Introduction
CAMEPA
, in which the letter R
from its base word (CAMERA
) was replaced with the visually similar letter P
, produces the same lexical decision times (in processing cycles) as the pseudoword CAMESA
, in which R
was replaced with the visually dissimilar letter S
(i.e., 118 cycles in both cases). Empirical evidence with misspelled common words corroborates this view. For instance, in lexical decision experiments, response times and error rates are similar for pseudowords like viotin
(where the letter “l
” from violin
was replaced with the visually similar letter “t
”) and viocin
(e.g., Perea & Panadero, 2014, Perea et al., 2022; see Gutierrez-Sigut et al., 2022, for electrophysiological evidence). If the visual appearance of letters played a role in how quickly and accurately people recognize words, we would have expected to observe a visual similarity effect. This effect would have resulted in more errors or slower “no” responses for visually similar pseudowords, such as viotin
compared to visually dissimilar pseudowords, like viocin
. Indeed, the lack of differences between the response times to pseudowords like viotin
versus viocin
has often been considered a key marker of abstract, orthographic processing (see Ziegler et al., 2013). tacebook
, where “f
” from facebook
was replaced with the visually similar letter “t
”) than to visually dissimilar misspelled logotypes like xacebook
(see Perea et al., 2022, for a replication with a different set of logotypes). This pattern was explained by a key feature of logotypes: their homogeneity in format. Unlike common words, which can vastly vary in appearance, logotypes are usually presented in a consistent font, style, and layout across different applications. After all, logotypes are designed to be easily recognizable and rely on visual information to achieve this recognition (Foroudi et al., 2017). Notably, visual letter similarity effects are also strong when presenting the misspelled logotypes in plain format (e.g., using Times New Roman; Perea et al., 2022), thus suggesting that the lexical representations of brand names may retain some perceptual elements (see Perea et al., 2021, for a comparison of transposed-letter effects in brand names vs. logotypes). Indeed, brand names are identified faster when presented in their typical letter case (e.g., IKEA
faster than ikea
; amazon
faster than AMAZON
; see Gontijo & Zhang, 2007; Perea et al., 2015). In contrast, this pattern does not occur for common words: PHARMACY
, although often capitalized, is no more readily identified than pharmacy
(Perea et al., 2018).Barcelona
). Furthermore, city names are encountered more frequently in printed materials, such as maps, travel guides, news, forums, or road signs, which tend to use a restricted set of fonts, colors, or layouts. For instance, the city names used in the present Experiments 1–2 had a considerably higher average frequency in the book/web database than in the subtitle database—where the latter is thought to reflect “everyday language” (M = 28.5 vs. 3.9 per million, respectively in the Duchon et al., 2013, EsPal databases in Spanish). In contrast, for the common words in Experiments 3–4, the frequency of appearance in books vs. informal contexts was comparable (book/web database: M = 54.7 and 57.01 per million in Experiments 3 and 4, respectively; subtitle database: M = 60.6 and 44.7 per million in Experiments 3 and 4, respectively; Duchon et al., 2013).Barcetona
from Barcelona
) or a visually dissimilar letter (e.g., visually dissimilar pseudoword, Barcesona
). We employed Simpson et al. (2013) letter visual similarity matrix to establish the degree of visual similarity of the misspelled letter with the original letter. The participants' task was to decide whether a given sequence of letters formed a correctly spelled word or not—they were also told that the set of words was composed of city names. The visual similarity effect was operationalized as the difference in response times and accuracy between visually similar pseudowords and visually dissimilar pseudowords. Experiment 1 used a standard setup in which each stimulus was displayed until the participant responded. In Experiment 2, we shortened the exposure duration to 200 ms to induce participants to make “word” decisions without a perfect match between the stimulus and the lexical entries. The logic of Experiment 2 was to maximize the chances of finding visual letter similarity effects for misspelled words (if any) in the absence of a careful post-access spelling check—note that visual letter similarity effects can occur in the first moments of processing (e.g., obiect
-OBJECT
producing faster responses than obaect
-OBJECT
in masked priming: Marcet & Perea, 2017, 2018; see also Lally & Rastle, 2022, for evidence with the Reicher–Wheeler task). Experiments 3 and 4 examined whether a limited viewing time could elicit a visual letter similarity effect with misspelled common words (e.g., votumen
vs. vosumen
; base word: volumen
[volume]).Barcetona
than for a visually dissimilar pseudoword like Barcesona
(i.e., a visual similarity effect). This result would support the idea that visual information may be preserved in the word recognition system for certain types of stimuli, thus posing problems for accounts of visual-word recognition that assume that lexical access is derived only from abstract representations. Alternatively, if there is nothing special about city name identification, we expect similar response times and accuracy for pseudowords like Barcetona
and Barcesona
. In this latter scenario, we would expect a null effect of visual letter similarity, as occurs with misspelled common words and as predicted by the leading models of visual-word recognition.Experiment 1 (misspelled city names, standard setup)
Method
Participants
Materials
Barcetona
from Barcelona
) and one in which we replaced the same internal consonant with a visually dissimilar consonant letter (e.g., Barcesona
). The average similarity of the original letter with the visually similar and visually dissimilar condition was 4.13 (range 2.77–5.33) and 1.31 (range 1.07–1.83), respectively, in the Simpson et al.'s (2013) visual letter similarity matrix. The average mean log bigram frequency in Spanish was similar for the two sets of misspelled items (2.11 vs. 2.06, respectively, p = 0.26; Davis & Perea, 2005). All pseudowords were pronounceable and orthographically legal. The base word was the only neighboring city name for the pseudowords. For the task, we also selected another set of 52 city names—of similar length as the initial dataset—that was presented correctly written. Their mean word-frequency in the Spanish book/web database (Duchon et al., 2013) was 27.86 (range 1.34–516.5) and their mean OLD20 was 2.40 (range 1.00–4.00). To ensure counterbalancing of the two types of pseudowords across participants, we created two stimulus lists using a Latin-square design. For instance, Barcetona
would be assigned to List 1 and Barcesona
to List 2. In each list, participants received 26 visually similar misspelled city names and 26 visually dissimilar misspelled city names—note that they were presented with 52 correctly written city names.Procedure
Results and discussion
City name | Visually similar misspelled city name | Visually dissimilar misspelled city name | Visual similarity effect |
---|---|---|---|
770 (9.9) | 794 (9.8) | 793 (3.8) | 1 (6.0) |
Response times
Accuracy
votume
≈ RT [accuracy] to vosume
; see Gutierrez-Sigut et al., 2022; Perea & Panadero, 2014; Perea et al., 2022). These results are consistent with the idea that lexical access for city names, like common words, is primarily driven by abstract representations (e.g., Dehaene et al., 2005; Grainger et al., 2008).Barcetona
than Barcesona
(i.e., an effect of visual similarity). Alternatively, if city names are represented similarly as common words (i.e., as abstract representations), we would expect a null effect of visual letter similarity for misspelled city names (i.e., as in Experiment 1).Experiment 2 (misspelled city names, brief presentation)
Method
Results and discussion
City name | Visually similar misspelled city name | Visually dissimilar misspelled city name | Visual similarity effect |
---|---|---|---|
738 (12.0) | 784 (18.4) | 771 (8.7) | 13 (9.7) |
Response times
Accuracy
Barcetona
than to Barcesona
(see the right panel of Fig. 2). We found an effect in the latency data in the same direction, but it was statistically weaker.votumen
vs. vosumen
; base word: volumen
[volume]). Gutierrez-Sigut et al. (2022) found that these misspelled words did not show any effects of visual similarity with the standard setup in neurotypical readers in behavioral or electrophysiological measures. Notably, for deaf readers, Gutierrez-Sigut et al. (2022) also reported that misspelled common words like vosumen
elicited more negativity in the N400 time window than votumen
. Thus, this set of misspelled words may produce visual letter similarity effects, at least for specific populations. (We discuss why some populations may be more sensitive to visual letter similarity effects in the General Discussion.)votumen
than for vosumen
). Alternatively, if the recognition of common words is primarily based on the rapid activation of abstract letter representations —which would likely occur in less than 200 ms—we would expect similar response times and error rates for visually similar and visually dissimilar misspelled common words (e.g., votumen
≈ vosumen
).Experiment 3 (misspelled common words, brief presentation)
Method
Participants
Materials
volumen
[volume]), 40 visually similar pseudowords created by changing one middle consonant letter from a base word by a visually similar letter (e.g., votumen
; the letter l
from volumen
was replaced with the letter t
; M = 3.28 in the Simpson et al.,’s, 2013, visual similarity matrix), 40 visually dissimilar pseudowords created by replacing the same internal letter as above with a visually dissimilar consonant letter (e.g., vosumen
; where the letter l
was replaced with the visually dissimilar letter s
; M = 1.52 in the Simpson et al., 2013, visual similarity matrix). None of the pseudowords had any word neighbors other than the base word, and the mean log bigram frequencies for the visually similar and visually dissimilar pseudowords were 2.30 and 2.35, respectively. The stimuli were counterbalanced across three lists—the base words were presented intact in the Gutierrez-Sigut et al. (2022) experiment to compare the ERP waves in all three conditions. As in the Gutierrez-Sigut et al. study, each list included 40 filler words to keep a 50% word/pseudoword ratio. We also created ten words and ten pseudowords for the practice phase.Procedure
Results and discussion
Word | Visually similar misspelled word | Visually dissimilar misspelled word | Visual similarity effect |
---|---|---|---|
637 (4.7) | 691 (5.0) | 699 (5.7) | 8 (− 0.7) |
votumen
[base word: volumen
]). The differences in visual letter similarity, as evidenced by Simpson et al.'s (2013) visual letter similarity matrix were large (M = 3.28 vs. 1.52, for visually similar and visually dissimilar pseudowords, respectively; t(119) = 13.0, p < 0.001). However, the difference between the visually similar and dissimilar conditions in Experiments 1 and 2 was more pronounced (average similarity scores were 4.13 and 1.31 respectively). circuilo
derived from circuito
[circuit]) consistently had high values in visual letter similarity (M = 4.17). In contrast, visually dissimilar pseudowords had low values in visual letter similarity (M = 1.18). In addition, we reduced the stimulus exposure duration from 200 to 150 ms in order to maximize the likelihood of detecting an effect of visual letter similarity for misspelled words, if present.Experiment 4 (misspelled common words, brief presentation of 150 ms)
Method
Participants
Materials
circuito
[circuit]), we created two pseudowords by replacing an internal consonant letter: (1) with a visually similar letter (visually similar pseudoword; e.g., circuilo
; replacing “t
” with “l
”); (2) with a visually dissimilar letter (visually dissimilar pseudoword; e.g., circuiso
; replacing “t
” with “s
”). The mean similarity of the original letter of the base letter and its visually similar and visually dissimilar counterparts was 4.17 (range 2.53–5.60) and 1.18 (range 1.07–1.40), respectively, in the Simpson et al. (2013) visual letter similarity matrix. None of the pseudowords had any word neighbors other than the base word, and the two sets of pseudowords were matched in mean log bigram frequency (M = 2.26 vs. 2.27 for the visually similar and the visually dissimilar pseudowords; p = 0.77). circuilo
would be in List 1 and circuiso
would be in List 2). Each list was composed of 53 visually similar pseudowords, 53 visually dissimilar pseudowords, and 106 words.Results and discussion
Word | Visually similar misspelled word | Visually dissimilar misspelled word | Visual similarity effect |
---|---|---|---|
609 (8.4) | 662 (13.2) | 676 (11.9) | − 14 (− 1.3) |
General discussion
amazom
[from amazon
] is often identified as a legit brand name; see Pathak et al., 2019; Perea et al., 2022) poses problems for those models of visual-word recognition proposing purely abstract codes. To explain this dissociation, one might argue that the processing of logotypes and the context in which they occur are very different from other categories of words because of the consistency in their format. In the present experiments, we investigated whether another type of stimuli—city names—is sensitive to visual similarity effects by comparing visually similar pseudowords like Barcetona
vs. visually similar pseudowords like Barcesona
in lexical decision. The logic was that, although to a lesser degree than brand names, city names are usually presented in a more homogeneous format than common words (e.g., initial capitalization, often in print). Using the standard setup (i.e., stimulus presentation until the participant's response), Experiment 1 found no evidence of visual similarity with misspelled city names (e.g., Barcetona
vs. Barcesona
). Crucially, when post-access verification mechanisms were restricted via a limited viewing time (i.e., stimulus presentation of 200 ms), we observed more errors for visually similar pseudowords (e.g., Barcetona
) than for visually dissimilar pseudowords (e.g., Barcesona
). Importantly, we also conducted two other experiments (Experiments 3 and 4) to test whether a limited viewing presentation would elicit a parallel effect with misspelled common words. Using different sets of items and varying stimulus presentation duration (200 ms in Experiment 3 and 150 ms in Experiment 4), we found no evidence of a visual letter similarity effect for misspelled common words.anazon
would activate amazon
more than atazon
). In addition, the representation of stimuli that keep some homogeneity in the visual format, like city names can be susceptible (to a lesser degree) to visual elements (e.g., Barcetona
would activate Barcelona
more than Barcesona
, at least with relatively brief exposure durations). Interestingly, this interpretation can easily explain why misspellings in common words in braille produce a tactile letter similarity effect (Baciero et al., 2022): braille letters have a characteristic homogeneous format (e.g., see UK Association for Accessible Formats, 2017), thus making them more sensitive to perceptual effects.forcet
would be more confusable with word forget
than the word forxet
(see Fig. 2 in Agrawal et al., 2020). Thus, this model can easily explain the presence of visual letter similarity effects for misspellings in lexical decision. The problem is that, for common words, these visual similarity effects only appear with heavily masked stimuli (e.g., masked primes: Marcet & Perea, 2017, 2018; Reicher-Wheeler task: Lally & Rastle, 2022). Furthermore, visual letter similarity effects are substantially greater for misspelled brand names than for misspelled city names for unprimed lexical decision. Future implementations of the models based on compositional codes need to consider the role of the variability across the visual input in their learning regime. We must keep in mind that Agrawal et al.’s model was trained with stimuli with the same font, thus creating a scenario similar to that of brand names and, thereby, especially susceptible to the influence of visual factors. More recently, Hannagan et al. (2021) implemented a deep convolutional neural network that was originally trained to perform object recognition and was later trained with thousands of images of words varying in case, font, and size. This model predicted a number of phenomena previously attributed to the emergence of abstract letter representations (see also Yin et al., 2022, for another convolutional network model that simulates masked form priming; see also Bowers et al., 2022, for discussion of these networks). However, unlike Agrawal et al.’s (2020) proposal, Hannagan et al. (2021) did not test visual similarity effects in lexical decision. While beyond the scope of the present paper, further simulations with these models are necessary to examine their plasticity to the various types of items (e.g., brand names, city names, common words) or groups of participants (e.g., readers with dyslexia).Barcetona
were more difficult to reject as words than Barcesona
under relatively brief exposure durations (200 ms). This pattern did not occur with misspelled common words in two additional experiments using different sets of items. This dissociation suggests that, at least for some types of words, visual codes are used during word processing flow, challenging biologically-inspired models of visual word recognition that rely solely on the activation of abstract letter codes (e.g., Dehaene et al.,’s 2005, Local Combination Detector model). Future implementations of models of visual-word recognition should consider that words that are often presented in a homogeneous format (e.g., brand names and, to a smaller degree, city names) can be more sensitive to visual codes than common words.