and Grainger (1994). Using the masked priming task in French, they replicated the phonological priming effect previously discussed: They showed that it is possible to prime the French word MERE (mother) with the homophonic pseudoword mert. In addition, however, they observed that the priming effect was larger for homophonic primes sharing many letters with the target word than for homophonic primes sharing few letters. This is possible in French, because both mert and mair are pronounced like MERE. The priming effect of mert on MERE was larger than the effect of mair on MERE. Furthermore, this orthographic priming effect was particularly strong at short prime durations, whereas the phonological priming effect needed longer prime durations to reach its maximum. These findings point to independent activation of orthographic and phonological information in visual word recognition (see also Adelman et al., 2014; Grainger et al., 2006; Kinoshita et al., 2018; Grainger, this volume).
A second problem for the strong phonological theory is that phonological effects in visual word recognition tend to be small (Brysbaert, 2003; Rastle & Brysbaert, 2006; Vasilev et al., 2019). Powerful experiments are needed to observe phonological priming, whereas orthographic priming is more robust.
A final problem for the strong phonological view is that transposing letters in written words does not affect word processing much, even if the transposition results in large changes to phonology (Perea & Carreiras, 2006; Perea et al., 2011). A striking example comes from Perea and Lupker (2003) who reported that in masked priming, the target word ALWAYS is primed almost as much by neevr as by never, despite the large difference in assembled phonology. This is difficult to explain in a model relying on phonology alone.
In summary, the empirical evidence points to the involvement of both orthography and assembled phonology in the processing of visual words, at least in alphabetic scripts. This view is called the weak phonological theory.
Computational Models of Visual Word Recognition
So far, we have discussed empirical findings indicating that phonology is involved in visual word recognition and reading. Psycholinguistic research is more than documenting empirical effects, however. Researchers want to integrate the findings in a coherent theory.
Computational models are the most informative to understand the mechanisms involved in word processing as implementation requires a detailed description of the operations involved (see Seidenberg et al., this volume). This is in contrast to verbal models which often resort to analogies with human intuition (Ward, 1998).
We limit our discussion to three computational models that include phonology. There are other models of visual word processing without assembled phonology, not discussed here (see Davis, 2010; Norris & Kinoshita, 2012; Whitney, 2001). So far, no computational model has been proposed for the strong phonological theory. The models we discuss are limited to English, and to the processing of monosyllabic words (some 6 000 in total).
The Dual Route Cascaded (DRC) model
The DRC model of visual word naming (Coltheart et al., 2001) was the culmination of decades of behavioral experiments with skilled readers and detailed analysis of reading in patients with various types of acquired dyslexia following brain damage (see Woollams et al., this volume). The model is summarized in Figure 4.2. It assumes two main routes for naming printed words (actually three, as we will see in the following text). The first route is based on addressed phonology; the second, on assembled phonology. Both routes start from an orthographic analysis stage in which the visual stimulus is translated into abstract letter identities. This means that the input is independent of the letter font used. Evidence for this stage comes from the observation that changes in font do not impact the magnitude of priming in masked priming: The target word television is processed equally fast after the primes television, TELEVISION, or TeLeViSiOn (Brysbaert et al., 2009; Forster, 1998; Rayner et al., 1980; see Grainger, this volume).
Figure 4.2 The DRC model of visual word naming
(Coltheart et al. 2001/With permission of American Psychological Association)
includes separate routes for addressed phonology (left part) and assembled phonology (right part).
The addressed phonology route within the DRC model assumes that the abstract letter identities activate entries in an orthographic input lexicon. A lexicon is a dictionary of known word forms. In many models of word processing it is separate from word meanings (called the semantic system) because the mappings between words and meanings are not one‐to‐one (e.g., synonyms refer to similar meanings although they have different forms; spoken and written words have different forms for the same meaning, and multilinguals know even more word forms referring to the same meaning).
The orthographic input lexicon in the DRC model contains word nodes that receive activation when the orthographic input includes overlapping letter identities at the right place. So, the orthographic input work activates all orthographic word nodes with the letter w at the beginning (want, went, why, work,…), all word nodes with the letter o in second position (soak, body, work,…), all word nodes with the letter r in third position (air, germs, work,…), and all nodes with the letter k in fourth position (bank, drake, work, …). There is competition between the activated word nodes until one wins. Usually this is the one corresponding to the input, as it receives the most activation.
Importantly, the word nodes in the orthographic input lexicon interact with corresponding word nodes in the phonological output lexicon (the mental dictionary of spoken word forms). As a result, the phonological node of the word is activated in parallel with the orthographic node. The lexicon‐mediated route of the DRC model makes it possible to name the words the model knows correctly, regardless of their orthographic transparency.
The authors of the DRC model argued for the existence of direct interactions between the orthographic input lexicon and the phonological output lexicon. This followed the observation that some patients with advanced dementia sometimes remain able to name words with irregular grapheme‐phoneme correspondences correctly, even though they no longer understand the words (e.g., Blazely et al., 2005). So, semantic involvement does not seem to be needed for addressed phonology to operate. At the same time, some patients with acquired deep dyslexia produce synonym errors in reading, saying tree to the target word bush (Coltheart et al., 1987). Because synonyms have no form overlap, the DRC maintained an extra route between the orthographic input lexicon and the phonological output lexicon via the semantic system. Word nodes in the orthographic lexicon activate meanings in the semantic system, which in turn activate word nodes in the phonological output lexicon. In the absence of direct interactions between the orthographic and phonological lexicons, this will result in synonym naming errors (but see Seidenberg et al., this volume, for an alternative account).
The second route of the DRC model implements assembled phonology. This is done by means of grapheme‐phoneme conversion rules. The abstract letter identities are translated one by one into phonemes. The phonemes are sent to a response buffer, which makes it possible for the model to name new words or pseudowords (e.g., briss, gert, tuise). Grapheme‐phoneme conversion is governed by a set of rules (e.g., b ‐> /b/, ea ‐> /i/). In case of multiletter graphemes, a wrong phoneme may be initially activated to some extent, which has to be corrected afterward (e.g., the letters ph will first activate /p/ before they activate /f/). This
