1. Introduction
For those learning a second language (L2), certain grammatical forms of the target language often present consistent challenges, regardless of first language (L1) background. Functional inflectional morphemes have received particular attention as a bottleneck in L2 acquisition (Slabakova, Reference Slabakova2008, Reference Slabakova, Ionin and Rispoli2019). Among functional morphemes, the third-person singular present -s (henceforth, 3SG -s) has been identified as one of the most challenging morphemes, with persistent non-target-like performanceFootnote 1 being observed even among advanced-level L2 learners (e.g., Lardiere, Reference Lardiere1998).
Previous L2 research has attributed the difficulty of 3SG -s to a range of factors, from representational deficits – namely, underspecified abstract morphosyntactic knowledge (Li & Yang, Reference Li and Yang2022; see Hawkins & Liszka, Reference Hawkins, Liszka, van Hout, Hulk, Kuiken and Towell2003 for a foundational account) – to processing-related constraints. These include computational difficulties in mapping morphosyntactic features onto their morphological realization (Haznedar & Schwartz, Reference Haznedar, Schwartz, Hughes, Hughes and Greenhill1997; Prévost & White, Reference Prévost and White2000), phonological interference from the learner’s L1 (Goad et al., Reference Goad, White and Steele2003), and processing biases that deprioritize functionally redundant morphology such as 3SG -s (Slabakova, Reference Slabakova, Ionin and Rispoli2019; VanPatten, Reference VanPatten, VanPatten, Keating and Wulff2020). The current study, in line with a processing-based account, examines whether sentence-position-related perceptual and processing factors contribute to L2 learners’ difficulty with this morpheme.
As has been widely discussed in the literature (Barcroft & VanPatten, Reference Barcroft, VanPatten, Pérez-Leroux and Glass1997; Carroll, Reference Carroll, O’Brien, Shea and Archibald2006; VanPatten, Reference VanPatten, VanPatten, Keating and Wulff2020), sentence position can affect the processing of grammatical morphemes, with potential consequences for their perceptual salience. In particular, previous research has reported reduced sensitivity to target morphemes when they occur in sentence-medial position. Sundara et al. (Reference Sundara, Demuth and Kuhl2011) examined monolingual English-acquiring children aged between 22 and 27 months, investigating how sentence position influences both the perception and production of 3SG -s in early L1 acquisition. Although their study did not involve L2 learners, their experimental paradigm provides an important baseline for assessing whether comparable perceptual asymmetries emerge in adult L2 processing. Building on and adapting the design of Sundara et al. (Reference Sundara, Demuth and Kuhl2011), the current study examines how sentence-position-related perceptual constraints affect L2 processing, thereby clarifying the role of real-time processing limitations in the bottleneck in L2 acquisition.
2. Literature review
2.1. 3SG -s as a bottleneck in L2 acquisition
Typically developing L1-English children acquire 3SG -s before the age of 4. Brown’s (Reference Brown1973) longitudinal study of three children showed that they produced 3SG -s in 90% of obligatory contexts between 26 and 46 months. Two- and three-year-old children often fail to produce 3SG -s in obligatory contexts; however, they seldom produce it when it is not called for (Lukyanenko & Fisher, Reference Lukyanenko, Fisher, Orman and Valleau2014). Unlike in L1 acquisition research, where 3SG -s is considered relatively unproblematic, in L2 acquisition research it has long been recognized as challenging (e.g., Ionin & Wexler, Reference Ionin and Wexler2002; Jensen et al., Reference Jensen, Slabakova, Westergaard and Lundquist2020; Lardiere, Reference Lardiere1998). A striking example comes from Lardiere’s (Reference Lardiere1998) examination of Patty, an adult Chinese L2-English speaker who had lived in the USA for about 20 years after arriving at the age of 22. In spontaneous speech, Patty produced 3SG -s only 4.5% of the time in obligatory contexts. Ionin and Wexler (Reference Ionin and Wexler2002) also report similar non-target-like performance with 3SG -s among 20 L1-Russian L2-English children (ages 3;9–13;10). In a similar vein, L2 morpheme acquisition studies consistently indicate that 3SG -s is more difficult to acquire than other functional morphemes. Specifically, both adult and child L2 learner data indicate that 3SG -s is among the least accurately produced morphemes in L2 English (see Goldschneider & DeKeyser, Reference Goldschneider and DeKeyser2001, for a meta-analysis; Murakami & Alexopoulou, Reference Murakami and Alexopoulou2016, for a large-scale learner-corpus study). Collectively, previous L2 studies highlight the difficulty of 3SG -s in L2 English acquisition.
2.2. Explaining non-target-like L2 performance on 3SG -s
There are two main approaches to explaining L2 learners’ non-target-like performance with grammatical morphemes: representational-deficit accounts and processing-based accounts. Representational deficit accounts attribute such performance to underspecified morphosyntactic knowledge (Hawkins & Liszka, Reference Hawkins, Liszka, van Hout, Hulk, Kuiken and Towell2003; Tsimpli & Dimitrakopoulou, Reference Tsimpli and Dimitrakopoulou2007). We use representational deficit in this narrower, morphosyntactic sense throughout. For example, Hawkins and Liszka (Reference Hawkins, Liszka, van Hout, Hulk, Kuiken and Towell2003) argued that L1-Chinese speakers’ difficulty in producing English past tense -ed in obligatory contexts reflects a failure to acquire the abstract morphosyntactic features associated with tense. In contrast, processing-based accounts hold that non-target-like performance arises, at least in part, from limitations in real-time processing. On this view, L2 learners’ variable production does not necessarily reflect impaired morphosyntactic representations but instead results from difficulties in accessing and mapping morphosyntactic knowledge onto surface forms, suggesting that learners may possess more underlying knowledge than their overt performance reveals (e.g., the Missing Surface Inflection Hypothesis; Haznedar & Schwartz, Reference Haznedar, Schwartz, Hughes, Hughes and Greenhill1997; Prévost & White, Reference Prévost and White2000).
Previous L2 research on non-target-like 3SG -s performance has been divided between representational deficit and processing-based accounts (e.g., Goad et al., Reference Goad, White and Steele2003; Ionin & Wexler, Reference Ionin and Wexler2002; Lardiere, Reference Lardiere1998; Li & Yang, Reference Li and Yang2022). For example, Li and Yang (Reference Li and Yang2022) found that L1-Chinese participants produced 3SG -s only 29.4% of the time in spoken tasks. No significant differences were detected in 3SG -s performance across IELTS proficiency levels, suggesting that higher proficiency does not necessarily eliminate non-target-like 3SG -s performance. In an argument consistent with the representational deficit account, Li and Yang (Reference Li and Yang2022) claimed that their L1-Chinese learners struggled to encode abstract morphosyntactic features through inflectional morphology.
Studies aligned with the processing-based accounts consider the effects of input processing and L1 phonology on L2 performance (Goad et al., Reference Goad, White and Steele2003; Slabakova, Reference Slabakova, Ionin and Rispoli2019; VanPatten, Reference VanPatten, VanPatten, Keating and Wulff2020). These accounts are grouped here because none attributes the difficulty to a deficient morphosyntactic representation; each instead locates the source of variability in how an intact morphosyntactic representation is realized in production or perceived in comprehension. VanPatten (Reference VanPatten, VanPatten, Keating and Wulff2020) states in his Preference for Nonredundancy Principle that redundant grammatical markers are more difficult to process than non-redundant ones (p. 109). In proposing this principle, VanPatten (Reference VanPatten, VanPatten, Keating and Wulff2020) cites 3SG -s as a prime example of a redundant marker. For example, in a sentence such as The duck swims, singular and third-person reference can be inferred from the subject noun phrase, while present tense is recoverable from the verb form itself. As a result, the 3SG -s morpheme contributes little additional semantic information, making it particularly vulnerable during real-time processing. Similarly, in her account of the Bottleneck Hypothesis, Slabakova (Reference Slabakova, Ionin and Rispoli2019) argues that functional redundancy contributes to non-target-like performance with 3SG -s, stating that ‘[t]he marker of S–V agreement is truly redundant in English from a functional perspective; therefore, it may be dropped more often than more meaningful morphemes’ (p. 332).
Goad et al. (Reference Goad, White and Steele2003) propose a phonology-based account of inflectional variability in L2 English, arguing that the apparent omission of 3SG -s by L1-Mandarin speakers reflects constraints in mapping grammatical features onto phonological form, rather than deficits in morphosyntactic representations. Specifically, they argue that producing English inflectional consonants requires phonological structures that allow these suffixes to be integrated into a phonological word, structures that are not readily available in Mandarin. Consequently, even when agreement features are represented in the syntax, learners may fail to map these features onto phonological form, yielding non-target-like production that resembles inflectional deletion. This argument has been further developed in later work on the Prosodic Transfer Hypothesis (Goad & White, Reference Goad and White2006, Reference Goad and White2019), which holds that L2 learners initially rely on L1 prosodic representations in the realization of L2 morphology, leading to persistent variability at the phonological interface despite otherwise intact grammatical knowledge. According to this hypothesis, although the segment /s/ is perceptually salient given its high spectral center of gravity and amplitude in the noise portion (Goad & White, Reference Goad and White2019), difficulties with 3SG -s in L2 English are attributed primarily to its prosodically weak representation rather than to purely perceptual limitations.
From this perspective, similar phonology-based constraints may be expected for Korean learners of English. In Korean, obstruents in coda position undergo neutralization, such that /s/ is not phonetically realized as [s] word-finally but instead surfaces as [t] (Jun, Reference Jun1993). Given this L1 background, English word-final -s may pose challenges for the perception and encoding of word-final [s, z] in English, consistent with evidence that L1 phonological categories and coda constraints shape learners’ perception and production of L2 consonants in word-final position (Flege & Wang, Reference Flege and Wang1989). Within phonology-based accounts, difficulties with 3SG -s in L1-Korean learners’ L2 English can thus be interpreted as arising from L1-specific constraints on the phonological realization of inflectional endings, rather than from impoverished morphosyntactic knowledge.
A related but partially distinct line of processing-based work attributes the vulnerability of 3SG -s more directly to perceptual salience. Goldschneider and DeKeyser (Reference Goldschneider and DeKeyser2001) examined the role of perceptual salience – defined as ‘how easy it is to hear or perceive a given structure’ (pp. 22–23) – across a range of inflectional morphemes. They operationalized perceptual salience as a composite measure consisting of (i) the number of phones in the functor (phonetic substance), (ii) the presence or absence of a vowel in the surface form (syllabicity), and (iii) the total relative sonority of the functor (see Kim et al., Reference Kim, O’Grady and Schwartz2018; Sagarra et al., Reference Sagarra, Sánchez and Bel2019, for a similar view). On the basis of these metrics, they argued that bound morphemes such as 3SG -s are inherently less perceptually salient than free morphemes, a property that may contribute to their comparatively slower development in L2 acquisition. In a related vein, Johnson (Reference Johnson1992) reported that L2 learners performed more accurately on grammaticality judgment tasks in the written modality than on comparable tasks in the auditory modality used in Johnson and Newport (Reference Johnson and Newport1989). Notably, morphology-related items were particularly susceptible to modality effects. Johnson (Reference Johnson1992) further argued that ‘processing or phonological properties must have contributed to adult learners’ poor performance on these structures in the auditory version’ (p. 236). Taken together, these findings suggest that non-target-like performance with inflectional morphology may, at least in part, reflect perceptual and phonological constraints in online auditory processing.
Two studies that are particularly relevant to the perceptual accessibility of 3SG -s are Johnson et al. (Reference Johnson, de Villiers and Seymour2005) and Sundara et al. (Reference Sundara, Demuth and Kuhl2011). Johnson et al. (Reference Johnson, de Villiers and Seymour2005) involved a comprehension task in which children aged 3–6 needed to perceive whether 3SG -s was present or absent on the verb in order to select the image with the correct number of animals. For instance, in sentences like The duck swims in the water (singular) and The ducks swim in the water (plural), accurately perceiving the presence/absence of 3SG -s is necessary for determining whether the subject is singular or plural. Crucially, the stimuli masked plurality on the noun by using verbs that began with an /s/ consonant cluster, such that the plural /s/ on the noun is coarticulated with the initial /s/ on the verb in ordinary speech. Notably, while children aged 3–4 struggled with the comprehension task, it has been shown by other researchers that children of a similar age reliably produce the 3SG -s in obligatory contexts (e.g., Brown, Reference Brown1973), which suggests that the challenge at this stage lies not in acquiring the morpheme, but in perceiving and processing it in real time.
The second study, Sundara et al. (Reference Sundara, Demuth and Kuhl2011), reported a sentence-position effect in both the production and perception of 3SG -s. In the production task, monolingual English-speaking children listened to and repeated sentences in which 3SG -s appeared either in sentence-medial position (e.g., He cries now) or in sentence-final position (e.g., There he cries). Both 22- and 27-month-olds were more accurate in producing 3SG -s in sentence-final than in sentence-medial position. This difference was attributed to perceptual salience associated with phonetic prominence, as the mean duration of 3SG -s in the stimuli was 125 ms in medial position compared to 202 ms in final position. In the perception task, the authors examined whether children’s listening behavior was sensitive to the presence or absence of 3SG -s in obligatory contexts. The results showed that both groups performed at chance in the sentence-medial condition but not in the sentence-final condition: 22-month-olds listened longer to grammatical sentences containing 3SG -s, whereas 27-month-olds showed longer listening times for ungrammatical sentences without the morpheme. Taken together, these findings suggest that perceptual factors – potentially related to phonetic prominence – play an important role in early patterns of 3SG -s production in English.
2.3. The present study and research questions
To review, previous research has documented persistent non-target-like performance with 3SG -s in L2 acquisition and has proposed a range of explanations for this difficulty, spanning representational deficit accounts and processing-based accounts. Research on L1 acquisition further shows that sensitivity to 3SG -s is systematically modulated by perceptual factors. In particular, studies with English-speaking children demonstrate that 3SG -s is more difficult to perceive and produce in sentence-medial position than in sentence-final position, where it is typically more perceptually prominent. However, despite extensive discussion of phonological constraints and input processing factors in L2 acquisition, to the best of our knowledge, no previous L2 study has directly examined whether, and to what extent, sentence-position-related perceptual constraints affect the processing of 3SG -s.
The current study addresses this research gap by examining the real-time processing of 3SG -s in an L2 context, testing whether processing difficulty is greater in auditory comprehension when the morpheme appears in sentence-medial than in sentence-final position. The study focuses on L1-Korean learners, given that Korean does not permit [s, z] in syllable-final (coda) position, which may influence the perception and processing of word-final fricatives in English. In addition, unlike learners in English-as-a-Second-Language contexts, who are immersed in the target language, learners in English-as-a-Foreign-Language (EFL) contexts such as those in Korea receive substantially less spoken input outside the classroom. As a result, the participants in the present study may be particularly sensitive to challenges in the auditory perception and real-time processing of the 3SG -s morpheme.
Using the visual-world eye-tracking paradigm, the present study adapts the experimental designs of Johnson et al. (Reference Johnson, de Villiers and Seymour2005) and Sundara et al. (Reference Sundara, Demuth and Kuhl2011) to an adult L2 population. Rather than relying on the preferential listening paradigm developed for very young children, we assess sensitivity to 3SG -s by combining eye-movement measures, which index the speed with which participants orient to the target picture, with accuracy measures derived from mouse-click responses. In addition, the study compares participants’ performance in an online processing task with their performance in an offline knowledge task. This comparison allows us to examine whether non-target-like performance with 3SG -s is better explained by representational deficits or by processing limitations, as reflected in a dissociation between online processing and offline knowledge. Based on the preceding discussion, the present study addresses the following research questions:
RQ1: Does sentence position affect the real-time processing of 3SG -s in L2 learners, such that processing is more difficult in sentence-medial position than in sentence-final position?
RQ2: Do the effects observed across the online auditory task and the offline written task support a morphosyntactic representational-deficit account or a processing-based account of non-target-like performance in L2 acquisition?
3. Methods
3.1. Participants
The final sample for this study included 24 self-identified L1-English speakers (7 males and 17 females) and 23 L1-Korean L2 learners of English (8 males and 15 females).Footnote 2 These participants were all recruited and tested at a large university located in Seoul, Korea, which has a significant international student population. The L1 group had a mean age of 22.17 years (SD = 2.62), and the L2 group had a mean age of 24.83 years (SD = 4.46). In terms of proficiency, the L1 participants scored an average of 38.92 (SD = 5.93) out of 50 on the English cloze test used in the study (Brown, Reference Brown1980), and the L2 participants scored an average of 31.65 (SD = 9.29) out of 50. A statistical analysis revealed that the L1 group had a significantly higher score than the L2 group (t = 3.18, p = .002, Cohen’s d = 0.93). Based on their cloze test scores, we considered the L2 participants to be at a high-intermediate to advanced level of English proficiency. The mean age at which participants in the L2 group started learning English was 6.91 years (SD = 1.62).
3.2. Materials
All test materials are available at https://osf.io/jvgq4.
3.2.1. Visual-world eye-tracking task
The visual-world eye-tracking task focused on the real-time processing of 3SG -s in sentence-medial versus sentence-final position. This task comprised 56 audio stimuli, including 16 critical sentences and 40 filler sentences (see below). The sentences were presented in isolation, without preceding discourse context. Participants listened to each sentence individually while viewing the visual display, in order to minimize contextual predictability and help them focus on online processing of the target morphosyntactic cue. A 2 × 2 Latin-square design was used for the critical sentences, with verb position (henceforth, Position; Medial versus Final) and subject number (henceforth, Number; Singular versus Plural) as the two factors. We modeled our design on Sundara et al. (Reference Sundara, Demuth and Kuhl2011), who employed a comparable manipulation of sentence-medial versus sentence-final position. This resulted in four distinct experimental conditions, as outlined in (1a–d). In the Medial-Singular condition, a verb marked with 3SG -s appears in the middle of the sentence, following a singular subject. In the Final-Singular condition, a verb marked with 3SG -s appears at the end of the sentence, also following a singular subject. In the Medial-Plural condition, a verb appears in a sentence-medial position, but without 3SG -s because the subject is plural. In the Final-Plural condition, a verb appears in a sentence-final position, again without 3SG -s due to the presence of a plural subject. The two plural conditions served as control conditions, allowing us to isolate effects of sentence position on the perception of 3SG -s. Accordingly, our primary analyses focus on the singular conditions, in which the target morpheme is present.
(1)
a.
Medial-Singular: e.g., The duck swims in the water.
b.
Final-Singular: e.g., In the water, the duck swims.
c.
Medial-Plural: e.g., The ducks swim in the water.
d.
Final-Plural: e.g., In the water, the ducks swim.
Crucially, following Johnson et al. (Reference Johnson, de Villiers and Seymour2005), we masked plurality on the noun by using verbs with an initial sibilant, most commonly [s]. One item, however, used [ʃ] (the stars shine in the sky). Because both [s] and [ʃ] are sibilants and thus phonetically similar, we expected this masking strategy would be effective for all participants, not only L1-Korean speakers, for whom [s] and [ʃ] are allophones of /s/. In addition, while some items yielded a perfect [s] – [s] sequence across the noun – verb boundary (e.g., ducks swim), others had a [z] – [s] sequence (e.g., birds sing). Although these cases did not involve an exact coda – onset match, they still involved adjacent sibilants, which we expected would provide sufficient masking of the plural marker.
All stimuli were recorded by a female native speaker of American English and checked for naturalness and clarity. An analysis of verb durations for experimental sentences using Praat (Boersma & Weenink, Reference Boersma and Weenink2017) showed that verbs in sentence-final position were approximately 22% longer (M = 864.70 ms; SD = 98.64 ms) than those in sentence-medial position (M = 708.29 ms; SD = 174.91 ms). As for the duration of 3SG -s, it was about 68% longer in sentence-final position (M = 237.69 ms, SD = 30.32 ms) than in sentence-medial position (M = 141.56 ms, SD = 26.92 ms), consistent with the results of Sundara et al. (Reference Sundara, Demuth and Kuhl2011). These results suggest that in the present study, sentence position should be treated as a composite cue involving both position and duration rather than position alone. Detailed information on the duration of 3SG -s in each critical sentence is outlined in Table 1.
Duration of 3SG -s in the Medial-Singular versus Final-Singular conditions
Table 1. Long description
The table consists of four columns: Sentence-medial 3 S G -s, Length in m s, Sentence-final 3 S G -s, and Length in m s.
* The duck swims in the water (143 m s) vs In the water, the duck swims (195 m s).
* The skunk stinks in the grass (157 m s) vs In the grass, the skunk stinks (262 m s).
* The cat sleeps in the bed (153 m s) vs In the bed, the cat sleeps (251 m s).
* The cop smokes in the street (129 m s) vs In the street, the cop smokes (286 m s).
* The ape swings in the tree (144 m s) vs In the tree, the ape swings (192 m s).
* The goat stands behind the fence (145 m s) vs Behind the fence, the goat stands (268 m s).
* The star shines in the sky (84 m s) vs In the sky, the star shines (214 m s).
* The bird sings in the tree (183 m s) vs In the tree, the bird sings (223 m s).
* The horse sprints across the field (169 m s) vs Across the field, the horse sprints (268 m s).
* The tourist sightsees around in the city (101 m s) vs Around in the city, the tourist sightsees (242 m s).
* The student studies in the library (127 m s) vs In the library, the student studies (254 m s).
* The ship sails across the ocean (151 m s) vs Across the ocean, the ship sails (206 m s).
* The guest socializes at the party (176 m s) vs At the party, the guest socializes (206 m s).
* The teenager skateboards in the park (123 m s) vs In the park, the teenager skateboards (218 m s).
* The dancer spins on the stage (117 m s) vs On the stage, the dancer spins (242 m s).
* The balloon soars into the air (163 m s) vs Into the air, the balloon soars (276 m s).
The 40 items that served as our fillers for this study featured a range of sentence types to introduce variation and were originally developed for separate, independent studies. There were 20 sentences targeting the verb’s semantic constraints and the frequency of the accompanying nouns (e.g., The {doctor/surgeon} {opened/moved} the blue box; see Hwang & Kim, Reference Hwang and Kim2026), and another 20 sentences containing there-expletive constructions (e.g., There {is/are/has been/have been} just one apple for three boys; see Hwang & Kim, Reference Hwang and Kim2025).
Stimuli were distributed across four running lists such that individual participants saw each item in only one of the four conditions. While listening to the sentences, participants were presented with two pictures on a computer screen, as shown in Figure 1. One of these pictures was the target response representing a correct interpretation of the sentence, and the other was a competitor representing an incorrect interpretation of the sentence. In the Medial-Singular condition (e.g., The duck swims in the water), the target response is the picture on the left-hand side of Figure 1 showing a scene with a singular subject. Participants were instructed to select the picture that matched the sentence they heard after the sentence finished playing. Throughout the task, an eye tracker recorded participants’ eye movements to capture their real-time processing during sentence comprehension. This procedure allowed for the analysis of both behavioral responses (picture selection accuracy) and online processing patterns (eye movements).
Example Medial-Singular trial in the visual-world eye-tracking task.
3.2.2. Fill-in-the-blank task
The purpose of the fill-in-the-blank task was to assess participants’ offline knowledge of both grammatical number and the lexical semantics of the verbs used in the eye-tracking task. In this task, participants were instructed to complete each sentence by selecting the most suitable option. The task had 16 critical items and 28 fillers. For critical items, a 2 × 2 Latin-square design was used to cross the factors Position (Medial versus Final) and Number (Singular versus Plural). Four response choices were provided below each sentence, as shown in (2). The expectation was that participants with a firm grasp on grammatical number would use verbs with 3SG -s (e.g., stinks) to complete sentences with singular subjects (e.g., skunk) and verbs without 3SG -s (e.g., stink) to complete sentences with plural subjects (e.g., skunks). We probed for knowledge of lexical semantics by providing response options with a semantically appropriate verb (e.g., stink/stinks) versus a semantically inappropriate verb (e.g., like/likes).
(2)
The {skunk/skunks} _______ in the grass. / In the grass, the {skunk/skunks} _______.
1. like
2. likes
3. stink
4. stinks
3.3. Procedure
Prior to the experiment, all participants completed an online cloze test (Brown, Reference Brown1980; see https://osf.io/jvgq4) at a location of their convenience. The test consisted of a written passage containing 50 gaps, which participants were required to fill with the appropriate missing words. This test was selected as an independent measure of L2 English proficiency due to its well-documented reliability and validity across diverse learner populations (Brown & Grüter, Reference Brown and Grüter2022). The test is widely regarded as an appropriate tool for measuring general language ability because it assesses not only grammatical accuracy but also discourse competence and lexical prediction. Consequently, even native speakers do not typically perform at ceiling on this measure (see Section 3.1).
About a week later, participants visited our eye-tracker lab. This second experimental session started with the visual-world eye-tracking task, which was created and administered using SR Research Experiment Builder. The task was conducted on a 15.6-inch computer screen with a resolution of 1,440 × 900 pixels. Participants’ eye movements were monitored using an EyeLink Portable Duo, which sampled data from the right eye at 1,000 Hz. To ensure precise eye-tracking, a 13-point calibration procedure was performed before the task began. For the main task, participants were given the following instructions in English: ‘In this task, we shall ask you to look at some pictures on a computer screen while listening to English sentences. After each sentence, you will click on one of the two pictures on the screen that best matches the sentence that you heard’.
Each trial began with a fixation cross appearing in the center of the screen for 1,000 ms. This was followed by a 3,000 ms silent preview of the visual scene, which allowed participants to preprocess the visual information in the scene and thereby reduced cognitive load (e.g., Kaan & Grüter, Reference Kaan, Grüter, Kaan and Grüter2021). After the preview, the stimulus sentence was played through the computer’s speaker, and then participants had to click on the image that matched their interpretation of the sentence.
Participants started with a practice session consisting of three trials, after which they proceeded to the main session. The presentation order was randomized to avoid consecutive trials in the same condition. The placement of the images was also counterbalanced across trials so that the target image appeared with equal frequency on the left and right sides of the screen.
The eye-tracking task was followed by the fill-in-the-blank task, a language background questionnaire, and a working-memory task, in that order. All tasks were hosted on PCIbex Farm (https://farm.pcibex.net/). Randomization was also applied to the sentences in the fill-in-the-blank task. Although all participants took part in the working-memory task, its outcomes are not within the scope of this study and will therefore not be discussed. The entire experiment lasted between 60 and 90 minutes.
3.4. Analysis
All statistical analyses for this study were conducted in R (R Core Team, 2025). The R analysis scripts along with the datasets are available at https://osf.io/jvgq4.
3.4.1. Fill-in-the-blank task
To investigate the effects of Group, Position, and Number on offline knowledge of grammatical number, we performed a logistic mixed-effects regression analysis on the accuracy data collected from the fill-in-the-blank task. The model was fit using the ‘lme4’ package (Bates et al., Reference Bates, Mächler, Bolker and Walker2015). The fixed effects were determined by the experimental design and included Group (L1 versus L2), Position (Medial versus Final), and Number (Singular versus Plural). These factors were contrast-coded such that L1, Medial, and Singular were set to −.5 and L2, Final, and Plural were set to .5. For the random-effects structure, we began with the maximal structure justified by the design and simplified it as necessary to obtain convergence, following Barr et al.’s (Reference Barr, Levy, Scheepers and Tily2013) recommendations concerning random-effects specification in mixed-effects modeling. Specifically, highly correlated random slopes were removed iteratively until the model converged. The final model included by-Participant and by-Item intercepts, as well as a by-Item slope for Number.
3.4.2. Eye-tracking task
The mouse-click response data were analyzed using a logistic mixed-effects model with the same fixed effects and contrast coding as the model used for the fill-in-the-blank task. For the random-effects structure, we again began with the maximal structure justified by the design and simplified it as necessary to obtain convergence. The final converged model included by-Participant intercepts and slopes for Position and by-Item intercepts and slopes for Group. To further examine the interaction effects, we performed post-hoc tests using the ‘emmeans’ package (Lenth, Reference Lenth2018).
For analysis of the eye-tracking data, we first removed trials with inaccurate mouse clicks from our dataset: 34 trials (8.85%) from L1 data and 90 trials (24.46%) from L2 data. Fixations for each 20-ms time bin were then transformed into logits to account for their bounded nature (e.g., Barr, Reference Barr2008). To improve data interpretation, we normalized the verb offset to 0 ms across all trials, and this served as the start of our time window of interest. Recall that the verbs in the critical stimuli always started with an [s] or [ʃ] sound and always immediately followed the subject noun (see Section 3.2.1). This design was meant to mask any morphological number information on the subject, thereby making the verb the earliest reliable cue to number. An independent statistical analysis (i.e., divergence point analysis; see below) confirmed that participants did not fixate the singular target picture before verb offset. We therefore deemed it appropriate to use the interval from 0 ms to 1,381 ms as the time window of interest for our analysis. The upper bound of 1,381 ms was determined by taking the average offset of all critical sentences in the Medial conditions (i.e., 1,181 ms) and adding an extra 200 ms buffer corresponding to the estimated latency of a typical eye movement.
To analyze the eye-tracking data within our time window of interest, we utilized a dual-method approach that drew on two statistical techniques to provide a more comprehensive understanding of the data. The first of these techniques, a growth curve analysis (e.g., Henry et al., Reference Henry, Jackson and Hopp2022), allowed us to investigate how our factors of interest interacted over time. The second technique, a divergence point analysis (Ito & Knoeferle, Reference Ito and Knoeferle2023; Stone et al., Reference Stone, Lago and Schad2021), helped us identify the specific time point at which participants’ gaze patterns toward the Target and Competitor began to exhibit a statistically reliable difference in each condition. Specifically, we employed a nonparametric bootstrapping method to derive a temporal confidence interval (CI) for the estimated onset of divergence. This bootstrap-based method is particularly well suited to handling the multiple-comparison and temporal-autocorrelation issues that arise in densely sampled time-course data (Stone et al., Reference Stone, Lago and Schad2021). By strategically combining these two analytical approaches, we were able to gain a more nuanced understanding of real-time processing patterns, uncovering both statistically reliable interaction effects and temporal changes, which would not have been possible if we had relied on a single method.
Our growth curve analysis model was constructed on the target advantage scores within the predefined time window. The target advantage scores were calculated by subtracting the logit-transformed fixations for the Competitor from those for the Target. The growth curve model included three key fixed effects: Group (L1 versus L2), Position (Medial versus Final), and Number (Singular versus Plural), which were contrast-coded (−.5 assigned to L1, Medial, and Singular; .5 assigned to L2, Final, and Plural). To model the complex and dynamic patterns in the data, we used second-order orthogonal polynomial terms. A linear term describes the overall trend in fixation rates, and a quadratic term captures more nuanced changes in this trend, such as changes in curvature. Model specification began with a maximal random-effects structure justified by the experimental design, but convergence problems required a reduction in model complexity, leading to a simplified random-effects structure. The final model retained the by-Participant and by-Item intercepts as well as the by-Participant slope for Number.
A growth curve model was initially estimated using a standard linear mixed-effects approach implemented with the ‘lme4’ package (Bates et al., Reference Bates, Mächler, Bolker and Walker2015). However, the residual diagnostics indicated significant departures from the expected Q-Q distribution and detected more outliers than expected under the fitted model, although the dispersion test did not indicate overdispersion. We therefore refitted the growth curve model using a robust mixed-effects regression in R with the ‘robustlmm’ package (Koller, Reference Koller2016), an approach designed to reduce the influence of extreme residuals and random effects (see also Koller & Stahel, Reference Koller, Stahel, Yi and Nordhausen2023). This robustness is achieved by modifying the estimating equations so that residuals and random effects are passed through bounded weighting functions, which automatically downweight highly extreme values rather than allowing them to unduly influence the estimates. Although this approach is computationally demanding, it is well suited to data that may contain contamination or outliers. Because the robust estimation procedure does not provide conventional p-values, inference for the robust model was based on 95% confidence intervals. Effects whose confidence intervals did not include zero were considered consistent with evidence for a statistically reliable effect.
After confirming statistical effects of the three factors with a growth curve model, we proceeded to conduct a divergence point analysis to identify when these effects emerged over time. Following prior work (e.g., Hwang & Han, Reference Hwang and Han2026; Ito & Knoeferle, Reference Ito and Knoeferle2023; Stone et al., Reference Stone, Lago and Schad2021), we first isolated the data for the target conditions to be compared (e.g., L1 versus L2 in Medial-Singular data, L1 versus L2 in Final-Singular data, Medial-Singular versus Final-Singular in L1 data, Medial-Singular versus Final-Singular in L2 data) and then generated 2,000 bootstrap samples by resampling with replacement from the original dataset. For each resampled dataset, we conducted t-tests at every time bin to determine the time point at which fixation patterns began to differ. The divergence point was defined as the earliest time bin that initiated a run of at least four consecutive bins with statistically significant t-values. The outcome of this procedure was a distribution of estimated divergence onsets, and we used the mean of this distribution as our final divergence estimate. Each run also produced a distribution of the difference in divergence points between conditions being compared. We considered the difference statistically reliable when the 95% CI of this distribution excluded zero.
4. Results
In this section, we mainly discuss the statistically significant results, while providing all statistical outcomes in the tables.
4.1. Fill-in-the-blank task
Table 2 summarizes the results of the logistic mixed-effects regression model analysis of the fill-in-the-blank task data. The analysis did not show significant effects of Group, Number, or Position. Accuracy rates were high for L1 speakers (Medial-Singular: 94.79%; Final-Singular: 96.88%; Medial-Plural: 97.92%; Final-Plural: 97.92%) as well as for L2 learners (Medial-Singular: 90.22%; Final-Singular: 92.39%; Medial-Plural: 95.65%; Final-Plural: 95.65%), suggesting that both groups performed very accurately on grammatical number and the lexical semantics of the verbs in the test items.
Output from the logistic mixed-effects model for the fill-in-the-blank task data
Table 2. Long description
The table consists of six columns: Predictor, b (coefficient), 95% C I (Confidence Interval), S E (Standard Error), z-value, and p-value.
* (Intercept): b = 3.93, 95% C I [3.06, 4.81], S E = 0.45, z = 8.79, p < .001.
* Group: b = -0.82, 95% C I [-1.93, 0.28], S E = 0.56, z = -1.47, p = .143.
* Position: b = 0.23, 95% C I [-0.57, 1.03], S E = 0.41, z = 0.57, p = .570.
* Number: b = 0.77, 95% C I [-0.37, 1.92], S E = 0.58, z = 1.33, p = .185.
* Group x Position: b = -0.04, 95% C I [-1.64, 1.56], S E = 0.82, z = -0.05, p = .961.
* Group x Number: b = 0.10, 95% C I [-1.49, 1.69], S E = 0.81, z = 0.12, p = .903.
* Position x Number: b = -0.59, 95% C I [-2.19, 1.01], S E = 0.82, z = -0.72, p = .472.
* Group x Position x Number: b = 0.13, 95% C I [-3.07, 3.33], S E = 1.63, z = 0.08, p = .937.
Footnote: Marginal R super 2 = 0.06; Conditional R super 2 = 0.41.
Note: Marginal R2 = 0.06; Conditional R2 = 0.41.
4.2. Eye-tracking task
A logistic mixed-effects model fitted to the mouse-click accuracy data with Group (L1 versus L2), Position (Medial versus Final), and Number (Singular versus Plural) as fixed effects revealed significant main effects of Group (b = −2.40, SE = 0.67, p < .001) and Number (b = 1.90, SE = 0.49, p < .001), as well as a significant Group × Number interaction (b = −1.90, SE = 0.81, p = .019), as shown in Table 3. In contrast, no significant main effect of Position or related interactions emerged. To further examine the Group × Number interaction, we conducted Bonferroni-adjusted post-hoc comparisons. These comparisons showed that the L1 group was more accurate in the plural conditions than in the singular conditions (b = −2.85, SE = 0.81, p = .003). In addition, the L1 group significantly outperformed the L2 group in the plural conditions (b = 3.35, SE = 0.94, p = .002) (L1: Medial-Plural: 96.88% [SD = 17.49]; Final-Plural: 98.96% [SD = 10.21] versus L2: Medial-Plural: 80.43% [SD = 39.89]; Final-Plural: 84.78% [SD = 36.12]). By contrast, the corresponding group difference in the singular conditions did not reach significance after Bonferroni correction (p = .071), although it was numerically in the same direction (L1: Medial-Singular: 85.42% [SD = 35.48]; Final-Singular: 83.33% [SD = 37.46] versus L2: Medial-Singular: 67.39% [SD = 47.13]; Final-Singular: 69.57% [SD = 46.27]).
Output from the logistic mixed-effects model for the mouse-click response data
Table 3. Long description
The table consists of six columns: Predictor, b, 95% C I, S E, z, and p. Significant results are indicated in bold.
* (Intercept): b = 2.83, 95% C I [2.07, 3.58], S E = 0.39, z = 7.33, p < .001.
* Group: b = -2.40, 95% C I [-3.70, -1.09], S E = 0.67, z = -3.60, p < .001 (Bold).
* Position: b = 0.58, 95% C I [-0.49, 1.64], S E = 0.54, z = 1.07, p = .286.
* Number: b = 1.90, 95% C I [0.94, 2.85], S E = 0.49, z = 3.89, p < .001 (Bold).
* Group x Position: b = -0.36, 95% C I [-1.99, 1.27], S E = 0.83, z = -0.43, p = .665.
* Group x Number: b = -1.90, 95% C I [-3.48, -0.31], S E = 0.81, z = -2.35, p = .019 (Bold).
* Position x Number: b = 0.95, 95% C I [-0.90, 2.79], S E = 0.94, z = 1.01, p = .315.
* Group x Position x Number: b = -1.46, 95% C I [-4.49, 1.56], S E = 1.54, z = -0.95, p = .343.
Note: Marginal R super 2 = 0.32; Conditional R super 2 = 0.63.
Notes: Marginal R2 = 0.32; Conditional R2 = 0.63. Statistically significant effects are shown in bold.
Figure 2 presents a clear visual pattern for our eye-tracking data, with both L1 and L2 participants exhibiting a robust Target advantage in fixations in the Plural conditions even before the verb offset, regardless of whether the verb appeared in Medial or Final position. However, with singular verbs, the two groups had faster Target fixations in Final position compared to Medial position. Notably, the L2 learners showed a slight delay in processing compared to the L1 speakers.
Logit-transformed fixations to target, competitor, and other area over time by Group, Position, and Number. The shaded area depicts 95% confidence intervals; verb offset set to 0 ms across all trials; noun offset, verb offset, and sentence offset lines represent mean offsets.
Figure 2. Long description
The graph is organized into two rows (L 1 English and L 2 English) and four columns (Singular Medial, Singular Final, Plural Medial, Plural Final). The y-axis represents logit-transformed fixations ranging from negative 4 to 3. The x-axis represents time in milliseconds from negative 4000 to 2000.
Three data series are shown in each panel:
* Target: A solid dark blue line.
* Competitor: A dashed light blue line.
* Other: A dotted olive green line.
Each line is surrounded by a light gray shaded area representing 95 percent confidence intervals. Vertical dotted lines mark three temporal anchors: noun offset, verb offset (at 0 ms), and sentence offset.
General trends across panels:
* In the Singular conditions, the Target line rises after the verb offset in both groups, with target–competitor divergence occurring earlier in sentence-final than in sentence-medial position. The L2 group shows a later divergence than the L1 group in both positions.
* In the Plural conditions, both groups show a Target advantage before the verb offset, with the Target line already elevated relative to the Competitor line.
* Across conditions, the Target trajectory generally becomes more dominant over time, while Competitor and Other fixations decline or remain relatively low.
Table 4 summarizes the results of the growth curve analysis, which showed a statistically reliable effect of the linear term (95% CI = [9.67, 10.82]), indicating that logit-transformed fixations underwent significant temporal changes. The analysis revealed that Group had a significant impact on the overall fixations (95% CI = [−2.49, −0.31]), as well as on the linear changes over time (95% CI = [−3.62, −1.32]). Similarly, Position showed statistically reliable effects on the intercept (95% CI = [0.86, 2.43]) and quadratic term (95% CI = [−3.93, −1.78]). Also, Number showed statistically reliable effects on the intercept (95% CI = [1.34, 4.03]), linear term (95% CI = [−12.22, −9.91]), and quadratic term (95% CI = [1.36, 3.52]). The analysis further showed several statistically reliable interaction effects, including interactions between Group and Position on the linear term (95% CI = [3.65, 8.24]) and quadratic term (95% CI = [7.01, 11.33]), between Group and Number on the linear term (95% CI = [−7.49, −2.88]) and quadratic term (95% CI = [−8.30, −3.98]), and between Position and Number on the linear term (95% CI = [1.96, 6.54]) and quadratic term (95% CI = [−6.61, −2.30]). Crucially, there was a three-way interaction among Group, Position, and Number on the intercept (95% CI = [−2.31, −1.18]), linear term (95% CI = [−18.49, −9.32]), and quadratic term (95% CI = [−21.27, −12.63]), suggesting that the temporal patterns of logit-transformed fixations are influenced by a complex interplay between these factors.
Output from the growth curve analysis model for the eye-tracking data
Table 4. Long description
The table contains 24 rows of data across five columns: Parameter, b (coefficient), 95% C I (Confidence Interval), S E (Standard Error), and t (t-statistic). Significant effects are highlighted in bold.
* (Intercept): b = 3.22, 95% C I [2.56, 3.89], S E = 0.34, t = 9.52.
* Linear term: b = 10.24, 95% C I [9.67, 10.82], S E = 0.29, t = 34.78.
* Quadratic term: b = -0.32, 95% C I [-0.86, 0.22], S E = 0.28, t = -1.15.
* Group: b = -1.40, 95% C I [-2.49, -0.31], S E = 0.56, t = -2.53.
* Group Linear term: b = -2.47, 95% C I [-3.62, -1.32], S E = 0.59, t = -4.20.
* Group Quadratic term: b = -0.55, 95% C I [-1.63, 0.53], S E = 0.55, t = -1.00.
* Position: b = 1.64, 95% C I [0.86, 2.43], S E = 0.40, t = 4.12.
* Position Linear term: b = 0.42, 95% C I [-0.72, 1.57], S E = 0.58, t = 0.72.
* Position Quadratic term: b = -2.85, 95% C I [-3.93, -1.78], S E = 0.55, t = -5.18.
* Number: b = 2.69, 95% C I [1.34, 4.03], S E = 0.68, t = 3.92.
* Number Linear term: b = -11.06, 95% C I [-12.22, -9.91], S E = 0.59, t = -18.78.
* Number Quadratic term: b = 2.44, 95% C I [1.36, 3.52], S E = 0.55, t = 4.43.
* Group x Position: b = -0.01, 95% C I [-0.29, 0.27], S E = 0.14, t = -0.05.
* Group x Position Linear term: b = 5.95, 95% C I [3.65, 8.24], S E = 1.17, t = 5.09.
* Group x Position Quadratic term: b = 9.17, 95% C I [7.01, 11.33], S E = 1.10, t = 8.32.
* Group x Number: b = 0.71, 95% C I [-1.51, 2.92], S E = 1.13, t = 0.63.
* Group x Number Linear term: b = -5.19, 95% C I [-7.49, -2.88], S E = 1.18, t = -4.41.
* Group x Number Quadratic term: b = -6.14, 95% C I [-8.30, -3.98], S E = 1.10, t = -5.57.
* Position x Number: b = -0.55, 95% C I [-2.12, 1.01], S E = 0.80, t = -0.69.
* Position x Number Linear term: b = 4.25, 95% C I [1.96, 6.54], S E = 1.17, t = 3.64.
* Position x Number Quadratic term: b = -4.45, 95% C I [-6.61, -2.30], S E = 1.10, t = -4.04.
* Group x Position x Number: b = -1.75, 95% C I [-2.31, -1.18], S E = 0.29, t = -6.07.
* Group x Position x Number Linear term: b = -13.91, 95% C I [-18.49, -9.32], S E = 2.34, t = -5.95.
* Group x Position x Number Quadratic term: b = -16.95, 95% C I [-21.27, -12.63], S E = 2.20, t = -7.69.
Notes: Effects whose confidence intervals did not include zero were considered consistent with evidence for a statistically reliable effect; these effects are shown in bold.
Results from the divergence point analysis were consistent with our visual inspection of Figure 2. They indicated that, in the Medial-Singular condition, target – competitor fixation divergence arose at 875.71 ms (95% CI = [860, 900]) for L1 speakers and 1,112 ms (95% CI = [1100, 1120]) for L2 learners. In the Final-Singular condition, the corresponding onset times were 439.43 ms (95% CI = [420, 440]) for L1 speakers and 970 ms (95% CI = [680, 1,000]) for L2 learners. These group differences were statistically reliable: Compared to the L1 speakers, the L2 learners exhibited delays of 234.55 ms (95% CI = [200, 260]) in the Medial-Singular condition and 530.57 ms (95% CI = [240, 580]) in the Final-Singular condition. Despite these group differences, both groups exhibited delays in the Medial-Singular condition (L1: 875.09 ms, 95% CI = [860, 900]; L2: 1,110.77 ms, 95% CI = [1,100, 1,200]) compared to the Final-Singular condition (L1: 439.59 ms, 95% CI = [420, 440]; L2: 971.11 ms, 95% CI = [680, 1,000]). These Medial-Final differences were also statistically robust (L1: 435.50 ms, 95% CI = [420, 460]; L2: 126.15 ms, 95% CI = [100, 180]). For the Medial-Plural and Final-Plural conditions, divergence onset was 0 ms for both groups. We interpret this pattern cautiously as an early bias toward the plural image, likely driven by the greater visual complexity of plural objects (see Section 5).
As an anonymous reviewer pointed out, proficiency is an important individual difference variable that warrants examination. To investigate a potential proficiency effect, we constructed an additional growth curve model for L2 eye-tracking data within the critical time window that incorporated proficiency as a fixed effect (converged model formula: target_advantage ~ (poly1 + poly2) * positionCtr * numberCtr * proficiency + (1 + numberCtr | participant) + (1 | item)). This model revealed statistically reliable three-way interactions among Position, Number, and Proficiency on both linear (95% CI = [−1.54, −0.77]) and quadratic (95% CI = [−0.91, −0.19]) terms. Given the small L2 sample size, which limited our ability to perform separate divergence point analyses by proficiency level, we instead conducted visual inspection after stratifying the data using a median-split approach (median = 34) to create higher-proficiency and lower-proficiency groups. During the critical post-verb offset window, higher-proficiency L2 speakers demonstrated faster and more robust Target fixation in the Medial-Singular condition, yet slower Target fixation in the Final-Singular condition relative to lower-proficiency L2 speakers (see Figure 3). This finding is somewhat difficult to interpret and may not represent generalizable L2 processing patterns given our limited sample size. We therefore regard this proficiency analysis as exploratory and refrain from drawing strong conclusions from it. It is possible that proficiency effects may become more explainable with a larger, more proficiency-diverse L2 sample, such that more advanced learners exhibit faster and more robust processing patterns in all conditions.
Logit-transformed fixations to target, competitor, and other area over time by Proficiency group, Position, and Number. The shaded area depicts 95% confidence intervals; verb offset set to 0 ms across all trials; noun offset, verb offset, and sentence offset lines represent mean offsets.
Figure 3. Long description
The multi-panel line graph is organized into two rows by proficiency (Higher Proficiency on top, Lower Proficiency on bottom) and four columns by number and position (Singular Medial, Singular Final, Plural Medial, Plural Final). The x-axis represents Time in m s from negative 4000 to positive 2000, with 0 m s marked as the verb offset. The y-axis represents Logit-transformed fixations from negative 3 to 3. Three lines are plotted in each panel: a solid dark blue line for Target, a dashed light blue line for Competitor, and a dotted olive line for Other. Shaded gray areas indicate 95 percent confidence intervals. Vertical dotted lines mark the noun offset, verb offset at 0, and sentence offset.
* Higher Proficiency Row: In all four panels, the Target line shows a sharp upward trend following the verb offset, peaking near the sentence offset. The Competitor line generally mirrors the Target line early in the time course but drops as the Target line rises. In the Plural conditions, the Target line begins its ascent earlier than in Singular conditions.
* Lower Proficiency Row: Proficiency-related patterns differ by condition. In the post-verb-offset window, the lower-proficiency group shows a later and weaker target preference in the Medial-Singular condition, but an earlier target preference in the Final-Singular condition, relative to the higher-proficiency group. These patterns are descriptive and exploratory.
* Across all panels, the Other category remains relatively stable and low, generally fluctuating between negative 1 and negative 3 on the logit scale.
The same reviewer further pointed out that it is important to determine whether the processing effect stems from position (i.e., sentence-medial versus sentence-final), phonetic duration (short versus long), or both. To address this question, we first examined the Variance Inflation Factor between Position and Verb Duration and found it to be 1.37—well below the multicollinearity threshold of 4 (e.g., Fox, Reference Fox1991), indicating that multicollinearity was not a concern. We then fitted two growth curve models including both Position and Verb Duration so that the significance of each factor could be evaluated while controlling for the other. Both models included the same random-effects structure as the main model reported in Section 4.2 (formula of the model where Verb Duration was controlled: target_advantage ~ (poly1 + poly2) * group * position * number + verb_duration + (1 + number | participant) + (1 | item); formula of the model where Position was controlled: target_advantage ~ (poly1 + poly2) * group * verb_duration * number + position + (1 + number | participant) + (1 | item)).
The results showed that when Verb Duration was controlled for, Position remained statistically reliable on the intercept (95% CI = [0.60, 2.44]) and quadratic term (95% CI = [−3.93, −1.78]), together with its interactions with other factors: Group × Position on the linear term (95% CI = [3.65, 8.24]) and quadratic term (95% CI = [7.01, 11.33]); Position × Number on the linear term (95% CI = [1.96, 6.55]) and quadratic term (95% CI = [−6.61, −2.30]); Group × Position × Number on the intercept (95% CI = [−2.31, −1.18]), linear term (95% CI = [−18.49, −9.33]), and quadratic term (95% CI = [−21.27, −12.63]). A similar pattern of results was observed when Position was controlled for. Specifically, Verb Duration remained statistically reliable on the linear term (95% CI = [−21.84, −14.02]) and quadratic term (95% CI = [−12.44, −4.85]), along with its interaction with other factors: Group × Verb Duration on the intercept (95% CI = [0.17, 2.41]), linear term (95% CI = [22.34, 37.92]), and quadratic term (95% CI = [23.44, 38.61]); Verb Duration × Number on the linear term (95% CI = [11.41, 27.05]) and quadratic term (95% CI = [−19.27, −4.09]); Group × Verb Duration × Number on the intercept (95% CI = [−7.74, −3.26]), linear term (95% CI = [−55.20, −24.04]), and quadratic term (95% CI = [−51.18, −20.83]). These results suggest that both Position and Verb Duration contribute to the temporal dynamics of target fixation. Because the two cues co-vary by design, we do not claim that their effects are fully independent; rather, each accounts for variance that the other does not fully absorb, consistent with our treatment of sentence position as a composite cue with partially separable positional and durational components.
5. General discussion
The present study used the visual-world eye-tracking paradigm to investigate how sentence position influences the real-time processing of English 3SG -s among L2 learners. There are several significant findings in this study. The first important finding concerns the discrepancy between offline and online performance, both of which consistently manipulated Position and Number. In the fill-in-the-blank task, the L2 group performed at near-ceiling levels, whereas accuracy in the eye-tracking task was notably lower—despite the fact that the eye-tracking trials presented only two response options (chance performance = 50%), compared to the four response options in the fill-in-the-blank task (chance performance = 25%). These findings indicate that the relatively advanced L2 learners performed very well on the offline written task while showing greater vulnerability in the online auditory task. This pattern is consistent with the written-task advantage reported in Johnson (Reference Johnson1992), who found that adult L2 learners tend to perform better on written than on aural measures of knowledge of English inflectional morphemes, suggesting that target linguistic knowledge can be masked in aural tasks.
Several processing-related factors likely contributed to the lower performance on the eye-tracking task. For one, the two tasks differed in modality: the fill-in-the-blank task presented written stimuli, whereas the eye-tracking task presented auditory stimuli. The lower accuracy rates observed in the eye-tracking task therefore suggest that sensitivity to 3SG -s is less stable under spoken real-time conditions than in the offline written task. In addition, the auditory stimuli in the eye-tracking task were presented only once, which may have further strained participants’ attentional resources, resulting in more incorrect responses.
While an auditory processing account provides one plausible explanation for the results, alternative interpretations must also be considered. For example, the offline task may have allowed greater reliance on reflective or explicit strategies, whereas the online task required rapid integration of spoken input under time pressure, with limited opportunity for such strategies.
This contrast between reflective, untimed conditions and demanding, real-time conditions connects to a conceptual distinction proposed in second language acquisition literature between explicit and implicit knowledge. Explicit knowledge is commonly defined as conscious, declarative, and potentially verbalizable knowledge about language, whereas implicit knowledge is typically characterized as less conscious, not readily verbalizable, and more closely associated with automatic processing (Hulstijn, Reference Hulstijn2005). We do not, however, interpret the fill-in-the-blank and visual-world tasks as pure measures of two categorically distinct knowledge types. As noted above, the two tasks differ along several dimensions (modality, timing, and task demands), so the dissociation cannot be mapped solely onto a knowledge-type contrast. More fundamentally, the dissociation we observe is better characterized as one between knowledge accessible under offline task conditions and its real-time deployment (cf. Hopp, Reference Hopp2009). Near-ceiling offline accuracy suggests that the relevant grammatical knowledge was available, whereas weaker online performance reflects the cost of detecting and integrating a brief auditory morpheme under real-time processing demands. On this view, tasks intended to tap rapidly deployed, implicit-like knowledge can underestimate learners’ target knowledge when perceptual and processing load is high; reduced online performance, particularly when paired with strong offline performance, should therefore not be read as evidence that such knowledge is absent.
Second, our results indicate a sentence-position effect in the processing of 3SG -s in the critical singular conditions, with earlier Target fixations in sentence-final than in sentence-medial position. In the singular conditions, the timing of fixation divergence between the Target and Competitor emerged earlier when verbs appeared in sentence-final position than in sentence-medial position. A similar pattern was observed among L1 participants in the singular conditions, who showed earlier fixation divergence in sentence-final position than in sentence-medial position. This pattern is consistent with previous findings by Sundara et al. (Reference Sundara, Demuth and Kuhl2011), who likewise reported a sentence-final advantage in the perception of 3SG -s among L1-English children.
While we found that sentence-final position led to earlier target-consistent fixations, sentence position should not be treated as a unitary explanatory factor. Sentence-final position bundles together multiple cues, including phrase-final lengthening, prosodic boundary marking, and reduced competition from subsequent input (cf. Sundara et al., Reference Sundara, Demuth and Kuhl2011). In sentence-medial contexts, the 3SG -s is often followed by vowel-initial material, potentially allowing resyllabification or coarticulation, whereas in sentence-final position the morpheme may be more acoustically salient. Inspection of our stimuli suggests that medial-position verbs in our critical items were indeed followed by vowel-initial prepositional phrases, a context that could in principle support such resyllabification or coarticulation, though we did not manipulate this factor systematically. Thus, the observed position effect likely reflects a combination of structural and phonetic – prosodic factors.
One factor that this study highlights is the longer duration of 3SG -s in sentence-final position, which is likely to enhance perceptual prominence (Sundara et al., Reference Sundara, Demuth and Kuhl2011). In our auditory stimuli, which were recorded by a native speaker of English, 3SG -s was consistently longer in sentence-final than in sentence-medial position, mirroring the pattern reported in Sundara et al. (Reference Sundara, Demuth and Kuhl2011) and likely enhancing its phonetic prominence. At the same time, the combined model analysis reported in Section 4.2 showed that both Verb Duration and Position contributed to the temporal trajectory of fixation patterns, suggesting that the observed effect reflects the joint contribution of multiple cues associated with sentence position.
Beyond the nature of the position effect itself, the relation between time-course and outcome measures within the eye-tracking task also bears on the theoretical interpretation of our findings. The logistic mixed-effects model fitted to mouse-click accuracy revealed no main effect of Position and no interactions involving Position, indicating that final response outcomes did not differ reliably across sentence-medial and sentence-final conditions. In contrast, the growth curve analysis revealed robust effects of Position, both as a main effect and in interaction with Group and Number, on the temporal trajectory of fixations. This accuracy – time-course dissociation is informative: even on trials in which participants ultimately converged on the correct interpretation, sentence position modulated the temporal dynamics by which target-consistent fixations emerged, without altering whether that interpretation was reached by the end of the trial. This pattern is consistent with a processing-based account, under which transient difficulty during online integration need not propagate to final response outcomes, and it converges with the broader dissociation observed between online performance in the eye-tracking task and near-ceiling performance in the offline fill-in-the-blank task. By contrast, this broader dissociation is less straightforwardly compatible with strong morphosyntactic representational-deficit accounts, on which the relevant features are absent, and difficulty should therefore persist in the offline task as well.
Third, the current study reveals that processing of 3SG -s is slower in L2 learners than in L1 speakers. Analysis of data from relevant trials in the eye-tracking task showed that the amount of time it took for a significant difference in fixations to emerge between the Target and Competitor was longer for the L2 group than for the L1 group. This delay occurred in both final and medial positions. These results are consistent with previous research showing slower real-time processing in L2 learners than in L1 speakers (e.g., Hwang & Han, Reference Hwang and Han2026). The larger final – medial difference observed for the L1 group may reflect native speakers’ more efficient use of the perceptual and processing advantages associated with sentence-final position. However, the L2 group remained reliably delayed relative to the L1 group even in sentence-final position.
Although the focus of this study is on the 3SG -s, it is noteworthy that both L1 speakers and L2 learners consistently demonstrated an early attentional bias toward plural images, fixating on them more often than singular images across all four conditions from the beginning of the trial. This phenomenon, which is well-documented in the visual-world eye-tracking literature (Brown et al., Reference Brown, Fox and Strand2022; Hwang & Kim, Reference Hwang and Kim2025; Rusk et al., Reference Rusk, Paradis and Järvikivi2020), has been explained as a consequence of the heightened visual complexity inherent in images with plural subjects. This visual salience may naturally capture attention, as exemplified by Brown et al.’s (Reference Brown, Fox and Strand2022) work where L1-English speakers tended to look at plural images more than singular images irrespective of linguistic agreement cues.
As an anonymous reviewer noted, the sentence position effects observed in this study may also be related to information-structural factors. Although information structure in the strict given – new sense presupposes discourse context, the present study did not manipulate context, as target sentences were presented in isolation. Nevertheless, because sentence-final position is typically associated with focus and informational prominence in English (Carroll, Reference Carroll, O’Brien, Shea and Archibald2006), information-structural prominence may contribute to the perceptual salience of low-salience morphemes such as 3SG -s. Future research should directly manipulate discourse context and information structure to examine how these factors interact with sentence position in real-time processing and acquisition.
In addition, the same reviewer noted that the variety of English may influence how 3SG -s is perceived. The stimuli in the present study were recorded by an American English speaker, and English instruction in Korea is generally oriented toward American English, with widely used instructional materials typically adopting American English norms. Korean EFL learners are therefore likely to have greater exposure to American English than to other varieties, which limits the likelihood of an accent-mismatch effect here. Whether the variety of input affects the perception of 3SG -s remains a question for future research.
6. Conclusion
The present study indicates a dissociation between near-ceiling offline performance and weaker online auditory performance, suggesting that L2 difficulty with 3SG -s may lie less in whether learners possess knowledge of this morpheme and more in how efficiently they can detect and integrate it under real-time spoken processing demands. The findings further show that sentence position modulates the real-time processing of 3SG -s in L2 learners, with sentence-final forms facilitating earlier processing than sentence-medial forms. This sentence-final advantage appears to reflect the joint contribution of multiple cues associated with sentence position, including the longer phonetic duration of the morpheme. Taken together, these results suggest that, for intermediate-to-advanced L2 learners, difficulty with 3SG -s is more plausibly explained by processing-based constraints than by strong morphosyntactic representational-deficit accounts.
Data availability statement
All materials, data, and R analysis scripts are available at https://osf.io/jvgq4.
Acknowledgments
We thank Amber Camp and Pui Yee ‘Ruby’ Lee for their help in creating the audio and visual stimuli for the eye-tracking task. We are also grateful to Kook-Hee Gil, Heather Goad, Theres Grüter, Holger Hopp, Tania Ionin, Jason Rothman, Bonnie D. Schwartz, Roumyana Slabakova, and Neal Snape, as well as to audiences at BUCLD 50, GASLA 18, the Language Acquisition Research Group at the University of Hawai‘i at Mānoa, and the Research Colloquium in Linguistics at TU Braunschweig, for their invaluable comments. Finally, we thank the anonymous reviewers and all our participants. AI tools were used for proofing. No original content was generated by AI.
Funding statement
This work was supported by the Ministry of Education of the Republic of Korea and the National Research Foundation of Korea (NRF-2023S1A5A8078900).
Competing interests
The authors declare none.
Open access
