Highlights
-
• Adult L2 learners show abstract structural priming in real-time comprehension.
-
• Prediction error drives structural priming in L2 comprehension.
-
• Grammatical knowledge constrains prediction-error-driven L2 structural priming.
-
• Prediction-error-based priming in L2 comprehension persists into later production.
-
• Prediction (error) constitutes a learning mechanism in late L2 acquisition.
Introduction
In life, we learn best from our mistakes, and learning a second language (L2) is no exception in this regard. What’s more, L2 learners frequently make mistakes, so that they have ample opportunity to learn from them in principle. However, learners can only learn from mistakes provided they receive appropriate feedback, be it in the form of explicit correction or some response that affords learners the opportunity to notice the gap between their knowledge of the L2 and the target. Although it is uncontested that error-driven learning must constitute a key learning mechanism in SLA (for review, VanPatten et al., Reference VanPatten, Keating and Wulff2020), it is an open question how consistently feedback to learners’ errors is provided and whether learners actually learn from feedback to their errors, in particular when these errors are grammatical mistakes that typically do not disrupt or impede understanding and communication.
Recent advances in language processing research suggest that error-driven learning may not need to rely on explicit correction or noticing of the error; instead, it may be an integral characteristic of language comprehension in that listeners and readers continuously make predictions about how a sentence or utterance continues, including its grammatical structure (for review, Kuperberg & Jaeger, Reference Kuperberg and Jaeger2016). For instance, when encountering a sentence with a ditransitive verb that takes two complements (e.g., “The man will send …”), listeners will predict either a double-object structure (DO, “the woman the letter”) or a prepositional-object structure (PO, “the letter to the woman”) to follow, depending on the probabilistic verb biases that make one structure more likely to occur than the other. When these predictions are not met and listeners encounter the opposite structure, they experience a prediction error that highlights the gap between their expectations and the input, affording them an opportunity to learn from such implicit feedback via structural priming.
Structural priming refers to the tendency of language users to reuse the syntactic structures of sentences they have recently encountered (for review, Mahowald et al., Reference Mahowald, James, Futrell and Gibson2016). According to implicit learning models of priming (Bock & Griffin, Reference Bock and Griffin2000; Chang et al., Reference Chang, Dell and Bock2006; Dell & Chang, Reference Dell and Chang2014), priming arises precisely as a consequence of learners experiencing prediction errors when the prime structure they encounter in the input does not conform to the expectations they have had about it. As a result, they will adjust their knowledge and processing to match the unexpected input, i.e., update their expectations to minimize future prediction errors. In these models, priming is driven by an error-driven learning mechanism. Several characteristics of structural priming support error-based learning models. First, priming effects for lower-frequency structures, e.g., PO orders in English, are more robust than for more frequent structures, e.g., DO orders in English (e.g., Jaeger & Snider, Reference Jaeger and Snider2013). Second, PO structures with verbs that typically occur with DO structures (e.g., “pay”), show larger priming effects than PO structures following verbs that preferentially occur in this structure (e.g., “send”; Bresnan et al., Reference Bresnan, Cueni, Nikitina, Baayen, Bouma, Krämer and Zwarts2007; Jaeger & Snider, Reference Jaeger and Snider2013). Both inverse frequency effects according to the frequency distribution of structures and inverse preference effects according to different verb biases suggest that the magnitude of the priming effect is the inverse of the size of the prediction error, also known as the surprisal, experienced in processing the prime (for review; Fazekas et al., Reference Fazekas, Sala and Pine2024). Third, structural priming persists beyond immediate trial-to-trial priming across many trials and cumulates over the course of an experiment and beyond, leading to longer-term syntactic adaptation (for review, Kaan & Chun, Reference Kaan and Chun2018). The key evidence in favor of error-driven implicit learning models comes from studies observing online effects of prediction error on priming in comprehension. Several visual world eye-tracking studies on structural priming in comprehension, conducted among child L1 learners and adult speakers, found that a less expected prime structure leads to larger priming effects during the processing of the target (e.g., Chen et al., Reference Chen, Wang and Hartsuiker2022; Fine & Jaeger, Reference Fine and Jaeger2013). Such findings demonstrate that prediction error drives priming, and clearly support error-driven implicit learning theories of structural priming.
It is an open question whether structural priming among adult L2 learners can be captured as an equivalent form of implicit learning. On the one hand, some studies on production priming have extended error-based models of priming from monolinguals to bilinguals on the basis of comparable priming patterns seen in within-L2 priming (for review, Jackson, Reference Jackson2018) and in cross-linguistic priming (Khoe et al., Reference Khoe, Tsoukala, Kootstra and Frank2023; for review, van Gompel & Arai, Reference van Gompel and Arai2018). For instance, structural priming in an L2 often shows inverse frequency effects as does L1 priming (e.g., Jackson & Hopp, Reference Jackson and Hopp2020), and some studies also report longer-term priming (e.g., Nitschke et al., Reference Nitschke, Kidd and Serratrice2010; Wei et al., Reference Wei, Boland, Cai, Yuan and Wang2019). On the other hand, there are also differences in priming patterns between L2 learners and monolinguals that challenge error-based learning models. L2 learners often show only immediate, short-term structural priming, yet no or attenuated longer-term priming or syntactic adaptation (e.g., Jackson & Hopp, Reference Jackson and Hopp2020; Jackson & Ruf, Reference Jackson and Ruf2017; yet see Hwang & Shin, Reference Hwang and Shin2019). Finding that L2 learners show less longer-term priming suggests that structural priming in an L2 may rely more on the explicit memory of primes or the local reactivation of lexical-syntactic nodes (for discussion, see Jackson & Hopp, Reference Jackson and Hopp2020).
These differences in L2 compared to L1 structural priming may follow from independent factors characteristic of L2 sentence comprehension (for review, Hopp, Reference Hopp, Kaan and Grüter2021). Across many studies, adult L2 learners have been found to make grammatical predictions to a lesser degree in sentence comprehension than L1 users (e.g., Grüter et al., Reference Grüter, Lau and Ling2020; Kaan, Reference Kaan2014; for review, Bovolenta & Marsden, Reference Bovolenta and Marsden2022; Schlenter, Reference Schlenter2023). In turn, L2 learners experience fewer prediction errors, so that they may make recourse to different mechanisms in L2 structural priming, which then potentially constrains grammatical learning in an L2 (e.g., Hopp, Reference Hopp, Kaan and Grüter2021; Kaan et al., Reference Kaan, Futch, Fuertes, Mujcinovic and de la Fuente2019; Phillips & Ehrenhofer, Reference Phillips and Ehrenhofer2015).
Against this backdrop, this study tests if prediction errors occasioned by verb bias in the comprehension of the primes affect the magnitude of L2 structural priming, and, if found, such priming extends beyond the priming phase in the form of longer-term priming. This way, we investigate whether error-based implicit learning models extend to L2 learners. If so, differences in priming patterns between L1 and L2 speakers cannot be reduced to differences in the availability of the underlying priming mechanisms, specifically, learning from prediction error. Few studies have directly tested whether prediction error drives L2 structural priming in similar ways as it appears to underlie L1 structural priming (but see Montero-Melis & Jaeger, Reference Montero-Melis and Jaeger2020). Especially studies on structural priming in L2 from comprehension to comprehension are few and far between. The few studies that have tested structural priming in L2 comprehension suggest that priming is lexically mediated but does not survive when prime and target verbs are different (e.g., Fujita & Cunnings, Reference Fujita and Cunnings2021; Weber & Indefrey, Reference Weber and Indefrey2009; Wei et al., Reference Wei, Boland, Brennan, Yuan, Wang and Zhang2018, Reference Wei, Boland, Cai, Yuan and Wang2019; but see Wei et al., Reference Wei, Boland, Zhang, Yang and Yuan2023).
We use visual world eye-tracking to examine (a) whether adult L2 learners show abstract structural priming in comprehension and whether L2 structural priming is affected by prediction error, operationalized in terms of verb bias. We also investigate (b) whether structural priming occasioned by verb bias violations, if found, feeds into longer-term learning in that it leads to changes in (unprimed) L2 production. Finally, to delineate the consequences of prediction error via priming for L2 learning, we test (c) how verb biases in the prime sentences interact with verb biases in the target sentences.
Effects of prime verb bias: Prediction error and structural priming
One way to examine the role of prediction error in priming is by studying inverse preference effects in priming in terms of surprisal. In priming from comprehension to production, Jaeger and Snider (Reference Jaeger and Snider2013) find that the magnitude of structural priming depends on the degree of surprisal of the structure of the prime sentence relative to its verb’s structural biases. Specifically, PO primes with verbs biased to the DO structure occasioned more PO targets in both spoken and written production than PO primes with PO-biased verbs. Peter et al. (Reference Peter, Chang, Pine, Blything and Rowland2015) replicated such verb bias effects on structural priming in the production of 3- to 6-year-old English-speaking children and adults (see also Fazekas et al., Reference Fazekas, Jessop, Pine and Rowland2020), and Bernolet and Hartsuiker (Reference Bernolet and Hartsuiker2010) reported comparable verb bias effects on the priming of DO structures among Dutch adults. Taken together, these studies suggest that inverse preference effects characterize structural priming in both L1 learners and L1 mature speakers across languages. Importantly, in these studies, the inverse preference effect of verb bias interacted with the inverse frequency effect of the structure in the language. The English speakers in Jaeger and Snider (Reference Jaeger and Snider2013) demonstrated significant priming and verb bias effects only for the overall less frequent PO structure in English, while Dutch speakers showed priming for the overall less frequent DO dative in Dutch (Bernolet & Hartsuiker, Reference Bernolet and Hartsuiker2010). When seen in conjunction, the inverse preference and frequency effects that generate surprisal support the idea of implicit learning models of priming are driven by prediction error.
In comprehension-to-comprehension priming, inverse preference effects on priming can be directly linked to prediction and prediction error in visual world eye-tracking studies. Here, the degree of surprisal from the prime structure can be observed in the eye movements during the processing of the target sentences. In a seminal study, Fine and Jaeger (Reference Fine and Jaeger2013) reanalyzed results from a priming study on DO versus PO-datives with L1 English children and adults by Thothathiri and Snedeker (Reference Thothathiri and Snedeker2008). In this study, participants acted out instructions to move objects (e.g., “Give the fish the dog bone” versus “Give the fishbowl to the bear”) while their eye movements were being tracked. Both 3- and 4-year-old children and adults showed priming by launching more looks in the ambiguous region (“the fish”) to the postverbal referent corresponding to the structure of the prime sentence. Fine and Jaeger (Reference Fine and Jaeger2013) independently assessed the verb biases of English verbs used in Thothathiri and Snedeker (Reference Thothathiri and Snedeker2008) in a written production study in which adults needed to complete sentence fragments (e.g., “A nice bookstore clerk gave …”). When reanalyzing the priming data from Thothathiri and Snedeker (Reference Thothathiri and Snedeker2008) by factoring in the verb biases of the primes, they found the strength of structural priming interacted with the structural biases of the verbs in the prime sentences: Participants were the more strongly primed by the structure of the prime sentence in launching looks to the postverbal referents, the more the structure of the prime sentence went against its verb’s bias, so when its surprisal was higher. Since these effects were found in anticipatory looks to the postverbal referent in the target sentence, the inverse preference effects clearly reflected prediction arising from the prior processing of the prime sentence. In a similar design, Chen et al. (Reference Chen, Wang and Hartsuiker2022) found comparable verb bias effects on structural priming in Mandarin, in that the strength of structural priming was inversely related to the structural biases of verbs in the prime sentences, as assessed in an independent sentence completion study with Mandarin speakers. Adapting the design to Dutch, Chen and Hartsuiker (Reference Chen and Hartsuiker2023) failed to elicit priming effects and verb bias effects with Dutch adults. However, they argue that the artificially slow speech rate and the high predictability of the target structure may have reduced priming. In sum, visual world eye-tracking studies on priming in L1 comprehension demonstrate direct effects of prediction error on structural priming across languages among L1 speakers. In Experiment 1, the present study directly tests effects of verb bias in the comprehension-to-comprehension priming among L2 users. This way, our study examines whether prediction error constitutes a mechanism of structural priming also in a later-learnt L2.
Longer-term effects of verb biases: Structural priming from comprehension to production
To investigate whether structural priming according to prediction errors constitutes implicit learning, this study also addresses whether structural priming in L2 comprehension leads to longer-term effects in L2 production, seeing that the term learning typically refers to changes in the output that L2 learners produce. Insights into the longevity of structural priming in production come from cumulative priming, that is, increases in priming magnitude over the course of a priming task, as well as from long-term priming in terms of priming effects persisting into (unprimed) language use after the priming task. L1 users demonstrate both cumulative effects of structural priming and long-term effects, which have been shown to persist for days, weeks and even months after the priming task (e.g., Heyselaar & Segaert, Reference Heyselaar and Segaert2022; Kaschak et al., Reference Kaschak, Kutta and Schatschneider2011). For L2 structural priming, relatively few L1 or L2 priming studies consider cumulative or long-term effects of structural priming beyond the immediate priming phase of an experiment (though see McDonough, Reference McDonough2006; Ruf, Reference Ruf2011). For datives, Kaan and Chun (Reference Kaan and Chun2018) report cumulative priming of DO structures in L1-Korean-L2-English speakers when they typed sentence completions to ditransitive verbs in a web-based production task. In contrast to L1 English speakers, the L2 learners did not show immediate trial-by-trial priming effects. In other studies, structural priming among intermediate to advanced-level L2 learners did not persist beyond the priming phase itself (e.g., Jackson & Hopp, Reference Jackson and Hopp2020; Jackson & Ruf, Reference Jackson and Ruf2017; but see Nitschke et al., Reference Nitschke, Kidd and Serratrice2010; Wei et al., Reference Wei, Boland, Cai, Yuan and Wang2019).
All of these studies considered production-to-production priming in that participants actively produced target sentences in response to the primes in the priming tasks. Hence, it is unclear whether the cumulative and long-term priming effects observed result from the comprehension of the primes or reflect the persistence of the primed structures in the learners’ own productions of the targets (auto-priming). Our study on comprehension-to-comprehension priming isolates priming effects coming from the primes. To our knowledge, no study on L2 structural priming has tested effects of comprehension priming to production (for L1, see Bock et al., Reference Bock, Dell, Chang and Onishi2007). Yet, this seems particularly relevant for understanding L2 learning mechanisms, as L2 acquisition occurs predominantly from the input and is not always mediated by production practice (VanPatten et al., Reference VanPatten, Keating and Wulff2020). Investigating comprehension-to-production priming allows us to directly examine how input processing during comprehension contributes to the development of productive language skills in L2 learners. In both L1 and L2 users, the effects of comprehension priming are smaller than in production priming and tend to be more lexically specific (e.g., Fujita & Cunnings, Reference Fujita and Cunnings2021; Wei et al., Reference Wei, Boland, Cai, Yuan and Wang2019). Against this backdrop, it is an open question whether structural priming in L2 comprehension will lead to learning as reflected in L2 production. To this end, the present study combines a comprehension-to-comprehension priming experiment with (unprimed) production tasks. This way, our study investigates the longevity of learning from prediction error.
Effects of target verb bias on structural priming
Finally, this study investigates the scope of prediction error and structural priming for L2 learning. To this end, we investigate how structural priming occasioned by verb biases in the prime sentences interacts with verb biases in the target sentences. Over and above verb bias effects in the prime sentences, previous studies found that verb biases in target sentences modulate structural priming in L2 learners (e.g., Kaan & Chun, Reference Kaan and Chun2018; Kootstra & Doedens, Reference Kootstra and Doedens2016). Such target verb bias effects illustrate that structural priming is affected by the learners’ knowledge of the L2, in that priming effects are mediated by learners’ structural preferences in the L2. In Experiment 2, we test how far the knowledge of L2 users affects priming by investigating whether structural priming is constrained by the grammatical options in the target sentences (see also Bernolet & Hartsuiker, Reference Bernolet, Hartsuiker, Miller, Bayram, Rothman and Serratrice2018; Hopp & Jackson, Reference Hopp and Jackson2023, for cross-linguistic priming). Specifically, we test whether structural priming from grammatical prime sentences with alternating verbs (e.g., “pay” and “send”) that allow both PO and DO structures extends to target sentences with non-alternating verbs (“donate”) that are limited to the PO structure (see also Ivanova et al., Reference Ivanova, Pickering, McLean, Costa and Branigan2012). If verb biases in the target sentences mediate priming, effects of prediction error and priming should be reduced or cancelled for non-alternating versus alternating verbs, since non-alternating verbs do not allow for structural optionality. This way, our study tests how far prediction errors interact with learners’ knowledge of grammatical constraints in the L2.
The present study
In two priming studies on L1-German-L2-English learners that combine structural priming and visual world eye-tracking in comprehension (Arai et al., Reference Arai, van Gompel and Scheepers2007; Chen et al., Reference Chen, Wang and Hartsuiker2022) as well as sentence production, we test if L2 structural priming constitutes a type of (prediction-)error-based implicit learning. We pose the following research questions:
RQ1) Does prediction error affect L2 structural priming in comprehension?
Experiment 1 investigates the effects of verb bias on structural priming among L1 German advanced learners of English to test if prediction error by virtue of the violation of a prime verb’s structural bias occasions structural priming. The primes consist of sentences with alternating verbs that are either biased to a PO or DO structure (PO-bias verbs, e.g., “send” versus DO-bias verbs, e.g., “pay”), and the target sentences have alternating verbs that do not have strong biases (equi-biased (EQ-bias) verbs, e.g., “show”). Following error-driven implicit learning models of priming (e.g., Dell & Chang, Reference Dell and Chang2014), we predict that structural priming in the target sentences is stronger if the prime structure goes against the verb’s structural bias.
RQ2) Does L2 priming in comprehension persist into production?
We also examine whether effects of structural priming in comprehension, if found, persist into a subsequent production task without primes. To this end, we compare the production of PO and DO structures between a baseline production task before the priming experiment and a posttest production task following the priming experiment. In line with implicit learning models of priming, we hypothesize that structural priming effects found in comprehension should persist into later (unprimed) production.
RQ3) Do target verb biases constrain structural priming by prediction error?
Experiment 2 examines whether the effects of prediction error as per the verb biases of alternating verbs in the prime sentences also affect the comprehension of target sentences with strongly biased verbs, namely, non-alternating verbs (e.g., “donate”). For native speakers of English, Ivanova et al. (Reference Ivanova, Pickering, McLean, Costa and Branigan2012) found that production priming effects are cancelled when the target sentences have non-alternating verbs (see their Experiments 1 and 4). This suggests that strong target verb biases restrict structural priming among native speakers. We investigate whether L2 speakers similarly recruit their knowledge of the target language to constrain structural priming.
In this study, we investigate only L2 learners, since our primary interest is to detect evidence of prediction error in within-L2 structural priming and to delineate its scope, rather than to assess relative differences in the strength of structural priming between L1 users and L2 learners (see, e.g., Fujita & Cunnings, Reference Fujita and Cunnings2021; Jackson & Hopp, Reference Jackson and Hopp2020).
Experiment 1: Prediction-error-based priming
Method
Participants
Forty-eight L2 learners of English (34 females) participated in Experiment 1. We determined the number of participants on the basis of the study by Şafak and Hopp (Reference Şafak and Hopp2023), in which a group of 36 comparable learners showed robust prediction effects in visual world eye-tracking. To account for potentially inflated effect sizes in that study, we upped the number of participants to 48. In addition, we focus the analyses on main effects rather than interactions. According to Brysbaert (Reference Brysbaert2019, p. 27; see also Brysbaert, Reference Brysbaert2021), the sample size of 48 provides a sufficient number of observations to investigate main effects and reach 90% power. All participants had normal or corrected-to-normal vision and reported having German as their L1 and English as their L2. All had acquired English in an instructional setting. They gave informed consent and received 20 Euros for taking part in the experiment. All procedures contributing to this work were covered by the laboratory’s ethics clearance by the German Society for Linguistics (DGfS Lab vote 2023-04), and they complied with the ethical standards of the relevant national and institutional committees on human experimentation and the Helsinki Declaration of 1975, as revised in 2008.
To establish proficiency levels in English, we administered the Lexical Test for Advanced Learners of English (LexTALE), a standardized vocabulary task requiring participants to make a lexical decision without any time limitation (Lemhöfer & Broersma, Reference Lemhöfer and Broersma2012). Stimulus presentation and response collection were controlled by E-Prime, Version 2.0 (Schneider et al., Reference Schneider, Eschman and Zuccolotto2002). The participants attained a mean score of 83 out of 100 (Table 1). This score corresponds to the C1 level (“advanced”) in the Common European Framework of Reference for Languages (CEFR).
Table 1. Participant characteristics in Experiments 1 and 2
Materials
Priming task. Following the design in Chen et al. (Reference Chen, Wang and Hartsuiker2022), we created 48 sets of experimental items (for the full set, see Supplementary Table S1). Each set consisted of four dative prime sentences and two dative target sentences; see (1a–d) and (2a–b) in Table 2.
Table 2. Example set of experimental items from Experiments 1 and 2
Note. In Experiment 1, both DO and PO target sentences had equi-bias alternating verbs, while in Experiment 2, DO target sentences remained the same, but PO target sentences were constructed with non-alternating verbs.
The experimental prime and target sentences contained an animate subject noun phrase, the modal verb “will,” a ditransitive verb, and two complements, one being an animate recipient and the other an inanimate theme. The prime sentences were constructed with four DO-bias alternating verbs (i.e., “pay,” “tell,” “owe” and “charge”) and four PO-bias alternating verbs (i.e., “send,” “whisper,” “transfer” and “hand”), whereas the target sentences were constructed with four EQ-bias alternating verbs (i.e., “show,” “give,” “serve” and “teach”). All prime and target verbs were selected on the basis of a sentence-completion norming study among native English speakers reported in Şafak and Hopp (Reference Şafak and Hopp2023). In this norming task, native speakers were presented with sentence fragments that included a subject noun phrase, the modal verb “will,” and a ditransitive verb. They were asked to complete these fragments by using two noun phrases as complements that were provided in parentheses. Responses were then grouped into categories: ‘DO-dative,’ ‘PO-dative’ or ‘other,’ with corresponding percentages calculated for each. A verb was categorized as alternating if both ‘DO-dative’ and ‘PO-dative’ responses surpassed 5%, and as non-alternating if one fell below this threshold. Verbs were classified as DO-bias if their frequency with a DO-dative structure exceeded that with a PO-dative structure by more than 15%, and conversely for PO-bias (for details, see Şafak & Hopp, Reference Şafak and Hopp2023, pp. 1240–1241). They were followed by either the DO-dative or the PO-dative structure. All recipients and themes were preceded by the definite article “the” in order to ensure that any preference for the DO-dative (the Recipient>Theme order), and the PO-dative structure (the Theme>Recipient order) was not confounded with differences in the definiteness status of verb complements (Bresnan et al., Reference Bresnan, Cueni, Nikitina, Baayen, Bouma, Krämer and Zwarts2007).
For each target sentence, a visual display was created using commercially available ClipArt packages. As shown in Figure 1, each display depicted three entities, i.e., an agent entity (e.g., “the tailor”), a recipient entity (e.g., “the model”) and a theme entity (e.g., “the dress”). The three entities were arranged in a triangular fashion and counterbalanced.
Figure 1. Example visual display from Experiments 1 and 2.
Note. Participants saw the visual display while listening to the corresponding DO or PO target sentence; see (2a versus 2b) and (3a versus 3b) in Table 2.
There were eight experimental conditions resulting from a 2×2×2 design with the factors prime verb bias (DO-bias versus PO-bias), prime structure (DO versus PO) and target structure (DO versus PO). The 48 sets of experimental items were distributed over eight lists according to a Latin square design, with each list containing exactly one prime-target sentence pair from each set. Each participant thus saw all eight conditions and six items per condition.
In addition to the 48 experimental items, each list included 96 fillers with a range of structures (e.g., adjunct clauses, relative clauses and passive structures). Half of the fillers involved prime-target sentence pairs with the same structure, while the other half did not. All experimental and filler items were presented in a pseudo-randomized order, with no more than two consecutive occurrences of the same type of experimental and filler items.
Baseline and posttest tasks. We designed two spoken picture description tasks following Ivanova et al. (Reference Ivanova, Pickering, McLean, Costa and Branigan2012). Each task contained 24 black and white line drawings of events involving persons and objects with the roles of agent, recipient, and theme. The positions of the agent and recipient entities were counterbalanced so that each appeared an equal number of times on the left (preceding the theme) and on the right (following the theme). Below each picture was a ditransitive verb in its bare form (Figure 2). The verbs were the same as in the priming task, and four non-alternating verbs that were semantically similar to the target verbs used in the priming task were added.
Figure 2. Example picture from the baseline task in Experiments 1 and 2.
Each task also included 24 filler picture stimuli with transitive or intransitive verbs, resulting in a total of 48 pictures per task. The pictures were either modified from previous studies (Kootstra et al., Reference Kootstra, van Hell and Dijkstra2010; Szekely et al., Reference Szekely, Jacobsen, D’Amico, Devescovi, Andonova and Herron2004) or newly constructed.
Procedure
Priming task. We conducted a structural priming task using visual world eye-tracking. Participants were seated at a distance of approximately 70 cm from a 22-inch computer screen. The movements of participants’ right eyes were recorded by an SR Research EyeLink Portable Duo at 250 Hz. The task was preceded by instructions and a practice block of four items. This was followed by a 9-point calibration and validation procedure. In the task, each item started with drift correction; then, the experimenter pressed the space bar to trigger the prime sentence. All prime sentences were presented on a single line of text in Times New Roman 22 point with black letters on a white background on the screen. When participants read the written prime sentence aloud, the experimenter pressed the space bar to move on to the target sentence. All target sentences had been recorded by a male native speaker of American English at a slow to moderate pace and were presented via two loudspeakers. While listening to the target sentence, participants looked at the corresponding visual display. All visual displays were presented 500 ms before the onset of the spoken target sentences and remained visible for 2000 ms after sentence offset. To ensure that participants listened to the target sentences attentively, a third of the filler items were followed by a written prompt that required participants to verbally describe the previous visual display. The entire task took approximately 80 minutes to complete.
Baseline and posttest tasks. Participants completed the baseline task before the priming task, and they performed the posttest task following the priming task. The tasks were administered in E-Prime 2.0 (Schneider et al., Reference Schneider, Eschman and Zuccolotto2002), and participants’ responses were recorded with a digital recorder.
The baseline task began with instructions and a practice session of three items. For each trial, a fixation cross appeared in the center of the screen for 500 ms, and then a picture stimulus was presented on the screen. Participants were instructed to describe the picture with one sentence using the verb printed below the picture. There was no time limit, but participants were encouraged to describe each picture in 10 seconds. The posttest task followed the same procedure as the baseline task, yet did not include any practice items.
Results
Baseline task
For the experimental picture stimuli in the baseline (and posttest) task, all responses were transcribed verbatim and scored as either ‘DO-dative,’ ‘PO-dative’ or ‘other’. Figure 3 shows the proportion of PO-datives by verb type; see Supplementary Figure S1 for the percentage of sentences produced for each structure by verb type (DO-bias versus PO-bias versus non-alternating versus EQ-bias) and task (baseline versus posttest). Descriptive results showed that participants had a general preference for the PO-dative structure.
Figure 3. Experiments 1 and 2. Proportions of PO-dative sentences produced by verb type and task.
Note. Error bars represent standard errors of the means.
To assess whether verb bias modulated the production of PO-dative sentences in the baseline task, we conducted a logistic mixed-effects model (see Jaeger, Reference Jaeger2008) using the lme4 package in R 4.2.2 (R Development Core Team, 2022). The model included verb type as a sum-coded fixed effect, random intercepts for participants and items, and verb type as a by-participant random slope. The dependent variable was the use of the PO-dative structure in the baseline task of Experiment 1 (see Figure 3, left panel, blue bars). The analysis showed that the production of PO-dative sentences varied according to verb type (for model output, see Supplementary Table S2). The overall proportion of PO-dative sentences was significantly lower for DO-bias verbs compared to the grand mean of all four verb types, while PO-bias verbs patterned with the general PO preference. For non-alternating verbs, the production of PO-dative sentences was significantly higher than the grand mean. These findings show that – over and above a general PO preference in their production – participants were sensitive to the verbs’ biases in English.
Priming task
We created three equal-sized areas of interest for the entities on the experimental target visual displays, labelled the agent, the recipient and the theme. Using the VWPre package in R (Porretta et al., Reference Porretta, Kyröläinen, van Rij and Järvikivi2016), we calculated the empirical logit of looks to the recipient and the theme in three time windows in 50-ms bins: the target verb (VERB), the determiner of the first postverbal argument (DETERMINER1) and the noun of the first postverbal argument (NOUN1). Following Chen et al. (Reference Chen, Wang and Hartsuiker2022, Reference Chen, Wang and Hartsuiker2023), we used the difference score of gaze probability to the recipient and to the theme (i.e., empirical logits of looks to the recipient minus looks to the theme) as the dependent variable. All time windows were aligned relative to the onset of the noun of the first postverbal argument and shifted forward 200 ms to account for the time required to initiate and implement a saccade according to spoken input (Matin et al., Reference Matin, Shao and Boff1993).
Following Chen et al. (Reference Chen, Wang and Hartsuiker2022, Reference Chen, Wang and Hartsuiker2023), we carried out both traditional time-window analyses and cluster-based permutation analyses on the empirical logit of looks to the recipient (predicting the DO-dative structure) versus looks to the theme (predicting the PO-dative structure) in three time windows. Traditional time window analyses were conducted using linear mixed-effect models (Baayen et al., Reference Baayen, Davidson and Bates2008) and the lme4 package in R (Bates et al., Reference Bates, Mächler, Bolker and Walker2015). In the first step of the eye-tracking data analysis, we conducted omnibus linear mixed-effect models that included both DO and PO target sentences. The initial models contained a maximal random-effect structures, incorporating prime structure (DO versus PO), prime verb bias (DO-bias versus PO-bias), and their interactions as sum-coded fixed effects. Additionally, random intercepts for participants and items, as well as random slopes for all fixed effects and their interactions, were included (Barr et al., Reference Barr, Levy, Scheepers and Tily2013). When a maximal model failed to converge, we simplified the random-effect structures by removing the random slopes of the fixed effects.
The results of the omnibus linear mixed-effects models are presented in Table 3 (see also Figure 4A). Results did not show any main effects, but the interaction of prime structure and prime verb bias became significant in the NOUN1 window, which disambiguates whether the target sentence would be followed by the DO-dative (e.g., “The tailor will show the model the dress”) or the PO-dative structure (e.g., “The tailor will show the dress to the model”).
Table 3. Experiments 1 and 2. Omnibus traditional time-window analyses for the eye-movement data
Note. SE, Standard error. Levels of prime structure are +1 for PO and − 1 for DO. Levels of prime verb bias are +1 for DO-bias and − 1 for PO-bias.
Figure 4. Experiment 1. Differences in gaze probability between recipient and theme by time window (empirical logits of looks to recipient minus looks to theme, by prime type (panel A), prime structure (panels B and C) or prime verb bias (panel D).
Note. Time (on x-axis) plotted with 0 ms corresponding to 200 ms after the Noun1 onset. Shaded areas denote significant clusters in cluster-based permutation analyses, including onset and offset times. The gray area around the lines represents 95% confidence intervals.
Given that significant effects emerged in the temporal region that distinguishes between target sentences with the DO-dative structure and those with the PO-dative structure, we subsequently analyzed the eye-tracking data separately by target structure to fully capture the effects of priming and verb bias. To test whether prediction error drives priming when ditransitive verbs biased to either the DO-dative or the PO-dative structure occur in the respective other structure, we made the following planned comparisons in the analysis.Footnote 1
First, to assess priming effects, we compared the DO prime and the PO prime conditions (e.g., (1a&c) versus (1b&d) in Table 2). When significant priming effects were found, we performed subsequent pairwise comparisons to establish whether prediction error, occasioned by verb bias, modulated these priming effects: For DO target sentences, comparisons were run between the DO-bias and the PO-bias DO prime conditions (e.g., (1a) versus (1c) in Table 2). For PO target sentences, we compared the DO-bias and the PO-bias PO prime conditions (e.g., (1b) versus (1d) in Table 2). The model outputs are shown in Table 4 for each comparison.
Table 4. Experiments 1 and 2. Traditional time-window analyses for the eye-movement data
Note. SE, Standard error. Levels of prime structure are +1 for PO and − 1 for DO. Levels of prime verb bias are +1 for DO-bias and − 1 for PO-bias.
To analyze the eye-movement data while controlling for family-wise error rate and autocorrelation and maintaining statistical power (Ito & Knoeferle, Reference Ito and Knoeferle2022; Stone et al., Reference Stone, Lago and Schad2021), we also conducted cluster-based permutation tests with the clusterperm.lmer function in the R package permutes (Voeten, Reference Voeten2022). We tested the main effects of prime structure (DO versus PO) or prime verb bias (DO-bias versus PO-bias) using permutation tests based on linear mixed-effect models with a maximal random-effect structures in each 50-ms time bin. To detect the start and end time of significant clusters for these effects, we employed 1000 permutations and the sum as a clustermass statistic.
Analyses for DO target sentences. Figure 4B graphs the time course of looks to the recipient and theme in DO target sentences following PO prime (red line) versus DO prime (blue dashed line) sentences across the three critical time windows. The mean durations of these time windows were 510 ms for the VERB, 637 ms for DETERMINER1 and 831 ms for NOUN1.
Visual inspection of the gaze data suggests that participants looked more often at the recipient than at the theme after both DO prime and PO prime sentences in all analysis windows, particularly in the NOUN1 window. Critically, participants did not make more looks to the recipient following DO primes compared to PO primes. As seen in Table 4, the traditional time-window analyses revealed no significant main effects of prime structure for DO target sentences. Consistent with these results, the cluster-based permutation analyses also did not detect any significant clusters for the main effects of prime structure in any time windows.
Analyses for PO target sentences. Figure 4C shows the time course of looks to the recipient and theme in PO target sentences following PO prime (red line) versus DO prime (blue dashed line) sentences for each time window. Note that, unlike for DO target sentences, the noun of the first postverbal argument in PO target sentences is the theme, not the recipient. The mean durations of time windows were 510 ms for the VERB, 644 ms for DETERMINER1 and 782 ms for NOUN1. The gaze plot illustrates that there was a higher proportion of looks to the theme than to the recipient following PO prime (versus DO prime) sentences in the NOUN1 window. In the traditional time-window analyses, we did not find any significant main effect of prime structure in any time windows (Table 4). In contrast, the cluster-based permutation tests showed one significant negative cluster in the NOUN1 window, indicating that participants looked more at the theme after PO prime than DO prime sentences (Figure 4C: 0–100 ms, cluster-mass = 15.32, p = .008). This finding establishes structural priming for PO target sentences.
The planned follow-up analysis tested whether the PO priming effect was modulated by verb bias. Figure 4D displays the time course of looks to the recipient and theme in PO target sentences that were preceded by PO prime sentences with either DO-bias (dark red line) or PO-bias (light red line) verbs. In the NOUN1 window, the main effect of prime verb bias was significant in the time-window analysis.Footnote 2 There was a higher proportion of looks to the theme than to the recipient after PO prime sentences with DO-bias verbs (Table 4). Similarly, the cluster-based permutation test detected a negative cluster in the NOUN1 window, reflecting a greater increase in looks to the theme in the DO-bias PO prime conditions than in the PO-bias PO prime conditions (Figure 4D: 150–500 ms, cluster-mass = 57.71, p = .001). These results suggest that prediction error, occasioned by verb bias, underlies the structural priming effect in PO target sentences.
Posttest
Following Kootstra and Doedens (Reference Kootstra and Doedens2016), we excluded ‘other’ responses from the analyses and focused on PO- and DO-dative responses. Our dependent variable was computed to show the tendency to use a PO-dative structure among all PO- and DO-dative responses. This approach enabled us to analyze both the tendency to use a PO-dative structure and, as a complementary measure, the tendency to use a DO-dative structure. In a logistic mixed-effect models, we examined the effects of verb type (DO-bias, PO-bias, EQ-bias and non-alternating), task (baseline versus posttest) and their interactions. Fixed effects were sum-coded, except for task, which was treated using treatment coding to contrast the posttest with the baseline, set as the reference level. Random intercepts were included for participants and items, with verb type and task as by-participant random slopes. In addition to significant main effects of verb type, our analysis revealed a significant main effect of task, indicating an increase in participants’ use of the PO-dative structure from baseline to posttest (for model output, see Table 5; see also Figure 3, left panel). However, there were no interactions of verb type and task, suggesting similar patterns in PO-dative sentence production across verb types, at baseline and posttest.
Table 5. Experiments 1 and 2. Models for baseline versus posttest production data
Note. SE, Standard error. Levels of Verb Type set to +1 for DO-bias, 0 for PO-bias, 0 for non-alternating, and − 1 for EQ-bias.
Discussion
In Experiment 1, L1 German advanced learners of L2 English made significantly more looks to the theme than to the recipient in PO target sentences following PO prime (versus DO prime) sentences. Since the target sentences were identical across the prime conditions, and the looks to theme started to differ before the onset of the post-verbal nouns, these differences in looks reflect prediction effects of the prime sentence. They indicate that L2 English learners generated more predictions for the PO-dative structure (i.e., the Theme>Recipient order) after PO prime sentences; that is, they exhibited a PO priming effect. Since priming was observed even though the prime and target sentences contained different verbs, priming in L2 comprehension appears to draw on lexically independent, abstract structural representations, similar to priming among L1 speakers (Chen et al., Reference Chen, Wang and Hartsuiker2022; Thothathiri & Snedeker, Reference Thothathiri and Snedeker2008; though see Arai et al., Reference Arai, van Gompel and Scheepers2007). Moreover, among the PO-prime–PO-target sentences, the L2 learners made significant differences in the proportion of looks to the theme and recipient as a function of prime verb bias: DO-bias PO prime sentences elicited more looks to the theme than PO-bias PO prime sentences in the processing of PO target sentences. These findings indicate that PO priming effects were larger when the prime structure was unexpected as per the bias of the prime verb, that is, after PO prime sentences with DO-bias verbs, when the surprisal of the PO structure was higher.
Although German-speaking learners of L2 English exhibited lexically independent structural priming and inverse preference priming for PO target sentences according to verb bias, they did not show any priming or prediction-error effects for DO target sentences. This finding replicated the pattern of results with English native speakers in Jaeger and Snider (Reference Jaeger and Snider2013), who also demonstrated effects of structural priming and verb bias in the production of PO but not DO-dative sentences. This asymmetry has been argued to reflect inverse frequency effects across the two dative structures in English, with priming and verb bias effects for the overall less frequent structure in English (PO-dative), yet not for the more frequent structure in English (DO-dative; for other languages, see Bernolet & Hartsuiker, Reference Bernolet and Hartsuiker2010; Chen et al., Reference Chen, Wang and Hartsuiker2022, Reference Chen, Wang and Hartsuiker2023).Footnote 3 Of note, the L2 learners in our study patterned with native speakers in showing inverse frequency effects of PO priming in comprehension – consistent with frequency differences they experience in the English input – even though they themselves produced an overall higher frequency of PO than DO structures in baseline and posttest (for discussion, see below). This finding suggests that experience with the target-language usage frequencies – rather than their own production preferences at baseline – underlies structural priming among advanced L2 learners.
In the baseline and the posttest, participants produced English-dative structures before and after the priming task. In the baseline production task, L1 German learners of L2 English revealed a strong preference for the PO-dative structure. This finding is in line with previous studies on the L2 acquisition of the dative alternation that reported a general preference for PO-dative over DO-dative constructions among L2 learners (e.g., Callies & Szczesniak, Reference Callies, Szczesniak, Walter and Grommes2008; Jäschke & Plag, Reference Jäschke and Plag2016). Such a preference has been related to the higher accessibility and lower markedness of the PO-dative structure (see Chang, Reference Chang2004; Mazurkewich, Reference Mazurkewich1984). After the priming task, the L1 German learners of L2 English produced significantly more PO-dative sentences in the posttest task than in the baseline task. This increase in the use of the PO-dative structure from the baseline to the posttest suggests that the PO priming effects found in comprehension priming persist into later unprimed production.Footnote 4 In these respects, Experiment 1 furnishes evidence that prediction errors according to verb biases occasion structural priming of the PO structure also among late L2 learners. Moreover, such structural priming extends beyond the priming phase, leading to persistent increases in the production of the PO structures that had been primed.
In order to isolate the effects of prime verb biases, Experiment 1 used equi-biased verbs in the target sentences. To test how prime verb biases interact with target verb biases, Experiment 2 changes the target verbs in the PO sentences to non-alternating verbs, such as “donate.” This way, we can test how prime and target verb biases interact in priming and prediction-error effects and whether the use of different verb biases in target sentences affects learning via longer-term priming.
Experiment 2: Effects of target verb bias on prediction-error-based priming
Method
Participants
We recruited 48 further participants (27 females) from the same population and on the same terms as in Experiment 1. Their mean LexTALE score was 79.77 (Table 1), and an independent samples t-test indicated that participants in Experiments 1 and 2 did not differ significantly in their proficiency scores (t(94) = 1.62, p = .109).
Materials
Priming task. The materials were the same as in Experiment 1, except that the verbs in the PO target sentences were replaced by four non-alternating verbs (i.e., “display,” “donate,” “carry” and “explain”), which are grammatically restricted to the PO-dative structure (Table 2).
Baseline and posttest tasks. The materials were identical to those of Experiment 1.
Procedure
This experiment involved the same procedure as Experiment 1.
Results
Baseline task
As seen in Figure 3 (right panel, blue bars), participants in Experiment 2 also showed a clear preference for the PO-dative over DO-dative structure and verb-type differences at baseline (for model output, see Supplementary Table S2).
Priming task
In Experiment 2, as in Experiment 1, we initially ran omnibus linear mixed-effect models that included both DO and PO target sentences. We found significant main effects of prime verb bias in the VERB window. These effects approached significance in the DETERMINER1 and NOUN1 windows. Furthermore, we observed significant interactions of prime structure and prime verb bias in the VERB window, where DO target sentences contained an equi-bias alternating verb (e.g., “The tailor will show the …”) and PO target sentences had a non-alternating verb (e.g., “The tailor will display the …”). To further explore these interactions, we performed all subsequent analyses separately for DO target sentences and PO target sentences. The results of the omnibus linear mixed-effect models are given in Table 3 (see also Figure 5A).
Figure 5. Experiment 2. Differences in gaze probability between recipient and theme by time window (empirical logits of looks to recipient minus looks to theme, by prime type (panel A), prime structure (panels B and C) or prime verb bias (panel D).
Note. Time (on x-axis) plotted with 0 ms corresponding to 200 ms after the Noun1 onset. Shaded areas denote significant clusters in cluster-based permutation analyses, including onset and offset times. The gray area around the lines represents 95% confidence intervals.
Analyses for DO target sentences. Figure 5B plots the time course of looks to the recipient and theme in DO target sentences following PO prime (red line) versus DO prime (blue dashed line) sentences for each critical time window. The mean durations of time windows were 510 ms for the VERB, 637 ms for DETERMINER1 and 831 ms for NOUN1.
Visual inspection of the gaze patterns shows that there was a general trend with more looks toward the recipient than toward the theme for both DO prime and PO prime conditions. Importantly, there was no increase in looks toward the recipient after DO primes compared to PO primes. Neither the time window nor the permutation-based analyses returned any significant differences between the two prime structure conditions (Table 4).
Analyses for PO target sentences. Figure 5C plots the time course of looks to the recipient and theme in PO target sentences following PO prime (red line) versus DO prime (blue dashed line) sentences for each time window. The mean durations of time windows were 625 ms for the VERB, 622 ms for DETERMINER1 and 752 ms for NOUN1. Figure 5C indicates that, in the VERB window, looks to the theme began to increase in the PO prime but not in the DO prime condition. In the DETERMINER1 window, participants continued looking more at the theme than at the recipient in the PO prime condition. In the NOUN1 window, however, there was an overall increase in the proportion of looks to the theme until around 600 ms.
The traditional time-window analyses returned a significant main effect of prime structure in the DETERMINER1 window (Table 4). The permutation tests also detected one significant negative cluster in the DETERMINER1 window, indicating a higher proportion of looks to the theme in the PO prime relative to the DO prime condition (Figure 5C: −123– −23 ms, cluster-mass = 13.89, p = .005). These findings replicate structural priming for PO target sentences, as found in Experiment 1. For PO target sentences following PO prime sentences, Figure 5D breaks down the PO prime conditions into DO-bias (dark red line) and PO-bias (light red line) verbs. It shows that the DO-bias PO prime condition occasioned a greater increase in the proportion of looks to the theme in the VERB and DETERMINER1 windows. Consistent with this, the traditional time-window analyses revealed a significant main effect of prime verb bias in the VERB window (Table 4). In the cluster-based permutation tests, we found two significant negative clusters, one in the VERB window (Figure 5D: −699– −649 ms, cluster-mass = 17.89, p = .002) and another in the DETERMINER1 window (Figure 5D: −623– −473 ms, cluster-mass = 28.57, p = .001). These results point to effects of prediction error on structural priming.
Posttest
As seen in Figure 3 (right panel), participants in Experiment 2 increased their PO-dative production from baseline to posttest. We ran the same models as in Experiment 1. As seen in Table 5, alongside significant main effects of verb type, the model returned only a significant main effect of task because participants’ use of the PO-dative structure increased from baseline to posttest.
In these respects, the findings for longer-term priming in Experiment 2 replicate the learning effects in Experiment 1 (for a model with experiment as a fixed effect, see Supplementary Table S3).
Discussion
In Experiment 2, we replaced the alternating unbiased verbs in PO target sentences by non-alternating verbs. Similar to Experiment 1, a comparable group of L2 English learners also made more predictive looks to the theme than to the recipient in PO target sentences following PO prime sentences; that is, they showed PO priming effects. They also generated stronger expectations for the PO-dative structure after PO prime sentences with DO-bias versus PO-bias verbs, demonstrating inverse preference effects of verb bias. In addition, the learners showed longer-term PO priming from baseline to posttest.
In all these respects, the results from Experiment 2 were comparable to those of Experiment 1. However, the analyses of the eye-tracking data by experiment showed that effects of prime structure and prime verb bias emerged in the NOUN1 window for PO sentences with alternating unbiased verbs (Experiment 1), yet earlier in the VERB and DETERMINER1 windows for PO sentences with non-alternating verbs (Experiment 2). To determine whether differences in the timing of effects were driven by the differences in the target verbs, we performed further analyses on the combined PO-target datasets of both experiments.
Experiments 1 and 2: Cross-experiment analyses for immediate PO priming
To identify effects of target verbs on structural priming, we analyzed the difference in the empirical logits of looks to the recipient versus theme for the PO prime conditions, where the target sentences following PO primes contained alternating EQ-bias verbs (e.g., “show” in Experiment 1) versus non-alternating verbs (e.g., “display” in Experiment 2). All analyses included experiment (Exp1 versus Exp2) as a sum-coded fixed effect, as well as participants and items as random intercepts. The traditional time-window analyses revealed that, in the DETERMINER1 window, the proportion of looks to the theme was significantly higher in Experiment 2 than in Experiment 1 (estimate = −0.43, SE = 0.16, t = −2.73, p = .008); for full models, see Supplementary Table S4. In the permutation analyses, we found negative clusters according to the experiment in the DETERMINER1 (−434–166 ms, cluster-mass = 9.78, p = .001) and in the NOUN1 windows (250–450 ms, cluster-mass = 26.35, p = .001). These differences underscore that the time course of priming was different in Experiments 1 and 2.
To assess the effects of verb constraints on prediction-error-based priming, we compared the DO-bias PO prime conditions in Experiments 1 and 2. The traditional time-window analyses returned a significant main effect of experiment in the DETERMINER1 window, showing that there were more looks to the theme versus recipient in Experiment 2 than in Experiment 1, i.e., when PO prime sentences with DO-bias verbs were followed by the target sentences with non-alternating verbs, for which the DO order is not available (estimate = −0.66, SE = 0.22, t = −3.05, p = .003), see Supplementary Table S2. The cluster-based permutation analyses detected corresponding significant negative clusters in the DETERMINER1 window (−634– −284 ms, cluster-mass = 110.20, p = .001) as well as in the earlier VERB window (−703– −603 ms, cluster-mass = 27.10, p = .001). These effects substantiate cross-experiment differences in the time course of effects of prediction error.
Discussion
Cross-experiment analyses revealed differences in the time course of prediction effects between PO sentences with non-alternating verbs in Experiment 2 and sentences with alternating unbiased verbs in Experiment 1. In Experiment 2, PO priming and surprisal verb bias effects occurred in earlier time windows of the target verb and the determiner of the first postverbal argument and did not extend to the time window of the noun of the first postverbal argument, in which they were significant in Experiment 1. One interpretation of this difference is that encountering a non-alternating verb in the target sentence in Experiment 2 led listeners to cancel the effects of the prime sentences, since the verb does not allow for structural optionality in the realization of its complements. With regard to priming, these comprehension-based effects of verb type add to previous L1 and L2 production studies, which show that – over and above effects of verb biases in the prime sentences – the strength of structural priming is modulated by the verb-selectional preferences in the target sentence (e.g., Bernolet & Hartsuiker, Reference Bernolet and Hartsuiker2010; Kaan & Chun, Reference Kaan and Chun2018; see also Gries, Reference Gries2005).
General discussion
In two experiments, we examined whether prediction error, occasioned by verb bias, affects structural priming of English-dative constructions during L2 sentence comprehension (RQ1), whether L2 structural priming in comprehension persists into later (unprimed) production (RQ2), and how verb biases in the target sentences affect prediction-error-based priming among adult L2 learners (RQ3). Experiment 1 showed that (1) prediction error systematically amplifies abstract structural priming and (2) prediction-error-based priming in comprehension translates into longer-term production. Experiment 2 indicated that (3) effects of prediction error and priming are smaller and shorter-lived when target verb biases restrict structural options. In the following, we discuss these findings and outline their implications for understanding how prediction can serve as a learning mechanism in L2 acquisition.
Consistent with the claims of error-based implicit learning models (Chang et al., Reference Chang, Dell and Bock2006; Dell & Chang, Reference Dell and Chang2014), the L2 learners in Experiments 1 and 2 showed priming effects that were stronger for the less frequent structure in English (PO versus DO structures) and stronger for verbs that dispreferred to occur in this structure (DO-bias versus PO-bias verbs). Finding such abstract structural priming together with inverse frequency effects between DO and PO structures as well as inverse preference effects according to verb bias suggests that the mechanisms assumed to underlie priming in (prediction-)error-based implicit learning accounts of priming are also operative among late L2 learners.Footnote 5
In addition, both experiments also found evidence of longer-term priming in production that went in the same direction as the immediate priming effects in comprehension. In these respects, our findings were in line with previous studies that reported longer-term priming in an L2 (e.g., Nitschke et al., Reference Nitschke, Kidd and Serratrice2010; Wei et al., Reference Wei, Boland, Cai, Yuan and Wang2019). In this study, we could not test whether such longer-term priming might have been reduced compared to priming in native speakers or early bilinguals (e.g., Jackson & Hopp, Reference Jackson and Hopp2020). Nevertheless, these longer-term priming effects further bolster the contention of error-based models of priming that structural priming is a form of learning.
As seen in Figure 3, German-speaking learners increased their PO-dative productions irrespective of verb type from baseline to posttest, while they continued to make differences in the number of POs produced according to verb type in the posttest. In other words, the longer-term priming effects were not restricted to the DO-biased verbs that caused the most priming. We suggest that this overall increase in POs is likely due to the overall ratio of PO- and DO-dative primes in the priming experiment: Participants encountered the PO-dative structure in 50% of the prime sentences and the DO-dative structure in the other 50%. Overall, this ratio exposed them to more PO-dative sentences than they would encounter in native English input. Consequently, learners appear to update their (frequency) preferences by upping the relative number of PO-dative sentences in their production, independently of the much lower native baseline for PO sentences (e.g., Jaeger & Snider, Reference Jaeger and Snider2013). Since they encountered an equal number of PO and DO prime sentences across verbs and their biases, they also increased their PO productions across verb types. To establish whether prediction-error-based priming can lead to more targeted learning relative to verb biases, a future priming study could create asymmetric prediction errors, e.g., by violating biases only for PO-bias verbs in the prime sentences, so as to test whether targeted prediction error leads to selective increases, e.g., in DO-datives for PO-bias verbs in production.
In Experiment 2, we found that the priming and verb bias effects from the primes were more restricted than in Experiment 1. These differences, backed up by the cross-experiment models, suggest that L2 learners immediately recruited the verb biases of the target verbs to reign in and cancel prediction error effects from the prime sentence when the verb in the target sentence could only occur in one order. Hence, L2 learners were sensitive to verb-based constraints on the dative alternation and applied them to gate effects of prediction error on priming. Accordingly, prediction-error-based priming in an L2 does not appear to operate as a blanket mechanism in L2 structural priming that generally leads to increased activation of a surprising prime structure in L2 learners’ processing of the target. Instead, L2 learners immediately apply grammatical verb constraints to circumscribe prediction-error-based priming as per their knowledge of the L2. In this respect, structural priming among advanced L2 learners can be modeled as the interplay of verb biases in the prime sentences and verb biases in the target sentences. Accordingly, error-based learning leads L2 learners to adjust their predictive processing and production to the input against the backdrop of their language experience and within the constraints afforded by their knowledge of the target grammar.
In establishing a loop from prediction error via comprehension priming to production, the present study suggests that language processing mechanisms, in particular predictive processing, scale up to be learning mechanisms also among late L2 learners (Bovolenta & Marsden, Reference Bovolenta and Marsden2022). Learners can capitalize on feedback provided by prediction error relative to their expectations about how the input will unfold during comprehension. Such implicit feedback translates into learning via priming and gradually changes learners’ production. In this regard, the study corroborates and adds to research on prediction and learning among child L1 learners (e.g., Reuter et al., Reference Reuter, Borovsky and Lew-Williams2019), suggesting there is continuity in the availability of language processing as learning mechanisms from early L1 to late L2 acquisition. Learning from prediction error via priming may thus constitute a powerful learning tool for L2 learners, next to other implicit and explicit learning mechanisms. At the same time, the scope of learning the L2 grammar via prediction may be circumscribed in various ways (for discussion, Hopp, Reference Hopp, Kaan and Grüter2021). First, learning from prediction error presupposes that learners make predictions in the first place. The advanced L2 learners in this study did engage in making predictions about the likely complement order of dative verbs (see also Şafak & Hopp, Reference Şafak and Hopp2023). In this sense, L2 learning does not seem to be constrained by lower degrees of prediction in principle (for discussion, Hopp, Reference Hopp, Kaan and Grüter2021; Phillips & Ehrenhofer, Reference Phillips and Ehrenhofer2015). However, it is an open question whether less proficient learners would also make predictions. After all, it is these learners that have the most to learn in an L2. Second, prediction-based grammatical learning via priming has typically been studied for optional word orders, such as DO- versus PO-datives and active versus passive voice. However, most grammatical learning involves non-optional aspects of the L2 grammar, such as basic word order or interpretive constraints on word order variation. It remains to be studied whether and how the shifting of grammatical preferences via prediction-error-based learning also underlies the acquisition of categorical constraints in the L2 grammar. Third, prediction-based learning potentially works well for forward contingencies, such as verb-complement ordering and agreement; yet, many grammatical phenomena, such as anaphoric relations and long-distance dependencies, are ‘backward’ contingencies and cannot be learnt from prediction. In sentence processing, adult L2 learners seem to struggle particularly with backward contingencies (e.g., Felser, Reference Felser2015), so that prediction-based learning may have a limited reach. Finally, a relevant question for future studies is whether comprehension priming to production extends across different, related grammatical structures (e.g., Jackson & Ruf, Reference Jackson and Ruf2017), which would allow learners to hone predictive processing to make generalizations (e.g., Hopp et al., Reference Hopp, Schimke, Öwerdieck, Gastmann and Poarchin press). For these reasons, delineating the scope of prediction-based learning across different learners and phenomena will be essential to situate learning from prediction alongside other types of learning mechanisms in L2 acquisition.
In conclusion, the present study shows that adult L2 learners adapt their production to the structure of the recently processed input through error-based implicit learning from their predictions. At the same time, learners’ grammatical knowledge of target verb biases constrains learning via prediction error, preventing learners from overgeneralizations. Hence, sentence processing can guide learners to the target grammar by virtue of priming via prediction error also in a later-learnt L2. Learning from prediction error may thus be a key mechanism in L2 acquisition, allowing L2 learners to process to learn the L2 grammar.
Supplementary material
To view supplementary material for this article, please visit http://doi.org/10.1017/S1366728925000033.
Data availability statement
The data that support the findings of this study are openly available in Open Science Framework at https://osf.io/hej2g/.
Acknowledgments
The research reported in this study was funded by the German Research Foundation, DFG (grant no. HO 5923/7-1) and carried out within the Research Unit “SILPAC – Structuring the Input in Language Processing, Acquisition and Change” (2022–2026). We would like to thank Julia Grubert, Svenja Polonyi and Kimberly Last for assistance with stimulus creation, data collection and transcription. We are grateful to the anonymous reviewers for BLC, the audiences at PiF 2023 at Ghent University, LingCologne 2023 at the University of Cologne, AMLaP 29 in San Sebastián as well as EuroSLA 32 in Birmingham for comments and questions.
Competing interest
The authors declare none.
Open access
