Holistic word processing in dyslexia

Aisling Conway; Nuala Brady; Karuna Misra

doi:10.1371/journal.pone.0187326

. 2017 Nov 9;12(11):e0187326. doi: 10.1371/journal.pone.0187326

Holistic word processing in dyslexia

Aisling Conway ¹, Nuala Brady ^1,^*, Karuna Misra ¹

Editor: Philip Allen²

PMCID: PMC5679523 PMID: 29121046

Abstract

People with dyslexia have difficulty learning to read and many lack fluent word recognition as adults. In a novel task that borrows elements of the ‘word superiority’ and ‘word inversion’ paradigms, we investigate whether holistic word recognition is impaired in dyslexia. In Experiment 1 students with dyslexia and controls judged the similarity of pairs of 6- and 7-letter words or pairs of words whose letters had been partially jumbled. The stimuli were presented in both upright and inverted form with orthographic regularity and orientation randomized from trial to trial. While both groups showed sensitivity to orthographic regularity, both word inversion and letter jumbling were more detrimental to skilled than dyslexic readers supporting the idea that the latter may read in a more analytic fashion. Experiment 2 employed the same task but using shorter, 4- and 5-letter words and a design where orthographic regularity and stimuli orientation was held constant within experimental blocks to encourage the use of either holistic or analytic processing. While there was no difference in reaction time between the dyslexic and control groups for inverted stimuli, the students with dyslexia were significantly slower than controls for upright stimuli. These findings suggest that holistic word recognition, which is largely based on the detection of orthographic regularity, is impaired in dyslexia.

Introduction

Developmental dyslexia is characterised by marked difficulties in learning to read that cannot be attributed to low intelligence, poor motivation or to lack of education and opportunity to read [1]. In standardised measures of reading ability the scores of children with dyslexia are found to scatter on the lower tail of a normal distribution, suggesting that dyslexia is not a discrete or easily defined diagnostic entity [2]. Clinical reviews paint a similarly complex picture with noted overlap between dyslexia, language impairment and speech sound disorder [3].

Fluent word recognition—the effortless mapping between the sight of a printed word and its sound or pronunciation—is central to reading and develops through the acquisition of two distinct skills, phonological and orthographic coding [4]. Phonological coding refers to the representation of grapheme-phoneme correspondences–the matching between letters and their associated sounds—and is crucial to the reading of new and unfamiliar words. Orthographic coding refers to the representation of the visual form of words which includes letter combinations at both coarse and fine scales that signal spelling regularities and that help identify words without the need to access phonological information at the sub-word level [5]. Reading fluency reflects expertise in this visual or direct-access route to word sounds. For example, whereas the time to name a word increases with word length for children who are learning to read and still ‘sounding out’ the words, word naming latency is largely independent of word length in fluent adult readers [6].

Bruck [7]argues that poor phonological processing—a central characteristic of dyslexia [8]—prevents dyslexic readers from accurately noting spelling-sound correspondences in reading which, in turn, prevents them from learning precise orthographic information about words. Ironically, it is their poor knowledge of spelling-sound correspondences that keeps dyslexic readers using these very cues to recognize words, arresting their reading development and preventing them from eventually achieving fluency via the automatic use of orthographic or visual word form cues to recognize words and to access their sounds directly.

Bruck’s idea is an interesting one that long predates current knowledge of the neural basis of visual expertise in word recognition. The visual word form area (VWFA), located in mid-fusiform gyrus in the left hemisphere, is a region of the brain that is activated during reading and that responds preferentially to words over consonant strings, chequerboard patterns [9] and line drawings of objects [10]. Activation of the VWFA is constant across changes in the retinotopic location of word stimuli [9]and across changes in their typographic case and size [11,12]. This response invariance to stimulus properties that are unimportant to word recognition coupled with a robust response to pseudowords—non-words that mimic the spelling structure of real words—suggests that the VWFA represents words in a way that captures their orthographic regularity rather than their physical form or associated semantics [12,13]. In fact, in studies using fMRI adaptation the VWFA shows greater selectivity—to whole words rather than to sub-lexical orthographic patterns—suggesting that the region may serve as a form of ‘visual dictionary’ [14,15].

Decreased brain activity in the VWFA has been reported in dyslexia [16–18]. Using a cross-sectional design to investigate changes associated with reading development, Shaywitz and colleagues [17] report that while normal readers show increased activation in the VWFA with age, dyslexic readers instead show increased activation of a nearby (posterior medial occipitotemporal) region. They suggest that increased VWFA activation in normal readers reflects changes associated with increasing fluency in mapping word forms to associated sounds, whereas increased activation in the neighbouring region in dyslexic children reflects increasing reliance on memorization of the visual word form. Others [18] show that while the VWFA is specialized for orthographic regularity at both a coarse scale (distinguishing print from non-print fonts) and at a finer scale that taps into spelling regularity in normal readers, the VWFA in dyslexic readers does not make this distinction. Together these studies point to anomalous development of the VWFA in dyslexia and implicate impaired sensitivity to orthographic regularity as a source of impaired fluency in reading.

We suggest that this insensitivity to orthographic regularity in dyslexia might usefully be conceptualized as a problem in acquiring a particular type of perceptual expertise, one that relies on holistic processing of words. Broadly speaking, perceptual expertise involves discriminating instances of a category—such as recognizing individual words or faces—and is thought to engage distinct processes from those involved in more basic categorization as when we identify a visual object as a word or as a face [19]. As noted by a number of researchers, words and faces share a characteristic that poses a particular challenge for their individual identification, specifically, their high self-similarity [20,21]. Individual words are made by arranging a fixed number of letters from a limited alphabet and individual faces share a common arrangement of features with each other.

In the case of faces, identification of individuals is thought to involve holistic or configural processing [19] and a number of experimental paradigms provide support for this idea. In the ‘composite face effect’ careful alignment of the top half of one face with the bottom half of another leads us to perceive a novel facial configuration that slows or prevents recognition of the component faces [22]. Central to this effect is an inability to selectively attend to one part of a face and ignore information from other, a type of ‘obligatory attention’ to all parts of the stimulus [21,23]. The ‘face inversion effect’ links holistic processing to the upright orientation of a face and shows that face recognition [24] and sensitivity to distortions of facial shape [25] are compromised when faces are inverted. Finally, Tanaka and Farah’s variant of the ‘face superiority effect’—which shows that facial features are more easily recognized within the context of a face than when presented alone or within a face with spatially scrambled features [26]–also points to the importance of facial configuration in face recognition.

These hallmarks of expertise in face recognition–an obligatory attention to the configuration of facial features that is specific to the typical, upright orientation of faces, and a superior ability to recognize facial features within the context of a face—have parallels in the study of word recognition. The ‘obligatory attention’ to all parts of the stimulus that is central to the composite face effect has recently been demonstrated for words. Wong and colleagues [21] asked participants to match the right or left halves of a pair of words while ignoring the other halves on both congruent trials, (both halves of the pair of words are matched), or incongruent trials (one half of the pair is matched, the other is mismatched). They found that performance was better on congruent than incongruent trials, showing interference from the supposedly ignored halves of the words. This ‘composite word effect’ was stronger for highly familiar words, i.e., those from participants’ first compared to their second language and for real words compared to pseudo-words [21].

Secondly, there is evidence that words are read in a holistic manner over the range of orientations at which we normally encounter them. A seminal study of reading rotated words suggests that words are read ‘holistically’ at orientations at or close to the upright but in a more piecemeal or analytic fashion when rotated by 180 degrees [27]. Response times to recognize words are independent of word length between 0° and ~ 60°, increase with increased word length in a manner suggesting piecemeal reading of sub-lexical units, with a final shift to very slow letter by letter reading between beyond ~120°. This ‘word inversion effect’ holds for words presented in their familiar format but not for words written backwards or for mirror reversed words [28]. Finally, the face superiority effect has obvious parallels with the ‘word superiority effect’ which shows that the recognition of single letters is more efficient when they are embedded in real words than when presented alone or embedded in meaningless string of letters [29].

The current study utilizes design aspects of both the word inversion and the word superiority paradigm to investigate the idea that adult dyslexics are somehow less tuned to the visual form of words than skilled adult readers using a task that does not require reading. We asked participants to judge whether a pair of real words or a pair of letter strings (formed by jumbling the letters of the words) were the same or differed by a single letter. The word-superiority effect [29] predicts that performance should be best in the real word condition. Additionally, we asked participants to make this perceptual judgement for both upright words and letter strings and for the same stimuli which had been inverted. The ‘word inversion effect’ predicts for our study superior performance in the upright condition. To preview our results, we found that both adults with dyslexia and skilled adult readers preformed as predicted, showing fastest response times for upright real words and slowest for inverted letter strings. Most interestingly, the effects of inversion and degradation of orthographic regularity were more detrimental to the skilled readers than to those with dyslexia, supporting the idea that dyslexic readers may continue to read in a more analytic fashion into adulthood and never attain full automaticity through the exclusive use of visual word form cues [7].

Experiment 1

Methods

Participants

The experimental group comprised 30 students (17 females) with a mean age of 23.6 years (SD = 6.07 years) and age range of 18 to 43 years who were registered with University College Dublin (UCD) Disability Support Services as dyslexic, having provided evidence of diagnosis by an educational or clinical psychologist on registration. Diagnoses of dyslexia accepted by the university are based on analysis of performance on a battery of standardized norm-based tests including (a) word reading: fluency, speed and accuracy (b) word comprehension and spelling, (c) phonological awareness, (d) working memory and (e) rapid automatic naming. To qualify for special arrangements that are put in place by Disability Support Services students must score at or below the 10^th percentile on at least two literacy tests from the following (reading accuracy, reading speed, single word reading, spelling, reading comprehension, writing speed and pseudo-word decoding). Of the 30 participants recruited, 4 reported other diagnoses, 2 for Specific Language Impairment in childhood, and 1 for ADHD and 1 for Tourette Syndrome, both concurrent. The control group comprised 30 students (15 females) with a mean age of 23.1 years (SD = 5.07 years) and age range of 18 to 44 years, also from UCD. The groups did not differ significantly in age, t = 0.35, df = 56.23, p = 0.73. The study was approved by the Human Research Ethics Committee (HREC) at UCD (http://www.ucd.ie/researchethics); in accordance with the Declaration of Helsinki all participants gave written, informed consent and were advised of their right to withdraw from the study at any time without prejudice.

Stimuli

The word stimuli were generated from 15 high-frequency 6-letter and 15 high-frequency 7-letter single morpheme American English words taken from http://slpath.com/ and based on the Kučera & Francis database [30]. The 30 words are shown in Table 1 alongside their SUBTLEX frequency per million words [31] where they all, bar one, also qualify as ‘high-frequency’ using the criterion of a frequency higher than 10 per million. The stimuli were presented in pairs across 4 conditions (upright-real, upright-jumbled, inverted-real, inverted-jumbled) in a task in which participants judged whether the stimuli were the same or different.

Table 1. The 30 high frequency words used in Experiment 1 to generate the experimental stimuli alongside their SUBTLEX frequency per million words.

Six letter words	SUBTLEX frequency per million words	Seven letter words	SUBTLEX frequency per million words
action	61.08	against	201.73
almost	193.27	already	318.55
Always	655.25	another	509.22
coarse	1.35	certain	85.37
either	182.51	country	161.84
enough	501.33	general	115.39
family	354.25	history	83.92
number	240.94	perhaps	136.06
public	71.08	present	89.45
rather	114.22	problem	330.06
result	19.76	quality	18.57
second	284.57	several	38.78
social	33.39	service	79.92
should	1061.94	through	549.53
system	91.51	weather	34.24

Open in a new tab

In the upright-real condition, each word was either paired with itself onscreen (e.g., ‘another’, ‘another’) on the ‘same’ trials, or with a variant of itself in which a single letter was changed (e.g., ‘another’, ‘arother’) on the ‘different’ trials, the original word always appearing to the left and the modified or misspelled word to the right. Following [32], changes were made that best preserved the overall word shape using letter substitutions from their Table 2, with a further stipulation that substitutions were always made within the word, the first and last letter never changing. For the upright-jumbled condition, the same 30 words were used to create jumbled or scrambled letter strings, with the same letter substitutions being used on ‘different’ trials as were used in the upright-real condition, thus (‘another’, ‘another’) became (‘hanrtoe’, ‘hanrtoe’) for the equivalent ‘same’ trial and (‘hanrtoe’, ‘hahrtoe’) for the equivalent ‘different’ trial. Finally, in the inverted-real and inverted-jumbled conditions, the entire stimulus was rotated through 180° so that the letters were upside-down and the words reversed left to right.

Table 2. The 30 high frequency words used in Experiment 2 to generate the experimental stimuli alongside their SUBTLEX frequency per million words.

Four letter words	SUBTLEX frequency per million words	Five letter words	SUBTLEX frequency per million words
area	74.92	argue	19.75
clue	17.61	baron	13.47
date	141.53	chase	32.80
drop	130.61	clown	15.82
dump	28.82	cloud	11.75
duck	24.76	enemy	48.51
dish	11.45	empty	47.24
exam	13.41	fault	104.12
foul	14.47	honor	96.31
fund	10.61	nurse	44.98
gold	78.94	prize	22.39
gate	32.04	pound	13.88
lawn	12.35	roses	14.18
rent	34.55	stage	45.57
soul	76.96	shell	13.22

Open in a new tab

Each word pair was typed in Times New Roman lower case font, 80 pts, black on a white background, using Microsoft Powerpoint^®, in two text boxes symmetrically spaced to either side of the slide centre. These were saved as JPEG images and the words subtended between ~6 and ~10 degrees of visual angle when presented onscreen at a comfortable viewing distance of ~60cm. The stimuli were presented and response time (RT) recorded using Presentation^® on a Dell PC with the display running at 60Hz and a spatial resolution of 1024 by 768 pixels.

Procedure

Each participant completed 240 trials, comprising the 30 distinct words by 4 conditions (upright-real, upright-jumbled, inverted-real, inverted-jumbled) by two trial types (same, different) and pressed either the M or Z key on the keyboard to indicate whether the stimulus pairs were the same or different. All 4 conditions were mixed within the block of 240 trials. After a short practice run of 16 trials in which participants were acquainted with all 4 conditions and prompted as to the correct response, they completed a further 24 practice trials without feedback. The 240 experimental trials were then presented in pseudo-random order with a short break after the first 120 trials. Participants were asked to respond as quickly as possible on each trial once they were reasonably sure of the correct answer, i.e., both speed and accuracy were stressed. They were assured that the task did not involve reading aloud, something that can be stressful for those with dyslexia.

Results

The response time (RT) data were analysed in R [33] using ANOVA with a between-subjects factor of Group (dyslexic, control) and within-subjects factors of Orientation (upright, inverted), Orthography (real, jumbled) and Word Length (6-letter, 7-letter). A further measure of performance, sensitivity (d’), was analysed with factors of Group and Condition. Greenhouse-Geisser corrections were used when Mauchly's Test for Sphericity was significant and effect sizes are given by generalized eta squared (η²_G) [34].

Response time

Response time (RT) analyses were carried out for correct trials only; 89.5% and 91.4% of trials for the experimental and control groups respectively. Exploratory data analyses highlighted a small number of outliers and 14 RTs (< 0.1% of all trials) greater than 16000ms or less than 600ms were removed.

Mean response time (RT) is plotted for both dyslexic and control participants in all conditions in Fig 1. For both groups RT increases with word length and with increased task difficulty, being faster for upright than for inverted stimuli and for real words than for jumbled words. Control participants are notably faster than those with dyslexia across these conditions.

ANOVA with dependent variable of RT showed significant main effects of Group, F(1,58) = 7.51, p < 0.01, η2G = 0.09, of Orientation, F(1,58) = 329.81, p < 0.01, η2G = 0.41, of Orthography, F(1,58) = 272.61, p < 0.01, η2G = 0.15, and of Word Length, F(1,58) = 322.05, p < 0.01, η2G = 0.09. The analysis also revealed three significant two-way interactions: Orientation*Orthography, F(1,58) = 13.54, p < 0.01, η2G = 0.005, Orientation*Word Length, F(1,58) = 30.79, p < 0.01, η2G = 0.004, and Orthography*Word Length, F(1,58) = 11.08, p < 0.01, η2G = 0.002. These are easily interpreted from Fig 2, where the increase in RT with stimulus inversion is somewhat steeper for real than jumbled words and for 7-letter than 6-letter words and where jumbling of the letters within a word leads to steeper increase in RT for 7-letter than 6-letter words. As the effect of word length did not differ between the dyslexic and control participants, it was not explored in further analyses.

To further explore the effects of ‘inversion’ and ‘orthography’, we calculated the proportionate increase in RT due to these experimental manipulations as follows. For the ‘inversion effect’ we calculated for each participant their mean RT in the inverted-real and inverted-jumbled conditions and divided by their corresponding mean RT for the upright-real and upright-jumbled conditions. For the ‘orthography’ effect we calculated for each participant their mean RT in the upright-jumbled and inverted-jumbled conditions and divided by their corresponding mean RT for the upright-real and inverted-real conditions. The mean proportionate increase in RT, for dyslexic and control participants, due to stimulus inversion and due to change in orthography is plotted in Fig 3. By these measures participants with dyslexia are, on average, less affected by stimulus inversion than are control participants and, similarly, they are less affected by changes in ‘orthography’—the jumbling of letters in the word- than are control participants.

Also evident from Fig 3 is that inversion is more detrimental to the processing of real than jumbled words and, similarly, that jumbling the letter order is more detrimental to the processing of upright than inverted words.

For the ‘inversion effect’ data ANOVA with between-subjects factor of Group and within-subjects factor of Orthography and dependent variable of proportionate increase in RT showed significant main effects of Group ( $\bar{X}$ = 1.60 [1.51, 1.70] for dyslexics, $\bar{X}$ = 1.86 [1.75, 1.97] for controls), F(1,58) = 10.74, p < 0.01, η2G = 0.12, and of Orthography ( $\bar{X}$ = 1.53 [1.46, 1.60] for jumbled words, $\bar{X}$ = 1.93 [1.82, 2.05] for real words), F(1,58) = 76.42, p ~ 0, η2G = 0.26. The Group* Orthography interaction was not significant (p = 0.45).

For the ‘orthography effect’ data ANOVA with between-subjects factor of Group and within-subjects factor of Orientation and dependent variable of proportionate increase in RT showed significant main effects of Group, ( $\bar{X}$ = 1.28 [1.22, 1.33] for dyslexics, $\bar{X}$ = 1.41 [1.34, 1.48] for controls), F(1,58) = 15.80, p < 0.001, η2G = 0.12, and of Orientation, [ $\bar{X}$ = 1.19 [1.15, 1.24] for inverted words, $\bar{X}$ = 1.49 [1.44, 1.55] for upright words), F(1,58) = 75.57, p ~ 0, η2G = 0.40. The Group*Orientation interaction was not significant (p = 0.65).

Sensitivity

Fig 4 plots d’—a measure of sensitivity that is independent of response bias calculated from the rates of ‘hits’ and ‘false alarms’ [35]–at each of the four stimulus presentation conditions with separate traces for the dyslexic and control participants, and where the d’ data have been scaled to a maximum value of 1.0. Sensitivity declines gradually with increased task difficulty and while slightly lower for the dyslexic than control participants there is considerable overlap of the 95% CIs error bars for the two groups. ANOVA showed a significant main effect of Condition, F(3,174) = 59.52, p ~ 0, η2G = 0.25. Neither the main effect of Group (p = 0.25) nor the Group*Condition interaction (p = 0.88) were significant.

Planned post-hoc comparisons showed significant differences between the upright-real and upright-jumbled conditions, F(1,59) = 28.98, p ~ 0.0, η2G = 0.11, and between the inverted-real and inverted-jumbled conditions, F(1,59) = 10.76, p = 0.002, η2G = 0.02, showing that the effect of jumbling was significant for both upright and inverted words. Similarly, there were significant differences between the upright-real and inverted-real conditions, F(1,59) = 99.36, p ~ 0, η2G = 0.30, and between the upright-jumbled and inverted-jumbled conditions, F(1,59) = 43.63, p ~ 0, η2G = 0.14, showing that the effect of inversion was significant for both real and jumbled words. Bonferroni corrected alpha is 0.0125 for four comparisons.

Discussion

Two general observations can be made about the results. First, the experimental task clearly engages perceptual expertise in word recognition, as response time increases with both inversion and with letter jumbling, establishing a gradient of difficulty whereby participants are fastest for upright-real words, next fastest for upright-jumbled words, slower again for inverted-real words and slowest for inverted-jumbled words. Although word length only varied by one letter, response times showed word length by orientation and word length by orthography interaction effects so that the latency differences for 6- and 7-letter words was greater for inverted than upright and for jumbled than real words. Therefore, the experimental manipulations served, as expected, to shift processing from an expert, largely holistic style to an increasingly analytic style and likely induced letter by letter reading in the most difficult condition. If participants were relying exclusively on an analytic strategy then word length effects would not be expected to vary with changing orientation and orthographic regularity.

Secondly, we stress that even the easiest condition likely involves some degree of analytic processing. This is because participants must scan between a pair of words in deciding whether they are the same or different. And, as spelling regularities exist at both a coarse, whole word scale and at finer sub-lexical scales [5], a strategy that capitalises on expertise for orthographic regularity may also involve some analytic processing.

Regarding the performance of dyslexic and normal readers, four aspects of the results are notable. First, the performance of both groups is consistent with the use of holistic processing, showing clear evidence of both superiority and inversion effects, i.e., participants from both groups are faster to detect a difference between the stimulus pairs when these are real words rather than jumbled words and when they are upright than inverted. Secondly, and central to this study, we find that college students with dyslexia are less hindered than their peers by changes in orientation and orthography that make the task more difficult. Specifically, the proportionate change in response time that accompanies both word inversion and word jumbling is significantly greater for normal readers than their dyslexic peers. This result is consistent with the idea that ‘compensated dyslexics’–adults who have successfully learned how to read–may read in a manner that still relies more on analytic and less on holistic processes. As such they may be less affected by the experimental conditions that force a shift from holistic to analytic processing. The third and fourth observations from our data are related, namely that the dyslexic college students are significantly slower than their peers in all conditions but show comparable sensitivity in the task. This is consistent with a recent study comparing the cognitive profile of dyslexic college students with their peers [36]; here differences in various performance measures, apart from spelling, reflect response time rather than accuracy differences.

In Experiment 2 we modified the task to address two limitations of the first study. First, there is the issue of word length. While the 6- and 7-letter words we used might not be classified as particularly ‘long words’ by some accounts [37], using shorter words may increase the use of holistic processing in the easier conditions. A second limitation of Experiment 1 relates to the use of a non-blocked design, i.e., all four conditions were intermixed so that participants, insofar as they used different strategies in the different conditions, could not predict in advance the optimal strategy from trial to trial. Therefore, Experiment 2 utilized shorter words of 4 or 5 letters and a blocked design with the expectation of revealing more marked differences between the groups.