Unpacking word segmentation processes by L2 Chinese readers: Evidence from eye movements

Abstract

Sentences written in Chinese are composed of continuous sequences of characters, without spaces or other visual cues to mark word boundaries. While skilled L1 readers can efficiently segment this naturally unspaced text into words, little is known about the word segmentation capabilities of L2 readers, including whether they employ the same strategies to process temporary segmental ambiguities. Accordingly, we report two eye movement experiments that investigated the processing of sentences containing temporarily ambiguous “incremental” three-character words (e.g., “体育馆,” meaning “stadium”) whose first two characters could also form a word (“体育,” meaning “sport”), comparing the performance of 48 skilled L1 Chinese readers and 48 high-proficiency L2 Chinese readers in each experiment. Our findings reveal that both groups can process this ambiguity efficiently, employing similar word segmentations strategies. We discuss our findings in relation to models of eye movement control and word recognition in Chinese reading.

1. Introduction

Most sentence processing research focuses on monolingual reading, yet there is growing interest in how second-language (L2) users interpret sentences in real time. A central question is whether L2 readers rely on the same or different cognitive mechanisms for sentence processing and ambiguity resolution (Hopp, 2022).

Current L2 reading research spans lexical access (Kroll et al., 2022), syntactic processing and ambiguity resolution (Brothers et al., 2021; Cunnings, 2017; Roberts & Felser, 2011) and predictive processing (Kaan & Gruter, 2021; Schlenter, 2023). Evidence suggests L2 readers, like native (L1) readers, process sentences incrementally, integrating each new word as part of their current interpretation of text. Moreover, both groups experience difficulty when this new information conflicts with prior context (Lee & Witzel, 2023; Williams, 2006). L2 readers, however, are typically slower and more error-prone in processing syntactic ambiguities, possibly reflecting weaker lexical–syntactic representations and higher cognitive load effects (Cop et al., 2015; Conklin et al., 2020).

L2 readers’ ability to process ambiguities appears to depend strongly on their L2 proficiency, while L2 processing also appears to be shaped by first-language (L1) knowledge. For example, studies on interlingual homographs, words with similar spelling but different meanings across languages (e.g., coin, meaning “corner” in French, and “metal currency” in English), shows that L1–L2 orthographic overlap affects meaning activation (Cop et al., 2017; Fernandez et al., 2021; Libben & Titone, 2009; Schwartz & Kroll, 2006; Titone et al., 2011; Tiffin-Richards, 2024a,b; Van Assche et al., 2011). Similarly, L1 syntax can influence L2 sentence processing, with facilitative or inhibitory effects depending on the cross-language structural similarity (Bardovi-Harlig & Sprouse, 2018; Choi & Ionin, 2021; Schwartz & Sprouse, 1996). When L1 and L2 share syntactic patterns, learners process sentences more efficiently (Kidd et al., 2015), whereas structural differences can cause misinterpretations (Hatzidaki et al., 2011; Roberts & Felser, 2011; Sasson et al., 2024).

This study extends this research into L2 sentence processing by examining how L2 readers segment and identify words during Chinese reading, focusing on processing of segmental ambiguities that occurs when a sentence segment supports competing lexical interpretations. Resolving such ambiguities is essential for fluent reading. We therefore investigated whether native (L1) and proficient L2 Chinese readers use similar or different segmental processing strategies to resolve ambiguities. To test this, we manipulated the contextual plausibility of alternative interpretations of a segmental ambiguity. Therefore, as well as providing insights into L2 segmental processing, this approach was likely to be informative about how L2 readers handle conflicts between new and prior contextual information (Lee & Witzel, 2023; Williams, 2006).

1.1. Word segmentation in Chinese reading

Chinese uses a logographic script composed of equally spaced monomorphemic (and monosyllabic) characters (see Li et al., 2015; Shen, 2018; Zang et al., 2011). While some words are single characters, most are multicharacter compounds, with around 70% of common words being two-character compounds (Beijing Language Institute, 1986). As Chinese lacks spaces between words, readers must infer word boundaries. However, skilled L1 readers are not hindered by this (Bai et al., 2008; Zang et al., 2013), and read at about 250 words per minute, comparable to reading rates for spaced alphabetic scripts like English (Liversedge et al., 2016). This suggests that L1 readers possess efficient strategies for grouping characters into words and establishing word boundaries without relying on spaces or other visual cues.

Little is known about the segmentation capabilities of L2 Chinese readers, however; including whether they employ the same segmental processing strategies as L1 readers. Beginning L2 learners typically use textbooks that employ intercharacter spaces to mark word boundaries in sentences (Cui et al., 2012), with studies showing that these visual cues help L2 readers process text more efficiently (Bai et al., 2013; Bassetti & Lu, 2016; Cui, 2023; Ma et al., 2019; Shen et al., 2012; Yu, 2022; Zhou, Ma, et al., 2018; Zhou, Wang, et al., 2018, 2020; but see Bassetti, 2009; Bassetti & Masterson, 2012). However, less is known about how L2 readers process words in unspaced text. Offline studies suggest that L2 readers make frequent segmentation errors that can impair comprehension when text is unspaced (Shen & Jiang, 2013; Shen & Dai, 2024; Yang, 2021). However, little work to date has examined how L2 readers segment words in naturally unspaced text during online sentence processing.

Accordingly, we used eye-movement measures reading to compare the online segmental processing strategies of L1 and L2 Chinese readers. Eye movements are highly informative about word identification processes during reading (Liversedge & Findlay, 2000; Rayner, 1998, 2009) and widely used in L2 reading research (see Frenck-Mestre et al., 2005; Roberts & Siyanova-Chanturia, 2013). They also are informative about L1 Chinese readers’ processing of segmental ambiguities (Huang & Li, 2020; Li et al., 2024, 2009; 2022; Ma et al., 2014; Wang et al., 2021; Zhou & Li, 2021; Zhou et al., 2018; Yao et al., 2025). This includes research comparing serial and parallel accounts of how readers process incremental segmental ambiguities in multicharacter sequences such as 体育馆, where all three characters form a word meaning “stadium,” but the first two characters also constitute a standalone embedded word, 体育, meaning “sport.”

1.2. Theoretical accounts of Chinese word segmentation

Serial accounts propose that Chinese readers process characters in strict linear order from left to right, making incremental decisions about how they combine to form words. A version of this account by Perfetti and Tan (1999) suggests that skilled readers rely on their knowledge that most words are two characters long to guide segmentation decisions, by first grouping adjacent characters into two-character units. Their account therefore predicts that an incremental three-character word such as 体育馆 (“stadium”) will initially be processed as the two-character standalone word 体育 (“sport”) followed by an additional character, leading readers to first identify the embedded word and then backtrack to reanalyze the sequence as a three-character compound, resulting in a processing cost.

Parallel accounts, implemented by the Chinese Reading Model (CRM; Li & Pollatsek, 2020), assume that alternative segmental analyses are considered simultaneously (Inhoff & Wu, 2005; Li et al., 2009). Other eye movements models such as E-Z Reader (Reichle et al., 1998, 2003), SWIFT (Engbert et al., 2005) and OB1-reader (Snell & Grainger, 2019) depend on interword spaces for eye-movement guidance and word identification and are difficult to adapt for Chinese (Zhou & Li, 2021; Reichle & Yu, 2018; Yu et al., 2021), whereas the CRM was designed specifically to account for word segmentation and identification in unspaced Chinese. The CRM assumes that, on each fixation, readers process a small group of characters within their perceptual span (the region of visual information that a reader can effectively process during a single fixation, McConkie & Rayner, 1975, 1976; Rayner et al., 1980). This comprises about one character to the left and two or three characters to the right of fixation for skilled L1 Chinese readers (Inhoff & Liu, 1998), but is likely to be smaller for L2 readers (see Fernandez et al., 2023; Jordan et al., 2014; Leung et al., 2014; Paterson et al., 2014; Whitford & Titone, 2015). Within the CRM, all possible character groupings within this region are assumed to be processed in parallel, with alternative segmental analyses competing for selection via mechanisms inspired by models of visual word recognition (e.g., McClelland & Rumelhart, 1981).

When reading a multicharacter sequence like 体育馆, it is assumed that lexical entries for both whole-word (“stadium”) and embedded word (“sport”) analyses are activated simultaneously and compete for selection. However, the whole-word form usually wins because it receives stronger bottom-up input from more characters. Once selected, this word is identified and segmented from the text, while activation of its embedded word declines rapidly. The CRM therefore predicts holistic processing of the ambiguity without reanalysis costs. The model further assumes that, though multiple candidates are activated in parallel, only one word ultimately is recognized at a time, preserving serial word recognition (Reichle et al., 2009; White et al., 2018, 2020). The model also incorporates eye-guidance mechanisms which assume readers subconsciously estimate how many characters they can identify on each fixation and program their next eye movement to land beyond these characters (see Li et al., 2011; Wei et al., 2013).

Zhou and Li (2021) compared these accounts in two eye-movement experiments with L1 readers. In the first, participants read sentences containing either an incremental word (e.g., 体育馆 “stadium”) or a control target identical to its embedded word (体育 “sport”). While the incremental word was always plausible, the control target word (and therefore the embedded word) could be plausible or implausible depending on the characteristics of a preceding verb. The control word should therefore elicit longer fixations when implausible (the word plausibility effect, see Staub, 2015), and a similar effect may also be observed for the embedded word if it is segmented first during processing. As Zhou and Li observed plausibility effects for control but not embedded words, this suggested that readers did not first identify the embedded word. A second experiment independently manipulated the plausibility of incremental and embedded word analyses of an ambiguity. Plausibility effects were observed only for the incremental words, further indicating that embedded words are not initially identified during sentence processing.

1.3. This study

Segmental ambiguities are frequent in Chinese, making it vital to understand how L1 and L2 readers resolve them. This study therefore extends Zhou and Li (2021) by investigating whether skilled L1 and high-proficiency L2 Chinese readers use the same or different segmental processing strategies. If they use the same strategies, both groups should replicate Zhou and Li’s findings. However, if they use different strategies, L1 readers may replicate Zhou and Li’s findings, while L2 readers exhibit processing differences.

One possibility is that higher cognitive load in L2 reading (Hopp, 2010) reduces the perceptual span (Fernandez et al., 2023; Jordan et al., 2014; Leung et al., 2014; Paterson et al., 2014; Whitford & Titone, 2015), causing fewer characters to be processed per fixation. L2 readers may also have reduced familiarity with three-character words, and these factors together may lead L2 readers to adopt a more serial segmentation strategy in which they initially identify embedded words. In this case, embedded word plausibility effects may be observed for L2 but not L1 readers. It is also possible that L2 readers evaluate plausibility less efficiently (Lee & Witzel, 2023; Williams, 2006), leading to later or slower detection of implausible segmentations and greater difficulty recovering from errors.

Following Zhou and Li (2021), we examined fixation times for target regions containing the incremental or control target words to test sensitivity to embedded word plausibility. If both groups process incrementally and holistically, plausibility effects should be observed for both groups for control word but not for embedded words. However, if L2 readers process more serially, fixation times may reflect both embedded and incremental word plausibility effects, while differences in the timing and magnitude of these effects may offer insights into differences in L1–L2 plausibility processing.

We also analyzed fixation times for the pretarget verb used to manipulate plausibility. One possibility is that, while fixating this verb, readers begin acquiring semantic information from the forthcoming target word, allowing them to start evaluating this word’s contextual plausibility. This would represent an instance of parafoveal processing, where linguistic information is obtained from a word in the reader’s peripheral vision before it is fixated (see Cutter et al., 2015; Schotter et al., 2012). Zhou and Li (2021; also Yao et al., 2025) observed no such effects, suggesting their readers did not process upcoming word meaning in advance. However, other work suggests readers can extract parafoveal meaning information, including about a word’s contextual plausibility (Yan et al., 2009; Yang et al., 2012; Pan et al., 2016). Accordingly, examining this issue further could clarify whether parafoveal processing influences segmentation. Crucially, L1 and L2 readers may differ in parafoveal processing ability. While some studies show proficient L2 readers use parafoveal information efficiently (Tiffin-Richards, 2024a,b), others report limited use by L2 Chinese readers (Cong & Chen, 2022; Xiao et al., 2021). Accordingly, we report analyses comparing the parafoveal processing of segmental ambiguity across L1 and L2 groups.

2. Experiment 1

Experiment 1 compared eye movements of skilled L1 and high-proficiency L2 Chinese participants during sentence reading. The sentences included an incremental three-character target word that was always contextually plausible, which included a two-character embedded word that was either contextually plausible or implausible. These sentences were compared with control sentences where the embedded word replaced the incremental word as the target word.

Using a similar design, Zhou and Li (2021) found that skilled L1 readers showed plausibility effects for control but not embedded words, as reflected in total reading times at the target region. We therefore examined whether a similar effect would be observed for L1 and L2 readers in out experiment, while also running exploratory analyses using additional eye-movement measures to identify possible group processing differences.

Von der Malsburg and Angele (2017) highlight that using multiple dependent variables in eye-movement research substantially increases the risk of Type I errors. To mitigate this, they recommend specifying hypotheses in advance, interpreting effects from a single measure with caution unless predicted a priori, and applying Bonferroni or similar corrections. However, they also caution that Bonferroni adjustments may be overly conservative for highly correlated fixation time variables. As an alternative, they suggest considering an effect reliable if it reaches significance in two or more related measures.

This study addressed this increased risk of Type I error by clearly distinguishing between hypothesis-driven and exploratory analyses. This enabled testing of specific theoretical predictions while exploiting the richness of eye-movement data. For hypothesis-driven tests, we applied Von der Malsburg and Angele’s (2017) criteria, using Bonferroni-corrected thresholds for multiple comparisons and a convergence heuristic deeming effects reliable if significant in two or more dependent measures. These criteria guided interpretation without serving as absolute cut-offs.

Ethics statement. The study was approved by the ethics committee of the Academy of Psychology and Behavior, Tianjin Normal University, and conducted in accordance with the Declaration of Helsinki (World Medical Association, 2013).

2.1. Method

Participants. Participants were 48 L1 Chinese readers aged 20–29 years (M = 22; 30 females), all undergraduates at Tianjin Normal University, and 48 L2 Chinese readers aged 18–32 years (M = 24; 27 females) enrolled in language courses at Tianjin Normal University or Fujian Normal University. All L1 readers were native Mandarin speakers with some English knowledge, as English is taught from primary school in mainland China. L2 readers were native users of diverse mainly alphabetic languages differing in word-spacing conventions (see Supplementary Table S5). All L2 participants had at least 2 years of Chinese learning experience (M = 7.6 years, SD = 4.7) and had achieved at least HSK level 4 (the standard six-level Chinese proficiency test administered by the Ministry of Education of China). Their average proficiency was HSK-5 (SD = 1), mean total score 226/300 (SD = 32) and reading score 75/100 (SD = 13), indicating moderate–high proficiency.

L2 participants completed the LHQ 3.0 (Language History Questionnaire; Li et al., 2019), which assesses second-language proficiency (Zhang et al., 2023) and collects self-report data on learning experience, usage and skill levels via its web interface (LHQ 2.0; Li et al., 2014). L2 readers self-rated moderate to high overall and reading proficiency (M = 0.59, SD = 0.13; reading M = 4, SD = 0.94).

Prior eye-movement studies provide limited effect size data for L1 Chinese segmental processing and none for L2 readers. To ensure adequate power, we used larger samples and stimuli than previous work, increasing observations per condition from 400 (Zhou & Li, 2021) to 960. We employed two strategies to ensure our study was adequately powered effects. First, we increased the participant sample size and stimulus set relative to previous research, raising the number of observations per condition from 400 in Zhou and Li (2021) to 960 in this study. We also ran a priori power analyses (powerSim, powerCurve; Green & MacLeod, 2016, in R: R Core Team, 2022) using data from a 30-participant pilot later included in the final sample (Huang & Li, 2023). Full details are in the Supplemental File (Section S1). Analyses confirmed sufficient power to detect key effects reported previously.

Stimuli and design. Stimuli consisted of 80 sentence sets, each containing a region of interest with either a two-character control word or a three-character incremental target. Twenty sets were adapted from Zhou and Li (2021) for L2 suitability and 60 new sets were created using the same procedures. For each set, the verb preceding the target was manipulated to yield four experimental conditions (see Table 1). We ensured all three-character incremental words appeared in a current standard dictionary, confirming their formal recognition in Chinese. A complete list of incremental and embedded words with English translations is provided in the Supplemental File (Section S6).

Table 1. Example sentence stimuli for Experiment 1

Note. In Experiment 1, characters in pretarget verbs are underlined and bold ( 关注/修建 ), characters in target words are italicized and bold ( 体育/体育馆 ).

^a This refers to the plausibility of the embedded words in incremental target word contexts and the plausibility of two-character target words in the control sentences. Note that the control words were identical to the corresponding embedded word. In Experiment 2, characters in pretarget verbs are underlined and bold ( 关注/修建/促进/查阅 ), characters in target words are italicized and bold( 体育馆 ).

In one condition (plausible embedded), the verb (e.g., 关注, “pay attention to”) was followed by a three-character incremental target (e.g., 体育馆, “stadium”) that was always plausible in the sentence context, and where its first two characters formed an embedded two-character word (e.g., 体育, “sport”) that also was plausible. In a second condition (implausible embedded), the verb was changed (e.g., to 建设, “build”) so that the incremental word stayed plausible, but the embedded word became implausible. Two further control conditions used the two-character embedded word itself as a control target word, producing plausible (e.g., 关注体育, “pay attention to sport”) and implausible (e.g., 建设体育, “build sport”) variants. Details of stimulus characteristics appear in the Supplemental File (Section S7).

Stimuli were placed in four lists. Each began with four practice sentences and included 80 experimental and 20 plausible filler sentences. Each list contained 20 items per condition, assigned via a Latin square so each sentence appeared once per list with balanced condition representation. The experimental and filler sentences were presented in random order. Equal numbers of L1 and L2 participants were pseudo-randomly assigned to each list.

Apparatus and procedure. An EyeLink 1000 Plus eye-tracker (SR Research, Canada) recorded each participant’s right-eye gaze during binocular viewing, with <.01° RMS spatial and 1000 Hz temporal resolution. A headrest and chinrest minimized movement. Sentences appeared on a 24-inch high-resolution LCD monitor (1920 × 1080 pixels, 150 Hz) in Song font as black text (RGB 0,0,0) on a light gray background (RGB 220,220,220). At a 75 cm viewing distance, each character subtended ≈1° horizontally, typical for reading (Xu & Jordan, 2009). Participants read individually and were instructed to read normally for comprehension. At the start, a 3-point horizontal calibration was performed along the same line as each sentence presentation, ensuring ≤.35° spatial accuracy. Calibration was checked before each trial and repeated as necessary to ensure the same high spatial accuracy throughout the experiment. Each trial began with a fixation square equal to one character on the left side of the screen. Once fixated, the sentence appeared, with its first character replacing the square. Participants pressed a key when they finished reading. On 25% of trials, a yes/no comprehension question followed, with responses recorded. The experiment lasted about 40 minutes per participant.

2.2. Results

Accuracy answering comprehension questions was high for all participants >80%, native, M = 95.5%; L2, M = 88.1%, p < .001), indicating that though the L1 group had higher accuracy, both groups understood the sentences well.

Data analysis procedure. Following standard procedures, fixations less than 80 ms and longer than 1200 ms were removed (affecting 4.9% of fixations for L1 participants and 3.1% of fixations for L2 participants). Trials in which participants had less than 5 fixations were removed, resulting in 0.3% loss of trials for L1 readers and 0.1% for L2 readers, attributable to track loss or failure to read sentences sufficiently carefully. The remaining data were analyzed by linear mixed effects models (LMEMs; Baayen et al., 2008) using R (R Core Team, 2022) and the lme4 package (Bates et al., 2011). The p values were estimated using the lmerTest package (Kuznetsova et al., 2014). For binomial variables, generalized LMEMs were conducted with the Laplace approximation. A maximal random effects structure was used (Barr et al., 2013), with participants and stimuli as crossed random effects. If models did not converge, the random effects structure was trimmed, first by removing correlations then slopes for stimuli, then correlations and slopes for participants, until the model converged. The models are listed in the Supplemental File (Section S2).

Sentence-level analyses. In addition to word-level analyses reported next, we examined sentence-level eye movements for the L1 and L2 groups. Full details are provided in the Supplemental File (Section S3). The results show that L2 readers exhibited longer sentence reading times, made more and longer fixations, shorter forward saccades and a higher number of regressions, consistent with greater reading difficulty in the L2 group.

Word-level analyses. We conducted word-level analyses for a target region comprising either the incremental word or control word and pretarget region comprising the preceding verb. Target word analyses were informative about fixational processing of the target words, while pretarget analyses were informative about both parafoveal processing of these words and subsequent re-reading effects.

Following Zhou and Li (2021), we used linear mixed effects models to test theoretically motivated hypotheses, using a single variable encompassing all the experimental conditions as a fixed effect. Customized contrasts (Schad et al., 2020) were used to test effects of (1) target word type (incremental word vs. control word); (2) control word plausibility (plausible vs. implausible) and (3) embedded word plausibility (plausible vs. implausible). These are reported separately for L1 and L2 groups, following which we report further customized contrasts that include the variable group to compare L1 versus L2 effects. Continuous variables (fixation times) were log-transformed (Wagenmakers et al., 2012). Contrasts were defined using the contr.sdif function in the MASS package (Venables & Ripley, 2002). As each variable had two levels, these produced main effect coding equivalent to other methods (e.g., contr.sum function). We coded contrasts using deviation coding (−1/1), though sum-to-zero coding (−0.5/0.5) produced the same patterns of effects.

Word-level analyses used standard eye movement measures informative about first-pass reading; that is, the initial processing of a word prior to a fixation to its right or a regression (e.g., Rayner, 2009). These comprised word skipping (probability of not fixating a word during first pass), first-fixation duration (FFD; duration of the first progressive fixation on a word) and gaze duration (GD; sum of all first-pass fixations on a word). Note that single-fixation durations (duration of the first progressive fixation for a word receiving only one first-pass fixation) are not reported as L2 readers rarely made only one first-pass fixation on words. We additionally report word-level measures informative about later processing; namely, total reading time (TRT; sum of all fixations on a word) and regressions-in (RI; probability of a regression back to a word). Tables 2 and 3 present means for the target and pretarget regions and summarize their statistical effects, respectively.

Table 2. Mean eye movement measures for target and pretarget regions in Experiment 1

Note. The standard error of the mean is given in parentheses. Fixation time measures are shown in ms.

^a These labels indicate whether the embedded words were plausible or implausible.

Table 3. Summary of statistical effects for Experiment 1

Note. Asterisks indicate p < .05. FFD = First-fixation duration, GD = gaze duration, TRT = total reading time, RI = probability of a regression-in.

Target region analyses. Target word type (incremental vs. control) influenced gaze durations and total reading times for both groups, with longer reading times for the incremental words (gaze duration: L1, 329 vs. 279 ms; L2, 892 vs. 592 ms; total reading time: L1, 495 vs. 440 ms; L2, 1418 vs. 947 ms). L1 first fixations also were longer for incremental words (238 vs. 227 ms), though no L2 effect was observed (298 vs. 302 ms). L2 regressions-in probabilities were higher for the incremental than control words (.31 vs. .26), with no L1 effect (.25 vs. .26). Significant Group × Word Type interactions reflected larger L2 differences, likely due to their reduced familiarity with three-character words. Assuming three main dependent measures (gaze duration, total reading time, regressions-in), Bonferroni correction set the threshold at p < .017. Gaze-duration and total-time effects exceeded this (p < .001) and appeared in two measures, confirming robustness.

We next examined control word plausibility effects to test whether L1 and L2 groups showed total reading time effects, as in Zhou and Li (2021). L1 readers showed longer total reading times for implausible targets (463 vs. 415 ms); while L2 readers showed a similar but nonsignificant trend (978 vs. 916 ms). An L2 plausibility effect appeared in regressions-in (.29 vs. .22), with no L1 counterpart (.28 vs. .24). As plausibility effects were absent in first-pass measures, these total reading time and regression effects likely reflect later, postinitial processing. Bonferroni correction (p < .017) confirmed that the L1 total reading time and L2 regressions-in effects were reliable. Overall, the findings replicate Zhou and Li (2021) for L1 readers and show that L2 readers show have greater difficulty computing word plausibility during online sentence processing. Finally, we examined embedded word plausibility effects in sentences with plausible incremental targets. No effects were found for either group, consistent with Zhou and Li (2021). This suggests that neither group processed the embedded word interpretation during fixations on incremental target words.

Pretarget region analyses. The interchangeable verbs used to manipulate plausibility were matched for word and character frequency and visual complexity (stroke count). To account for any residual variation, pretarget region models included log-transformed and centered word frequency plus centered frequencies and stroke counts for both characters. Models with and without these covariates yielded the same fixed effects, with no group differences, confirming that verb characteristics did not influence pretarget eye movements.

We first examined target word type effects (incremental vs. control), which appeared only in late measures. Total reading times were longer for verbs followed by a control than an incremental word (L1: 467 vs. 428 ms; L2: 1025 vs. 977 ms). L1 regressions-in were also higher for verbs preceding a control than incremental target (.34 vs. .26), with no L2 effect. As these occurred only in late measures, they reflect re-reading effects. Bonferroni correction (p < .017) confirmed that the L1 effects were robust (p < .001), though the L2 total reading time effect was not. These re-reading effects therefore appear stronger in L1 than L2.

Next, we assessed control word plausibility effects. Consistent with Zhou and Li (2021), both groups showed total reading time and regressions-in effects. Total reading times were longer (L1: 498 vs. 435 ms; L2: 1077 vs. 972 ms) and regressions-in were higher (L1: .38 vs. .29; L2: .32 vs. .26) for verbs followed by an implausible than plausible control word. All effects exceeded p < .01 and Bonferroni p < .017, indicating robustness. No gaze-duration effects emerged, suggesting control word plausibility was not processed parafoveally. The total-time and regression effects therefore reflect re-reading effects following both L1 and L2 readers experiencing difficulty integrating an implausible control target word.

We next considered embedded word plausibility for sentences with incremental target words (which were always plausible). Both groups showed longer total reading times for verbs when the embedded word was implausible (L1: 439 vs. 417 ms; L2: 1011 vs. 942 ms). These exceeded p < .01 and Bonferroni p < .017, indicating robust effects. L1 gaze durations also were longer for verb followed by implausible embedded words (303 vs. 285 ms), with a marginal L2 trend (643 vs. 604 ms). The L1 effect did not reach the Bonferroni-corrected threshold, but appeared in both gaze duration and total time, satisfying the convergence heuristic, and therefore, was considered reliable. It is noteworthy that Zhou and Li (2021) did not report a comparable parafoveal processing effect. Figure 1 illustrates these effects.

Figure 1. Embedded word plausibility effects for the pretarget region in Experiment 1, in (A) gaze durations and (B) total reading times.

Further exploratory analyses examining effects of L2 native language characteristics and L2 Chinese reading proficiency are reported in the Supplemental File (Section S5).

2.3. Discussion

Experiment 1 replicated and extended findings reported by Zhou and Li (2021), while providing new evidence about how L1 and L2 readers process segmental ambiguity. Participants read sentences containing a three-character incremental target word whose first two characters could form a standalone embedded word, which was either plausible or implausible in the sentence context. Eye movements for these sentences were compared with control sentences in which the embedded word served as a control target word, replacing the incremental word. We examined effects for both a target word region and a pretarget region containing a two-character verb.

At the target region, both groups showed increased gaze durations and total reading times for incremental versus control target words, with larger effects for L2 readers. These likely reflect greater difficulty in processing three-character incremental words, especially among L2 readers, who may be less familiar with such lexical forms. The robustness of these findings, confirmed across multiple measures, underscores the cognitive cost of processing segmentally ambiguous words during reading. In line with Zhou and Li (2021), we also observed an L1 plausibility effect for control target words in both total reading times and regressions-in, indicating the effectiveness of our plausibility manipulation, and suggesting that L1 readers re-evaluate contextual fit with earlier sentence content when encountering an implausible continuation. The L2 group showed a plausibility effect in regressions-in only, suggesting more limited contextual sensitivity. Crucially, neither group exhibited an embedded word plausibility effect at the target region for sentences with incremental target words, suggesting neither group gained lexical access to the embedded word’s lexical representation during fixational ambiguity processing. These findings mirror those of Zhou and Li for L1 readers. Zhou and Li interpreted their results as supporting a parallel activation account of segmental ambiguity resolution, as implemented by the CRM (Li & Pollatsek, 2020). According to this account, embedded and whole-word analyses are activated in parallel, with the whole-word interpretation usually winning the competition for selection, without providing full lexical access for the embedded word. Our target region findings align with this explanation, while similarity between L1 and L2 patterns suggests the two groups exhibit a similar whole-word preference when resolving segmental ambiguity.

Pretarget region analyses also revealed word plausibility effects. Implausible control target words led to increased regressions and longer total reading times for both groups, suggesting re-reading due to integration difficulty. Crucially, we observed similar pretarget region effects of embedded word plausibility in total reading time for sentences containing incremental target words. This was consistent with both groups re-reading the verb when the embedded word analysis was contextually implausible. The L1 group showed a similar effect in gaze durations, indicating that the effect also occurred during first-pass processing for the pretarget verb. This suggests that early, parafoveal access to embedded word meaning additionally produced contextual mismatch effects for L1 readers.

The re-reading effects for both groups at the pretarget region indicate lexical access to embedded word representations during sentence processing by both L1 and L2 readers, with this causing integration difficulty, similarly to effects observed for control words. L1 gaze-duration effects additionally suggest these readers lexically access the embedded word representation during an early stage of reading associated with the parafoveal processing of target words. Crucially, this pattern challenges the interpretation of segmental ambiguity processing offered by Zhou and Li (2021; also Li & Pollatsek, 2020), who propose that embedded words are activated but not fully lexically accessed during ambiguity resolution. We consider the implications of these findings in the General Discussion.

3. Experiment 2

A potential limitation of Experiment 1 was that the incremental target words were always plausible. Readers likely take account of plausibility when segmenting words and may be more likely to segment an embedded word when the incremental word does not fit plausibly with the prior context. To take account of this, Experiment 2 manipulated the contextual plausibility of incremental target words and their embedded words orthogonally.

As in Experiment 1, we sought to minimize the likelihood of false positive effects by focusing hypothesis-driven tests on measures that previously showed L1 plausibility effects in Zhou and Li’s (2021) study. For the target word region, hypothesis testing assessed whether: (1) whole-word plausibility influenced total reading times and (2) embedded word plausibility affected total reading times when the whole word was plausible, but not when it was implausible. We also conducted exploratory analyses across a broader range of eye movement measures to examine L1 and L2 differences in ambiguity processing. For the pretarget verb region, we conducted hypothesis-driven tests to assess parafoveal effects of embedded or target word ambiguity, as reflected in gaze durations. As with the target region, these exploratory analyses were performed to gain a fuller understanding of processing differences between L1 and L2 groups.

3.1. Method

Participants. Participants were 48 L1 Chinese readers aged 19–25 years (M = 22; 29 females) and 48 L2 Chinese readers aged 18–30 years (M = 24; 29 females) from Tianjin Normal University and Fujian Normal University. The L1 group were native Mandarin speakers with some English knowledge from schooling. No participants took part in Experiment 1. L2 readers were native users of diverse alphabetic languages differing in word-spacing conventions (Supplementary Table S5). They had studied Chinese for 2 years minimum (M = 6.4, SD = 3.7), having passed HSK-4 on average at level 5 (SD = 1), with total M = 233/300 (SD = 31) and reading M = 76/100 (SD = 14). The L2 group also completed the LHQ, confirming moderate to high proficiency (overall M = 0.58, SD = 0.10; reading M = 4, SD = 0.8).

As in Experiment 1, we employed two strategies to ensure adequate power. First, participant and stimulus numbers were increased relative to previous work, raising observations per condition from 400 (Zhou & Li, 2021) to 960. Second, a priori power analyses were conducted using data from a 30-participant pilot later included in the final sample. Full details are provided in the Supplemental File (Section S1). Results confirmed sufficient power to detect key effects observed previously.

Stimuli and design. Eighty sentence sets were used in Experiment 2, each containing an incremental word target (see Table 2). Following Zhou and Li (2021), the preceding verb was manipulated to yield four experimental conditions. In two, the target word was plausible, and its embedded word was either plausible or implausible depending. In the remaining two, the target was implausible while its embedded word varied in plausibility. Detailed stimulus characteristics are reported in the Supplemental File (Section S7).

Sentences were distributed across four presentation lists. Each began with four practice sentences and included 80 experimental and 20 plausible filler sentences. Each list contained 20 items per condition, assigned via a Latin square so every sentence appeared once per list with balanced condition numbers Sentences were presented in random order, with equal numbers of L1 and L2 participants pseudo-randomly assigned to each list.

Apparatus and procedure. This was the same as Experiment 1.

3.2. Results

Accuracy on comprehension questions was high for all participants (>80%; L1: M = 94.6%, L2: M = 86.9%, p < .001), though significantly higher for L1 readers, confirming that both groups understood the sentences.

Data analysis procedures. Following standard procedures, fixations shorter than 80 ms or longer than 1200 ms were removed (3.8% for L1 and 2.6% for L2). Trials with fewer than five fixations or track loss were excluded (0.7% L1; 0.3% L2). The remaining data were analyzed with LMEMs in R using the lme4 package, with p values from lmerTest (Kuznetsova et al., 2014).

Sentence-level analyses. We conducted sentence-level analyses of eye movements for sentences containing plausible or implausible incremental targets to assess reading of globally plausible versus implausible sentences. L2 readers had longer reading times, and made more and longer fixations, shorter saccades and more regressions. Consistent with Experiment 1, these findings indicate greater reading difficulty for L2 readers. Reading times were also longer for globally implausible than plausible sentences, with this effect larger for L2 readers. Full details appear in the Supplemental File (Section S3).

Word-level analyses. As in Experiment 1, we report word-level eye-movement measures for two regions of interest: the target region (three-character incremental word) and the pretarget verb region. Effects at the target region indicate fixational processing of incremental words, whereas pretarget effects reflect parafoveal processing or re-reading.

We used linear mixed effects models (LMEMs) to test theoretically motivated hypotheses, with a single condition variable as a fixed effect and customized contrasts (Schad et al., 2020) assessing: (1) main effects of incremental word plausibility; (2) embedded word plausibility for plausible incremental words and (3) embedded word plausibility for implausible incremental words. These were reported separately for L1 and L2 groups, followed by contrasts including group to compare L1 versus L2.

We distinguished between hypothesis-driven and exploratory analyses to limit Type I error risk from multiple measures. For hypothesis-driven tests, we evaluated robustness under conservative criteria, applying Bonferroni correction and the convergence-based heuristic of von der Malsburg and Angele (2017). As in Experiment 1, these criteria assessed reliability but were not treated as strict cutoffs.

Continuous variables (fixation times) were log-transformed (Wagenmakers et al., 2012). Contrasts were defined using the contr.sdif function in MASS (Venables & Ripley, 2002), yielding effect coding equivalent to contr.sum. We used deviation coding (−1/1), though sum-to-zero coding (−0.5/0.5) produced identical results. We employed maximal random-effects structures where possible (Barr et al., 2013), including participants and stimuli as crossed effects. If models failed to converge, random structures were simplified stepwise, first removing correlations, then slopes for stimuli and participants, until convergence was achieved. The Supplemental File (Section S2) lists the final models used. Tables 4 and 5 report mean eye movements and statistical results; Figures 2 and 3 illustrate key effects at target and pretarget regions.

Table 4. Mean eye movements for pretarget and target regions in Experiment 2

Note. The standard error of the mean is shown in parentheses. Fixation time measures are shown in ms.

Table 5. Summary of statistical effects for Experiment 2

Note. Asterisks indicate p < .05. FFD = first-fixation duration, GD = gaze duration, TRT = total reading time, RI = probability of a regression-in.

Figure 2. Embedded word plausibility effects for plausible incremental target words in Experiment 2, in (A) gaze durations for the pretarget region and (B) total reading times for the target region.

Figure 3. Embedded word plausibility effects for implausible incremental target words in Experiment 2, in (A) gaze durations for the pretarget region and (b) total reading times for the target region.

Target region analyses. Hypothesis-driven contrasts testing main effects of incremental word plausibility showed longer total reading times for implausible than plausible incremental words for both groups (L1: 679 vs. 558 ms; L2: 1616 vs. 1533 ms), and a higher probability of regressions-in (L1: .40 vs. .33; L2: .34 vs. .30). Using three measures (gaze duration, total reading time, regressions-in) and Bonferroni correction (p < .017), all effects were significant (p < .01) and appeared in at least two measures, confirming robustness. These results replicate Zhou and Li (2021), confirming the effectiveness of the incremental word plausibility manipulation. However, as effects emerged only in late measures, readers likely detected implausibility during re-reading rather than first-pass processing. We also found a significant Group × Plausibility interaction in total reading time, driven by a larger L1 effect, indicating greater L1 sensitivity to word plausibility.

Contrasts of embedded word plausibility for plausible incremental words showed no significant effects. However, when incremental words were implausible, both groups showed embedded word plausibility effects in total reading time (L1: 710 vs. 648 ms; L2: 1649 vs. 1584 ms). These findings indicate that both groups considered the alternative embedded word analysis when the preferred whole-word interpretation was implausible. Effects exceeded the Bonferroni-corrected threshold (p < .017) and appeared in multiple measures, suggesting robustness.

In sum, both L1 and L2 readers not only showed plausibility effects consistent with whole-word processing but also activated embedded word meanings when the whole-word interpretation failed. These effects appeared late during processing, in total reading times, with larger L1 effects suggesting higher L1 sensitivity to contextual plausibility.

Pretarget region analyses. As in Experiment 1, statistical models for the pretarget region included each verb’s log-transformed and centered lexical frequency, first- and second-character frequencies and stroke counts as covariates. Including or excluding these did not change fixed effect patterns. Moreover, no group differences appeared for these covariates, unsurprising given that verbs were closely matched across groups.

Analyses revealed main effects of incremental word plausibility in total reading times and regressions-in for both groups. Verbs preceding implausible incremental words elicited longer total reading times and more regressions-in (L1: 663 vs. 487 ms; L2: 1249 vs. 1087 ms; L1 regressions-in: .50 vs. .31; L2: .33 vs. .28). These effects suggest both groups reinspected the pretarget verb when encountering an implausible incremental word.

An L2-only gaze-duration effect also was observed (704 vs. 661 ms) but was absent for L1 (331 vs. 318 ms), indicating L2 parafoveal detection of upcoming incremental word plausibility. L2 gaze durations were generally longer, allowing more parafoveal processing time. These effects exceeded the Bonferroni-corrected threshold (p < .017) and appeared in multiple measures, confirming robustness.

When incremental words were plausible, embedded word plausibility produced no L1 effects but yielded L2 effects in gaze duration (680 vs. 642 ms) and total time (1130 vs. 1045 ms). This indicated L2 sensitivity to embedded word plausibility even when the whole-word was appropriate, with the gaze-duration effect implying that this effect occurred during parafoveal processing. The L2 total-time effect exceeded p < .017; the gaze-duration effect met it exactly, indicating strong but partly ambiguous reliability.

When incremental words were implausible, L1 readers showed embedded word plausibility effects in gaze duration (350 vs. 312 ms), but L2 readers did not (703 vs. 704 ms). Both groups, however, showed effects in total reading time (L1: 726 vs. 559 ms; L2: 1303 vs. 1196 ms) and regressions-in (L1: .58 vs. .42; L2: .35 vs. .30). These results imply increased re-reading for implausible embedded and whole-word interpretations, with only L1 readers showing parafoveal sensitivity. Additionally, a significant Group × Embedded Word Plausibility interaction in total time and regressions-in reflected greater L1 reanalysis effort. The L1 effects exceeded p < .017 and appeared across all three measures. For L2 readers, only total-time effects met this criterion. Thus, L1 readers showed both parafoveal and re-reading effects, whereas L2 readers showed only later re-reading effects.

In summary, L1 gaze-duration effects of embedded word plausibility appeared only when the whole-word interpretation was implausible, whereas L2 readers showed little comparable parafoveal processing. Both groups showed embedded word effects in re-reading, but L1 effects occurred only when the whole-word was implausible, whereas L2 effects appeared regardless of plausibility. Overall, these findings suggest more limited L2 parafoveal processing and greater difficulty using context to resolve segmental ambiguity.

3.3. Discussion

Experiment 2 compared L1 and L2 readers’ processing of sentences containing a three-character incremental word, manipulating the plausibility of both whole-word and embedded word interpretations. L2 readers showed slower and more effortful reading, with longer fixations, shorter saccades and more regressions.

At the target region, both groups showed incremental word plausibility effects. For L1 readers, the effect emerged in early (gaze duration) and late (regressions-in, total-time) measures, indicating sensitivity to word plausibility during lexical processing and integration. For L2 readers, effects appeared only later, suggesting delayed or less efficient contextual integration. No embedded word plausibility effects appeared in first-pass reading. However, L1 readers showed late effects when the incremental word was implausible, implying reanalysis when the preferred interpretation conflicted with context. This pattern indicates lexical access to the embedded word during reanalysis, matching Experiment 1 evidence that embedded words influence processing when context or preview supports activation.

Findings from the pretarget region also showed sensitivity to word plausibility. Both groups had longer total times and more regressions-in for implausible parafoveal words, reflecting integration difficulty. Notably, L2 readers additionally showed a gaze-duration effect, suggesting parafoveal plausibility processing, which may have been enabled by these readers making generally longer fixations. Group differences were clearer for embedded word plausibility. L2 readers showed reliable effects even when the incremental word was plausible, including a gaze-duration effect implying parafoveal access to embedded meaning (again potentially supported by their generally longer fixations). In contrast, L1 readers showed such effects only when the whole word was implausible, consistent with context-driven reanalysis. Under these conditions, L1 readers displayed both early (gaze duration) and late (total-time, regressions-in) effects, indicating rapid detection of ambiguity and flexible access to the embedded word when context disfavored the whole-word meaning. This parallels Experiment 1, where L1 readers showed parafoveal sensitivity to embedded word plausibility before direct fixation. Together, these findings support dynamic L1 engagement with embedded meanings, which may be activated early, following parafoveal preview, and later when integration fails.

Overall, both L1 and L2 readers were sensitive to contextual plausibility and re-read when expectations were violated. However, L1 readers showed greater flexibility and efficiency in accessing embedded meanings, especially when context contradicted whole-word interpretations. In contrast, L2 readers appeared to maintain multiple candidate segmentations longer, suggesting weaker suppression of implausible alternatives.

4. General discussion

Two eye-movement experiments investigated L1 and L2 processing of segmental ambiguities in Chinese reading involving incremental three-character words (e.g., 体育馆, “stadium”) whose first two characters form an embedded word (e.g., 体育, “sport”). In Experiment 1, experimental sentences contained a plausible incremental word, while the plausibility of its embedded word was manipulated; with these sentences compared with control sentences where the incremental word was replaced by its embedded word. This allowed us to investigate whether L1 and L2 readers show similar plausibility effects, by making longer fixations for implausible control target words, and perhaps also implausible embedded words, across conditions. Experiment 2 further investigated these effects by manipulating the plausibility of whole and embedded words independently. Prior work (Zhou & Li, 2021) showed holistic L1 processing of segmental ambiguity, observing control word plausibility effects in their Experiment 1 and whole-word plausibility effects in their Experiment 2, but an absence of embedded word plausibility effects in either experiment. They took these findings to support the parallel activation account of segmental processing implemented in the Chinese Reading Model (CRM; Li & Pollatsek, 2020).

Our goal was to establish whether these effects could be replicated for L1 readers and whether high-proficiency L2 readers adopt comparable segmentation strategies. L2 reading was slower in both experiments, consistent with prior evidence of slower L2 reading without word boundaries (Bai et al., 2012; Bassetti & Lu, 2016; Cui, 2023; Ma et al., 2019; Shen et al., 2012; Yu, 2022; Zhou et al., 2018, 2020). Following Zhou and Li, we reported effects for both a region containing the target word and a pretarget region comprising the verb used to manipulate contextual plausibility.

4.1. Target region effects

Results from both experiments indicate that L1 and L2 readers favor holistic interpretations of segmentally ambiguous words. In Experiment 1, plausibility effects were observed for the control target words, especially among L1 readers, but not for the embedded words in incremental targets, as would be expected if these were segmented first during sentence processing. These findings replicated Zhou and Li (2021) and so supported the CRM’s assumption that whole-word interpretations dominate when context allows. In Experiment 2, embedded word plausibility effects emerged only when the whole-word interpretation was implausible, particularly for L1 readers, suggesting context-driven reanalysis via embedded word access. Thus, while both groups appear to rely primarily on the holistic processing of segmental ambiguity, it also appears that they can lexically access the embedded words during reanalysis.

Both groups processed sentences incrementally, integrating input with context and showing difficulty when this input conflicted with contextual expectations (Lee & Witzel, 2023; Williams, 2006). Word plausibility effects were weaker for L2 readers, suggesting less efficient integration despite these readers’ high language proficiency. Nonetheless, the L2 readers used segmentation strategies broadly similar to L1 readers, demonstrating native-like resolution of segmental ambiguities.

This research is among the first to use eye movements to examine L2 processing of segmental ambiguity (see also Yao et al., 2025). Crucially, the evidence suggests that both L1 and L2 readers employ holistic segmentation strategies but can reanalyze ambiguities flexibly when this is required by context, supporting the view that high-proficiency L2 readers can achieve near-native word segmentation. Future work should explore how these strategies develop across proficiency levels.

One possibility is that L2 readers draw on their L1 experience. Previous studies show cross-linguistic influences in sentence processing, such as orthographic overlap (Libben & Titone, 2009) and syntactic transfer (Bardovi-Harlig & Sprouse, 2018; Schwartz & Sprouse, 1996). In the Supplemental File, we reported analyses comparing L2 readers whose L1 employs interword spaces with those whose L1 does not, but found no subgroup differences. Similarly, analyses comparing effects in terms of HSK or self-rated language proficiency revealed no clear proficiency effects. These null results suggest that once high proficiency is achieved, segmentation strategies converge. However, further work is required as early-stage learners may still exhibit L1-driven segmentation differences.

4.2. Pretarget verb region effects

Analyses for the pretarget verb region enabled assessment of both parafoveal processing effects and re-reading effects. Parafoveal processing is critical for fluent reading, allowing extraction of information from upcoming words before fixating them (Cutter et al., 2015; Schotter et al., 2012). Prior research shows orthographic and phonological information can be accessed parafoveally. However, whether higher level semantic information, including contextual plausibility, can be parafoveally processed remains debated (for discussion, see Pan et al., 2024; Rayner et al., 2014; Schotter, 2013; Schotter et al., 2023). Some studies suggest that parafoveal processing of semantic information is possible in compact logographic scripts like Chinese (Yan et al., 2009; Yan & Sommer, 2015; Pan et al., 2016). Moreover, the studies additionally suggest that this semantic preview information can influence processing of the currently fixated word, that is, a parafoveal-on-foveal effect. However, such findings are controversial (see Li et al., 2014; Drieghe, 2011).

In Experiment 1, an L1 gaze-duration effect of embedded word plausibility suggested parafoveal semantic access during first-pass processing of the verb. A similar L2 trend did not reach significance. Both groups also showed total reading time effects, while L1 readers showed regressions-in effects. These late effects reflect re-reading of the verb after fixating the incremental word, indicating later lexical access to the embedded word interpretation during reanalysis. The regressions-in effect confirms that this lexical access triggered reanalysis. The findings therefore indicate that embedded words were lexically and semantically accessed during segmental ambiguity processing and that this influenced subsequent processing, despite the holistic analysis ultimately being strongly preferred.

In Experiment 2, both groups again showed embedded word plausibility effects at the pretarget region, but with distinct profiles. L2 readers showed early (gaze duration) and late (total-time) effects even when the incremental word was plausible, reflecting parafoveal access, which may have been facilitated by long L2 fixation times on the pretarget verb. In contrast, L1 readers showed such effects only when the incremental word was implausible, with both early and late measures affected. This pattern suggests that L1 readers selectively accessed embedded word candidates when the whole-word analysis conflicted with context, whereas L2 readers maintained multiple interpretations even without conflict. Together, these results confirm parafoveal access to embedded word meanings in both groups, which may differ with how efficiently context guides segmentation and reanalysis.

These group-level differences raise broader questions about L2 parafoveal processing. Both L1 and L2 readers accessed embedded word meanings, including parafoveally. High-proficiency L2 readers therefore appear capable of semantic parafoveal processing, possibly supported by their longer first-pass fixations at the verb. These results nevertheless align with findings showing proficiency in L2 readers parafoveal processing (Tiffin-Richards, 2024a,b), and contrast with other reports of more limited L2 parafoveal processing in Chinese reading (Cong & Chen, 2022; Xiao et al., 2021). This may reflect variation in L2 proficiency across studies, and future research should assess how parafoveal processing develops with L2 proficiency.

It is also possible that some parafoveal effects reflect mislocated fixations, which occur when eye movements either undershoot or overshoot the intended target word (Drieghe, 2011; Drieghe et al., 2008). In our data, some fixations on the pretarget verb may have occurred while readers were processing initial characters of the incremental word. This could mimic parafoveal preview while actually reflecting foveal processing. Crucially, as Chinese lacks interword spaces, readers are less likely to use word boundaries to guide eye movements, so that readers are less likely to fixate optimally on words and likely to fixate a broad range of fixation locations (see Li et al., 2009; Yan et al., 2010). As a result, characters from upcoming words sometimes may fall within foveal vision, with the potential to create apparent parafoveal-on-foveal effects (Yin et al., 2025). Such effects may be especially common among L2 readers, who exhibit less precise saccade-targeting and greater fixation variability (Gnetov & Kuperman, 2024). While this does not undermine the present evidence for embedded word activation, it suggests such activation may sometimes arise from foveal rather than parafoveal processing. Nevertheless, whether through parafoveal preview or mislocated fixation, our results clearly show embedded word meanings are accessed during segmental processing.

4.3. Implications for the Chinese reading model

The present findings have implications for the CRM. As embedded word meanings appear to be accessed during both early and late processing, we propose that the model should be refined to include a mechanism for transient, context-sensitive activation of embedded constituents. This may arise from parafoveal preview, mislocated fixation or variability in fixation location. As noted earlier, fixation position during word recognition is likely more variable in Chinese than in alphabetic scripts, due to the absence or reduction of word-based eye-guidance effects (Li et al., 2011; Yan et al., 2010; see also Li et al., 2018). As a result, readers sometimes may fixate near the end of the preceding word or within different parts of an incremental word, bringing its embedded portion into foveal or parafoveal vision depending on fixation location. These spatial factors likely influence when and whether embedded word meanings are activated during early processing and this influence of fixation location should be investigated further.

However, early activation alone cannot explain the late embedded word plausibility effects we observed. These suggest that embedded meanings can also be reactivated during postlexical reanalysis, particularly when a whole-word interpretation is implausible. We therefore propose a dual-route CRM: one route supporting early, fixation dependent activation of embedded words and another supporting later, context-driven reactivation. This dual-route approach may also clarify some L1–L2 reading differences, as between-group variability in eye-movement control, parafoveal efficiency and contextual use could modulate both early and late access to embedded words meanings.

4.3. Conclusions

In conclusion, our findings show that both L1 and high-proficiency L2 readers resolve segmental ambiguity in Chinese using broadly similar processing strategies. We also provide novel evidence that embedded word meanings are accessed both parafoveally and during reanalysis by both groups. While the underlying mechanisms appear similar, the timing and strength of effects suggest that L2 readers engage in less efficient or less selective word segmentation processing. Future research should examine how word segmentation mechanisms and parafoveal processing develop during L2 Chinese learning and are shaped by proficiency and reading experience.