Too harts, won sole: Using Dysgraphia treatment to address homophone
representation

Polly Barr1,2, Britta Biedermann3,2, Marie-Joseph Tainturier4, Saskia Kohnen2,5, Lyndsey
Nickels2,5
1

The Brain and Mind Centre, University of Sydney, Sydney, Australia

2

ARC Centre of Excellence in Cognition and its Disorders, Macquarie University, Sydney,

Australia
3

School of Occupational Therapy, Social Work and Speech Pathology, Curtin University,

Perth, Australia
4

School of Psychology, Bangor University, Bangor, Wales

5

Department of Cognitive Science, Macquarie University, Sydney, Australia

Corresponding Author:
Dr Polly Barr, The Brain and Mind Centre, University of Sydney, Sydney, NSW 2000 Australia
Email: polly.barr@sydney.edu.au
Word count: 7825
Abstract word count: 195

1

Abstract
Previous spoken homophone treatment in aphasia found generalisation to untreated
homophones and interpreted this as evidence for shared phonological word form
representations. Previous written treatment of non-homophones has attributed
generalisation to orthographic neighbours of treated items to feedback from graphemes to
similarly spelled orthographic word forms. This feedback mechanism offers an alternative
explanation for generalisation found in treatment of spoken homophones.
The aim of this study was to investigate the mechanism underpinning generalisation (if
any) from treatment of written homophones. To investigate this question a participant with
acquired dysgraphia and impaired access to orthographic output representations undertook
written spelling treatment. Generalisation to untreated items with varying degrees of
orthographic overlap was investigated. Three experimental sets included homographs (e.g.
bank-bank), heterographs (e.g., sail-sale), and direct orthographic neighbours (e.g., bathpath). Treatment improved written picture naming of treated items. Generalisation was
limited to direct neighbours. Further investigation of generalisation found that items with a
greater number of close neighbours in the treated set showed greater generalisation. This
suggests that feedback from graphemes to orthographic word forms is the driving force of
generalisation. The lack of homograph generalisation suggests homographs do not share a
representation in the orthographic lexicon.
Key words: Homophone, Language production, Spelling, Treatment, Dysgraphia

2

Introduction
There is general agreement that each word in our language requires a stored
representation of meaning (semantics), and phonological/orthographic form (e.g.,
Caramazza, 1997; Dell, 1986; Levelt, Roelofs, & Meyer, 1999). However, there are
circumstances when a word’s form completely overlaps with that of another word (e.g.,
cricket the insect, and cricket the sport). These types of words (homophones) clearly have
separate meanings and hence separate semantic representations but perhaps their form
need not be separately represented? The nature of homophone representation has been
investigated over the past 20 years using various different techniques (e.g., frequency
effects on reaction times, neuropsychological treatment, error rates) yet it is still unclear
how homophones are represented in the lexicon. This study explores the orthographic
representation and written production of homophones, which treatment factors may
influence generalisation and the nature of orthographic processing.
Representations in the lexicon
There are currently no models that focus solely on the orthographic representation
of homophones. In contrast, there are three competing models of homophone
representation in spoken production. These can therefore assist conceptualisation of written
homophone production. Figure 1 depicts these three theories of spoken homophone
production, whereby in (1a) homophones share a lexical representation as in, for example,
Levelt et al.'s Two Stage model (1999; Figure 1a). According to (1b) homophones have
independent representations with strictly feed forward activation (Figure 1b; Caramazza,
Costa, Miozzo, & Bi, 2001), while in (1c) homophones have independent representations
with interactive activation between the lexical representations and segment level (Figure 1c;
Middleton, Chen & Verkuilen's 2015 adaptation of Schwartz, Dell, Martin, Gahl, & Sobel,
2006). Due to similarities between spoken and written production (Damian, Dorjee, &
3

Stadthagen-Gonzalez, 2011) it seems plausible that phonological and orthographic
architecture are similar. Therefore, we can assume that the orthographic representation of
homophones is one of the figures in Figure 1, however the lexical representation would be
orthographic (not phonological) and the segment level would be graphemes (not
phonemes). Previous psycholinguistic studies have offered competing evidence for the
different types of models shown in Figure 1 (e.g., Caramazza et al., 2001; Dell, 1990;
Jescheniak & Levelt, 1994), however here we focus on neuropsychological evidence.
Figure 1 about here

Treatment studies of spoken homophone production
One method of investigating homophone representation and the nature of
orthographic processing involves exploring generalisation of the effects of language therapy
with people with aphasia. Previous studies have found that treating one homophone (e.g.,
knight) improves spoken naming of the untreated homophone partner (e.g., night).
Following Blanken’s (1989) landmark study, Biedermann and colleagues replicated the
pattern of homophone generalisation in both German (one person) and English people with
aphasia (two people) with lexical-access impairments. Biedermann and Nickels (2008b)
found this effect for both homographic homophones (e.g., cricket [insect]/cricket [sport])
and heterographic homophones (e.g., knight/night), with no difference in the extent of
generalisation explained by orthographic similarity. As there was no generalisation to
phonologically related items, the authors interpreted generalisation as being due to shared
representations at the phonological word form level (support for Figure 1a).
Biran et al. (2013) replicated this work in Hebrew using a phonological cueing therapy
with two people with aphasia who also had word retrieval deficits. Both participants showed
improvement in naming treated homographic homophones (e.g., mapa; tablecloth) and

4

their untreated homophone partners (e.g., mapa; map), but not phonologically related
controls (e.g., maca; Matzah).
In sum, to date, every phonological treatment study that has addressed spoken
homophone production has found generalisation of treatment to untreated homophones.
Furthermore, the effects have been interpreted as evidence for shared representations at
the phonological word form (Biedermann et al., 2002; Biedermann & Nickels, 2008a, 2008b;
Biran et al., 2013). These findings support Levelt et al.'s (1999) Two-Stage Model of spoken
language production (Figure 1a) where homophones share a phonological word form.
However, an alternative explanation for effects described above, is that homophone
generalisation is caused by feedback from phonemes to independent representations (as in
Figure 1c). Treatment of one word form would result in increased activation of the target’s
phonemes. Activation from these phonemes could feedback and activate other word forms
that also contain these phonemes, therefore the word form with the greatest amount of
phonological overlap would show the most generalisation, with homophones showing
greatest generalisation given their 100% overlap. However, Biedermann and Nickels (2008b)
rejected this feedback account as they found no evidence for differing degrees of
generalisation to phonologically related non-homophones dependent on number of shared
phonemes (indeed, there was no generalisation to phonologically similar items regardless of
their phonological similarity; see also Biedermann et al., 2018).
The architecture shown in Figure 1b does not predict generalisation to untreated
items, and therefore it cannot explain the homophone-specific generalisation found in the
phonological lexicon (Biedermann et al., 2002; Biedermann & Nickels, 2008a, 2008b; Biran et
al., 2013). Consequently, the major debate that remains is as to whether homophone
generalisation is caused by a shared word form (Figure 1a) or by feedback to independent

5

word form representations (Figure 1c). This disentangling of homophone representation is
one motivation for studying generalisation in the orthographic lexicon.
Treatment studies of written production
It is well established that there are separate and independent (but related)
orthographic and phonological output lexicons (e.g., Shelton and Weinrich, 1997). However,
as noted above, most research investigating the production of homophones has involved the
spoken modality- indeed few models address the representation of written homophones at
all (but see Caramazza, 1997). Therefore, in order to gain a comprehensive understanding of
homophone representation in the language system, it is critical to investigate whether the
orthographic lexicon is structured similarly to the phonological lexicon. If homophones share
a phonological word form, do they also share an orthographic word form? To our
knowledge, there has been only one previous study investigating orthographic treatment of
written homophone naming. Behrmann (1987) conducted a treatment study with a
participant (CCM) who had impaired access to the orthographic output lexicon, resulting in
poor irregular spelling with phonologically plausible errors and homophone confusions.
Before treatment, CCM scored 49% on homophone spelling to dictation (with a
disambiguating sentence) and a large proportion (57%) of errors were the homophone
partner (e.g. writing ‘sail for ‘sale’). It was hypothesised that CCM had difficulty retrieving
the homophone spelling on the basis of semantic information (accessing the orthographic
output lexicon from semantics). CCM was treated using a series of tasks that involved
contrasting the different spellings of homophones, using pictures and written sentence
completion tasks (e.g., she was shown a picture of a knight, with the written word knight,
which was orthographically contrasted with the word night, and then completed a written
sentence requiring the spelling of the target knight). The treatment improved homophone
spelling of the treated homophones and this generalised to better spelling of untreated
6

irregular words but not homophone partners. Behrmann (1987) suggested this was due to
improved orthographic lexical access (for the treated homophones) as well as an improved
visual-checking mechanism (generalisation to irregular words). This was supported by the
pattern of responses to untreated homophone partners: the percentage of non-word errors
was reduced (from 36% to 18%) and a larger percentage of errors were the (treated)
homophone partner (from 57% to 82%).
This is contradictory to previous (spoken) homophone treatment findings that found
homophone generalisation (e.g. Biedermann et al., 2002; Biedermann & Nickels, 2008a,
2008b; Biran et al., 2013). However, Behrmann's (1987) written treatment investigated
written orthographic production of items that were orthographically different (i.e.,
heterographic homophones, e.g., flour/flower). In contrast, the spoken homophone
treatments focused on phonological generalisation to items shared phonology (regardless of
orthography). It remains to be investigated whether treatment of words that share
orthography and phonology (homographic homophones), results in generalisation in written
naming. Further, some methodological weakness in Behrmann's (1987) treatment design
could also have led to reduced homophone generalisation. In particular, items selected for
treatment were spelled incorrectly at baseline whereas the majority of correctly spelled
items were assigned to the untreated group (30/50). Therefore, the untreated items had less
room to improve to show generalisation (and the treatment effects may have also been
magnified by ‘regression to the mean’, Howard, Best & Nickels, 2015).
There are other instances of differing patterns of generalisation across phonological
and orthographic treatment modalities. For example, while there has been a lack of
generalisation to phonological neighbours (e.g. cricket-ticket) in some spoken treatment
(e.g., Biedermann & Nickels, 2008b), generalisation to orthographic neighbours (e.g. clockblock) has been found in some written treatment (e.g., Harris et al., 2012).
7

Harris et al. (2012) and Sage and Ellis (2006) both conducted treatment studies with
individuals with aphasia who presented with a graphemic buffer impairment (the level at
which activation of graphemes is maintained before processing for output). Both studies
found that treatment generalised to untreated non-homophonic direct neighbours (pairs of
words with one grapheme different e.g. clock-block). The authors suggested this
generalisation could have been caused by the treatment increasing activation of the
orthographic representations of the treated items (e.g. clock), which in turn activated the
graphemes stored in the graphemic buffer (e.g., c-l-o-c-k). These activated graphemes would
have fed activation back to the word forms that shared these letters (e.g., the graphemes lo-c-k- would feed activation back to ‘block’) resulting in subsequent improved production of
both the target and neighbours. This suggests an orthographic network similar to the one
shown in Figure 1c. However, Krajenbrink, Nickels, and Kohnen (2017) failed to replicate this
direct neighbour generalisation in two similar cases with a graphemic buffer profile of
impairment. One possible explanation was that, in these individuals, perhaps there was
reduced or absent feedback from the graphemic buffer to the lexicon (Krajenbrink et al.,
2017).
Generalisation from treated to untreated spelling has also been found in cases of
developmental dysgraphia with no graphemic buffer profile impairment (e.g., Brunsdon,
Coltheart, & Nickels, 2005; Kohnen, Nickels, Coltheart, & Brunsdon, 2008). Kohnen et al.
(2008) found that this generalisation was dependent on orthographic neighbourhood. When
a word is spelled similarly to many of other words it has a large orthographic neighbourhood
(e.g., line has 22 ‘neighbours’, lime, lane, lint, pine etc), whereas a word has a small
orthographic neighbourhood if very few words differ in spelling by one letter (e.g., skull has
only one neighbour; skill). Similar to the studies with acquired dysgraphia, Kohnen et al.
(2008) suggested that untreated words with large orthographic neighbourhoods are more
8

likely to improve due to increased feedback from graphemes to orthographic word forms
that are repeatedly activated due to the large number of graphemes shared with other
words, including, most likely, the treated words.
This feedback mechanism, as the cause of generalisation to direct neighbours in
written production, is analogous to the alternative (feedback) explanation for generalisation
to untreated homophones in spoken production mentioned earlier (Biedermann & Nickels,
2008b, 2008a). However, if it were the case that generalisation in the spoken modality was
due to feedback, there should have also been generalisation to phonological neighbours
(which was not found), as was found to orthographic neighbours (e.g., Harris et al., 2012).
In sum, it is still unclear whether the effects of generalisation found for phonological
homophone treatment are due to feedback or shared-word forms, nor is the nature of the
orthographic representation of homophones clear. Therefore, we carried out a treatment
study that aimed to further investigate whether any generalisation in orthographic
homophone treatment is due to improved shared lexical entries, or feedback to separate
entries and how this differs across heterographic and homographic homophone spelling. We
thus treated three groups of stimuli (homographs, heterographs, non-homophonic controls)
and investigated generalisation to five untreated groups (homograph partners, heterograph
partners, direct neighbours of the controls, unrelated high orthographic neighbourhood
words and unrelated low orthographic words).
If the mechanism for generalisation is feedback from graphemes, then we would
expect to find generalisation to homographs as well as (less) generalisation to items that
have high (but not 100%) orthographic overlap (heterographs and direct neighbours), as in
Figure 1c. If we find generalisation only for homographic homophones, this will suggest they
share a word form representation, and there is no generalisation due to feedback from

9

graphemes as in Figure 1a. If no generalisation is found, this is consistent with homographs
having separate representations with strictly feed forward activation as in Figure 1b.
Case History
The participant in this study, CWS, was a 67-year-old right-handed, high school
educated, former builder from North Wales. He learnt both Welsh and English before the
age of six and still used both regularly. He reported that both pre- and post-stroke, he was
equally proficient in English and Welsh (i.e., he had no ‘dominant’ language), however, as
this treatment investigates English, only his English naming performance is reported (see
Roberts, 2013, for a comprehensive report of his bilingual language abilities). CWS suffered a
right frontal infarct in 1997 (18 years prior to this experiment). This resulted in left-sided
hemiplegia and crossed-aphasia (aphasia due to right hemisphere damage despite righthandedness) resulting in agrammatic, non-fluent speech. Table 1 shows CWS’s language
performance on a range of standardised (on English monolinguals) tests. CWS’s spoken and
written comprehension remains intact along with word and non-word repetition. CWS
performed within the control range for spoken object and action naming (Druks &
Masterson, 2000). His visual word recognition was just below ceiling and within control
performance range. Regular and irregular reading aloud were intact, although he showed
severely impaired non-word reading, a symptom pattern that is consistent with phonological
dyslexia (see Tainturier, Roberts, & Leek, 2011, for detailed analysis of his reading).
Table 1 about here
Table 2 about here
CWS’s spelling performance is reported in further detail in Table 2. Although CWS was
impaired in all aspects of spelling, he had a significantly larger deficit when spelling irregular
words compared to regular words and non-words (Chi squared, X2(1, N = 80) = 12.99,
p<.001, X2(1, N=80)= 10.90, p=.001). During written picture naming tasks (e.g. PALPA 53),
10

despite being asked not to name or spell aloud, CWS would attempt to spell by breaking a
word down into phonemes (which he could do accurately) and spelling one phoneme at a
time. This pattern of performance suggests attempted use of a sub-lexical strategy
secondary to damage to the orthographic output lexicon or access to this lexicon. This
strategy results in better performance on regular word spelling compared to irregular word
spelling and phonologically plausible (regularisation) errors. Table 3 shows examples of
errors taken from Roberts (2013), indeed 39% of CWS’s spelling errors were phonologically
plausible. However, CWS also produced a large number of phonologically implausible errors
(33%). To summarise, CWS presented with a mixed dysgraphia profile. He had a clear lexical
impairment, however the presence of a length effect and impaired nonword spelling also
suggests possible additional graphemic buffer and/or sub-lexical impairments (for a detailed
analysis of this see Roberts; 2013).
Table 3 about here
CWS’s orthographic word form level impairment made him a suitable candidate to
investigate generalisation of homophone treatment at the orthographic word form level.
Intervention Study
A copy and recall treatment (CART) (e.g. Beeson, 1999) in the presence of the picture
was conducted. This task was chosen to ensure that the treatment did not improve spelling
via a focus on phoneme to grapheme conversion but instead increased the accessibility of
the orthographic representations from the semantic system (as this was the representation
we were investigating). Previous treatment studies have found CART to be an effective
strategy to strengthen orthographic representations (Beeson, Hirsch, & Rewega, 2002).
Method
Stimuli.

11

Stimuli were picturable nouns presented as photographs 300 x 300 pixels in size,
displayed in the centre of a computer screen with written descriptions underneath.
Descriptions were designed to clarify picture identity to facilitate written naming. They were
not ‘stand-alone’ descriptions and did not contain any semantic competitors of the target
(e.g. the definition used with knife was ‘used for eating'). All stimuli (picture with the
description) had over 70% name agreement when named by 10 control participants (mean
age 29.20 years).
Stimuli belonged to one of eight experimental subsets: 1) Homographic homophones
(e.g., cricket [insect]), (2) homographic partners of 1 (e.g., cricket [sports]), (3) heterographic
homophones (e.g., sale), (4) heterographic partners of 3 (e.g., sail), (5) non-homophonic
controls with direct neighbours (e.g., bath), (6) direct orthographic neighbours of 5 (e.g.,
path), (7) non-homophonic control words with high orthographic neighbourhoods (e.g., line)
(8), and non-homophonic control words with very low orthographic neighbourhoods (e.g.,
church). The direct neighbours in subsets 5 and 6 consisted of words with one grapheme
substituted (they differed by one grapheme in the same position e.g., cake-cave; Coltheart,
Davelaar, Jonasson, & Besner, 1977). They did not include additions or subtractions.
Homophone and words with direct neighbours (subsets 1-6) were randomly assigned to two
sets (per condition i.e. 1 or 2, 3 or 4, 5 or 6). These sets were then adjusted to ensure
matching on the variables presented in Table 4 before randomly being assigned to treated or
untreated conditions. Values for all the psycholinguistic variables except frequency and
regularity were obtained from N-watch (Davis, 2005). CELEX (Baayen, Piepenbrock, &
Gulikers, 1996) was used to obtain frequency counts per million, then log10 transformations
were performed. An item was listed as irregular if it had at least one grapheme that was
classed as not the most frequent use by Perry, Ziegler and Coltheart (2002) or rare by Fry
(2004).
12

Table 4 about here
The treated and untreated subsets were matched for accuracy across the three
baseline sessions (see Table 4). The treated subsets were matched with their paired
untreated subset (i.e., subsets, 1 with 2, 3 with 4, 5 with 6) for all the psycholinguistic
variables shown in Table 4 (paired-sample t-tests, p=> .05; e.g., subset 1 had the same
frequency as subset 2). The untreated orthographic neighbourhood sets (subsets 7,
untreated high orthographic neighbourhood, and 8, untreated low orthographic
neighbourhood) were also matched to subset 5 (treated non-homophone, and therefore
also 6, untreated direct neighbours of subset 5) on all variables except the orthographic
neighbourhood variables for subset 8. As orthographic neighbourhood was manipulated in
subsets 7 and 8, the untreated low orthographic neighbourhood subset (subset 8), was
significantly different to subset 7 (and therefore 5 and 6) for any variables associated with
orthographic neighbourhood (regularity and phonological neighbourhood frequency).
Subset 7 also had significantly lower phonological neighbourhood density. None of the
subsets had any direct substitution neighbours in another subset apart from subset 5 that,
by design, had exclusively direct neighbours in subset 6. It was impossible to match number
of items (and variables) across pairs of sets, given the nature of the experimental items (i.e.,
the need for stimuli to be picturable homophones and picturable non-homophones with
picturable direct neighbours).
Procedure.
All subsets were presented for picture naming over three pre-treatment baseline tests
and three post-tests. Only stimuli from subsets 1, 3 and 5 were treated. As indicated in the
timeline shown in Figure 2, both written and spoken word production were tested during
pre-treatment baselines, once within the treatment phase and at post-tests. All assessments

13

were two weeks apart, except for the final two post-tests. The experimenter read out the
description as the picture was presented. In each session items were presented in a different
randomised order.
Spoken naming familiarisation.
Two days prior to baseline testing, two sessions of spoken naming familiarisation were
conducted, two days apart from each other. This was to ensure that CWS was familiar with
the picture names and to rule out any incorrect written responses due to ambiguous
pictures and to ensure that the phonology of each item was equally available. The spoken
familiarization phase consisted of presentation of the stimulus picture and spoken name for
CWS to repeat. The stimuli were split into two equal sets. During Session 1, Set 1 was
presented first with the correct name for CWS to repeat. The same items were subsequently
presented for CWS for uncued spoken naming, if he named an item correctly, he was given
feedback (e.g., ‘well done, that is correct’). If he produced the wrong name or no response,
the correct name was given for repetition (he was always able to repeat the item correctly).
Set 2 was then presented using the same procedure. Session 2 consisted of the same
procedure as Session 1 with Set 2 items presented. Overall, in total, each item was repeated
and named four times (twice in each session).
Assessment sessions.
Because of the large number of items, each assessment was split into two sessions.
Each set contained 84 experimental items and 16 filler items (which were included for data
collection for a separate study). During the first session of Baseline 1, CWS wrote the names
of Set 1, and completed spoken naming of the pictures of Set 2. During the second session of
Baseline 1 CWS was asked to write the names of Set 2 and complete spoken naming of Set 1.
Whether Set 1 or Set 2 was given for written naming in the first or second session of an
assessment alternated across time points. The two sessions assessing both modalities of one
14

set were at least one day apart. As we were primarily interested in investigating the effect of
treatment on written naming performance, and only secondarily in the effect of spelling
treatment on spoken naming (which was close to ceiling), all of our baselines and post-tests
assessed written naming before spoken naming. Each set was given in a different
randomised order at each presentation time point.
Figure 2 about here
Treatment consisted of six sessions over two weeks before an interim assessment and
then another six sessions over two weeks. Overall, CWS received 12 treatment sessions over
four weeks. Each treatment session contained 60 experimental items (i.e. the homograph,
heterograph and direct neighbour treatment sets) as well as 15 filler items and took roughly
1 and half hours. CWS also completed six homework sessions, one after every second
treatment session. The first post-test occurred two days after the last treatment session.
Due to CWS’ unplanned hospitalisation, post-test 2 and 3 were conducted later than planned
at three- and four-weeks post treatment. Consequently, analysis was conducted only
including one post-test, although all three post-tests are depicted in the figures below.

Treatment.
In order to promote lexical (rather than sublexical) spelling, treatment was based on
the copy and recall (CART) approach from Beeson (1999). CWS was presented with a
stimulus picture and correct spelling of the target name and asked to copy the word while
the word stayed in sight. This immediate copying was excellent, and this stage of the
treatment was error-free. The experimenter then covered both the presented correct
spelling and CWS’s immediate copy and counted to ten aloud. After this ten second delay
CWS was asked to ‘try and spell the name of the picture again’. The correct spelling was then
presented, and CWS was asked to judge whether he had correctly spelled the target item. If
he (correctly) confirmed that his delayed copy was correctly spelled, feedback was given,
15

and the next item was presented. If he misspelled the item and realised this was the case, he
was asked what part he thought was wrong. Then the target word was presented and
contrasted to his incorrect spelling, before he copied it once again. On the very few
occasions that he incorrectly judged his spelling as correct, the experimenter presented the
target item and pointed out the contrast, before he copied the correct target. On the rare
occasion he was unable to produce anything from delayed copying, the target was
presented again to copy. Therefore, each item was written correctly twice in each treatment
session. Homework also consisted of immediate copying and delayed recall. CWS was given
a booklet with one picture per page. On one page the written word was presented below
the stimulus picture for copying. This would be followed by another stimulus item for
copying, and then the first picture presented for delayed recall after this intervening item,
followed by the second item for delayed recall. As the homework was conducted
independently no feedback was given.
Analysis.
We examined the effects of treatment on performance using Weighted Statistics
(Howard, Best, & Nickels, 2015). We initially established if there was overall improvement
over the course of the study by conducting a trend (WEST-Trend) analysis. If there was
significant improvement, then we also conducted a Rate of Change (WEST-ROC) analysis to
investigate if this trend could be attributed to improvement during the treatment phase.
Only if both these tests showed significant positive change did we conclude there was
treatment-specific improvement (Howard et al., 2015). For written naming, we analysed
both whole word accuracy and letter accuracy using letter accuracy scoring adapted from
Buchwald and Rapp (2009). Each letter attracted a score of one if it was correctly produced
in the correct position. Between 0.25-1 points were deducted for each letter position that

16

was either transposed, migrated, substituted, missed or included an additional letter1. We
present the average letter accuracy in the results (following e.g. Buchwald & Rapp, 2009;
Krajenbrink et al., 2017 ; Sage & Ellis, 2006), however we also conducted analysis on the sum
of the letter scores (see Appendix A; Figure A1 and Table A1). We were concerned that the
average letter analysis would diminish improvement in longer words. For example, one more
letter correct in a three-letter word represents 33% improvement, but improvement by one
letter in a four-letter word results in 25% increase. However, as shown in Appendix A, there
was no difference in the results between the analysis using the average and that using the
sum letter scoring.
Results
We only report the spelling analysis below, as our primary interest is on written
naming. The results and analysis of spoken naming are presented in Appendix B (Figure B1
and Table B1). In brief however, spoken naming was close to ceiling (as was intended by the
familiarisation phase), and there was no statistically significant treatment-related change.
Due to the unplanned extended period between the last treatment session and posttests two and three, analysis with only one post-test is reported below (see Table 5).
However, analysis including all three post-tests is presented in Appendix C. All the effects
were in the same direction in both analyses; however, improvement was not maintained for
all items (as shown in Figures 3-6).

1 For example, when scoring KNIFE spelled as neafh the E has migrated from fifth position to the second, so it

was scored 0.5. When dealing with multiple errors in a single response, following Buchwald and Rapp (2009)
and Krajenbrink et al., (2017), the ‘visible’ transformation is scored:, we did not assume migration of
substituted letters, nor penalise the same position twice. For example, spelling WORD as whod, was scored by
penalising the addition of H between W and O (i.e. 0.5 score for W and the O) and deletion of R (0 points for R).
We would NOT, for example, assume that O and R transpose [(0.75 points each), and the R is also replaced with
H (reducing the 0.75 score to 0) as there is no way of knowing this. Instead only what is seen (that O has moved
and there is an additional H) is scored.

17

Figure 3 about here

There was significant improvement due to treatment both for word and letter
accuracy for the treated homographs (Figures 3a & b; and Table 5). However, there was no
treatment-related improvement for the untreated homographs, for either word or letter
accuracy analyses.
Figure 4 about here

The same pattern was found for the heterographs; significant improvement for the
treated items (both whole word, and letter analyses; Figure 4 and Table 5) but no
generalisation to the untreated items.

Figure 5 about here

The treated non-homophonic controls also improved due to treatment in both whole
word and letter analyses (as shown in Figure 5 and Table 5). In the whole word analysis, the
untreated direct neighbours of controls did show significant improvement, but this was not
replicated across the letter level analysis.
Figure 6 about here

There was no treatment-related improvement for whole word or letter accuracy for
either of the untreated orthographic neighbourhood control sets (Figure 6 and Table 5).

Table 5 about here

Discussion
The aim of this study was to investigate whether homophones have shared or separate
representations in the orthographic output lexicon by investigating generalisation of
treatment-related improvement to items with differing degrees of orthographic overlap. We
18

found that written naming of treated items improved in both whole word and letter analyses
and there was some generalisation to untreated items (in the whole word analysis). If the
effect of treatment was due to improved knowledge or application of phoneme-graphemeconversion rules or improved functioning of the graphemic buffer, generalisation to all items
that were not treated would have been expected. This did not occur, ruling out these nonlexical mechanisms underlying treatment-related improvement.
One means by which treatment can improve written naming is through strengthening
of orthographic representations (e.g. Rapp & Kane, 2002). This strengthening makes these
representations more accessible following treatment (Krajenbrink et al., 2017). This predicts
improvement of any orthographic representation that was activated during treatment. An
alternative account of treatment-related improvement proposes that prior activation (i.e.
during treatment) can improve access to the lexical representation by strengthening the
links to these representations from semantics (e.g., Howard, Hickin, Redmond, Clark, & Best,
2006; Howard, Nickels, Coltheart & Cole-Virtue, 2006; Wheeldon & Monsell 1992). We have
no means of distinguishing between these two accounts as the mechanism for improvement
in treated items for CWS, or indeed whether both mechanisms are at play. However, some
insights may be gained from the pattern of generalisation (see later).
The only untreated items that showed improvement were direct orthographic
neighbours of treated items (which were non-homophones), and this was restricted to the
whole word analysis. Sage and Ellis (2006) suggest that generalisation to direct orthographic
neighbours originates from treatment increasing activation of the treated word form (e.g.,
bath) which activates the letters (e.g., B, A, T, H). During treatment, letters that are shared
with a neighbour (e.g., for the neighbour path, the letters A, T, H), feed their activation back
to the orthographic word form of this neighbour (e.g., path). This in turn activates the
neighbour’s word-form representation and improves subsequent access to this item.
19

However, if feedback from shared letters was the mechanism underpinning generalisation to
direct neighbours in the present study, then treating homographs should also result in
generalisation to their untreated homophone partners (as homographs share 100% of
letters with their treated partners). This was not the case. Consequently, in the next section
we present the results of supplementary analyses that aimed to explore potential reasons
for this apparent contradiction: why might feedback result in generalisation to items that do
not overlap by 100% of their letters but not to items that do?
Further Analyses: Exploring generalisation to untreated items
In common with other similar investigations (e.g. Sage & Ellis, 2006, Harris et al.,
2012), we conceptualised orthographic overlap as overlap between matched pairs of items,
then subsequently predicted the extent of expected generalisation based on this measure.
However, any untreated item will not just receive feedback from its treated ‘mate’ but also
from any treated item with which it has orthographic overlap. In other words, the amount of
priming for an untreated lexical item following treatment relates to the number of times
that item is activated over the treatment set as a whole, not just from its experimentally
assigned partner. Hence, coal will not only be activated (and primed) when its
experimentally assigned direct neighbour partner coat is treated but also when cone, cow
and court are treated (but presumably to a lesser degree).
Consequently, to further understand what was driving generalisation in this study, we
carried out additional analyses examining whether there was an influence of the similarity to
other items that were treated (number of treated neighbours).
Method
To investigate the effect of similarity to all treated items on generalisation we
investigated two measures of orthographic overlap. ‘Orthographic Overlap’ was the average
of the orthographic similarity between each untreated word and all the items in the treated
20

set using the ‘ends-first’ spatial coding system within the match calculator application (Davis,
2007). For a more detailed explanation of the coding and algorithms used see Davis and
Bowers (2006), however, in brief, the match calculator outputs a weighting for each
untreated word compared to each treated word (one being an exact match e.g. bank-bank,
and zero having no letters in common e.g. drain-shoe). The second measure, ‘Number of
Treated Near Neighbours’ was the number of treated items that were zero, one and two
Levenshtein distance neighbours of the untreated items2 using the vwr package (Keuleers,
2013) in R. Levenshtein distance is the number of single-character steps (addition,
subtraction or substitution) it takes to transform one word into another. Therefore, for the
word bank (money), bank (river) is a zero Levenshtein distance neighbour, whereas band is a
one Levenshtein distance neighbour (substitute the k with d) and bat is a two Levenshtein
distance neighbour (remove the k and substitute the n with t).
When comparing the untreated direct neighbour and untreated homograph sets, the
direct neighbour set had significantly higher average Number of Treated Near Neighbours
than the homograph set (two-sample t-test, t(44)=1.98, p=.05, 4.33 vs 3.045 respectively).
However, this was not true for Orthographic Overlap (t(44)=-0.04, p=.96, 0.11 vs 0.11
respectively).
The dependent variables, the amount of treatment-related improvement, were the
WEST-ROC coefficients for whole word accuracy and average letter accuracy across the word
for every untreated item. Correlations were performed between these dependent variables
and the two new variables (orthographic overlap, and number of treated near neighbours)
and written lemma frequency, length (number of letters), regularity, number of
orthographic neighbours, average orthographic neighbour frequency, number of

2 We had controlled for the effect of orthographic neighbourhood by including a high and low orthographic

control set. However, this did not consider orthographic neighbourhood within our set of treated items.

21

phonological neighbours and average phonological neighbourhood frequency. All
continuous variables were centred. Any variables that significantly correlated with
improvement (n=108, p≤ .05, r ≥0.2) were entered into a stepwise linear regression using the
package gdata in R (Warnes et al., 2014) alongside any variables these (correlated) variables
also correlated with. No items that were correlated above .80 were included within one
regression model, to avoid potential multicollinearity (Hutcheson & Sofroniou, 1999).
Results
Word accuracy
On the basis of the correlations (see Appendix D) the first model for the whole word
accuracy regression included the effect of the variables length, regularity, number of
orthographic neighbours, number of phonological neighbours, orthographic overlap and
number of treated near neighbours on treatment-related improvement for whole word
accuracy (WEST-ROC; see Appendix E). Table 6 shows that after the backwards step-wise
regression, the only significant variable to predict WEST-ROC word accuracy scores was the
number of treated near neighbours. Figure 6 shows that as the number of treated near
neighbours increased, so did the amount of treatment-related improvement indexed by the
WEST-ROC coefficients.
Table 6 about here

Figure 7 about here
Letter scoring
As there were no significant correlations between improvement in letter accuracy and
any variable (see Appendix D), regressions were not performed for the average letter
scoring.
22

Discussion
The results of the regression analyses showed that, for word accuracy, only the
number of near neighbours in the treated set predicted generalisation for untreated items.
As direct neighbours had significantly more near neighbours than homographs in the treated
set, this likely accounts for why only generalisation to direct neighbours was significant.
These results suggest that when, during treatment, treated items activate their
corresponding letters, these letters feedback activation to other items in the orthographic
lexicon that share some of those letters. This results in the representations of those items
also being strengthened and more accessible. This mechanism of feedback from shared
letters leading to improved access to untreated orthographic representations is in line with
the interpretations of generalisation in previous studies (e.g., Kohnen et al., 2008; Sage &
Ellis, 2006). Moreover, this suggests that one mechanism underpinning improved spelling of
treated items is also improved accessibility of lexical representations, although it is not
possible to rule out additional treatment-related strengthening of semantic-lexical links.

General Discussion
This study examined homophone representation in the orthographic output lexicon;
specifically, whether homophones share a word form representation. This was investigated
by treating written picture naming using copy and recall treatment (in the presence of a
picture) and investigating generalisation to untreated written picture naming of items with
varying degrees of orthographic overlap. All treated items improved, however generalisation
only occurred in whole word analysis to direct neighbours of the treated items and not to
untreated homophone partners. Further analyses revealed that generalisation was
determined by the degree of orthographic overlap (number of near neighbours) with whole
treated set rather than simply the degree of orthographic overlap with the treatment
23

partner. In other words, an item was originally categorised by its similarity to its
experimentally assigned ‘treatment partner’ (e.g. homophones, bank-bank share all of the
same letters whereas direct neighbours car-ear only share 2/3 letters). However, our results
suggest that whether an untreated item will benefit from generalisation of treatment
depends not so much on this similarity, but rather on how much orthographic overlap it has
with all of the treated items regardless of experimental set (i.e. all 58 of the treated
heterographs, homographs and controls). For example, the untreated item chest had 100%
orthographic overlap with its experimentally assigned partner chest but had only 1 other
item in the whole treated set (treated heterographs, homographs and controls combined)
that was a ‘near orthographic neighbour’. In contrast, although the untreated item car only
had 66% overlap with its experimentally assigned treated partner ear, it had 11 near
orthographic neighbours in the whole treated set.
What do the results mean for homophone representation?
We found no generalisation from improved spelling of treated homophones to their
untreated partners. Instead, generalisation was dependent on how many near neighbours
were in the treated set, suggesting feedback from graphemes that increased accessibility of
word forms resulted in generalisation. This implies that homophones do not share an
orthographic word form, as if they did, improved accessibility of the word form would
predict both homophone partners (treated and untreated) should improve following
treatment.
This is not in line with previous work investigating homophone representations in
spoken production (Biedermann et al., 2002; Biedermann & Nickels, 2008a, 2008a; Biran et
al., 2013; Jescheniak & Levelt, 1994) which concluded that there was a shared homophone
representation in the phonological lexicon. However, there is some research to suggest that
having one close neighbour (i.e. a homophone) can be detrimental to production, whereas
24

larger numbers of more distant neighbours (as in this study) can be beneficial (Mirman,
Kittredge, & Dell, 2010). It is unclear how this theory for spoken word production (using
depth and width of phonological ‘basins’ as a method of word selection) would apply to
generalisation of orthographic therapy. Nevertheless, it does imply that having several near
neighbours can facilitate production more than having one identical word as a ‘neighbour’.
Why would homophone retrieval generalise in the phonological but not the
orthographic lexicon?
Previous treatment of homophones in spoken production has resulted in
generalisation (Biedermann et al., 2002; Biedermann & Nickels, 2008a, 2008a; Biran et al.,
2013), however, the present study did not find generalisation in written homophone
production. This could be due to one of three reasons.
Firstly, perhaps the phonological and orthographic lexicons are not organised
identically as we originally hypothesised: it is possible that (homographic) homophones
share a representation in the phonological output lexicon but have separate representations
in the orthographic output lexicon. At first it seems inconsistent to have different
architectures for phonology and orthography, however, Best, Herbert, Hickin, Osborne and
Howard (2002) point out that this seems more plausible when considering the considerable
differences between acquiring spoken and written language (e.g. age, method and ease of
acquisition). Nevertheless, there seems no logical reason as to why homographic
homophones would share phonological but not orthographic representations, in contrast
heterographic homophones require separate representations in the orthographic lexicon
(but not in the phonological lexicon), due to their different spellings.
Secondly it is possible that the differences between written homophone production in
the current study and spoken homophone production in the previous studies are in fact due
to differences in the participants’ responses to treatment. In order to fully test this
25

hypothesis, both a phonological and orthographic treatment should be carried out with the
same participant. If indeed homophones are represented as shared representations in the
spoken modality, but independent representations in the written modality, then
generalisation should occur for spoken naming treatment but not for written naming
treatment. However, this design was not an option for CWS as his spoken naming was close
to ceiling.
One crucial difference between CWS and the participants that undertook the
phonological homophone treatment in Biedermann et al.'s (2002; Biedermann & Nickels,
2008b; 2008a) studies is the number of languages spoken by each participant. While the
latter studies only included monolingual speakers with aphasia, CWS was an early bilingual,
highly proficient in Welsh and English. Bilinguals are known to have smaller vocabulary sizes
within each of their languages compared to monolingual speakers (e.g., Bialystok & Feng,
2009). Therefore, it is conceivable that a homophone representation is more likely to be a
non-homophone for a late bilingual speaker who might not know both homophone
meanings (e.g., due to their different frequencies), than a monolingual who has a larger
vocabulary and is familiar with both word forms. However, this seems unlikely for CWS who
was an early bilingual, growing up in a Welsh dominant household, but exposed frequently
to English in the community from an early age and using both languages daily throughout his
life. Nonetheless, it would be worthwhile investigating both orthographic and phonological
homophone treatment within a monolingual participant to rule out that the possible lack of

26

generalisation was not caused by some undetected non-native-like differences in the English
lexicon3.
Thirdly, it is possible that previous homophone generalisation in spoken production
was due to feedback from phonemes to independent word representations (as in Figure 1c
earlier). However, unlike the current study, previous spoken homophone generalisation
studies may have included more near neighbours of homophones in the treated set overall,
and relatively few for the phonologically related controls, hence, resulting in exclusively
homophone generalisation. Currently, no study has investigated spoken homophone
treatment generalisation as a function of overlapping phonemes with all treated items, but
our research suggests that this is vital.
Unfortunately, we are unable to currently determine which of the above possibilities
caused the difference in homophone generalisation depending on the production modality.
However, our results suggest that homophones do not share a representation in the
orthographic lexicon.
Why is there a difference between whole word and letter accuracy?
As we only found generalisation in the whole word analysis, it seems logical that
predictors of generalisation were only found in the whole word correlations and regressions.
However, letter analysis has often been suggested to be a more sensitive measure of change
(e.g. Krajenbrink et al., 2017), so why was no significant generalisation found using letter
scoring when there was with whole word scoring? It is possible that CWS was already
spelling the majority of the letters correctly (compared to whole words) resulting in less

3Gvion, Biran, Sharabi, and Gil (2015) conducted a phonological homophone treatment with a bilingual

participant, however, as this participant suffered from phonological output buffer impairment (not
phonological word form impairment), homophone generalisation was not predicted. In fact, no treatment
effects at all were found in this individual. Therefore, this particular case is uninformative in terms of
homophone representations and whether being bilingual can influence homophone representation and
generalisation.

27

room for improvement. Indeed, CWS’s scoring on the letter analysis was higher (averaged
over the three baselines was 64.07%) than the word analysis (16.67%) prior to treatment.
Perhaps one letter improvement in a few items was enough to result in spelling more items
100% correctly, therefore improving whole word accuracy (to 37.50%- an improvement of
20.83 %), but this was not enough to significantly improve letter accuracy (to 71.74%; an
improvement of only 9.08%). Indeed, of the 62% of the words CWS spelled incorrectly
(across all baselines) had one letter wrong (i.e., the majority of errors were one letter
incorrect) 23% of these were spelled correctly at post-test. This subset of words with one
letter wrong, had an average word score of 0% correct and an average letter score of 74%
correct before treatment. At post-test one 23% of these words are spelled correctly (23%
improvement) with a letter accuracy score of 80% (6% improvement). This shows how a
small letter improvement (6%) can result in a larger whole word improvement (23%)
explaining the difference between whole word and letter analysis. Clearly, it is not always
the case that letter scoring is a more sensitive measure of improvement. This also supports
Konhen et al ‘s (2008) finding that items that are more accurate before treatment were most
likely to show generalisation.
Clinical implications
This study was conducted first and foremost to provide insights into theories of
language production (homophone representation and models of written language
production). However, it also carries clinical implications. First, it replicates the previous
literature demonstrating that the Copy and Recall (CART) method can, in the short term,
improve spelling (e.g., Beeson, 1999; Beeson et al., 2002) and in this case with a participant
with impaired access to the orthographic lexicon. Secondly, our findings support previous
research that suggests that those items that are most likely to generalise are those from
larger neighbourhoods ( Brunsdon et al., 2005; Kohnen et al., 2008). However, while
28

replication is required, our research suggests that in order to maximise generalisation,
untreated items should be selected that have many neighbours within the treatment set
Conclusion
To conclude, this single case study shows that copy-and-recall-treatment can (at
least transiently) improve spelling of treated items in a participant with impaired access to
the orthographic lexicon. These effects generalised to items which have many near
neighbours in the treated set. While replication is required across a case series, our results
have implications for theories of orthographic processing and potential mechanisms
underlying treatment effects. The lack of homophone generalisation and the fact that
generalisation was predicted by number of near neighbours in the treatment set, suggests
that the orthographic lexicon is organised with separate homophone representations, but
that feedback from graphemes to word form representations that share graphemes is
possible (as implemented in models with interactive activation, e.g., Dell, Lawler, Harris, &
Gordon, 2004; Middleton et al., 2015, Figure 1c). If enough items are treated that share
graphemes with untreated items, this can result in generalisation of treatment via feedback
from these shared graphemes. More research is needed to determine if feedback also drives
generalisation in spoken word production.

29

Acknowledgement
The authors would like to thank CWS for his participation in this research. During the
preparation of this paper, LN was funded by an Australian Research Council Future
Fellowship [FT120100102 (2012-2018) and PB was funded by Macquarie University’s
international HDR scholarship].
Declarations of interest
The authors have no potential conflict of interest.

30

References
Baayen, R. H., Piepenbrock, R., & Gulikers, L. (1996). Celex2. Linguistic Data Consortium,
Philadelphia.
Beeson, P. M. (1999). Treating acquired writing impairment: strengthening graphemic
representations. Aphasiology, 13(9–11), 767–785.
https://doi.org/10.1080/026870399401867
Beeson, P. M., Hirsch, F. M., & Rewega, M. A. (2002). Successful single-word writing
treatment: Experimental analyses of four cases. Aphasiology, 16(4–6), 473–491.
https://doi.org/10.1080/02687030244000167
Behrmann, M. (1987). The rites of righting writing: Homophone remediation in acquired
dysgraphia. Cognitive Neuropsychology, 4(3), 365–384.
https://doi.org/10.1080/02643298708252044
Best, W., Herbert, R., Hickin, J., Osborne, F., & Howard, D. (2002). Phonological and
orthographic facilitation of word-retrieval in aphasia: Immediate and delayed effects.
Aphasiology, 16(1–2), 151–168. https://doi.org/10.1080/02687040143000483
Bialystok, E., & Feng, X. (2009). Language proficiency and executive control in proactive
interference: Evidence from monolingual and bilingual children and adults. Brain and
Language, 109(2–3), 93–100. https://doi.org/10.1016/j.bandl.2008.09.001
Biedermann, B., Blanken, G., & Nickels, L. (2002). The representation of homophones:
Evidence from remediation. Aphasiology, 16(10–11), 1115–1136.
https://doi.org/10.1080/02687030244000545
Biedermann, B., Cholin, J., Jorschick, A., Croot, K., Hameau, S., & Nickels, L. (2018). Is the
homophone advantage influenced by post-lexical effects? Cortex; a Journal Devoted
to the Study of the Nervous System and Behavior.
https://doi.org/10.1016/j.cortex.2018.01.014
31

Biedermann, B., & Nickels, L. (2008a). Homographic and heterographic homophones in
speech production: Does orthography matter? Cortex, 44(6), 683–697.
https://doi.org/10.1016/j.cortex.2006.12.001
Biedermann, B., & Nickels, L. (2008b). The representation of homophones: More evidence
from the remediation of anomia. Cortex, 44(3), 276–293.
https://doi.org/10.1016/j.cortex.2006.07.004
Biran, M., Gvion, A., Sharabi, L., & Gil, M. (2013). The Representation of Homophones in the
Phonological Lexicon: Additional Evidence from Hebrew. Procedia - Social and
Behavioral Sciences, 94, 86–87. https://doi.org/10.1016/j.sbspro.2013.09.041
Brunsdon, R., Coltheart, M., & Nickels, L. (2005). Treatment of irregular word spelling in
developmental surface dysgraphia. Cognitive Neuropsychology, 22(2), 213–251.
https://doi.org/10.1080/02643290442000077
Caramazza, A. (1997). How Many Levels of Processing Are There in Lexical Access? Cognitive
Neuropsychology, 14(1), 177–208. https://doi.org/10.1080/026432997381664
Caramazza, A., Costa, A., Miozzo, M., & Bi, Y. (2001). The specific-word frequency effect:
Implications for the representation of homophones in speech production. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 27(6), 1430–1450.
https://doi.org/10.1037/0278-7393.27.6.1430
Coltheart, M., Davelaar, E., Jonasson, T., & Besner, D. (1977). Access to the internal lexicon.
In S. Dornic (Ed.), Attention and Performance VI. Lawrence Erlbaum Associates.
Davis, C. (2007). Match calculator. Retrieved 18 April 2018, from
http://www.pc.rhul.ac.uk/staff/c.davis/utilities/matchcalc/index.htm
Davis, C. J. (2005). N-Watch: A program for deriving neighborhood size and other
psycholinguistic statistics. Behavior Research Methods, 37(1), 65–70.
https://doi.org/10.3758/BF03206399
32

Davis, C. J., & Bowers, J. S. (2006). Contrasting five different theories of letter position
coding: Evidence from orthographic similarity effects. Journal of Experimental
Psychology: Human Perception and Performance, 32(3), 535–557.
https://doi.org/10.1037/0096-1523.32.3.535
Dell, G. S. (1986). A spreading-activation theory of retrieval in sentence production.
Psychological Review, 93(3), 283–321. https://doi.org/10.1037/0033-295X.93.3.283
Dell, G. S. (1990). Effects of Frequency and Vocabulary Type on Phonological Speech Errors.
Language and Cognitive Processes, 5(4), 313–349.
https://doi.org/10.1080/01690969008407066
Dell, G. S., Lawler, E. N., Harris, H. D., & Gordon, J. K. (2004). Models of errors of omission in
aphasic naming. Cognitive Neuropsychology, 21(2–4), 125–145.
https://doi.org/10.1080/02643290342000320
Druks, J., & Masterson, J. (2000). An action and object naming battery. Hove: Psychology
Press.
Fry, E. (2004). Phonics: A Large Phoneme - Grapheme Frequency Count Revised. Journal of
Literacy Research, 36(1), 85–98. https://doi.org/10.1207/s15548430jlr3601_5
Gvion, A., Biran, M., Sharabi, L., & Gil, M. (2015). Learning from naming treatment on the
phonological representation of homophones. Presented at the American Speech
Hearing Association (ASHA) convention, Colorado, USA.
Harris, L., Olson, A., & Humphreys, G. (2012). Rehabilitation of spelling in a participant with a
graphemic buffer impairment: The role of orthographic neighbourhood in
remediating the serial position effect. Neuropsychological Rehabilitation, 22(6), 890–
919. https://doi.org/10.1080/09602011.2012.709872

33

Howard, D. (2000). Cognitive neuropsychology and aphasia therapy. In Acquired Neurogenic
Communication Disorders: A Clinical Perspective. Retrieved from
http://www.wiley.com/WileyCDA/WileyTitle/productCd-1861561113.html
Howard, D., Best, W., & Nickels, L. (2015). Optimising the design of intervention studies:
critiques and ways forward. Aphasiology, 29(5), 526–562.
https://doi.org/10.1080/02687038.2014.985884
Howard, D., Hickin, J., Redmond, T., Clark, P., & Best, W. (2006). Re-Visiting “Semantic
Facilitation” of Word Retrieval for People with Aphasia: Facilitation Yes But Semantic
No. Cortex, 42(6), 946–962. https://doi.org/10.1016/S0010-9452(08)70439-8
Hutcheson, G. D., & Sofroniou, N. (1999). The multivariate social scientist: Introductory
statistics using generalized linear models. Sage.
Jescheniak, J. D., & Levelt, M., J. (1994). Word frequency effects in speech production:
Retrieval of syntactic information and of phonological form. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 20(4), 824–843.
https://doi.org/10.1037/0278-7393.20.4.824
Jescheniak, J. D., Meyer, A. S., & Levelt, W. J. M. (2003). Specific-word frequency is not all
that counts in speech production: Comments on Caramazza, Costa, et al. (2001) and
new experimental data. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 29(3), 432–438. https://doi.org/10.1037/0278-7393.29.3.432
Keuleers, E. (2013). vwr: Useful functions for visual word recognition research. R package
version 0.3. 0.
Kohnen, S., Nickels, L., Coltheart, M., & Brunsdon, R. (2008). Predicting generalization in the
training of irregular-word spelling: treating lexical spelling deficits in a child. Cognitive
Neuropsychology, 25(3), 343–375. https://doi.org/10.1080/02643290802003000

34

Krajenbrink, T., Nickels, L., & Kohnen, S. (2017). No effect of orthographic neighbourhood in
treatment with two cases of acquired dysgraphia. Aphasiology, 31(5), 595–628.
https://doi.org/10.1080/02687038.2016.1218435
Levelt, W. J. M., Roelofs, A., & Meyer, A. S. (1999). A theory of lexical access in speech
production. Behavioral and Brain Sciences, 22(01), 1–38.
Middleton, E. L., Chen, Q., & Verkuilen, J. (2015). Friends and foes in the lexicon:
Homophone naming in aphasia. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 41(1), 77–94. https://doi.org/10.1037/a0037778
Mirman, D., Kittredge, A., K., & Dell, G. S. (2010). Effects of near and distant phonological
neighbors on picture naming. Proceedings of the 32nd Annual Conference of the
Cognitive Science Society.
Perry, C., Ziegler, J. C., & Coltheart, M. (2002). A dissociation between orthographic
awareness and spelling production. Applied Psycholinguistics, 23(1), 43–73.
https://doi.org/10.1017/S0142716402000036
Roberts, J. (2013). Cross-Lingisitic treatment generalisation in Welsh-English Bilingual
Aphasia (Unpublished doctoral dissertation). Bangor University, UK.
Sage, K., & Ellis, A. (2006). Using orthographic neighbours to treat a case of graphemic buffer
disorder. Aphasiology, 20(9), 851–870. https://doi.org/10.1080/02687030600738945
Schwartz, M. F., Dell, G. S., Martin, N., Gahl, S., & Sobel, P. (2006). A case-series test of the
interactive two-step model of lexical access: Evidence from picture naming. Journal
of Memory and Language, 54(2), 228–264. https://doi.org/10.1016/j.jml.2005.10.001
Shelton, J., R., & Weinrich, M. (1997). Further Evidence of a Dissociation between Output
Phonological and Orthographic Lexicons: A Case Study. Cognitive Neuropsychology,
14(1), 105–129. https://doi.org/10.1080/026432997381637

35

Tainturier, M.-J., Roberts, J., & Leek, E. C. (2011). Do reading processes differ in transparent
versus opaque orthographies? A study of acquired dyslexia in Welsh/English
bilinguals. Cognitive Neuropsychology, 28(8), 546–563.
https://doi.org/10.1080/02643294.2012.698986
Warnes, G. R., Bolker, B., Gorjanc, G., Grothendieck, G., Korosec, A., Lumley, T., … Rogers, J.
(2014). gdata: Various R programming tools for data manipulation. R Package
Version, 2(3), 35.

36

Appendices
Appendix A

Figure A1. Graphs to show written naming of the a) the homographs (subsets 1&2), b) the heterographs
(subsets 3 &4) c) non-homophonic controls (5&6) and d) the orthographic controls (7&8) for the summed letter
accuracy

Table A1. Results of the WEST t-tests and their significance for treatment across the three baselines and first post-test
for the summed letter accuracy.
Sum letter accuracy
Subset (degrees of freedom)

West trend

West-ROC

TSI

Consistent across
sum/average
letter
yes
yes
yes
yes
yes
yes
yes
yes

Homographs: 1. Treated (1,21)
1.88*
3.82**
✓
2.Untreated (1,21)
-0.42
✗
Heterographs: 3. Treated (1,13)
2.92**
4.38**
✓
4.Untreated (1,13)
1.01
✗
Non-homophone: 5. Treated (1,23)
2.91**
4.73**
✓
6.Direct N (1,23)
0.14
✗
Controls:
7.High ON (1,25)
-1.24
✗
8. Low ON (1,22)
-1.65
✗
*
p<.05 **p<.001
TSI= Treatment specific improvement
✓= both West-trend and West-roc were significant suggesting significant improvement over the course of the study and
during the treatment phase signifying treatment related improvement
✗= any change over the study could not be attributed to the treatment
Consistent across sum/average letter: is the same TSI shown if the analysis is conducted with the sum of letters correct
per word or the average letter percentage correct per word.

37

Appendix B

Figure B2. Graphs to show the spoken production of the a) the homographs (subsets 1&2), b) the heterographs
(subsets 3 &4) c) non-homophonic controls (5&6) and d) the orthographic controls (7&8)

Table B2. Results of the WEST one sample t-tests and their significance for treatment for spoken naming.
One Post-test
Three Post-test
Subset (degrees of freedom)

West trend
(t)
-0.22
0.33
2.28*
-0.41
-1.57
-0.95
-0.84
-1.23

West-ROC
(t)

TSI?

West trend
(t)
0.49
-0.25
2.25*
-0.78
-0.97
-1.52
-0.59
-1.18

West-ROC
(t)

TSI?

Consistent across
one/three posttests

Homographs: 1. Treated (1,21)
yes
✗
✗
2.Untreated (1,21)
yes
✗
✗
Heterographs: 3. Treated (1,13)
10.20**
1.68
no
✓
✗
4.Untreated (1,13)
yes
✗
✗
Non-homophone: 5. Treated (1,23)
no
✗
✗
6.Direct N (1,23)
yes
✗
✗
Controls:
7.High ON (1,25)
yes
✗
✗
8. Low ON (1,22)
yes
✗
✗
*
**
p<.05 p<.001 (one-tailed)
TSI= Treatment specific improvement
✓= both West-trend and West-roc were significant suggesting significant improvement over the course of the study and
during the treatment phase signifying treatment related improvement
✗= any change over the study could not be attributed to the treatment
Consistent across one/three post-tests: is the same TSI result shown in analysis that include one or three post-tests.

38

Appendix C
Table C1. Results of the WEST one sample t-tests and their significance for treatment across the three baselines and three
post-tests for written naming.
Whole word accuracy
Average letter accuracy
Consistent across
word/letter
Subset (degrees of freedom)
West trend West-ROC TSI? West trend West-ROC
TSI?
(t)
(t)
(t)
(t)
Homographs: 1. Treated (1,21)
2.53**
4.62**
1.72
no
✓
✓
2.Untreated (1,21)
1.11
0.29
yes
✗
✗
Heterographs: 3. Treated (1,13)
2.74*
4.60**
1.41
no
✓
✓
4.Untreated (1,13)
-0.12
-1.00
yes
✗
✗
Non-homophone: 5. Treated (1,23) 2.87 **
6.15 **
0.77
no
✓
✓
6.Direct N (1,23)
1.30
-1.72
yes
✓
✗
Controls:
7.High ON (1,25)
0.04
-0.49
yes
✗
✗
8. Low ON (1,22)
-0.38
-1.07
yes
✗
✗
*
**
p<.05 p<.001 (one-tailed)
TSI= Treatment specific improvement
✓= both West-trend and West-roc were significant suggesting significant improvement over the course of the study and
during the treatment phase signifying treatment related improvement
✗= any change over the study could not be attributed to the treatment
Consistent across word/letter: is the same TSI result shown across both word and letter accuracy

39

Appendix D
+

Table A4. The Pearson’s correlation r value and significance for all the psycholinguistic variables and the orthographic overlap, number of neighbours in the treated set and treatment
effect for the accuracy analysis. Items in italics indicate which variables significantly correlated with generalisation (WEST-ROC) and their co-variables and therefore were included in the
regression
Frequency

Length

Regularity

OrthN-freq

OrthN

PhonoN-freq

PhonN

Length

0.02

Regularity

-0.07

-0.26*

OrthN-freq

0.01

-0.67**

0.35**

OrthN

-0.11

-0.3*

0.1

PhonN-freq

-0.03

**

-0.49

*

0.22

0.66**

0.05

PhonN

0.1

-0.19

-0.02

0.01

0.65**

0.02

0.11

*

*

-0.09

0.35**

-0.06

**

Orth-O

-0.22

Orth-O

Neighbours

0.14

0.26

0.28

**

*

**

Neighbour

0.02

-0.51

0.27

0.61

0.18

0.44

0.15

0.29*

WEST-ROC - Whole word

0.1

-0.08

0.02

0.15

-0.04

0.07

0.00

0.06

0.20*

0.01

0.05

-0.04

WEST-ROC – Average letter
-0.04
0.13
0.03
-0.06
0.06
-0.06
Due to negative correlation the reverse end of the p-value was taken
*
r ≥ 0.2 plus p<.05
**
r ≥ 0.2 plus p<.001
Frequency= written lemma frequency
Length= number of letters
Regularity= whether a word was spelled regularly or not
OrthN-freq= the number of orthographic neighbours
OrthN= the average frequency of the orthographic neighbours
PhonN-freq= the number of phonological neighbours
PhonN= the average frequency of the phonological neighbours
Orth-O= measure of orthographic similarity between each untreated word and every item in treated set using the ‘ends-first’ coding system
Neighbours= the number of items that were zero, one and two Levenshtein distance neighbours of the treated set in the untreated set
+

40

Appendix E
First regression model: WEST-ROC (whole word accuracy) ~ length + regularity + orthN-freq + PhonoNfreq + Orth-O + Neighbours
Sum of sq
<none>

RSS

AIC

325.49

137.15

F value

P value

length

0.82

328.36

132.09

0.25

0.62

regularity

0.72

328.27

132.06

0.22

0.64

orthN-freq

2.06

329.60

132.50

0.63

0.43

PhonoN-freq

0.41

327.96

131.96

0.13

0.72

Orth-O

0.06

327.60

131.84

0.02

0.89

Neighbours
7.51
335.06
134.27
2.32
0.13
Length= number of letters
Regularity= whether a word was spelled regularly or not
OrthN-freq= the number of orthographic neighbours
PhonN-freq= the number of phonological neighbours
Orth-O= measure of orthographic similarity between each untreated word and every item in treated set
using the ‘ends-first’ coding system
Neighbours= the number of items that were zero, one and two Levenshtein distance neighbours of the
treated set in the untreated set

41

List of figure captions
Figure 1: Three architectures for homophone representation in the phonological lexicon a) separate lemmas
and shared (Levelt et al., 1999) b) no lemma level and separate modality specific word forms (Caramazza et al.,
2001) and c) lemmas but no word form representations and interactive activation (Middleton et al., 2015;
Schwartz et al., 2006).
Figure 2: Timeline of baselines (B1, B2, B3), treatment sessions (T1-12) and post-tests (P1, P2, P3).
Figure 3: Accuracy for homographs on whole word accuracy (A), and average letter accuracy (B)
Figure 4: Accuracy for heterographs on whole word accuracy (Panel A), letter percent accuracy (Panel B)
Figure 5: Accuracy for non-homophonic controls and direct neighbours) on whole word accuracy (Panel A), and
average letter accuracy (Panel B).
Figure 6: Accuracy for orthographic neighbourhood controls on whole word accuracy (Panel A), and average
letter accuracy (Panel B).
Figure 7. A boxplot to show the relationship between the number of 0, 1 and 2 Levenshtein distance
neighbours and treatment related-improvement

42

Tables
Table1. English Language background assessment for CWS
Task

Number
of items

CWS
correct
%

Control
Mean
%

Control
Min.
%

Control
SD
%

N of
Control
s

Comprehension
PALPA 47 oral word-picture matching
40
100%
98
87
2.67
31
PALPA 48 written word-picture matching
40
100%
99
87
1.53
32
Pyramids and Palm Trees Test; three pictures
52
96%
98
94
13
Single word Repetition
Word repetition (Bangor University)
80
100%
English Non-word repetition (Bangor University)
40
100%
Spoken Naming
Object Naming battery (list B)
81
91%
91
40
Action Naming battery (list B)
50
86%
86
40
Visual Word Recognition
PALPA 25: Real and non-words lexical decision++
120
98%
99
0.54
26
PALPA 3 minimal pairs: Written word selection
72
97%
97
2.35
23
Reading
PALPA 19 upper case to lower case letter matching
26
96%
100
96
0.77
26
PALPA 32 grammatical class reading
80
95%*
100
1.45
32
PALPA 34 lexical morphology and reading
90
90%*
40
98%
99
95
1.65
20
English Reading: Regular words (Bangor University)+
40
100%
99
95
1.73
20
English Reading: Irregular words (Bangor University)+
+
40
43%**
95
83
5.10
20
English Reading: Non-words (Bangor University)
PALPA 53: Cross Modality Comparisons
Oral picture naming
40
65%**
99
0.87
29
Irregular
20
80%
Regular
20
60%
Written picture naming
40
15%**
97
3.33
27
Irregular
20
5%
Regular
20
25%
Spelling to dictation
40
7.5%**
99
2
Irregular
20
0%*
Regular
20
15%*
Repetition
40
97.5%
99
2.05
28
Irregular
20
100%
Regular
20
95%
+
Control scores taken from aged matched control monolingual participants from Bangor University. All other control data is
from the appropriate published test.
++
Average mean and standard deviations across the subsets
Bold represents scores which are impaired (2.5 standard deviations below control mean)
** Scores that are at least two standard deviations below control mean.
* Scores that are thought to be impaired to some degree, but normative data is not available.

43

Table 2. In-depth spelling-to-dictation assessments
Task

Number
of
stimuli

CWS %
correct

Control
Mean
%

Control
Range
%

Control
SD
%

Regularity (Bangor University)
Regular words
80+
41
97
80-100
5.23
Irregular words
80+
18
91
55-100
12.40
Non-words
80+
39
74
53-93
12.07
Frequency
(Words collapsed across JHU# lists)
High-frequency words
147
21*
Low-Frequency words
146
12*
Length (JHU list length)
4-5 Letters
27
30*
99
6 letters
15
7*
92
7+ letters
28
7*
93
Grammatical Category (JHU part-ofspeech)
Nouns
28
4*
Verbs
28
4*
Adjectives
28
7*
Nonwords
34
12*
Concreteness (JHU)
Concrete words
21
19*
98
Abstract words
21
0*
91
Copy
Direct copy (PALPA 44)
40
98
Delayed copy transcoding
Regular words
20
80*
Irregular words
20
45*
#
JHU= John Hopkins University Dysgraphia Battery (Goodman & Caramazza, 1985)
+
Double administration for CWS, therefore control number of items = 40
Impaired Scores in bold (2.5 standard deviations below control mean)
* Scores that are thought to be impaired to some degree, but normative data is not available.

44

Control
N
20
20
20
5
5
5
5
5
-

Table 3. Errors made by CWS in spelling to dictation taken from Roberts (2013)
Error type

Example

Phonologically plausible errors
Real word error
Phonologically implausible nonwords (50% or more letters
correct)
Phonologically implausible nonwords (less than 50% target
letters correct)
Cross language errors

45

Into-> INTU
Work->WORD
Hotel->HOTOL

Words %
(N= 599)
39
7
33

Non-words %
(N=45)
1
36
7

Feather->FAFARA

8

1

Nine->NAIN

12.52

2.22

Table 4. Matching of experimental subsets on accuracy, log frequency and other psycholinguistic variables.
Homographs
Heterographs
Subset
1
2
3
4
5
Treated
Untreated
Treated
Untreated
Treated
(N=22)
partners
(N= 14)
partners
(N=24)
(N=22)
(N=14)
Written accuracy Baseline 1
9.09
9.09
14.29
7.14
16.67
Written accuracy Baseline 2
13.64
9.09
21.43
21.43
25.00
Written accuracy Baseline 3
0.00
4.55
14.29
14.29
12.50
Spoken accuracy Baseline 1
95.45
77.27
100.00
85.71
83.33
Spoken accuracy Baseline 2
59.09
63.64
64.29
85.71
70.83
Spoken accuracy Baseline 3
92.86
66.67
81.82
72.73
71.43
Frequency: written lemma (log10)
2.06
2.20
2.41
2.62
2.67
Frequency: written word form (log10)
0.57
0.71
1.06
1.19
1.05
Frequency: spoken lemma (log10)
0.69
0.84
1.17
1.29
1.17
Frequency: spoken word form (log10)
1.88
2.01
2.23
2.50
2.48
Syllables
1.14
1.14
1.07
1.07
1.00
Phonemes
3.64
3.64
3.00
3.00
3.09
Letters
4.32
4.32
4.21
4.36
3.95
Orthographic neighbourhood density
7.41
7.41
7.36
8.57
10.91
Orthographic neighbourhood freq.
106.61
106.61
62.27
385.25
86.53
Phonological neighbourhood density
16.05
16.05
21.29
21.29
21.09
Phonological neighbourhood freq.
86.14
86.14
451.21
451.21
184.22
Regularity
0.91
0.91
0.57
0.50
0.92
*Significantly different from matched subset
Orthographic neighbourhood density = number of words with one letter difference
Orthographic neighbourhood freq. = average frequency of all the orthographic neighbours
Phonological neighbourhood density = number of words with one phoneme difference
Phonological neighbourhood freq.= average frequency of all the phonological neighbours

46

Non-homophonic controls
6
7
Untreated Direct
Untreated
neighbours of 5
High ON
(N=24)
(N=26)
16.67
25.93
16.67
22.22
8.33
23.63
79.17
88.89
70.83
77.78
75.00
81.48
2.76
2.86
0.96
1.15
1.14
1.25
2.59
2.74
1.00
1.00
3.25
3.38
4.00
3.88
10.21
8.62
129.83
310.67
20.50
14.92*
306.20
279.22
0.92
0.81

8
Untreated
Low ON
(N=22)
4.55
4.55
4.55
72.73
72.73
68.81
2.67
0.85
1.00
2.49
1.00
3.19
5.05*
0.57*
27.82*
11.62*
286.50
0.55*

Table 5. Results of the WEST one sample t-tests and their significance for treatment across the three baselines and first posttest.
Whole word accuracy
Average letter accuracy
Consistent across
word/letter
Subset (degrees of freedom)
West trend West-ROC TSI? West trend West-ROC
TSI?
(t)
(t)
(t)
(t)
Homographs: 1.Treated (1,21)
3.25**
5.25**
2.38*
3.91**
yes
✓
✓
2.Untreated (1,21)
0.62
-0.66
yes
✗
✗
Heterographs: 3.Treated (1,13)
3.38**
5.12**
3.25**
4.41**
yes
✓
✓
4.Untreated (1,13)
1.10
0.94
yes
✗
✗
Nonhomophone: 5.Treated (1,23)
4.26 **
8.26 **
3.20**
5.07**
yes
✓
✓
6.Direct N(1,23)
1.73*
1.83 *
1.00
no
✓
✗
Controls:
7.High ON(1,25)
0.46
-1.19
yes
✗
✗
8.Low ON(1,22)
-1.00
-1.81
yes
✗
✗
*
p<.05 **p<.001 (one-tailed)
TSI= Treatment specific improvement
✓= both West-trend and West-roc were significant suggesting significant improvement over the course of the study and
during the treatment phase signifying treatment related improvement
✗= any change over the study could not be attributed to the treatment
Consistent across word/letter: is the same TSI result shown across both word and letter accuracy

47

Table 6. Final regression model for the whole word analysis
Variable
DF
Sum of sq
RSS
Intercept
330.32
Neighbours
1
13.42
343.74

AIC
124.74
127.04

F value

P value

4.31

.04*

* Significant p-value of <.05
Neighbours= the number of items that were zero, one and two Levenshtein distance neighbours of the treated set in the untreated set

48

Figures

Figure 1

Figure 2

49

Figure 3

Figure 4

Figure 5

50

Figure 6

Figure 7

51