Masked orthographic neighbor priming effects in Chinese two-character words

Abstract

In masked priming lexical decision tasks in alphabetic or syllabic script languages, latencies are longer when a word target is primed by a higher frequency neighbor (e.g., blue–BLUR) than when primed by an unrelated word of equivalent frequency (e.g., care–BLUR) – an “inhibitory neighbor priming effect.” In contrast, Zhou et al. (1999) demonstrated facilitatory orthographic neighbor priming for two-character Chinese words (e.g., 华丽–华贵). However, Zhou et al. did not control for relative prime-target frequency, which has been shown to be important in experiments when examining languages with other scripts. In the present Experiment 1 word neighbor primes (e.g., 容易-容貌) produced an inhibitory neighbor priming effect when the prime was higher frequency than the target, paralleling effects in other script languages. In further experiments, paralleling those in other script languages, two-character targets primed by nonword neighbors (容待-容貌) or single constituent characters matching the target in either position (容-容貌) showed significant facilitation. These results suggest that lexical activation/competition processes for two-character Chinese words are reasonably similar to those for words in alphabetic/syllabic script languages.

1. Introduction

Language researchers have used a broad range of experimental paradigms to investigate the mechanisms involved in visual word recognition. The masked priming paradigm (Forster & Davis, 1984) has been a common tool in this research. The masked priming presentation sequence begins with a mask (typically a row of hash marks), then a brief (<60 ms) display of the prime in lowercase, followed by the target in uppercase. Participants are typically unaware of the prime because of its brief presentation and the presence of the mask. As such, in contrast to unmasked priming, this approach is assumed to prevent participants from using processing strategies based on the existence of a relationship between primes and targets (Forster et al., 1987; Forster & Davis, 1984).

A variety of data suggests that the display of a word prime activates not just its orthographic representation but also representations of orthographically similar words, in particular, the orthographic neighbors of that word as well as words containing that word (e.g., Adelman et al., 2014; Peressotti & Grainger, 1999; Stinchcombe et al., 2012). Orthographic neighbors, according to Coltheart et al. (1977), are words that can be formed from a target word by substituting one letter for another while preserving letter positions and word length, for instance, “case,” “ease,” and “vast” are all orthographic neighbors of “vase” (see Davis & Taft, 2005, for a similar, but expanded, definition of orthographic neighbors).

The effect of presenting a word’s orthographic neighbor as a prime in a masked priming experiment has been widely investigated in alphabetic script languages such as French (Segui & Grainger, 1990), Dutch (Brysbaert et al., 2000), Spanish (Duñabeitia et al., 2009), and English (Andrews & Hersch, 2010; Davis & Lupker, 2006; Nakayama et al., 2008). When the prime is more frequent than the target (e.g., blue-BLUR), an inhibitory effect (in comparison to when an unrelated prime is presented, e.g., come-BLUR) is typically obtained. This result, which also emerged in Japanese for words written in syllabic Katakana (Nakayama et al., 2011) is explained by proposing that, although the brief presentation of the prime activates its orthographic neighbors, it most strongly activates the prime’s lexical representation. That representation then competes with and inhibits activation in the representations of orthographic neighbors, most relevantly, that of the target word. What is important to note, however, is that these inhibition effects typically only emerge (or only emerge strongly) when the prime is a high-frequency word and the target is a low-frequency word (e.g., Davis & Lupker, 2006). Essentially, in order for clear evidence of inhibition to be observed, the resting activation level of the prime needs to be higher than the resting activation level of the target.

Inhibitory orthographic neighbor priming effects, the core issue in the present Experiment 1, are, therefore, typically explained in terms of the involvement of mechanisms of lexical competition and inhibition with competition/inhibition mechanisms being incorporated into the majority of localist activation-based models of visual word identification, including the interactive-activation (IA) model (McClelland & Rumelhart, 1981) and the multiple read-out model (Grainger & Jacobs, 1996), as well as other more recent IA-type models (e.g., Davis, 2010). These models specifically assume that during the early stages of processing, the lexical representations of a presented word and those of orthographically similar words (e.g., the word’s “neighbors”) are simultaneously activated, and that, once activated, they compete with one another through a process involving mutual inhibition with the more activated representation successfully inhibiting its competitors.

Consistent with these ideas is the fact that when a nonword neighbor prime is used, the result is typically facilitation rather than inhibition (e.g., Andrews & Hersch, 2010; Davis & Lupker, 2006; Forster et al., 1987; Forster & Veres, 1998). That is, nonword neighbor primes would partially activate the target’s lexical representation (as well as representations of other orthographically similar words). However, because a nonword lacks a lexical representation there is no competition between its lexical representation and that of the word target. Note finally that there is typically a null priming effect when a word prime is a lower frequency neighbor of the target. Because the prime word would, presumably, have a lower resting activation level than the target, it would have less ability to compete with the target when the target is presented. Essentially, whatever activation is provided by the orthographic similarity of the prime and target is canceled by the lower ability of the prime’s representation to compete, leading to a latency that is often no different than that produced following an unrelated prime.

2. Orthographic neighbor priming effects in logographic scripts

In contrast to letters in alphabetic script languages, characters in logographic script languages are more than simple orthographic symbols. Characters are frequently morphemes as well (morphemes are defined as the smallest meaning-based units in a language). At this point, less is known about the nature of orthographic neighbor priming effects in logographic script languages despite the fact that neighbor priming effects (from both word and nonword primes) are well established in alphabetic script languages. There has, however, been some research on this issue in both Japanese and Chinese. In one example, Nakayama et al. (2014) investigated neighbor priming of two-character Kanji words, applying Coltheart et al.’s (1977) definition of orthographic neighbors and treating individual Kanji characters as the basic orthographic units, a method that has often been used in previous Japanese (e.g., Hino et al., 2011; Kawakami, 2002) and Chinese (e.g., Huang et al., 2006; Tsai et al., 2006) studies. (In Japanese, two-character words make up around 80% of all Kanji words (e.g., Hino et al., 2011; Hino & Lupker, 1998).) Specifically, Kanji neighbors share a character in the same position. For example, the Kanji words 企業 and 仕業 are considered orthographic neighbors that differ at the first character position, and the Kanji words 会議 and 会話 are considered orthographic neighbors that differ at the second character position. This definition of orthographic neighbors for Kanji two-character words was also used in our investigation of orthographic neighbor priming in Chinese (a language with a logographic script quite similar to that of Kanji).

Using two-character Japanese Kanji words, Nakayama et al. (2014) obtained inhibitory neighbor priming effects. Specifically, inhibitory neighbor priming effects were obtained in four studies for low-frequency targets primed by higher-frequency Kanji word neighbors (情報-情緒, Experiments 1A, 1B, 3A, and 3B). Targets primed by Kanji nonword neighbors, in contrast, showed a noticeable facilitation effect (情門-情緒, Experiments 2 and 3). Further, when targets were primed by individual constituent Kanji characters (情-情緒, Experiment 4), a significant facilitation effect was also found. These findings imply that, as with words in alphabetic script languages, lexical competition influences the identification of Kanji words. The results of Nakayama et al.’s Experiment 4, showing that subset primes also produce facilitation, are also consistent with the results in alphabetic script languages (e.g., Peressotti & Grainger, 1999; Stinchcombe et al., 2012). However, because the primes and targets in their Experiment 4 also shared a morphological unit, it is possible that the priming observed in Experiment 4 was morphologically based, an interpretation that those authors preferred.

3. Morphological considerations

As noted, Kanji (or Chinese) neighbors are not only orthographically related in that they share a character, but they are also typically morphologically related since each character is a morpheme, even though Kanji or Chinese neighbors are not always semantically related due to the multiple meanings of each morpheme (see Gu et al., 2015; Yang et al., 2022). This characteristic of Kanji/Chinese neighbors distinguishes them from most orthographic neighbors in alphabetic languages, such as blue-blur. Therefore, masked neighbor priming in Kanji/Chinese involves a possible facilitation element that is not normally present when priming neighbors in alphabetic script languages (Nakayama et al., 2014).

It is now clear that morphologically-based priming effects do appear in masked priming experiments in alphabetic script languages (e.g., Crepaldi et al., 2010; Heathcote et al., 2018; Rastle et al., 2004). Most relevant to the present research, morphological facilitation effects have been found for compound word prime-target pairs. For example, Crepaldi et al. (2013) found that transposed morpheme primes (e.g., moonhoney) facilitated the processing of multi-morphemic targets (HONEYMOON), whereas monomorphic primes transposed in a similar way (e.g., ticeprac) did not facilitate target processing (PRACTICE). Therefore, it is possible that Nakayama et al.’s (2014) neighbor priming effect in their Experiment 4 as well as any neighbor priming effect in Chinese may be at least somewhat morphologically driven.

In fact, some models of visual word recognition in alphabetic script languages are based on the idea that morphological representations play an important role in word recognition, hence, predicting priming from morphologically related primes. For example, Taft and Nguyen-Hoan (2010) proposed that morphemes are represented at both the form and lemma level, with the lemma level being an abstract level coding consistent form–meaning linkages. Higher-level representations are activated as a result of activating the lemma-level units. Similarly, in the hybrid dual-route model, Diependaele et al. (2009) suggested that morpho-orthographic decomposition exists with the activation of morphological units driving visual word recognition although holistic processing occurs at the same time.

Although these models differ somewhat in how morphological representations are assumed to be activated, they both suggest that morphological representations can be activated automatically early in word recognition in alphabetic script languages, producing morphological priming effects. As the logographic script used in Chinese appears to be even more morphologically based than alphabetic scripts, one would imagine that morphological priming effects would be even more obvious in Chinese language experiments. Indeed, there are experiments that do seem to document facilitation from morphologically related primes (e.g., Tsang & Chen, 2014; Wu et al., 2020; Zhou et al., 1999). However, the issue is not clear-cut. For example, Yang et al. (2020) reported that facilitatory morphological priming effects were minimal for four-character Chinese words. Nonetheless, the potential impact of morphological priming in Chinese cannot simply be dismissed and, therefore, needs to be considered when evaluating the impact of orthographic neighbor primes.

4. Orthographic neighbor priming effects in Chinese

The general question we addressed concerned the nature of orthographic neighbor priming effects for two-character Chinese words and what impact relative prime-target frequency has on those effects. At present, there does not appear to have been a direct examination of this issue with what appears to be the most extensive examination of Chinese neighbor priming being provided by Zhou et al. (1999) using a definition of orthographic neighbors similar to that used in Nakayama et al.’s (2014) Kanji experiments. That is, Zhou et al. examined the priming effect between word primes and targets that share one character in the same position (e.g., 华丽-华贵). They found that, with a 57-ms prime duration, there was a significant facilitation effect across four experiments, with targets processed faster when primed by a neighbor, an effect that was independent of shared character position. Further, that facilitation occurred both when the shared character had the same meaning in the prime and target (their MORPH condition: 华丽-华贵, where the character 华 means “splendid”) and when the shared character had different meanings in the prime and target (their CHAR condition: 华侨-华贵, where the character 华 means “Chinese” and “splendid”, respectively) with the MORPH condition consistently producing a larger facilitation effect than the CHAR condition. Based on these results, Zhou et al. concluded that their facilitation effect was at least partially due to morphological overlap, paralleling the morphological priming effects in alphabetic languages (Duñabeitia et al., 2009; Rastle et al., 2004).

Zhou et al.’s (1999) experiments, however, were not ideal for testing for lexical competition (i.e., inhibition) effects between orthographic neighbors because they did not attempt to control relative prime-target frequency. Most of their stimulus pairs involved low-frequency primes (e.g., mean frequency = 22 occurrences per million in their Experiment 1) and higher-frequency targets (mean frequency = 47 occurrences per million in their Experiment 1). According to the literature in other languages, lower-frequency neighbor primes typically do not create much lexical competition for their higher-frequency targets (Andrews & Hersch, 2010; Davis & Lupker, 2006; Nakayama et al., 2008). Therefore, the prime-target frequency relationship would appear to be a factor that must be controlled when investigating the lexical competition process.

One additional issue that is problematic for present purposes is that Zhou et al. (1999) did not examine the effect of the lexicality of the neighbor primes. Their primes were always words. As previously mentioned, nonword neighbor primes in alphabetic languages do not produce inhibition effects (e.g., Davis & Lupker, 2006; Forster & Veres, 1998), presumably because nonword primes do not strongly activate any specific neighbor of the target. Everything considered, Zhou et al.’s data would not provide a full view of the nature of orthographic neighbor priming effects in Chinese. Essentially, when examining the nature of orthographically-driven lexical processing in Chinese in order to contrast it with that in other script languages, relative prime-target frequency needs to be manipulated and the distinction between priming from word neighbor primes and priming from nonword neighbor primes needs to be examined.

Two other issues concerning Zhou et al.’s (1999) experiments should also be mentioned. First, 50% of their prime-target pairs were designed to have a strong whole-word semantic relationship as the shared character is associated with the same morpheme in the prime and target (e.g., 华丽 and 华贵 meaning “magnificent” and “luxurious”, respectively, their MORPH condition). That fact makes that condition potentially problematic for present purposes because Zhou and Marslen-Wilson (2000) have shown semantic priming effects with a short SOA (57 ms) for Chinese compound words. Any semantic priming effects would have masked any (inhibitory) effects of lexical competition. In the other 50% of the primes and targets (their CHAR condition), the shared character did not represent the same morpheme. However, it is not entirely clear that all of the pairs were not semantically related. Second, the number of observations in the relevant experiments was considerably smaller than what is currently deemed necessary (see Brysbaert & Stevens, 2018) with there being 6–10 stimuli per condition and 30–50 participants per experiment.

5. The present research

The goal of the present research was to provide a more comprehensive investigation of orthographic neighbor priming effects for two-character Chinese words by examining potential interactions with prime-target frequency and prime lexicality. In Experiment 1, we used the masked priming lexical decision task and tested whether inhibitory neighbor priming effects obtain for two-character Chinese words when prime-target frequency is varied systematically. Half of the pairings were high-frequency neighbor primes with low-frequency targets, and the other half were low-frequency neighbor primes with high-frequency targets. The expectation is that any inhibition effect should be more evident with high-frequency primes than with low-frequency primes.

In Experiment 2, the influence of prime lexicality on neighbor priming effects was examined in order to determine whether nonword neighbor primes would produce facilitation using those same targets. If so, that result could provide additional evidence that any inhibitory neighbor priming effect obtained in Experiment 1 was due to lexical competition, supporting the idea that the nature of lexical activation in Chinese is similar to that in other script languages.

Experiment 3 tested one more aspect of orthographic priming, whether there is a facilitation for Chinese two-character words using the first-position character or second-position character of the target words as the prime. These primes should produce evidence of facilitation, as is typically reported in alphabetic language experiments (e.g., Peressotti & Grainger, 1999; Stinchcombe et al., 2012).

6. Experiment 1

6.1. Method

6.1.1. Participants

We conducted power calculations based on the significant interaction obtained in a similar priming experiment (partial η ² = 0.248; Segui & Grainger, 1990, Experiment 2). Using G*Power 3.1 software, it was determined that a power of 0.80 would be achieved for a similar-size interaction by using 27 participants. We also noted the sample sizes used in Nakayama et al.’s (2014) Experiment 1 and Zhou et al.’s (1999) Experiment 1, which were 44 and 40, respectively. Ultimately, 46 undergraduate students from Zhejiang Gongshang University participated in this experiment. Each received a small gift for their participation. All were native speakers of Chinese and had normal or corrected-to-normal vision with no reading disorder.

6.1.2. Materials

All the stimuli were two-character Chinese words. One hundred twenty pairs of orthographic neighbors were selected to create the word target pairs. None of the neighbor pairs were semantically related, as judged by the first author (a native Chinese speaker). For each pair, each neighbor could serve as either a prime or a target depending on the frequency condition that the pair was assigned to. One member of the neighbor pair was a high-frequency word and the other was a low-frequency word, according to the SUBTLWF frequency counts in the SUBTLEX-CH database (Cai & Brysbaert, 2010). The frequency for high-frequency members (occurrences per million: M = 151.85, SD = 2.01) was significantly larger than that for low-frequency members (occurrences per million: M = 2.01, SD = 2.18), t = 8.97, p < 0.001. The orthographic neighborhood size for high-frequency members (M = 56.99, SD = 41.06) was similar to that for low-frequency members (M = 59.06, SD = 51.29), t = −0.47, p = 0.641. The orthographic neighborhood size for a two-character Chinese word is the number of other two-character Chinese words that share one of its two characters in the SUBTLEX-CH database in the same position. Sixty neighbor pairs shared a character in the first position, and the other 60 neighbor pairs shared a character in the second position.

The words forming the 120 neighbor pairs were used twice for each participant, once as a target and once as a (related or unrelated) prime. When the high-frequency member of the pair was used as a target, its low-frequency neighbor was used as its related prime (e.g., 容貌 (Appearance) primed 容易 (Easy)) and a low-frequency word from another pair was used as its unrelated prime (e.g., 参差 (Uneven) primed 容易(Easy)). Similarly, when the low-frequency member of the pair was used as a target, its high-frequency neighbor was used as the related prime (e.g., 容易 (Easy) primed 容貌 (Appearance)) and a high-frequency neighbor from another pair was used as the unrelated prime (e.g., 号码 (Number) primed 容貌 (Appearance)). In the experiment, each participant responded to each target word only once, with one member of the pair presented following a related prime and the other member presented following an unrelated prime. Assignment of targets to conditions was counterbalanced across participants. Note that, in this design, all the primes in the related and unrelated conditions were the same across participants. The unrelated primes shared no characters with their targets.

Four hundred and eighty-two-character Chinese nonwords were selected to serve as the nonword primes (240) and targets (240). These nonword stimuli were derived from the nonwords found in the Chinese Lexicon Project (Tse et al., 2017). The manipulation of prime type for the nonword targets was done similarly to that for word targets. However, there was no counterbalancing of lists for nonword targets (only one list of 120 neighbor nonword primes, 120 unrelated primes, and 240 targets were used).

As the assignment of word targets to relatedness conditions was counterbalanced across participants, we created two lists of size 480 (with 120 targets in each condition in each list and with each target being preceded by its neighbor prime in one list and an unrelated prime in the other list and 240 nonword targets). All primes were presented in 25-pt Dengxian light font typeface, whereas the targets were presented in 30-pt SimSun font typeface.

6.1.3. Procedure

The data were collected using Eprime 2.0 software (Psychology Software Tools, Pittsburgh, PA; see Schneider et al. (2002)). The stimulus color was black and the background color was white. We used a standard 60 Hz refresh rate in a 1,068 × 768-pixel CRT screen.

Each trial began with a pattern mask (####) for 500 ms, followed by a 50 ms prime, and then the target appeared for 3,000 ms or until the participant responded. The stimuli were all displayed in the middle of the screen. Participants were asked to decide if each character string was a real Chinese word or not and to press the “J” button if it was a word, and the “F” button if it wasn’t. They were instructed to respond as rapidly and accurately as possible. Each participant’s trial order was randomized. Sixteen practice trials not involving any stimuli from the experiment proper preceded presentation of the experimental stimuli. Participants had a short break after completing every 160 trials. ¹

6.2. Results

For word targets, latencies less than 250 ms or greater than 1,800 ms (2.2% of the data) and incorrect responses (9.7% of the data) were excluded from the latency analyses. The mean RTs and percentage error rates, as a function of Relatedness (neighbor vs. unrelated) and Frequency Type (high-frequency vs. low-frequency targets) based on subject means for the word targets are shown in Table 1. The data from nonword targets were not analyzed due to the fact that the nonword targets were not counterbalanced across prime types. Generalized Linear Mixed-effects Models (GLMM) from the lme4 package (version 1.1.23) in R software (version 3.6.1) were used to analyze the latency and error rate data (Bates et al., 2015; Lo & Andrews, 2015; R Core Team, 2015).

Table 1. Mean lexical decision latencies (RTs, in milliseconds) and percentage error rate for words in experiment 1

Abbreviation. RT = reaction time; %E = percentage error rate.

In the GLMM analysis, Relatedness, Frequency Type, and their interaction were entered as fixed factors. Subjects and items were entered as random factors. For the latency analysis of word targets, the model was: RT = glmer (RT ~ Relatedness * Frequency Type + (Frequency Type |subject) + (1 |item), family = Gamma(link = “identity”), control = glmerControl (optimizer = “bobyqa”)). For the error rate analysis of word targets, the model was: Accuracy = glmer (accuracy ~ Relatedness * Frequency Type + (Frequency Type |subject) + (1 |item), family = “binomial”, control = glmerControl(optimizer = “bobyqa”)). More complex models that included all relevant random structures were used in our initial analyses. Ultimately, we had to use the models noted above due to convergence failures with the more complex random slope models (Barr et al., 2013). Before running the model, R-default treatment contrasts were altered to sum-to-zero contrasts (Levy, 2014; Singmann & Kellen, 2019). Follow-up analyses were conducted using the emmeans package, version 1.5.0 (Russell, 2020). The raw data used for the analyses and word stimuli used in all different experiments are publicly available at https://osf.io/8cdnm/.

6.2.1. Latency and error rate analyses

The main effects of Relatedness, ß = 2.48, SE = 1.20, z = 2.07, p = .038, and Frequency Type were significant, ß = 62.57, SE = 3.31, z = 18.93, p < 0.001, indicating that, overall, the related condition latencies were longer than those in the unrelated condition (i.e., an inhibitory priming effect) and low-frequency target latencies were longer than the latencies for high-frequency targets. Importantly, there was an interaction between Relatedness and Frequency Type, ß = 5.01, SE = 1.21, z = 4.13, p < 0.001. In a follow-up analysis, only the inhibitory priming effect from the high-frequency neighbor primes for low-frequency targets was significant (z = 3.99, p < 0.001). The 6-ms facilitation effect from the low-frequency neighbor primes for high-frequency targets was marginal (z = −1.68, p = .094).

In the error rate analysis, a similar pattern emerged. The main effect of Relatedness was significant, ß = −0.15, SE = 0.05, z = −3.16, p = .002, as was the main effect of Frequency Type, ß = −1.14, SE = 0.12, z = −9.42, p < 0.001, indicating that, overall, targets in the related condition and low-frequency targets produced more errors than targets in the unrelated condition and high-frequency targets, respectively. There was, again, an interaction between Relatedness and Frequency Type, ß = −0.11, SE = 0.05, z = −2.36, p = .019. In a follow-up analysis, only the inhibitory effect from high-frequency primes for low-frequency targets was significant (z = −6.19, p < .001).

6.2.2. Bayes Factor analyses

We also conducted a Bayes Factor analysis for word latencies to quantify the statistical evidence in favor of or against the interaction between Relatedness and Frequency Type. The Bayes factor analysis was calculated using the “lmBF” function and “compare” function from the BayesFactor package (version 0.9.12.4.2) with the default JZS type being used to calculate the Bayes factor (Morey et al., 2015). The analysis was based on subject-averaged latencies. Model 1 (the full model with an interaction) was compared with Model 0 (the null model with no interaction). The contrast between these two models produced a BF ₁₀ of 7.59 ± 4.40%, favoring the hypothesis that there was an interaction between Relatedness and Frequency Type in the latency analysis.

6.2.3. Supplementary analysis

Although efforts were made to make sure that the prime-target pairs were not semantically similar, for 10 pairs of neighbors the shared character had the same morphemic meaning in the prime and target words (i.e., those primes and targets were morphologically similar). Therefore, in a supplementary analysis, we removed those 10 pairs of neighbors from the analysis in order to make certain that the effects reported in Experiment 1 did not reflect any meaning similarity. In this analysis, the main effects of Relatedness and Frequency Type were significant (all p’s < .05). Importantly, there was still a significant two-way interaction between Relatedness and Frequency Type, ß = 4.93, SE = 1.25, z = 3.94, p < 0.001. In follow-up analyses, only the 16 ms inhibitory priming effect from the high-frequency neighbor primes for low-frequency targets was significant (z = 3.82, p < 0.001). The 5-ms facilitation effect from the low-frequency neighbor primes for high-frequency targets was not (z = −1.50, p = .133). Essentially, therefore, removing the pairs in which primes and targets had some morphological similarity had no impact on our data pattern. ²

6.3. Discussion

Experiment 1 explored the orthographic neighbor priming effect for two-character Chinese neighbor word pairs. If lexical competition occurs during Chinese word recognition, one would expect to see indications of inhibition in the latency or error rate data. If lexical competition plays no role in reading Chinese words, however, a facilitatory priming effect for Chinese neighbor pairs would be more likely. Our results were consistent with the first possibility. We found an inhibition effect for pairs involving a high-frequency neighbor prime and a low-frequency target in both the latency and error rate data, with no significant priming effect being obtained for pairs involving low-frequency neighbor primes and high-frequency targets. This pattern is consistent with the results from other languages. High-frequency primes cause inhibition for low-frequency targets, whereas observing inhibition effects with low-frequency primes and high-frequency targets is rare. Our lexical competition effect was slightly smaller than that often obtained in alphabetic script language experiments (e.g., Davis & Lupker, 2006; Nakayama et al., 2008; Segui & Grainger, 1990), but slightly more robust than that obtained in Kanji script (Nakayama et al., 2014).

7. Experiment 2

A prime lexicality effect in masked form priming is an important prediction of the localist models such as the IA model (McClelland & Rumelhart, 1981). According to these IA-type models, facilitatory and inhibitory priming effects combine to produce form priming effects. All primes provide activation to the representations of all words that are orthographically similar. However, in the case of word primes, the prime is a competitor of the target. As such, because a high-frequency word prime like “able” will most powerfully activate the lexical representation of itself, a representation that has a high level of activation in general because of its high frequency, that representation will become a powerful competitor. Hence, when the low-frequency target word AXLE is presented, its processing will be slowed due to the competition from the highly activated representation for able. Such is not the case when the prime-target relationship is reversed due to the fact that the representation for axle has a low resting activation level and the representation for able has a high resting activation level. The result is a priming effect that is typically around 0. What is important here, however, is that the facilitatory component of form priming is assumed to be comparable for orthographically similar word and nonword primes. Accordingly, response latencies to the target AXLE would be facilitated by nonword neighbor primes (e.g., azle), even though they are slowed down by high-frequency word neighbor primes (e.g., able).

This assumed impact of the lexicality of orthographic neighbor primes was the main issue in Experiment 2. Numerous studies using alphabetic script languages have consistently demonstrated facilitatory form priming from nonword primes and inhibitory form priming from related word primes (Andrews & Hersch, 2010; Davis & Lupker, 2006; Forster et al., 1987; Forster & Veres, 1998). As well, Kanji experiments (Nakayama et al., 2014) have shown a significant facilitation effect when targets are primed by Kanji nonword neighbors. Zhou et al.’s (1999) experiments, as noted, did not include nonword primes (nor does it appear that this issue has been examined elsewhere in the literature on Chinese word recognition). From a theoretical standpoint, it would be important to test the effect of lexicality on neighbor priming using the set of targets from Experiment 1 in order to determine whether Chinese produces that same pattern as produced in experiments in other languages.

Essentially, if IA-type models are valid for Chinese, one would expect that when using nonword neighbor primes (e.g., two-character Chinese nonword primes that share one character in the same position as in the two-character Chinese word targets), there should be a facilitation effect for low- (as well as high-) frequency targets since nonword primes do not have lexical representations. This is the question addressed in Experiment 2 using the same target words as used in Experiment 1.

7.1. Method

7.1.1. Participants

We conducted power calculations based on the significant interaction obtained in a similar priming experiment (partial η ² = 0.137; Nakayama et al., 2014, Experiment 2). Using G*Power 3.1 software, it was determined that a power of 0.80 would be achieved for a similar-size interaction by using 53 participants. Ultimately, 58 undergraduate students from Zhejiang Gongshang University participated in this experiment. Each received a small gift for their participation. All were native speakers of Chinese and had normal or corrected-to-normal vision with no reading disorder.

7.1.2. Materials

The same 120 pairs of orthographic neighbors used in Experiment 1 were used as word targets. As noted, one member of the neighbor pair was a high-frequency target and the other was a low-frequency target, according to the SUBTLWF frequency counts in the SUBTLEX-CH database (Cai & Brysbaert, 2010).

One hundred twenty nonword neighbors that shared one character with a target in the same position as in the word neighbor were generated to use as nonword primes. The same nonword neighbor primed both members of each neighbor pair when that target appeared in the related condition. As in Experiment 1, both members of the pair were shown to each participant as targets, once in the related condition, and once in the unrelated condition in a counterbalanced fashion. The unrelated prime-target pairs were created by re-pairing the primes and targets used in the unrelated condition for that participant group.

The nonword targets and primes were the same as in Experiment 1. The other details were the same as those in Experiment 1.

7.1.3. Procedure

The basic procedure was the same as that in Experiment 1.

7.2. Results

For word targets, latencies less than 250 ms or greater than 1,800 ms (2.3% of the data) and incorrect responses (8.5% of the data) were excluded from the latency analyses. The mean RTs and percentage error rates, as a function of Relatedness (neighbor vs. unrelated) and Frequency Type (high frequency vs. low frequency) based on the subject means for the word targets are shown in Table 2. For the latency analysis, the model was: RT = glmer (RT ~ Relatedness * Frequency Type + (Frequency Type |subject) + (1 |item), family = Gamma(link = “identity”), control = glmerControl (optimizer = “bobyqa”)). For the error rate analysis, the model was: Accuracy = glmer (accuracy ~ Relatedness * Frequency Type + (Frequency Type |subject) + (1 |item), family = “binomial”, control = glmerControl(optimizer = “bobyqa”)). The other details were the same as those in Experiment 1.

Table 2. Mean lexical decision latencies (RTs, in milliseconds) and percentage error rate for words in experiment 2

Abbreviation. RT = reaction time; %E = percentage error rate.

7.2.1. Latency and error rate analyses

The main effect of Relatedness was significant, ß = −2.64, SE = 1.05, z = −2.51, p = .012, as was the main effect of Frequency Type, ß = −54.64, SE = 3.56, z = −15.37, p < 0.001, indicating that there was a facilitatory priming effect (i.e., latencies were faster following related primes) and that there was a standard target frequency effect (low-frequency target latencies were longer than high-frequency target latencies). There was no interaction between Relatedness and Frequency Type, ß = −0.14, SE = 1.07, z = −0.13, p = .897.

In the error rate analysis, the main effect of Relatedness was significant, ß = −0.11, SE = 0.04, z = −2.55, p = .011, indicating a small (1.2%) inhibitory effect (the related condition produced slightly more errors than the unrelated condition). The main effect of Frequency Type (in the expected direction) was also significant, ß = 1.01, SE = 0.11, z = 9.21, p < 0.001. Again, there was no suggestion of an interaction between Relatedness and Frequency type, ß = −0.02, SE = 0.04, z = −0.48, p = 0.634.

7.2.2. Bayes Factor analyses

We also conducted a Bayes Factor analysis for word latencies to quantify the statistical evidence in favor of or against the interaction between Relatedness and Frequency Type. Model 1 (the full model with an interaction) was compared with Model 0 (the null model with no interaction). The contrast between these two models produced a BF ₁₀ of 0.22 ± 6.05%, favoring the hypothesis that there was no interaction between Relatedness and Frequency Type in the latency analysis.

We also conducted a Bayes Factor analysis for word latencies to quantify the statistical evidence in favor of or against the main effect of Relatedness. Model 1 (the full model with a main effect) was compared with Model 0 (the null model with no main effect). The contrast between these two models produced a BF ₁₀ of 3.02 ± 0.75%, favoring the hypothesis that there was a Relatedness effect in the latency analysis.

Finally, we conducted a Bayes Factor analysis for error rates to quantify the statistical evidence in favor of or against the main effect of Relatedness. Model 1 (the full model with a main effect) was compared with Model 0 (the null model with no main effect). The contrast between these two models produced a BF ₁₀ of 0.29 ± 3.06%, favoring the hypothesis that there was no Relatedness effect in the error analysis.

8. Discussion

Experiment 2’s results using nonword neighbor primes were different from Experiment 1’s results using word neighbor primes with respect to the low-frequency targets, in that the nonword primes produced a significant facilitation effect for those targets, at least in the latency data. That is, even the targets that showed an inhibition effect from high-frequency neighbor word primes in Experiment 1 produced a small but significant facilitation effect in Experiment 2. These results are consistent with IA-type models. That is, because nonwords do not have lexical representations, one would not expect them to produce much lexical competition/inhibition. Therefore, any facilitation obtained due to either orthographic similarity (or morphological relatedness due to the shared constituent character) would be expected to have a facilitatory impact. These results support the idea that lexical competition effects do exist in word recognition for Chinese two-character words, with those effects being noticeable when the primes are high-frequency words, just as is typically observed in other script languages.

One potential puzzle is that our nonword neighbor priming effects were smaller than parallel effects obtained in Kanji (30 ms from nonword neighbor primes for low-frequency word targets, Nakayama et al., 2014). One possibility is that the reduced nonword neighbor priming effects could be a “shared neighborhood” effect, as the Chinese target words have a large number of orthographic neighbors (the mean number of orthographic neighbors is above 56). That is, with a large number of orthographic neighbors, there is more potential for any activation produced by the prime to be divided among the neighbors, leading to a small priming effect (Forster et al., 1987; Nakayama et al., 2008; Van Heuven et al., 2001). However, even if this argument is viable, it may seem a bit surprising that the impact of orthographic similarity was so small. A second (and not mutually exclusive) possibility is that, in Experiment 2, there might have been some inhibition from the unrelated character in the related primes (i.e., the character that, combined with the character shared between the prime and the target, made up the related nonword prime). Experiment 3 was an attempt to examine this issue more closely.

9. Experiment 3

In Experiment 3 we used the same word targets as used in Experiments 1 and 2, and each word target was primed by the single overlapping character of the orthographic neighbor used in those experiments. For example, the word target “容貌” was primed either by a single constituent character “容” or by an unrelated character “怀”. Thus, there is no information in any related prime that is inconsistent with the target and that might inhibit processing of that target. Only low-frequency targets were used as they were the targets that showed different priming effects across the two previous experiments.

Note also that the position of the prime character in the target was manipulated. In left-to-right languages, sequential models of letter processing in word reading propose an initial left-to-right processing sequence, with each letter requiring 10 to 25 milliseconds to process (Whitney, 2001). Furthermore, in paradigms concentrating on full word recognition (e.g., Scaltritti & Balota, 2013) and even sentence reading (e.g., Johnson & Eisler, 2012; Jordan et al., 2003), a distinctive status for letters in the first position has been demonstrated. Chinese is also a left-to-right language, so in this experiment, we introduced position as a factor in order to test whether there is any first character priming advantage. That is, the question was would there be a larger priming effect when Chinese compound targets were primed by their first character than by their second character?

9.1. Method

9.1.1. Participants

We conducted power calculations based on the significant main effect obtained in a similar priming experiment (partial η ² = 0.459; Nakayama et al., 2014, Experiment 4). Using G*Power 3.1 software, it was determined that a power of 0.80 would be achieved with 13 participants. We also noted the sample sizes used in Nakayama et al.’s (2014) Experiment 4 and Zhou et al.’s (1999) Experiment 1, which were 36 and 40, respectively. Ultimately, 40 undergraduate students from Zhejiang Gongshang University participated in this experiment. Each received a small gift for their participation. All were native speakers of Chinese and had normal or corrected-to-normal vision with no reading disorder.

9.1.2. Materials

The word targets were the 120 low-frequency words used in Experiment 1. Each word target was primed by a constituent character or an unrelated character. The constituent character was the same character that was shared in the orthographic neighbor pair in Experiment 1. The unrelated prime-target pairs were again created by re-pairing the primes and targets. Relatedness was again counterbalanced across participants, resulting in two lists with 120 word targets (with 60 related primes and 60 unrelated primes in each list). In both lists, half of the related primes shared the constituent character in the first position and the other half of the related primes shared the constituent character in the second position, and each target was preceded by its constituent character prime in one list and an unrelated character prime in the other list.

Another 120 two-character Chinese nonword targets, selected from those targets used in Experiments 1 and 2, were also included, but only half of the nonwords from those experiments were used in Experiment 3. The manipulation of prime type for the nonword targets was done in a similar fashion as for word targets. However, there was no counterbalancing of lists for nonword targets (there was only one list of 60 constituent character primes and targets and one list of 60 unrelated character primes and targets). The other details were the same as in Experiment 1.

9.1.3. Procedure

The procedure was the same as that in Experiment 1. Note that the one-character prime was presented centrally.

9.2. Results

For word targets, latencies less than 250 ms or greater than 1,800 ms (2.7% of the data) and incorrect responses (11.4% of the data) were excluded from the latency analyses. The mean RTs and percentage error rates, as a function of Relatedness (neighbor vs. unrelated) and Position in the target (first vs. second), from the subject-based analysis for the word targets are shown in Table 3. For the latency analysis of word targets, the model was: RT = glmer (RT ~ Relatedness * Position + (Position |subject) + (1 |item), family = Gamma(link = “identity”), control = glmerControl (optimizer = “bobyqa”)). For the error rate analysis of word targets, the model was: Accuracy = glmer (accuracy ~ Relatedness * Position + (Position |subject) + (1 |item), family = “binomial”, control = glmerControl(optimizer = “bobyqa”)). The other details were the same as those in Experiment 1.

Table 3. Mean lexical decision latencies (RTs, in milliseconds) and percentage error rate for words in experiment 3

Abbreviation. RT = reaction time; %E = percentage error rate.

9.2.1. Latency and error rate analyses

The main effect of Relatedness was significant, ß = −5.48, SE = 2.05, z = −2.67, p = 0.008, indicating faster latencies for related than unrelated primes, whereas the main effect of Position was not significant, ß = −10.57, SE = 6.41, z = −1.65, p = 0.099. The interaction between Relatedness and Position was also not significant, ß = −2.47, SE = 2.02, z = −1.22, p = 0.222.

In the error rate analysis, the main effect of Relatedness was not significant, ß = 0.08, SE = 0.05, z = 1.61, p = 0.107, however, the main effect of Position was significant, ß = 0.39, SE = 0.15, z = 2.62, p = 0.009. Targets primed by related primes overlapping in the second position produced more errors than targets primed by related primes overlapping in the first position, averaged across conditions. Again, there was no significant interaction between Relatedness and Position, ß = 0.08, SE = 0.05, z = 1.53, p = 0.127.

9.2.2. Bayes Factor analyses

We also conducted a Bayes Factor analysis for word latencies to quantify the statistical evidence in favor of or against the null interaction between Relatedness and Position. Model 1 (the full model with an interaction) was compared with Model 0 (the null model with no interaction). The contrast between these two models produced a BF ₁₀ of 0.30 ± 5.97%, favoring the hypothesis that there was no interaction between Relatedness and Position in the latency analysis.

We also conducted a Bayes Factor analysis for word latencies to quantify the statistical evidence in favor of or against the main effect of Relatedness. Model 1 (the full model with a main effect) was compared with Model 0 (the null model with no main effect). The contrast between these two models produced a BF ₁₀ of 1.44 ± 0.71%, only very slightly favoring the hypothesis that there was a main effect of Relatedness in the latency analysis.

9.3. Discussion

A facilitatory priming effect was observed in this experiment, with targets primed by one of their constituent characters being responded to faster and more accurately than targets primed by an unrelated character, even though this effect was weak according to the Bayes factor analysis. These results are consistent with previous priming results using a longer prime duration (Gao et al., 2022; Peng et al., 1994) as well as those observed in alphabetic script languages (Adelman et al., 2014; Peressotti & Grainger, 1999; Stinchcombe et al., 2012) and in Kanji (Nakayama et al., 2014). Moreover, the size of the facilitatory priming effect was not significantly affected by the character position being primed, a result that is also consistent with the findings in other similar masked priming studies (e.g., Shoolman & Andrews, 2003; Nakayama et al., 2014).

Importantly, these results support two further conclusions. First, the small priming effect in Experiment 2 was not due to the existence of a second character in the primes that was incompatible with the target. As noted, it may have been the case that the small size of the priming effects in Experiment 2 was due to the large size of the neighborhood activated by the primes. The single-character primes used in Experiment 3, of course, do not have neighbors in the classical sense. However, the basic argument here would be that they do activate representations of words containing those characters (e.g., see Davis & Taft, 2005). Therefore, the primes in Experiment 3 would have activated a very large set of word representations, potentially preventing any given word’s representation that contained that character from becoming strongly activated.

The second, and more central, conclusion, based on the fact that the identical targets produced different effects in Experiments 3 (where the related primes were constituents of their targets) and 1 (where the related primes were orthographic neighbors of those targets) is that the inhibition effect in Experiment 1 was due to lexical competition among the representations of two-character Chinese words.

10. General discussion

The empirical goal of the present experiments was to examine masked orthographic neighbor priming effects for two-character Chinese words in an effort to contrast those effects with the effects found in other script languages. The specific focus of Experiment 1 was whether there would be an inhibitory neighbor priming effect when prime-target frequency was manipulated in the same way that it is typically manipulated in experiments examining other scripts and when there was essentially no semantic/morphological relationship strength between the primes and targets. Indeed, Experiment 1 showed an inhibition effect for pairs involving a high-frequency neighbor prime and a low-frequency target in both the latency and error rate data, whereas a small, nonsignificant priming effect was obtained for pairs involving a low-frequency neighbor prime and high-frequency target. In Experiment 2, the same targets produced a small but significant facilitatory priming effect in the latency data for both the high-frequency targets and low-frequency targets when primed by nonword primes, again, as typically obtained in other language/script experiments. In Experiment 3, a facilitatory priming effect was obtained in the latency data when the related prime was a component (i.e., one character) of the (low-frequency) target with the size of the priming effect not being significantly affected by character position. In general, the conclusion seems to be that whereas orthographic neighbor priming does obtain for Chinese two-character words, lexical competition plays a role in the recognition of those words.

These results are consistent with localist IA-type models of visual word identification (Davis, 2003, 2010; Grainger & Jacobs, 1996; McClelland & Rumelhart, 1981). One of the fundamental assumptions of the majority of those models is that there is competition among the lexical units of orthographically similar words (i.e., orthographic neighbors). This prediction has been supported by prior research on masked priming with alphabetic script languages (e.g., Davis & Lupker, 2006; Nakayama et al., 2008; Segui & Grainger, 1990). The fact that Nakayama et al. (2011) found a similar inhibitory neighbor priming effect with Japanese Katakana words suggests that lexical processing in nonalphabetic script languages also involves lexical competition. To this point, however, there had been no demonstration that such was also the case in Chinese.

As noted, there was, however, one prior direct examination of inhibitory priming in a logographic script language. Nakayama et al. (2014) using Japanese Kanji stimuli, demonstrated inhibitory neighbor priming effects, although those effects were typically limited to error rates (see their Experiments 1A, 1B, and 3A). Only in their Experiment 3B, when participants were instructed to prioritize accuracy over speed when responding, was an inhibitory neighbor priming effect shown in the latency data. Therefore, evidence for inhibitory effects in that set of experiments was a bit weak. In the present Experiment 1, participants were given standard lexical decision instructions (i.e., to respond as quickly and accurately as possible) with the result being that inhibitory priming effects were obtained in both latencies and error rates, meaning that our data were a bit more similar to data obtained in previous alphabetic script language studies (e.g., Andrews & Hersch, 2010; Brysbaert et al., 2000; Davis & Lupker, 2006; Nakayama et al., 2008; Segui & Grainger, 1990).

Our findings differ to a greater degree from those of the previous investigations of neighbor priming effects for two-character Chinese words, particularly the extensive set of experiments reported by Zhou et al. (1999). Zhou et al. reported a significant facilitatory priming effect in essentially all relevant orthographic priming conditions. As noted, however, a crucial factor that was not controlled in Zhou et al.’s studies was the relative frequencies of the primes and targets, with the average prime frequencies being either less than or equal to the average target frequencies. In that situation, it would be difficult for the primes to produce strong lexical competition for the targets. As noted, the inhibitory priming effect in our Experiment 1 only emerged when the prime frequencies were considerably higher than the target frequencies. In addition, Zhou et al. did not endeavor to make sure that prime-target pairings (both neighbor pairs and unrelated pairs) were not semantically related at the whole word level, as was done in our Experiment 1. It is plausible that the differences in findings between that experiment and Zhou et al.’s experiments may have been at least partially due to there being some level of semantic relatedness between primes and targets in their experiments even in their CHAR condition.

10.1. Disentangling the impact of morphological and orthographic similarity

Zhou et al.’s (1999) experiments are not the only experiments that seem to show morphological priming in Chinese. As noted, in their experiments, the orthographically related prime-target pairs in their MORPH condition were also morphologically (and semantically) related in the sense that the overlapping character represented the same morpheme in the prime and target. When there was no morphological overlap (i.e., their CHAR condition), there was reduced priming compared to their MORPH condition. One interpretation of this pattern is that both effects are orthographic with the additional effect in the MORPH condition being due to morphological similarity. (As the primes were all low-frequency words, one would not have expected them to have produced much lexical competition.) However, an alternative interpretation is that the effects were completely due to orthographic similarity and there is an inhibition of processing whenever a shared character in the prime and target must activate different morphemes.

This latter (morphological competition) interpretation gains some support from the results of Yang et al. (2022) and Tsang and Chen (2014). Yang et al. examined masked transposed character priming of two-character Chinese words. They showed that priming effects were reduced when the two characters represented morphemes that were unrelated to the full meaning of the target word (English example: petcar priming CARPET). Tsang and Chen (2014) also investigated masked priming of two-character Chinese words. In their related condition, the first character of the prime and target matched. In their crucial condition (their “transparent-opaque” condition), the shared character in the prime represented a morpheme consistent with the meaning of the whole word while the shared character in the target did not. This condition, which, in many ways, resembles the CHAR condition in Zhou et al.’s (1999) experiments, produced no priming effect, suggesting that it was difficult to process the meaning of the shared character once the prime had activated the more common morpheme.

The idea of morpheme competition is not a new one as it has already been incorporated into Taft and Nguyen-Hoan’s (2010) model. However, coupled with the present results, what it leaves us with is the challenge of teasing apart what is being facilitated and what is being inhibited in masked priming experiments in Chinese. One would assume that the best way to separate the impact of orthography and morphology is to contrast priming effects of one factor when there is no potential for priming from the other. Doing so is almost impossible when trying to investigate morphological priming since morphologically related primes in Chinese are, by definition, orthographically identical. Doing so is seemingly possible when investigating orthographic priming as was done in the present experiments in which the primes and targets were only minimally morphologically similar. However, even in that situation, the issue is not simple. If the prime and target are orthographically similar (i.e., sharing a character) but those two characters are associated with different morphemes, the potential for morphological competition becomes a real one. In contrast, there would likely not be competition in the unrelated prime condition. The result would be a reduction in the size of any facilitation effect. To the extent that there might have been morphological competition in the present experiments, that fact, together with the issue of the large neighborhood sizes of the primes, may provide an explanation for why our facilitatory priming effects in Experiments 2 and 3 were so small. ³

11. Conclusions

The present research offers new evidence concerning how orthographic neighbors affect reading Chinese two-character words based on the results of three masked priming experiments. Our theoretically-based goal was to test whether the nature of orthographically-driven lexical processing when reading two-character Chinese words paralleled that in other languages with a key question being whether lexical competition occurs when reading those words. We found the typical (in other scripts) inhibition effect only for high-frequency neighbor prime and low-frequency target word pairs in both the latency and error rate data when the contribution of morphology/semantics was minimized. Further, when the same targets were primed by their nonword neighbors or constituent characters, the inhibitory priming effects for the low-frequency word targets changed to become a facilitation effect. Together, these findings support the idea that orthographically-driven lexical processing involved in reading Chinese is reasonably similar to that in other script languages in that it involves a lexical competition process, consistent with the assumptions of localist IA-type models.