1. art and entertainment
2. visual art and design
3. painting

Experimental
M.©Psychology
Welvaert
2008 Hogrefe
et2008;
al.:&Sup
Vol.
Huber
erset
55(1):54–63
Publishers
Priming
Graded Effects of Number of Inserted
Letters in Superset Priming
Marijke Welvaert1, Fernand Farioli2, and Jonathan Grainger2
1
Ghent University, Belgium, 2CNRS & University of Provence, France
Abstract. Three masked priming experiments investigated the effects of target word length and number of inserted letters on superset
priming, where irrelevant letters are added to targets to form prime stimuli (e.g., tanble–table). Effects of one, two, three, and four-letter
insertions were measured relative to an unrelated prime condition, the identity prime condition, and a condition where the order of letters
of the superset primes was reversed. Superset primes facilitated performance compared with unrelated primes and reversed primes, and
the overall pattern showed a small cost of letter insertion that was independent of target word length and that increased linearly as a
function of the number of inserted letters. A meta-analysis incorporating data from the present study and two other studies investigating
superset priming, showed an average estimated processing cost of 11 ms per letter insertion. Models of letter position coding are examined
in the light of this result.
Keywords: orthographic priming, letter position, visual word recognition
There has been a recent increase in the number of studies
examining low-level orthographic processing in visual
word recognition. This upsurge is mainly because one key
question has been isolated and has served as the focus for
quite intense empirical investigation over the last five
years. The question is how information concerning the positions of a word’s component letters is stored in memory
and processed during word recognition. This question has
been investigated over the years using a number of different
paradigms involving manipulations of letter order with real
word anagrams (e.g., bale–able: Chambers, 1979), nonword anagrams (e.g., salior: Andrews, 1996), and illusory
word identifications caused by letter migration (Mozer,
1983; Davis & Bowers, 2004). However, the most recent
investigations have turned to the masked priming paradigm
as a primary tool for investigating early automatic perceptual processing of printed strings of letters (see Forster,
1998, for a presentation of the merits of masked priming).
Relative-position priming is one empirical result obtained with masked priming that has played a key role in
attracting attention to the issue of letter position coding. In
experiments investigating relative-position priming, the
position of letters shared by prime and targets is manipulated, typically by removing some of the target’s letters to
form the prime stimulus (e.g., grdn–garden). In a seminal
study, Humphreys, Evett, and Quinlan (1990) found evidence for such relative-position priming in participants’
percent correct word identification in a masked priming
paradigm in which both primes and targets were briefly
presented and pattern-masked. Peressotti and Grainger
(1999) replicated and extended these results using Forster
and Davis’ (1984) variant of masked priming with the lexical decision task. One important result in the Peressotti and
Experimental Psychology 2008; Vol. 55(1):54–63
DOI 10.1027/1618-3169.55.1.54
Grainger study is that the recognition of 6-letter target
words such as the French word “balcon” was facilitated by
a 4-letter prime formed by concatenating a subset of the
target’s letters (e.g., blcn), but not by the same subset of
letters in a different order (e.g., bcln, nlcb). Grainger, Granier, Farioli, Van Assche, and van Heuven (2006) have provided a further exploration of relative-position priming effects, examining the role of letter contiguity and phonological overlap between prime and target. Another key result
reported by Peressotti and Grainger (1999) and Grainger et
al. (2006) is that priming effects are not affected by inserting hyphens, such that letters shared by prime and target
occupy the same absolute, length-dependent position in
each string (e.g., “g-rd-n” as a prime for “garden”). Priming
effects in this condition are statistically equivalent to the
priming effects obtained from concatenated primes (e.g.,
grdn–garden).
These relative-position priming effects allowed Grainger and van Heuven (2003) to reject all the methods of
letter position coding implemented in the classic models of
visual word recognition (Coltheart, Rastle, Perry, Langdon,
& Ziegler, 2001; McClelland & Rumelhart, 1981; Seidenberg & McClelland, 1989). All of these coding schemes
predicted little or no priming from concatenated subset
primes. Grainger and van Heuven (2003) then went on to
describe a number of promising new accounts of letter position coding. In one of these new approaches, traditional
context-sensitive coding schemes (e.g., the wickelgraph
scheme used in the model of Seidenberg and McClelland,
1989) were given more flexibility by introducing coding
for noncontiguous letter sequences. This type of coding
was first used in the work of Mozer (1987) and more recently in the work of Whitney (2001) and Grainger and van
© 2008 Hogrefe & Huber Publishers
M. Welvaert et al.: Superset Priming
55
Table 1. Summary of the priming conditions tested in the present study. Numbers refer to the position of letters in the
target stimulus that are maintained in the prime stimulus; the letter “d” refers to a letter that is not present in the
target stimulus
Experiment 1
Insert 1
Insert 2
Insert 3
Unrelated
5-letter targets
12d345
123d45
12dd345
123dd45
12ddd345
123ddd45
ddddddd
7-letter targets
123d4567
1234d567
123dd4567
1234dd567
123ddd4567
1234ddd567
ddddddddd
Experiment 2
Identity
Insert 1
Insert 2
Unrelated
7-letter targets
1234567
123d4567
1234d567
123dd4567
1234dd567
ddddddddd
9-letter targets
123456789
1234d56789
12345d6789
1234dd56789
12345dd6789
dddddddddd
Experiment 3
Insert 2
Insert 3
Insert 4
Unrelated
7-letter targets
123dd4567
1234dd567
123ddd4567
1234ddd567
123dddd4567
1234dddd567
dddddddddd
Identity
Reverse 2
Reverse 3
Reverse 4
7-letter targets
1234567
165dd4327
1654dd327
165ddd4327
1654ddd327
165dddd4327
1654dddd327
Heuven (2003). Grainger and van Heuven coined the term
“open-bigram” to refer to ordered sequences of adjacent
and nonadjacent letters. For example, the word CART
would be coded as the following set of bigrams: CA, CR,
CT, AR, AT, and RT. Grainger and van Heuven’s (2003)
open-bigram model and Whitney’s (2001) SERIOL model
differ mainly in terms of the mechanism used to activate
open-bigrams. Furthermore, Grainger et al. (2006) drew a
distinction between an unconstrained open-bigram model
that codes for all possible sets of ordered letter pairs without consideration of the distance separating the letters, and
a more plausible version of open-bigram coding (referred
to as the overlap open-bigram model) with graded bigram
activations, as in the SERIOL model. In this version of
open-bigram coding, the activation level of open-bigram
units is a function of the distance between their constituent
letters (see Dehaene, Cohen, Sigman, & Vinckier, 2005, for
a proposed neural implementation of this type of coding,
and Grainger et al., 2006, for a more detailed discussion of
these different models). Finally, Davis’ (1999) SOLAR
model uses an activation gradient to code for relative letter
position. In this model, the orthographic input layer includes letter units that are position-independent and context-independent. The relative order of the letters in a string
is coded by the relative activity of the letter nodes. The
activation input for each word detector is calculated by
evaluating the match between two spatial codes: (1) the
spatial code corresponding to the word represented by this
detector, and (2) the spatial orthographic code corresponding to the input stimulus. The similarity of these two spatial
codes is a function of the number of shared letters and the
extent to which the order of the shared letters is respected
across the two strings.
All of these new accounts of letter position coding not
only captured another recently reported phenomenon,
transposed-letter priming (Perea & Lupker, 2004; Schoon© 2008 Hogrefe & Huber Publishers
Table 2. Match values for the overlap open-bigram model
(OOB), the SERIOL model, and the SOLAR model for the different priming conditions tested in Experiments 1–3, for each target word length
Condition
OOB
SERIOL
SOLAR
Insert 1
0.80
0.83
0.92
Insert 2
0.68
0.72
0.74
Insert 3
0.66
0.67
0.62
Insert 1
0.87
0.89
0.92
Insert 2
0.79
0.81
0.73
Insert 3
0.78
0.77
0.60
Insert 4
0.78
0.77
0.57
Reverse 2
0.11
0.06
0.36
Reverse 3
0.10
0.06
0.34
Reverse 4
0.10
0.06
0.33
Insert 1
0.90
0.91
0.92
Insert 2
0.85
0.85
0.72
5-letter targets
7-letter targets
9-letter targets
baert & Grainger, 2004), but they also predicted the existence of another form of relative-position priming called
superset priming. Van Assche and Grainger (2006) provided the first direct test of the predicted effects of superset
primes by inserting unrelated letters in target words to create superset primes (e.g., gafrdlen–garden). These authors
found that 1-letter and 2-letter insertions generated practically as much priming as identity primes. Three-letter insertions, on the other hand produced significantly less
priming than identity primes, but still facilitated target
word recognition compared with unrelated primes. Van
Assche and Grainger’s finding of robust effects of superset
Experimental Psychology 2008; Vol. 55(1):54–63
56
M. Welvaert et al.: Superset Priming
primes confirms the predictions of the model of letter position coding presented above. However, none of these
models predicted the precise pattern observed by Van Assche and Grainger. All models, except for a completely unconstrained open-bigram model with no bottom-up inhibition (i.e., from letters to words), predicted graded effects of
letter insertion, with lower levels of priming as the number
of inserted letters increases.
Van Assche and Grainger (2006) suggested that the pattern of superset priming effects they observed might be due
to the limitations of the human visual system when processing long prime stimuli. The relatively small superset priming effect found with 3-letter insertions could be due to a
limitation in the number of letters that can be processed in
parallel from a masked prime stimulus. This account generates the clear prediction that shorter target words should
show stronger superset priming. The present study was designed to test this prediction, and to provide a further test
of the different accounts of letter position coding described
above. Table 1 provides a summary of the priming conditions tested in the present study. Table 2 summarizes the
predictions of different letter position coding schemes.
Experiment 1
Method
nonwords 1.8 (0–9), 7-letter words 1.1 (0–5), 7-letter nonwords 0.2 (0–3). These 256 items formed the targets. Each
target was associated with four different prime stimuli:
three superset prime conditions: (1) a superset prime
formed by inserting one letter in the corresponding target
(insert 1 condition, e.g., chafrbon as a prime for charbon),
(2) a superset prime in which two letters were inserted (insert 2 condition, e.g., chafmrbon as a prime for charbon),
and (3) a superset prime in which three letters were inserted
(insert 3 condition, e.g., chafmvrbon as a prime for charbon), and an unrelated prime condition in which the prime
had no letters in common with the target (e.g., stieulmaf as
a prime for charbon). The unrelated primes were the same
length as primes in the insert 2 condition. In order to simplify the design there were only two possible positions of
insertion for each type of related prime condition. The letters were inserted one position before or after the central
letter of the target (7-letter targets: insert 1 – 123d4567 and
1234d567, insert 2 – 123dd4567 and 1234dd567, insert 3
– 123ddd4567 and 1234ddd567; 5-letter targets: insert 1 –
12d345 and 123d45, insert 2 – 12dd345 and 123dd45, insert 3 – 12ddd345 and 123ddd45). All inserted letters were
grouped. The word targets and corresponding prime stimuli
are given in Appendix A1. Target Length and Prime Type
were within-participant factors in a 2 × 4 factorial design.
Four lists of material were constructed so that each target
appeared once in each list, but each time in a different priming condition. Different groups of participants were tested
with each list.
Participants
Thirty-six students at the University of Provence, Marseille, voluntarily took part in this experiment. They all reported being native speakers of French with normal or corrected-to-normal vision.
Stimuli and Design
Sixty-four 7-letter words and sixty-four 5-letter words in
French were selected as critical targets in a masked priming
lexical decision experiment. The mean printed frequency
of the 7-letter words was 25 per million (range 1–302). For
the 5-letter words the mean frequency was 23 per million
(range 3–133; New, Pallier, Ferrand, & Matos, 2001). The
words were nouns, adjectives, or verbs in infinitive form.
One hundred-and-twenty-eight pronounceable, orthographically regular nonwords were selected, half of which
were seven letters long and the other half five letters long.
The nonwords were created by changing one or two letters
in a real French word while respecting the orthotactic and
phonotactic constraints of French. The mean and range (in
parentheses) of the number of orthographic neighbors (N)
for these stimuli were: 5-letter words 4.5 (0–11), 5-letter
1
Procedure
Each trial consisted of four stimuli presented one after the
other at the center of a CRT display. The first was a row of
twelve hash marks (############), which served as a forward mask, and was presented for 500 ms together with two
vertical lines positioned above and below the center of the
mask and serving as a fixation mark. Second, the prime was
displayed on the screen for 50 ms and was followed immediately by a backward mask (identical to the forward mask)
for 16 ms. The target was presented immediately after the
backward mask and remained on the screen until participant’s response or for a maximum duration of 4000 ms. The
intertrial interval was 666 ms. Presentation of the visual
stimuli and recording of the RTs were controlled by DMDX
software Version 3.0 (Forster & Forster, 2003) on a desktop
computer. All stimuli were presented in Arial lowercase
font, as white characters (luminance = 12 cd/m2) on a black
background. Primes and targets had different sizes in order
to minimize physical overlap in the identity prime condition, Arial 16 for primes and Arial 12 for targets. Mask
stimuli were presented in the same font size as the primes.
Participants were instructed to focus on the center of the
Appendices can be downloaded from http://www.up.univ-mrs.fr/wlpc/grainger – click on “Publications” in the menu bar and then on
“ExpPsych-appendices.”
Experimental Psychology 2008; Vol. 55(1):54–63
© 2008 Hogrefe & Huber Publishers
M. Welvaert et al.: Superset Priming
row of hash marks when they appeared (indicated by the
two vertical lines), and to decide whether the following
string of letters that remained on the screen was a French
word or not. They were requested to make this decision as
quickly and as accurately as possible. The presence of a
prime was not mentioned. They responded “Yes” by pressing the right response button and “No” by pressing the left
response button of a Logitech Wingman Precision GamePad. Before the experimental trials there were 20 practice
trials where the prime consisted of a blank screen and the
targets fulfilled the same criteria as the experimental stimuli.
Results
Mean response times and percentage errors are presented
in Table 3. Errors and RTs that were smaller than 200 ms
or larger than 1200 ms (0.9% of the data for word targets)
were excluded from the latency analysis.
Table 3. Mean RTs (in ms) and percentage of errors for
word and nonword targets in Experiment 1; standard errors of the mean (SE) are given in parentheses
57
trend was the linear trend, F1(1, 35) = 12.19, p < .001,
F2(1, 126) = 27.51, p < .001. This was also the case for the
three related prime conditions considered alone, F1(1, 35)
= 5.96 p < .02, F2(1, 126) = 12.19, p < .001. Thus, increasing the number of unrelated letters in prime stimuli causes
a significant increase in RT to word targets, and the variance of the linear contrast accounted for 98.5% of the overall variance across participants and 98.3% across items for
the insert 1, 2, and 3 conditions. The ANOVA on error percentages revealed a significant effect of Length by participant, F1(1, 35) = 5.58, p < .03, F2(1, 126) = 1.64. The effect
of Prime Type was not significant, but there was a significant interaction between Prime Type and Length,
F1(3, 105) = 3.55, p < .02, F2(3, 378) = 3.79, p < .02, reflecting the greater effects of Prime Type in 7-letter words.
Nonword Analyses
A 4 (Prime Type) × 2 (Length) ANOVA on the latencies
showed both a significant main effect of Length, F1(1, 35)
= 28.81, p < .0001, F2(1, 126) = 5.81, p < .02, and Prime
Type, F1(3, 105) = 3.90, p < .02, F2(3, 378) = 3.67, p < .02,
but no interaction. The ANOVA on the error percentages
revealed a significant main effect of Length, F1(1, 35) =
24.10, p < .0001, F2(1, 126) = 11.95, p < .0001.
Type of prime
Insert 1
5-letter
words
Insert 2
Insert 3
Unrelated
577
(11.8)
584
(11.6)
597
(10.4)
5.7
(1.0)
3.4
(1.0)
2.4
(0.7)
Mean RT 551
(SE)
(13.1)
566
(13.1)
572
(11.7)
598
(12.3)
%Errors
(SE)
2.0
(0.7)
2.4
(0.9)
3.2
(0.8)
676
(16.2)
681
(15.3)
691
(16.8)
4.7
(1.1)
5.1
(1.0)
6.1
(1.2)
654
(16.6)
661
(16.1)
667
(15.4)
1.7
(0.6)
2.2
0(.6)
3.0
(0.6)
Mean RT 569
(SE)
(10.8)
%Errors
(SE)
7-letter
words
2.5
(0.7)
1.7
(0.6)
5-letter
Mean RT 667
nonwords (SE)
(17.8)
%Errors
(SE)
6.6
(1.4)
7-letter
Mean RT 643
nonwords (SE)
(16.2)
%Errors
(SE)
2.4
(0.6)
Word Analyses
A 4 (Prime Type) × 2 (Length) ANOVA on mean correct
RTs showed a significant main effect of Prime Type,
F1(3, 105) = 12.50, p < .001, F2(3, 378) = 16.40, p < .001.
The main effect of Length and the interaction were not significant (all F values < 1). Pairwise comparisons using
Dunnett’s test showed that the related prime conditions
were all significantly faster than the unrelated condition by
participants and by items (p < .001). Polynomial contrasts
showed that across all prime conditions the highest order
© 2008 Hogrefe & Huber Publishers
Discussion
Experiment 1 shows robust superset priming in 5-letter and
7-letter words that varied as a function of the number of
inserted letters but not as a function of target word length.
We therefore successfully replicated the superset priming
effects reported by Van Assche and Grainger (2006). However, contrary to the predictions of one account of Van Assche and Grainger’s (2006) superset priming results, we
found no evidence that superset priming was greater in
shorter words. Thus, the effect of number of inserted letters
found by Van Assche and Grainger cannot be accounted for
by a limit in the number of prime letters that can be processed simultaneously from a masked prime stimulus, since
we expected to observe stronger superset priming with
shorter target words. This was definitely not the case in
Experiment 1, since the 7-letter words actually showed numerically larger priming effects than the 5-letter words.
Furthermore, the results of Experiment 1 show more
graded effects of number of inserted letters compared with
the pattern reported by Van Assche and Grainger. The different patterns could be due to the fact that number of inserted letters was manipulated across experiments in Van
Assche and Grainger’s study, whereas a within-participant
design was used here. Experiment 2 was designed as a further test of superset priming across words of different
length, this time with 7-letter and 9-letter targets, and 1-letter and 2-letter insertions. An identity priming condition
was included in this experiment in order to test whether the
Experimental Psychology 2008; Vol. 55(1):54–63
58
M. Welvaert et al.: Superset Priming
presence of this condition might influence the amount of
priming generated by superset primes (note that there was
an identity prime condition in all of Van Assche & Grainger’s experiments).
Procedure
The procedure was the same as in Experiment 1.
Results
Experiment 2
Method
Mean response times and percentage of errors are presented
in Table 4. Incorrect responses and RTs smaller than 200 ms
or larger than 1200 ms (1.1% of the data for word targets)
were excluded from the latency analysis.
Participants
Thirty-six students at the University of Provence, Marseille, participated voluntarily in the experiment. All of
them reported being native speakers of French and had normal or corrected-to-normal vision. None had participated
in the previous experiment.
Table 4. Mean RTs (in ms) and percentage of errors for
word and nonword targets in Experiment 2; standard errors of the mean (SE) are given in parentheses
Type of prime
Insert 1
7-letter
words
Stimuli and Design
A new set of stimuli was created (see Appendix B, Footnote
1). Sixty-four French 7-letter words and sixty-four French
9-letter words were selected as critical targets. The two lists
were matched for mean printed frequency: for 7-letter
words, 38 per million (range 11–145) and for 9-letter
words, 37 per million (range 10–197; New et al., 2001).
The words were nouns, adjectives, or verbs in infinitive
form. An equal number of pronounceable, orthographically
regular nonwords were selected half of which were 7 letters
long and the other half 9 letters long. The mean and range
(in parentheses) of the number of orthographic neighbors
(N) for these stimuli were: 7-letter words 1.2 (0–8), 7-letter
nonwords 0.1 (0–3), 9-letter words 0.6 (0–3), 9-letter nonwords 0. These 256 items formed the targets. For all these
targets, two superset primes were constructed: (1) a superset prime formed by inserting one different letter before or
after the central letter of the target (insert 1 condition, e.g.,
réudnion or réundion as a prime for réunion), and (2) a
superset prime in which two different letters were inserted
before or after the central letter (insert 2 condition, e.g.,
réudsnion or réundsion as a prime for réunion). In addition
to the superset prime conditions, there was an unrelated
prime condition in which the prime had no letters in common with the target (e.g., dydalèel – reunion) and an identity prime condition in which prime and target were identical (e.g., reunion – reunion). The unrelated primes were
the same length as the insert 1 primes. Target Length and
Prime Type were within-participant factors in a 2 × 4 factorial design. Four lists of material were constructed so that
each target appeared once in each list, but each time in a
different priming condition. Different groups of participants were tested with each list.
Experimental Psychology 2008; Vol. 55(1):54–63
9-letter
words
Insert 2
Insert 3
Unrelated
Mean RT 516
(SE)
(16.4)
534
(17.9)
548
(17.3)
564
(16.7)
%Errors
(SE)
1.2
(0.6)
1.6
(0.5)
2.5
(0.8)
Mean RT 529
(SE)
(18.4)
529
(16.3)
545
(17.8)
560
(15.6)
%Errors
(SE)
2.3
(0.9)
2.3
(0.7)
2.7
(1.0)
635
(21.2)
640
(24.6)
646
(25.0)
4.3
(1.0)
4.9
(1.1)
6.6
(1.7)
685
(26.4)
669
(28.5)
700
(31.0)
10.7
(2.3)
9.6
(2.0)
9.4
(2.0)
1.0
(0.4)
1.8
(0.5)
7-letter
Mean RT 645
nonwords (SE)
(24.2)
%Errors
(SE)
5.9
(1.2)
9-letter
Mean RT 684
nonwords (SE)
(24.6)
%Errors
(SE)
9.8
(2.0)
Word Analyses
A 4 (Prime Type) × 2 (Length) ANOVA on mean correct
RTs revealed a main effect of Prime Type, F1(3, 105) =
16.37, p < .0001, F2(3, 378) = 12.56, p < .0001. The effect
of Length and the interaction were not significant. Pairwise
comparisons with Dunnett’s test showed that RTs in the
unrelated condition were significantly larger by participant
and by item than in the identity condition (p < .0001), the
insert 1 condition (p < .01), and the insert 2 condition (p <
.01). There was a significant linear trend across all prime
conditions, F1(1, 35) = 35.60, p < .0001, F2(1, 126) = 52.65,
p < .0001, and this was the highest order trend. The linear
trend was also highly robust when the unrelated prime condition is removed, F1(1, 35) = 16.10, p < .001, F2(1, 126)
= 11.39, p < .001, with the linear component accounting
for 97.7% of variance across participants and 99.7% of the
variance across items. Thus RT is almost perfectly predicted by the number of inserted letters (0, 1, or 2). The
© 2008 Hogrefe & Huber Publishers
M. Welvaert et al.: Superset Priming
ANOVA on the error percentages showed no significant
effects.
59
ported being native speakers of French and had normal or
corrected-to-normal vision. None had participated in the
previous experiments.
Nonword Analyses
Stimuli and Design
A 4 (Prime Type) × 2 (Length) ANOVA on mean correct
RTs yielded a significant main effect of Length, F1(1, 31)
= 40.94, p < .0001, F2(1, 126) = 31.82, p < .0001, and the
effect of Prime Type was significant by participants,
F1(3, 93) = 3.23, p < .02. The interaction was not significant
(F1 < 1, F2 < 1). The ANOVA on error percentages also
revealed a significant main effect of Length, F1(1, 31) =
15.60, p < .001, F2(1, 126) = 13.41, p < .001. The effects
of Prime Type and the interaction were not significant.
Discussion
The results of Experiment 2 once again show graded effects
of number of inserted letters on superset priming, with RTs
gradually increasing when going from the 0-letter insert
condition (i.e., identity priming) to 1-letter and 2-letter insertions. The within-participant manipulation of number of
inserted letters in Experiments 1 and 2 of the present study
therefore suggests that effects of letter insertion in superset
primes are more of a graded nature, with a small processing
cost associated with each additional letter added. This graded effect is particularly clear in Experiment 2 where the
linear component accounts for practically all of the variance across the 0-letter, 1-letter, and 2-letter insert conditions.
Experiment 3
Experiment 3 provides a further investigation of superset
priming while examining the importance of the order of
letters in superset primes. Peressotti and Grainger (1999)
and Grainger et al. (2006) have shown that subset priming
effects disappear when the order of the letters is disrupted.
Thus, while “slene” primes “silence” the prime “snele”
does not, compared with an unrelated prime condition. In
Experiment 3 we manipulated the order of the letters that
are shared by prime and target in superset primes (see Table
1). Experiment 3 also provides a further investigation of
the upper limit of superset priming by testing prime stimuli
with four inserted letters.
Method
Participants
Forty students at the University of Provence, Marseille,
participated voluntarily in the experiment. All of them re© 2008 Hogrefe & Huber Publishers
Targets were the 7-letter words and nonwords tested in Experiment 1. Each target word was tested in 8 priming conditions: Three superset prime conditions with 2, 3, and 4
inserted letters which were inserted before or after the central letter of the target word (Insert 2, Insert 3, Insert 4), and
three reversed prime conditions (Reverse 2, Reverse 3, Reverse 4) in which the related inner letters (all letters shared
by prime and target except for the first and last letter) of
the matched insert prime were reversed in order. For example, the target word “courage” was primed by “coubpkage”
in the insert 3 condition, and by “cgabpkuoe” in the reverse
3 condition. In addition there was an identity prime condition where primes were the same as targets, and an unrelated prime condition, where primes were composed of ten
letters that were not present in the target. The different
priming conditions are summarized in Table 1, and the
complete list of stimuli shown in Appendix C (Footnote 1).
The eight priming conditions were separated into two subgroups formed of the four insert conditions plus the unrelated prime condition, and the four reverse conditions plus
the identity condition. Four lists of stimuli were constructed
so that each target appeared twice in each list, in one of the
priming conditions of each subgroup. Each list was tested
with different participants. All other factors were manipulated within participants.
Procedure
The procedure was the same as in Experiments 1 and 2.
Results
Mean response times and percentage of errors are presented
in Table 5. Incorrect responses and RTs smaller than 200 ms
or larger than 1200 ms (1.1% of the data for word targets)
were excluded from the latency analysis.
Word Analyses
First of all, a 3 (Number of Letters) × 2 (Letter Order)
ANOVA on mean correct RTs was performed in order to
directly compare insert primes with the reverse primes. The
effect of letter order was significant, F1(1, 39) = 4.62, p <
.05, F2(1, 63) = 31.59, p < .00001. Superset primes with
letters in the correct order generated faster RTs than superset primes in which the order of shared letters was disrupted. There was also a significant interaction between NumExperimental Psychology 2008; Vol. 55(1):54–63
60
M. Welvaert et al.: Superset Priming
Table 5. Mean RTs (in ms) and percentage of errors for the 7-letter word and nonword targets in Experiment 3; standard
errors of the mean (SE) are given in parentheses
Words
Nonwords
Identity
Insert 2
Insert 3
Insert 4
Reverse 2
Reverse 3
Reverse 4
Unrelated
Mean RT
554
571
579
587
600
608
594
602
(SE)
(12.2)
(11.9)
(11.9)
(12.2)
(10.3)
(12.5)
(11.4)
(11.6)
%errors
1.3
2.0
2.3
2.3
3.0
3.4
1.7
1.9
(SE)
(0.4)
(0.6)
(0.7)
(0.6)
(0.8)
(0.7)
(0.6)
(0.8)
Mean RT
668
678
672
668
683
675
681
689
(SE)
(15.2)
(13.8)
(14.3)
(13.4)
(16.1)
(14.0)
(14.2)
(14.2)
%errors
4.5
3.3
3.3
2.7
2.8
3.3
3.6
3.4
(SE)
(1.2)
(0.7)
(0.7)
(0.7)
(1.1)
(0.7)
(0.8)
(0.7)
ber of Letters and Letter Order, F1(2, 78) = 4.47, p < .05,
F2(2, 126) = 3.23, p < .05. As can be seen in Table 5, the
difference between insert primes and reverse primes is
greatly reduced in the case of 4-letter insertions. As in the
previous experiments, there is a strong linear effect of number of inserted letters in the superset prime condition
(which is not significant by-items), F1(1, 39) = 5.47, p <
.025, F2(1, 63) = 1.38, p = .24, accounting for 99.9% of
variance across these three conditions in the means by participant. Again, the linear trend was the highest order trend
in these data. Next, we performed pairwise comparisons
using Dunnett’s test comparing the insert and reversed
primes against the unrelated condition and the identity condition. All prime conditions produced longer RTs than the
identity prime condition (all p values < .05) except for the
insert 2 condition (nonsignificant in the by-participant
analysis). The insert 2 and insert 3 primes produced significantly faster RTs than the unrelated prime condition (all p
values < .01), but neither the insert 4 nor any of the reverse
prime conditions differed significantly from this baseline.
There were no main effects or interaction in the analysis of
errors percentages.
Nonword Analyses
There were no main effects or interaction in the analyses
of the RTs and percentage errors to nonword targets.
Discussion
The results of Experiment 3 show significant superset
priming effects when measured relative to the effects of
primes containing exactly the same letters but in a different order. The position of the first and last letter was maintained in these reverse primes, but the order of the inner
letters shared with the target was reversed. These results
therefore confirm and extend the findings of Peressotti
and Grainger (1999) and Grainger et al. (2006) showing
the importance of letter order in relative-position priming.
They further demonstrate that having the target word’s external letters (i.e., the first and last letter) in the correct
position is not sufficient to generate significant priming
relative to primes having no letters in common with targets.
Figure 1. Linear regression between
number of inserted letters and priming
effect size (in ms) calculated relative
to the unrelated prime condition in the
meta-analysis.
Experimental Psychology 2008; Vol. 55(1):54–63
© 2008 Hogrefe & Huber Publishers
M. Welvaert et al.: Superset Priming
Meta-Analysis of Superset Priming
Given the observed variability in the size of superset priming effects across experiments within the present study and
across different studies, we decided to combine the data
from all experiments that have examined superset priming
in our laboratory. This meta-analysis therefore includes the
data from Experiments 1–3 of the present study, Experiments 1, 2, and 4 from Van Assche and Grainger (2006),
and Experiment 2 reported in Welvaert, Granier, Farioli,
and Grainger (2006). We chose to only analyze the data
concerning 7-letter words, since this word length was tested in all of these experiments. This meta-analysis therefore
involves data from 7 experiments testing 248 participants.
The results of the meta-analysis are clear-cut. Figure 1
shows the linear regression between number of inserted letters (0, 1, 2, 3) and priming effect size calculated relative
to the unrelated condition in each experiment. The regression is significant (p < .0001) with 62% of the variance in
priming effect sizes explained by the number of inserted
letters. The identity condition (0 inserted letters) generates
an average effect of 55 ms, and there is a cost of 11 ms per
letter inserted. Figure 1 also shows how, given the variability across different groups of participants, different patterns
can emerge from one experiment to another. What is important from this meta-analysis is that the overall pattern is
consistent with the findings of the present study. Furthermore, the cross-over on the Y-axis (55 ms) fits well with
the fact that these experiments all used a 50 ms prime duration, and identity priming effects are typically in the
range of 40–60 ms for that prime duration. Finally, we did
not include the 4-letter insertion condition of Experiment
3 in the meta-analysis, since there was only one data point
for this condition. The regression equation obtained in this
analysis predicts an 11 ms priming effect with 4-letter insertions. The effect reported in Experiment 3 was 14 ms.
General Discussion
The results of the present study provide a replication of the
superset priming effects recently reported by Van Assche
and Grainger (2006). Experiments 1–3 show that superset
priming effects are not greater in shorter words, thus contradicting one account of Van Assche and Grainger’s results. This account attributed the relatively small size of
superset priming effects with 3-letter insertions (compared
with 1-letter and 2-letter insertions) in Van Assche and
Grainger’s study, as being due to a limitation in the number
of letters that can be processed in parallel from briefly presented masked prime stimuli. Shorter target words were
therefore expected to generate superset priming effects in
conditions where longer words did not show an effect. Such
a pattern was not found in the present study.
The use of a within-participant manipulation of number
© 2008 Hogrefe & Huber Publishers
61
of inserted letters in the present study has revealed a rather
graded influence of this factor on the size of superset priming effects. This is good news for the models of letter position coding put to test in the present study, all of which
predicted such graded effects (see Table 2). The difference
with respect to the pattern reported by Van Assche and
Grainger (2006) could simply reflect a problem in measurement sensitivity for such fine-grained manipulations. Even
within the present study the size of superset priming effects
fluctuated somewhat, probably reflecting the limits in the
precision of the particular paradigm we employed. In order
to overcome these limits, at least to some extent, a metaanalysis was performed on the data of three different studies. The results of this meta-analysis show a highly robust
significant linear relation between number of inserted letters and the magnitude of superset priming effects (see Figure 1).
It is important to remember that these superset priming
effects once again highlight the shortcomings of letter position coding schemes implemented in many models of visual word recognition (Coltheart et al., 2001; McClelland &
Rumelhart, 1981; Seidenberg & McClelland, 1989). According to all of these coding schemes, inserting irrelevant
letters in prime stimuli should be much more damaging
than is actually observed. In these coding schemes, the insertion of irrelevant letters disrupts the coding of letter position information such that a prime like “silpmence” has
very little orthographic overlap with the target word “silence.” In line with the evidence from subset priming
(Grainger et al., 2006; Humphreys et al., 1990; Peresotti &
Grainger, 1999) and transposed-letter priming (Perea &
Lupker, 2004; Schoonbaert & Grainger, 2004), superset
priming further demonstrates the flexible nature of orthographic coding.
Evaluation of the Models
In Table 2 we presented the predictions of some prominent
recent accounts of letter position coding in visual word recognition with respect to the priming conditions tested in the
present study. All of these predicted the graded influence
of number of inserted letters on superset priming effects
that was found in the present study and confirmed in the
meta-analysis. We now examine whether the overall pattern of priming effects found in the meta-analysis allows
one particular model to emerge as notably superior to the
others in predicting effects of superset priming. To do so,
the match values of the models were correlated with the
average effect sizes (measured against the unrelated prime
condition) for the identity prime condition, and 1-letter, 2letter, and 3-letter insertions from the meta-analysis (N =
4). Larger match values predict greater priming effects. The
SERIOL model, the SOLAR model, and the overlap openbigram model (OOB) all provide excellent fits to the pattern found in the meta-analysis (OOB: r = .97, p < .05;
SERIOL: r = .99, p < .05; SOLAR: r = .95, p < .05).
Experimental Psychology 2008; Vol. 55(1):54–63
62
M. Welvaert et al.: Superset Priming
All of these models predicted the graded effects of letter
insertion that we found in the present study and that were
also present in the meta-analysis. However, as can be seen
Table 2, the models differ in terms of the way priming effects vary as a function of target length and whether or not
order information is respected (the “reverse” primes of Experiment 3). If one now calculates the correlation between
priming effect sizes and model match values for all 14 conditions tested in the present study (i.e., the different superset priming conditions and the identity priming condition
for different word lengths, plus the three reverse priming
conditions, see Table 1 and Table 2), then the SOLAR model emerges as the superior approach (r = .93, p < .001, for
the SOLAR model; r = .83, p < .001, for the SERIOL and
OOB models). This result therefore contrasts with the recent analysis of Grainger et al. (2006) who found that the
OOB model provided better fits than the SOLAR model
with respect to relative-position priming effects obtained
with subset primes (i.e., primes formed by removing some
of the target’s letters, e.g., slne–silence). The problem is
that all three models do relatively well in capturing these
two types of relative-position priming (superset and subset
priming), and the differences in performance across the
models is relatively minor. Furthermore, the match values
generated by these models are not expected to perfectly
capture masked orthographic priming effects in all their details. For example, none of the match calculations include
a term via which unrelated letters could influence the similarity between prime and target strings. We now turn to
examine the possible role of such unrelated letters in orthographic priming.
Effects of Unrelated Letters
Van Assche and Grainger (2006) discussed a simple account of superset priming whereby the target word representation is maximally activated by a prime containing all
of the target’s component letters in the correct order, and
independently of the presence of irrelevant letters. Furthermore, given the nonsignificant difference between the identity prime condition and the 1-letter and 2-letter insertion
conditions in their study, Van Assche and Grainger argued
that irrelevant letters in the prime do not inhibit target processing. However, in the light of the present results and the
meta-analysis performed across three different studies, it
appears that this simple model would best account for the
data when combined with an inhibitory influence of unrelated letters in the prime. Each unrelated letter in the prime
stimulus would generate a constant inhibition, and increasing the number of unrelated letters would increase the total
inhibitory effect on target word processing. The results of
the meta-analysis suggest that the inhibitory cost per letter
is 11 ms.
It is trivial to show that this “letter inhibition” model can
generate predictions for average priming effect sizes in the
three superset priming conditions examined in the metaExperimental Psychology 2008; Vol. 55(1):54–63
analysis, by subtracting the inhibitory constant (11 ms)
from the size of identity priming effects (55 ms). This subtraction results in a significant correlation between predicted effects and mean priming effects found in the meta-analysis (r = .99, p < .05). The success of the letter inhibition
model follows from the strong linear relation found between number of inserted letters and priming effect size
found in the meta-analysis. In an arguably stronger test of
this simple model, the same analysis can be applied using
the priming effect sizes found in the different superset
priming conditions tested at each word length in the present
study (N = 9). Again the identity priming condition serves
as the baseline for estimating superset priming with an
11 ms cost per inserted letter. This correlation is also significant (r = .75, p < .01).
The reasonable success of the simple letter inhibition
model suggests that feedforward inhibition from letters or
letter clusters to whole-word orthographic representations,
may be all that is necessary in order to account for superset
priming effects. However, this model requires a mechanism
for coding letter order independently of the number of intervening letters. An unconstrained open-bigram model, as
described by Grainger et al. (2006) does exactly this. Unfortunately for this approach, Grainger et al. also showed
that an unconstrained open-bigram model did not provide
a satisfactory account of subset priming effects, compared
with the other models described above. Since there are no
unrelated letters in subset primes, the simple model would
therefore require a further mechanism to enable it to provide a satisfactory account of both subset and superset
priming effects. An alternative, and arguably more promising avenue for developing a more complete account of
relative-position priming effects, would be to augment a
model such as the overlap open-bigram model with a mechanism for bottom-up inhibition. This can easily be done be
inserting the open-bigram coding scheme within an interactive-activation model of visual word recognition (Grainger & Jacobs, 1996; McClelland & Rumelhart, 1981), as
proposed by Grainger and van Heuven (2003).
Conclusions
The present study has provided further evidence for a novel
form of orthographic priming obtained with briefly presented prime stimuli. Superset primes are formed by inserting irrelevant letters in the target stimulus in order to generate an orthographically related nonword prime (e.g., tanble-table). In these priming conditions the letters shared by
prime and target occupy different absolute (length-dependent) positions. Superset priming therefore likely reflects a
level of orthographic processing where some form of position invariance has been achieved, such that the relative
positions of letters and not the absolute, length-dependent
position is retained. The models of letter position coding
investigated in the present study describe such relative-po© 2008 Hogrefe & Huber Publishers
M. Welvaert et al.: Superset Priming
sition coding. These models predicted a graded effect of
number of inserted letters on the size of superset priming
effects. The present study found clear evidence in support
of this prediction.
Acknowledgments
This work was performed while the first author was visiting
the Laboratoire de Psychologie Cognitive on an Erasmus
exchange program with Ghent University. We thank André
Vandierendonck for his continuing support in arranging
such visits. We also thank Colin Davis, Walter van Heuven,
and Carol Whitney for providing us with the match scores
of the SOLAR model, open-bigram model, and SERIOL
model respectively, and Sachiko Kinoshita and William
Owen for constructive comments on an earlier version of
this work.
References
Andrews, S. (1996). Lexical retrieval and selection processes: Effects of transposed-letter confusability. Journal of Memory
and Language, 35, 775–800.
Chambers, S.M. (1979). Letter and order information in lexical
access. Journal of Verbal Learning and Verbal Behavior, 18,
225–241.
Coltheart, M., Rastle, K., Perry, C., & Ziegler, J. (2001). DRC: A
dual route cascaded model of visual word recognition and
reading aloud. Psychological Review, 108, 204–256.
Davis, C.J. (1999). The self-organising lexical acquisition and
recognition (SOLAR) model of visual word recognition. Unpublished doctoral dissertation, University of New South
Wales, Australia.
Davis, C.J., & Bowers, J.S. (2004). What do letter migration errors reveal about letter position coding in visual word recognition? Journal of Experimental Psychology: Human Perception and Performance, 30, 923–941.
Dehaene, S., Cohen, L., Sigman, M., & Vinckier, F. (2005). The
neural code for written words: A proposal. Trends in Cognitive
Sciences, 9, 335–341.
Dunnett, C.W. (1955). A multiple comparison procedure for comparing several treatments with a control. Journal of the American Statistical Association, 50, 1096–1121.
Forster, K.I. (1998). The pros and cons of masked priming. Journal of Psycholinguistic Research, 27, 203–233.
Forster, K.I., & Davis, C. (1984). Repetition priming and frequency attenuation in lexical access. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10, 680–698.
Forster, K.I., & Forster, J. (2003). DMDX: A Windows display
program with millisecond accuracy. Behavioral Research
Methods: Instruments and Computers, 35, 116–124.
Grainger, J., Granier, J.P., Farioli, F., Van Assche, E., & van Heuven, W. (2006). Letter position information and printed word
perception: The relative-position priming constraint. Journal
© 2008 Hogrefe & Huber Publishers
63
of Experimental Psychology: Human Perception and Performance, 32, 865–884.
Grainger, J., & Jacobs, A.M. (1996). Orthographic processing in
visual word recognition: A multiple read-out model. Psychological Review, 103, 518–565.
Grainger, J., & Van Heuven, W. (2003). Modeling letter position
coding in printed word perception. In P. Bonin (Ed.), The mental lexicon (pp. 1–24). New York: Nova Science.
Humphreys G.W., Evett, L.J., & Quinlan, P.T. (1990). Orthographic processing in visual word identification. Cognitive
Psychology, 22, 517–560.
McClelland, J.L., & Rumelhart, D.E. (1981). An interactive activation model of context effects in letter perception: Part 1. An
account of basic findings. Psychological Review, 88, 375–407.
Mozer, M.C. (1983). Letter migration in word perception. Journal
of Experimental Psychology: Human perception and performance, 9, 913–916.
Mozer, M.C. (1987). Early parallel processing in reading: A connectionist approach. In M. Coltheart (Ed.), Attention and performance XII (pp. 83–104). Hillsdale, NJ: Erlbaum.
New B., Pallier C., Ferrand L., & Matos R. (2001). Une base de
données lexicales du français contemporain sur internet: LEXIQUE [A lexical database of contemporary French on the
internet: LEXIQUE]. L’Année Psychologique, 101, 447–462.
Perea, M., & Lupker, S.J. (2004). Can caniso activate casino?
Transposed-letter similarity effects with nonadjacent letter positions. Journal of Memory and Language, 51, 231–246.
Peressotti, F., & Grainger, J. (1999). The role of letter identity and
letter position in orthographic priming. Perception and Psychophysics, 61, 691–706.
Schoonbaert, S., & Grainger, J. (2004). Letter position coding in
printed word perception: Effects of repeated and transposed
letters. Language and Cognitive Processes, 19, 333–367.
Seidenberg, M.S., & McClelland, J.L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96, 523–568.
Van Assche, E., & Grainger, J. (2006). A study of relative-position
priming with superset primes. Journal of Experimental Psychology: Learning, Memory, and Cognition, 32, 399–415.
Welvaert, M., Granier, J.P., Farioli, F., & Grainger, J. (2006). Effects of removing, inserting, and substituting letters in masked
form priming. Manuscript in preparation.
Whitney, C. (2001). How the brain codes the order of letters in a
printed word: The SERIOL model and selective literature review. Psychonomic Bulletin and Review, 8, 221–243.
Received October 31, 2006
Revision received December 4, 2006
Accepted December 6, 2006
Jonathan Grainger
Laboratoire de Psychologie Cognitive
Université de Provence
3 place Victor Hugo
F-13331 Marseille cedex 1
France
E-mail [email protected]
Experimental Psychology 2008; Vol. 55(1):54–63