123 Writing systems vary: Frost, Katz, & Bentin (1987).
123 The myriad inconsistencies: The English Spelling Society wants to modify spelling (http://spellingsociety.org). For PG-17-rated spelling humor, watch “‘The Impotence of Proofreading,’ by Taylor Mali,” video uploaded to YouTube by Taylor Mali, August 14, 2008, https://goo.gl/PdZlYZ.
127 His first, landmark article: Van Orden (1987).
127 The subjects in these studies: The figure illustrates conditions in the series of experiments reported in Van Orden (1987) and Van Orden, Johnston, & Hale (1988). They weren’t all included in a single experiment, and some control conditions are omitted.
127 I’ve created a toy version: On seidenbergreading.net.
128 reliably make more errors: Van Orden, Johnston, & Hale (1988).
129 unreplicable “false positives”: On replications in psychology, see the several articles in Perspectives on Psychological Science,7, 2012.
129 Phonological effects: Pollatsek et al. (1992) and Lee et al. (1999) are representative studies.
129 the taxis delivered the tourists: Filik & Barber (2011); McCutchen & Perfetti (1982).
129 And so on: Ellis (1979); Jared, Levy, & Rayner (1999).
129 Van Orden effect: Goswami et al. (2001).
129 third- and fifth-grade children: Perfetti & Hogaboam (1975).
130 Is skill in reading aloud: Spurious correlation fun! “Spurious Correlations,” tylervigen.com, http://www.tylervigen.com/spurious-correlations.
130 The second Perfetti experiment: Perfetti & Roth (1981).
Examples of the materials used in Perfetti’s study of good and poor readers’ use of information concerning the predictability of words in context. Each sentence ended with the word dump.
130 Good readers were faster: Perfetti and his colleagues conducted many such studies in this period, including ones in which children either heard the context over headphones or read it themselves; the target words appeared at intermittent intervals in a story; and the words were preceded by unrelated words or read in isolation. They also looked at factors such as word length and frequency. The conclusions rest on this set of studies and others they inspired.
131 a good reader rapidly identifies: Keith Stanovich put these pieces together in 1980 (Stanovich 1980).
131 “I’ve got the eye of the tiger”: For “Roar,” see “Katy Perry—Roar (Official),” https://goo.gl/b7pwlG. Pointed out to me by Maryellen MacDonald.
131 Korean alphabet: On Korean spelling reform, see “Korean,” Omniglot, http://www.omniglot.com/writing/korean.htm.
131 The problem is: Joshi et al. (2008) is a fine, accessible discussion of English spelling.
133 Sign is an oddity: Chomsky & Halle (1968).
133 Skilled readers are aware: The classic work on this topic is J. Winter, “How I Met My Wife,” New Yorker, July 25, 1994. It begins, “It had been a rough day, so when I walked into the party I was feeling very chalant, despite my efforts to appear gruntled and consolate.” For cartoonist Mark Stivers’s witty illustrations of morphological inconsistencies, see http://goo.gl/qiIdeV.
135 Some of the most interesting work: Spencer and Hanley (2003); Hanley et al. (2004).
135 Researchers have concluded: Hoxhallari, van Daal, & Ellis (2004).
135 As comprehension is clearly the goal: Hanley et al. (2004).
136 Spelling-sound inconsistencies: Seidenberg (2011).
138 An alternative approach: Thomas & McClelland (2008); Rogers (2009).
138 My colleagues and I have used: For overviews, see Seidenberg (2005) and Plaut (2005).
139 The models I’ll describe: Deep learning models are solving hard problems such as speech recognition and scene analysis, and what matters is how well they work, not whether they work the same way as people. Our goal was to try to understand some basic properties of human behavior and their brain bases, and for that purpose the simpler models have proved helpful.
139 Basic properties of these neural networks: Flusberg & McClelland (2014).
Schematic of a simple feed forward network
139 Think of the card section: A card section creates a distributed representation of a word.
Each card is a binary “unit” that is either on or off (light or dark). Each letter is represented by a unique pattern of on and off units. Each card contributes to the representations of many letters. Words can be represented by using cards to represent successive letters.
144 Rule-governed but inconsistent words: The breakthrough study was Glushko (1979), which launched many others.
144 the system isn’t rule governed: Seidenberg (2012).
144 Language is said to be rule-governed: Our first model could correctly pronounce simple nononwords like nust but mispronounced hard ones like jinje. The errors arose from the fact that it didn’t adequately represent phonology. Once that was improved, so did generalization.
145 The idea that generalization: The application of these models to language was controversial: Seidenberg & Plaut (2014).
146 This developmental progression: Pugh et al. (2010).
146 It has long been thought: Baron & Strawson (1976) called them “Chinese” and “Phonenician” readers. At the time it was thought that Chinese characters could only be read visually, but we now know that Chinese readers make use of the phonological cues in the radical + phonetic characters.
147 At a given level of experience: The empirical jury is still out on how much skilled readers vary. Behavioral research has tended to focus on a narrow demographic (the “convenience sample” of mainly white, middle-class good readers) that does not represent the full range of backgrounds and experience. The models also focused on only a few of the factors that affect outcomes. Neuroanatomical variation relevant to skilled reading is only starting to be studied, but here is one provocative finding. In a recent study (Graves et al. 2014), we found individual differences among highly skilled readers in their use of the orthographyàsemanticsàphonology pathway in reading aloud. Reliance on this pathway was found to be correlated with neuroanatomical variation in the degree of connectivity (density of white matter tracts) between brain regions for semantics and phonology. Further studies of this sort should settle questions about how much skilled readers vary in division of labor. Then we will have to figure out why.
147 The most interesting evidence: Yang et al. (2009).
147 Such sequential models exist: Thomas & McClelland (2008).