Letter location versus letter order
In a previous section, I described a mechanism for encoding the location of different letter identities along a line of text. This first level of orthographic processing involves the location‐specific encoding of letter identities via a horizontally aligned (for written languages with horizontally aligned scripts) bank of letter detectors. These are the gaze‐centered letters in Figure 3.4.2 The second level of orthographic processing in this model is a location‐invariant sublexical orthographic code that provides information about the order of letters in a word independently of where readers’ eyes are looking at the word and prior to activating whole‐word orthographic representations. These are the word‐centered sublexical orthographic representations in Figure 3.4, which illustrates one specific scheme for encoding location‐invariant letter‐in‐word order. This particular scheme encodes letter order via a bag (i.e., an unordered set) of ordered contiguous and noncontiguous letter pairs (Whitney, 2001), referred to as open‐bigrams in Grainger and van Heuven’s (2004) model. It should be noted that this level of sublexical orthographic processing, the encoding of letter‐in‐word order, is the starting point of all other models of orthographic processing. A number of alternative accounts have been proposed for how this is achieved, such as the noisy spatial coding scheme (Davis, 2010) and the noisy length‐dependent ordinal encoding applied in the overlap model (Gomez et al., 2008). Positional noise is added to these coding schemes in order to account for transposed‐letter effects (see previous section). In Grainger and van Heuven’s model on the other hand, positional noise operates on the location‐specific letter detectors of the first level of processing.3 That is, transposed‐letter effects are a direct result of the location‐invariant order encoding scheme, and not just the result of positional noise.
Effects of the number and frequency of orthographically similar words
The impact of orthographic similarity with other words on the ability to read a given word is typically investigated by manipulating the number of orthographic neighbors a word has. The traditional definition of an orthographic neighbor is Coltheart’s N (Coltheart et al., 1977). This is a simple count of the number of words that can be formed from a given word by substituting a single letter with a different letter while respecting letter position (e.g., FARM: harm, firm, form, etc.). The classic finding observed with both lexical decision and word naming (speeded reading aloud) tasks is that words with more orthographic neighbors are easier to process (e.g., Andrews 1989; see Andrews, 1997 for a review of the evidence). This observation appears to be another challenge to the interactive‐activation model, given that the lateral inhibitory connections between word representations lead it to predict inhibitory effects of orthographic neighbors (Jacobs & Grainger, 1991). Given the clear evidence against a rigid position coding of letter identities, it follows that Coltheart’s N is not an ideal metric to capture orthographic overlap. More recent work has substituted N with OLD20, a measure of orthographic similarity that includes letter deletions, substitutions, and additions (Yarkoni et al., 2008). OLD20 counts the number of such changes required to transform a given word into another word (the Orthographic Levensthein Distance), retains the 20 most similar words (i.e., those requiring the fewest changes), and takes the average of the number of changes across these 20 words. The facilitatory effect of neighborhood size on lexical decision speed is also seen with this measure of orthographic similarity (e.g., Balota et al., 2007; Ferrand et al., 2018; Keuleers et al., 2012) and is expressed as a facilitatory effect of decreasing OLD20 (the smaller the value of OLD20 for a given word, the greater the similarity of this word to the 20 most similar words).
While facilitatory effects of orthographic similarity are typically found in lexical decision and naming tasks, other tasks often show a different pattern – most notably when isolating a unique word identity is requisite for the task, such as the perceptual identification task in which participants are asked to report the identity of a word that was presented to them under data‐limited conditions (i.e., with brief stimulus durations and/or masking: Carreiras et al., 1997; Grainger & Jacobs, 1996), or in a natural sentence reading context (Pollatsek et al., 1999). These findings led Grainger and Jacobs (1996) to propose their multiple read‐out model (MROM) according to which the type of information used to generate a response depends on the nature of the task. Most important is the idea that lexical decisions can be based on the global lexical activity generated by a stimulus, rather than having to uniquely identify that stimulus. Since increasing similarity with other words leads to higher levels of global lexical activity, this accounts for the facilitatory effects of orthographic neighborhood on lexical decision within an interactive‐activation framework. That is, according to the MROM, when responses in a given task can be generated on the basis of global lexical activity, then facilitatory effects of number of orthographic neighbors can be found, but when a word must be identified for the task at hand then inhibitory effects emerge, as predicted by the lateral inhibitory connectivity of the interactive‐activation model.
Another means to reveal inhibitory influences of orthographic relatedness is to use the masked priming paradigm (Forster & Davis, 1984), with orthographically similar words as prime stimuli. Several studies have reported inhibitory effects of word neighbor primes, particularly when the prime word is higher in frequency than the target word, such as the prime blue for the target blur (De Moor & Bysbaert, 2000; Davis & Lupker, 2006; Segui & Grainger, 1990). Moreover, this pattern typically reverses (becomes facilitatory) when the orthographically‐related primes are nonwords rather than words, such as the prime blun for the target blur (Forster & Davis, 1991, van Heuven et al., 2001). Furthermore, both neighborhood density and participants spelling ability appear to modulate these inhibitory effects, with the strongest effects arising in target words with many orthographic neighbors and participants with good spelling abilities (Andrews & Hersch, 2010; Meade et al., 2018). One possible explanation for these findings is that during the process of learning to read the orthographic representations of words are shaped by the number of orthographically similar words there are in a given reader’s lexicon, with the co‐existence of similar words forcing a more precise representation of which letters are where in a word (the “lexical tuning” hypothesis: Forster & Taft, 1994). Since accurate spelling requires exactly this kind of information, better spellers would on average have more precise orthographic representations. This greater precision would then increase the capacity of words to suppress other words that enter in the competition for identification (i.e., orthographically similar words).
Bridging the Gap with Sentence Reading
Orthographic processing provides the central interface between visual processing and higher‐level syntactic and semantic processing, with a key role played by whole‐word orthographic representations in this process. In this final section, I consider the crucial interface between whole‐word orthographic representations and the syntactic and semantic representations required for sentence comprehension. The central hypothesis here is that once a whole‐word orthographic representation is activated (not necessarily identified), that word identity (i.e., a specific whole‐word orthographic representation) is associated, not necessarily unambiguously, with a syntactic category (i.e., part‐of‐speech, Snell et al., 2017; see Declerck et al., 2020, for supporting evidence) and a meaning. Our understanding of this critical interface is limited, not the least because word reading and sentence reading tend to be investigated within two independent lines of research (see Liversedge et al., this volume). Important evidence concerning the word‐sentence interface has been obtained with three recently developed (or redeveloped) paradigms: a reading version of the classic Eriksen flankers task, the Rapid Parallel Visual Presentation (RPVP) of written words, and the grammatical decision task (the sentence‐level equivalent of the lexical decision