Comparing and validating methods of reading instruction using behavioural and neural findings in an artificial orthography.

Reading Instruction in Alphabetic Writing Systems

In England, the provision of systematic phonics instruction is a legal requirement in state-funded primary schools. To ensure compliance in all schools, children are required to participate in a national “phonics screen” in their second year of reading instruction (when they are five or six years-old), which measures word and nonword reading aloud. Since its implementation in 2012, the results of this assessment have shown dramatic year-on-year gains in the percentage of children reaching the expected standard—from 58% in 2012 to 81% in 2016. However, despite the apparent success of this policy, there continues to be resistance to it among teachers’ unions and others, who argue in favor of a less-prescriptive approach consisting of a variety of phonic- and meaning-related skills ( Association of Teachers & Lecturers, 2016 ; National Union of Teachers, 2015 ). One frequent objection is that while phonics may assist reading aloud, it may not promote (and may even erode) reading comprehension ( Davis, 2013 ).

The provision of systematic phonics instruction is also part of the Common Core State Standards Initiative in the U.S. ( http://www.corestandards.org/ ). However, at present, not all U.S. states have adopted the Common Core standards. Further, unlike in England, there is no national assessment of phonic knowledge in young children that could reveal the success of the standards, or individual schools’ compliance with them. Similarly, even among states that have adopted the Common Core standards, there are reports that particular school boards promote “balanced literacy” approaches that include a variety of meaning-related as well as phonic-related skills ( Hernandez, 2014 ; Moats, 2000 ).

Finally, although reading using phonic knowledge is included in the Australian curriculum, it is suggested as only one strategy, alongside a multicuing approach based on contextual, semantic, and grammatical information ( Snow, 2015 ), including guessing the pronunciation of a word based on a picture or the word’s first letter ( Neilson, 2016 ). The use of systematic phonics instruction is even less widespread in New Zealand classrooms, where text-based information (e.g., predictions based on pictures, preceding context, and prior knowledge) is regarded as more important than word-level phonic information for reading acquisition ( Tunmer, Chapman, Greasy, Prochnow, & Arrow, 2013 ).

This brief review suggests that there is considerable variability in how reading is taught in English-speaking countries. Some prioritize print-to-sound knowledge, others prioritize print-to-meaning knowledge, and still others teach a variety of sound-based and meaning-based skills in the initial periods of reading instruction. Though there is strong evidence for the importance of learning to appreciate print-to-sound relationships in reading acquisition (e.g., National Reading Panel, 2000 ; Rayner et al., 2001 ; Rose, 2006 ), there is limited data on the behavioral and neural consequences of the relative difference in emphasis that characterizes reading instruction in the classroom. In the next section, we consider what might be the cognitive foundations of this differential emphasis.

Print-to-Sound and Print-to-Meaning Pathways for Reading

The Simple View of Reading ( Gough & Tunmer, 1986 ) has had substantial impact on policy and practice in literacy education (e.g., Rose, 2006 ). The Simple View proposes that reading comprehension arises from the combination of print-to-sound decoding plus oral language skill (i.e., sound-to-meaning mappings), and hence emphasizes the importance of phonic knowledge in reading instruction. However, the Simple View is not a processing model, so is silent as to the actual mechanisms that underpin the discovery of meaning from the printed word. In order to capture these mechanisms, we must look to computational models of reading such as the Dual Route Cascaded model (DRC model, Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001 ) and the Triangle model ( Harm & Seidenberg, 2004 ; Plaut, McClelland, Seidenberg, & Patterson, 1996 ).

Like the Simple View of Reading, both the DRC and the Triangle model propose that reading comprehension can be achieved via an indirect pathway that maps from print-to-sound, and then from sound-to-meaning using preexisting oral language. The importance of the print-to-sound-to-meaning pathway is demonstrated by decades of research showing that phonological information is computed rapidly and as a matter of routine in the recognition and comprehension of printed words ( Frost, 1998 ; Rastle & Brysbaert, 2006 ). In alphabetic and syllabic writing systems, print-to-sound mappings are largely systematic, allowing words to be broken down into symbols that correspond to sounds. The print-to-sound-to-meaning pathway has therefore sometimes been termed the sub-word pathway because of the componential nature of print-to-sound mappings ( Wilson et al., 2009 ; Woollams, Lambon Ralph, Plaut, & Patterson, 2007 ). However, because the models vary in how they accomplish this mapping, we refer to the indirect pathway from print-to-sound-to-meaning simply as the phonologically mediated pathway .

Unlike the Simple View of Reading, both the DRC and Triangle model also propose a direct pathway from print-to-meaning. In alphabetic and syllabic writing systems, in contrast to the relationship between print and sound, the relationship between print and meaning is largely arbitrary and holistic. Similarly spelled words do not have similar meanings (at least when they are morphologically simple), and thus there is no sense in which a word can be broken down into component parts in order to access its meaning. This pathway is therefore sometimes termed the whole-word pathway ( Wilson et al., 2009 ; Woollams et al., 2007 ). However, the distributed nature of these mappings means that the Triangle model can also capture subword regularities between print and meaning where these exist, for example, in polymorphemic words ( Plaut & Gonnerman, 2000 ). Thus, throughout this article we will refer to the mapping from print-to-meaning as the direct pathway , because both models propose that written words can be comprehended without phonological mediation.

Plaut et al. (1996) argued that the “division of labor” between the phonologically mediated versus the direct pathway between print and meaning depends on the necessity of these processes for producing the appropriate response, given the task being performed and the characteristics of the orthography. This has led to suggestions that the direct pathway may only be necessary in orthographies with some degree of inconsistency between spelling and sound, because otherwise the phonologically mediated pathway can support accurate reading aloud and comprehension ( Share, 2008 ; Ziegler & Goswami, 2005 ). However, this overlooks the fact that comprehension of written words should be more efficient using the direct than the phonologically mediated pathway, irrespective of spelling–sound consistency ( Seidenberg, 2011 ). In the current research we therefore sought to determine how reading instruction that emphasizes print-to-sound versus print-to-meaning mappings impacts on the development of these pathways, and thus on reading aloud and comprehension of written words, in a regular orthography.

Neuroimaging and Neuropsychological Evidence for Dual Reading Pathways

Neuroimaging and neuropsychological evidence offer strong support for the notion of dual pathways to meaning proposed in cognitive models. This evidence has yielded a model in which phonologically mediated reading is underpinned by a dorsal pathway including left posterior occipitotemporal cortex, inferior parietal sulcus, and dorsal portions of the inferior frontal gyrus (opercularis, triangularis). Data from fMRI experiments in alphabetic languages reveal that these regions consistently show greater activation for nonwords than words ( Taylor, Rastle, & Davis, 2013 ). Left inferior parietal cortex has also been found to be more active when reading alphabetic relative to logographic writing systems ( Bolger, Perfetti, & Schneider, 2005 ). Furthermore, patients with damage to left posterior occipitotemporal cortex show slow and effortful reading in alphabetic scripts ( Roberts et al., 2013 ), and those with damage to left inferior parietal cortex and dorsal inferior frontal gyrus show poor nonword relative to word reading ( Rapcsak et al., 2009 ; Woollams & Patterson, 2012 ).

Conversely, neural evidence suggests that ventral pathway brain regions underpin direct print-to-meaning processes. Vinckier et al. (2007) showed that from left posterior to anterior occipitotemporal cortex there is an increasingly graded response to the word-likeness of written stimuli, with the mid-fusiform/inferior temporal gyrus responding more strongly to words and pseudowords than to stimuli containing frequent bigrams, followed by consonant strings, then false fonts. This processing hierarchy is supported by analyses of anatomical connectivity ( Bouhali et al., 2014 ), with posterior occipitotemporal cortex connecting to speech processing regions such as left inferior frontal gyrus and posterior middle and superior temporal gyri, whereas anterior fusiform shows connectivity with more anterior temporal regions that are important for semantic processing. Supporting the idea that direct print-to-meaning processes are underpinned by anterior fusiform, meta-analytic fMRI data reveal greater activation for words than nonwords in this region, in addition to more lateral temporal lobe regions such as left middle temporal and angular gyri 1 ( Taylor et al., 2013 ). Similarly, patients with left anterior fusiform lesions display poorer performance in reading and spelling words than nonwords ( Purcell, Shea, & Rapp, 2014 ; Tsapkini & Rapp, 2010 ). Further evidence for the correspondence between the direct pathway and ventral brain regions comes from semantic dementia patients, who have atrophy in anterior temporal lobes, including anterior fusiform gyrus ( Mion et al., 2010 ), and often show particular problems reading aloud words with atypical spelling-to-sound mappings ( Woollams et al., 2007 ). These irregular or inconsistent words depend more heavily on the direct pathway than words with typical or consistent print-to-sound relationships, which primarily rely on the phonologically mediated pathway.

The Development of Neural Reading Pathways

Like adults, children show neural activity in both dorsal and ventral pathways for a simple contrast of reading words and/or nonwords relative to rest ( Martin, Schurz, Kronbichler, & Richlan, 2015 ). Supporting the involvement of dorsal stream regions in children’s phonologically mediated reading skills, activity in left dorsal inferior frontal gyrus and inferior parietal cortex during a rhyme judgement task was positively correlated with change in nonword reading skill, between the ages of 9 and 15 ( McNorgan, Alvarez, Bhullar, Gayda, & Booth, 2011 ). Providing further longitudinal evidence, Preston et al. (2016) showed that in left fusiform gyrus, inferior parietal cortex, and dorsal inferior frontal gyrus, the degree of convergence between neural activity during print and speech tasks at age 8 predicted reading skill at age 10.

Computational models suggest that the involvement of the direct pathway increases with reading skill ( Harm & Seidenberg, 2004 ; Plaut et al., 1996 ). In line with this, several authors have proposed that as children become better readers, reliance shifts from the dorsal to the ventral pathway ( Pugh et al., 2000 ; Rueckl & Seidenberg, 2009 ; Sandak et al., 2012 ). This conceptualization is supported by longitudinal data showing that areas of the ventral pathway increase in sensitivity to written words between the ages of 9 and 15, and that this increasing sensitivity is associated with speeded word reading ability but not with nonword reading or phonological processing skill ( Ben-Shachar, Dougherty, Deutsch, & Wandell, 2011 ).

Overall, although the data from children are somewhat limited, they support the proposed distinction between the phonologically mediated dorsal pathway and the direct print-to-meaning ventral pathway. However, we are unaware of any evidence linking instructional methods to changes in these neural systems.

Laboratory Approaches to Studying Language Learning

Ultimately, the questions being addressed in this manuscript need to be investigated in child populations. However, to provide an initial investigation into the impact of teaching method on neural mechanisms for reading, we used an artificial language approach with adults. There has been a surge of interest in recent years in using these approaches to model the acquisition of different types of linguistic information ( Bowers, Davis, & Hanley, 2005 ; Clay, Bowers, Davis, & Hanley, 2007 ; Fitch & Friederici, 2012 ; Gaskell & Dumay, 2003 ; Hirshorn & Fiez, 2014 ; Tamminen, Davis, & Rastle, 2015 ; Taylor, Plunkett, & Nation, 2011 ). In contrast to studying children learning their first language (who vary in their prior experience with both spoken and written forms), artificial language approaches provide total control over participants’ prior knowledge of a new language or writing system. They also make it possible to manipulate what participants are taught and how they are taught in a way that could never be achieved in a naturalistic learning setting with children. Finally, working with adult learners permits collection of more extensive behavioral and brain imaging evidence during different stages of acquisition than would be possible with children. Artificial language learning studies consistently show that participants can learn sets of novel linguistic materials to a high degree of accuracy in a single training session, that this knowledge is sufficient to promote generalization to untrained materials ( Tamminen et al., 2015 ; Taylor et al., 2011 ; Taylor, Rastle, & Davis, 2014a ), and that this knowledge is long lasting ( Havas, Waris, Vaquero, Rodríguez-Fornells, & Laine, 2015 ; Laine, Polonyi, & Abari, 2014 ; Merkx, Rastle, & Davis, 2011 ; Tamminen & Gaskell, 2008 ).

This body of research has yielded interesting insights into the mechanisms that underpin the learning and abstraction of different types of linguistic information. However, significant questions remain over the extent to which artificial language learning reflects natural acquisition processes. Indeed, in the related field of artificial grammar learning, there is a long history of debate over the nature of knowledge acquired ( Perruchet & Pacteau, 1990 ; Reber, 1967 ; see Pothos, 2007 , for a review). For example, if these paradigms reflect strategic, problem solving operations rather than the development of long-term abstract knowledge, then this would undermine their usefulness in understanding the acquisition of linguistic knowledge. Similarly, it could be that adults who have fully developed language and/or reading systems solve these laboratory learning tasks in fundamentally different ways than children acquiring these linguistic skills for the first time. One approach to countering these criticisms has been to demonstrate that similar behavioral effects emerge in artificial language learning studies as in natural languages (e.g., frequency and consistency effects on reading aloud; Taylor et al., 2011 ), while another has been to show that the constraints that underpin artificial language learning in adults also pertain to children ( Henderson, Weighall, Brown, & Gaskell, 2013 ). However, more direct evidence that the processes recruited in artificial language learning paradigms overlap with those used in natural language would be desirable.

In this article we propose that neuroimaging data may provide this more direct evidence. Specifically, we will use brain responses to gain information about the mechanisms underlying different methods of literacy instruction in alphabetic writing systems. This approach is appropriate and timely because, as outlined earlier, the neural systems that underpin the phonologically mediated and direct pathways to reading in adults are well understood and appear to be similar in children ( Martin et al., 2015 ; Taylor et al., 2013 ). We propose that we can capitalize on knowledge of these pathways to assess (a) whether training people to read new words printed in artificial scripts engages these neural reading pathways, and (b) whether and how different parts of the reading system are affected by different forms of training. These observations would provide direct evidence not only of the value of artificial learning paradigms for investigating reading, but also of the mechanism by which a particular training intervention operates. We believe that these inferences would be very difficult to draw from a purely behavioral outcome (e.g., accuracy or learning rate in a particular training condition), or through naturalistic study of children learning to read in their first language. This knowledge gained from our study will therefore complement that acquired from studies conducted on children, to inform the development of new interventions that target specific reading pathways.

Laboratory Approaches to Studying the Neural Basis of Reading Acquisition

In the present study we aimed to uncover the neural consequences of reading instruction that prioritizes print-to-sound versus print-to-meaning mappings. Previous training studies suggest that learning and retrieving componential print-to-sound associations for novel words (written either in familiar or artificial letters) modulates neural activity in dorsal pathway brain regions such as left inferior parietal cortex and inferior frontal gyrus ( Mei et al., 2014 ; Quinn, Taylor, & Davis, 2016 ; Sandak et al., 2004 ; Taylor et al., 2014a ). However, modulation of ventral pathway activity in artificial language learning studies has been somewhat elusive. Taylor et al. (2014a) found that learning whole object names activated the ventral pathway (left anterior fusiform gyri and ventral inferior frontal gyrus), more than learning letter-to-sound associations (see also, Quinn et al., 2016 , for a similar finding when retrieving object names). However, no studies have reported ventral pathway activity for training on whole written words; instead, left angular and middle temporal gyri are more often implicated ( Mei et al., 2014 ; Takashima et al., 2014 ). The failure to observe ventral pathway activity for trained words may be the result of relatively short and/or superficial training regimes, or because trained words were meaningless. In the current study, we therefore trained novel words extensively, all items had associated meanings, and we examined dorsal and ventral pathway activation during both phonological and meaning based tasks.

The Present Study

We used an artificial language paradigm underpinned by fMRI measures of brain activity to reveal the behavioral and neural consequences of an emphasis on print-to-sound versus print-to-meaning mappings as adults learned to read new alphabetic orthographies. We used a within-subject design in which 24 adults learned to read two sets of novel words (henceforth referred to as languages) written in two different sets of unfamiliar symbols (orthographies), over a 2-week training period. Figure 1 provides some examples of the stimuli, and Appendix A shows the stimuli learned by one participant. Participants were first preexposed to the sounds (phonology) and meanings (semantics) of the novel words in each language ( Figure 2 , row A). They then learned both orthography-to-phonology (O–P, print-to-sound) and orthography-to-semantic (O–S, print-to-meaning) mappings for each language over a 2-week training period ( Figure 2 , row C). Each orthography had a systematic one-to-one correspondence between print and sound, and an arbitrary whole-word correspondence between print and meaning. Our artificial languages therefore had writing systems that were similar to those of natural languages with transparent orthographies, such as Spanish or Italian. However, we manipulated the focus of learning: for one language participants received three times as much training on O–P mappings and for the other they received three times as much training on O–S mappings.

Examples of stimuli used and illustration of how the learning focus manipulation was implemented. Note that this represents the experience of one participant as the assignment of orthography to spoken word set, noun set, and learning focus was counterbalanced across participants, as detailed in Appendix B . See the online article for the color version of this figure.

Overview of procedures, task details, key questions, data sets, and results for each part of the experiment. Column 1 gives an overview of each training procedure. Column 2 gives further details of the behavioral and MRI protocols. In rows B and E, which show the trial format and timing during MRI scans, dotted lines indicate correspondence between stimulus presentation and scan onset. Black outlined boxes show what participants were viewing, and what they were hearing, thinking, or saying is shown above each box. Column 3 delineates the key questions addressed by each part of the experiment, column 4 shows where the results can be found, and in column 5 ticks and crosses indicate whether each prediction was confirmed by the data. See the online article for the color version of this figure.

All aspects of the stimulus sets were counterbalanced across subjects, including which set of meanings was associated with which set of spoken words, which set of spoken words was written in which orthography, and which training focus was associated with which orthography (full counterbalancing details provided in Appendix B ). Thus, any observed influences of training focus on learning could not be attributed to any inadvertent differences between the sets of spoken words, symbol sets, or meanings.

We measured behavioral performance throughout the course of training, when the relative amount of exposure to orthography–phonology and orthography–semantic mappings varied between the two languages. Literacy acquisition in this laboratory model thereby enabled us to assess the behavioral consequences of an emphasis on print-to-sound versus print-to-meaning relationships for reading aloud and comprehension of printed words. We also used brain imaging to assess the neural impact of the different training protocols in three different ways.

The existing literature suggests that learning print-to-sound and print-to-meaning mappings should engage the dorsal and ventral pathways of the reading network respectively. To determine whether this was indeed the case, we measured neural activity while participants learned print-to-sound mappings for the O–P focus language and print-to-meaning mappings for the O–S focus language. This was participants’ first exposure to the two artificial orthographies. Figure 2 , row B provides further details about MRI Scan 1, which also included an English word and pseudoword reading task.

Following 2-weeks of intensive behavioral training, participants underwent a second scanning session (MRI Scan 2), in which they generated pronunciations (reading aloud) and meanings (reading comprehension) of trained items from both languages. Details of these tasks are provided in Figure 2 , row E. This enabled us to examine whether the two training regimes (O–P vs. O–S focus) differentially impacted activity in the dorsal and ventral pathways of the reading network during both reading aloud and reading comprehension. We anticipated that training focused on print-to-sound, rather than print-to-meaning, mappings should increase the efficiency of the phonologically mediated pathway. Thus, by the end of training, activity in brain regions along the dorsal pathway (e.g., inferior parietal sulcus, inferior frontal gyrus) should be reduced during reading aloud, reflecting less effortful processing for the print-to-sound than the print-to-meaning focused language. In the context of this artificial language, we can also address whether the converse is true—that is, whether training focused on print-to-meaning mappings increases the efficiency of the direct pathway. If so, then by the end of training, we would expect activity in brain regions along the ventral pathway (e.g., anterior fusiform gyrus) to be reduced during reading comprehension, indicating less effortful processing for the print-to-meaning than the print-to-sound focused language. Such an outcome could indicate that there are positive neural consequences of “balanced literacy” reading instruction programs that emphasize print-to-meaning relationships.

In MRI Scan 2, participants also read aloud untrained items from both the artificial

Artificial orthographies

Two sets of 20 unfamiliar alphabetic symbols were selected from two different archaic orthographies (Hungarian Runes, Georgian Mkhedruli). Each phoneme from the two languages was associated with one symbol from each orthography, for example, the phoneme /b/ was associated with one Hungarian symbol and one Georgian symbol. Thus, there was an entirely regular, or one-to-one correspondence between symbols and sounds in each orthography. Each participant learned to read a set of trained items from one language written in Hungarian Runes and a set of trained items from the other language written in Georgian Mkhedruli. The assignment of language to orthography was counterbalanced across participants, as detailed in Appendix B . Some examples of items written in the artificial orthographies are shown in Figure 1 , and the full set of items learned by one participant is shown in Appendix A .