Orthographic Neighbourhood Size Effects in Chinese Character Recognition: Small, Inconsistent, and Theoretically Ambiguous

The Chinese writing system

At least three features of the Chinese writing system are relevant in connection to orthographic neighbourhood measures. First, Chinese has a large set of characters, around 8,000 (List of Commonly Used Standard Characters; Language and Text Information Management Department, 2013), among which 3,500 are frequently used. These characters are mostly words themselves and combine to form words of up to four characters. In contrast, alphabetic languages’ vocabularies are expressed in writing using a much smaller number of letters (e.g., 26 letters in English). A written word, in this case, will often resemble many other words closely, reflecting the fact that many words in natural languages have a high degree of phonological similarity. This makes the concept of orthographic neighbourhood evident and easily interpretable for alphabetic scripts. In Chinese, the concept is less straightforward: due to the vast character inventory, on average, the orthographic overlap of words is much lower.

Second, the Chinese script is morphosyllabic, meaning each character has its own pronunciation(s). As a result, there is no single unit in Chinese that maps onto the concept of letter. This implies that there is no single clear unit over which orthographic neighbourhood size should be calculated.

Third, Chinese characters are hierarchical. In its composition, a character can be decomposed into components at different levels of granularity: radicals, stroke patterns, or strokes.¹ See Table 1 for definitions and Figure 1 for an example character. At each level, components can be arranged in spatial relationships that are often complex. For example, at the radical level, components can be positioned such that one semi-encloses the other, such as 凵 and 㐅 of the character 凶 xiong1 ‘unfortunate’. At the stroke pattern level, the spatial arrangements are even more intricate, due to finer-grained components involved, such as 口, 力, and 刂, of the character 别bie2 ‘farewell’. At the most basic level, strokes within a single character can remain parallel (e.g., 二), connect (i.e., 丁), or intersect (e.g., 十). This spatial dynamics of strokes is clearly manifested in, for instance, character 于 yu2 ‘a common surname’. In contrast, for alphabetic languages, the spatial relation between letters or letter clusters is much simpler; it is linear. This makes the computation of string distances straightforward in alphabetic scripts but not in Chinese.

Figure 1

Previous studies on orthographic neighbourhood size in Chinese script have often sidestepped these complexities by treating characters as simple combinations of two components, typically at the radical level, such as radicals and residual components, or semantic and phonetic components. These studies often adopt the one-component substitution method (analogous to Coltheart’s N in the alphabetic writing systems) to define orthographic neighbours. For radical-substituting neighbours, as the non-substituted component can vary widely in visual complexity, neighbours tend to differ greatly in their similarity to the target character. For instance, 凝 ning2 ‘solidify’ and 冰 bing1 ‘ice’ share the radical 冫 but they are not orthographically similar overall, due to the varying visual complexity of the residual components. This issue can be partially resolved by defining neighbours using residual components: as they usually occupy more salient parts of characters, the corresponding neighbours tend to be more consistent in their similarity level to the target (e.g., 咤 zha4 ‘roar’ and 挓 zha1 ‘expand’ sharing the residual component 宅). However, these methods at the radical level share at least two limitations. They miscateogorise some similar characters as unrelated, such as those that differ below the radical level (e.g., 托 tuo1 ‘trust’ & 挓) or those that form part-whole relationships (e.g., 刘 liu2 ‘a common surname’ & 浏 liu2 ‘clear stream’). More importantly, as neighbours are defined by component inclusion alone, orthographic similarity is reduced to a binary judgment (similar/dissimilar), rather than a graded, continuous measure closer to human perception (Wang and Keuleers, 2024). Ignoring the hierarchy of characters thus results in an imprecise judgment of orthographic similarity and a theoretically compromised account of the effects of orthographic neighbourhood measures. This gap calls for a more appropriate conceptualisation of orthographic neighbourhood measures, which motivates the current paper. We will return to the empirical findings of the neighbourhood size effects in Chinese later; for now, we trace the neighbourhood size research back to its origins in alphabetic scripts.

Orthographic neighbourhood size effects in alphabetic scripts

Havens and Foote (1963) defined a word as having competitors if there are other words of the same length that differ in only one letter. In a recognition task under tachistoscopic conditions, they found that the presence of competitors inhibited word recognition after word frequency was controlled for. Landauer and Streeter (1973, Study 1) later investigated the relation between the number of competitors a word has and its frequency. Rather than using the term competitors, they referred to these words as neighbours and found that more frequent words tend to have more neighbours than less frequent words.

The number of neighbours is now commonly referred to as Coltheart’s N. In a lexical decision task, Coltheart and colleagues (1977) divided stimuli into a group with high-N and another with low-N. They found that N had inhibited nonword processing but had no effect on word processing. Andrews (1989) examined the interaction between word frequency and neighbourhood size in a series of lexical tasks using four-letter stimuli. They found a facilitatory effect in lexical decision and speeded naming, but only for low-frequency words. Subsequent studies (Andrews, 1997; Hendrix and Sun, 2021; Holcomb et al., 2002; Pollatsek et al., 1999; Sears et al., 1995; Ziegler and Perry, 1998) confirmed this, although Snodgrass and Mintzer (1993) found inconsistent effects across tasks.

The neighbourhood size effect has also been investigated for other alphabetic languages. In French, Peereman and Content (1995; Study 1) found that N facilitated both speeded and delayed naming, but only for low-frequency words and nonwords. Grainger et al. (1989; Study 1) found the evidence for the neighbourhood frequency effect: words with higher-frequency neighbours were responded to more slowly in lexical decision and progressive demasking paradigms (Grainger and Segui, 1990). Similar studies have also been done for Dutch (e.g., Frauenfelder et al., 1993), Spanish (e.g., Carreiras et al., 1997), German (e.g., Fiebach et al., 2007), Portuguese (e.g., Justi and Jaeger, 2017), and so on. For a more detailed review of N in alphabetic scripts, see for example Andrews, 1997 and Perea, 2015.

Coltheart’s N overlooked word similarity via letter deletions or additions. Yarkoni et al. (2008) then introducedOLD20, which measures the average Levenshtein distance to a word’s 20 nearest neighbours (Levenshtein et al., 1966). While these two measures were highly correlated (r = –.925) when measuring short, monosyllabic words, OLD20 was advantageous for longer words which typically have neighbours via deletions. In addition, recently, the one-letter distance approach has been used to build orthographic networks (e.g., Siew, 2018), which extended the orthographic analysis beyond local neighbourhoods to the scale of the whole lexicon. These networks are often constructed by representing each word as a node and connecting words that differ by a letter. In such networks, degree, a commonly reported node measure, counts immediate connections and therefore provides complementary evidence to N. Instead of representing words as strings of letters, Tulkens et al. (2018) represented them as vectors and introduced RD20—the average cosine distance to the 20 nearest vectors. They found that RD20, using one-hot encoded words (where each word is represented as a unique high-dimensional binary vector of a fixed length), slightly outperformed OLD20 in explaining variance in lexical decision response latencies across Dutch, English, and French.

Nonwords are usually created with the clear purpose of resembling real words. Keuleers and Brysbaert (2011) noted that such similarity might introduce bias into experiments, as was demonstrated by a nearest-neighbour model that computed the likelihood of recognising a stimulus based solely on its similarity to what was presented before it.

Orthographic neighbourhood size effects in Chinese script

Studies on two-character words have commonly used the one-character substitution method to determine orthographic neighbours, although they have varied with respect to character position. Huang et al. (2006) defined NS as the number of two-character words that share either the first or the second character with a target word. They found that NS facilitated recognition of high-frequency words but inhibited recognition of low-frequency words in a lexical decision task. In contrast, NS counted on words sharing the first characters has been consistently shown to facilitate lexical decision (Tsai et al., 2006; Xiong et al., 2024), naming (Li et al., 2015), and sentence reading (Tsai et al., 2006). Similarly, NS defined as sharing the second characters also facilitated lexical decision, particularly when neighbourhoods had high semantic consistency (Hsieh et al., 2024; referred to as ‘family size’).

Orthographic neighbours between single-character words are mainly assessed via approaches that rely on sub-character structural units, along with pixel-based approaches that measure visual similarity directly from character images (e.g., Sun et al., 2018). At the finest granularity, strokes are used, and two characters are considered orthographic neighbours if they differ by one or more strokes (e.g., Yu et al., 2023). Using this definition, Shen and Foster (1999) found that characters were named faster if they were primed by neighbours. However, in priming lexical decision tasks, both the presence of neighbours (Wang et al., 2014; Yu et al., 2023) and their average character frequency (Liu et al., 2023) inhibited recognition time. Using ERP data, characters with large neighbourhoods were found to have stronger N400 effects, which are typically associated with the processing of meaning (Dong et al., 2015; Wang and Zhang, 2011). Yet, there is currently no literature addressing the impact of stroke-based neighbourhood size on response latencies in psycholinguistic tasks.

At the radical–residual level, neighbourhood size can be defined in various ways targeting radicals or residual components. Radical frequency (or combinability) is one neighbourhood size measure. It refers to the number of characters that share a radical, in other words, differ by a residual component). Parallelly, residual component frequency (or combinability) refers to the number of characters that share a residual component, or, differ by a radical. The broader term subcomponent frequency refers to the size of neighbourhoods defined by either.

Using characters with horizontal structures whose radical and residual component are arranged side-by-side, e.g., 别 (⿰另刂), Taft and Zhu (1997, Study 1) found that the residual component frequency facilitated character recognition in lexical decision, whereas the effect of radical frequency was not significant. In contrast, Wang (2002) reported an inhibitory effect of radical frequency in a component detection task, where frequency by token was used. More recently, Wang et al. (2025) demonstrated a more complex picture, revealing that the effects of radical frequency and residual component frequency depended on character frequency. For frequent characters, both high radical frequency and high residual component frequency slowed down recognition, suggesting competition from orthographic neighbours. For less frequent characters, however, high residual component frequency sped up recognition as radical frequency increased, suggesting that neighbours provide supports. These interactions were observed for both position-general and position-specific subcomponent frequency measures. Collecting EEG data in a lexical decision task, Wu et al. (2012) found that characters with low position-general subcomponent frequency elicited a greater P200 effect, while characters with low position-specific subcomponent frequency elicited stronger P150, P200, and N400 effects. These findings suggest that sublexical processing was involved in both early lexical processing and later semantic processing, while the effect of position-specific subcomponent frequency lasted longer.

The role of radicals and residual components extends beyond their function as orthographic constituents. Approximately 80% of characters are known as phonograms (Yu and Reichle, 2017). A phonogram consists of a semantic component indicating its semantic category (usually a semantic radical) and a phonetic component indicating its pronunciation (usually a phonetic residual component). Feldman and Siok (1997) in a lexical decision task found that semantic radical NS facilitated character recognition, but in a later follow-up study, the authors (1999) found that this facilitation could be task-dependent and time-sensitive. For example, in a primed lexical decision task, the effect disappeared with long lags between characters and their orthographic neighbours sharing a semantic component (e.g., 冷 ‘cold’ and 冰 ‘ice’ both contain the semantic component 冫, which is related to ‘water’). In a naming task with ERP data collected, Hsu et al. (2021) found that, at the early stage of visual element discrimination (N170 time window), the effect of semantic radical NS was significant in reading low-consistency words. The authors explained that the activation of semantic radicals seemed to depend on phonetic components.²

The phonetic component NS is defined as the number of phonograms that share a phonetic component (e.g., 澋 and 憬 share the phonetic component 景 and both characters are pronounced as jing3). A phonetic component determines two phonological features of its character: regularity (whether the pronunciation of the phonetic component agrees with the character’s pronunciation) and consistency (to what extent the character’s pronunciation agrees with other characters that share its phonetic component). Li et al. (2020) conducted naming and lexical decision to investigate NS. To account for phonological effects, they selected phonograms that were either consistent-regular or inconsistent-irregular. Their results showed that a larger phonetic component NS inhibited the recognition of both types of phonograms, even after controlling for character frequency and number of strokes. The authors offered two explanations for this: first, the pronunciation derived from the phonetic component was often inaccurate, and second, the presence of more orthographically similar candidates with unreliable phonological values led to confusion at the orthographic level. These findings were in line with Li et al. (2011). For detailed reviews on phonetic component NS, see Hsieh et al. (2021).

Linearising character composition for neighbourhood measures

Recently, Wang and Keuleers (2024) proposed using Ideographic Description Sequences (IDSs) to compute character similarity. As a standardised system, IDSs encode the compositional structure of CJKV³ characters. In the original IDS database (Morioka and Wittern, 2002; kawabata, 2018), an IDS describes a character’s composition specifying its components and the spatial relationships between these components using 12 layout descriptors (⿰⿱⿲⿳⿴⿵⿶⿷⿸⿹⿺⿻).⁴ Each layout descriptor specifies the number of components, mostly two, except for ⿲ and ⿳ which take three, as well as their spatial arrangement. For example, 捌 is described as ⿰扌别, where ⿰ indicates a left-right arrangement of its two components, 扌 and 别. A challenge in using the original IDSs for similarity computation is that characters are decomposed to inconsistent levels of granularity. For example, 别 is described as ⿰另刂, meaning it is decomposed more deeply than 捌 (⿰扌别), making direct similarity comparison between the two challenging.

Wang and Keuleers (2024) addressed this by recursively replacing all components that could be further decomposed with their corresponding IDS representations in the database. As a result, each character obtained a ‘flattened’ representation, to which string-based methods can be applied (see Methods in Study 2). These representations consist of layout descriptors and stroke patterns (Figure 2). The number of stroke patterns in the dataset is 510, including 213 stroke patterns not yet assigned a code point, such as ‘&CDP-8CE6;’, and this size is comparable to the 514 stroke patterns identified in a Chinese national standard (National Language Commission, 2009). The final IDS representations of 20,830 characters used in the paper are provided in Data Availability.

Figure 2

Compared to the radical-based neighbourhood size measures, using IDSs to construct neighbourhood measures directly address their limitations. First, by decomposing characters to the stroke-pattern level, sub-radical similarity can be captured. Second, by applying string-based operations, i.e., substitutions, additions, and deletions, the part–whole relationships can also be captured. Third, by computing distance metrics over the linearised IDS representations, orthographic similarity between characters becomes a graded, continuous measure, more closely approximating human perception. It should be recognised that the stroke-based measures could even represent similarity at an even more nuanced level. Yet, no existing system encodes strokes and their spatial relationships. In addition, given that few characters have neighbours by a single stroke difference, such measures, even if existing, would be difficult to operationalise at the scale of a full lexicon. The IDS-based measures therefore constitute an optimal solution.

However, despite proposing the methods, Wang and Keuleers (2024) have not run statistical models to study how these measures influence character recognition. In what follows, we then present a series of studies in which we predict lexical decision response times (RTs) based on various neighbourhood size measures. First, we attempt to replicate previous findings measuring neighbourhood sizes using semantic and phonetic components. The effects of radical frequency and residual component frequency have been addressed in a separate study (see Wang et al., 2025). Second, we propose a set of distance-based neighbourhood size measures using characters encoded as IDSs.

Method

Results

Table 5 shows the correlations between response times and predictors. While response times exhibited a moderate correlation with frequency and NStrokes, their correlation with the other measures was limited. A moderate correlation between N-pc and consistency was found (r_s = –.579): consistency decreases as neighbourhood size increases. Consistency also had a moderate correlation with regularity (r_s = .516): consistency increases as regularity increases. Both consistency and regularity relate to the pronunciation of phonetic components. Regularity measures the agreement between a character and its phonetic component, while consistency measures the between-character agreement on shared phonetic components.

Table 6 shows the facilitatory effects of character frequency and the inhibitory effect of NStrokes. None of the other variables seemed to strongly affect reaction times. To assess collinearity among the phonological variables (Dormann et al., 2013), we computed the variance inflation factors (VIFs; O’brien, 2007) for the predictors in the LMM. Common cutoffs for serious collinearity include 10 and 5 (see, for example, Kutner et al., 2004; O’brien, 2007), while Graham (2003) shows that a VIF as low as 2 already suggests considerable collinearity. Although the VIF of consistency is close to the conservative cutoff of 2, we still accepted the possible mild variance inflation in favor of its theoretical importance and thus retained the variable in the model. To address concerns of collinearity, we applied PCA with varimax rotation, following the approach of Hendrix and Sun (2021). The results are provided in the Appendix.

Method

Design

In the representation of Chinese characters, an IDS starts with a layout descriptor followed by arguments, which can be either stroke patterns or another layout descriptor. For example, the IDS of 英 ying1 ‘blossom’ is represented as ⿱艹央 and that of 嚻 xiao1 ‘a mythic beast’ is ⿲⿱口口頁⿱口口 (examples from Wang and Keuleers, 2024).

Analogous to Coltheart’s N (Coltheart et al., 1977), we introduce N-Chinese (N-ch). N-ch is defined as the number of characters that differ from a target character by the substitution of one component (either a layout descriptor or a stroke pattern). For example, character 器 has a N-ch of 4, pointing to its four neighbours by one substitution (噐, 嚚, 嚣, and 囂).

Parallel to Yarkoni’s OLD20 (Yarkoni et al., 2008), we introduce WLD-k. WLD-k is defined as the mean of the weighted Levenshtein distance between a character and its k nearest neighbours. In the current implementation, we assign a cost of 2 to a substitution and a cost of 1 to an addition or deletion. This corrects an asymmetry in the calculation of raw distance that arises from the use of IDS representations. In an IDS representation, adding or deleting a stroke pattern affects both the component set and the layout of a character. For example, to add a stroke pattern 艹 to 央 (IDS: 央), we must first specify their spatial relationship with a layout descriptor. With ⿱, this produces 英 (IDS: ⿱艹央) and the distance between the two characters is 2. However, replacing the bottom component of 英 to form 苗 counts as a single substitution, leading to a distance of only 1. This unweighted scheme thus treats characters with a subpart–whole relationship (i.e., 英 vs. 央) less similar than characters differing by a substitution (i.e, 英 vs. 苗), although both character pairs differ by one stroke pattern perceptually. Weighting substitution at a cost of 2 while keeping addition and deletion at a cost of 1 resolves this issue, ensuring that swapping a component and adding/deleting a component result in equal distance. In the above example, both pairs end up having the same distance of 2 using the weighted scheme (see Figure 3 for an illustration of the calculation steps).

Figure 3

Levenshtein distance is an absolute measure of the number of operations required to transform one sequence into another. It is preferable to use a normalised version of edit distance, because absolute edit distance is not independent of sequence length (Wang and Keuleers, 2024). As Figure 4 shows, different pairs of characters with the same absolute distance can occur at widely differing normalised distances: the same absolute distance becomes a smaller normalised distance as the sequence length of the pairs increases. Table 7 illustrates this further, by comparing absolute and normalised distances for a few examples in English and Chinese.

Figure 4

Table 7

Comparison of absolute and normalised edit distance measures. The steps of WLD calculation for the English word examples are illustrated in Figure 5. Chinese characters: 鬰 yu4 ‘lush and growing abundantly’, 礬 fan2 ‘alum’, and 匕 bi3 ‘spoon’. The IDS of 鬰 is ⿳⿲木⿱㐅⿻丿乀 木冖⿰⿱⿶凵⿻ 㐅⿳丶⿰丶 丶丶⿺乚丿彡, of 礬 is ⿱⿱⿲ 木⿱㐅⿻丿乀木⿻一人⿸⿱一 丿口, and of 匕 is ⿺乚丿.

	WEIGHTED LEVENSHTEIN DISTANCE	WEIGHTED NORMALISED DISTANCE
D(hair, pen)	7	1.000
D(hair,hailstone)	7	.540
D(鬰, 匕)	23	.793
D(鬰, 礬)	24	.545

Figure 5

Following Wang and Keuleers (2024), normalised edit distance is calculated by dividing the absolute Levenshtein distance between two strings by the maximum possible distance between them. Given a weighted Levenshtein distance as used in WLD-k, the corresponding weighted maximum distance is:

$\text{max distance}=\min (N_i,N_j)\times 2+|N_i–N_j|$

where $N_i$ and $N_j$ are the lengths of IDSs of characters $C_i$ and $C_j$. The resulting weighted normalised distance has a range from 0 (characters are identical) to 1 (characters have no shared components in their IDSs).

Results

Table 9 shows the correlations between the predictors and response times. Both N-ch and WLD-10 were derived from the absolute Levenshtein distance and they were inversely related: larger N-ch reflects a higher degree of orthographic similarity to others through the larger size of immediate neighbourhood in the orthographic lexicon, while larger WLD-10 reflects lower orthographic similarity through longer distance to its ten nearest characters. Reflecting this, they were strongly negatively correlated and were therefore used for separate models. In comparison, although the normalised neighbourhood density measure WND was computed based on WLD, these two measures had a very low correlation, suggesting that they either capture different neighbourhoods altogether, or characterise similar neighbourhoods with different levels of nuance. Similarly, WND-10 had a negligible correlation with N-ch.

Like Study 1, the correlations between neighbourhood size measures (N-ch, WLD-10, and WND-10) and response times were weak. Of note, however, is that the correlation between NStrokes and WND-10 was smaller and had a different sign than the correlation between NStrokes and WLD-10. This suggests that WLD-10 partly reflects the number of strokes, with higher values being more likely for characters with more strokes. WND-10, on the other hand, should be insensitive to number of strokes. The smaller correlation between WND-10 and RT does not mean that WND-10 is a worse measure of neighbourhood density. Rather, it is a purer version of neighbourhood density, which incorporates aspects of neighbourhood density as well as character length.

Table 10 shows that character frequency had a facilitatory effect and NStrokes had an inhibitory effect and the pattern remain unchanged across models. All neighbourhood measures had reached significance.⁷ To check for collinearity, as in Study 1, we calculated the VIFs and the results imply that collinearity was not a large concern for all models. However, the results for marginal R² show that these neighbourhood measures explained only a minimal share of the total variance in the response times (at most 0.01%), despite being statistically significant.

A reviewer raised a concern regarding different control variables and character sets used across Studies 1 and 2, which may compromise direct comparisons between the radical-based and the IDS-based neighbourhood measures. To ensure differences in results cannot be attributed to differences in the data sets, we refitted three models of Study 2 (for N-ch, WLD-10, and WND-10, respectively) on the Study 1 dataset, with identical control variables (i.e., regularity, consistency, character frequency, and NStrokes). Despite the presence of additional phonological variables, the effects of IDS-based neighbourhood measures remained statistically significant, showing qualitatively similar results to what was reported above (See Appendix for full model results).

Discussion

In this study, we used Ideographic Description Sequences to represent characters. Unlike N-pc and N-sc in Study 1, where each character was represented by a combination of two subcomponents, characters using IDS representations encode fine-grained stroke patterns and their spatial relations. As a result, character similarity and orthographic neighbourhood can be more precisely measured.

Both N-ch and WLD-10 were computed based on absolute distance and showed inhibitory effects in the models. As WLD-10 averages distance, its larger values indicate that a character is further from other characters, which means it has greater orthographic distinctiveness. Its inhibitory effect therefore implies a facilitatory density effect: characters in denser neighbourhoods (lower WLD-10) are recognised faster. This may seem contradictory at first glance, given the inverse relation between the two measures, but they capture different aspects of orthographic similarity. N-ch tells us the size of the immediate neighbourhood. In this neighbourhood, neighbours are highly similar to a target. They differ by only one stroke pattern from the target, or they share the same set of stroke patterns but differ in structural arrangement. More neighbours slow down recognition, suggesting possible competition due to high similarity in visual forms. In contrast, WLD-10 tells us how dense a character’s region (constrained by a size of 10) is in the whole orthographic lexicon. Our finding suggests that characters more isolated in the hypothetical space are recognised more slowly, while characters embedded in a denser region are faster to process, possibly due to faster global access.

The role of WLD-10 is, in theory, particularly meaningful for complex characters, which tend to have relatively few one-substitution neighbours. Table 11 illustrates this distinction using two complex characters (Nstrokes of around 15) that have the same N-ch of 4 but differ in WLD-10. One character has a higher WLD-10 value of 3.3 (above Q3 = 3.2), indicating higher orthographic distinctiveness beyond its immediate neighbours. The other has a lower value of 2 (equal to the median), indicating that it is located in a denser cluster and is therefore less distinctive even beyond its immediate neighbours.

An alternative explanation relates to different types of neighbours included to compute these two measures. Levenshtein distance includes three operations: substitution, addition, and deletion. N-ch only counts the number of neighbours by substitution, ignoring neighbours formed by the other two operation types. By contrast, WLD-10 employs all three and thus reflects an additional whole-subpart relationship between neighbours (Table 11 shows an example with 憬 and 景). The different effects then possibly suggest that orthographic relationships may point to different aspects of orthographic processing, which requires further investigation.

We also introduced a normalised measure of neighbourhood density, WND-10. By using proportional distance to compute neighbourhood density, WND-10 avoids the problem that, for instance, a one-component difference between two-component characters equals a one-component difference between four-component characters. In doing so, it becomes less affected by the character complexity, as can be seen the correlations and VIFs from both models. Despite the conceptual difference, we found a similar result pattern as for WLD-10, with recognition times for characters increasing as a character’s neighbourhood becomes less dense. The substitution of WLD-10 by WND-10 also did not alter the effect of the other variables. These results do not indicate that both measures are interchangeable or equally valid. They do show that, even with a conceptually better measure, neighbourhood density effects remain weak. In addition, while the three measures in the present study capture a character’s level of orthographic distinctiveness in the lexicon, they are still local measures without a clear mechanism showing how the global organisation of characters shapes their influence. Conceptualising such a clear global organisation of characters based on orthographic relations, though currently outside the scope of this paper, constitutes our future research direction.