1. Introduction
As we comprehend sentences, our minds often predict upcoming information (Huettig, Reference Huettig2015; Kuperberg & Jaeger, Reference Kuperberg and Jaeger2016). Research has shown that first language (L1) speakers generate predictions based on various types of linguistic information, including phonological, morphosyntactic, and semantic cues (e.g., Aumeistere et al., Reference Aumeistere, Bultena and Brouwer2022; Itzhak & Baum, Reference Itzhak and Baum2015). Previous work on second language (L2) prediction has focused on whether the underlying cognitive processes involved in L1 and L2 are similar or different. Some research has identified L1-L2 differences (e.g., Mitsugi, Reference Mitsugi2017; Mitsugi & MacWhinney, Reference Mitsugi and Macwhinney2016; Perdomo & Kaan, Reference Perdomo and Kaan2021), suggesting that L2 learners exhibit a reduced ability to generate expectations, which has given rise to the RAGE hypothesis (Grüter et al., Reference Grüter, Rohde and Schafer2017; Grüter & Rohde, Reference Grüter and Rohde2021). However, other research has shown L1–L2 similarities (e.g., Dijkgraaf et al., Reference Dijkgraaf, Hartsuiker and Duyck2017; Fang & Wu, Reference Fang and Wu2024; Ito et al., Reference Ito, Corley and Pickering2018; Kim & Grüter, Reference Kim and Grüter2021; Mitsugi, Reference Mitsugi2022), suggesting that L2 learners have a sound ability to generate expectations, which has led to the development of the SAGE hypothesis (Hwang & Kim, Reference Hwang and Kim2025).
While the RAGE and SAGE hypotheses present competing explanations, they converge on a crucial point that various factors influence predictive processing. The utility account, which originated in the L1 context (Kuperberg & Jaeger, Reference Kuperberg and Jaeger2016) and has since been extended to the L2 context (Kaan & Grüter, Reference Kaan, Grüter, Kaan and Grüter2021), is relevant to our research purpose as it emphasizes the dynamic and adaptive nature in predictive processing. This account posits that individuals strategically analyze the utility of prediction during comprehension, such that they will engage in prediction when the benefits of generating expectations exceed the costs, but not when the costs outweigh the benefits. According to this account, the prediction process is influenced by a range of factors, including linguistic cue validity and individual differences (see Section 2.1). Yet, the L2 context remains an understudied area in this regard, which highlights the need for an examination of various factors to uncover the underlying mechanisms that shape L2 predictive processing. In particular, some of the most pressing research questions are: What specific linguistic and individual factors drive the dynamics of prediction? Do these factors have similar or different effects on the predictive mechanisms of L1 and L2 speakers? And how do they influence the robustness and/or timing of prediction?
To explore these questions, the current study examines the roles of the linguistic cue validity, language proficiency, and immersion experience as potential factors that may influence the utility of prediction, using a visual-world eye-tracking paradigm. Cue validity refers to the informational value or predictive power of a specific linguistic feature within a particular context (MacWhinney et al., Reference MacWhinney, Bates and Kliegl1984). In this study, we operationally define cue validity as the likelihood that a predictive cue will accurately forecast an upcoming element. One linguistic phenomenon that can serve as a useful probe to investigate this factor is number agreement in expletive sentences, as shown in (1), where a there-associated NP appears post-verbally, and so the verb (i.e., the predictive cue) should agree with the head noun (i.e., the element to be predicted) in terms of number. From this point on, we will refer to this head noun as the “there-associated noun.”
Crucial to our study is the fact that the validity of the predictive cue varies depending on the number and aspect of the verb, allowing us to examine the effects of cue validity on predictive processing of the grammatical number of upcoming there-associated nouns. Whereas plural verbs (e.g., are, have been) are consistently followed by number-matching plural nouns, regardless of aspect, singular verbs (e.g., is, has been) exhibit greater variability, as they can be followed by a range of noun types, including number-matching singular nouns, uncountable nouns, and even plural nouns, albeit less frequently (e.g., Munn, Reference Munn1999); in the case of singular verbs, the aspect further plays a significant role, with those in the perfect aspect being even less likely to be followed by singular nouns than those in the simple aspect (see Section 4.2). These differences in verb–noun associations suggest that verbs in expletive sentences have varying levels of cue validity, with singular verbs in the perfect aspect having the lowest cue validity for predicting number-matching nouns, followed by singular verbs in the simple aspect, and then plural verbs in either aspect. A reasonable expectation is that cue validity will influence the speed of predictive processing, such that cues with lower validity will decrease the benefits of prediction, leading to weaker and slower prediction.
While examining the cue validity in there-expletive sentences among L1 speakers and L2 learners, we also investigate how two key individual factors – L2 proficiency and L2 immersion experience – influence the L2 group’s prediction patterns. For the current study, we operationally define L2 proficiency as an individual’s control over the L2 system for communication, and L2 immersion experience as experience residing in an English-speaking country for a minimum of 3 months with consistent daily interaction in English. It is worth noting that the validity of predictive cues discussed above cannot be influenced by our L2 participants’ L1, Korean, because Korean presents neither there-expletives nor number agreement. As a result, not all L1-Korean L2-English learners may exhibit predictive processing of there-associated nouns. However, increased proficiency in English and/or increased immersion experience and exposure to diverse, authentic English input could reduce the cognitive costs associated with prediction for these learners, potentially resulting in more robust and rapid prediction. Regarding the validity of cues for prediction, those with higher proficiency and/or more extensive immersion are likely to exhibit greater sensitivity, which may lead to a stronger cue validity effect compared to those with lower proficiency and/or no immersion. In this sense, the effect of cue validity may be modulated by individual differences in proficiency and L2 experience. While the role of proficiency in L2 prediction has been widely studied, with mixed findings (e.g., Hwang & Kim, Reference Hwang and Kim2025; Kim & Grüter, Reference Kim and Grüter2021; Perdomo & Kaan, Reference Perdomo and Kaan2021), the influence of immersion experience has not yet been a focus of research in this area. Moreover, previous research has not considered the interaction of these individual factors with cue validity. Therefore, examining the impact of both proficiency and immersion experience on L2 learners’ predictive processing – and how these relate to cue validity – can provide useful insights into how these factors shape their predictive processing mechanisms.
2. Literature review
2.1. Factors involved in second language prediction
The utility-based account of predictive processing proposes that language users weigh the costs and benefits of making predictions in order to achieve their communicative goals (Kaan & Grüter, Reference Kaan, Grüter, Kaan and Grüter2021; Kuperberg & Jaeger, Reference Kuperberg and Jaeger2016). Based on this cost–benefit analysis, individuals dynamically adapt their predictive behavior, adjusting both whether they engage in prediction and what they choose to predict, depending on factors like the validity of predictive cues, their own goals, task demands, and available resources. Regarding the validity of predictive cues, the focus of this study, language users may rely on valid cues for active prediction. When predictive cues are not valid, however, they may adjust by using different cues or refraining from prediction altogether to maximize their processing efficiency (see Kaan & Grüter, Reference Kaan, Grüter, Kaan and Grüter2021).
Although research that has explicitly manipulated the validity of predictive cue in sentence-level predictive processing is scarce (with some research exploring syntactic and semantic anomalies in the context of prediction [e.g., Coulson et al., Reference Coulson, King and Kutas1998; Haeuser & Kray, Reference Haeuser and Kray2022]), a limited number of studies provide suggestive evidence highlighting its importance in lexical-level predictive processing. Notably, Lau et al.’s (Reference Lau, Holcomb and Kuperberg2013) event-related potential (ERP) study investigated the impact of cue validity on L1-English predictive processing using a semantic category probe detection task. In order to vary the predictive validity of the prime for the target, the researchers manipulated the proportion of related words between the prime and target, while maintaining the words’ semantic association constant. Specifically, for the target word square, they used the prime rectangle which had a higher proportion of related words, as well as the prime table which had a lower proportion of related words. The results demonstrated that a higher proportion of related words led to a larger and faster priming effect, indicating that the presence of more valid cues enables more robust and rapid predictive processing. Given that sentence-level prediction relies on integrating early words and anticipating upcoming information based on these words, it is conceivable that the effects of cue validity on lexical processing carries over to more complex, sentence-level predictive processing. The current study thus builds upon existing research to provide new empirical evidence on how cue validity operates in sentence-level prediction.
In the domain of L2 processing, Henry et al. (Reference Henry, Jackson and Hopp2022) examined cue coalitions, focusing on how L1-English L2-German learners integrate multiple linguistic cues, such as case and prosody, during predictive processing. Their visual-world eye-tracking task included sentences featuring case markers to distinguish between subjects and objects, as in Den Wolf tötet gleich der Jäger [The wolfACC kills soon the hunterNOM] “The hunter will soon kill the wolf.” The results showed that while L2 learners did not successfully predict post-verbal arguments using case cues alone, their prediction success improved when both case and prosodic cues were available, suggesting that multiple cues can jointly contribute to predictions.
The previous discussion suggests the crucial role of cue validity in shaping the predictive behaviors of language users, where more valid cues can facilitate prediction, and emphasizes the complex interplay between multiple cues in processing. However, the study of cue validity in (L2) predictive processing is still in its infancy. In particular, it remains unclear how language users utilize morphosyntactic cues with differing levels of validity to inform their predictions. This study aims to address this gap by focusing on the validity of morphosyntactic cues in there-expletive sentences.
In addition to cue validity, this study investigates the influence of two individual differences factors on L2 learners’ predictive processing patterns. The first factor, L2 proficiency, was measured using a Cloze proficiency test in this study (see Section 4.3). The second factor, immersion experience, was determined through a language background questionnaire asking participants about the duration of their stay in an English-speaking country (see Section 4.3).
The impact of L2 proficiency has long been a topic of interest in the field, with the expectation that L2 learners’ predictive processing abilities will converge with those of L1 speakers as their overall proficiency improves. However, previous research has yielded inconsistent results regarding the role of proficiency in L2 prediction (Kaan & Grüter, Reference Kaan, Grüter, Kaan and Grüter2021). There is some evidence to suggest that L2 learners with higher proficiency demonstrate predictive abilities that are more native-like, sometimes with a slight delay in processing (e.g., Chambers & Cooke, Reference Chambers and Cooke2009; Dussias et al., Reference Dussias, Valdés Kroff, Guzzardo Tamargo and Gerfen2013; Henry et al., Reference Henry, Jackson and Hopp2022; Hopp, Reference Hopp2013; Hopp & Lemmerth, Reference Hopp and Lemmerth2018; Hwang & Kim, Reference Hwang and Kim2025). Several studies, on the other hand, have failed to find any L2 proficiency effects in prediction or found that even advanced L2 learners differ from L1 speakers in their use of predictive cues (e.g., Hopp, Reference Hopp2015; Kim & Grüter, Reference Kim and Grüter2021; Mitsugi, Reference Mitsugi2020; Perdomo & Kaan, Reference Perdomo and Kaan2021). While the role of L2 proficiency in L2 predictive processing remains a topic of ongoing debate, it is essential to exercise caution when interpreting the absence of proficiency effects. One possible reason for the lack of such effects may be the challenge of defining what constitutes high proficiency, as well as the need to ensure a sufficient range of proficiency levels in the data collection process.
Regarding L2 immersion experience, while research is scarce in the context of predictive processing, Klassen et al. (Reference Klassen, Ferreira and Schwieter2023) showed that L1-English L2-Spanish speakers exhibited target-like predictive processing patterns following a 25-day study-abroad program in Spain (for the role of immersion in integrative processing,Footnote 1 see Pliatsikas & Marinis, Reference Pliatsikas and Marinis2013). This result can be attributed to the quality of the input. L2 immersion experience provides L2 learners with a wide range of authentic language experiences across both written and spoken modalities. In contrast, classroom exposure is often limited to textbook-based input, which may not offer the same level of linguistic authenticity and diversity. This difference suggests that immersion experience may lead to more L1-like predictive processing routines, enabling learners to automatically and effortlessly process complex morphosyntactic information to anticipate upcoming information. However, it is still largely unknown how L2 proficiency and L2 immersion independently influence L2 predictive processing within the same individuals.
To shed new light on the debate surrounding the individual differences in L2 predictive processing, the current study collected data from L2 learners with varying levels of proficiency and immersion experience. By doing so, this study aims to provide a more detailed examination of whether, and if so, how each of these two individual factors contributes to facilitating more efficient predictive processing strategies in L2 learners.
2.2. Number features in (second language) processing
While number has been a well-studied grammatical phenomenon in language processing research, there is a notable lack of studies on how L2 learners use this feature to make predictions. To date, only a single study has investigated the predictive processing of number, and this is limited to L1-English speakers. In a visual-world eye-tracking paradigm, Brown et al. (Reference Brown, Fox and Strand2022) examined whether L1-English speakers utilize the grammatical number of copular verbs within questions (e.g., Where is/are the …?) to predict the upcoming objects. The results showed that the L1-English speakers fixated more frequently on plural objects prior to the target word when they encountered the phrase Where are the versus Where is the. These results suggest that L1-English speakers make use of the verb’s number information to anticipate the upcoming referent, thereby guiding their sentence processing. However, the study also revealed a general preference for plural objects even in the Where is the context, which was attributed to the visual complexity of plural objects (see also Rusk et al., Reference Rusk, Paradis and Järvikivi2020).
The majority of L2 processing studies on grammatical number have been done within an integrative processing context. Research on this phenomenon has sparked debate about whether L2 learners are sensitive to number features or agreement during real-time processing. Some studies have found that L2 learners lack this sensitivity (e.g., Chen et al., Reference Chen, Shu, Liu, Zhao and Li2007; Jiang, Reference Jiang2004, Reference Jiang2007; Rusk et al., Reference Rusk, Paradis and Järvikivi2020; Tamura et al., Reference Tamura, Fukuta, Nishimura and Kato2023). This finding has been taken as support for the Shallow Structure Hypothesis (Clahsen & Felser, Reference Clahsen and Felser2006, Reference Clahsen and Felser2018), according to which L2 processing is different from L1 processing such that L2 learners construct less detailed morphosyntactic representations in real time.
Jiang’s (Reference Jiang2004) study was the first to investigate number processing in L2-English learners, where “proficient” (p. 611) learners completed a self-paced reading task that manipulated subject–verb number agreement (e.g., The *bridge/bridges to the island were about ten miles away). In the study, unlike L1-English speakers, L1-Chinese L2-English learners did not show sensitivity to agreement violation. This group difference led Jiang to conclude that the English nominal plural morpheme -s “is not an integrated part of [the L1-Chinese L2 learners’] L2 competence” (p. 624; see also Jiang, Reference Jiang2007). However, the conclusions of Jiang (Reference Jiang2004) remain to be confirmed, as the results of their L2 learners followed the same numerical pattern as those of the L1 controls (see Jiang, Reference Jiang2004, p. 613). In addition, it is debatable whether the tested L2 learners can fairly be deemed “proficient” because their proficiency varied widely, as shown in the range of the reported TOEFL scores (i.e., 570–667). It is conceivable that the high variability in proficiency in these L2 learners gave rise to a null effect of their sensitivity to subject-verb number (dis)agreement.
More relevant work to the current study is that of Tamura et al. (Reference Tamura, Fukuta, Nishimura and Kato2023), which examined two sentence types related to number agreement, including the there construction, using a self-paced reading task. In regular sentences, their L1-Japanese L2-English learners were found to fail to identify number agreement violations (e.g., The mother and her son *is/are in the cottage now). In there-expletive constructions (e.g., there *is/are a pen and an eraser), on the other hand, the learners exhibited longer reaction times (RTs) at the first noun phrase (e.g., a pen) when it followed a plural verb, likely perceiving this as a number agreement violation. Also, with repeated exposure to similar patterns, they began to anticipate a coordinated structure even before encountering the conjunction. The researchers interpreted their findings as showing shallow syntactic processing among L1-Japanese L2-English learners. However, in our view, the results from the there construction may not necessarily be considered as evidence of a processing deficit in L2 learners; rather, it seems possible that the tested learners were sensitive to number agreement initially, and capable of reanalyzing sentence structure following exposure. Moreover, the shift in their expectations alludes to the potential role of cue validity in guiding predictive processing – an issue that our study focuses on.
Another body of research has demonstrated that L2 learners are indeed sensitive to number agreement between subject and verb (e.g., Cheng et al., Reference Cheng, Rothman and Cunnings2022; Lee & Phillips, Reference Lee and Phillips2023; Lim & Christianson, Reference Lim and Christianson2015; Wen et al., Reference Wen, Miyao, Takeda, Chu, Schwartz, Franich, Iserman and Keil2010). For example, Lim and Christianson showed that, like L1-English controls, L1-Korean L2-English learners with higher proficiency (measured with a Cloze test) were able to process number agreement violations (e.g., *The teacher who instructed the students were very strict) in an eye-tracking-while-reading experiment. Similarly, Cheng et al. found that L1-Chinese L2-English learners, who had immersion experience of studying in the UK at the time of testing (length of immersion: M = 31.6 months, SD = 30.1) and had relatively high proficiency scores on the quick Oxford Placement Test (M = 43 [out of 60], SD = 6.33), also showed sensitivity to number agreement in a self-paced reading task.
In summary, research on L2 number processing has primarily focused on the integrative processing of number agreement between subject and verb, yielding inconsistent findings. The current study builds on the existing literature by exploring the predictive processing patterns of L2 learners in regard to the number cues encoded in verbs in there-expletive sentences.
3. The present study
The following research questions (RQs) frame this study:
RQ 1: Do linguistic cues that vary in validity based on verb number and aspect affect the robustness or timing of predictions about the number of upcoming there-associated nouns in L1-English speakers and L1-Korean L2-English learners?
RQ 2: Do language proficiency and/or immersion experience affect the robustness or timing of predictions about the number of upcoming there-associated nouns in L1-Korean L2-English learners?
Grounded in the utility account, we formulate the following hypotheses for each research question. For RQ 1, we predict that increased cue validity will enhance the benefits of prediction in both L1-English speakers and L1-Korean L2-English learners. Conversely, decreased cue validity will increase the costs of prediction, resulting in weaker and slower prediction. As for RQ 2, we hypothesize that higher L2 proficiency and/or greater L2 immersion experience will reduce the cognitive costs associated with prediction, thereby enabling stronger and faster prediction in L1-Korean L2-English learners. Additionally, we consider that greater L2 proficiency and/or L2 immersion may strengthen the cue validity effect in L2 learners’ predictive processing.
4. Method
4.1. Participants
The study involved the participation of 26 self-identified L1-English speakers (L1 speakers, hereafter) and 40 self-identified L1-Korean L2-English learners (L2 learners, hereafter). The study received ethical approval from the first author’s institution, and each participant provided written consent before the commencement of the study. Calibration problems with the eye-tracking equipment led to the exclusion of three participants (one L1 speaker and two L2 learners) from the analysis. Upon initial analysis of the L2 proficiency data obtained from a Cloze test, we identified one L1 speaker’s proficiency score (13 out of 50) and two L2 learners’ scores (1 and 7 out of 50) as outliers, which exceeded 1.5 times the interquartile range for their respective groups, and thus excluded these participants as well. This led to a final sample of 24 L1 speakers and 36 L2 learners. Table 1 provides an overview of the demographic information for these participants.
Background information of participants
4.2. Materials
The visual-world eye-tracking task featured a set of 60 audio stimuli, including 20 critical sentences (Supplementary Appendix S1) and 40 filler sentences. As shown in (1), the 20 critical sentences were constructed using a Latin square design, with two factors – Verb number (Singular; Plural) and Aspect (Simple; Perfect) – crossed to systematically vary the predictive cue validity. In addition to be, we chose to use remain as the main verb for lexical variations based on consultation with two L1-English linguists. These two verbs were judged to be more natural within the syntactic frame of our interest, particularly when used with the perfect aspect (e.g., (1b), (1d)), compared to other verbs initially considered (e.g., appear, seem).
As an independent measure of the cue validity, we used faith (Gries, Reference Gries2005), which is one of the association strength indices calculating the conditional probability of two items occurring together. In our case, this measure determined the likelihood of a target phrase including a predictive cue (there is/remains; there has been/remained; there are/remain; there have been/remained) being followed by a certain grammatical number of there-associated nouns (e.g., singular, plural, uncountable). In this way, the faith measure provides an estimate of how validly each cue predicts a particular grammatical number of there-associated nouns.
To measure the faith values, we used the News On the Web (NOW) corpus (https://www.english-corpora.org/now/), which contains 20.3 billion words. We opted for this corpus because it produced the largest number of target phrases among the publicly available native English corpora at the time of our analysis. In the NOW corpus, we first searched for eight specific phrases (there is/remains; there has been/remained; there are/remain; there have been/remained) using the “List” feature, aiming to extract the first 100 concordance lines for each phrase (e.g., Latić, Reference Latić, Sadeghpour and Sharifian2021). However, the search engine was unable to display the concordance lines for there is, there are, there has been, and there have been, due to their extremely high frequency; it instead returned an error message saying, for example, “The word/phrase < there is > occurs more than 500,000 times in the corpus, and this is too many to see them in the ‘Keyword in Context’ display. Please limit the search (e.g., < * there is > or < there is * >) to generate a new list of strings.” After consulting with the corpus creator, we implemented the suggested workaround to overcome the issue. Specifically, we searched for the target phrase followed by an asterisk, which generated various types of subsequent words (e.g., there is a , there is no, there is an ) instead of concordance lines. We then clicked on each of the 100 types and selected the first concordance line for each type. If the type under search still yielded an error message, we repeated the same process using the asterisk (e.g., there is a *) to obtain the first concordance line.Footnote 2
While we were able to obtain the full 100 instances for most phrases, we encountered limitations with there has remained, which provided 92 instances, and there have remained, which yielded only 32 instances. Our review of the instances also revealed that there in some cases was used as an adverbial to indicate location rather than as an expletive (e.g., many people out there remain deeply proud to be American): Two instances for there is; two instances for there has been; two instances for there are; one instance for there have been; one instance for there remains; 32 instances for there has remained; three instances for there remain; and 24 instances for there have remained. Moreover, we encountered instances where the search phrases appeared in constructions that were challenging to code in terms of the number feature of the verb or noun involved, such as conditionals (e.g., Should there remain any in the Jewish community not taking appropriate or mandated actions, […]): one instance for there is; one instance for there are; four instances for there have been; and 18 instances for there remain. Additionally, we found two instances where there are was followed by collective nouns, namely Chinese and Almajirai. After excluding all these instances, our final dataset consisted of a total of 631 instances: 97 for there is; 98 for there has been; 95 for there are; 95 for there have been; 99 for there remains; 79 for there remain; 60 for there has remained; and 8 for there have remained. The searches for the target phrases were conducted between 27 December 2024 and 16 January 2025.
The trimmed data were systematically organized in an Excel sheet according to six categories: Search phrase, Verb number (singular; plural), Aspect (simple; plural), There-associated noun, Noun number (singular; plural; uncountable) and concordance line. In the There-associated noun column, we identified the specific noun associated with there; and coded its number feature in the Noun number column. For example, we categorized list as a singular noun, players as a plural noun, and peace as an uncountable noun. We then computed the cue validity based on this information; the results are summarized in Table 2. The predictive cue validity of the plural number of verb for plural number of there-associated nouns was extremely high, exceeding 95%, across all aspects (Plural-Simple: 99.43%; Plural-Perfect: 97.09%). In contrast, the cue validity of the singular number of verb for singular there-associates was far lower, falling below 50%, and was even lower for the perfect aspect (33.54%) than for the simple aspect (46.94%).
Cue validity of Verb number and Aspect for the number of there-associated noun measured by faith
All critical sentences for the visual-world eye-tracking task were 11–12 words long and employed a particular syntactic frame in which a locational preposition phrase was followed by there, verb(s), adverb, numeral, there-associated noun, and another preposition phrase, as shown in (1). The adverbs used were just, only, exactly, and precisely. The locational phrase and the there-associated noun always differed across the items. All items were recorded on Praat (Boersma & Weenink, Reference Boersma and Weenink2017) by an L1-English speaker in natural prosody.
Crucially, the predictive region was defined as the time span from the offset of the main verb to the onset of the numeral – which corresponds to the adverb (e.g., just in (1)). This region is where participants may predict the number of the upcoming there-associated noun either as singular (e.g., apple) or plural (e.g., apples) based on the morphosyntactic number cue available at the first verb region (e.g., is/are/has/have). While it is technically possible for participants to make predictions right after the auxiliary verb has or have in the case of the perfect conditions (e.g., (1b), (1d)), we used the same adverb region for the predictive region across all four conditions in order to control its duration. The average duration of this period for all critical items was approximately 583 ms. To ensure that the auditory stimuli had comparable predictive region durations across the four conditions, we conducted a linear regression analysis with Verb number and Aspect as independent variables. The results showed no significant main or interaction effects of Verb number and Aspect (all ps > .1), indicating that the predictive region durations did not differ across the four conditions.
For each sentence, there was also a visual scene presented to participants, as shown in Figure 1. Each scene had clipart images of two objects in color, which served as two areas of interest in the analysis: The target (i.e., one apple for (1a) and (1b)) and its competitor (i.e., two apples for (1a) and (1b)).
Example visual scene in the visual-world eye-tracking task.
4.3. Procedure
The participants completed five tasks in the following sequence: A Cloze test (Brown, Reference Brown1980) as a measure of English proficiency, a visual-world eye-tracking task, a fill-in-the-blank task, a language background questionnaire (Supplementary Appendix S2), and a working memory task. The Cloze test was administered remotely, at a location of the participants’ choice, approximately 1 week prior to their visit to the eye-tracker lab. This test involved a passage with 50 blank spaces that participants were required to fill in with the correct words. With the exception of the Cloze test, all tasks were conducted in the eye-tracker lab. The results of the fill-in-the-blank task and the working memory task are not pertinent to this study and will not be reported.
The creation of the visual-world eye-tracking task was done using the SR Research Experiment Builder software. Participants completed this task on a 15.6-inch screen with a resolution of 1,440 × 900 pixels; and their eye movements were recorded using an EyeLink Portable Duo eye tracker with a sampling rate of 1,000 Hz.
To start the task, participants completed eye calibration procedures using a 13-point array shown on the computer screen. During the task, they were asked to look at the pictures on the screen while listening to the English sentences, and then, after each sentence, to click on the picture that best matched what they had heard. The start of each trial was signaled by the appearance of a fixation cross, which remained on screen for 1,000 ms. This was followed by a 3,000 ms preview of the visual scene, after which the audio stimuli were played through the computer speakers. The task protocol began with a practice session with three trial runs. In the main session, the recorded sentences and corresponding visual scenes were presented in a pseudo-randomized order. Furthermore, the positions of the image objects were counterbalanced across trials, so the target image appeared equally often on the left and right sides of the screen.
4.4. Analysis
Prior to analyzing the eye-tracking data, we first excluded the trials where the participants provided incorrect mouse-click responses (L1-English: 0%; L2-English: 0.14%). Next, we transformed the fixation counts in 20-ms time binsFootnote 3 into empirical logits (elogits). Lastly, we computed a target advantage score by taking the difference between the elogit-transformed fixations for the target and the competitor.
For efficient interpretation of the data, the predictive region was redefined. The offset of this region was reset at 0 ms, corresponding to the onset of the numeral before the there-associated noun; additionally, a 200 ms offset was applied to account for the latency involved in planning and executing eye movements (Matin et al., Reference Matin, Shao and Boff1993; e.g., Grüter & Rohde, Reference Grüter and Rohde2021; Mitsugi & MacWhinney, Reference Mitsugi and Macwhinney2016). The predictive region’s starting point, or onset, was set at the end of the main verb.
To analyze eye-tracking data, this study utilized two complementary statistical models: a growth curve analysis model (e.g., Henry et al., Reference Henry, Jackson and Hopp2022) and a generalized additive mixed-effects model (GAMM; e.g., Rusk et al., Reference Rusk, Paradis and Järvikivi2020; Wieling, Reference Wieling2018).Footnote 4 Through the growth curve analysis, we were able to test any statistical interaction effects among the factors of interest over time, which is not feasible with GAMMs. Through the GAMMs, we could explore a possible source of such interaction effects and pinpoint the exact onset point when participants’ gaze patterns toward the target and competitor began to diverge in each condition, as revealed by the resulting difference plots, which is a level of detail that growth curve analysis is unable to offer. Our strategic deployment of these two complementary analysis methods thus enabled us to delve deeper into the complexities of predictive patterns than we could have by relying on either method alone, providing greater statistical robustness and revealing temporal characteristics.
To begin our analysis, we constructed growth curve analysis models on the target advantage scores during the predictive region. The model for RQ 1 included Group (L1; L2), Verb number (Singular; Plural), and Aspect (Simple; Perfect) as fixed effects. To address RQ2, we first divided the L2 participants into two proficiency groups using a median split (score of 35), excluding five participants who scored exactly 35. This resulted in two groups: Higher (n = 14), with proficiency scores of 41.07 on average (SD = 7.06), and Lower (n = 17), with proficiency scores of 26.47 on average (SD = 3.77). We then divided the same 31 L2 participants into two groups based on their immersion experience: those with immersion experience (i.e., the Yes group; n = 12) and those with no immersion experience (i.e., the No group; n = 19). On average, the Yes group had an immersion experience of 27.58 months (SD = 24.46). Next, we constructed two models: one model including Proficiency group (Higher; Lower), Verb number, and Aspect as fixed effects; and another model including Immersion experience, Verb number, and Aspect as fixed effects. The two fixed effects were contrast-coded, with −0.5 assigned to L1, Higher, Yes, Singular, and Simple, and 0.5 assigned to L2, Lower, No, Plural, and Perfect, and then they were centered. Consistent with the standard practice for accurately modeling dynamic, time-dependent changes in data, we employed second-order orthogonal polynomials (e.g., Henry et al., Reference Henry, Jackson and Hopp2022). The linear time term, which represents the first-order polynomial, captures the overall trend in fixations, whether it is increasing or decreasing. The quadratic time term, which represents the second-order polynomial, accounts for the rate of change in this trend, including any acceleration, deceleration, or changes in curvature. Initially, all our models incorporated Participant and Item as random effects, but we encountered convergence issues. Following the approaches suggested by Barr et al. (Reference Barr, Levy, Scheepers and Tily2013) and Bates et al. (Reference Bates, Maechler, Bolker and Walker2015), we simplified the models; and in the case of the model for the entire dataset, we optimized it using the “bobyqa” optimizer when necessary. The final model formulas are listed below the results table in the “Results” section for readers’ reference.
Building on the findings from the growth curve analyses, which confirmed the statistical effects of the tested factors and their interaction, we developed GAMMs using the same datasets. The central purpose of constructing the GAMMs in this study was to create difference plots which are able to reveal time-related differences in fixation patterns between the target and competitor in each condition of our interest. All GAMMs included the elogit-transformed fixations within the predictive region as the dependent variable. Given the unique manner in which GAMMs encode critical factors, compared to standard regression models, we created a new composite variable termed Hyperparameter, following previous studies (e.g., Rusk et al., Reference Rusk, Paradis and Järvikivi2020; Wieling, Reference Wieling2018). For RQ 1, this Hyperparameter was created by crossing Group, Verb number, Aspect, and Fixation area, and included as a fixed effect. For RQ2, the Hyperparameter was created by crossing Proficiency group or Immersion experience, Verb number, and Aspect. Like conventional mixed-effects regression models, a GAMM model can also incorporate random effects for Participant and Item. However, all our full models failed to converge, and we had to remove the Item effect.
GAMMs produce two types of output: Parametric coefficients and smooth terms. The parametric coefficients represent the average fixations for each condition, but since we are interested in time-related processing patterns, these values are not relevant to our discussion. The smooth terms reveal how each condition affects fixations over time. A significant p-value for a smooth term indicates a time-related effect but does not show the specific pattern. To gain a deeper understanding of the pattern, it is essential to visually inspect the data through difference plots generated by the GAMM (e.g., Ito & Knoeferle, Reference Ito and Knoeferle2023; Rusk et al., Reference Rusk, Paradis and Järvikivi2020; Wieling, Reference Wieling2018). We will focus on these plots to discuss timing variations, while providing all of the statistical results in Supplementary Appendices S3–S5. In terms of interpreting these plots, the x-axis denotes the time during the predictive region, and the y-axis illustrates the estimated difference in elogit-transformed fixations between the target and competitor. A positive y-axis value means that more fixations were directed toward the target compared to the competitor, whereas a negative value indicates the reverse. The difference plots feature vertical dashed lines that identify the points in time when the target and competitor fixation patterns are statistically different.
R 4.5.1 (R Core Team, 2025) was used to analyze our entire dataset, with the growth curve analysis models conducted using the “lme4” package (version 1.1.37; Bates et al., Reference Bates, Maechler, Bolker and Walker2015) and the GAMMs implemented via the “itsadug” package (version 2.4.1; van Rij et al., Reference van Rij, Wieling, Baayen and van Rijn2022).
5. Results
5.1. Effects of cue validity on prediction
As depicted in Figure 2, both L1 and L2 groups demonstrated a stronger tendency to fixate on the Target compared to the Competitor during the predictive region across all conditions. However, the speed at which this tendency emerged was modulated by the validity of the predictive cue. When Plural verbs were heard, which had high predictive cue validity, both groups showed robust and rapid evidence of predictive processing of the upcoming there-associated noun, independent of Aspect. The prediction of both groups was delayed when Singular verbs were encountered, with a more pronounced delay observed when the Aspect was Perfect compared to when it was Simple, the result which is consistent with the weakened cue validity that characterizes these conditions. However, it should be noted that both groups tended to direct their gaze towards a plural object from the beginning of the audio stimuli. In fact, this pattern aligns with the previously documented plural object preference (Brown et al., Reference Brown, Fox and Strand2022; Rusk et al., Reference Rusk, Paradis and Järvikivi2020), which will be discussed further in Section 6.
Elogit-transformed fixations on the target versus competitor by Group, Verb number, and Aspect.
Notes. The shaded area around the lines indicates 95% confidence intervals. The section bounded by dotted lines represents the predictive region, whose end point is offset by 200 ms.
The growth curve analysis on elogit-transformed fixations, summarized in Table 3, showed significant effects of both linear and quadratic time terms (all ps < .001), indicating that fixations changed over time and that the rate of these changes also varied during the predictive region. The analysis also exhibited significant effects of Group on the linear time term (p < .001), Verb number on the intercept (p = .001), the linear time term (p < .001), and the quadratic time term (p = .008), and Aspect on both the linear term and the quadratic term (all ps < .001). There was also a significant interaction between Group and Verb number on the quadratic time term (p = .001), between Group and Aspect on the linear time term (p < .001), and between Verb number and Aspect on the linear time term (p < .001). Notably, a significant three-way interaction effect of Group, Verb number, and Aspect (p = .034) was found on the linear term, thus indicating that the two groups’ fixation patterns changed over time in a way that depended on both Verb number and Aspect.
Output from the growth curve analysis model for all data
Notes: Model formula: lmer(Target advantage score ~ (poly1 + poly2) × Group × Verb number × Aspect + (1 + Verb number × Aspect | Participant) + (1 + Group | Item), control = lmerControl(optimizer = “bobyqa”)). Effect sizes: R 2 m = 0.04; R 2 c = 0.35.
To further investigate the time-related interaction effects among the three factors of interest, we conducted a GAMM analysis, which enabled us to produce difference plots that visualize the changes over time. The statistical results of this analysis are provided in Supplementary Appendix S3. The results of our pair-wise comparisons between Target and Competitor in the L1 group indicated significant differences in all four conditions (Figure 3). The differences emerged between −235 ms and 200 ms in the Singular-Simple condition, between 192 ms and 200 ms in the Singular-Perfect condition, and between −583 ms and 200 ms in both the Plural-Simple and Plural-Perfect conditions. Similarly, the results obtained from the L2 group showed significant differences in all four conditions, with the differences arising between −440 ms and 200 ms in the Singular-Simple condition, between −401 ms and 200 ms in the Singular-Perfect condition, and between −583 ms and 200 ms in both the Plural-Simple and Plural-Perfect conditions.
Difference between the elogit-transformed fixations on the target and competitor over time during the predictive region by Group, Verb number, and Aspect in all data.
Note. Vertical dashed lines indicate significant differences between the elogit-transformed fixations on the target and competitor.
Overall, the results show that in both L1 and L2 groups, the predictive patterns emerged in the following temporal order: The two Plural conditions arose first, followed by the Singular-Simple condition, and finally the Singular-Perfect condition. These patterns are consistent with the relative strengths of cue validity.
5.2. Effects of L2 immersion experience and L2 proficiency on second language prediction
Table 4 summarizes the results from the growth curve analysis of the L2 learners’ data where the immersion factor was incorporated. The results showed statistically significant effects of the linear term (p < .001) and the quadratic term (p = .005). Immersion was also found to have significant effects on the intercept and the quadratic term over time (all ps < .001). Additionally, Verb number had significant effects on the intercept (p = .012) as well as the linear term (p < .001), and Aspect had significant effects on both the linear term (p = .004) and the quadratic term (p < .001). The analysis also yielded significant interaction effects. Specifically, Immersion interacted with both Verb number and Aspect on the linear term (all ps < .001), while Verb number and Aspect interacted on both the linear term (p < .001) and quadratic term (p = .032). Crucially, there was a three-way interaction effect on the linear term (p < .001), indicating that the two L2 groups with and without immersion experience exhibited distinct patterns of prediction over time depending on Verb number and Aspect.
Output from growth curve analysis model for L2 data with the Immersion factor added
Notes: Model formula: lmer(Target advantage score ~ (poly1 + poly2) × Immersion × Verb number × Aspect + (1 + Verb number × Aspect | Participant) + (1 | Item)). R 2 m = 0.21; R 2 c = 0.43.
To further explore the three-way interaction effects found in the growth curve analysis, we employed a GAMM on the same dataset (for results, see Supplementary Appendix S4). The difference plots generated from the GAMM (Figure 4) showed distinct patterns between the immersed and non-immersed L2 learners. The immersed group demonstrated strong predictive patterns throughout the predictive region (−583 ms to 200 ms) across all four conditions. The non-immersed group showed predictive processing in the two Plural conditions (starting at −583 ms) and the Singular-Simple condition (starting at 152 ms) but no evidence of prediction in the Singular-Perfect condition. The results suggest an effect of immersion experience on L2 prediction, as well as its interaction with cue validity. Notably, non-immersed learners exhibited slower or no predictive processing when cue validity was weakest.
Difference between the elogit-transformed fixations on the target and competitor over time during the predictive region by Immersion experience, Verb number, and Aspect in L2 data.
Note. Vertical dashed lines indicate significant differences between the elogit-transformed fixations on the target and competitor. “Yes”: Immersed group; “No”: Non-immersed group.
The growth curve analysis performed on our L2 dataset, with Proficiency included as a fixed factor, is outlined in Table 5. This analysis showed significant effects of the linear term (p < .001) and the quadratic term (p = .002). Also shown were significant effects of Proficiency on the quadratic term (p = .006), Verb number on the linear term (p < .001), and Aspect on both the linear term (p = .010) and the quadratic term (p < .001). Furthermore, there was a significant interaction effect on the linear term between Proficiency and Verb number (p < .001), Proficiency and Aspect (p = .016), and Verb number and Aspect (p < .001). Additionally, a significant three-way interaction effect involving Proficiency, Verb number, and Aspect emerged on the quadratic term (p = .013), suggesting that fixation patterns of the two proficiency groups over time differed based on both Verb number and Aspect. However, it has been noted that without significant effects from the intercept and linear term, interpreting the quadratic term becomes problematic and requires caution (Henry et al., Reference Henry, Jackson and Hopp2022).
Output from growth curve analysis model for L2 data with the Proficiency factor added
Notes: Model formula: lmer(Target advantage score ~ (poly1 + poly2) × Proficiency × Verb number × Aspect + (1 + Verb number × Aspect | Participant) + (1 + Proficiency | Item)). R 2 m = 0.03; R 2 c = 0.49.
To delve deeper into the statistical effects identified in the growth curve analysis, we developed a GAMM using the same dataset (for results, see Supplementary Appendix S5). The difference plots produced from the GAMM (Figure 5) revealed consistent trends in both the Higher and Lower proficiency groups, mirroring the observations made for the entire group of L2 participants. Both proficiency groups exhibited strong predictive patterns throughout the predictive region from −583 ms to 200 ms for the Plural conditions, irrespective of aspect; and demonstrated faster predictions in the Singular-Simple condition (Higher: −369 ms; Lower: −61 ms) compared to the Singular-Perfect condition (Higher: 57 ms; Lower: −53 ms). However, an unexpected group difference was also observed: the Lower group showed faster prediction in the Singular-Perfect condition, while the Higher group exhibited faster prediction in the Singular-Simple condition. These results seem to suggest that in the context of prediction of the grammatical number of there-associated nouns, proficiency may not be a key factor.
Difference between the elogit-transformed fixations on the target and competitor over time during the predictive region by Proficiency, Verb number, and Aspect in L2 data.
Note. Vertical dashed lines indicate significant differences between the elogit-transformed fixations on the target and competitor. “Higher”: Higher proficiency group; “Lower”: Lower proficiency group.
6. Discussion and conclusion
The results of this study lend support to the utility-based account of prediction, demonstrating its dynamic and adaptive nature, as well as underscoring the significance of linguistic and individual factors that shape the prediction process. In addition, the finding that the L2 learners immersed in the target language were able to make predictions using morphosyntactic information during the short predictive region aligns with previous studies showing that L2 learners have a sound ability to generate expectations (Hwang & Kim, Reference Hwang and Kim2025). As Kaan (Reference Kaan2014) argued, the underlying mechanisms of prediction and how they are used seem to be similar between L1 speakers and L2 learners; the key difference lies in the factors that trigger and shape these predictive processes.
Regarding cue validity, our analysis showed that this factor affected the timing of predictions in both L1 and L2 groups, who exhibited slower predictions as cue validity decreased. Singular verbs with the Perfect aspect triggered the slowest predictions, followed by Singular verbs with the Simple aspect, and then Plural verbs with either aspect. However, the results from the Perfect conditions should be interpreted with caution, as our visual inspection (Figure 3) revealed the two groups’ general preference for a plural object from the start of the audio stimulus presentation. This preference has also been reported in previous visual-world eye-tracking studies (Brown et al., Reference Brown, Fox and Strand2022; Rusk et al., Reference Rusk, Paradis and Järvikivi2020). For example, Brown et al.’s L1-English participants fixated on plural objects more frequently than singular ones before they heard the target, regardless of whether they were presented with the phrase where is the or where are the. This preference can be explained by the inherent visual complexity of plural objects, which is likely to draw more attention from human observers.
Importantly, this study is the first to examine the there-expletive construction in a predictive processing context. By taking this novel focus, the study extended prior work on the integrative processing of subject–verb number agreement to explore how verb number cues influence the prediction of there-associated nouns. Our findings align with some previous studies on integrative processing, which have shown that L2 learners are sensitive to number agreement (e.g., Cheng et al., Reference Cheng, Rothman and Cunnings2022; Lim & Christianson, Reference Lim and Christianson2015; Wen et al., Reference Wen, Miyao, Takeda, Chu, Schwartz, Franich, Iserman and Keil2010) and that they seem to adjust their anticipations of the number of there-associated nouns to facilitate efficient processing (e.g., Tamura et al., Reference Tamura, Fukuta, Nishimura and Kato2023).
When it comes to individual differences, this study found that immersion, rather than proficiency, plays a more significant role in L2 predictive processing. Specifically, immersed L2 learners demonstrated predictive processing across all conditions, whereas non-immersed learners exhibited predictive processing only in certain conditions, such as Singular-Simple, Plural-Simple, and Plural-Perfect, but not Singular-Perfect. In contrast, both higher and lower proficiency learners showed evidence of prediction. This result seems to indicate that naturalistic exposure is a more influential factor in L2 learners’ successful predictive processing. However, Hwang and Kim (Reference Hwang and Kim2025) present conflicting evidence suggesting that proficiency also plays a crucial role. Using a visual-world eye-tracking task, the researchers found that only L1-Korean L2-English learners with higher proficiency, but not those with lower proficiency, were able to anticipate upcoming object referents based on the verb’s semantic information (e.g., The doctor opened the {box/*chair}). The inconsistencies in these findings may stem from differences in the target linguistic phenomenon (i.e., number agreement versus selectional restriction of verbs) and the type of linguistic information needed to be processed (i.e., morphosyntactic versus semantic). Future research should systematically control and vary these factors to provide a more precise understanding of L2 prediction.
Furthermore, our results shed light on the complex interaction between cue validity and individual differences. Contrary to expectations that greater immersion/proficiency would heighten sensitivity to cue validity and yield more L1-like prediction, the results presented a somewhat different pattern. Both higher- and lower-proficiency learners showed L1-like predictive behaviors as they were slower to predict when cue validity decreased (see also Table 5, where no interpretable interaction was found between Proficiency, Number, and Aspect). The non-immersed learners showed a cue validity effect as well, though with a distinct pattern – predictive processing was absent when cue validity was at its lowest. However, the cue validity effect did not appear with the immersed L2 learners. They demonstrated consistently strong predictive processing across all four conditions and outpaced L1 speakers in prediction speed.
One possible, though tentative, explanation for this result is that the L2 learners’ rich linguistic experience with two languages may have supported their development of automatized language processing in both their L1 and L2. This, in turn, could have enabled them to process the target language more efficiently in real time, consistent with observations in previous studies on bilingual advantages in processing (Gardner-Chloros, Reference Gardner-Chloros2025; e.g., Costa et al., Reference Costa, Santesteban and Ivanova2006; Lee & Phillips, Reference Lee and Phillips2023). For instance, Lee and Phillips found that L2-English learners outperformed L1-English speakers in a speeded acceptability judgment task, demonstrating greater sensitivity to number disagreement in sentences where the subject and verb disagree in number, but the intervenor and verb agree (e.g., *The artist with the tall sculptures are very talented). The researchers argue that this is due to the fact that “the learners have additional control of the L2 that helps them overcome the interference and avoid judgment errors” (p. 11). According to them, high proficiency and extensive experience with the L2 likely contribute to the learners’ ability to take advantage of this additional control. However, this explanation of bilingual advantages remains speculative, and it would be beneficial for future research to test this possibility by comparing learners with rich linguistic experience in both their L1 and L2 with monolingual speakers of each language.
Interestingly, but unexpectedly, lower proficiency learners showed faster prediction in the Singular-Perfect condition than their higher proficiency counterparts. This seemingly suggests an advantage for lower proficiency, but this interpretation is undermined by the opposite pattern observed in the Singular-Simple condition, where higher proficiency learners were faster. The inconsistency may be attributed to the difference in proficiency score variability between the two proficiency groups, with the higher proficiency group exhibiting a broader range of scores (SD = 7.06) compared to the lower proficiency group (SD = 3.77). That is, the median split at a score of 35, along with the resulting imbalance in group variability, may have limited our ability to capture possible proficiency effects accurately.
Finally, we acknowledge that our experimental stimuli have certain limitations. Regarding the use of be and remain as the main verbs, combining these verb types in the analysis could have obscured potentially meaningful effects. To address this concern, we ran an additional growth curve analysis. This analysis showed that adding Verb type as a fixed effect significantly improved the fit of the growth curve model built on the full dataset reported in Table 3 (χ 2 (24) = 203.21, p < .001). Further visual inspection suggested that the observed improvement may be driven by L1-English speakers, who demonstrated stronger and faster prediction in the Singular-Simple condition when the verb was be rather than remain. This result highlights the importance of considering the differences between the verbs being tested in terms of morphological inflections (e.g., suppletion [e.g., be: is–are] versus regular inflection [e.g., remain: remains–remain]), frequency, and cue validity (e.g., Table 2), which should be addressed in future research.
While we utilized cue validity estimates from the corpus to inform our design, our stimuli were not directly derived from the corpus. Instead, we employed controlled sentences to minimize potential confounding variables. As a result, the link between our corpus-based approach and experimental stimuli is not as robust as it could be. We believe that further research incorporating there-expletive sentences from the corpus would offer more nuanced insights. As for the placement of just within our stimuli, an anonymous reviewer noted that has just been sounds more natural than has been just. Although we chose the latter to ensure comparability across the critical conditions, with just appearing before the numeral, we acknowledge that the placement of just may have introduced noise into the experiment. Moreover, using the same adverb as the predictive region across the critical conditions to control its duration may have created another confound because in the perfect aspect conditions, predictions could potentially start right after the auxiliary verb has/have. Future work should address all these issues.
In conclusion, this study makes both theoretical and methodological contributions. Theoretically, our findings shed light on the underlying mechanisms guiding L1 speakers’ and L2 learners’ predictive processing, as well as the factors that modulate these mechanisms. Notably, we found that cue validity affects both L1 and L2 prediction, and that the underlying mechanisms are fundamentally similar between L1 and L2 predictive processing, while L2 learners show individual differences. Methodologically, we employed a more rigorous approach, combining growth curve analysis and GAMMs to provide a more fine-grained analysis of the eye-tracking data. This study’s theoretical and methodological approaches are expected to provide deeper insights into predictive processing mechanisms.
Supplementary material
The supplementary material for this article can be found at https://osf.io/nwev5.
Data availability statement
All of our materials (Supplementary Appendices S1–S2), supplementary results (Supplementary Appendices S3–S5), data, and analysis scripts are available at https://osf.io/nwev5.
Acknowledgments
This study was funded by the Direct Grant for Research awarded to the first author by the Faculty of Arts, The Chinese University of Hong Kong (project code: 4051221). We would like to thank Jihyeon Baeg, Joonhee Kim, Yukyung Kim, and Hyunjoo Lee for the help they provided during the data collection phase. We are also indebted to Amber Camp and Pui Yee “Ruby” Lee, who helped create the audio and visual stimuli needed for our eye-tracking task.
Open access
