Frost presents a compelling argument for constraining models of visual word recognition by the linguistic properties of language, and attempting to find universals across languages. As noted in the target article, there is a large body of research focussing on linguistic elements that modulate the speed of word recognition (e.g., semantic, phonological, and morphological priming) and several models, such as the CDP++ (Perry et al. Reference Perry, Ziegler and Zorzi2010), and the bi-modal interactive activation model (Diependaele et al. Reference Diependaele, Ziegler and Grainger2010), accommodate these factors within a structural framework of orthographic word recognition. There are, however, several problems with these models, the most striking being that for the majority the first explicit stage of processing does not begin until after perceptual information has been encoded into an abstract form, that is, abstract letter representations. Hence, the crucial “visual” aspect of “visual word recognition” is absent from these models. Although linguistic factors are likely to shape the development and functioning of abstract letter units, visual processes must constrain these representations, given that, for sighted individuals, vision is the sensory input system for processing printed text. The fundamental role of early visual processing in reading research has been largely overlooked, but there is an emerging body of research showing the effects of visual regularities on recognising letters and words. For example, the overall shapes of words, not just the constituent letters, influence word recognition in children (e.g., Webb et al. Reference Webb, Beech, Mayall and Andrews2006), skilled readers (e.g., Healy & Cunningham Reference Healy and Cunningham1992), and dyslexics (e.g., Lavidor Reference Lavidor2011), by constraining the set of potential lexical candidates. This is most evident in lowercase scripts with ascenders and descenders that appear in the middle of words (e.g., Kelly et al. Reference Kelly, van Heuven, Pitchford and Ledgeway2011). We propose that these factors originate from general mechanisms associated with visual object perception, such as pattern recognition and episodic memory, which become specialised for processing text across development.
We agree with Frost that the cognitive system is tuned to pick up statistical regularities of the language. A series of studies have shown that letter and word recognition is mediated by positional letter frequency but only for some positions, such as the first letters (e.g., English and Greek) and last letter (shown in English) (Ktori & Pitchford Reference Ktori and Pitchford2009; Pitchford et al. Reference Pitchford, Ledgeway and Masterson2008). In contrast to Frost's assertion that, in English, all letters have a more or less similar contribution to word identity and the distribution of transitional probabilities is more or less flat, our data suggest that different letter positions may be more or less informative for constraining word recognition. Also, this varies across different alphabetic languages and within bilinguals, perhaps reflecting the statistical properties of letter distributions of each language (Pitchford et al. Reference Pitchford, Ktori and van Heuven2011). Interestingly, insensitivity to statistical regularities may be indicative of failure to become a skilled reader, such as in developmental dyslexia (Pitchford et al. Reference Pitchford, Ledgeway and Masterson2009). Sensitivity to statistical orthographic regularities could emerge implicitly within structural models of word recognition, for example, in interactive activation models (Dijkstra & van Heuven Reference Dijkstra, van Heuven, Grainger and Jacobs1998; McClelland & Rumelhart Reference McClelland and Rumelhart1981) that have feedforward and feedback connections between multiple levels of representation at the feature, letter, and word level. If top-down processing can modify activation at other levels, where relative letter position is encoded, this would provide a possible mechanism whereby orthographic regularities present in the lexicon could influence letter and word recognition.
Frost also argues that learning models have more to offer than the popular structural models of orthographic processing that have recently been proposed. However, the argument is not as straightforward as Frost presents (e.g., Page Reference Page2000; Thomas & van Heuven Reference Thomas, van Heuven, Kroll and de Groot2005). Learning models account for the acquisition of reading processes when a child is also developing spoken language skills. In contrast, structural models account for skilled reading processes that are optimised for recognising familiar words, rather than learning new words. Ideally, each type of model should inform the other regarding the key parameters needed to construct a cogent and “ecologically valid” theory of visual word recognition and reading. The end point of learning models should be structural representations, and structural models should emerge through learning (Johnson & Karmiloff-Smith Reference Johnson and Karmiloff-Smith1992).
One of the criticisms that Frost raises in relation to structural models of orthographic processing is that letter-order insensitivity is not a general property of the cognitive system. This is based on the lack of orthographic priming in non-alphabetic scripts, such as Hebrew (Frost et al. Reference Frost, Kugler, Deutsch and Forster2005). However, letter-order insensitivity in non-alphabetic scripts may be more pervasive than Frost suggests. For example, orthographic priming by transposed characters (Wang & Peng Reference Wang and Peng2000) and radicals (Ding et al. Reference Ding, Peng and Taft2004) has been shown in Chinese monolinguals. This demonstrates that insensitivity to order operates across different levels of representation even within a logographic writing system. This raises the possibility that analogous effects exist in Hebrew at some level of representation not yet reported. Additionally, the finding that English–Hebrew bilinguals only show form priming when tested in English but not in Hebrew (Frost et al. Reference Frost, Kugler, Deutsch and Forster2005) may not be as revealing about the encoding of letter position as first appears. The lack of form priming when reading Hebrew may have resulted from automatic activation of the English translations of printed Hebrew words, which do not necessarily share any orthographic features and/or interfere with the encoding of Hebrew. Indeed, it has recently been shown that when bilinguals read in one language they automatically translate into the other language, even with two distinctive scripts, such as English and Chinese (e.g., Zhang et al. Reference Zhang, van Heuven and Conklin2011). Furthermore, Chinese is not simply a logographic language as implied by Frost, as its phonology in Mandarin is represented using Hanyu Pinyin, a transparent alphabetic script, used in learning to read Chinese. This has been shown to impact on recognising printed English words in Chinese–English bilinguals (van Heuven & Conklin Reference van Heuven and Conklin2007). So, letter-order insensitivity may not be a strategy of optimising encoding resources as Frost suggests, but rather may reflect a fundamental property of the cognitive system, and therefore the brain.
Frost presents a compelling argument for constraining models of visual word recognition by the linguistic properties of language, and attempting to find universals across languages. As noted in the target article, there is a large body of research focussing on linguistic elements that modulate the speed of word recognition (e.g., semantic, phonological, and morphological priming) and several models, such as the CDP++ (Perry et al. Reference Perry, Ziegler and Zorzi2010), and the bi-modal interactive activation model (Diependaele et al. Reference Diependaele, Ziegler and Grainger2010), accommodate these factors within a structural framework of orthographic word recognition. There are, however, several problems with these models, the most striking being that for the majority the first explicit stage of processing does not begin until after perceptual information has been encoded into an abstract form, that is, abstract letter representations. Hence, the crucial “visual” aspect of “visual word recognition” is absent from these models. Although linguistic factors are likely to shape the development and functioning of abstract letter units, visual processes must constrain these representations, given that, for sighted individuals, vision is the sensory input system for processing printed text. The fundamental role of early visual processing in reading research has been largely overlooked, but there is an emerging body of research showing the effects of visual regularities on recognising letters and words. For example, the overall shapes of words, not just the constituent letters, influence word recognition in children (e.g., Webb et al. Reference Webb, Beech, Mayall and Andrews2006), skilled readers (e.g., Healy & Cunningham Reference Healy and Cunningham1992), and dyslexics (e.g., Lavidor Reference Lavidor2011), by constraining the set of potential lexical candidates. This is most evident in lowercase scripts with ascenders and descenders that appear in the middle of words (e.g., Kelly et al. Reference Kelly, van Heuven, Pitchford and Ledgeway2011). We propose that these factors originate from general mechanisms associated with visual object perception, such as pattern recognition and episodic memory, which become specialised for processing text across development.
We agree with Frost that the cognitive system is tuned to pick up statistical regularities of the language. A series of studies have shown that letter and word recognition is mediated by positional letter frequency but only for some positions, such as the first letters (e.g., English and Greek) and last letter (shown in English) (Ktori & Pitchford Reference Ktori and Pitchford2009; Pitchford et al. Reference Pitchford, Ledgeway and Masterson2008). In contrast to Frost's assertion that, in English, all letters have a more or less similar contribution to word identity and the distribution of transitional probabilities is more or less flat, our data suggest that different letter positions may be more or less informative for constraining word recognition. Also, this varies across different alphabetic languages and within bilinguals, perhaps reflecting the statistical properties of letter distributions of each language (Pitchford et al. Reference Pitchford, Ktori and van Heuven2011). Interestingly, insensitivity to statistical regularities may be indicative of failure to become a skilled reader, such as in developmental dyslexia (Pitchford et al. Reference Pitchford, Ledgeway and Masterson2009). Sensitivity to statistical orthographic regularities could emerge implicitly within structural models of word recognition, for example, in interactive activation models (Dijkstra & van Heuven Reference Dijkstra, van Heuven, Grainger and Jacobs1998; McClelland & Rumelhart Reference McClelland and Rumelhart1981) that have feedforward and feedback connections between multiple levels of representation at the feature, letter, and word level. If top-down processing can modify activation at other levels, where relative letter position is encoded, this would provide a possible mechanism whereby orthographic regularities present in the lexicon could influence letter and word recognition.
Frost also argues that learning models have more to offer than the popular structural models of orthographic processing that have recently been proposed. However, the argument is not as straightforward as Frost presents (e.g., Page Reference Page2000; Thomas & van Heuven Reference Thomas, van Heuven, Kroll and de Groot2005). Learning models account for the acquisition of reading processes when a child is also developing spoken language skills. In contrast, structural models account for skilled reading processes that are optimised for recognising familiar words, rather than learning new words. Ideally, each type of model should inform the other regarding the key parameters needed to construct a cogent and “ecologically valid” theory of visual word recognition and reading. The end point of learning models should be structural representations, and structural models should emerge through learning (Johnson & Karmiloff-Smith Reference Johnson and Karmiloff-Smith1992).
One of the criticisms that Frost raises in relation to structural models of orthographic processing is that letter-order insensitivity is not a general property of the cognitive system. This is based on the lack of orthographic priming in non-alphabetic scripts, such as Hebrew (Frost et al. Reference Frost, Kugler, Deutsch and Forster2005). However, letter-order insensitivity in non-alphabetic scripts may be more pervasive than Frost suggests. For example, orthographic priming by transposed characters (Wang & Peng Reference Wang and Peng2000) and radicals (Ding et al. Reference Ding, Peng and Taft2004) has been shown in Chinese monolinguals. This demonstrates that insensitivity to order operates across different levels of representation even within a logographic writing system. This raises the possibility that analogous effects exist in Hebrew at some level of representation not yet reported. Additionally, the finding that English–Hebrew bilinguals only show form priming when tested in English but not in Hebrew (Frost et al. Reference Frost, Kugler, Deutsch and Forster2005) may not be as revealing about the encoding of letter position as first appears. The lack of form priming when reading Hebrew may have resulted from automatic activation of the English translations of printed Hebrew words, which do not necessarily share any orthographic features and/or interfere with the encoding of Hebrew. Indeed, it has recently been shown that when bilinguals read in one language they automatically translate into the other language, even with two distinctive scripts, such as English and Chinese (e.g., Zhang et al. Reference Zhang, van Heuven and Conklin2011). Furthermore, Chinese is not simply a logographic language as implied by Frost, as its phonology in Mandarin is represented using Hanyu Pinyin, a transparent alphabetic script, used in learning to read Chinese. This has been shown to impact on recognising printed English words in Chinese–English bilinguals (van Heuven & Conklin Reference van Heuven and Conklin2007). So, letter-order insensitivity may not be a strategy of optimising encoding resources as Frost suggests, but rather may reflect a fundamental property of the cognitive system, and therefore the brain.