Generally, models of skilled word reading are constructed via a process of reverse engineering: (i) A body of findings concerning a relatively small set of phenomena is identified (e.g., effects of word frequency, orthographic-phonological regularity, letter transpositions); (ii) an internal organization is hypothesized in the form of a system of computational or neural mechanisms; and (iii) the model is evaluated in terms of whether the hypothesized organization would generate the patterns of behavior that it was designed to explain. This form of theorizing is not entirely circular: The models are also evaluated in terms of their capacity to generate accurate predictions about new facets of the phenomena of interest and, less often, their capacity to address other kinds of phenomena. The reverse-engineering approach is not specific to the study of word reading, but theorists in the domain of skilled word reading are especially adept practitioners of this approach; there are many word reading models, and as a group they are perhaps as detailed and mechanistically explicit as can be found in any subfield of cognitive science.
Frost's article reveals the bankruptcy of the reverse-engineering approach. At one level, his article is largely a criticism of the “new age of orthographic processing” (sect. 1.1, para. 2) – the proliferation of models inspired by the discovery that letter position is coded far less rigidly (in some languages!) than previous models would have led us to believe. Frost demonstrates that the effects of letter transpositions (and other manipulations) are quite different for Hebrew readers than for readers of English (and Spanish and French), and thus, that flexible position coding is not a necessary consequence of how our minds/brains are predisposed to represent strings of letters, but instead depends on the interaction of the reader and his or her linguistic environment.
At a broader level, Frost's article is not simply about how readers represent the orthographic properties of printed words; rather, it is an exploration of how cross-language differences (and commonalities) in word reading should be explained more generally. To the extent that reverse-engineering models can account for these differences, it is by stipulating language-specific differences in the organization of the reading system (in the simplest case, differences in parameterization; in the more extreme case, by positing different sets of underlying mechanisms). In this approach, the impact of the structure of the writing system on the organization of the reading system is more a matter of rationalization than explanation; that is, the model provides no explanation of how experience with a given writing system results in the reading system having a particular organization. Relatedly, although reverse-engineering models can serve to generate hypotheses about the relationship between the organization of skilled and beginning readers, or about the relationship between skilled and disordered reading, they provide little insight about the processes that transform a beginning reader to a skilled reader or how these processes differ in typically developing and reading-disabled individuals.
Given these considerations, Frost's endorsement of learning models over the reverse-engineering approach (“structured models,” in his terms) is precisely the right move. I would add to his analysis two key points: First, I believe the field has generally failed to appreciate that these two kinds of approaches represent different understandings of what counts as scientific explanation. For the reverse-engineering approach, the question is how to explain the behavior exhibited by readers in word recognition experiments, and the answer is the organization stipulated by the theorist, which describes the millisecond-scale processes by which the meaning and pronunciation of a printed word are computed. For learning models, the organization of the reading system plays a dual role. It describes the millisecond-scale processes by which a written word is read, and thus provides an explanation of the same kinds of phenomena addressed by reverse-engineering models. At a slower time scale, the organization itself changes as a consequence of learning, and the theory must explain how and why this happens. Thus, organization is both the explanation and the explanandum.
The second point I would add to Frost's analysis is that the acknowledgment that organization must itself be explained, and that learning is central to understanding this explanation, raises a new set of theoretical challenges. (1) We need to understand the nature of the learning process. For example, to what extent is reading acquisition a form of statistical learning? Are the mappings among orthography, phonology, and semantic learned independently, or does knowledge of one mapping constrain how the other mappings are learned? (2) How should the properties of a language or writing system be characterized? It has proven constructive to think that writing systems vary in their phonological transparency (Frost et al. Reference Frost, Katz and Bentin1987), the grain size of the mapping between orthography and phonology (Ziegler & Goswami Reference Ziegler and Goswami2005), and the richness of their morphology (Plaut & Gonnerman Reference Plaut and Gonnerman2000). But these characterizations are imprecise; we need much better ways to quantify these and other dimensions of statistical structure. (3) The properties of an orthography are determined by the properties of the language it represents. Frost hypothesizes in the target article that orthographies are optimized to provide “maximal phonological and semantic information by the use of minimal orthographic units” (sect. 3.1, para. 1, italics in the original). Similarly, Seidenberg (Reference Seidenberg, McCardle, Miller, Lee and Tzeng2011) proposed that languages with complex inflectional morphologies generally have phonologically transparent orthographies. Our theories should provide a basis for understanding how and why orthographic systems are constrained by the properties of spoken languages. (4) Knowledge is not an all-or-none thing. Stipulated models typically assume otherwise: For example, a lexical unit either exists or not. But an impressive array of evidence indicates that the quality of lexical representations (their precision and stability) can vary substantially, even for skilled readers (Perfetti Reference Perfetti2007). Our theories must provide the means to capture these “in-between” states. (5) The organization of the reading system differs for readers of different languages, but also among readers of the same language (Andrews & Hersch Reference Andrews and Hersch2010; Yap et al. Reference Yap, Balota, Sibley and Ratcliff2012). On what dimensions do these individual differences occur, and what gives rise to them?
Generally, models of skilled word reading are constructed via a process of reverse engineering: (i) A body of findings concerning a relatively small set of phenomena is identified (e.g., effects of word frequency, orthographic-phonological regularity, letter transpositions); (ii) an internal organization is hypothesized in the form of a system of computational or neural mechanisms; and (iii) the model is evaluated in terms of whether the hypothesized organization would generate the patterns of behavior that it was designed to explain. This form of theorizing is not entirely circular: The models are also evaluated in terms of their capacity to generate accurate predictions about new facets of the phenomena of interest and, less often, their capacity to address other kinds of phenomena. The reverse-engineering approach is not specific to the study of word reading, but theorists in the domain of skilled word reading are especially adept practitioners of this approach; there are many word reading models, and as a group they are perhaps as detailed and mechanistically explicit as can be found in any subfield of cognitive science.
Frost's article reveals the bankruptcy of the reverse-engineering approach. At one level, his article is largely a criticism of the “new age of orthographic processing” (sect. 1.1, para. 2) – the proliferation of models inspired by the discovery that letter position is coded far less rigidly (in some languages!) than previous models would have led us to believe. Frost demonstrates that the effects of letter transpositions (and other manipulations) are quite different for Hebrew readers than for readers of English (and Spanish and French), and thus, that flexible position coding is not a necessary consequence of how our minds/brains are predisposed to represent strings of letters, but instead depends on the interaction of the reader and his or her linguistic environment.
At a broader level, Frost's article is not simply about how readers represent the orthographic properties of printed words; rather, it is an exploration of how cross-language differences (and commonalities) in word reading should be explained more generally. To the extent that reverse-engineering models can account for these differences, it is by stipulating language-specific differences in the organization of the reading system (in the simplest case, differences in parameterization; in the more extreme case, by positing different sets of underlying mechanisms). In this approach, the impact of the structure of the writing system on the organization of the reading system is more a matter of rationalization than explanation; that is, the model provides no explanation of how experience with a given writing system results in the reading system having a particular organization. Relatedly, although reverse-engineering models can serve to generate hypotheses about the relationship between the organization of skilled and beginning readers, or about the relationship between skilled and disordered reading, they provide little insight about the processes that transform a beginning reader to a skilled reader or how these processes differ in typically developing and reading-disabled individuals.
Given these considerations, Frost's endorsement of learning models over the reverse-engineering approach (“structured models,” in his terms) is precisely the right move. I would add to his analysis two key points: First, I believe the field has generally failed to appreciate that these two kinds of approaches represent different understandings of what counts as scientific explanation. For the reverse-engineering approach, the question is how to explain the behavior exhibited by readers in word recognition experiments, and the answer is the organization stipulated by the theorist, which describes the millisecond-scale processes by which the meaning and pronunciation of a printed word are computed. For learning models, the organization of the reading system plays a dual role. It describes the millisecond-scale processes by which a written word is read, and thus provides an explanation of the same kinds of phenomena addressed by reverse-engineering models. At a slower time scale, the organization itself changes as a consequence of learning, and the theory must explain how and why this happens. Thus, organization is both the explanation and the explanandum.
The second point I would add to Frost's analysis is that the acknowledgment that organization must itself be explained, and that learning is central to understanding this explanation, raises a new set of theoretical challenges. (1) We need to understand the nature of the learning process. For example, to what extent is reading acquisition a form of statistical learning? Are the mappings among orthography, phonology, and semantic learned independently, or does knowledge of one mapping constrain how the other mappings are learned? (2) How should the properties of a language or writing system be characterized? It has proven constructive to think that writing systems vary in their phonological transparency (Frost et al. Reference Frost, Katz and Bentin1987), the grain size of the mapping between orthography and phonology (Ziegler & Goswami Reference Ziegler and Goswami2005), and the richness of their morphology (Plaut & Gonnerman Reference Plaut and Gonnerman2000). But these characterizations are imprecise; we need much better ways to quantify these and other dimensions of statistical structure. (3) The properties of an orthography are determined by the properties of the language it represents. Frost hypothesizes in the target article that orthographies are optimized to provide “maximal phonological and semantic information by the use of minimal orthographic units” (sect. 3.1, para. 1, italics in the original). Similarly, Seidenberg (Reference Seidenberg, McCardle, Miller, Lee and Tzeng2011) proposed that languages with complex inflectional morphologies generally have phonologically transparent orthographies. Our theories should provide a basis for understanding how and why orthographic systems are constrained by the properties of spoken languages. (4) Knowledge is not an all-or-none thing. Stipulated models typically assume otherwise: For example, a lexical unit either exists or not. But an impressive array of evidence indicates that the quality of lexical representations (their precision and stability) can vary substantially, even for skilled readers (Perfetti Reference Perfetti2007). Our theories must provide the means to capture these “in-between” states. (5) The organization of the reading system differs for readers of different languages, but also among readers of the same language (Andrews & Hersch Reference Andrews and Hersch2010; Yap et al. Reference Yap, Balota, Sibley and Ratcliff2012). On what dimensions do these individual differences occur, and what gives rise to them?