourse topic. Journal of Experimental Psychology: Learning, Memory and Cognition, 21(5), 1186–1196. Bodner, G. E., & Masson, M. E. J. (2003). Beyond spreading activation: An influence of relatedness proportion on masked semantic priming. Psychonomic Bulletin & Review, 10(3), 645–652. Chen, L., & Boland, J. E. (2008). Dominance and context effects on activation of alternative homophone meanings. Memory & Cognition, 36(7), 1306–1323. COSMAS-I Corpora (2002). Corpus search, management and analysis system (on-line). Mannheim: Institut für deutsche Sprache.
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1343
Neely, J. H. (1977). Semantic priming and retrieval from lexical memory: Roles of inhibitionless spreading activation and limited-capacity attention. Journal of Experimental Psychology: General, 106(3), 226–254. Neely, J. H. (1991). Semantic priming effects in visual word recognition: A selective review of current findings and theories. In D. Besner & G. Humphreys (Eds.), Basic processes in reading: Visual word recognition (pp. 264–336). Hillsdale, NJ: Erlbaum. Neill, W. T. (1997). Episodical retrieval in negative priming and repetition priming. Journal of Experimental Psychology: Learning, Memory and Cognition, 23, 1291–1305. Norris, D. (1994). Shortlist: A connectionist model of continuous speech recognition. Cognition, 52, 189–234. Onifer, W., & Swinney, D. A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory and Cognition, 9(3), 225–236. Paul, S. T., Kellas, G., Martin, M., & Clark, M. B. (1992). Influence of contextual features on the activation of ambiguous word meanings. Journal of Experimental Psychology: Learning, Memory and Cognition, 18(4), 703–717. Rodd, J. M., Gaskell, M. G., & Marslen-Wilson, W. D. (2002). Making sense of semantic ambiguity: Semantic competition in lexical access. Journal of Memory and Language, 46, 245–266. Schvaneveldt, R. W., Meyer, D. E., & Becker, C. A. (1976). Lexical ambiguity, semantic context, and visual word recognition. Journal of Experimental Psychology: Human Perception and Performance, 2(2), 243–256. Shapiro, K. L., Driver, J., Ward, R., & Sorensen, R. E. (1997). Priming from the attentional blink: A failure to extract visual tokens but not visual types. Psychological Science, 8, 95–100. Simpson, G. B. (1981). Meaning dominance and semantic context in the processing of lexical ambiguity. Journal of Verbal Learning and Verbal Behavior, 20(1), 120–136. Simpson, G. B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Bulletin, 96, 316–340. Simpson, G. B., & Burgess, C. (1985). Activation and selection processes in the recognition of ambiguous words. Journal of Experimental Psychology: Human Perception and Performance, 11, 28–39. Simpson, G. B., & Kang, H. (1994). Inhibitory processes in the recognition of homograph meanings. In D. Dagenbach & T. H. Carr (Eds.), Inhibitory processes in attention, memory and language (pp. 359–381). San Diego, CA: Academic Press. Simpson, G. B., & Krueger, M. A. (1991). Selective access of homograph meanings in sentence context. Journal of Memory and Language, 30, 627–643. Tabossi, P., & Zardon, F. (1993). Processing ambiguous words in context. Journal of Memory and Language, 32, 359–372. Tanenhaus, M. K., Leiman, J. M., & Seidenberg, M. S. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18(4), 427–440. Till, R. E., Mross, E. F., & Kintsch, W. (1988). Time course of priming for associate and inference words in a discourse context. Memory & Cognition, 16(4), 283–298. Tipper, S. P. (2001). Does negative priming reflect inhibitory mechanisms? A review and integration of conflicting views. Quarterly Journal of Experimental Psychology: Human Experimental Psychology, 54A(2), 321–343. Treisman, A. M., & Gelade, G. (1980). A feature-integration theory of attention. Cognitive Psychology, 12, 97–136. Van den Bussche, E., & Reynvoet, B. (2008). The asymmetry between top-down effects and unconscious cognition: Additional issues. Consciousness and Cognition, 17(4), 1361–1363. doi:10.1016/j.concog.2008.05.003. Van den Bussche, E., Segers, G., & Reynvoet, B. (2008). Conscious and unconscious proportion effects in masked priming. Consciousness and Cognition, 17(4), 1345–1358. Van den Bussche, E., Van Den Noortgate, W., & Reynvoet, B. (2009). Mechanisms of masked priming: A meta-analysis. Psychological Bulletin, 135, 452–477. Vu, H., Kellas, G., & Paul, S. T. (1998). Sources of sentence constraint on lexical ambiguity resolution. Memory & Cognition, 26(5), 979–1001. Weber, M. R. (1970). First-graders’ use of grammatical context in reading. In H. Levin & J. P. Williams (Eds.), Basic studies on reading (pp. 147–163). New York: Basic Books. Wentura, D., & Frings, C. (2005). Repeated masked category primes interfere with related exemplars: New evidence for negative semantic priming. Journal of Experimental Psychology: Learning, Memory and Cognition, 31(1), 108–120.
Contents lists available at ScienceDirect
Consciousness and Cognition journal homepage: www.elsevier.com/locate/concog
Visible homonyms are ambiguous, subliminal homonyms are not: A close look at priming Doris Eckstein a,⇑, Matthias Kubat a,b, Walter J. Perrig a a b
Department of Psychology, University of Bern, Switzerland Distance Learning University Switzerland, Brig, Switzerland
a r t i c l e
i n f o
Article history: Received 9 February 2010 Available online 12 June 2011 Keywords: Homonym Polysemy Ambiguity Priming Awareness Consciousness Masked priming
a b s t r a c t Homonyms, i.e. ambiguous words like ‘score’, have different meanings in different contexts. Previous research indicates that all potential meanings of a homonym are first accessed in parallel before one of the meanings is selected in a competitive race. If these processes are automatic, these processes of selection should even be observed when homonyms are shown subliminally. This study measured the time course of subliminal and supraliminal priming by homonyms with a frequent (dominant) and a rare (subordinate) meaning in a neutral context, using a lexical decision task. In the subliminal condition, priming across prime-target asynchronies ranging from 100 ms to 1.5 s indicated that the dominant meaning of homonyms was facilitated and the subordinate meaning was inhibited. This indicates that selection of meaning was much faster with subliminal presentation than with supraliminal presentation. Awareness of a prime might decelerate an otherwise rapid selection process. Ó 2011 Elsevier Inc. All rights reserved.
1. Introduction One of the characteristics of intelligent thinking is flexibility, which is needed when ambiguous information is processed. For example, owing to our ability to resolve ambiguity, we intuitively assign different meanings to the word ‘score’ when talking about exams compared to when talking about music. Ambiguous words like ‘score’ or ‘paper’ are called homonyms, i.e. words that have at least two distinct meanings which cannot be distinguished by either the word’s writing or pronunciation. Due to their ambiguity, homonyms are ideal for investigating presumed stages of word recognition, including access to meaning, selection of meaning and the role of context in this process. Although most studies on homonyms have focused on the influence of context on meaning selection (e.g., Binder & Morris, 1995; Chen & Boland, 2008; Gottlob, Goldinger, Stone, & Van Orden, 1999; Marslen-Wilson, 1987; Moritz, Mersmann, Quast, & Andresen, 2001; Onifer & Swinney, 1981; Paul, Kellas, Martin, & Clark, 1992; Schvaneveldt, Meyer, & Becker, 1976; Simpson, 1981; Simpson & Krueger, 1991; Tabossi & Zardon, 1993; Tanenhaus, Leiman, & Seidenberg, 1979; Vu, Kellas, & Paul, 1998), one strand of research has concentrated on measuring meaning selection when no context or a neutral context is given. In single-word priming, neutral trials are usually formed by pairing homonym primes with unrelated target words (e.g., ‘calf’–‘wine’) and experimental trials are formed by pairing homonym primes with targets related to one of their meanings (e.g., ‘calf’–‘cow’). The difference in reaction times between experimental and biased trials indicates the amount of priming related to the prime-target relationship. A couple of single-word priming time course studies using polar homonyms, i.e. homonyms with a dominant (more frequent) (e.g., ‘calf’–‘cow’) and a subordinate (less
⇑ Corresponding author. Address: Institut für Psychologie, Muesmattstr. 45, 3000 Bern 9, Switzerland. Fax: +41 316318212. E-mail address: [email protected] (D. Eckstein). 1053-8100/$ - see front matter Ó 2011 Elsevier Inc. All rights reserved. doi:10.1016/j.concog.2011.05.010
1328
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343 Table 1 Examples of possible prime-target pairs within a test. In Experiment 1, control primes were neutral words. In Experiment 2, homonym primes and their subordinate/dominant related words were recombined to form unrelated pairs in the control condition.
Experiment 1 Subordinate Dominant Control Experiment 2 Experimental – Subordinate Experimental – Dominant Control – Subordinate Control – Dominant
Prime
Target
PORT BANK POPPY
WINE ACCOUNT SCENE
PORT BANK BANK PORT
WINE ACCOUNT WINE ACCOUNT
Remark: English examples are given for illustration. See the Appendices for the word set and stimulus rotations.
frequent) meaning (e.g., ‘calf’–‘leg’) (e.g., Marcel, 1980; Simpson & Burgess, 1985), reported positive priming for both meanings of a homonym at short SOAs (i.e., SOAs between 0 and 300 ms), indicating that all meanings of a word are more accessible first, whereas later on, the dominant meaning has a clear advantage over other meanings (cf. Duffy, Morris, & Rayner, 1988; Simpson, 1981; Simpson & Burgess, 1985). A similar time course was found with sentence studies and single word studies with normal words (e.g., Tanenhaus et al., 1979; Till, Mross, & Kintsch, 1988). These findings of short-SOA priming were taken as evidence for an early parallel access to meanings (cf. Simpson, 1984; Simpson & Burgess, 1985), occurring within 300 ms after word presentation, followed by selection of one meaning and suppression of nonselected meanings. Whereas researchers seem to agree on the conclusion that parallel access is automatic in nature (given that priming at short SOAs is likely to be automatic, cf. Neely, 1977; Neely, 1991), it is unclear whether the late selection/suppression pattern of priming is reflecting automatic competitive selection (interactive activation model, McClelland & Rumelhart, 1981; TRACE model, McClelland & Rumelhart, 1986; Contextual-Feature-Sensitive Model, Vu et al., 1998), or strategic and expectationrelated controlled selection (e.g., Neely, 1977; Neely, 1991; Simpson, 1984). Modular selection accounts (e.g. Forster, 1976) identify this stage with context integration, during which one is in a state of openness for multiple interpretations of a word, or a readiness for revision of the first choice. In this study, we used subliminal priming to determine the extent to which processes of meaning selection are indeed automatic. Subliminal presentation of primes was used to reduce the contribution of strategic, controlled processes in priming.1 By using a design with several prime-target SOAs (100, 300, 600, 900 and 1500 ms), we sought to determine whether there was a time point at which priming for different meanings of the homonyms diverged. More specifically, the time point of automatic selection of a meaning would be indicated by the inhibition of one meaning (the subordinate meaning) compared to the other meaning of the homonym (the dominant meaning). For comparison, the same test was also used with visible primes. The time course of subliminal and supraliminal homonym priming in a neutral context was measured in three experiments. Polar homonyms like ‘port’ were used, which have one dominant meaning (harbour) which is statistically chosen more frequently than the subordinate meaning (beverage) in neutral contexts. In Experiment 1, prime-target SOA was varied between 100 ms and 900 ms SOAs. Experiment 2 varied the prime-target SOA across a range of 100–1500 ms. 2. Experiment 1 In this experiment, priming of subordinate and dominant meanings of polar homonyms was measured at four prime-target SOAs (100 ms, 300 ms, 600 ms, 900 ms) using a lexical decision task. In the subliminal priming test, primes were sandwich masked using individually titrated presentation timing in order to ensure subliminal presentation. In the visible priming test, primes were shown without masks and were therefore clearly visible. In each test, two experimental conditions and one control condition was measured, whereby each prime and target word was shown not more than once across both tests. In the experimental conditions, homonym primes were paired with a target word related to the subordinate or the dominant meaning of the homonym. In the control condition, neutral primes were paired with unrelated targets, which were related to the subordinate or the dominant meaning of the homonyms used in the experimental condition (cf. Table 1). Priming was computed as the difference in lexical decision times to target words preceded by a related homonym and the same words preceded by an unrelated word.
1 When using subliminal perception to reduce controlled processing, one has to be careful in controlling influences of attentional and strategic processes, which have been shown to still have influence on subliminal priming (e.g. Dehaene, Changeux, Naccache, Sackur, & Claire Sergent, 2006; Eckstein & Perrig, 2007; Van den Bussche, Van Den Noortgate, & Reynvoet, 2009). In this study, attentional requirements were identical in the masked and the visible priming tests and strategic influences were minimized by reducing across-trial repetitions of words.
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1329
2.1. Method 2.1.1. Participants Subjects were 85 students (27 men, 58 women) aged 23.6 years (SD = 4.5) of the Institute of Psychology of the University of Bern, who received credit for an introductory psychology course. All of them were naive as to the hypothesis of the experiment. 2.1.2. Design Two 4 3 mixed-models ANOVA with the factor SOA (100 ms, 300 ms, 600 ms and 900 ms) varied between subjects and the factor ‘Prime-Target Relation’ (homonym prime – subordinate target, homonym prime – dominant target, neutral prime – unrelated target) varied within subjects were implemented separately for the masked and nonmasked tests. Examples of prime-target pairs are given in Table 1. 2.1.3. Material and apparatus For the experimental set, 80 German polar homonym words and their associates were selected from our own standardisation of homonyms and a list of homonyms published in Moritz et al. (2001), in which the naming frequency of different meanings of homonyms was measured, i.e. the number of times a specific meaning comes to mind first when asking a large sample of participants. All homonym words were true homonyms, i.e. their subordinate and dominant meaning represented not merely different senses of the same word but described different meanings. The minimum naming frequency of dominant associates was 52% (M = 77%), the maximum naming frequency of subordinate associates was 39% (M = 20%) and the difference between the two ranged between 23% and 93% (M = 57%). Then, 80 additional neutral words were selected and assigned to the homonyms such that they were unrelated to any of the homonym’s meanings. The 80 homonym-subordinate-dominant-neutral word quadruples (see Appendix 2) were divided into four groups equated in word length, syllable length, naming frequency, log word frequency (using the ‘‘Tagesanzeiger 1996–2000’’ Cosmas corpus of text words containing 75 million words, COSMAS-I Corpora, 2002) and orthographic similarity of prime and target. Subordinate words and dominant words did not differ in average in all the controlled variables, and prime-target pairs had very low average orthographic similarity (M = 65 as computed according to Weber, 1970). The four groups of homonyms and their associates were used to counterbalance word pairs over conditions, tests and participants. Finally, 80 additional words with linguistic properties similar to the homonyms were selected to be used as primes in the nonword target condition, and 30 additional words were selected for the threshold-setting conditions (see Appendices 1 and 2 for the word set and its characteristics). A total of 210 pronounceable nonwords were constructed by recombining consonant/vowel- and vowel/consonant-pairs occurring in each position in the words of the word set. Nonword length and number of syllables was equated to the word set. Of these, 120 nonwords were used as targets in the masked and nonmasked test, and the remaining 90 nonwords were used in the threshold-setting test. A standard IBM-compatible PC with CRT monitor (refresh frequency at 70 Hz) was used for display of the stimuli and an external button box was used for response. 2.1.4. Procedure SOA was determined for each participant before test, and remained the same throughout the whole test session. The experiment started with a prime visibility threshold-setting procedure, which was followed by the masked priming test and then the nonmasked priming test. Each of the three parts of the experiments was preceded by a short practice with a separate set of words and nonwords in which response feedback was given. 2.1.4.1. Trial sequence. All stimuli were written in black on a light grey background. The timeline is illustrated on the left in Fig. 1. Appearance of the fixation point indicated that participants could start the trial, which they did by simultaneously pressing both response keys. Display of premask, prime, postmask and target was synchronized with the screen refresh signal. The prime and postmask durations were individually set in the threshold-setting task. Participants had to give their response within a 1.5 s response window and the target was cleared when a response was given or after 1.5 s, whichever shorter. The intertrial interval lasted 500 ms. 2.1.4.2. Threshold-setting procedure. Participants were informed as to the trial structure and also informed on the sequence of the threshold-setting procedure. Their task was to decide whether the masked letter string was a word or not. Trial sequence and display was the same as in the test trials except that (a) words and nonwords were taken from another stimulus set and (b) all targets were nonwords. Prime duration was set to 57 ms at the start. Blocks of 10 trials each were presented, whereby each time when the mean accuracy over a block was higher than 60%, prime duration was lowered by 14.3 ms and postmask duration was increased by 14.3 ms, so that prime-target SOA remained constant. Accuracy feedback was given to participants at the end of each block. If accuracy was lower or equal to 60% in two succeeding blocks or if 10 blocks had been done, the threshold-setting procedure was concluded. After threshold setting, prime and postmask durations were held constant over the masked priming and nonmasked priming tests. Presentation of prime words and nonwords was randomized within the constraint that 50% of primes within a block should be words. Presentation of target nonwords was random.
1330
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
Fig. 1. Trial timing in Experiment 1 (left) and Experiment 2 (right). Prime and postmask presentation times were individually set in a threshold-setting test. In the visible conditions, pre- and postmasks were replaced by empty screens.
2.1.4.3. Masked priming test. Participants were instructed to concentrate on the target from now on and to decide whether the target was a word or not. A total of 120 trials were shown separated by a short break; 50% of the trials presented nonword targets and 50% presented word targets. Of the word trials, 20 trials presented a homonym-subordinate prime-target pair, 20 trials presented a homonym-dominant prime-target pair and 20 trials presented a neutral-subordinate or a neutraldominant prime-target pair. In order to counterbalance the prime-target pairs across conditions and the masked and visible tests such that none of the prime and target words was shown more than once during the masked and visible priming tests, four lists of homonym-associate pairs were rotated across participants such that (a) each homonym occurred only once in a list, (b) each homonym occurred once with its dominant associate and once with its subordinate associate across the four lists, (c) each neutral prime occurred once in a list (half in the control condition, half in the nonword condition), (d) each homonym’s subordinate and dominant associates were paired once with a neutral word across the four lists and (e) none of the targets occurred more than once in each list (see Appendix 3 for the counterbalancing procedure). The sequence of trial conditions was randomised. 2.1.4.4. Visible priming test. Participants were informed that although they would now see the primes they should carry on doing the lexical decision on targets. Here, 60 nonword target and 60 word target trials were shown with randomisation similar to the masked test. Each homonym prime was shown once across the masked and visible priming tests. None of the word targets shown in the masked test was shown in the visible test. Assignment of word pairs to conditions and tests was counterbalanced across participants as described in the masked priming test paragraph (cf. Appendix 3). 2.1.5. Analysis For each subject, session, test and condition, trials with accurate word responses in trials that were within 2.5 standard deviations from the mean were used for RT analysis. Nonword trials were not analysed. Mean RTs by condition and subject were used as dependent variables in the RT analyses, and arcsine transformed accuracies were used in the accuracy analyses. Greenhouse–Geisser corrections were used when sphericity of data was violated, and a 5% type I error criterion was used unless stated otherwise. 2.2. Results Data of 11 participants (3, 2, 2 and 4 in each of the 100–900 ms SOA conditions) had to be excluded from analysis due to low diligence (less than 80% correct responses), leading to valid data from 23, 22, 20 and 20 participants in the 100 ms, 300 ms, 600 ms and 900 ms SOA conditions, in that order. Distribution of prime presentation durations across participants was comparable in each SOA condition: Prime duration was set to 57.1 ms, 42.9 ms, 28.6 ms and 14.3 ms for 77%, 19%, 2% and 2% of the participants, respectively. Accuracy was high with 94.4% correct responses in the masked priming test and 94.8% correct responses in the visible priming test. 2.2.1. Analyses across the masked and visible priming tests 2.2.1.1. Reaction times and accuracies. The hypothesis that related primes would lead to different classification times compared to unrelated primes was tested with two separate omnibus ANOVAs on RTs and arcsine transformed accuracies, using
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1331
the factors Prime-Target Relation, Test and SOA (see Fig. 2, Tables 2 and 3 for descriptives and Table 4 for the outcome of statistical analyses). In both RTs and accuracies, a highly significant main effect of Prime-Target Relation indicated that prime-target relationship and masking affected RTs. Furthermore, Prime-Target Relation and Test interacted with each other, suggesting that priming was different in the masked compared with the visible priming test. Neither the main effect of SOA, nor any interaction with SOA reached significance. The masked and visible tests were then analysed separately. 2.2.2. Masked primes 2.2.2.1. Reaction times. Descriptives of the results are given in Table 2 (see also Fig. 2). Outlier RTs and errors led to removal of 7.2% of trials from analysis. The hypothesis that related primes would lead to different classification times compared to unrelated primes was tested with an omnibus ANOVA on RTs across Prime-Target Relation and SOA (see Table 4). A highly significant main effect of Prime-Target Relation indicated that the prime-target relationship affected RTs. Separate ANOVAs comparing the experimental conditions with the control condition indicated that dominant priming was significantly greater than zero and that subordinate priming was significantly negative. Neither the main effect of SOA nor the interaction of Prime-Target Relation with SOA reached significance. Descriptives and outcomes of t-tests for comparisons at single SOAs are given in Table 2: only in the 300 ms SOA condition was subordinate priming significantly negative. 2.2.2.2. Power analyses. Given the clear effect of priming across SOAs, post hoc power analyses were done on RT results in order to determine the achieved power in testing for priming of experimental vs. control trials at single SOAs (using the GPower 3 program by Faul, Erdfelder, Lang, & Buchner, 2007). The priming effects of about 10 ms corresponded to a small effect size (d = 0.3), where average power was at 70% with 75 participants, and at 30% with 23 participants. Hence, tests at single SOAs lacked the power to detect 10 ms-priming. In contrast, the 20 ms difference between homonym-subordinate and homonym-dominant RTs represented a medium effect (d = 0.5), which corresponded to an achieved power of over 90% for 75 participants, and 70% for 23 participants. These analyses indicate that the statistical tests used to test for simple priming effects at single SOAs lacked power. 2.2.2.3. Accuracies. Mean errors by condition are given in Table 3. In the omnibus ANOVA on arcsine transformed accuracies across Prime-Target Relation and SOA, a significant main effect of Prime-Target Relation indicated that the prime-target relationship affected accuracies (cf. Table 4). Neither the main effect of SOA nor the interaction of Prime-Target Relation with
Fig. 2. RT priming in Experiments 1 and 2. Grey lines indicate priming for the subordinate meaning, i.e. RTs to target words related to the less frequent meanings of homonyms when preceded by the homonym itself, compared to an unrelated word. Dark lines indicate priming for the dominant meaning, i.e. RTs to target words related to the more frequent meaning of homonyms when preceded by the homonym itself, compared to unrelated word. ⁄⁄p < .01, ⁄ p < .05. Bars indicate standard errors.
1332
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
Table 2 Reaction times (ms) and priming in Experiments 1 and 2. SOA
RT neutral
RT subordinate
RT dominant
Subordinate priming
Dominant priming
M
SE
M
SE
M
SE
M
SE
Experiment 1 – Masked primes 100 ms 610 15 300 ms 621 15 600 ms 627 14 900 ms 635 16 All SOAs 623 8
621 638 638 634 632
17 17 18 16 9
608 613 616 613 612
14 12 12 13 6
11 17* 11 1 10*
8 7 11 12 5
1 8 11 22 10*
8 7 10 12 5
Experiment 1 – Visible primes 100 ms 621 17 300 ms 607 18 600 ms 595 14 900 ms 605 16 All SOAs 608 8
614 588 593 604 600
16 16 11 15 7
589 578 582 584 583
14 18 13 14 7
7 19* 2 1 8*
8 10 10 7 5
33** 29** 13 22 24**
8 11 9 14 5
Experiment 2 – Masked primes 100 ms 634 15 300 ms 554 11 900 ms 542 13 1500 ms 564 12 All SOAs 571 7
657 566 549 572 583
16 12 13 13 8
636 545 535 551 564
16 11 14 11 7
23** 12** 7 8a 12**
6 4 5 5 3
2 9** 7 13** 7**
5 3 5 3 2
Experiment 2 – Visible primes 100 ms 604 13 300 ms 534 12 900 ms 538 11 1500 ms 552 11 All SOAs 555 6
615 528 537 547 554
16 12 13 11 7
599 532 522 538 546
14 12 10 11 6
11 7 1 6 1
7 4 3 4 2
5 2 17** 14** 10**
5 5 4 4 2
M
SE
Remark: One-tailed t-tests were used in the visible conditions. a p < .10. * p < .05. ** p < .01.
Table 3 Errors (in%) and error priming in Experiments 1 and 2. SOA
RT neutral
RT subordinate
RT dominant
Subordinate priming
Dominant priming
M
SE
M
SE
M
SE
M
SE
Experiment 1 – Masked primes 100 ms 6.1 1.0 300 ms 3.9 0.4 600 ms 4.0 1.0 900 ms 5.3 0.8 All SOAs 4.8 0.4
7.6 3.4 5.5 6.0 5.6
0.8 0.5 0.9 0.9 0.4
5.4 3.4 4.3 2.8 4.0
1.1 0.7 1.1 1.0 0.6
1.5 0.5 1.5 0.7 0.8
1.4 0.7 1.6 1.4 0.5
0.7 0.5 0.3 2.5 0.8
1.9 1.0 2.0 1.6 0.8
Experiment 1 – Visible primes 100 ms 10.0 1.0 300 ms 4.8 0.5 600 ms 7.5 0.9 900 ms 7.3 0.9 All SOAs 7.4 0.5
7.0 6.1 6.0 6.3 6.4
0.8 0.9 1.0 0.9 0.4
3.5 4.5 3.8 3.3 3.8
0.8 0.8 1.3 0.7 0.4
3.0 1.4 1.5 1.0 1.1
1.6 0.7 1.6 1.4 0.8
Experiment 2 – Masked primes 100 ms 5.4 0.8 300 ms 2.3 0.4 900 ms 3.3 0.5 1500 ms 2.4 0.4 All SOAs 3.2 0.3
5.1 3.1 4.3 2.7 3.7
0.9 0.7 0.8 0.6 0.4
4.1 2.9 2.8 2.7 3.1
0.6 0.6 0.6 0.6 0.3
0.3 0.8 1.1 0.3 0.5
0.8 0.6 0.7 0.7 0.4
Experiment 2 – Visible primes 100 ms 5.3 1.1 300 ms 3.5 0.5 900 ms 3.1 0.6 1500 ms 2.9 0.5 All SOAs 3.6 0.3
5.5 4.9 3.2 3.1 4.1
1.0 0.8 0.6 0.6 0.4
3.6 1.6 2.8 1.6 2.3
0.8 0.4 0.5 0.4 0.3
0.2 1.4 0.1 0.2 0.5
1.0 0.8 0.5 0.7 0.4
M
SE
Remark: Analyses were done on arcsine transformed accuracies. * p < .05. p < .01.
**
6.5** 0.2 3.8** 4.0** 3.6** 1.3 0.6 0.5 0.3 0.2* 1.7* 1.9** 0.3 1.3** 1.3**
1.6 1.0 2.0 1.6 0.8 0.9 0.6 0.5 0.7 0.3 0.9 0.6 0.5 0.6 0.3
1333
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343 Table 4 ANOVA results of Experiment 1. Analysis/effect
df1
df2
Reaction times
Accuracies
F
MSE
p
F
MSE
p
11.49 6.99 3.54
0.041 0.032 0.034
.010 .035
3.41 1.72 1.78
0.040 0.096 0.061
.036 .19 .19
13.86 30.40 2.27
0.032 0.055 0.062
⁄⁄
Omnibus ANOVA Prime-Target Relation Test Prime-Target Relation Test
2 1 2
162 81 162
18.40 27.81 3.53
973 3074 922
⁄⁄
Masked test Prime-Target Relation Dominant vs. Control Subordinate vs. Control
2 1 1
162 81 81
8.77 5.12 16.59
955 1856 2016
⁄⁄
Visible test Prime-Target Relation Dominant vs. Control Subordinate vs. Control
2 1 1
162 81 81
13.59 20.67 2.52
940 2352 1766
⁄⁄
⁄⁄
.031
.026 ⁄⁄
⁄⁄
.11
⁄⁄
⁄⁄
.14
Remark: Significant p-values are printed in bold. ⁄⁄p < .001, ns: p > .20. The main effect of SOA and interactions with SOA did not reach significance in any of the analyses, F’s < 1.0, p > .20 (except for the interaction of Test SOA, F (3, 81) = 2.20, p = .09).
SOA reached significance. In contrast to RT analyses, priming effects were not significant in follow-up contrasts and at single SOAs. Significant priming effects were thus only observed in reaction times. 2.2.3. Visible primes 2.2.3.1. Reaction times. The overall ANOVA with the factors Prime-Target Relation and SOA yielded a highly significant main effect of Prime-Target Relation. Neither the main effect of SOA nor the interaction of Prime-Target Relation with SOA was significant. Only in the dominant condition were RTs significantly faster than in the control condition, whereas the difference between the subordinate and the control condition only approached significance (corresponding to a marginal p = .051 when using a one-sided t-test). As can be seen in the right panel in Fig. 2, separate t-tests at single SOAs indicated that at 100 ms and 300 ms SOAs, RTs to dominant targets were significantly faster than RTs in the control condition, whereas RTs to subordinate targets were faster than RTs in the control condition at 300 ms SOA (see also Table 2). 2.2.3.2. Accuracies. The overall ANOVA with the factors Prime-Target Relation and SOA yielded a highly significant main effect of Prime-Target Relation. Neither the main effect of SOA nor the interaction of Prime-Target Relation with SOA was significant. Only in the dominant condition were accuracies significantly higher than in the control condition, whereas the difference between the subordinate and the control condition was not significant. Separate t-tests at single SOAs indicated that responses to dominant targets were significantly better than in the neutral condition at SOAs of 100 ms, 600 ms and 900 ms, whereas responses to subordinate targets where not more accurate than in the neutral condition at any SOA (cf. Tables 3 and 4). 2.3. Discussion In visible priming, the following pattern was observed: Dominant priming was significant across SOAs, whereby priming was significant in RTs or accuracies at all SOAs. Subordinate priming was only marginally significant across SOAs, whereby there was an apparent maximum of priming at 300 ms SOA. This pattern is consistent with a model of parallel access of meanings, as suggested by previous research. Masked homonym RT priming, in contrast, indicated facilitation of dominant meanings and inhibition of subordinate meanings across SOAs. There was no apparent change of priming with time (i.e., no interaction with SOA). This pattern of results, and the finding that subordinate priming was significantly negative at 300 ms SOA, supported an automatic selection account: an early selection of the dominant meaning and simultaneous suppression of the subordinate meaning. However, there is an intriguing peculiarity in the data, which is that subordinate meanings of homonyms were actually inhibited and not just neutral. What could be the reason for the inhibition observed in subordinate priming? One known cause of negative priming in the priming literature has been investigated with negative priming paradigms (e.g., Neill, 1997; Tipper, 2001), in which a slowdown of responses to items is observed if these items had to be actively ignored earlier on. This might also be an issue in our study, as negative priming has been reported with masked primes in within-trial priming (e.g., Frings & Eder, 2009). Negative priming is thought to be due either to (a) attentional inhibition, whereby withdrawing attention from a to-be-ignored prime slows the processing of succeeding related stimuli (Tipper, 2001), or to (b) a conflict between episodic retrieval of prime (non) response and response to a target related to it (Neill, 1997). In the present study, both causes are conceivable: Firstly, because the mask-prime-mask sequence appears as a series of flashes, participants might withdraw attention from this sequence altogether, which would also inhibit any attentional processing of the prime word; secondly, the requirement
1334
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
that one is not to respond to an (unconsciously) perceived word might lead to retrieval of a no-response trigger in lexical decision to the target. In both cases, one would expect such negative priming by a masked prime to be stronger the more effective a prime is as a distractor, which is the case if it is related to the target word. Before we expand on this possibility of negative priming and other explanations, we present Experiment 2, which aimed to replicate the subordinate meaning inhibition effect found in this experiment.
3. Experiment 2 Experiment 2 consisted of two experiments, Experiment 2a and 2b, which were mainly designed to validate the results of Experiment 1. However, given that power was low for most between-SOA tests in Experiment 1, a couple of changes were made in the design used in Experiment 1: (a) The word set was reduced to include only homonyms whose dominant and subordinate meaning were more clearly separated in naming frequency, (b) the number of trials per condition was increased by using all targets of the word set in each test, (c) each homonym prime was shown once in the control condition and once in the experimental condition, (d) two control conditions were used in which homonym primes were paired with (subordinate or dominant) targets related to other homonyms of the experimental set, (e) the range of SOAs was extended in order to detect potentially slow decay of priming (SOAs were 100 ms, 300 ms, 900 ms and 1500 ms) and (f) a prime visibility test was added in order to measure whether prime visibility increased over the masked test. Experiment 2a included SOAs of 300, 900 and 1500 ms that were varied within participants, using three separate sessions. In Experiment 2b, priming at the 100 ms SOA was measured using the same procedure for the masked and visible priming tests, using an extended threshold-setting procedure and a longer discrimination test, and an additional forced-choice discrimination test at the end of the experiment. The reason for this was that it is quite uncommon to find significant priming effects with truly invisible primes, especially when using large stimulus sets as in the reported experiments. One likely explanation for the effects found in the previous experiments could be partial prime perception: when a participant sees a target that is related to the preceding prime, this might help him/her to recognise the otherwise illegible prime word, which directly leads to faster lexical decision. Because we used a relatively benign threshold-setting procedure and the resulting prime presentation times were relatively long, compared to other studies of this type, partial prime visibility might be the main driver for the priming effects. We addressed this hypothesis in Experiment 2b (a) by using a more restrictive threshold-setting procedure and testing whether this led to shorter prime presentations, (b) by doubling the trials in the prime visibility test in order to obtain a more sensitive measure of prime recognition and (c) by adding a further prime visibility test assessing partial prime visibility. Priming analyses are reported on the data collapsed across Experiments 2a and 2b; correlation analyses of priming and visibility measures are reported separately for Experiments 2a and 2b. Hypotheses were not changed, although we set out to validate results of Experiment 1, which were not in line with any of our a priori hypotheses, in particular (a) slower classification in homonym-subordinate trials than in unrelated trials in the masked priming test, and (b) no interaction of priming with SOA. The visible condition was expected to show the pattern of RTs found in Experiment 1. 3.1. Method 3.1.1. Participants Participants in Experiment 2a were 30 students (24 women, 6 men) aged 25 years on average (SD = 8.8). Participants of Experiment 2b were 25 students (22 women, 3 men) aged 22.7 years on average (SD = 3.7). All participants were students of the Institute of Psychology of the University of Bern who received credit for an introductory psychology course. All of them were naive as to the hypothesis of the experiment. 3.1.2. Design In Experiment 2a, the factors SOA (300 ms, 900 ms and 1500 ms) and Prime-Target Relation (homonym prime – subordinate target, homonym prime – dominant target, homonym prime – unrelated target collapsed across target types) were varied within subjects for both the masked and nonmasked tests. Experiment 2b implemented the same design except that SOA was fixed at 100 ms. Analyses on data collapsed across Experiments 2a and 2b thus used a mixed-model design with SOA varied between participants (four levels: 100 ms, 300 ms, 900 ms and 1500 ms) and Prime-Target Relation varied within participants. 3.1.3. Material and apparatus For the experimental set, 60 homonyms and their associates were selected from the word set of Experiment 1 such that the probability of naming the dominant associate when presented with the homonym was higher than 70%, and the probability of naming the subordinate associate when presented with the homonym was lower than 30%, according to the word norms mentioned in Experiment 1 (cf. Appendix 1). Thirty pairs of homonyms were formed such that their associates were not related to the other homonym of a pair. The 30 pairs were divided into two lists of 15 pairs equated in frequency, naming probability and orthographic similarity. These lists of homonym pairs and their associates were used to counterbalance word
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1335
pairs over tests, sessions and participants (see procedure and Appendix 3). Sixty words with similar linguistic properties as the homonyms were selected as primes in the nonword target condition. Fifty additional neutral words were selected for the threshold-setting conditions. PCs with same configuration as in Experiment 1 were used with a standard keyboard as input device. Two-hundred and seventy nonwords used in the threshold-setting and priming test were constructed in the same way as described in Experiment 1. 3.1.4. Procedure of Experiment 2a Each participant did three sessions that were 1 week apart from each other. In each session, one SOA condition was tested. The order of SOAs over sessions was counterbalanced across participants. Each session started with a prime visibility threshold-setting procedure, which was followed by the masked priming test. After this, prime visibility was measured, followed by the visible priming test. The threshold-setting, masked priming and visible priming test were each preceded by a short practice with a separate set of words and nonwords. Trial timing was the identical to Experiment 1, with the exception that the postmask was shown for 100 ms and then replaced by an empty screen (cf. Fig. 1 on the right). The threshold procedure was identical to Experiment 1, with the exception that prime duration was set to 71 ms at the start of the procedure. Trial sequence and display was the same as in the test trials, except that different words and nonwords were used for primes and targets and that all targets were nonwords. After threshold setting, the prime and mask presentation times were held constant over the masked priming, visibility and visible priming tests within a session. The masked priming test consisted of 240 trials separated by three short breaks. Instruction, trial timing and masks were otherwise identical to Experiment 1. Across every 120 trials, 60 nonword target trials, 15 homonym-subordinate, 15 homonym-dominant, 15 unrelated-subordinate and 15 unrelated-dominant trials were shown in a randomised sequence. In order to avoid repetition effects, homonyms were never shown with the same target within and across the masked and visible tests of one session. Each homonym was shown once in 120 trials, and over 240 trials, each homonym was shown once in an experimental condition and once in a control condition, and each target was shown once in each test (see Appendix 3 for the counterbalancing procedure). The visibility test consisted of a random sequence of 15 nonword prime and 15 word prime trials with identical trial timing and material as in the threshold-setting test. Participants were asked to decide whether the prime was a word or not. In the visible priming test, instruction and trial timing were identical to Experiment 1. A total of 120 nonword target and 120 word target trials were shown, whereby prime and target pairing was controlled as described in the masked test paragraph. The order of related and unrelated trials for a given homonym and dominance was counterbalanced across participants and sessions. 3.1.5. Procedure of Experiment 2b Participants did one session only with the SOA set at 100 ms, using the same procedure as in Experiment 2a, with the following adaptations: (a) an attempt to increase the sensitivity of the threshold procedure was made by increasing the threshold block length to 20 instead of 10 trials, and by setting the threshold exit criterion to smaller or equal to 55%, (b) the prime visibility test using lexical decision was extended to 30 nonword prime and 30 word prime trials and (c) at the end of the test, an additional forced choice test assessing partial awareness of primes was conducted. 3.1.5.1. Partial prime visibility test. In this final test, participants were again asked to try to recognise the prime in each trial, whereby the usual prime-target word trial sequence was extended by a two-stage decision task, i.e. a visibility rating and a forced-choice word selection. After each prime-target sequence, participants were asked to do a visibility rating, i.e. to choose the best fitting response of three alternatives describing what they had seen: (a) they had not seen the prime, (b) they had seen a part of the prime, as for instance single letters or letter combinations or (c) they had seen the prime word. In the second, forced choice selection task, they were asked to indicate which of two choice words matched the previously presented prime, whereby they were encouraged to guess the word if they were unsure. Stimulus presentation and input was controlled by the PC: After the mask-prime-mask sequence (cf. Fig. 1), the target was shown as usual, but was not erased after 1.5 s. Instead, a selection scale appeared at the bottom screen, indicating the three response alternatives of the visibility rating, corresponding to the outer left, middle and outer right keys of the button box. They were given 9.5 s for this decision. After their response, a second screen was presented for the forced choice response, with two choice words displayed on the left and on the right of the screen centre. The two choice words were the two homonyms that had been combined with the presented target word in the priming tests (i.e., the two homonyms of a homonym pair, one related and one unrelated with the target, see the Material section); the prime was thus one of the choice words. Participants could then give in their second response, using the outer left and right keys. Response window was set to 10 s. A total of 64 target word trials from the 120 word–word trials from the priming test were shown in a random sequence, whereby the following variables were balanced: Test half within which the prime-target pair occurred, prime-target relatedness, target dominance and display side of the prime word in the selection pair. 3.1.6. Analysis The main analysis was identical to Experiment 1. A priori power analyses showed that with a sample size of 35 participants, single-sided tests of priming effects of 10 ms (d = 0.3) at single SOAs would yield a power of 70% if alpha was increased to 0.10. P-values below .10 are therefore mentioned separately in the analyses.
1336
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
Responses in the visibility rating of the partial prime visibility test in Experiment 2b were used to classify responses in the forced choice test. Responses in the forced choice test in Experiment 2b were registered as correct or incorrect (i.e., correct if the selected word matched the prime). In order to test whether priming was driven by greater partial prime visibility for related than for unrelated prime-target pairs, a discrimination measure for related vs. unrelated prime-target pairs was computed. The hit rate was defined as the proportion of correct prime word choices in related prime-target trials and the false alarm rate was defined as the proportion of incorrect word choices in unrelated prime-target trials. 3.2. Results 3.2.1. Visibility and accuracy in Experiment 2a Four participants who responded to the prime instead to the target in the visible priming test and eight participants with average high prime discrimination (d0 P .80) in the final test for prime visibility at single SOAs were replaced. Average d0 was not significantly different from zero in any condition, M = 0.05 (SE = 0.08), M = 0.10 (SE = 0.09) and M = 0.09 (SE = 0.06) for the 300, 900 and 1500 ms SOAs, respectively. Prime duration was set to 71.4 ms, 57.1 ms, 42.9 ms and 28.6 ms in 62%, 26%, 10% and 2% of the sessions, respectively (distribution of prime durations was similar at all SOAs). Accuracy was high with 96.5% correct responses in the masked priming test and 96.3% correct responses in the visible priming test. 3.2.2. Visibility and accuracy in Experiment 2b No participant had to be replaced. Prime duration was set to 71.4 ms, 57.1 ms, 42.9 ms and 28.6 ms in 64%, 16%, 8% and 12% of the sessions, respectively. Because the threshold-setting procedure was stricter in this experiment, the distribution of prime durations across subjects was compared with that of the previous experiments. The distribution of prime durations was not significantly different from the one measured in Experiment 2a or in Experiment 1, v2’s < 1.0. Accuracy was high with 94.4% correct responses in the masked priming test and 94.5% correct responses in the visible priming test. Average d0 in the lexical decision prime visibility test was not significantly different from zero, M = 0.09 (SE = 0.07), p = .20. In contrast, average d0 in the partial prime visibility test was significant both when the prime was followed by a subordinate target, M = 0.38 (SE = 0.18), p = .045 and when the prime was followed by a dominant target, M = 0.53 (SE = 0.19), p = .009. 3.2.3. Analyses across the masked and visible priming tests The following priming analyses were done on the data collapsed across Experiments 2a and 2b. 3.2.3.1. Reaction times. The hypothesis that related primes led to different classification times compared to unrelated primes was tested with an omnibus ANOVA on RTs across Prime-Target Relation, Test and SOA (see Fig. 2 and Table 2 for descriptives and Table 5 for inference tests). The main effects of Prime-Target Relation and Test were highly significant, which indicated that prime-target relationship and masking affected RTs. Furthermore, Prime-Target Relation and Test interacted, suggesting that priming was different in the masked compared with the visible priming test. There was also a significant main effect of SOA, which was qualified by Prime-Target Relation and, marginally, by Test. No other interaction reached significance.2 3.2.3.2. Accuracies. In the omnibus ANOVA on arcsine transformed accuracies across Prime-Target Relation, Test and SOA, the main effect of Prime-Target Relation was highly significant, as was the effect of SOA (cf. Tables 3 and 5). There was also a marginal interaction of Prime-Target Relation and Test, suggesting that there was a tendency for priming to be different in the masked compared with the visible priming test. No other main effect or interaction reached significance. 3.2.4. Masked primes 3.2.4.1. Reaction times. The main hypotheses were tested with ANOVAs on average RTs. In the omnibus ANOVA with the within factor Prime-Target Relation (neutral, subordinate, dominant) and the factor SOA (100, 300, 900 and 1500 ms) varied between subjects, the main effects of Prime-Target Relation and SOA were significant. The interaction of SOA and Prime-Target Relation however did not reach significance, indicating that there was little variation in priming across SOAs. Follow-up contrasts between experimental and the neutral condition indicated that dominant priming was significantly positive and subordinate priming was significantly negative (see Fig. 2 and Table 5). Priming for the subordinate meaning was significant at SOAs of 100 ms and 300 ms and priming for the dominant meaning was significant at SOAs of 300 ms and 1500 ms (cf. Table 2). In order to determine whether prime visibility influenced priming (across Experiments 2a and 2b), we (a) analyzed the correlation of subordinate and dominant priming with d0 from the prime visibility task and (b) determined the intercept and slopes of priming vs. d0 using linear regression. The correlation of priming with d0 was neither significant for subordinate priming, r = .15, p = .11, nor for dominant priming, r = .18, p = .058. The scatterplots are given in Fig. 3. Because correlation of dominant priming approached significance, we also estimated the intercept and slope of the d0 -priming relation using linear regression: The intercept of subordinate priming was negative, b1 = 12.3 ms, SE = 2.6 ms, t(113) = 4.80, p < .001 and 2 Because SOAs were varied within participants in Experiment 2a, a first analysis was done using a repeated measures ANOVA across SOA conditions 300 ms, 900 ms and 1500 ms only. Results were comparable to the one reported here, except that the main effect of SOA was not significant in any of the RT and accuracy analyses.
1337
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343 Table 5 ANOVA results of Experiment 2. Analysis/effect
df1
df2
Reaction times
Accuracies
F
MSE
p
F
MSE
p
13.69
0.018
⁄⁄
4.60 2.87
0.019 0.020
1.77 4.10
0.022 0.034
4.42
0.012
14.74 24.50
0.017 0.028
2.71 1.29
0.014 0.017
Omnibus ANOVA Prime-Target Relation Test SOA Prime-Target Relation Test Prime-Target Relation SOA Test SOA 3-way interaction
1.84 1 3 1.81 6 3 6
204 111 111 200 222 111 222
37.08 37.01 10.91 7.03 3.13 2.38 1.28
328 2154 7804 439 301 2154 396
⁄⁄
Masked test Prime-Target Relation Dominant vs. Control Subordinate vs. Control SOA Prime-Target Relation SOA
1.67 1 1 3 6
187 111 111 111 222
29.30 11.05 24.14 11.74 1.28
440 460 740 4600 370
⁄⁄
Visible test Prime-Target Relation Dominant vs. Control Subordinate vs. Control SOA Prime-Target Relation SOA Dominant-Control SOA Subordinate-Control SOA
2 1 1 3 6 1 1
222 111 111 111 222 111 111
9.52 16.52 0.10 8.37 2.98 2.33 3.09
327 305 282 3922 327 305 282
⁄⁄
⁄⁄ ⁄⁄
.002 .006 .074 ns
.001 ⁄⁄ ⁄⁄
ns
⁄⁄
ns ⁄⁄
.008 .08 .030
ns .005 .059 ns ns ns .18 .045 ns .006 ns ⁄⁄ ⁄⁄
ns .049 ns
Remark: Significant p-values are printed in bold. SOA was treated as a between-factor. ⁄⁄p < .001, ns: p > .20. F-values and MSE’s are not listed for F’s smaller than 1.0.
Fig. 3. Scatterplots of subordinate (left) and dominant (right) masked priming in function of lexical decision prime visibility for the data collapsed across Experiments 2a and 2b. Grey lines are regression lines (correlations were not significant).
the slope was not significantly different from zero, b2 = 10.38 ms, SE = 6.4 ms, t(113) = 1.62, p = .11. The intercept of dominant priming was positive, b1 = 6.8 ms, SE = 2.0 ms, t(113) = 3.40, p = .001 and the slope was not significantly different from zero, b2 = 9.7 ms, SE = 5.1 ms, t(113) = 1.92, p = .058.3 This indicates that subordinate priming was negative and dominant priming was positive even when conscious prime perception was completely suppressed. 3.2.4.2. Accuracies. The main effect of Prime-Target Relation in the omnibus ANOVA was not significant but SOA had a significant effect on accuracies. The interaction of SOA and Prime-Target Relation did not reach significance, however. Follow-up contrasts between experimental and neutral conditions indicated that dominant priming was significantly positive and subordinate priming was not significant. Priming was not significant when tested at single SOAs (see Table 3). 3.2.5. Visible primes 3.2.5.1. Reaction times. In the omnibus ANOVA, the main effects of Prime-Target Relation and SOA were significant. More important, the interaction of SOA and Prime-Target Relation was also significant, indicating that priming varied with SOA. Two follow-up ANOVAs on the single experimental and the neutral condition indicated that the variation in dominant 3 It has been suggested that an adapted regression method be used in order to correct for the measurement error of the predictor variable d0 (Klauer, Draine, & Greenwald, 1998). This method leads to slightly greater intercepts when the slope is negative, which is the case with the present data.
1338
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
priming across the whole range of SOAs was small, given that the interaction of SOA and Prime-Target Relation did not reach significance. In contrast, subordinate priming was not significant across the whole range of SOAs but changed with SOA, as indicated by a significant interaction of Prime-Target Relation and SOA (see Fig. 2 and Table 5). When tested at single SOAs, priming for the subordinate meaning was not significant, whereas priming for the dominant meaning was significant at the 900 ms and1500 ms SOAs (see Table 2). 3.2.5.2. Accuracies. The main effects of Prime-Target Relation and SOA were significant. However, the interaction of SOA and Prime-Target Relation was not significant, indicating that variation of priming across SOAs was small. Follow-up contrasts on the single experimental and neutral conditions indicated that dominant priming was significantly positive across the whole range of SOAs, whereas subordinate priming was not significant. Priming for the dominant meaning was significant at the 100 ms, 300 ms and1500 ms SOAs (see Table 3). 3.2.6. Correlation of prime visibility and priming in Experiment 2b In order to determine whether prime visibility influenced priming in Experiment 2b, we analysed the correlation of subordinate and dominant priming with d0 from the two prime visibility tasks (see Fig. 4 for scatterplots). Correlation of RT priming with lexical decision d0 was not significant either for subordinate priming, r = .02, p > .20 or for dominant priming, r = .05, p > .20. Similarly, subordinate priming did not correlate with partial visibility d0 for homonyms when subordinate targets were shown, r = .13, p > .20 and dominant priming did not correlate with partial visibility d0 when dominant targets were shown, r = .26, p > .20. This suggests that in this experiment, prime visibility did not have a measurable influence on masked priming. The correlation of accuracy priming with lexical decision d0 was not significant for subordinate priming, r = .08, p > .20, but significantly negative for dominant priming, r = .49, p = .013. Correlations of accuracy priming with partial visibility d0 were not significant, r’s < .18, p’s > .20. 3.2.7. Partial prime visibility test in Experiment 2b The additional partial prime visibility test in Experiment 2b provided more detailed data on whether and how often participants were able to recognise parts of the primes, and whether the target following the prime facilitated prime perception. In the prime visibility task, all participants declared they had not perceived anything of the prime in some of the 64 trials (average number of trials: M = 40.5, SE = 3.8), while 16 participants reported partial perception of the prime in some trials (M = 22.5, SE = 2.9) and eight participants reported seeing the prime word in some trials (M = 24.6, SE = 4.9). Participants’ subjective reports were in good agreement with the objective measure of discrimination: The accuracy (ac) in selecting the prime word in the partial visibility task was indeed at chance in the ‘‘not seen’’ category, acunrelated = 0.52, acrelated = 0.52,
Fig. 4. Scatterplots of subordinate (left) and dominant (right) masked priming in function of prime visibility measures for Experiment 2b. Lexical decision d0 was assessed with a task asking for lexical decisions with respect to the prime; partial prime visibility d0 was assessed with a task wherein participants had to decide which of two choice words was the prime. Grey lines are regression lines (correlations were not significant).
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1339
d0 = 0.13, SE = 0.12, t(24) = 1.01, p > .20, but better than chance in the ‘‘partially seen’’ category, acunrelated = 0.65, acrelated = 0.71, d0 = 1.5, SE = 0.38, t(15) = 3.91, p = .001 and in the ‘‘seen a word’’ category, acunrelated = 0.78, acrelated = 0.84, d0 = 2.44, SE = 0.66, t(7) = 3.70, p = .008.4 Hence, when participants indicated perceiving part of the prime or the whole prime, they indeed more often selected the correct word in the forced choice task than what would be expected by chance. Prime discrimination was not better for dominant than for subordinate meanings across response categories, Dd0 = 0.15, SE = 0.15, t(24) = 1.0, p > .20 and within response categories, t’s < 1.4, p > .20 for every category. Furthermore, discrimination in the partial prime visibility task was not significantly correlated with discrimination in the lexical decision task, r = .16, p > .20. In summary, when participants directed their full attention to the primes, they were better than chance at identifying which of two words was the prime, whereby there were large differences between participants. Eight participants were able to read whole words in about a third of the trials; across all participants, primes were not perceived at all in two thirds of the trials. It is thus conceivable that partial prime perception occurred in the priming test; however, it appears unlikely that partial prime visibility was the main driver of priming, given that correlations of priming with d0 were nonsignificant (or negative). 3.3. Discussion In Experiment 2, the basic masked priming pattern seen in Experiment 1 was replicated. When lexical decisions in trials with related homonym prime-target word pairs were compared to a control condition, responses were slower for subordinate meanings and faster for dominant meanings. Differences between SOAs did not reach significance, which indicates that this pattern of priming remained fairly stable across SOAs. As in Experiment 1, the pattern suggests some fast selection of the (dominant) word meaning, whereby the subordinate meaning is simultaneously inhibited. In the visible priming test, dominant priming was positive across all SOAs, whereby the priming effect at single SOAs was observed in RTs or errors. In contrast, there was no significant subordinate priming across SOAs, but subordinate priming varied with SOA. Visible subordinate priming was thus weaker in Experiment 2 than in Experiment 1 and did not replicate the qualitative time course found in Experiment 1. One possible explanation for this difference comes from the partial within-variation of SOA: participants in Experiment 2a did three sessions using the identical word set (but different prime-target SOAs), wherein primes and targets were combined using two variants. In the visible test of the first session, participants usually noticed that the words were ambiguous; it is possible that this led to a bias in the second and third sessions which favoured the dominant compared to the subordinate meanings of homonyms. In Experiment 2b, which assessed priming at 100 ms SOA, masked subordinate priming was again negative, but masked dominant priming was only positive in accuracies. In the visible test, the finding that subordinate priming was not significant indicates that RT effects were smaller than in Experiment 1; the finding that dominant priming was significant in errors only supports this contention. The more sensitive threshold-setting procedure of Experiment 2b, compared with Experiments 1 and 2a, did not produce shorter prime presentation times. Prime visibility was very low (mean d0 = 0.09) and nonsignificant when assessed with lexical decision, but higher (mean d0 = 0.5) and significant when assessed with a partial prime visibility test in which participants had to identify the prime out of a list of two words in each trial. Given that it is difficult to focus essentially on the prime in the priming test, it is likely that prime visibility during the priming test was even lower. The results thus indicate that participants probably perceived some letters of the primes in a few trials during test. However, we think that it is unlikely that such perception led to the observed priming effects, because masked priming did not correlate with prime visibility measures. 4. General discussion The characteristic time course of supraliminal homonym priming indicates early parallel access of all meanings of a word, and later selection of one of the meanings. This study endeavoured to measure subliminal priming by polar homonyms in order to gain more insight into the automatic processes of access and selection of word meaning. If selection of meaning were automatic, a difference in priming for different meanings of homonyms should be seen at some point when measuring priming in a range of prime-target SOAs, even when the primes are shown subliminally. Results in the visible prime conditions indicated that meaning selection occurred after 300 ms, which confirmed previous research with visible primes (e.g., Simpson, 1981; Simpson & Burgess, 1985). Conversely, in the subliminal condition, the dominant meaning was facilitated and the subordinate meaning was inhibited not only with respect to the dominant meaning, but also with respect to a neutral condition at prime-target SOAs ranging from 100 ms to 1500 ms. In both experiments, subliminal and dominant priming did not increase or decrease significantly across SOAs. The reliable difference between priming for the dominant meaning and priming for the subordinate meaning indicates that word meanings of masked homonyms were not only accessed, but one of the meanings was selected early on. The results therefore support an automatic early selection account. However, the finding that access to
4 Participants did not report seeing the prime more often when the prime and target were related, compared with when the prime and target were unrelated, as the average difference in the number of trials between the two conditions was not significant, Ndiff = 0.6, SE = 0.4, t(24) = 1.28, p > .20, Ndiff = 0.1, SE = 0.6, t(15) = 0.20, p > .20 and Ndiff = 1.4, SE = 0.7, t(7) = 1.95, p = .09 for the ‘‘not seen’’, ‘‘partially seen’’ and ‘‘seen a word’’ response categories in this order.
1340
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
the subordinate meaning of homonyms was actually inhibited and not just neutral deserves special consideration. In the following we therefore discuss the possible sources of such an inhibition. 4.1. Negative priming and inhibition In both masked priming experiments, priming by subordinate meanings was negative across SOAs, which indicates selective inhibition of the subordinate meaning. In principle, such type of inhibition is present either in automatic selection or in strategic checking. Automatic selective inhibition is found in most models that provide mechanisms of selection, such as inhibition between lexical nodes in spreading activation models (McClelland & Rumelhart, 1981; McClelland & Rumelhart, 1986), inhibition by lexical competition (Norris, 1994) or inhibition between competing attractors for lexical entries (Rodd, Gaskell, & Marslen-Wilson, 2002). There is also one theory of priming suggesting that inhibition is needed to enhance discrimination of weak traces (Dagenbach & Carr, 1994). Although the various models locate inhibition at different levels of representation, the main principle is similar for all of them: Inhibition is a means of counterweighting activation or influence by other, competing concepts during a selection phase. Mutual inhibition of competing concepts will lead to amplification of small differences between the concepts, which ultimately allows for selection of one of the candidates. Therefore, one would expect to find inhibition between competing meanings of a homonym at some point in the selection process. Therefore, inhibition might be a sign of selection or interrupted selection. The positive masked priming of dominant primes found in both experiments might indeed support this thesis. There are, however, other sources of automatic selective inhibition which are due to intertrial inhibition. Simpson and Kang (1994) for instance found evidence for long-term inhibition of unselected meanings at lags of 0, 1, 4 or 12 trials after presentation of a homonym. Because the inhibition was obviously linked to earlier conscious access to one of the meanings, the authors suggested that ‘‘an inhibitory effect is diagnostic of previous activation’’ of a homonym (p. 377). Following this line of reasoning, one could speculate that even at lags of hours, days or weeks, the memory trace associated with a homonym might be preferentially linked with the context associated with its last encounter, and therefore inhibition might be inherent in any word retrieval from memory. Another explanation for inhibition due to inter- and intratrial repetition might be deduced from the negative priming literature. As we mentioned in the discussion of Experiment 1, if the prime is ignored but relevant for a target decision, it can function as a distractor, provoking attentional inhibition or response interferences. Frings and Eder (2009) for instance observed that negative priming by a masked distractor that was repeated as a visible distractor in the probe display produced negative priming of about 10 ms in a probe naming task. This priming effect decayed over time: It was significant at 138 ms and 538 ms SOA, but not at 1038 ms SOA. Although primes were not repeated as targets in our study, the mere fact that primes resembled the targets semantically might be a cause of negative priming, because response to primes must be inhibited. Interestingly, the inhibition observed with subordinate meanings in our study had a (nonsignificant) tendency to decrease across SOAs, similar to the decrease reported in the Frings and Eder study. A data pattern similar to the one reported here was also reported by Wentura and Frings (2005) in four experiments on far and near category concepts, which Wentura and colleagues termed ‘‘dominant’’ and ‘‘nondominant’’ words. When they presented subliminal category words followed by words denominating elements of a category, they observed negative nondominant priming and (less reliably) positive dominant priming. Strategic sources of inhibition might have been an influence in Experiment 2a, because our participants probably realized that primes were homonyms in the visible test, which could have influenced their behaviour in sessions 2 and 3. According to Kahan (2000), participants will use their knowledge on the possible prime-target relationship to guess the prime’s content. Therefore, if participants realize that primes are sometimes related to targets, they might start to expect such a relationship between primes and targets. This expectation will lead to positive priming if perception of the prime is above a certain threshold. However, it might lead to negative priming if perception of the prime is below that threshold, because the weak match between prime and target creates interference in the decision process. In Experiment 2a, stimuli were repeated across sessions. That is, in sessions 2 and 3, participants probably guessed that related prime-target pairs were shown in the masked conditions, too. Therefore, any relationship between prime and target could have led to negative priming if prime visibility was sub-threshold, which was most likely the case in Experiment 2 (our detection rate was much lower than the ‘poor performer’ detection rate in Kahan’s study). However, according to this logic, RTs to dominant and subordinate targets would be expected to be inhibited due to a related prime, which is not exactly what we observed. Nevertheless, this is an interesting explanation which should be followed up. 4.2. Generalizability of results The above suggested interpretation of results rests on the assumption that word identification times were not systematically biased by task demands. In order to exclude systematic biases, care was taken to show equal proportions of trials in all relevant conditions of the reported experiments. Thus, the number of related and unrelated prime-target trials, the number of dominant and subordinate prime-target trials and the number of dominant and subordinate neutral trials were held equal in all parts of the experiments. The rationale for this approach was that unequal proportions of trials would contribute to the formation of biases within an experiment, such that lexical decision RTs for dominant vs. subordinate target words, for instance, would be shorter if dominant words were more frequent. There might be other sources for biases, e.g. biases could result from different salience of groups of word pairs, as for instance homonym-dominant target pairs, that are by definition
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1341
more readily associated than homonym-subordinate target pairs. Furthermore, it is conceivable that partial prime perception had a similar effect in the masked priming tests. Although we have no evidence that our pattern of results could be produced or strengthened by pre-existing or experiment-induced bias for dominant meanings, and by pre-existing or experiment-induced bias for related word pairs, future experiments might aim to directly influence such biases, in order to test the robustness of the observed pattern of results. 4.3. Were primes subliminal or not? There were clear qualitative differences between masked and visible priming. In the masked priming test, we aimed to obtain a measure of prime perception when the prime is not perceived, in order to exclude controlled processes that occur when participants have insight into the aim of the experiment and the type of prime-target relationships occurring during the tests. In order to reduce prime awareness to a minimum, a threshold-setting procedure was used which adapted presentation times to individual perception thresholds for prime words in each session. Were we successful, then, in reducing prime awareness in the masked test? As a general rule, prime visibility tests should use conditions that are as similar as possible to the ones used in the actual priming test. In Experiment 2, participants did a lexical decision prime visibility test in which they decided whether the prime was a word or not. Prime discrimination was not higher than zero and did not correlate with priming. However, the prime words used in this test were not the same as the ones used in the priming test and targets were nonwords (the latter change was made in order to facilitate the prime detection task). Therefore, one could argue that the lexical decision prime visibility test used here did not measure the relevant features of prime visibility that influence priming. As an additional measure of prime visibility, a partial prime visibility test was included in Experiment 2b, which assessed after each priming trial whether participants could identify the prime word in a list of two words, whereby primes and choice words were identical to primes and targets in the masked priming test. This test indicated (a) that partial perception of the primes was occasionally possible and (b) that between-subject differences in prime visibility were large. Some of the subjects, for instance, reported being able to read the prime words in some trials. It is therefore important to note that partial prime perception did not correlate with masked priming, which leads us to two important conclusions: Firstly, partial prime perception was not the main driver of the masked priming effects found here and, secondly, partial prime perception did not produce a pattern resembling the visible priming pattern (i.e., more positive priming for both meanings of homonyms). Given that subjective reports supported the validity and reliability of the partial priming test, it therefore appears that partial prime visibility was not informative enough to produce priming at the level of word meaning.5 From these analyses, we conclude (a) that prime awareness was not totally suppressed in the masked priming tests and (b) that the masked priming results presented are most probably not a consequence of spurious conscious prime perception. It is, however, important to bear in mind that our ‘subliminal’ primes were probably ‘periliminal’. 4.4. Awareness, lexical access and priming The differences between priming by subliminal words and priming by visible words give us insight into processing differences associated with perception with and without awareness. Priming by visible primes indicated that both meanings of homonyms were accessed (at least in Experiment 1), and the dominant meaning was selected between 300 ms and 600 ms after prime onset. Priming by invisible primes indicated that the dominant meaning was more easily accessed and the subordinate meaning less easily accessed than a control, even at short SOAs. The finding that positive priming for multiple meanings is only observed with visible primes, and only at short SOAs, indicates that the stage of meaning selection assumed to occur between 0 and 300 ms depends on awareness of both prime and target words. Indeed, it has been shown that it is difficult to mentally separate two stimuli under such conditions (e.g., Shapiro, Driver, Ward, & Sorensen, 1997; Treisman & Gelade, 1980), indicating that this pre-selection stage also reflects binding of prime and target properties. Accordingly, visible priming at short SOAs might be more sensitive to mental combination of prime and target, rather than being a measure of passive influence by the perception of a prime. In consequence, it is conceivable that the supraliminal homonym priming pattern reflects a post-lexical stage of context integration which is characterised by malleability of prime meaning due to awareness of both the prime and the target. This would imply that subliminal perception is selective and rigid and supraliminal perception is integrative and flexible (for a similar conclusion with respect to different levels of awareness, see Davis et al., 2007). The observed dissociation between subliminal and supraliminal priming might also have even more profound implications for visible priming effects in general. This is because the flexible choice of meaning when both prime and target are visible could have been partly or completely driven by strategic influences: When primes are visible, participants may notice that targets sometimes reflect the dominant and sometimes the subordinate meaning of the prime words, which could induce an inclination to consider the different meanings of the words throughout the test.6 If this were the case, visible priming 5 Strong influences of partial priming on masked ‘‘semantic’’ priming have been reported by Kouider and Dupoux (2004) with colour words. Compared with the Kouider and Dupoux study, the present study used a set of stimuli that was much larger and less easy to predict. Therefore, partial prime visibility was less useful here for second-guessing the prime. 6 We thank an anonymous reviewer for mentioning this possibility.
1342
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
effects assessed with a transparent design such as the one used here would overestimate the ‘real’ priming effect. Such an influence is less likely if primes are presented subliminally (e.g., Van den Bussche & Reynvoet, 2008; Van den Bussche, Segers, & Reynvoet, 2008; see also Bodner & Masson, 2003). In conclusion, we suggest that when the prime is shown subliminally, the initial choice of meaning cannot be revised and the resulting priming is based on within-lexicon interactions of prime and target representations. Because the dominant meaning of a homonym is its most frequently used meaning, the combination of a homonym prime with its subordinate meaning as a target leads to interference within the lexicon. 4.5. Conclusion Compared to other words, homonyms have a more variable meaning, which renders them more dependent on context information. Priming experiments with visible primes indicate that in the first moments during which we become aware of a homonym, we are open to different interpretations of it. The rise and decline of priming, which indicates access to and selection of meanings, is slow. In contrast, our results indicate that when we are not aware of seeing a (subliminal) homonym, we behave selectively. Access to meanings and selection of a meaning is fast and favours the most frequent meaning. Further research is needed to determine whether the observed dissociation between visible and masked priming is due to strategies or, rather, to qualitative differences in perception with and without consciousness. Acknowledgments This manuscript was partly funded by Swiss National Foundation Fellowship No. PA001 – 113106/1 granted to the first author and Swiss National Foundation Grant No. 1114-067180 granted to the third author. We thank Bettina Gerber, Melanie Hegi, Matthias Minder, Ajshe Ponxha and Helen Wyler for conducting parts of Experiment 1, and the University of Bern and the MRC Cognition and Brain Sciences Unit for granting use of infrastructure. Many thanks also to Matt Davis and Rik Henson, who supported the writing process with valuable comments, and to four anonymous reviewers for their helpful comments. Appendix A. Supplementary material Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.concog.2011.05.010. References Binder, K., & Morris, R. K. (1995). Eye movements and lexical ambiguity resolution: Effects of prior encounter and discourse topic. Journal of Experimental Psychology: Learning, Memory and Cognition, 21(5), 1186–1196. Bodner, G. E., & Masson, M. E. J. (2003). Beyond spreading activation: An influence of relatedness proportion on masked semantic priming. Psychonomic Bulletin & Review, 10(3), 645–652. Chen, L., & Boland, J. E. (2008). Dominance and context effects on activation of alternative homophone meanings. Memory & Cognition, 36(7), 1306–1323. COSMAS-I Corpora (2002). Corpus search, management and analysis system (on-line). Mannheim: Institut für deutsche Sprache.
D. Eckstein et al. / Consciousness and Cognition 20 (2011) 1327–1343
1343
Neely, J. H. (1977). Semantic priming and retrieval from lexical memory: Roles of inhibitionless spreading activation and limited-capacity attention. Journal of Experimental Psychology: General, 106(3), 226–254. Neely, J. H. (1991). Semantic priming effects in visual word recognition: A selective review of current findings and theories. In D. Besner & G. Humphreys (Eds.), Basic processes in reading: Visual word recognition (pp. 264–336). Hillsdale, NJ: Erlbaum. Neill, W. T. (1997). Episodical retrieval in negative priming and repetition priming. Journal of Experimental Psychology: Learning, Memory and Cognition, 23, 1291–1305. Norris, D. (1994). Shortlist: A connectionist model of continuous speech recognition. Cognition, 52, 189–234. Onifer, W., & Swinney, D. A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory and Cognition, 9(3), 225–236. Paul, S. T., Kellas, G., Martin, M., & Clark, M. B. (1992). Influence of contextual features on the activation of ambiguous word meanings. Journal of Experimental Psychology: Learning, Memory and Cognition, 18(4), 703–717. Rodd, J. M., Gaskell, M. G., & Marslen-Wilson, W. D. (2002). Making sense of semantic ambiguity: Semantic competition in lexical access. Journal of Memory and Language, 46, 245–266. Schvaneveldt, R. W., Meyer, D. E., & Becker, C. A. (1976). Lexical ambiguity, semantic context, and visual word recognition. Journal of Experimental Psychology: Human Perception and Performance, 2(2), 243–256. Shapiro, K. L., Driver, J., Ward, R., & Sorensen, R. E. (1997). Priming from the attentional blink: A failure to extract visual tokens but not visual types. Psychological Science, 8, 95–100. Simpson, G. B. (1981). Meaning dominance and semantic context in the processing of lexical ambiguity. Journal of Verbal Learning and Verbal Behavior, 20(1), 120–136. Simpson, G. B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Bulletin, 96, 316–340. Simpson, G. B., & Burgess, C. (1985). Activation and selection processes in the recognition of ambiguous words. Journal of Experimental Psychology: Human Perception and Performance, 11, 28–39. Simpson, G. B., & Kang, H. (1994). Inhibitory processes in the recognition of homograph meanings. In D. Dagenbach & T. H. Carr (Eds.), Inhibitory processes in attention, memory and language (pp. 359–381). San Diego, CA: Academic Press. Simpson, G. B., & Krueger, M. A. (1991). Selective access of homograph meanings in sentence context. Journal of Memory and Language, 30, 627–643. Tabossi, P., & Zardon, F. (1993). Processing ambiguous words in context. Journal of Memory and Language, 32, 359–372. Tanenhaus, M. K., Leiman, J. M., & Seidenberg, M. S. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18(4), 427–440. Till, R. E., Mross, E. F., & Kintsch, W. (1988). Time course of priming for associate and inference words in a discourse context. Memory & Cognition, 16(4), 283–298. Tipper, S. P. (2001). Does negative priming reflect inhibitory mechanisms? A review and integration of conflicting views. Quarterly Journal of Experimental Psychology: Human Experimental Psychology, 54A(2), 321–343. Treisman, A. M., & Gelade, G. (1980). A feature-integration theory of attention. Cognitive Psychology, 12, 97–136. Van den Bussche, E., & Reynvoet, B. (2008). The asymmetry between top-down effects and unconscious cognition: Additional issues. Consciousness and Cognition, 17(4), 1361–1363. doi:10.1016/j.concog.2008.05.003. Van den Bussche, E., Segers, G., & Reynvoet, B. (2008). Conscious and unconscious proportion effects in masked priming. Consciousness and Cognition, 17(4), 1345–1358. Van den Bussche, E., Van Den Noortgate, W., & Reynvoet, B. (2009). Mechanisms of masked priming: A meta-analysis. Psychological Bulletin, 135, 452–477. Vu, H., Kellas, G., & Paul, S. T. (1998). Sources of sentence constraint on lexical ambiguity resolution. Memory & Cognition, 26(5), 979–1001. Weber, M. R. (1970). First-graders’ use of grammatical context in reading. In H. Levin & J. P. Williams (Eds.), Basic studies on reading (pp. 147–163). New York: Basic Books. Wentura, D., & Frings, C. (2005). Repeated masked category primes interfere with related exemplars: New evidence for negative semantic priming. Journal of Experimental Psychology: Learning, Memory and Cognition, 31(1), 108–120.