I agree with many of the points made by Frost, most notably, that (a) visual word identification should be studied in the context of language processing more generally, (b) there are important differences between languages, and (c) learning matters. Taken together, these factors suggest that visual word identification may differ across languages, and, indeed, Frost and colleagues appear to have identified an important difference (that is, letter position uncertainty varies across languages). I think Frost downplays the insights that have been made by studying a restricted set of languages (most often English), but I agree that there is value in looking for general principles of visual word identification that apply across languages.

However, I think Frost has missed a universal property of visual word identification that inspired some of the models he criticizes, namely, position-invariant letter identification. According to Frost, the recent “revolution” in modeling word identification was motivated by the many demonstrations of letter-transposition priming, and this inspired the development of models with letter position uncertainty. However, contrary to this account, the revolution started with models designed, in large part, to support position-invariant identification of letters. This enables readers to identify familiar words in novel contexts (e.g., the word play in treeplay). Position-independent letter identification was central to the development of the self-organising lexical acquisition and recognition (SOLAR) model of visual word identification (Davis Reference Davis1999) and the more recent spatial coding model (Davis Reference Davis2010), both of which were inspired by the neural networks Grossberg (e.g., Grossberg Reference Grossberg1978) designed to solve the so-called LTM invariance principle – networks developed long before the literature on transposed priming was established. More generally, context-independent identification of items is central to symbolic theories of object identification (Hummel & Biederman Reference Hummel and Biederman1992), semantic knowledge (e.g., Fodor & Pylyshyn Reference Fodor and Pylyshyn1988), and short-term memory (Bowers et al. Reference Bowers, Damian and Davis2009), amongst other domains. This is not to deny that transposed letter (TL) priming effects provide a rich data source for constraining models of visual word identification, but, contrary to Frost, this was not the only (nor primary) motivation for some of the current theories of letter position coding.Footnote ¹

The task of supporting position-invariant identification led to theories that are quite different than characterised by Frost, and a better appreciation of these models is critical for interpreting the reduced TL priming effects in Hebrew. I agree that the reduced TL priming suggests that position uncertainty is reduced in Hebrew compared to alphabetic scripts, plausibly for the reasons that Frost advances. But what the findings do not show is that letters are more rigidly bound to position in Hebrew compared to other (alphabetic) languages. Such a solution can account for the absence of TL priming in Hebrew, but at the cost of achieving position-independent letter identification. The challenge is to develop a (universal) model of letter coding that under some conditions includes very little positional uncertainty and, at the same time, achieves positional invariance (in Davis's terminology, solves the alignment problem).

In fact, the combination of universal positional invariance and language-specific TL priming is highly constraining for theories. On the one hand, the position-invariant identification rules out rigid as well as noisy slot coding (noisy slot coding includes some positional uncertainty; Gomez et al. Reference Gomez, Ratcliff and Perea2008). That is, any model in which letters are bound to position, such that the word play is coded as P-in-position-1, L-in-position-2, L-in-position-3, and Y-in-position-4, will not recognize play in treeplay (noisy slot coding will not help, as the noise cannot be so great that P-in-position-1 in play and P-in-position 5 in treeplay are similar). On the other hand, the lack of TL priming in Hebrew appears to challenge the open bigram account of letter position coding. Open bigrams support position-invariant identification by relying on the positional uncertainty of the bigrams (bigrams are not coded by position at all), and, accordingly, TL priming and position-invariant identification go hand-in-hand (which is not the case in Hebrew).

The spatial coding scheme of Davis (Reference Davis2010), however, can accommodate the range of TL priming effects across languages, and at the same time, support position-independent identification. A key feature of this model is that it does not rely on positional uncertainty in order to achieve position-invariant identification, and, accordingly, it is possible for the model to accommodate both sets of results. Indeed, Table 5 of Davis (Reference Davis2010) describes the results of the model when the position-uncertainty parameter is set to zero. Under these conditions, the model predicts no TL priming (like Hebrew; e.g., see simulations 21 and 22) and, at the same time, supports positional invariance as reflected in robust position-invariant priming (e.g., when the prime is composed of the last four letters of an eight-letter target; prime = stic, target = PLASTIC; simulation 60).

Still, the current spatial coding model does not predict that letter position uncertainty is reduced in Hebrew compared to alphabetic languages. Accordingly, findings of Frost and colleagues do suggest that additional constraints need to be added to the spatial coding model so that the positional uncertainty varies as a function of the orthographic density of words or morphemes in a lexicon (or perhaps other idiosyncratic aspect of the a given language). This is an important conclusion, but it is even more important not to lose sight of a key universal constraint of visual word identification, namely, letter position invariance. The parallel distributed processing (PDP) models of word identification that Frost points to do not solve the alignment problem (despite their learning), nor do more recent PDP models that attempt to directly address this issue (Bowers & Davis Reference Bowers and Davis2009).