EPS mid-career prize lecture 2017: Writing systems,reading,and language

Spoken language basis of reading and writing

The concept of primacy dominates virtually all scholarly thinking about reading and writing. In the discipline of linguistics, spoken language is characterised as having primacy over written language (e.g., Sapir, 1921; Saussure, 1966). Written languages are viewed as reflections of spoken languages, as emerging from spoken languages, and have even been characterised as a form of language technology (Sproat, 2010), of little theoretical interest in their own right. Indeed, spoken languages are found in all societies, sometimes with and sometimes without written language, but the reverse never occurs. Similarly, in the domain of cognitive psychology, reading is characterised as an “accessory that must be painstakingly bolted on” to the spoken language system (Pinker, 1997, p. ix). Babies are born with the ability to acquire spoken language, but children learn to read only through years of formal instruction and practice. Even when the reading skill is fully acquired, leading scholars characterise it in terms of spoken language—“language at the speed of sight” (Seidenberg, 2017).

The notion that reading is dependent on spoken language has been at the heart of psychological approaches to understanding reading since its first treatment as a scientific phenomenon. Huey (1908) described silent reading as involving “inner speech” (p. 117) and of words being “mentally pronounced” (p. 118). In contemporary research, the most influential theories posit that “decoding” the printed word to a spoken language representation is an essential component of successful reading comprehension (e.g., Gough & Tunmer, 1986). On this perspective, the decoding process allows the reader to activate an existing spoken language representation and to derive meaning on the basis of that pre-existing vocabulary knowledge. Thousands of articles have been devoted to understanding the impact of phonological decoding ability on reading success (e.g., Lervåg, Hulme, & Melby-Lervåg, 2017), the impact of restricted access to spoken language representations on deaf children learning to read (e.g., Marschark & Harris, 1996), how irregularity in the spelling-to-sound mapping poses challenges for reading acquisition (e.g., Seymour, Aro, & Erskine, 2003), and how best to teach phonological decoding skills (e.g., Johnston & Watson, 2005). Similarly, debate in the skilled reading literature has focused on how phonological decoding affects reading comprehension (e.g., Van Orden, 1987), whether phonological decoding is obligatory (e.g., Frost, 1998) or “fast” (e.g., Rastle & Brysbaert, 2006), and the mechanisms that underpin the translation of letters into sounds (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001; Plaut, McClelland, Seidenberg, & Patterson, 1996).

The conceptualisation of reading and reading acquisition in terms of the primacy model is to a certain extent appropriate. There is a broad consensus that reading acquisition in alphabetic writing systems is strongly determined by a child’s ability to decode the printed word to a spoken language representation (see, for example, Castles, Rastle, & Nation, 2018, for discussion), and there is no doubt that such representations are computed and influence word recognition and comprehension in skilled readers (e.g., Leinenger, 2014, for review). Similarly, evidence from functional magnetic resonance imaging (MRI) reveals that brain systems underlying print and speech processing converge in very similar ways across a wide spectrum of contrasting languages, suggesting that reading acquisition may be constrained by the organisation of the brain network underlying speech (Rueckl et al., 2015). But I will argue that by viewing reading through the primacy lens, we miss much of what there is to see in written language and we fail to appreciate fully how the challenge of skilled, rapid reading is met.

Divergence of written and spoken language in antiquity

One theme of this article is that writing has diverged from spoken language—that it communicates different information to spoken language—and that this is key to understanding the problem of skilled reading. In antiquity, writing was seen as a direct representation of spoken language. Classical Greek and late Classical Latin texts were written in scriptura continua, mimicking the nature of spoken language, with no spacing between words or other graphic conventions (Saenger, 1997). Within the stream of text, there was a very high degree of transparency between the sounds of the spoken language and the symbols used to write those sounds. These texts were typically read aloud, in groups or individually in a soft voice (Saenger, 1997). Reading in this period thus consisted of overt translation to the spoken language, and because writing consisted of such a high-fidelity record of the spoken language, it was thought to facilitate this translation.

However, slowly the graphic conventions apparent in modern writing did emerge, including spaces between words, punctuation, and experimentation with the visual layout of text (Parkes, 1992). Palaeographic study reveals that the emergence of spacing and other graphic conventions began around the 7th century, and it has been argued that that this coincided with the emergence of silent reading (Saenger, 1997). It is not possible to find causal data linking these two milestones. However, the introduction of spacing to mark word boundaries is rather perplexing if written language is a simple record of spoken language. Indeed, the introduction of spacing actually broke the very tight link between written and spoken language, as word boundaries are not present in continuous speech. Thus, this innovation in writing practice introduced inconsistency in the mapping between spelling and sound, in what was otherwise a highly transparent relationship.

If written language is an expression of spoken language, and if reading is built on a spoken language foundation, then why should introducing inconsistency in the relationship between print and sound have occurred and been sufficiently successful to be retained? To answer this question, we need to consider more carefully the proposal that meaning can be derived from text through translation to spoken language representations. In particular, on considering the nature of the spoken language representations that can be accessed through a decoding process, it is immediately apparent that these are highly impoverished relative to the rich information characteristic of actual spoken language. When we compute meaning from spoken language, we take advantage of coarticulatory cues to predict upcoming sounds (e.g., McMurray & Jongman, 2016; Salverda, Kleinschmidt, & Tanenhaus, 2014), prosodic and intonation cues to predict aspects of meaning in sentences (e.g., Grosjean & Gee, 1987) and single words (e.g., Soto-Faraco, Sebastián-Gallés, & Cutler, 2001), information about the context and intent of the speaker (e.g., McQueen, 2007, for discussion), along with audio-visual (McGurk & MacDonald, 1976) and gesture (Kita, 2000) cues. No matter how tight the link between written and spoken language, these forms of information will always be absent in text.

This discussion reveals the weakness of viewing reading and writing through the primacy lens. The classical scriptura continua manuscripts provided the closest possible link between written and spoken language. But a simple analysis exposes the paucity of information in writing. Translating the written language back to sound-based representations presents the reader with many of the challenges of understanding spoken language, but without any of the usual cues that support this process. Although this form of analysis might allow slow, overt reading aloud by trained orators, it is unlikely that it could support the rapid analysis of text that characterises modern, skilled reading.

This insight suggests that while weakening the link between spoken and written language, perhaps the introduction of graphic conventions was related to the development of a more advanced form of reading. Indeed, we know from modern research that word spacing has a powerful impact on eye-movement control and word identification. Unsurprisingly, removing word spacing in typically spaced formats such as English sentences slows reading (Rayner, Fischer, & Pollatsek, 1998). But conversely, introducing word spacing to typically non-spaced formats such as Japanese Hiragana (Sainio, Hyönä, Bingushi, & Bertram, 2007) and Thai (Kohsom & Gobet, 1997) speeds reading. More recent research on Chinese (another non-spaced script) suggests that the addition of spacing facilitates word identification by reducing fixation time as well as the number of refixations in child and adult readers (Zang, Liang, Bai, Yan, & Liversedge, 2013), perhaps because physical demarcation of word boundaries influences the computation of the next fixation location (Zhou, Wang, Shu, Kliegl, & Yan, 2018). Spacing also assists children to learn new vocabulary words printed within Chinese sentences (Blythe et al., 2012).

Our consideration of the limits of information within the classical scriptura continua texts also helps to explain the geographical distribution of the introduction of spacing and other graphic conventions. Paleographic analysis suggests that the introduction of these conventions began in Ireland before spreading to the European continent (Parkes, 1992). This is significant because for the Irish monks who had embraced Christianity, Latin was a foreign language (Parkes, 1992). In the absence of lexical knowledge to guide segmentation of the text, seeking to access meaning even through overt reading aloud would have presented a very serious challenge indeed. The introduction of word spacing in these texts probably made reading possible, and as such, also provided a scaffold for these monks to learn the spoken language.

This brief introduction to the development of graphic conventions in antiquity serves two purposes. First, it reveals that breaking the link between spoken and written language may have been functional in the development of skilled, silent reading. This position is contrary to much of the recent theoretical work on reading acquisition, which posits that the convergence between spelling and sound in a writing system is a major determinant of reading success (e.g., Seymour et al., 2003; Spencer & Hanley, 2003). Second, it provides a simple example of how written language communicates different information to spoken language. Indeed, I have argued that these additional graphic conventions may have been necessary to overcome the paucity of information in the written signal, as reading transitioned from use in public oration to the silent, rapid analysis of printed text associated with contemporary usage (see also Saenger, 1997).

The information in English writing

The central challenge of reading is learning the relationship between visual symbols and their meanings. Recent empirical and theoretical work suggests that precisely how this challenge is met may depend on the writing system (Frost, 2012). Thus, achieving an understanding of reading and reading acquisition requires reference to the nature of information available in a particular writing system. In this article, I describe the relationship between the writing system and the reading system with reference to English. One might argue that this is an example of Anglocentricity in reading research (Share, 2008), and I do not disagree that the trajectory of reading research has been driven to some extent by the accessibility of English within research-intensive universities. However, as I hope will become clear, achieving an understanding of the reading system requires a deep understanding of the writing system, and the focus on a single writing system provides an opportunity to develop and address a number of issues of theoretical importance. Furthermore, English is particularly interesting to study because it is a self-organising system, allowed to vary freely over time (Berg & Aronoff, 2017). This contrasts with many major writing systems (e.g., Dutch, French, Welsh, Norwegian, Malay, Portuguese, English braille), in which spellings have been standardised by national or supra-national bodies. The self-organising nature of English writing may mean that it transmits information in a more complex or nuanced manner than in more regulated systems in which particular forms of information are deliberately emphasised. If reading systems are closely linked to writing systems, then this may have demonstrable consequences for reading and reading acquisition. However, although I focus on English, I hope to draw out more general lessons about the relationship between writing systems, reading, and language, in a manner that can be evaluated in future research.

Writing systems are all expressions of spoken language, but they vary considerably, even among the alphabetic writing systems that are the focus of this article (see, for example, Daniels, 2018). Drawing from the primacy model, a great deal of research concerned with alphabetic writing systems has focused on the consequences of orthographic depth (Katz & Frost, 1992). Orthographic depth refers to the transparency with which visual symbols relate to sounds in a writing system. Writing systems with high transparency between spelling and sound (e.g., Spanish) are known as shallow orthographies, while writing systems with low transparency (e.g., English) are known as deep orthographies. However, it is worth remembering that even shallow orthographies consist of additional information not available in the acoustic signal. For example, Korean has very high transparency between symbols and sounds, but the written language physically demarcates word and syllable boundaries, and the shapes of the letters themselves depict the articulatory organs that would be needed to pronounce them.

The characterisation of writing systems in terms of spelling-to-sound transparency has had a profound impact on reading research. Research has been concerned with how orthographic depth influences the trajectory of reading acquisition (e.g., Caravolas et al., 2012; Seymour et al., 2003), the nature of the representations acquired (e.g., Ziegler & Goswami, 2005), and the development of the reading network in the brain (e.g., Paulesu et al., 2000; Rueckl et al., 2015). In particular, decades of research have been devoted to challenges associated with the very low degree of spelling–sound transparency in English, in which there are multiple possible pronunciations for particular letters and multiple possible spellings for particular sounds (e.g., Baron & Strawson, 1976; Ziegler, Stone, & Jacobs, 1997). This phenomenon has been a driving force in defining the shape of the most important theoretical approaches to reading (e.g., Coltheart et al., 2001; Perry, Ziegler, & Zorzi, 2007; Plaut et al., 1996; Seidenberg & McClelland, 1989; see also Castles et al., 2018; Share, 2008, for discussion). The classification of words based on spelling–sound transparency has also been central to initial work on the genetic (Bates et al., 2007) and neural (Taylor, Rastle, & Davis, 2013) underpinnings of reading.

English has an alphabetic writing system and so its primary regularities are indeed between spelling and sound (more specifically, between graphemes and phonemes). However, the ultimate goal of reading is not to map symbols to sounds, but to map symbols to meanings, and this consideration brings different forms of regularity into view. These regularities arise largely because of the morphological structure of English. That is, the vast majority of English words are built by combining a finite number of stem morphemes (e.g., book) with other stem morphemes (e.g., bookworm) or affix morphemes (e.g., bookish). This combinatorial organisation means that stems reoccur in words with similar meanings (e.g., kindness, unkind, kindly) and that affixes modify the meanings of stems in a predictable manner (e.g., kindness, darkness, shyness; Plaut & Gonnerman, 2000; Rastle, Davis, Marslen-Wilson, & Tyler, 2000). The computational models of reading described above have focused largely on single stems in which the only regularity is between spelling and sound; the relationship between spelling and meaning for such words is arbitrary (e.g., pint and mint share three quarters of their letters but do not overlap in meaning). It is perhaps for this reason that regularities between spelling and meaning have been underappreciated in the reading literature. But if we consider the whole lexicon, made primarily of words with more than one morpheme, then regularity between spelling and meaning is abundant. ¹

The characterisation of spelling–meaning (morphological) regularities in written language, and their impact on reading acquisition and skilled reading has been treated as somewhat of a niche area of study. However, morphological regularity is intimately linked to the spelling—sound regularities that have been the focus of much of the work on reading. It has long been observed that meaningful information is more apparent in English spelling than in spoken language (e.g., Carney, 1994; Chomsky, 1970). For example, the words herded, kicked, and snored all relate to something that happened in the past, but the past-tense status of these words is far easier to appreciate in their spellings (i.e., the letters [-ed] denote the past tense) than in their spoken realisations (in which past tense is associated with three different sounds or sound sequences; /ԑd/, /t/, and /d/). Similarly, in words such as magician and health, it is far easier to extract meaningful components (i.e., that these words have something to do with magic and heal) from their spellings than from their spoken realisations. These examples show how written language can transmit meaningful information more effectively than spoken language. However, the consequence of preserving meaningful information in the spellings of these words is divergence between spelling and sound. That is, marking the past tense consistently through use of the word-final letters [-ed] necessarily means that the relationship between these letters and their corresponding sounds is inconsistent. If English writing were a perfect reflection of spoken language, the words herded, kicked, and snored might be spelled herdid, kickt, and snord, and the words magician and health might be spelled majishun and helth, losing the meaningful information that the English written forms convey. Thus, like the introduction of word spacing, these examples demonstrate how divergence from spoken language can be functional, by enhancing the transmission of meaningful information.

Anecdotal examples like these are often recounted as evidence of the visibility of meaningful information in English writing, but only recently has linguistic analysis begun to quantify the relationship between English spelling and meaning precisely. In their study of the past tense, Berg and colleagues noted that there are only 38 English words comprising a single morpheme that contains word-final [-ed]. Of these, 8 do not have a plausible stem (e.g., bed, shed) and 20 occur as part of a different vowel grapheme (e.g., breed, feed); there are only 10 instances that could support an erroneous morphemic reading (e.g., wicked, hatred, moped). The relative absence of this latter class of words means that the presence of [-ed] is a highly reliable cue to past tense (Berg, Buchmann, Dybiec, & Fuhrhop, 2014). Berg et al. (2014) further noted that the high fidelity of information denoted by the spelling [-ed] is achieved because words comprising a single morpheme that could legitimately be spelled with [-ed] are typically spelled another way (e.g., instead, salad, horrid; Berg et al., 2014). Thus, the spelling [-ed] appears to have become reserved for communicating the past. Berg et al. (2014) reported similar findings for word-final [-s]: use of word-final [-s] is very rare in words comprising a single morpheme (e.g., lens); such words that could be spelled using [-s] are typically spelled another way (e.g., please, sneeze, box).

Berg and Aronoff (2017) articulated a similar relationship between spelling and grammatical class achieved in derivational suffixes. For example, they observed that there are many possible spellings for the word-final sound sequence /әs/ (e.g., bonus, service, nervous, princess), but that the spelling [-ous] is virtually always used if the word is an adjective. Furthermore, while it could be that the spelling [-ous] is also used to indicate other grammatical classes, this is not the case; word-final spelling [-ous] always indicates adjective status (e.g., nervous, joyous, famous). Thus, as for the past-tense marker, the suffix [-ous] appears to be reserved to communicate adjective status. Berg and Aronoff (2017) studied four suffixes and observed similar regularities across all of them. Furthermore, a diachronic analysis revealed that these regularities appear to have become more visible over the last 1,250 years of English spelling change.

Recent computational work demonstrates that this strong relationship between suffix spelling and grammatical class is a general principle of English suffixation (Ulicheva, Harvey, Aronoff, & Rastle, 2018). Ulicheva et al. (2018) studied the relationship between suffix spelling and grammatical class for more than 150 English suffixes; this relationship is visualised in Figure 1 for two suffixes. The first example shows that there are many possible spellings of the word-final sound sequence /lәs/ (e.g., [-lace], [-less], [-liss]). But only one of those spellings is possible if the word is an adjective, and adjectives almost without exception are spelled with [-less] (e.g., fearless, helpless, joyless). The second example shows precisely the same pattern. There are many possible spellings of the word-final sound sequence /Ikәl/ (e.g., [-ickle], [-ycle], [-ickel], [-ical]). But again, with just a few exceptions, one spelling seems to be reserved for adjectives.

OPEN IN VIEWER

Figure 1.

Illustration of how English spelling communicates meaningful information.

These examples give us pause to consider the relationship between English spelling-to-sound and spelling-to-meaning regularities. Specifically, discussion of English writing has typically focused on the apparent disorder created by the proliferation of spellings for particular sound sequences. But it is only through this fractionation that particular spellings can be used reliably to transmit meaningful information. Furthermore, these meaningful regularities are not present in the spoken language; the very strong information about grammatical class arising through suffixation is present only in the written language. Thus, we can see that spelling–meaning regularity and spelling–sound irregularity are two sides of the same coin. The diversification of spelling maximises the transmission of meaningful information, but this comes at the cost of an absence of a tight link between spelling and spoken language.

The extent to which this trade-off between phonology and morphology is characteristic of writing systems in general is an open question that requires investigation. Certainly, morphology is a feature of virtually all languages, and many languages have far richer morphology than English. But for the purposes of understanding reading and reading acquisition, we are interested in the visibility of meaningful information in the writing, not in the spoken language. I’ve argued that the high visibility of meaningful morphological information in English writing is a consequence of the fractionation of English spelling (i.e., multiple potential spellings for the same sound sequence). It is certainly possible to find examples of writing systems that are reasonably transparent in terms of the spelling–sound mapping, but which nevertheless convey morphological information in spelling. For example, the past–present distinction in Dutch is sometimes conveyed through spelling alterations in a manner that distinguishes homophonic words (as in zij betwisten [they dispute] vs. zij betwistten [they disputed]; Brysbaert, Grondelaers, & Ratinckx, 2000), but this is not characteristic of Dutch spelling in general. Indeed, I would argue that the closer a writing system gets to a one-to-one mapping between spelling and sound, the less scope there is for highly reliable morphological cues to arise in the spelling, unless particular sound sequences themselves are reserved to communicate meaningful morphological information. However, this proposal is speculative; moving forward towards a more general understanding of how writing systems balance spelling–sound and spelling–meaning information will require much more detailed analyses of contrasting writing systems along the lines of those undertaken by Ulicheva et al. (2018).

More broadly, it seems almost inconceivable that the trajectory of reading acquisition, and the nature of representations characterising skilled reading, would not be substantially influenced by the nature of the information being learned (see also Frost, 2012). Yet, while there have been decades of research on reading and reading acquisition, research has only just begun to articulate the nature of information transmitted through the orthographies of the world’s languages. I would argue that understanding more about how orthographies convey meaning—formulating quantitative models of the nature of information communicated in written languages—will be an essential part of developing a deeper knowledge of reading and reading acquisition into the future.

The writing system reflected in the reading system

Becoming a skilled reader requires instruction and massive text experience over many years (see Castles et al., 2018, for review). Ultimately, all theories of skilled reading in alphabetic systems agree that the outcome of these years of text experience is a reading system whereby visual forms map onto meaning via two broad pathways (e.g., Coltheart et al., 2001; Harm & Seidenberg, 2004; Woollams, Ralph, Plaut, & Patterson, 2007). One pathway maps visual forms onto meanings via sound-based (phonological) representations, while the other pathway maps visual forms onto meanings directly. Recent neuroimaging studies suggest that this dual-pathway architecture is also observed in the brain, with a dorsal pathway representing spelling-to-sound (Taylor et al., 2013) and sound-to-meaning (Hoffman, Lambon Ralph, & Woollams, 2015) computations, and a ventral pathway representing direct access to meaning (e.g., Fischer-Baum, Bruggemann, Gallego, Li, & Tamez, 2017; Taylor et al., 2013). In the following, I describe how the spelling-to-sound and spelling-to-meaning regularities that I have described in English writing are represented in these pathways and the extent to which they contribute to skilled, silent reading (Figure 2).

OPEN IN VIEWER

Figure 2.

Dual-pathway architecture of skilled reading and its neural underpinnings.

Spelling–meaning (morphological) regularities and orthographic learning

I have argued that the acquisition of spelling–sound knowledge is necessary in reading acquisition in alphabetic writing systems, but that it is unlikely that this knowledge alone could support skilled, silent reading in English. Instead, contemporary theories of skilled reading propose that readers must also acquire a direct mapping between printed words and their meanings (e.g., Coltheart et al., 2001; Harm & Seidenberg, 2004; Woollams et al., 2007). This direct mapping is represented in ventral pathway brain regions (Taylor et al., 2013), which become increasingly tuned to printed words as readers develop greater skill (Dehaene-Lambertz, Monzalvo, & Dehaene, 2018) certainly into the period of adolescence (Ben-Shachar, Dougherty, Deutsch, & Wandell, 2011). Much less is known about the acquisition of direct links between spelling and meaning, but it is thought that this process of “orthographic learning” arises through an accumulation of experience with printed text (e.g., Castles & Nation, 2006; Nation, 2017), as a child uses their spelling–sound knowledge as a self-teaching device (Share, 1995). The self-teaching theory proposes that successful, repeated decoding of an unfamiliar word into a spoken language representation provides a mechanism through which to establish the orthographic lexical representations necessary for rapid word recognition (Share, 1995).

Most of the research on the acquisition of the direct pathway linking spellings to meanings has focused on item-level effects—the journey of a particular word to one that can be recognised rapidly (Castles & Nation, 2006; Nation, 2017). Indeed, theoretical work in this area suggests that at any point in time, there will be some words that a child can recognise rapidly in an item-based manner, while recognition of other words will require an analytic decoding process (Castles et al., 2018). Simulation models have investigated how this orthographic learning might arise within a self-teaching framework (Pritchard, Coltheart, Marinus, & Castles, 2018; Ziegler, Perry, & Zorzi, 2014). In the developmental model of Ziegler et al. (2014), rudimentary knowledge of a small set of grapheme–phoneme relationships is used to decode novel words; if a decoded pronunciation matches a representation in the phonological lexicon, a representation in the orthographic lexicon is learned, and the successful decoding experience is used as an internal teacher to improve grapheme–phoneme knowledge. The model of Pritchard et al. (2018) went on to explore the impact of context on orthographic learning under a variety of circumstances. However, although these models provide proof-of-concept that orthographic representations can be acquired through a self-teaching mechanism, they do not address the question of how these orthographic representations map onto semantic contents. This would seem to be a difficult problem in English (as in other alphabetic languages) because of the sheer scale of the challenge: It is estimated that the average 20-year-old English reader can recognise around 70,000 unique words (Brysbaert, Stevens, Mandera, & Keuleers, 2016). One counterargument might be that this type of item-based learning is similar to what is involved in learning Chinese characters. ² However, estimates of the number of characters that must be learned for full adult literacy in Chinese are an order of magnitude less than the requirement for English, ranging from 4,300 characters (Katz & Frost, 1992) to between 7,000 and 9,000 characters (Pine, Ping’an, & Song, 2003).

It turns out that the problem of how English readers can possibly acquire item-specific knowledge of so many tens of thousands of words is somewhat of a red herring, because the English writing system is structured such that orthographic learning does not need to arise in an item-based manner. That is, the challenge of learning the spelling–meaning mapping is dramatically reduced if the learner has an appreciation of how morphological relationships are expressed in the writing. This principle is illustrated in Figure 3. The first panel of the figure lists 15 unique English words; however, if the learner has an understanding of morphology, then these 15 words become variations of a single word. It is unlikely that these words would need to be learned in an item-specific manner, one at a time, because learning the meaning of one word assists learning the others. In fact, in the estimates of vocabulary size recently published by Brysbaert et al. (2016), the number of unique words that the average 20-year-old can recognise drops from 71,000 to 42,000 if inflections (e.g., develops, developing, developed) are not counted as unique, and this number drops further to a much more manageable 11,100 if derivations (e.g., developer, redevelop) are not counted as unique. It seems important that this basic fact about the writing system should feature in theories of orthographic learning.

OPEN IN VIEWER

Figure 3.

Illustration of the impact of morphological knowledge on orthographic learning and orthographic representation.

Taking advantage of these regularities requires that the reader has an understanding of morphological relationships. But what does this mean in mechanistic terms? It means that the orthographic input needs to be structured in such a way that reveals that they share the same stem [develop], as illustrated in the second panel of the figure. Yet, to acquire overlapping representations of this nature, the reader must appreciate that the letter groups at the beginnings and endings of these words are somehow significant and should be separated from the stem [develop] in a way that would not occur for other cases in which words are embedded in other words (e.g., shallow, formula). To understand how this might be achieved, we again need to go back to the information available in the writing system, and specifically to the observation that the large number of English spellings for particular sound sequences permits some spellings (affixes) to become reserved to communicate particular meanings. If learners become sensitive to this highly reliable information about the functions of particular letters and letter groups, then that ultimately will allow them to structure orthographic representations in such a way that captures the meaningful relationships between words (as in the second panel of Figure 3; see also Rastle & Davis, 2008). It seems plausible that limited morphological knowledge may act as a schema with which to interpret new words and conversely, that appropriately structured orthographic representations of these new words may further strengthen morphological knowledge. In this way, knowledge of this feature of the writing system may support a form of morphological self-teaching that facilitates ongoing orthographic learning, as readers encounter increasingly complex derivations through their text experience. However, further computational modelling work is required to develop this proposal, and of course it is also subject to empirical evidence.

Recent work using laboratory acquisition paradigms has begun to articulate how readers might discover this powerful morphological information (e.g., Tamminen, Davis, & Rastle, 2015). In these laboratory experiments, adult participants are taught novel vocabularies with a morphological structure (e.g., sleepnule, buildnule, teachnule), and then probed some days later in a variety of ways to assess their knowledge of the novel affix (e.g., [-nule]), including online tasks requiring generalisation to untrained items (e.g., sailnule). Tamminen et al. (2015) reported that learning of the novel affix is dependent on (a) the consistency with which the affix spelling maps to a particular meaning and (b) the number of unique stems to which the affix attaches (i.e., family size). These findings suggest that we acquire knowledge of spellings that consistently signify particular meanings across multiple diverse contexts. This conclusion seems consistent with work investigating morphological representation in skilled readers (e.g., Ford, Davis, Marslen-Wilson, 2010) but further work is required to assess the validity of these predictions fully. More generally, the development of a theory of how we acquire this form of knowledge will be enriched through contact with research on other forms of language learning, including statistical learning (e.g., Gómez, 2002) and novel word learning (e.g., Joseph & Nation, 2018; Nation, 2017), as well as research on the impact of explicit instruction on the development of long-term knowledge (e.g., Kirschner, Sweller, & Clark, 2006).

Conclusions and emerging questions

In this discussion, I have put forward the view that the study of reading and reading acquisition is the study of how information transmitted through writing becomes represented in the minds and brains of individuals, through an accumulation of instruction and text experience. This broad perspective seems uncontroversial, and yet, it is striking how little we understand of the nature of information transmitted through writing. Written language is an expression of spoken language, but the nature of information transmitted through these forms of language differs. Examples going back to antiquity show how written language has diverged from spoken language in ways that support skilled, silent reading. Understanding any aspect of reading acquisition requires a deep appreciation of what it is that is being learned; this knowledge provides the foundation for asking the right questions and helps us to make sense of behaviour. Yet, our field has produced thousands of journal articles and monographs on the problem of reading and reading acquisition, including differences in these processes across languages, while only just scratching the surface of characterising written language(s) in a quantitative manner. Addressing this gap presents a major challenge for future research.

I have conceptualised English writing (and therefore, reading) as reflecting a trade-off between spelling–sound regularity and spelling–meaning regularity. It is well known that English writing is a relatively poor reflection of English spoken language. Indeed, this theme has driven much of the scholarly thinking on reading instruction, reading acquisition, and skilled reading. However, I’ve argued here that spelling–sound irregularity and spelling–meaning regularity are two sides of the same coin. The former is a necessary condition of the latter. The proliferation of spellings for English sound sequences weakens the relationship between spelling and sound, but permits some spellings to become reserved for the transmission of specific aspects of meaning (Berg & Aronoff, 2017; Ulicheva et al., 2018). The transmission of this highly consistent meaningful information would not be possible if English spelling were a totally faithful record of the spoken language with a one-to-one mapping between spellings and sounds. I have argued that English readers capitalise on both of these forms of regularity, although in different ways, at different time points in the process of reading acquisition, and underpinned by different brain pathways. This perspective turns the dominant narrative of English reading and writing on its head. The disorder of the English spelling–sound mapping is not some nasty feature of the writing to be overcome, but a consequence of order in the English spelling–meaning mapping. I’ve argued that this order in the spelling–meaning mapping may be vital for orthographic learning as text experience accumulates through the period of adolescence and beyond, and for rapid word recognition in skilled, silent reading.

The proposal that systematicity in the spelling–meaning mapping may be particularly important in supporting skilled, silent reading in English raises an interesting question about whether readers of writing systems that prioritise systematicity in the spelling–sound mapping might somehow be at a disadvantage. It is tempting to argue that the answer is “no,” for the reason that writing systems all provide optimal solutions to representing spoken language—that “most languages get the orthography they deserve” (Katz & Frost, 1992, p. 67). On this view, although writing systems may differ widely, every writing system itself provides the most efficient path into the spoken language that it represents. The reading system then becomes tuned to the specific solution for that language through the process of reading acquisition. Various examples of the ways in which writing systems are optimally tuned to represent spoken language were described by Frost (2012). This is an attractive proposal, and it is easy to find examples that offer broad support. However, it is also easy to find examples in which writing systems do not provide optimal solutions for representing spoken language, and we therefore need to be cautious about accepting this proposal uncritically.

One particularly interesting example of the latter situation is found in English braille. The most common form of English braille is contracted braille, in which highly familiar words and letter clusters (e.g., EA, ST, ED) are represented in a single cell (a cell is the 2 × 3 matrix of raised dots that constitutes the basic unit of braille writing). Although uncontracted braille (a writing system that does not use contractions) is used in the very early stages of learning to read, contracted braille is preferred because it saves space in printing and is also thought to permit more rapid reading (see Fischer-Baum & Englebretson, 2016, for discussion). Braille contractions are language specific, and in English, are regulated by the International Council of English Braille (see, for example, Simpson, 2013). However, Fischer-Baum and Englebretson (2016) identified a potential problem with these contractions: They often straddle and thus obscure morpheme boundaries (e.g., miSTrust, rERun, milEAge; in these examples, the letters in capitals are expressed in a single Braille cell). Fischer-Baum and Englebretson (2016) went on to show that the obscuring of the morpheme boundary in these forms disrupts word recognition. Thus, here we have a case in which the writing system is clearly non-optimal in the way that it expresses spoken language. Of course, one could argue that this is an unusual case due to the intense regulation of English braille. But writing systems are altered through human intervention frequently and for many reasons, and it is unlikely that those alterations always lead to more efficient transmission of meaning.

Moving forward, a number of emerging questions come into view. Becoming a skilled reader requires case-specific but also general knowledge of regularities in the writing system being learned. This latter aspect of the acquisition process permits the learner to deal with unfamiliar words—to read new words aloud (e.g., blog), to determine the meanings of new words (e.g., untweetable), and to spell new words in a plausible manner (e.g., that word-final [-less] should not be used to spell a noun). The extent to which a learner comes to appreciate the item-specific and general information available in a writing system depends on text experience over many years, along with the nature of any instruction provided (e.g., training on spelling–sound relationships via systematic phonics). However, we do not have a good understanding of how these factors come together to affect adult reading skill. Research has begun to address these questions—how the knowledge underpinning reading builds up over time (e.g., Monaghan & Ellis, 2010), whether certain learning experiences are more important than others (e.g., Joseph & Nation, 2018; Zevin & Seidenberg, 2002), whether explicit training on aspects of a writing system improves learning (Powell, Plaut, & Funnell, 2006), and how pre-existing language ability may modulate the impact of that explicit training (Chang, Taylor, Rastle, & Monaghan, 2017). However, much further work on these dimensions of experience needs to be conducted to understand reading acquisition. This work would also need to include contact with the memory literature (e.g., McClelland, McNaughton, & Reilly, 1995) to ensure that theories of the accumulation of knowledge in reading acquisition are consistent with what is known of the nature of short- and long-term learning processes.

Finally, while I have discussed how the information in writing affects reading and reading acquisition, an interesting question is whether this relationship also goes in the reverse direction. Is it possible that cognitive constraints on reading acquisition affect the evolution of writing systems? No doubt, there are multiple social, historical, and political forces that can give rise to spelling change. However, the proposal that cognitive constraints might also affect the evolution of spelling seems worthy of consideration, particularly given research suggesting that properties of human learning may affect the structure of spoken languages (e.g., Christiansen & Chater, 2008; Smith et al., 2017). I have already described how writing in antiquity may have changed due to (a) lack of familiarity with the spoken language used in early Christian texts and (b) the change in the use of reading to a private, silent activity (see Parkes, 1992; Saenger, 1997). It seems possible that the introduction of spacing and other graphic conventions was driven by these cognitive challenges. The extent to which cognitive constraints may affect the evolution of modern writing systems is an open question. However, emerging evidence suggests that English spelling has changed in such a way as to make regularities between spelling and meaning more visible (Berg & Aronoff, 2017), regularities that I have suggested are critical for the rapid analysis of English words. This insight suggests that perhaps English spelling has evolved to facilitate skilled reading—and in a manner that permits greater efficiency in accessing meaning than if the spelling had been reformed to present a more direct representation of spoken language. Understanding how information transmitted through written language affects how the brain solves the problem of reading, and understanding whether the challenge of reading drives evolution of written language are important areas for future discovery.