JCPSLP Vol 22 No 1 2020

Given the abovementioned studies have revealed rich sources of regularities, and sensitivity to these cues in humans and connectionist computational models, it is interesting to consider whether SL might help us to understand why some children struggle to learn to read. Of course, it is well established that a number of factors can affect children’s progress when learning to read. It is unlikely that dyslexic individuals lack any capacity for SL whatsoever and unlikely that atypical SL would be found to be the overarching feature of dyslexia given what we already know about this disorder. However, we might ask whether SL plays some kind of role in atypical reading development. Group comparisons of SL in dyslexic versus non-dyslexic individuals Decades of research suggests that phonological processing difficulties play a significant role in dyslexia (e.g., Vellutino, Fletcher, Snowling, & Scanlon, 2004). However, difficulties in other areas could also contribute to dyslexia (e.g., Gathercole, Alloway, Willis, & Adams, 2006; Nicolson, Fawcett, Brookes, & Needle, 2010; Wolf & Bowers, 1999, among others). SL might be one of these other areas of difficulty. In its broadest sense, SL refers to the ability to detect regularities and is linked with implicit learning, procedural learning, motor learning, sequence learning (adjacent and non-adjacent dependencies), and serial order learning. Current theorising suggests that SL is an ability that emerges from multiple sub abilities (e.g., Arciuli, 2017; Reber, 2013; Sawi & Rueckl, 2019; Thiessen, Kronstein, & Hufnagle, 2013), although it does appear to be separable from what we know as “intelligence” (e.g., Conway, Baurnschmidt, Huang, & Pisoni, 2010; Kaufman et al., 2010; Kidd, 2012; Kidd & Arciuli, 2016; Siegelman & Frost, 2015; Tong, Leung, & Tong, 2019; Torkildsen, Arciuli, & Wie, et al., 2019). As explained by Arciuli and Conway (2018), there is no single agreed upon way to measure SL. Rather, there are a number of tasks that have been used to explore SL including the triplet task (Saffran, Aslin, & Newport, 1996), artificial grammar learning (Reber, 1967), the serial reaction time task (Nissen & Bullemer, 1987), and the Hebb repetition task (Hebb, 1961), among others. It is not clear whether each of these tasks draws on the same sub abilities thought to comprise SL. Those interested in a fuller explanation of each of these tasks and accompanying graphical depictions are referred to Arciuli and Conway (2018). As well as differences across tasks, each of these tasks can be altered in a myriad of ways (e.g., participant instructions, modality of presentation, presentation times, complexity of the regularities, number of times participants are exposed to the regularities, etc.). In addition, dependent variables from these tasks differ greatly (e.g., some measure learning as it is happening, some measure learning immediately after it has taken place, some dependent variables rely on motor processes or explicit judgements, etc.). Thus, different tasks, task alterations, and a multitude of dependent variables may draw on the underlying sub abilities of SL in different ways. A meta-analysis of studies using serial reaction time tasks by Lum, Ullman and Conti-Ramsden (2013) concluded that dyslexic individuals exhibit impaired SL relative to non-dyslexic individuals, although there was heterogeneity in effect sizes across studies and some studies showed no statistically significant group differences. There was evidence that the link between SL and dyslexia is affected by age with smaller effects reported in studies of adult participants.

patterns of lexical stress in English. Following Kelly’s (2004) suggestion that words’ beginnings might also hold cues to lexical stress, Arciuli and Cupples (2007) conducted a large- scale corpus analysis of CELEX and identified additional probabilistic orthographic cues to lexical stress. Subsequent cross-linguistic research has revealed pervasive probabilistic orthographic cues to stress in the initial and final parts of polysyllabic words in large-scale corpus analyses of Italian, Greek, Dutch, Spanish, and German as well as English (Monaghan, Arciuli & Ševa, 2016). Behavioural and electrophysiological studies have demonstrated participants’ sensitivity to these kinds of probabilistic orthographic cues (e.g., see studies of English Italian, Greek and Russian by Arciuli and Cupples [2006]; Arciuli and Cupples [2007]; Arciuli and Paul [2012]; Burani and Arduino [2004]; Grimini and Protopapas [2016]; Jouravlev and Lupker [2014]; Jouravlev and Lupker [2015]; Sulpizio and Colombo [2017], among others). As an aside, such probabilistic orthographic cues are also related to grammatical category membership (e.g., noun vs verb) in English disyllables (Arciuli & Cupples, 2006) and English trisyllables (Arciuli & Monaghan, 2009). Behavioural and neuroimaging studies show that people are sensitive to these cues (e.g., Arciuli, McMahon, & de Zubicaray, 2012; Kemp, Nisson, & Arciuli, 2009). The developmental study by Arciuli, Monaghan, and Ševa (2010) utilised a triangulation of methods to investigate probabilistic orthographic cues to lexical stress: corpus analysis of children’s reading materials, behavioural testing of 5–12 year olds to determine sensitivity to these cues at different ages, and computational modelling to explore incremental and implicit learning of these cues following exposure to age-appropriate input. Corpus analyses of disyllabic words in the Educator’s Word Frequency Guide database (Zeno, Ivens, Millard, & Duvvuri, 1995) examined 2959 words in reading materials for 5–6 year olds, 3814 words in materials for 7–8 year olds, 4430 words in materials for 9–10 year olds, and 4594 words in materials for 11–12 year olds. Results confirmed the presence of probabilistic orthographic cues to lexical stress in children’s reading materials. Nonwords were created to test children’s sensitivity to these cues at different ages. Behavioural results demonstrated that sensitivity to these probabilistic orthographic cues to lexical stress increases with age, presumably due to more exposure to text. A computational model, which advanced earlier modelling work on adult data by Ševa, Monaghan and Arciuli (2009), utilised a single route connectionist architecture based on the principles of SL and closely simulated the child data. A study by Mousikou, Sadat, Lucas, and Rastle (2017) compared three computational models: the rule-based algorithm by Rastle and Coltheart (2000), the connectionist CDP++ model by Perry, Ziegler, and Zorzi (2010), and the connectionist model by Ševa et al. (2009). The two connectionist models outperformed the rule-based model in simulating adult human data on assignment of lexical stress “thus providing support for a statistical-learning approach” (p. 188). Other studies have focused on different kinds of lesser known regularities, unrelated to lexical stress. As mentioned, Steacy et al. (2019) investigated regularities in context dependent vowel pronunciation in English monosyllables and showed that children were sensitive to these regularities when reading aloud. Those interested in related research are referred to Treiman (2018), Senechal et al. (2016) and Pacton et al. (2005).

4

JCPSLP Volume 22, Number 1 2020

Journal of Clinical Practice in Speech-Language Pathology