Theoretical models of reading - better informed but none the wiser?

For skilled readers, silently reading text feels no more demanding than listening to an audiobook and in many ways can be easier: passages can be reread, and unknown words’ spellings may indicate morphology and give clues to meaning. This automatic process, whereby the working memory is liberated to concentrate almost entirely on the meaning the author was attempting to convey contrasts with the stuttering, painful progress of the beginning reader. Contrast that again with the infant who cannot read but may be able to recognise the Coca-Cola logo and the complexity of learning to read becomes almost overwhelming. Add to that the responsibility of primary school teachers (and sadly many secondary teachers) of ensuring that all children attain automatic word recognition that gifts the capacity to read for meaning, and perhaps pleasure, and we have a cognitive gordian knot.

As a result of this complexity, allied to the fact that even when we look in the brain (more here), we can only see which areas are being utilised when reading occurs, as opposed to what is actually happening, and a model of reading development may help us in deciding which approach to instruction is the most effective.


A number of models have been developed to explain how words are read by proficient readers and how both skilled readers approach the reading of regular and irregular words and familiar and unfamiliar words. The two models that have received the most attention and generated the most research are the dual route model and the parallel distributed processing models, better known as the triangle models.

In the dual route model (Coltheart, 1978), the pronunciation of a word can be generated in two ways. This is either through the application of grapheme-to-phoneme correspondence rules by breaking the word into its phonological representations, or alternatively through a more direct mapping of the spelling of the word into its pronunciation. In other words, the reader either recognises the word as a whole or sounds out its graphemes to decode the word. In the further developed Dual Route Cascade model (Coltheart et, al., 2001), these two routes work in parallel with pronunciation of any given word being jointly determined by the products of both routes.

By the dual route model words read by the direct route are processed by the lexicon’s word recognition units. Words read via this method are processed holistically, and therefore, all words, whether regular or irregular are treated in the same manner. Nonwords cannot be recognised in this way as no representations of these words exist in the memory. The indirect route converts graphemes to phonemes to enable correct pronunciation. All nonwords will be read in this way along with unfamiliar words. Irregular words read in this manner may be mispronounced which results in the conclusion that these can only be read correctly from memory of the word in the lexical store and thus require learning. Regular words will therefore be read more rapidly and accurately (more on the word method of teaching reading here).

The fundamental assumptions of the triangle models (Harm and Seidenberg, 2004) are the opposite of the dual route model. These models posit the triple (hence triangle) interacting effects of orthography, phonology and meaning (Rayner et al., 2010) work in parallel to read the word. Thus, in contrast to the dual route models, the knowledge that allows a reader to identify printed words is contained in a single set of input and output connections that enables the reader to access all available knowledge to pronounce the word. The orthography layer contains a very large, but finite, store of legitimate letter patterns and the phonology layer contains a similarly large store of sounds and sound patterns. In addition, there is a semantic layer containing a store of word meanings. Thus, reading occurs as a word’s spelling activates the units corresponding to the word within the orthography layer and interacts with the output units in the phonology layer enabling pronunciation.

Triangle models do not, therefore, assume that lexical information is represented by discrete processing units in the lexicon but accessed through the connections formed between the orthographic input and phonological output. Although children initially learn mappings between orthography and phonology, they eventually utilise meaning as well. Repeated exposure to words strengthens these connections resulting in the recognised phenomenon that familiar words are pronounced more accurately than infrequent words. All words are, therefore, read in the same way through input of spelling and output of pronunciation.

Whilst dual route models treat regular and irregular words as distinctly different word types, triangle models take into account the regularity within irregular words and the consistency of letter sound patterns. Therefore, irregular words with consistent spelling patterns will be read more quickly than unusual words.

Both models have been implemented as computer programmes to simulate tasks that have been used to study word identification. However, as Rayner and Reichle (2010) warn, these are models for reading aloud, or how a single word displayed in isolation is identified and are not therefore models of reading , although they are often referred to as models of reading. Rayner and Reichle (2010) also add the caveat that these models seldom make contact with models that address the other aspects of the reading process.

Furthermore, both dual route models and triangle models are post facto propositions that attempt to explain what is occurring in the brain with regard to what is known in terms of what can be measured (eye-movements, presentation times and word recognition speed). They are thus both to explain the recognised phenomena reported in experimentation that frequent words are identified more rapidly than infrequent words and that words with irregular pronunciations are identified less rapidly than words with more regular pronunciations. Both models are able to provide cogent explanations for reading deficits encountered by readers with phonological dyslexia and surface dyslexia. However, both models rationalise the phenomena in different ways. As Rayner et al. (2010) point out, they are models that are consistent with and currently fit our understanding of word reading; they are not a definitive explanation of how words are read.

Triangle models are often invoked to support code-emphasis approaches to reading instruction that encourage children to practice letter-to-sound mappings and apply decoding skills. In contrast, as dual route models propose a direct access account of word recognition through an orthographic-semantic route to identify the majority of words, they are often used to justify a meaning-based approach to reading that de-emphases decoding instruction (Smith, 1973) - more here. The greater efficiency of the instant orthographic-semantic processing route associated with dual route models may feel intuitively obvious. Nonetheless, computer modelling has simulated human reading data based on parallel processing assumptions (Harm and Seidenberg, 2004).

In terms of guiding pedagogy and efficacious instruction, the existence of two apparently rational models enables the justification of specific systematic phonics instruction and whole language philosophy. You pays your money and you takes your choice…

So, once again, the academic world of reading research seems to leave the world of education, to misquote the 1st Earl of Birkenhead, ‘better informed, but none the wiser’.


Whereas language acquisition is a biologically determined skill (Geary, 2007), reading is the product of cultural evolution (Rayner et al., 2010). It may, therefore, seem specious to assign a sequence of stages to the process of learning to read. Although stage theories of reading attempt to explain reading acquisition and development by the passing through of a number of sequential stages by a developing reader on their journey to becoming a competent reader, they are ‘best viewed as convenient ways to describe how reading changes as children gain skill.’ (Rayner et al., 2010:279).

Chall (1996) identifies five stages of reading development from birth to adulthood grouped into three general phases: Learning to Read, Reading to Learn and Independent Reading. In the Learning to Read phase, which occurs usually between the ages of five to eight, initial reading and decoding builds reading fluency with listening comprehension far superior to reading comprehension. Vocabulary then expands through further reading in the Reading to Learn phase, particularly in subject areas, with reading comprehension improving to align or surpass listening comprehension. By the age of ten the Independent Reading phase is entered with continued vocabulary expansion, wider subject and genre reading and reading becoming a far more efficient method of acquiring knowledge than listening (more here).

Marsh et al.’s (1981) identified four stages of early reading that sit within Chall’s (1996) Learning to Read phase: linguistic guessing, discrimination net guessing, sequential decoding and hierarchical decoding. A child at the linguistic guessing stage approaches word recognition by a strategy of rote association; they make guesses of words based on the linguistic context. Guessed words will usually be semantically and contextually appropriate but may not contain any appropriate letter combinations. Children with strong memories may be able to memorise whole stories. At the stage of discrimination net guessing, children will utilise additional cues such as word shape, word length and some letter identity to distinguish words.

As children are exposed to a greater number of words, memory load increases and whole word memorisation has diminishing returns, as does the use of discrimination net guessing as greater word exposure results in more words looking similar (because the reader encounters more words). At the same time the cognitive capacity of children increases enabling the processing of the order of letters which can be coordinated with a series of sounds resulting in the recognition of predictable relationships between individual letters and sounds. Marsh et al.’s (1981) third stage of sequential decoding is then entered characterised by the use of letter-to-sound correspondence rules. Through explicit instruction or self-teaching (Share, 1995) children learn that many letters are pronounced the same way in different words and that unknown words can be identified by extrapolating and applying this knowledge. It is at this stage that systematic synthetic phonics is considered appropriate and the most efficient method of ensuring letter-to-sound knowledge. The estimated 25% of children that cannot self-teach (McGuiness 1999; Adoniou, 2017; Turner and Burkard, 1996) remain reliant on whole-word memorisation and with the limits of logographic memory restricted to between 2000 and 5000 words (Leong, 1973, Gough and Hillinger, 1980, Akamatsu, 2006) further development stalls - Chall’s (1983) ‘grade four slump’ ‘grade four slump’ is often cited as evidence of this along with the 20% of New Zealand six year old’s unable to read after a year of whole language immersion (Chapman et al., 2001).

In Stage three of the Marsh et al. (1981) theory, decoding is relatively basic with one-to-one correspondence predominately applied, but as the child gains experience with decoding letter-sound correspondences become context sensitive and the child will use analogy as an alternative strategy for decoding. This is the start of Marsh et al.’s fourth stage; the hierarchical decoding stage. However, as the complexity of the code unravels, decoding is often slow and not always accurate. Word recognition is dependent on letter-sound relationships which support spelling by enhanced memory for the order of letters within words beyond the memory of single words (Ehri, 2002). It is in this phase that children may appear to struggle the most, read slowly and tire quickly. It was probably when listening to children in this phase of development that aHM inspector noted the, ‘the dismal and unnatural monotony of sound which pervades every classroom…’ (quoted in Diack, 1965:77). Rather than identifying this as a necessary phase the rush to fluency lead to the word method becoming dominant with Russell Webb (1846) and his ‘The New Word Method’ delighting in the absence of ‘unpleasant tones and drawling,’ (1846:3) more here. It was only much later that it was realised by the inspectorate that by the age of fourteen a child could have attended three thousand reading lessons but would have as a result neither ‘the power or the desire to read’ (Holmes, 1911, 128).

Frith’s (1985) three stage model begins with the identification of the logographic stage whereby children recognise words as symbols (and they may even be symbols, like frequently advertised cola brands). Similar to Marsh et al.’s (1981) linguistic guessing stage, in the logographic stage, words are seen as wholes with familiarity and shape bearing significance. This is followed by the alphabetic stage where grapheme-phoneme correspondences become crucial and letters and groups of letters are associated with sounds. Words are starting to be read by increasing understanding of letter-sound precepts similar to the developments suggested in Marsh et al.’s (1981) decoding stages. As familiarity and practice build confidence, a child enters the orthographic stage whereby word recognition becomes automatic and letter patterns and experience enable words to be read almost instantly.

Ehri (1998) renamed Frith’s (1985) orthographic stage the pre-alphabetic phase. It was here that visual cues and surrounding context enabled the recognition of words. The second of the four phases identified by Ehri (1998) was the partial alphabetic phase whereby letters, particularly initial and final letters in words assisted in the generation of pronunciation. In this phase, Ehri (1998) found that children were able to utilise associated letter patterns to identify words rather more effectively than distinctive and memorable letter patterns corresponding with Marsh et al.’s (1981) discrimination net guessing stage. As readers develop the ability to form alphabetic connections and acquire the skill of mapping graphemes to phonemes, they begin to be able to identify new words by blending generated pronunciations. They also become more accurate at identifying familiar words and more automatic at identifying words almost instantly without the need for phoneme sounding and blending. This full alphabetic phase is considered by Ehri (1998) to be the most important with the development of sight word reading (Ehri defines a sight word as one that can be recognised instantly). However, crucially, Ehri (1998) differs from Frith (1985) in maintaining the importance of grapheme-phoneme connections in this phase. Frith (1985) regards sight word reading as non-phonological. The final of the four stages identified by Ehri (1998) is the consolidated alphabetic stage which corresponds with Frith’s (1985) orthographic stage. It is in this phase that letter patterns become consolidated and unitised reducing memory load. The accumulation of letter pattern knowledge and words that can be identified automatically without reference to a synthetic phonic strategy gifts the reader the facility to identify new words with increasing ease. By referring to the hierarchical elements as phases rather than stages, Ehri (1998) implies that contiguous processing chapters have an ill-defined vagueness and ambiguity at their edges.

Ehri (1998) suggests that reading development is a progression to automaticity but does not explicitly state that there is progression from one phase to the next and does not make clear whether a child can be in two phases in parallel. In addition, she offers no indication as to what method of instruction is the most efficacious for each phase and what underlying cognitive structures relate to the phases of reading development. Ehri (1992) was critical of the dual route model of reading as she maintained the importance of grapheme-phoneme connections in the full alphabetic phase. She argued that the lexical route, with its reliance on a word’s orthography and semantic representation, ignored what she considered a systematic relationship between spellings and their pronunciations. Her championing of systematic synthetic phonics as a necessary instruction strategy will come as no surprise.

All these stage models describe a very similar pathway of reading development. Children initially recognise high-frequency words by rote association and contextual cues, and they are unable to do this when the words appear out of context. If an unfamiliar word appears in a familiar context, a story for instance, the child will respond with a word that is semantically plausible. In the next stage children attend to the letters in the words and their associated sounds and reading errors usually occur as a result graphemic complexities with responses often being semantically implausible but having a close graphemic representation ( as opposed to ).

The breakthrough in reading occurs when children realise that many letters are pronounced the same way in different words and that an unknown word can be identified by the employment of grapheme-to-phoneme correspondence rules. The crossing of this developmental Rubicon results in the understanding that the alphabetic principle operates through patterns of sound-to-letter correspondences. Initially, only decoding of the simpler elements of the alphabetic code, where one-to-one letter-to-sound mapping is evident, is possible.

The complex morphology of the English writing system (more here) means that there are numerous many-to-one mappings of letters and sounds which requires knowledge of multi-letter units in reading and spelling and as this knowledge is acquired children can read new words by analogy. Between the ages of eight and ten most children are able to recognise words automatically and they are ready to move to Chall’s (1996) stage of Reading to Learn.

Where all of the stage model theories concur, is that the process of adult reading is very different to that of a beginning and developing reader. Emergent and developing readers are not reading the same way that adults do, only slower and less expertly. They, therefore, cannot be taught to improve their reading in the same manner that adults improve their reading – by greater and greater exposure to more and more complex texts. They are in a phase that will ultimately enable them to develop that adult capacity but require tailored instruction appropriate for their stage of development.