Arabic is a Semitic language with a rich historical and cultural heritage, distinguished by unique script and linguistic features that contribute to its complexity. Like other Semitic languages, Arabic is written from right to left along a horizontal line. It employs an ABJAD writing system wherein its fundamental script comprises consonants, with optional symbols for denoting short vowels and other morpho-phonemic features of the language (Daniels, 2013). The writing system is substantially more complex in Arabic than in languages using the Latin alphabet (Verhoeven and Perfetti, 2022). The Arabic alphabet consists of 28 letters, including two semi-vowels that can function as either a consonant or a long vowel depending on context. Written in a cursive style, Arabic orthography features connected letters with no uppercase counterparts. However, the form of Arabic letters varies based on their within-word position (initial, medial, or final). For example, the letter “Kaaf” /k/ is written as in the initial position, as in the medial position, and as in the final position. These distinct letter forms, known as “allographs,” encompass over one hundred variations in the Arabic script. Additionally, despite the cursive nature of the Arabic script, six out of the twenty-eight letters ( ) do not connect to the following letter, resulting in one or more spaces within words.

Furthermore, many Arabic letters share similar shapes, differing only in the presence and positioning of dots or points (known as primary diacritics), such as the letters /b/, /t/, and /θ/. Short vowels, in Arabic orthography, are represented by secondary diacritics above or below the letter, with three main vowels: /a/ or /fatḥa/, /u/ or /ḍamma/, and /i/ or /kasra/. The absence of a vowel is denoted by ° or /sukūn/. Certain diacritical marks, such as the /tanwīn/ (or nunation), indicating an indefinite noun through vowel doubling, and the /šadda/ (or gemination), representing consonant doubling, are considered morpho-phonemic (Saiegh-Haddad and Henkin-Roitfarb, 2014). It is important to note that vowelization in Arabic script is optional. Short vowel signs (secondary diacritics) may or may not be included, leading to two versions of the Arabic script: vowelized and unvowelized. The vowelized version is prevalent in classical Arabic texts like the Holy Qur’an, classical poems, and literacy books for young learners. The unvowelized version, devoid of short vowel markings, is used by proficient readers in books, novels, media, etc., and is introduced to children around the fourth grade to gradually familiarize them with reading in its unvowelized form. The two scripts impose different constraints on the cognitive system of reading. The vowelized script is fully transparent while the unvowelized script transcribes only part of the word phonological form (Abu-Rabia, 2001).

Another fundamental characteristic of the Arabic language is its diglossic nature, wherein it manifests in two distinct forms: the standard variety and the spoken variety. The Modern Standard Arabic (MSA) form adheres to defined rules and grammar, serving as a shared language among all Arabic speakers. It is predominantly used in written form and finds application in formal settings, religious discourse, and media communications (Versteegh, 2014). In contrast, the spoken variety is employed in everyday conversations and exhibits geographical variations across regions and even within the same country. This spoken form serves as the primary language of Arabic speakers, acquired naturally through familial interactions. Exposure to MSA typically begins during formal education, often in kindergarten. Notably, a linguistic gap exists between spoken Arabic and MSA, encompassing differences in phonology, lexicon, syntax, and morphosyntax. These disparities position MSA as a second language for young learners (Saiegh-Haddad and Schiff, 2016).

Last, Semitic morphology differs from that of European languages due to its unique non-concatenative derivational structure (Deutsch et al., 2018). Arabic morphology relies on a system of discontinuous morphemes known as roots and patterns. The root, typically composed of three consonants, indicates a semantic field, and serves as the foundation for deriving numerous words of the same semantic family. The specific meaning of each word results from the combination of the root with a pattern that corresponds to a set of vowels (and sometimes additional consonants). For instance, from the root KTB, denoting the realm of writing, arise words such as /KaTaBa/ (he wrote), KaaTiB (writer), KiTaaB (book), and maKTaBa (library). There is evidence that the morphological structure of Arabic words has an impact on reading accuracy and comprehension (Abu-Rabia, 2007; Boudelaa, 2014). Note that the impact of morphological processing on reading acquisition is beyond the scope of this paper.

Phonological awareness, RAN, LK and VAS have been identified as reading-related skills in Indo-European languages and as significant or potential predictors of reading in Arabic.

Rapid automatized naming requires the rapid and accurate naming of arrays of familiar items, such as letters, digits, colors, or objects. RAN is a robust predictor of reading across Indo-European languages (Araújo et al., 2015; Araújo and Faísca, 2019; Georgiou et al., 2016; Landerl et al., 2019). The predictive strength of RAN is particularly pronounced for RAN-letters or RAN-digits tasks, although RAN-colors and RAN-objects also significantly predict reading achievement in pre-readers (Lervåg and Hulme, 2009). Evidence from Arabic similarly highlights the role of RAN: it contributes to reading (Asadi et al., 2017; Hassanein et al., 2023; Ibrahim, 2015) and serves as an early predictor of reading speed, independently of PA (Asaad and Eviatar, 2014).

Phonological awareness, the ability to recognize and manipulate the phonological units of spoken language, is a key predictor of reading achievement in all languages. Strong PA skills facilitate accurate decoding and fluent reading (Duncan, 2018; Ehri et al., 2001; Melby-Lervåg et al., 2012). In Arabic as well, PA is consistently reported as a strong predictor of reading proficiency (Layes et al., 2023; Makhoul, 2017; Saiegh-Haddad and Taha, 2017; Taibah and Haynes, 2011; Tibi and Kirby, 2018). Good PA in preschool Arabic children is critical for literacy development (Mansour-Adwan et al., 2023) and PA has been shown to uniquely predict reading outcomes, independently of RAN (Asadi et al., 2017; Taibah and Haynes, 2011).

Letter knowledge (LK), defined as the ability to recognize letter shapes and associate them with their corresponding names or sounds, is one of the strongest predictors of early reading skills in Indo-European languages (Foulin, 2005). LK supports the development of decoding skills (Acha et al., 2023; Kim et al., 2010) and a LK deficit is associated with poor decoding and reduced fluency (Thompson et al., 2015). In Arabic, letter identification poses particular challenges for beginning (and even, more proficient) readers (Abdelhadi et al., 2011; Eviatar and Ibrahim, 2004). Several studies have emphasized the difficulty of recognizing and discriminating Arabic letters, suggesting that letter recognition may be more demanding on visual attention, thereby resulting in slower reading speed (Hansen, 2014; Ibrahim et al., 2002). However, to our knowledge, no study has investigated the potential unique contribution of LK to reading, beyond that of PA and RAN.

A further factor influencing reading acquisition is VAS. VAS is a measure of multi-element parallel processing in the visual modality (Valdois, 2022). It differs from the concepts of perceptual span, or visual span (Frey and Bosse, 2018), and reflects the total amount of visual attention available for the simultaneous processing of multiple elements. VAS contributes to decoding, word recognition and reading fluency, beyond PA and RAN (Bosse and Valdois, 2009; Valdois et al., 2021). It is also an early predictor of later reading skills (Valdois et al., 2019a). The few studies, that have investigated the contribution of VAS to reading in Arabic, have led to inconclusive findings. In expert readers, Awadh et al. (2016) reported no effect of VAS on reading unvowelized texts. By contrast, Lallier et al. (2018) found a relationship between VAS and text reading in Grade 4 Arabic children, but only among those with greater proficiency in reading unvowelized scripts. Last, Awadh et al. (2022) reported that VAS uniquely contributed not only to word and pseudoword reading but also to text fluency and comprehension in Grade 4–5 children. Taken together, these mixed findings underscore the need for further research into the VAS-reading relationship in the Arabic language.

Overall, RAN, PA, LK and VAS are well established as early and independent predictors of reading achievement in Indo-European languages. In Arabic, the respective roles of PA and RAN have also been investigated, with similar evidence for their unique contribution to reading. By contrast, the potential contribution of LK has yet to be examined independently of PA and RAN, and the status of VAS as an early and independent predictor of reading is still debated. Theoretical models of reading offer valuable insights into how these different skills influence reading acquisition, thereby enhancing our understanding of their unique contributions and complex interrelationships.

According to self-teaching models (Ziegler et al., 2014; Pritchard et al., 2018), stronger PA facilitates the acquisition of decoding skills and sight-word recognition, thereby supporting more efficient reading. By contrast, the cognitive mechanisms underlying the RAN-reading relationship are still debated (Decker et al., 2013; Georgiou and Parrila, 2020). Although RAN shares many visual, attentional, language and articulatory processes with reading, no specific mechanism has been conclusively identified to account for its unique contribution to reading development.

In contrast, the contribution of letter recognition and visual attention is more clearly accounted for in computational models of reading. All reading models (Coltheart et al., 2001; McClelland and Rumelhart, 1981; Perry et al., 2007; Phénix et al., 2025) posit that word recognition is letter-based. All assume that letters are recognized through sensory information extracted from the input word. Efficient and accurate letter identification enables rapid activation of words’ orthographic representation in long term memory and its mapping onto the corresponding phonological word, thus supporting fluent reading. However, these models do not postulate that the letter itself is the basic unit of processing. Rather, letter identification is assumed to be feature-based (Grainger et al., 2008; McClelland and Rumelhart, 1981). Incoming visual information from the letter is first analyzed by detectors specialized in the processing of visual features, such as horizontal and vertical lines, intersections, curvatures or points (Dehaene et al., 2005). Letter recognition occurs through the detection and integration of these features. Some theories assume that feature binding for letter identification is achieved through visual attention (Laberge and Samuels, 1974; Treisman and Gelade, 1980).

Recent models provide insights on the role of visual attention in letter and word processing (Ginestet et al., 2019, 2022; Phénix et al., 2025; Steinhilber et al., 2023). In the BRAID model of word recognition (Phénix et al., 2025), illustrated on Figure 1, visual attention intervenes between the sensory and perceptual levels of processing to promote letter recognition.

Letter recognition is challenging at the sensory level, due to both the characteristics of the alphabetic system and the properties of the visual system. Sensory processing primarily focuses on visual features, making letter identification more difficult when a letter shares numerous features with other letters in the alphabet. Additionally, letter visibility within word is degraded by the decline in visual acuity with increasing distance from fixation and by interference from adjacent letters. Visual attention has the potential to compensate for these deleterious effects, as allocating attention to letters improves their discriminability and visibility.

However, visual attention capacity (i.e., the total amount of visual attention available for processing) varies across readers. A high visual attention capacity allows simultaneous processing of multiple letters, facilitating rapid decoding and word recognition. Conversely, a lower visual attention capacity requires focusing on fewer letters for accurate recognition, resulting in more serial reading (Steinhilber et al., 2023).

This theoretical framework yields two key predictions. First, it predicts that letter processing requires more attentional resources when letters are visually complex and difficult to discriminate (i.e., sharing many features with other letters). Second, it predicts that within-word letter recognition depends on the amount of visual attention available for processing, a subject-dependent capacity that can be estimated through VAS tasks.

In the present study, our primary aim was to determine whether LK and VAS make unique contributions to the reading performance of Arabic beginning readers, independent of PA and RAN. While previous studies have examined the influence of some of these predictors, none have simultaneously considered all four. Oral vocabulary knowledge was also included, based on evidence that vocabulary has a positive influence on learning to read and the development of orthographic knowledge (Ziegler et al., 2014).

We began with the observation that most previous studies on reading in Arabic have used Arabic letters as stimuli to assess LK, VAS and RAN. However, RAN and VAS skills can be assessed using a variety of item types. Assuming that using the same stimuli across all three tasks would artificially strengthen the relationships among predictors and reduce their potential unique contribution to reading, we opted to use different items for assessing LK, RAN and VAS. LK is necessarily evaluated using the letters of the alphabet, whereas RAN can be measured using either alphanumeric or non-alphanumeric items to tap rapid access to phonological labels, fast visuo-verbal matching and/or processing speed. Participants were administered a RAN-objects task, based on prior evidence for a relationship between this version of the task and reading in Arabic (Layes et al., 2017). VAS is typically assessed through the oral report of alphanumeric stimuli to estimate the amount of visual attention allocated to the simultaneous processing of multiple visual elements (Valdois, 2022). Leveraging participants’ knowledge of English, the VAS report tasks were administered using Latin letters. This approach – using Arabic letters for the assessment of LK, objects for RAN and Latin letters for VAS – ensured that each fundamental skill was specifically targeted.

Consistent with previous behavioral evidence and current models of self-teaching (Ziegler et al., 2014; Pritchard et al., 2018), higher PA at the onset of literacy instruction should facilitate the development of decoding skills and thus predict higher performance in reading nonsense syllables and words in vowelized script. As previously reported, RAN should account for a significant proportion of variance in Arabic reading, over and above PA.

Theoretical models of visual attention posit that visual attention is involved in feature binding for letter recognition and in parallel multiletter processing for letter-chunk and word recognition. Given the particular challenge of Arabic letter identification for beginning readers, we predict a contribution of VAS to letter recognition (LK). According to the BRAID model, visual attention further facilitates letter-chunk and word recognition by compensating for the detrimental effects of visual acuity decline and crowding on letter identification within strings. Therefore, we predict a relationship between VAS and reading fluency.

However, visual attention is a capacity-limited resource. When greater attentional resources are required for letter feature identification, fewer resources remain available for letter-chunk and word recognition processing. Therefore, we expect VAS to exert either a direct residual effect on reading after controlling for LK, or only an indirect effect on reading, in case of full mediation by LK.

Reading skills were assessed through tasks of nonsense syllable (viewed as legal monosyllabic pseudo-words) and word reading. Evidence from Indo-European languages suggests that PA, RAN, LK and VAS contribute differently to word and pseudo-word reading (Georgiou et al., 2008; Shapiro et al., 2013). Accordingly, we will examine the predictive power of these different skills separately for each reading measure.

One hundred and thirty-four bilingual Lebanese first graders were recruited from four private schools with mild to low socio-economic levels. All children had normal hearing and either normal or corrected-to-normal vision. They had completed 3 years of kindergarten and began formal reading instruction in Grade 1. Their native language was Lebanese Arabic. From kindergarten onward, children were exposed to both Arabic and English. Oral instruction in Arabic consisted in a combination of Modern Standard Arabic (MSA) and Lebanese Arabic, adapted to the children’s comprehension and oral expression abilities. Written instruction in Grade 1 was exclusively in MSA. English reading instruction also began in Grade 1, and both Arabic and Latin letters had been formally introduced in KG3. Testing occurred after the second trimester of the academic year, after 4 months of formal literacy instruction.

The study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki. Ethical approval for the study (PASEM Project: E.T. as PI) was granted by the Ethics Committee of the Africa Institute for Research in Economics and Social Sciences (under grant: ECAIRESS-002-2024). Legal responsibility for the children during school hours was assumed by the school principals, who consented to the assessments. Written informed consent was obtained from all parents for their child’s participation. Additionally, verbal assent was secured from each child at the beginning of each testing session, with reassurance that they could withdraw at any time.

Due to absences during at least one data collection session, complete data were not available for all children: 18 data points were missing for reading tasks, 23 for VAS tasks and 10 for PA tasks, resulting in the exclusion of 32 children (24%). In addition, we excluded one extreme outlier, likely due to a measurement error during the reading task (120 syllables correctly read per minute). The final sample comprised 101 first graders (58 females) with a mean age of 6 years and 11 months (SD = 4.1 months).

Most of the tasks administered in this study were custom-designed due to the absence of standardized assessment for beginning Arabic readers.

Data collection was conducted by trained experimenters, all of whom had background in speech and language therapy. The tests were administered individually in a quiet room at school. To control for fatigue effects, the order of task administration was randomized across participants. Each child completed the assessment in 2-to-3 sessions, each lasting approximately 35–50 min. Following data collection, scoring was performed by the experimenters and cross-verified twice to ensure accuracy. Reliability estimates for the different measures were calculated using McDonald’s Omega coefficients (Hayes and Coutts, 2020) and are reported in Table 1.

For statistical analyses, we first computed correlation matrices, which revealed a strong association between two potential predictors of reading fluency: Visual attention span and Letter knowledge (LK). We then conducted two sets of regression analyses: one excluding Letter knowledge and one including it as a predictor. Finally, structural equation modeling (SEM) was employed to further examine the specific contributions of these two correlated predictors to reading fluency.

A substantial floor effect was observed on the Phoneme Deletion task, with 68% of children scoring 0 out of 8. Consequently, this measure was excluded from further analysis. Table 1 reports descriptive statistics for all remaining variables, as well as for the four composite measures of PA, LK, VAS and Word Reading. The PA composite score was theoretically justified and calculated by averaging performance on the Syllable Segmentation and Syllable Deletion tasks. The LK composite score was obtained by summing performance across the three tasks of Letter Naming, Allograph Naming and Allograph Designation. The Word Reading composite score corresponded to the average performance on the mono- and poly-syllabic words. To balance the contributions of the global and partial report tasks, the VAS composite score was calculated using the following formula:

As shown in Table 1, McDonald’s Omega coefficients for all variables, including the four composite constructs, ranged from 0.83 to 0.96 (with the exception of the Raven test: ω = 0.76), indicating good internal consistency of the measurement instruments.

The distributions of most variables (Raven, Vocabulary, LK, SLIT and VAS) approximated normality, with skewness values ranging from −0.56 to +0.48 and kurtosis values from −1.01 to 1.18. The distribution was moderately skewed for PA and highly skewed and peaked for the RAN, Syllable Fluency and Word Fluency measures. To increase symmetry and reduce kurtosis, square root transformations were applied to all variables except RAN, which underwent an inverse transformation. After transformation, skewness and kurtosis values fell within acceptable ranges (skewness: −0.56 to 1.01; kurtosis: −1.01 to 1.18; see Supplementary Table 1).

Before computing composite scores of LK, VAS and Word Reading, we verified that the component scores for each construct were significantly correlated (see Supplementary Table 2). Strong correlations were observed among the three LK tasks (from r = 0.82 to 0.87, all p_s > 0.001). Significant associations were also found between Partial and Global Report tasks [r(99) = 0.59, p < 0.001], and between monosyllabic and polysyllabic words [r(99) = 0.86, p < 0.001]. Furthermore, high internal consistency for aggregated items (ω_PA = 0.85; ω_LK = 0.96; ω_VAS = 0.88; ω_{Word Reading} = 0.95) justified the use of composite scores in subsequent analyses.

Table 2 presents the zero-order and partial correlation coefficients among all main variables. Partial correlations were computed while controlling for RAVEN, to ensure that the observed associations between variables were not solely attributable to general cognitive skills.

As shown on Table 2, the predictive variables of PA, LK and VAS were significantly correlated with the two dependent measures of nonsense syllable and word reading. RAN was significantly correlated with these two reading measures, but only when Raven’s scores were not controlled for. In contrast, Vocabulary did not show significant associations with any variable and was therefore excluded from subsequent predictive analyses. The strongest correlations were observed between the two (word and syllable) reading measures and LK. However, significant intercorrelations were also found among some predictors. In particular, VAS correlated moderately with LK [r(99) = 0.57, p < 0.001] and, to a lesser extent, with PA [r(99) = 0.37, p < 0.05]. VAS was also correlated with the control variable of single-letter identification threshold (SLIT).

Simple regression analyses were first performed with either Syllable Reading or Word Reading as dependent variables, and PA, RAN, VAS, Raven, and SLIT as independent variables. The analyses (see Supplementary Tables 3a, b) revealed that the contribution of RAN did not significantly predict Word Reading fluency (t = 1.56, p = 0.12, R_p² = 0.03), and its effect on Syllable Reading fluency was only marginally significant (t = 1.83, p = 0.07, R_p² = 0.03). Only PA and VAS emerged as significant predictors of reading skills (PA: t = 2.99, p = 0.004, R_p² = 0.09, and t = 4.41, p < 0.001, R_p² = 0.17; VAS: t = 3.21, p = 0.002, R_p² = 0.10, and t = 4.19, p < 0.001, R_p² = 0.16, for Syllable and Word Reading, respectively).

However, when LK was introduced as an additional predictor, PA remained a significant predictor of both syllable and word reading fluency (Syllables: t = 3.37, p = 0.001, R_p² = 0.11; Words: t = 4.92, p < 0.001, R_p² = 0.20), whereas the predictive contribution of VAS disappeared (Syllables: t = −0.95, p = 0.34, R_p² = 0.01; Words: t = 0.83, p = 0.41, R_p² = 0.01). A significant effect of RAN was found, but only for Syllable Reading (Syllables: t = 2.28, p = 0.03, R_p² = 0.05; Words: t = 1.72, p = 0.09, R_p² = 0.03). Notably, LK emerged as the strongest predictor for both Syllable Reading (t = 10.37, p < 0.001, R_p² = 0.53) and Word Reading (t = 7.98, p < 0.001, R_p² = 0.40). Together with the substantial correlation between LK and VAS (r = 0.57), these findings suggest that the effect of VAS on reading performance is likely indirect, mediated through LK.

To directly assess this hypothesis, we performed two structural equation models while controlling for RAVEN and Single (Latin) letter identification threshold (SLIT). These models were conceptually similar to previous regression analyses, but included both the direct and indirect (through LK) effects of VAS on reading performance. As shown in Figures 2, 3, the models accounted for 71% and 68% of variance in Syllable Reading and Word Reading, respectively (see Supplementary Tables 4a, b for detailed information about the two structural models). Consistent with our predictions, VAS exerted no significant direct effect on reading fluency (Syllables: b* = −0.07, p = 0.33, z = −0.98; Words: b* = 0.07, p = 0.36, z = 0.91). Instead, its contribution was entirely indirect through LK (Syllables: b* = 0.41, p < 0.001, z = 5.39; Words: b* = 0.33, p < 0.001, z = 5.46). VAS explained 33% of the variance in LK, which in turn accounted for 50% and 32% of the variance in Syllable and Word Reading, respectively.