RESEARCH ARTICLE
A VISUAL DISORDER PRODUCING HIGHLY SELECTIVE DELETION
OF RECURRING LETTERS
Rodger A. Weddell
(Neurosciences Directorate, Morriston Hospital, Swansea, UK, and Department of Psychology,
University of Wales, Swansea, UK)

ABSTRACT
A case of unusually selective deletion of recurring letters is described. The disorder was primarily visual, spelling
being unaffected. There was no evidence of neglect. Reading was fluent and rapid (not letter-by-letter), though words often
appeared as a series of resolving fragments. Selective deletions were entirely confined to adjacent letters, excluding
explanation in terms of repetition blindness (Kaniwisher, 1991). Mixing case and font, or increasing inter-letter distance,
did not alter rate of recurring letter deletion, provided the task required some form of perceptual grouping. The final 2
experiments establish the pre-lexical nature of the deficit; they show deletion of recurring symbols, for example, when
copying Navon-style global forms constructed from same or different Greek letters. This visual deficit is explained using
the neural network model developed by Rolls and Deco (2002).
Key words: vision disorders, alexia, agnosia, attention

INTRODUCTION
This report concerns a woman (TR) with an
unusually selective tendency to delete recurring
letters and/or same-letter gaps when reading.
Miceli et al. (1995) previously described selective
deletion of one doubled letter (e.g., bloccare →
blocare), but only spelling was affected, implying a
lexical deficit “... at the level of the graphemic
buffer ...”. TR’s deletions, which were even more
selective, only affected reading; consequently, it is
important to consider the possibility that her
difficulty reflected a recognised form of peripheral
dyslexia.
In neglect dyslexia, when reading a paragraph,
the subject might omit words on the neglected side
(most commonly the left side). They might omit
(e.g., spine → pine or ine) and/or substitute (e.g.,
spine → brine or apine) and/or add (e.g., spine →
shrine) letters on the neglected side (Young et al.,
1991). Patients with attentional dyslexia read single
letters better than letters flanked by other letters
(Shallice and Warrington, 1977). One model
suggests reading individual letters in a string
requires sequential restriction of an attentional
“spotlight” to one letter at a time: inadequate
narrowing of this attentional window causes
intrusion errors, which (a) impair identification of a
central target letter flanked by other letters, and (b)
produces within- and/or between-string letter
migrations (Mayall and Humphreys, 2000; Saffran
and Coslett, 1996). Saffran and Coslett (1996)
described an attentional dyslexic, NY, who
frequently deleted doubled letters (e.g., hoop →
hop). However, NY, and other attentional dyslexics
(Hall et al., 2001) frequently generated other errors
Cortex, (2005) 41, 471-485

(deletions, migrations, additions, and substitutions
of non-recurring letters).
In letter-by-letter reading, subjects identify one
letter at a time and latencies increase monotonically
with word length. Letter-by-letter reading has been
attributed to pre-lexical visual difficulties, including
impaired activation of individual letters (Behrmann
and Shallice, 1995), and a general deficit in the
rapid visual identification of multiple symbols
or pictures (Farah and Wallace, 1991). Moreover,
for Farah (1990), ventral simultanagnosia reflects
left occipito-temporal damage, which limits the
number of objects that can be seen at any one time,
yielding letter-by-letter reading. In contrast, in
dorsal simultanagnosia, occipito-parietal damage
restricts spatial attention and the spatial extent of
contour that can be explicitly resolved at any one
time. However, Humphreys and Price (1994)
attributed simultanagnosia to poor feature
discrimination in two cases, and argued that Farah’s
dichotomy does not explain all forms of
simultanagnosia.
Though TR might be called simultanagnosic as
objects and words appeared as a series of
perceptual fragments, she read long words rapidly.
Her error profile did not fit neglect or attentional
dyslexia, and her selective same-symbol and gap
deletions extended to non-lexical tasks.
Accordingly, TR’s difficulties are attributed to
impaired ventral stream grouping and restriction of
the number of form elements supportable in
explicit visual percepts. The primarily visual nature
of her deficit permitted explanation in terms of
primate ventral stream neurophysiology and a
recent biologically plausible model of visual object
perception (Rolls and Deco, 2002).

472

Rodger A. Weddell

Fig. 1 – Early CT scan, inset showing right hippocampalparahippocampal infarct.

CASE REPORT
On 10.1.93, TR, aged 29 and predominantly
right-handed, developed left-sided weakness and
sensory impairment. These deficits resolved quickly
leaving
slight
left-sided
weakness
and
somatosensory loss. A CT scan performed on
11.1.93 revealed a small right hippocampalparahippocampal hypodensity due to a stroke
(Figure 1). Visual fields were full to confrontation
(acuity, J1 in both eyes). Neuropsychological testing
on 16.6.93 found no evidence of visual neglect.
However, she had topographical disorientation,
marked prosopagnosia, and visual object agnosia for
real and drawn objects. Semantic memory was
markedly impaired; for example, she did not know
the meaning of the word lion. She named 1/5
colours, but correctly matched colours (5/5). She
made surface dyslexic and surface dysgraphic
errors. Her fluent speech lacked obvious dysphasic
errors. Autobiographical memory loss was
extensive. A CT scan on 20.9.93 found slight
prominence of the temporal horn of the right lateral
ventricle. Her initial marked story and design recall
deficits were relatively mild by 1994. WAIS-R
Verbal IQ was 61 and 81, while Performance IQ
was 68 and 87 in June 1993 and January 1994,
respectively. Corresponding National Adult Reading
Test (Nelson and Willison, 1991) scores increased
from 4/50 to 32/50, the final score yielding a
premorbid Full Scale IQ-estimate of 108. Repeat
testing in August 1997 produced essentially the

same results. The Birmingham Object Recognition
Battery (Riddoch and Humphreys, 1993)
demonstrated late recovery of object recognition
between 1995 and 1997: Association Matching
(measuring semantic memory) scores increased
from 20/30 to 28/30, corresponding Object Decision
scores being 70/128 and 109/128 (both initial scores
were impaired but final scores were not). However,
figure-ground formation, as measured by the test of
overlapping letters, geometric shapes, and objects,
was similarly impaired at both assessments. For
example, the ratio of naming latencies for adjacent:
overlapping pairs of objects was above normal
limits in the first (1:2) and second (1:1.7)
assessments. She remained prosopagnosic
throughout. All tests described subsequently were
administered between 1995 and 1999.
TR persistently merged letters on either side of
inter-word spaces if the same letter ended one
word and began the next, when reading (e.g., were
each → wereach) but not when spelling to
dictation. Similarly, TR correctly copied thirty 4- to
10-letter words lacking doubled letters (e.g.,
importance), incorrectly copying 15/30 words
containing doubled letters solely through doubled
letter deletions (e.g., commission → comission):
she correctly orally spelled all words to dictation.
Preliminary
studies
also
investigated
perceptions of briefly-exposed single and multiple
letters and letter strings. After presentation of a
central fixation square, 1 of 16 black on white
lower case letters subtending 0.9° appeared 2° or
4° from fixation for 17, 51, 84, 118, or 152 msec
on a computer monitor (VDU) in north (N), NE, E,
SE, S, SW, W, or NW positions. TR identified
significantly fewer letters (147/160) than one
uninjured control subject (158/160) but not the
other (151/160).
On disappearance of a fixation cross, 1.1° black
lower case consonants, centred 3.3° to the right or
left of fixation, were exposed against white for 17,
34, 68, 101, 135, or 168 msec on a VDU alone (N
= 120), or flanked by letters – one at 4.6° or two at
2° and 4.6° from fixation. Trigrams (N = 120)
comprised different letters (e.g., czq). Digrams were
different-letter (N = 120; e.g., wn) or same-letter (N
= 120; e.g., yy). Fully reported randomised stimuli
appeared in counterbalanced monocular blocks.
Excluding same-letter pairs, TR correctly identified
letters 3.3° from fixation when these appeared
singly (95%), in digrams (95%), or in trigrams
(90.8%). Side of presentation (χ2df = 1 = 0.42, NS),
exposure duration (χ25 = 2.83, NS), and the number
of letters presented (χ21 = 2.32, NS) did not affect
performance. However, she was less accurate than 2
uninjured control subjects (single, 100% and
98.3%; digram, 99.2% and 95.8%; trigram, 97.5%
and 98.3%). Scoring report of all letters as correct
or incorrect produced identical results. Thus, TR
processed different letters slightly less well than
controls. However, full trigram report did not

Selective recurring symbol deletions

require prolonged exposures and performance did
not decline sharply with (a) increasing letter
number (c.f., some simultanagnosic letter-by-letter
readers Behrmann and Shallice, 1995; Levine and
Calvanio, 1978) or (b) letter flanking (c.f.,
attentional dyslexia, Shallice and Warrington,
1977). Importantly, TR identified only 18.3% of
same-letter digrams (69.2% deletions, e.g., ww →
w; 12.5% visual, e.g., ww → wv) – considerably
fewer than controls (97.5% and 89.2%). Side did
not affect TR’s performance (χ21 = 0.89, NS),
exposures longer than 101 msec being error-free.
FURTHER EXPERIMENTAL INVESTIGATION
EXPERIMENT 1
Effects of Lexical Status and Spatial Distance
on Deletion of Recurring Letters
and/or the Intervening Gap
TR said written text was often incomprehensible
because she merged words if the same letter ended
the first and began the second (e.g., the edge →
thedge). This report distinguishes two types of
adjacent recurring letters. Repeated letters are same
identity letters straddling an inter-string gap (e.g.,
the edge, letters underlined to illustrate). Doubled
letters recur within a string (e.g., beer).
The present study asks if TR selectively deletes
doubled (e.g., noon → non) as well as repeated
letters and/or the separating gap (e.g., letter + gap,
reco ord → record and/or gap only, reco ord →
recoord). It also asks if letter and/or gap deletion
rates vary with (a) lexical factors and (b) gap size
and/or type.
Method
There were nine stimulus sets, each containing
20 items. Sets 1-5 included an inter-string gap
bounded by the same-letter (Sets 1-2) or differentletters (Sets 3-5). Set 1 gap deletion plus or minus
letter deletion generated a nonword (e.g., ston ner →
stoner). Set 2 gap and letter deletion converted 2
nonwords into a word (e.g., sug gar → sugar). Set 3,
gap deletion converted two words into a word with
at least one phoneme change (e.g., fat her → father).
Set 3 stimuli were reordered in Set 4 – merging
produced a nonword (e.g., her fat → herfat). Set 5
gap deletion yielded a word from a nonword and
word (e.g., dra wing → drawing). Set 6-9 items
contained a doubled letter, with deletions altering
phonemic structure: Set 6, nonword → nonword
(e.g., noom → nom); Set 7, nonword → word (e.g.,
doog → dog); Set 8, word → nonword (e.g., door →
dor); and Set 9, word → word (e.g., steep → step).
Ten 10-lines, 45-item paragraphs were
constructed from 450 items presented in random
order, half of Set 1-9 items appearing in three
paragraphs, the remainder in two. One further Set

473

1-5 item also straddled each line break, making
550 items and adding one “line” to each paragraph.
The following section from a paragraph starts
and ends with Set 1-5 items straddling line break
(alternate items italicised to illustrate). The items,
in order, are from Sets 5 (flow er), 2, 8, 7, 1, 9, 1,
2, 3, 4, 6, and 2:
… flow
er ques stion summer toon ase erp latter cag
ger div vide not ice go bin fopper insta
ance …
A further 10 treble-spaced paragraphs were
constructed by converting each single between- and
within-item space into a 3-character space. Line
breaks were doubled. Typographical errors
excluded a few stimuli, these errors being
duplicated in the treble-spaced paragraphs. TR read
4 blocks of 5 paragraphs, blocks 1 and 4 being
single-spaced. The experimenter (E) rated TR’s
tape-recorded responses to identify clear interstring pauses. He also identified phoneme changes,
signalling all Set 6-9 deletions (e.g., noom → nom,
doog → dog) and some (e.g., fam mous →
famous), but not all (e.g., seg gree → segree or
seggree) Set 1-5 letter deletions.
In all experiments, statistical tests were twotailed, while nonwords were orthographically legal
and phonologically plausible. Unless otherwise
stated, lower black case 12-point Times New
Roman or 10-point New Courier letters were used.
Results
Set 1 and 2 inter-string gaps were almost always
deleted in both gap size (single vs. treble spacing)
conditions (Table I). Table I shows they were also
usually deleted in both gap type conditions (i.e.,
character spaces vs. line breaks represent gaps). To
ascertain whether TR deleted gaps and letters or only
gaps, Table I considers instances where the
phonemic structure of the stimulus was changed only
if a letter was deleted as well as the gap (e.g., seco
ond → second). Apparently, she usually deleted
letters and gaps, as the rate of phonemic change in
Set 1 and 2 remained high in the type (character
space vs. line break) and size (single vs. treble
spaced) spacing conditions. As spacing factors did
not alter deletion rate, subsequent analyses pool data
from the type and size spacing conditions.
Deletions of gaps (Fisher Exact, p < 0.007) and
repeated letters (signalled by phonemic change,
Fisher Exact, p < 0.001) were slightly more
frequent when they produced words (Set 2) rather
than nonwords (Set 1). Table I pools results from
Set 3-5 different-letter stimuli, which were virtually
all pronounced as two strings. Set 1 and 2 vs. Set
3-5 contrasts were highly significant for gap
deletion (χ21 = 635.96, p < 0.001) and phonemic
change (χ21 = 488.30, p < 0.001).

474

Rodger A. Weddell
TABLE I

Effect of Spacing Conditions (size and type) on Deletions of a Gap and/or Letter from Repeated or Doubled Letter Stimuli
Single Spacing

Treble Spacing

Gap
Deletion

Phonemic
Change

Gap
Deletion

Phonemic
Change

Set 2 (sug gar → sugar)

Same line
Line ends
Same line
Line ends

41/46
18/19
47/47
19/19

14/18
6/7
31/31
13/13

45/46
19/19
47/47
19/19

15/18
6/7
31/31
13/13

Different Letter Gap
Set 3 (fat her → father);
Set 4 (her fat → herfat);
Set 5 (dra wing → drawing)

Same line
Line ends

0/145
0/58

1/145
0/58

1/145
0/58

1/145
0/58

Same Letter Gap
Set 1 (ston ner → stoner)

Table II shows Set 6-9 letters were infrequently
deleted (signalled by phonemic change). Indeed,
letter deletions were considerably more frequent
in Set 1 and 2 repeated letter strings than in
Set 6-9 doubled letter stimuli (χ21 = 409.42,
p < 0.001), suggesting that the gap separating
recurring letters might facilitate deletion.
Nevertheless, Set influenced double letter deletion
rate (Kruskal-Wallis, p < 0.003), deletions being
most common for nonword to word conversions
(Set 7), and least frequent for the opposite
conversion (Set 8).
TR rapidly read single-spaced (M = 72.4 sec,
SD = 6.7) and treble-spaced (M = 70.7 sec, SD =
4.3) paragraphs, taking an average of 0.8 sec to
read each (3- to 9-letter) string. In contrast to
frequent recurring letter deletions, she made only 8
other errors (e.g., hod → hot, liveer → livee, and
know → now).
Conclusions
1. This experiment confirmed TR’s observation
that she selectively deleted same-letter inter-string
gaps. The high proportion of Set 1 and 2 phonemic
change suggested that she often also deleted one of
the same-identity letters bounding the gap.
2. Gap size (1 or 3 spaces) did not influence
deletion rate. More importantly, same-letter
deletions occurred when one letter ended a line and
the other began the next line: consequently, TR’s
word/nonword groupings respected letter sequence,
but not other topographical features.
3. Deletions were more frequent for nonword
→ word conversions than for word → nonword
conversion, suggesting a slight lexical influence on
deletion rate.
TABLE II

Effect of Lexical Status on Doubled Letter Deletion Rate
Phonemic Change
Set 6 (noom → nom)
Set 7 (doog → dog)
Set 8 (door → dor)
Set 9 (steep → step)

3/100
10/94
0/100
4/100

4. She fluently read all strings. That is, she did
not read letter-by-letter. There was no evidence of
neglect or attentional dyslexia: she read all items,
and she produced very few of the addition,
substitution, or migration errors associated with
these conditions (Hall et al., 2001; Young et al.,
1991).
EXPERIMENT 2
Repetition Blindness cannot Explain Nonword
Letter and Gap Deletion
Repetition blindness might explain recurring
letter deletion. Thus, in sentences such as A brown
dog came in in the house, the repeated in is
missed. However, repetition blindness cannot
explain selective same-letter deletion of inter-string
gaps. Moreover, Kanwisher (1991) states that “...
there is no repetition blindness for letters that are
shared by two different words in a list (e.g., the t
in fault and heart).” In contrast, TR deleted
repeated letters (e.g., div vide → divide) more often
than doubled letters (e.g., spood → spod) in
Experiment 1.
Nevertheless, the differing deletion rates for
doubled and repeated letters in Experiment 1 may
have arisen because string lengths were not closely
matched in the different stimulus Sets.
Consequently, Experiment 2 asks if, as the
repetition blindness hypothesis predicts, the
repeated letter deletion rate is higher in singlestring stimuli such as burudde than in two-string
stimuli comprising the same letters with a gap
inserted between the doubled letters (e.g., burud
de). Moreover, Bavelier and Potter (1992) showed
that repeated letters were deleted more often when
separated by 1 than by 2 symbols: i.e., reducing lag
increases repeating letter deletions. Therefore, this
experiment asks if repeating vowel are sometimes
deleted (e.g., burude → burde or sopone → sopne).
Reading and oral spelling responses were
compared. Reading aloud affords an immediate
window into TR’s visual percepts, but there is
some latitude in the use of English phoneme-

Selective recurring symbol deletions

grapheme correspondences. Subsequently, TR was
asked to orally spell each target, as oral spellings
are unambiguous. However, they introduce a delay
between perception and production. To reduce the
perception-production delay, TR’s oral spellings
were preceded by one probe question asking her if
1 or 2 graphemes represented the target consonant
or vowel phoneme.
Method
Set 1 stimuli were 16 nonword pairs. A
repeating vowel separating 3 consonants made the
first word. The final consonant followed by e made
the second (e.g., sunul le and sopon ne). Removal
of the gap generated Set 2 items (e.g., sunulle and
soponne). Loss of a Set 2 doubled letter yielded
Set 3 (e.g., sunule and sopone). For Set 4, the first
Set 3 vowel was deleted (e.g., snule and spone).
Stimuli were presented in sixteen lines, each
containing one target from each set in
counterbalanced order. Targets alternated with filler
words in each line, e.g., (Set 3, 4, 1, and 2 targets
italicised to illustrate):
edge sunule bat clage tiger balat te lime
burudde dagger.
TR read individually presented lines, the line
being occluded immediately after she read the
target of that trial. One question was then posed.
Question 1 followed all targets and sought the
number of final consonant letters: e.g., “How many
b’s in corobbe?”. Question 2a determined recurring
vowel number in Set 1 and 2 items: e.g., “How
many o’s in corobbe?”. Question 2b solely aimed
to distribute attention evenly across target letters; it
asked about the number of letters representing Set
3 and 4 second consonants: e.g., “How many r’s in
crobe?”. Finally, she orally spelt each target. All 16
lines were presented 4 times in 2 sessions. Each of
the 64 items was the target once per session,
Question 1 being asked in one session and
Question 2 (a or b) in the other, question order
being random and counterbalanced across sessions.
Two Speech Therapists, blind to the stimulus,
independently transcribed TR’s tape-recorded

475

responses to the target of each trial. They then
agreed the most likely spelling of each response
guided by their phonetic transcriptions and a
written stimulus list, supplied with the misleading
suggestion that it may contain distracters.
Results
Raters independently transcribed all Set 1
utterances as a single nonword. Their phonetic
transcriptions mostly concurred for vowel number
(127/128) and identity (109/128) and for final
consonant (126/128) identity. Table III presents
data from raters’ subsequently agreed spellings
(NB: they did not agree 3 final consonants and 1
vowel number). The number of final consonant
letters varied with Set (χ23 = 29.04, p < 0.001)
because 2 letters most often represented the final
consonants of all but Set 4 items. Data from the
agreed spellings diverge from the repetition
blindness hypothesis in that Set 1 repeated and Set
2 doubled letters were equally often deleted.
Moreover, as indicated above, repetition blindness
cannot explain why gaps were selectively deleted
when reading Set 1 stimuli.
Probe responses agreed with subsequent oral
spellings in all but 2 occasions: e.g., one probe
response showed TR initially saw both t’s in balat
te, while her subsequent oral spelling was balate.
Table III presents oral spelling data. No Set 1 oral
spelling contained a gap. One letter represented the
final consonant of most Set 1, 3, and 4 (but not Set
2) items (χ23 = 76.87, p < 0.001). Against the
repetition blindness hypothesis, Set 2 doubled
letters were deleted less often than Set 1 repeated
letters, the gap between the letters actually
facilitating deletion (Table III). Moreover, in
contrast to frequent repeated letter deletions, no Set
1-3 repeated vowels were deleted (Table III).
TR’s spellings and those agreed by the raters
concurred for (a) Set 1 inter-string gap deletion,
and (b) vowel number. However, the rater spellings
doubled Set 3 but not Set 4 final consonant letters.
Presumably, TR applied different graphemephoneme correspondence rules when reading Set 3
and 4 items, the alternative explanation being
improbable (i.e., she actually saw 2 final consonant

TABLE III

Incidence of Inter-string Gaps, Number of Vowels and Number of Final Consonants in Raters’ Agreed Spellings and in TR’s Oral Spellings
of Set 1-4 Stimuli

Raters’ Agreed Spelling
Gap (No. Absent: Present)
No. Final Consonants (1:2)
No. Vowels (1:2)
TR’s Oral Spelling
Gap (No. Absent: Present)
No. Final Consonants (1:2)
No. Vowels (1:2)

Set 1
(e.g., balat te)

Set 2
(e.g., balatte)

Set 3
(e.g., balate)

Set 4
(e.g., blate)

32:0
6:26
0:32

32:0
8:24
2:30

32:0
12:19
0:32

32:0
24:6
27:4

32:0
31:1
0:32

32:0
8:24
0:32

32:0
32:0
0:32

32:0
30:2
31:1

476

Rodger A. Weddell

letters in Set 3 but not Set 4 items). Thus, TR’s
visual perceptions of Set 3 and 4 items are
probably more accurately represented by her oral
spellings than by rater agreed spellings. These
considerations imply that her oral spellings also
correctly indicate that she usually deleted 1 final
consonant letter from Set 1 items. Finally, again in
contrast to her frequent recurring letter deletions,
TR produced only 6 other oral spelling errors: 2
phonological errors (e.g., plide → plied), 2 single
letter duplications (e.g., stome → stomme), 1
addition, and 1 substitution.
Conclusions
1. TR’s visual percepts were more faithfully
described by oral spelling than by phonetic
transcriptions. Moreover, though the delay between
perception and response is shorter for reading than
for oral spelling, the neural representation of the
final consonant only occasionally transformed over
the delay. Consequently, subsequent experiments
rely on copying.
2. The present results contradict 2 predictions of
the repetition blindness hypothesis. First, withinstring Set 2 doubled letters should have been
deleted more frequently than between-string Set 1
repeated letters. Instead, inter-string gaps facilitated
repeated letter deletions. Second, repeated vowels
were not deleted in oral spellings.
3. As in Experiment 1, TR did not read letter by
letter. There was no evidence of neglect. Recurring
letter deletion errors predominated in her oral
spellings, which yielded little indication of the
migration or other errors associated with attentional
dyslexia.
EXPERIMENT 3
Mixing Font and Case does not alter Deletion Rate
Experiment 3 asks if TR only deletes physically
identical shapes including letters. Alternatively, she
may also delete abstract letter representations.
More specifically, abstract letter identities would be
deleted if TR deleted repeated letters from
different-case and/or different font pairs.
Alternatively, if she only deletes from same-case
same-font pairs, then physical identity would be
critical. In the latter case, deletion rate may decline
as feature differences increase: i.e., deletion may be
more common when same-letter features are
similar (e.g., rr) than when they differ (e.g., gG).
Method
The 4- to 8-letter stimuli were generated from
12 words and 12 nonwords. Both sets included 2
examples with the letters l, n, e, d, g, and r
doubled. A 2-character gap was next introduced

between doubled letters (e.g., din ner and stig ged).
Letter and gap deletion generated another word
from most (10/12) of the original words (e.g., spe
ed → sped), but not from nonwords. These 24
items were printed in 4 case-font combinations. All
4 combinations included lower case Times New
Roman (Font1) letters. Font2 (mostly Braggadocio)
was selected to maximise shape differences, font
size varying to ensure that all letters were roughly
the same height. Combinations were: same case
and font, all lower case Font1 letters; different-case
same-font, some upper case Font1 letters; samecase different-font, some lower case Font2 letters;
and different-case different-font, some upper case
Font2 letters. Three mixtures were blocked. Here,
the left segment of half of the stimuli (6 words and
6 nonwords) or the right segment of the other half
were in lower case Font1, the remaining segments
being in upper case and/or Font2 (e.g., bar RED,
bar red, and bar RED). Font1 lower case letters
also alternated with upper case and/or Font2 letters
(e.g., bAr ReD, bar red, and bAr ReD). These
7 conditions (same case and font plus the blocked
and alternate mixes) generated 168 stimuli.
Selected segments of these 168 items were
combined to produce 32 control items with
different-letter gaps (e.g., star low): a typographic
error necessitated deletion of one of these. The 199
stimuli
were
individually
presented
in
pseudorandom order for TR to pronounce then to
copy, a slash representing inter-string gaps.
Results and Discussion
Table IV shows TR deleted few control
different-letter gaps and almost all same-letter gaps
(χ27 = 133.22, p < 0.001). Moreover, different
letters were never deleted, while a recurring letter
was frequently deleted (χ27 = 29.21, p < 0.001),
the deletion rate being constant across case-font
conditions. The E categorised mixed case-font
target letters as almost identical (e.g., bar red),
similar (e.g., BeL lOw), or very different (e.g., lad
TABLE IV

Frequency of Gap and Letter Deletions: Manipulating the Casefont Mix does not alter Selective Deletion Rate
Deletion Incidence

Same case
Same font
Different font
Blocked
Alternate
Different case
Same font
Blocked
Alternate
Different font
Blocked
Alternate
Control
Different letters

Gap

Letter

24/24

12/24

23/24
23/24

14/24
14/24

23/24
23/24

11/24
10/24

23/24
22/24

11/24
12/24

2/31

0/31

Selective recurring symbol deletions
DY). Letter deletions were equally frequent in all 3
levels of shape similarity (χ22 = 0.32, NS). TR’s
only other errors were 6, probably phonological,
single letter substitutions. Table IV treats all 6
substitutions as correct because she made a single
substitution in the different-letter control condition
and in 5 experimental conditions (all 5 being, dor
re → doore).
As in Experiments 1 and 2, TR quickly read
each stimulus aloud: i.e., reading was not letter by
letter. Migration errors of attentional dyslexia are
not confined to repeated letter deletions. TR’s
deletions affected central letters, ruling out neglect.
In summary, selective gap and letter deletion
rates were comparable across same-letter conditions,
similar (e.g., r r ) and different (e.g., g G )
letterforms being equally vulnerable. It is concluded
that TR deleted abstract letter identities.

EXPERIMENT 4
Deletion and Fragmentation in Nonwords
Experiment 4 replicates evidence that differentcase repeated and doubled letters were deleted. It
further shows that she did not delete adjacent
similarly-shaped different-identity letters. This
study also confirms preliminary data showing that
TR only deleted single letters, not recurring lettersequences (e.g., she did not make errors such as
slor ort → slort).
Method
The first stimulus set comprised twenty-four
lower case 9-letter nonwords, which included left
(e.g., irrasteat), middle (e.g., inoddaist), or right
(e.g., doscrippe) doubled letters. Changes to each
of these first set items generated 7 further 24-item
stimulus sets. In an upper case set, an upper case
letter substituted one lower case doubled letter
(e.g., irrasteat → iRrasteat). A 2-recurring-letter
set was produced by converting the doubled and
immediately adjacent letters into a repeated 2-letter
sequence (e.g., irrasteat → irirsteat). In the shape
control set, a similar-feature letter (p, r, d, o → q,

477

n, b, e, respectively) substituted one doubled letter
(e.g., irrasteat → inrasteat). Insertion of a space
between the target letters of the items of the first 4
sets generated a further 4 sets (e.g., ir rasteat, iR
rasteat, ir irsteat, and in rasteat). One
typographical error reallocated a stimulus into
another set, and the other error necessitated
removal of that item. Stimuli were presented
individually and in random order. TR read each
aloud before transcribing it, representing interstring gaps with a hyphen.
Results
For with-gap strings, 1 gap was deleted (Table
V), no gaps were added, and 6 were relocated up
to two letters to the left (e.g., spetrale e → spetrale) or right.
For the 4 sets of without-gap stimuli, gaps were
inserted into between 16.7% and 33.3% items. For
those containing doubled letters, inserted gaps were
significantly (Binomial p < 0.007) more often
located between doubled letters (e.g., deloffots →
delof-ots) than off-target. The gap was inserted
between the target letters of 2-recurring-letter and
control strings, (e.g., spetralal → spetral-al, target
letters underlined) as often as not (Binomial, NS):
all off-target gaps appeared towards the centre of
strings with left-sided targets (N = 7, e.g., coesleats
→ coes-leats) or right-sided targets (N = 8, e.g.,
cirostees → ciros-tees).
The 6 target letter substitutions (e.g., asenanate
→ asenenate) were ignored, being evenly
distributed across conditions. Considering target
deletions, complete 2-recurring-letter sequences
were never deleted, though there were occasional
single-letter errors (e.g., ir irsteat → ir-isteat).
Letter deletion rates were comparable whether case
was mixed or unmixed within the repeated and
doubled subgroups (Table V), so mixed and
unmixed case data were pooled. Repeated letters
(χ21 = 60.92, p < 0.001) and doubled letters (χ21 =
21.60, p < 0.001) were deleted more often than
shape control letters with and without gaps,
respectively (Table V). Moreover, letter deletions
were more common (χ21 = 18.48, p < 0.001) in
repeated than in doubled targets (Table V).

TABLE V

Number of Gaps Perceived and Target-letters Deleted in Nonwords with rRepeated or Doubled Letters
Stimuli
(with-gap
examples
in brackets)
Lower case recurring
(e.g., inod daist)
Mixed case recurring
(e.g., uN niposta)
2-recurring-letter
(e.g., esporot ot)
Shape control
(e.g., nevar neal)

Gap

Letter Deletion

With Gap
(Deletion)

Without Gap
(Insertion)

With Gap

Without Gap

0/25

7/24

23/25

13/24

1/24

6/24

23/24

14/24

0/22

8/24

2/22

0/24

0/24

4/24

0/24

0/24

478

Rodger A. Weddell

Off-target
letter
errors
were
mostly
phonological and were equally distributed across
conditions: 12 single-letter additions (e.g.,
preorasin → preoraisin), 3 one or two letter
substitutions (e.g., cunslot ta → conslot ta), 1
deletion (i.e., meargar ar → meagar ar), and 1
migration (i.e., nevar-arel → nevar-eal).
Comparing the 4 leftmost vs. 4 rightmost letters,
off-target errors were more frequent on the right
(χ21 = 4.35, p < 0.05) perhaps because nonwords
are transcribed left-to-right yielding more rightsided phonological errors for long nonwords. More
importantly, left and right target rates were very
similar (χ21 = 0.00, NS); i.e., unilateral neglect
does not explain target letter deletions.
Discussion
TR’s deletion of different-case recurring letters
and faithful transcriptions of similar-feature
(different-letter) control pairs confirms that abstract
letter identities are deleted. The lateral reach of this
effect did not extend beyond adjacent single letters,
as 2-recurring-letter sequences were not deleted. It
is suggested that competitive lateral interactions
between
adjacent
same-identity
letter
representations weaken or “delete” one letter
representation but not letter sequences, especially
as she selectively deleted recurrent letters exposed
too briefly for saccade initiation (see Case Report).
In the present study, TR added gaps, while the
other 5 studies in this report demonstrate robust
gap deletion. This difference is explained using 3
assumptions. First, gap additions reflect TR’s
fragmented perceptions. Thus, she said that she
often read short letter strings as fully resolved
single units, but longer strings first appeared as an
indistinct dark blur, which sometimes included
uncategorized forms. The blur then subdivided into
2 or 3 segments separated by “illusory” gaps,
which she distinguished from real spaces. The
letterforms in 1 segment then resolved.
Subjectively, this process entailed matching letters
with candidate phoneme sequences. The segment
then returned to its blurred state as the next
segment resolved, the process repeating itself for
any third fragment. Finally, she mapped the sounds
of phoneme groups onto their remembered
locations in the residuum of the initial visual
representation of the whole string.
It is next assumed that she sometimes confused
illusory and real gaps. On one hand, same-letter
real and illusory gaps were particularly difficult to
distinguish because competitive lateral interactions
are particularly strong when adjacent letters are the
same. One letter representation might be weakened
or “deleted”. A weaker signal in the region of
doubled letters may have contributed to TR’s
present tendency to insert a gap between doubled
letters in unbroken strings. On the other hand,
while Experiments 1-3 indicate that TR most

readily distinguished real from illusory gaps
separating different letters, the present addition of
gaps to unbroken control strings suggests she
sometimes converted illusory into real differentletter gaps when resolving longer nonwords.
The third assumption is that resolved fragments
were short because TR could group only a severely
limited amount of information in one visual
percept. Remembering phoneme sequences and remapping them onto a residual visual representation
of the string creates more room for information
degradation and error. It is argued that TR
degraded (and, consequently, deleted) gap
information in the attempt to “squeeze” the shorter
strings of Experiments 1-3 into a single object
grouping. Conversely, representations of potential
inter-string spaces were facilitated in the pressure
to encode long 9-letter strings into at least 2 visual
percepts; consequently, TR introduced rather than
deleted gaps in the present study.
Finally, fragmentation might facilitate recurring
letter deletion. In the complex process of segment
resolution, it might be difficult to distinguish (a)
“erroneous” re-activation of a letter in a
previously-resolved segment from (b) an adjacent
identical letter in another segment. Additionally,
insertion of a gap between recurring letters might
enhance deletion because gaps trigger saccades,
necessitating integration of new visual input with
that gained from the previous saccade. This might
increase the difficulty of distinguishing single and
adjacent recurring letters.
EXPERIMENT 5
Non-lexical Symbol Groups are Deleted
Experiment 5 eliminates lexical influences,
which conceivably constrained letter and gap
deletions in Experiments 1-4. It demonstrates
selective same-symbol gap deletion from
meaningless strings of English consonants, Greek
letters, and mathematical symbols.
Method
Insertion of 4 line breaks into 8 randomlysequenced, 144 lower case consonant (c, r, n, and
x) strings yielded 5 different-length lines.
“Paragraphs” of variable length letter groups were
created by inserting inter-string spaces at random
points within each line: 117 spaces (39 spaces 1, 2,
and 3 additional characters wide) were distributed
between 8 different “paragraphs” which contained
11-19 spaces. Four versions of the 8 paragraphs
were created: same or different letters bounded all
(a) inter-string gaps and/or (b) line breaks.
Mathematical symbols (–, +, ≈, ÷) or Greek letters
(φ, σ, π, τ) were substituted for each letter to
produce 2 further 32-paragraph sets. The following

Selective recurring symbol deletions

2 lines are from a different-symbol inter-string gap,
same-symbol line break, Greek-letter paragraph:

479

mathematical symbol gaps. Gap approached
significance [F (1, 84) = 3.56, p < 0.07] because
she detected slightly more different-letter gaps than
same-letter gaps. However, break [F (1, 84) = 0.01,
NS] and all other terms were insignificant.

φπφσφτ φπτσπστ σφσ πσπσττσφ τπσφσττ φτπτπτ
τφσπφπτφσστ πσττπσσπσ φτππσφπφσφ

Discussion

TR counted either symbol groups/strings or
inter-string gaps. Line breaks were included in
group counts, and excluded in gap counts (the
above lines contain 9 symbol groups and 7 gaps).
Two 2 matched and randomly ordered 48paragraph blocks were presented in the same order
in 2 sessions. In session 1, TR counted groups and
gaps in the first and second blocks, respectively. In
session 2, gap counting preceded group counting.

When grouping symbols, TR selectively deleted
same-symbol spaces. This clearly non-lexical
deficit affected all symbols equally. In contrast,
same-symbol gaps minimally impeded gap
counting.
TR described different group and gap counting
strategies. For group counting, she said she
combined each successive symbol with the
previous few symbols until a break appeared. She
then (a) added 1 to the cumulating group count,
and (b) started assembling the next group. This
strategy narrowly focused on local features.
However, gap counting was more global: she said,
“I keep looking for a white gap and then I become
aware of the letters.” Here, her apparently more
global analysis might have reduced detection of
smaller gaps (about 5 per paragraph), but her
global strategy did not entirely eliminate local
analysis, since there was a weak tendency to delete
same-letter gaps.
A spontaneous comment fits the foregoing
local-global hypothesis rather well. After counting
1 group in a same-letter (gap and line break)
paragraph, TR pondered, “I say ‘one’ but it’s four
lines ... but it’s all one group? When you took it
away, it flashed that it was four lines.” More
specifically, her group counting focus was local
and confined to a few symbols: she counted 1
group because no clear break appeared (though she
probably perceived “illusory” gaps) despite large
eye movements at line breaks. In contrast, she
adopted a global perspective when the paragraph
was removed and she discerned 4 of the 5 lines.
This fits another comment that she often initially
perceived the 5-line paragraphs as a top and a
bottom line of dark blurred segments each separate
from a blurred region sandwiched between these
lines.

Results
Group counts (M = 18.5 sec, SD = 4.3, per
paragraph) took longer (t = 8.27, p < 0.001) than
gap counts (M = 13.2 sec, SD = 4.5). Proportion
correct (group count divided by correct group
number) was the dependent variable in a 3 × 2 × 2
ANOVA with symbol (English vs. mathematical vs.
Greek), gap (same- vs. different-symbol), and line
break (same- vs. different-symbol) as factors. Gap
[F (1, 84) = 9240.09, p < 0.001] and break [F (1,
84) = 694.89, p < 0.001] were the only significant
terms because TR identified virtually all differentsymbol (3 gap sizes and line breaks) boundaries,
and remarkably few same-symbol boundaries
(Table VI). Thus, her group counts were close to
the true value when all separations were differentletter. There were 4 line breaks, and she deleted
about 4 groups from same-symbol-break, differentsymbol-gap paragraphs. Conversely, each line in
same-symbol-gap, different-symbol-break paragraphs
was usually deemed a single group. She often saw
only one group when gaps and line breaks were
both same-symbol.
TR missed an average of 4.9 (SD = 2.2) gaps
(Table VI). Her gap count divided by the true gap
number was the dependent variable in a 3 × 2 × 2
ANOVA with symbol, gap, and break as factors.
Symbol was significant [F (2, 84) = 3.10, p < 0.05]
because she saw more (M = 1.2) consonant than

TABLE VI

Mean Number of Groups (true M = 19.6) and Gaps (true M = 14.6) counted in Paragraphs Composed of Different Symbols and Gap
Types (number deleted in brackets)

1. Count Groups
English letters
Greek letters
Math symbols
2. Count Gaps
English letters
Greek letters
Math symbols

DIFFERENT SYMBOL GAPS

SAME SYMBOL GAPS

Line break symbols

Line break symbols

Different

Same

Different

Same

19.1 (– 0.5)
19.4 (– 0.2)
18.9 (– 0.7)

15.4 (– 4.2)
15.5 (– 4.1)
15.3 (– 4.3)

5.1 (– 14.5)
5.0 (– 14.6)
5.3 (– 14.3)

1.0 (– 18.6)
1.5 (– 18.1)
1.4 (– 18.2)

11.5 (– 3.1)
10.0 (– 4.6)
9.1 (– 5.5)

10.4 (– 4.2)
9.9 (– 4.7)
9.8 (– 4.8)

9.5 (– 5.1)
9.5 (– 5.1)
8.8 (– 5.8)

9.8 (– 4.8)
8.8 (– 5.8)
8.8 (– 5.8)

480

Rodger A. Weddell

Fig. 2 – Selective deletion of same-element Navon-style contours.

EXPERIMENT 6
Non-lexical Navon-style Contours are also Deleted
Experiment 5 demonstrated deletions of gaps
separating non-lexical symbol groups that could not
be phonologically encoded. That observation is
extended here with Navon-style forms, which
reveal selective deletion of same-Greek-letter
contours and gaps. An initial formal study used
pairs of familiar forms that (a) were separated by a
gap, (b) shared one contour, or (c) overlapped.
Subsequent supplementary explorations probed this
phenomenon further.

σ). Model forms were separated by a gap (e.g., Figure
2a), shared a contour (e.g., Figure 2b), or overlapped
(e.g., Figures 2c and d). The pairs were: vertically
arranged C-shapes, contour and gap deletion
producing an E (e.g., Figure 1a); vertically arranged
D’s (see Figure 2b), contour and gap deletion
yielding B; adjacent 0 | (see Figure 2c), producing d
on contour and gap deletion; and vertically arranged
O’s (see Figure 2d), forming 8 after contour and gap
deletion. These randomly ordered 24 stimuli, which
were 2.5 cm to 4 cm high and 1.6 cm to 2.3 cm wide,
were presented individually for 10 sec and drawn
from memory. They were continuously exposed and
copied later on. Supplementary investigations relied
on direct copying.

Method
There were 4 pairs of meaningful forms, each
constructed from the same (e.g., Figure 2a and d) or
different (e.g., Figures 2b and c) Greek letter (Σ and

Results
Table VII shows that TR usually deleted a
contour and gap from separate (e.g., Figure 2a) and

Selective recurring symbol deletions

481

TABLE VII

Meaningful Navon-style Greek-letter Contours: Frequency of Contour Deletion and Continuity
Same Elements

Separate contours
Shared contour
Overlapping contours

Different Elements

Contour
Deletion

Contour
Continuity

Contour
Deletion

Contour
Continuity

4/4
0/4
3/4

4/4
4/4
4/4

0/4
0/4
0/4

4/4
0/4
0/4

overlapping (e.g., Figure 2d) same-Greek-letter
pairs, but never from different-letter pairs (e.g.,
Figure 2c). Unsurprisingly, contours were never
deleted from shared contour pairs (Figure 2b).
Table VII also shows that same-letter forms were
always drawn as continuous and without breaks
(Figures 2a and d). Different-letter forms were
continuous only for separate pairs (Table VII), and
breaks in her copies of shared and overlapping
contours often corresponded with the points of
transition from one letter to the other, presumably
reflecting lateral inhibition between competing
forms and/or groupings (see Figures 2b and c).
Finally, identical results were obtained when TR
copied from continuously exposed models.
Four pairs of nonsense forms were constructed
from the same or different Greek letters (e.g.,
Figure 2e). Separate and shared contour versions
were created. TR’s copies of these 16 continuously
exposed stimuli corresponded with her copies of
meaningful forms. Separate same-letter contours
merged and contours were continuous. Differentletter nonsense contours or the intervening spaces
were never deleted and shared contours were
discontinuous (Figure 2e).
The distance between the meaningful forms of
original separate contour stimuli was doubled
(4 mm to 8 mm). Whole stimuli were doubled in
size on another occasion. These size manipulations
did not reduce selective deletion of same-letter
contours.
The 4 separate contour pairs and a fifth (V V
→ W, adjacent V-shapes becoming a W-shape, if
merged) were constructed from various elements.
These were sometimes presented on more than one
occasion. She consistently merged only same-letter
contours. Her robustly selective same-letter
deletions extended to the following component
elements: English letters (10/10), dots (9/10),
asterisks (10/10), and dashes (5/5). This effect also
generalised to contour pairs both constructed from
the letter Σ, with the letter elements rotated through
180° (which usually, 9/10, merged) and 45° (which
often, 5/8, merged, Figure 2f).
There were limits to these effects. Thus, 90°
Greek-letter element rotations infrequently (2/12)
merged. Merging did not occur when contours
were continuous unbroken lines (0/5), suggesting
that grouping discrete elements into a contour
facilitated merging. Finally, 5 contour pairs
constructed from dots were presented with one set

of the pair coloured red, the other dots being blue.
Five similar mixed colour pairs were constructed
with dashes. TR rarely merged these contours
(1/10), indicating that colour or form differences in
the grouped elements prevent deletion.
Finally, TR sometimes described her
perceptions. After drawing Figure 2c, she said, “It’s
like a blur if I look at it all at once”, and “I home
in on the biggest part first…” She added that she
saw the left-sided arc first, then the vertical line
and, finally, the right-sided arc. In fact, when
asked, she always saw different-symbol forms as a
series of resolving fragments, while all parts of
same-letter stimuli appeared simultaneously in an
unfragmented percept.
Conclusions
1. Selective same-symbol deletions extended to
meaningful and meaningless Navon-style global
forms, emphasising the non-lexical nature of the
deficit. Moreover, merging of vertically arranged
components tends to suggest that neglect cannot
explain her difficulty.
2. TR deleted contours made from various
elements (English and Greek letters, dots, asterisks,
and dashes), which further suggests a non-lexical
grouping disorder.
3. Greek letters were used in the belief that they
would evoke neural shape representations rather than
abstract letter identities. However, she deleted the
same Greek letters rotated through 45° or 180°, and
she recalled using Greek letters in school algebra;
consequently, Greek letters may have evoked some
form of abstract representation. It would, therefore,
be interesting to see if deletions extend to Navonstyle forms composed of completely unfamiliar
shapes (e.g., Arabic letters), as that would indicate
that shapes per se (i.e., relatively low-level feature
combinations) can support deletion.
4. The fragmentation TR experienced during
word resolution extended to Navon-style global
forms. Some stimuli were fragmented into 2-3
parts; each was resolved individually as the other
segment(s) remained blurred.
GENERAL DISCUSSION
Experiments 1-4 demonstrate robust selective
deletion of recurring letters and/or associated gaps.

482

Rodger A. Weddell

This effect was limited to adjacent single letters, as
recurring 2-letter sequences (e.g., arar → arar)
were unaffected in Experiment 4. Experiment 2
excluded repetition blindness as a possible cause
(Kaniwisher, 1991). In Experiments 2&4, the gap
between repeated letters facilitated deletion
(doubled letters being less often deleted), and this
effect was linked with TR’s tendency to resolve
letter-strings as a series of fragments. Robust
deletions from case and font mixtures (e.g.,
dD → d) and no effect of letter shape similarity
on deletion (e.g., pq → pq) show she deleted
abstract letter identities (Experiments 3 and 4).
Experiments 5 and 6 confirm the deficit was nonlexical. Thus, same-symbols (e.g., Mathematical
symbols and Greek letters) were selectively deleted
when TR (a) assembled symbols into non-linguistic
groups and (b) copied meaningful and meaningless
Navon-style global forms.
Selective recurring letter deletions and the
virtual absence of substitution and migration errors
excluded explanation in terms of attentional
dyslexia. There was no evidence of neglect and she
did not read letter-by-letter. Moreover, the visual
non-lexical nature of TR’s reading difficulties
permits the present attempt to link TR’s deficits
with primate visual neurophysiology and the
biologically plausible neural network model
developed by Rolls and Deco (2002).
Implications of the Rolls and Deco (2002)
Model for Reading
Single cell recordings reveal steadily increasing
receptive field (RF) area as the signal feeds
forward through V1 → V2 → V4 → TEO
(posterior inferotemporal) → TE (inferotemporal)
cortices. Figure 3 schematises the increasing
receptive field size in ventral visual pathways. It
shows lateral geniculate body innervation of V1
RF1 to RF8. Projections from two V1 neurons
(RF1 and 2) converge onto one V2 neuron,
yielding a double-sized RF. Similarly, the left-most
V4/TEO neuron receives information from V1 RF’s
1-4, while the TE neuron is informed by RF’s 1 to
N. Thus, if a stimulus activating V1 RF1 moves (or
if the eyes move) so that RF2 is now activated,
then the same V2 and all subsequent neurons still
respond to that stimulus. Similarly, stimulus
movement to RF3 causes selective activation of the
next-in-line V2 neuron, but the same V4/TEO and
TE neurons still respond to that stimulus. VisNet
simulated this translation invariant effect for the
symbols L, T, and + (Rolls and Deco, 2002). Each
symbol activated nodes in a circumscribed region
of input layer 1, while layer 4 nodes responded
wherever the symbol was located. Rolls and Deco
(2002) also report the following actual progression
in RF size in the macaque: V1 (0.5°) → V2 (3.2°)
→ V4 (8°) → TEO (20°) → TE (40°). The
progressive coarsening of the spatially discrete

retinotopic organisation of V1 until its apparent
loss in TE (where cells respond to stimuli
appearing in central vision) corresponds with
increasing complexity of the forms evoking
selective neuronal responses. Thus, V1 neurons
respond selectively to oriented edges, while some
TE neurons respond to faces (Desimone et al.,
1985; Hasselmo et al., 1989). At intermediate
stages, the proportion of cells responding to
complex feature combinations increases as the
signal passes through V4 to TEO (Tanaka et al.,
1993).
In addition to feeding forward, extensive
backprojections feed the signal back. The resulting
re-entrant recycling of information presumably
partly underlies prolongation (200-300 msec) of TE
neuronal responses (Rolls and Tovee, 1994).
Importantly, monkeys successfully discriminated
backwardly masked faces, though TE neurons were
selectively responsive for only 20-30 msec, which
suggests that the early activity phase is often
sufficient for identification. Later phases of the
sustained TE activity may correlate with other
processes such as short-term visual memory (Rolls
et al., 1999) and/or may contribute to recognition
of ambiguous objects. Consider a possible
implication for visual representations of ambiguous
objects. One form of ambiguity might entail
insufficient activation of a single TE neuronal
ensemble selectively tuned to the stimulus.
Response selection often does not proceed until
one ensemble attains supra-threshold activation via
recycling activity in striate-extrastriate circuits. The
hypothesised dynamic recycling of information
would presumably include activity in nonretinotopic TE neurons plus increasingly retinotopic
activity as backprojections re-activate TEO, V4,
V2, and perhaps V1. Thus, the ambiguous
representation is a dynamic spatially distributed
zone of activity. It includes responses to (a) simple
features in neurons within a defined region of
retinotopic primary cortex, (b) more coarsely
retinotopic intermediate cortical (V4/TEO)
representations of complex feature combinations,
and (c) non-retinotopic activity in TE neurons
tuned to different objects.
Covert attentional shifts might resolve the
ambiguity through sequential and more detailed
analyses of different parts of the ambiguous object;
Rolls and Deco (2002) provide a neural mechanism
that may have been engaged when TR sequentially
resolved perceptual fragments. They suggest
parietal projections to V1 (and possibly up to
V4/TEO) selectively facilitate activity in discrete
retinotopic regions of vision (Figure 3). Here,
modulation would act on the retinotopic end of the
(V1 through to TE) dynamic recycling
representation of the ambiguous object. This
attentional activation is effectively equivalent to
allowing only a spatially limited segment of the
stimulus to be fed forward, thereby allowing

Selective recurring symbol deletions

483

Fig. 3 – Ventral stream retinotopic through to non-retinotopic cortical substrate for recycling visual representations modulated by
dorsal stream parietal spatial attentional processes.

“later” extrastriate networks to predominantly
process that segment. Once that part of the signal
has been processed, parietal modulations would
highlight another zone. These sequential attentional
shifts might glean sufficient information for suprathreshold activation of one TE, object-specific,
neuronal ensemble.
Extension of the Model to TR’s Reading
This will be achieved using the model Caramazza
and Hillis (1990) applied to a case of neglect in
lexical and non-lexical tasks. The model specifies 3
stages preceding access to the orthographic input
lexicon. In stage 1, feature contours (e.g., horizontal
and vertical lines) are represented retinotopically. In
stage 2, features are integrated into letter shapes –
the physical appearance of letters being maintained

(e.g., R, r, and R being distinct). In stage 3,
graphemes are computed from letter shapes. These
abstract letter identities (case- and font-independent)
are stimulus-centred, being spatially arrayed
according to ordinal positions relative to the central
letter (e.g., in CUP, <C>, <U>, and <P> occupy
positions – 1, 0, and 1, respectively).
It is suggested that V1 (and perhaps V2)
provide a retino-centric feature map of letter
strings, while neurons responding to letter shape
and abstract letter identities are in the human
homologues of intermediate ventral extrastriate
regions (V4/TEO). However, relatively more V4
(and perhaps V2) neurons respond to shape, while
TEO neurons are more often tuned to abstract letter
identities (Tanaka et al., 1993). V4 and TEO are
coarsely retinotopically-organised; consequently,
activated abstract letter identities would coarsely

484

Rodger A. Weddell

encode within-string left-right location. This model
diverges from the original in 3 ways. First,
retinotopy ends after stage 1 in the original. In the
present account, it progressively degrades from V1
through TEO, which affords a stimulus-centred
representation by gathering information from a
large number of V1 neurons. Second, Caramazza
and Hillis (1990) specify solely bottom-up
processing, but this assumption is biologically
implausible, given the evidence of extensive
backprojections and recycling of neural activity
between visual cortices (Rolls and Deco, 2002).
Moreover, the modified model resolves a puzzle
recognised by Caramazza and Hillis, who could not
reconcile 2 assumptions: graphemic information is
(a) abstract and it is (b) “… arrayed in a spatially
defined co-ordinate system.” Third, the original
model states that stage 2 letter representations CUP
& PUC both map onto one stage 3 word-centred
grapheme representation <C><U><P>. However,
mirror-reversed words are not automatically
mapped onto the stage 3 representation of the
modified model, which would assume application
of a strategy, e.g., the reversed stimulus is read
right-to-left and letter-by-letter to produce a left-toright image, which may be subject to neglect.
The orthographic lexicon is thought to occupy
non-retinotopic TE, which contributes neurons
selectively tuned to whole familiar letter-strings
(words and subword units) to the re-entrant
dynamic representation. In the manner suggested
by Plaut et al. (1996), TE units feed forward to
intact phonological lexical processing mechanisms,
perhaps in the left posterior superior temporal
gyrus (Small et al., 1996). Indeed, the weak lexical
influence on deletion rate in Experiment 1 does not
necessarily imply additional lexical impairments.
Thus, Behrmann et al. (1998) argue that the lexical
effects shown by letter-by-letter readers reflect
intact top-down lexical influences on degraded prelexical visual processes.
It has already been suggested that double letter
deletion partly reflected lateral inhibition between
adjacent letter representations. Briefly, competition
particularly increased ambiguity at points in the
signal where the identity of adjacent letters was the
same. This may end in “winner-take-all” deletion
of the weakest letter representation and it may
explain TR’s tendency to insert a gap between
doubled letters in Experiment 4. Competition is
hypothesised to occur at all levels of the dynamic
representation. On the one hand, if letter deletions
were due to dysfunction of the V1 cortical feature
map, then lateral inhibition between V1 neurons
(macaque RF area 0.5°) representing same letters
would be stronger when letters are separated by a
single character space (0.4°) than by a larger space.
However, the size of the inter-symbol gap did not
alter error rate in Experiments 1, 5, and 6. On the
other hand, symbol deletions cannot depend solely
on competition between damaged non-retinotopic

TE representations of word and subword units.
Also, as only adjacent letters are deleted, the
competing representations probably at least
coarsely encode retinotopic location. Therefore, it
is concluded that deletions at least partly reflect
competition between damaged ensembles tuned to
abstract letter identities (and perhaps letter-shape
representations) in the human homologues of
V4/TEO (i.e., in intermediate retinotopic cortex
with larger RF sizes).
To explain TR’s fragmented percepts, it is argued
that whole stimuli resolved without fragmentation if
the initial forward progression of the signal activated
one TE ensemble. However, anterior extrastriate
damage meant that one TE ensemble often failed to
achieve supra-threshold activation; that is, damage
to word/subword and letter identity representations
yielded an ambiguous response to the information in
the signal. (Though the parahippocampal damage
evident on TR’s CT scan is too small to explain her
initial severe visual agnosia and permanent
prosopagnosia, its anterior location suggests the
dysfunction encompassed TE, TEO and perhaps V4
circuits.) It might be argued that ambiguous anterior
extrastriate responses corresponded to TR’s blurred
percepts. She next divided the blurred region into 2
or 3 segments. (Unpublished experiments by the
author show that successive fragment resolution still
occurred when saccades were excluded by using
brief stimulus exposures). As we have seen, the Rolls
and Deco (2002) model permits the possibility that
covert shifts in spatial attention may be effected via
parietal projections to V1 (and perhaps V4/TEO).
Essentially, parietal modulation would sequentially
facilitate activity in different regions of the persisting
(200-300 msec) and dynamically re-cycling
representation of the string. This activation would
effectively feed forward data in the region of one
unresolved segment, and enable compromised
anterior extrastriate networks to largely dedicate
processing to that segment. Once that signal was
resolved, parietal projections would highlight
another zone.
Acknowledgements. I am grateful to TR, who willingly
and generously participated in the study. I also thank H.
Extence and A. Miles for transcribing TR’s audio-taped
responses to stimuli presented in Experiment 2. Finally, I
appreciate Dr A. Beaton’s comments on an earlier draft of
this manuscript.
REFERENCES
BAVELIER D and POTTER MC. Visual and phonological codes in
repetition blindness. Journal Experimental Psychology:
Human Perception and Performance, 18: 134-147, 1992.
BEHRMANN M, PLAUT DC and NELSON J. A literature review and
new data supporting an interactive account of letter-by-letter
reading. Cognitive Neuropsychology, 15: 7-51, 1998.
BEHRMANN M and SHALLICE T. Pure alexia: A nonspatial visual
disorder affecting letter activation. Cognitive Neuropsychology,
12: 409-454, 1995.
CARAMAZZA A and HILLIS AE. Levels of representation, coordinate frames, and unilateral neglect. Cognitive
Neuropsychology, 7: 391-445, 1990.

Selective recurring symbol deletions
DESIMONE R, SCHEIN SJ, MORAN J and UNGERLIEDER LG. Contour,
color and shape analysis beyond the striate cortex. Vision
Research, 25: 441-452, 1985.
FARAH MJ. Visual Agnosia: Disorders of Object Recognition and
What They Tell Us about Normal Vision. Cambridge, MA:
MIT Press, 1990.
FARAH MJ and WALLACE MA. Pure alexia as a visual impairment:
A reconsideration. Cognitive Neuropsychology, 8: 313-334,
1991.
HALL DA, HUMPHREYS GW and COOPER CG. Neuropsychological
evidence for case-specific reading: Multi-letter units in visual
word recognition. Quarterly Journal of Experimental
Psychology, 54A: 439-467, 2001.
HASSELMO ME, ROLLS ET and BAYLIS GC. The role of expression
and identity in the face-selective responses of neurons in the
temporal visual cortex of the monkey. Behavioural Brain
Research, 32: 203-218, 1989.
HUMPHREYS GW and PRICE CJ. Visual feature discrimination in
simultanagnosia: A study of two cases. Cognitive
Neuropsychology, 11: 393-434, 1994.
KANWISHER N. Repetition blindness and illusory conjunctions:
Errors in binding visual types with visual tokens. Journal
Experimental Psychology: Human Perception and
Performance, 17: 404-421, 1991.
LEVINE DN and CALVANIO R. A study of the visual defect in
verbal-alexia simultanagnosia. Brain, 101: 65-81, 1978.
MICELI G, BENVEGNU B, CAPASSO R and CARAMAZZA A. Selective
deficit in processing double letters. Cortex, 31: 161-171,
1995.
MAYALL K and HUMPHREYS GW. Presentation and task effects on
migration errors in attentional dyslexia. Neuropsychologia,
40: 1506-1515, 2002.
NELSON HE and WILLISON JR. National Adult Reading Test,
Second Edition. Windsor: NFER-NELSON, 1991.
PLAUT DC, MCCLELLAND JL, SEIDENBERG MS and PATTERSON K.
Understanding normal and impaired word reading:

485

Computational principles in quasi-regular domains.
Psychological Review, 103: 56-115, 1996.
RIDDOCH MJ and HUMPHREYS GW. Birmingham Object
Recognition Battery. Hove: Lawrence Erlbaum, 1993.
ROLLS ET and DECO G. Computational Neuroscience of Vision.
New York: Oxford University Press, 2002.
ROLLS ET and TOVEE MJ. Processing speed in the cerebral cortex
and the neurophysiology of visual masking. Proceedings of
Royal Society of London, Series B, 257: 9-15, 1994.
ROLLS ET, TOVEE MJ and PANZERI S. The neurophysiology of
backward visual masking: Information analysis. Journal of
Cognitive Neuroscience, 11: 300-311, 1999.
SAFFRAN EM and COSLETT H. ‘Attentional dyslexia’ in Alzheimer’s
disease: A case study. Cognitive Neuropsychology, 13: 205228, 1996.
SHALLICE T and WARRINGTON EK. The possible role of selective
attention in acquired dyslexia. Neuropsychologia, 15: 31-41,
1977.
SMALL SL, NOLL DC, PERFETTI CA, HLUSTIK P, WELLINGTON R
and SCHNEIDER W. Localizing the lexicon for reading aloud:
Replication of a PET study using fMRI. Neuroreport, 7: 961965, 1996.
TANAKA K, FUJITA I, KOBATAKE E, CHENG K and ITO M. Serial
processing of visual object-features in the posterior and
anterior parts of the inferotemporal cortex. In Ono T,
Squire LR, Raichle ME, Perret DI and Fukuda M (Eds),
Brain Mechanisms of Perception and Memory: From
neuron to behaviour. New York: Oxford University Press,
1993.
YOUNG AW, NEWCOMBE F and ELLIS AW. Different impairments
contribute to neglect dyslexia. Cognitive Neuropsychology, 8:
177-191, 1991.
Rodger A. Weddell, Neurosciences Directorate, Morriston Hospital, Swansea and
Department of Psychology, University of Wales, Swansea.
e-mail: linda.young@swansea-tr.wales.nhs.uk

(Received 29 October 2002; reviewed 14 February 2003; revised 9 September 2003; accepted 2 October 2003; Action
Editor Alan Beaton)