Many studies that investigate the use of semantic information for prediction involve the visual world paradigm, in which participants listen to target sentences while researchers track their eye gazes to different images on a computer screen (for review, see Huettig et al., 2011). Studies using this paradigm have shown that listeners use linguistic information in conjunction with the visual context to build strong expectations for upcoming referents, thereby looking at a target object even before it is mentioned (see Huettig, 2015, for review). In a seminal study, Altmann and Kamide (1999) found that L1 English listeners looked at an image of the target object (e.g., cake) more than unrelated distractors upon hearing a semantically-constraining verb in sentences like (1a), because from this set of objects, only the cake was a plausible object of the verb eat. In contrast, they looked at multiple objects upon hearing a semantically-neutral verb in sentences like (1b), because more than one object on the screen was a plausible object of the verb move.

Following Altmann and Kamide, many studies have further probed how listeners use verbal semantics to anticipate the upcoming object in sentences with SVO word order, but only a few have considered semantically-cued prediction in other word orders (see Bondoc & Schafer, 2022; Dahan & Tanenhaus, 2004; Sauppe, 2016, for exceptions).

Although the primary focus of their study was not the impact of syntax or word order on prediction, Dahan and Tanenhaus (2004) exploited grammatical properties of Dutch to study semantically-cued prediction in non-subject-initial AdvVS sentences, in which the verb precedes the subject. Dutch is an SOV language that requires verb-second (V2) word order in main clauses. As seen in examples (2a) and (2b), this means that the finite verb always appears in second position in main clauses, even when another constituent occurs sentence initially. In such cases, the subject appears after the verb. Further, when a clause contains only an inflected lexical verb, that verb appears in second position, as in (2a). In constructions with an auxiliary or modal verb, by contrast, the inflected auxiliary appears in second position, while the non-finite lexical verb remains in clause-final position, as in (2b).

Paralleling research with SVO sentences, Dahan and Tanenhaus found that L1 Dutch listeners looked at the target referent (i.e., bok ‘goat’) earlier in sentences with a semantically-constraining verb (i.e., klom ‘climbed’) versus a neutral auxiliary verb (e.g., is ‘is’), suggesting that they successfully anticipated a post-verbal subject, based solely on the semantic (thematic) constraints provided by the verb.

Turning to L2 research, a growing number of studies show that L2 speakers also reliably use semantic information to predict upcoming arguments in sentences with SVO word order (e.g., Chambers & Cooke, 2009; Dijkgraaf et al., 2017; Ito, Corley, et al., 2018), even though L2 processing, in general, is more resource-intensive (e.g., Declerck & Kormos, 2012; Francis et al., 2018). Notably, L1 Japanese-L2 English listeners can successfully use verbal semantics to anticipate an upcoming object in SVO sentences, even though their L1 Japanese has SOV word order (Ito, Pickering, et al., 2018). Further, Dijkgraaf et al. (2017) found that L1 Dutch-L2 English listeners used semantic information for prediction to similar degrees in both of their languages, and the strength of their predictions in the L2 were indistinguishable from those made by L1 English listeners. Ito, Corley, et al. (2018) also reported that L1 English and L2 English listeners with a variety of L1s used semantic information for prediction to similar degrees under normal listening conditions, and that both groups experienced a similar delay in prediction when they performed a concurrent memory task, which induced greater cognitive load. These parallels suggest that the mechanisms underlying the use of semantic information for prediction are largely similar in L1 and L2 speakers (see also Chambers & Cooke, 2009; Kaan, 2014).

However, more recent studies point to potential quantitative differences between L2 and L1 speakers in their use of semantic information for prediction when L2 speakers are faced with processing more complex sentences or integrating multiple types of information. Whereas the studies described above all relied on simple sentences with SVO word order, when processing SVO sequences embedded in more syntactically complex sentences, such as relative clauses, L2 listeners exhibited a delay in prediction relative to their processing of syntactically simple sentences (Chun et al., 2021) and a delay relative to L1 listeners (Chun & Kaan, 2019; see also Dijkgraaf et al., 2019; Ito, Pickering, et al., 2018). Paralleling these differences, Corps et al. (2023) found that while L2 English listeners engaged in simple associative semantic prediction (e.g., looking at an image of a tie after hearing the verb wear) with a speed similar to L1 English listeners, they were slower than L1 English listeners to use top-down contextual cues of a male versus female voice to focus their gaze on a stereotypically masculine object (e.g., tie) versus a stereotypically feminine object (e.g., dress). Together, such findings have led researchers to hypothesize that semantically-cued prediction may be qualitatively similar for L1 and L2 speakers, but that it may be more vulnerable to pressures from other resource demands, whether they be non-linguistic (e.g., working memory load) or linguistic (e.g., syntactic context, cross-linguistic influence) (see Ito & Pickering, 2021; Kaan & Grüter, 2021; Schlenter, 2023, for further discussion). The present study contributes to a better understanding of the characteristics of predictive processing in L1 and L2 speakers by investigating how semantically-cued prediction may vary across different syntactic contexts, namely, subject-first SVO versus adverb-first AdvVS(O), especially when one of these word orders is absent in the L1.

There is growing consensus in the psycholinguistic literature that multiple mechanisms may underlie predictive processing among both L1 and L2 speakers. Even proposals that argue for a single underlying mechanism acknowledge that multiple linguistic and non-linguistic factors influence the scope – and even the presence or absence – of predictive processing across different populations and in different contexts (e.g., Knoeferle & Crocker, 2007; Kuperberg, 2007; Pickering & Gambi, 2018; Pickering & Garrod, 2013; Ryskin & Niewland, 2023, among others. See Huettig, 2015, for a comprehensive review). While different models vary in their assumptions regarding the types and number of mechanisms that drive prediction during language processing, in broad terms, all models acknowledge that prediction can be driven by semantic priming through the spreading activation of related concepts. For instance, upon hearing the verb arrest, a listener would activate closely related concepts, such as robber, gunman, policeman, bank, jail, etc. In turn, this would prime listeners to look at images of these concepts in a visual world display (e.g., Kukona, et al., 2011; see also Chow, et al., 2016). This type of simple associative mechanism may be less resource intensive and more automatic than predictive mechanisms that take a larger number of linguistic and non-linguistic factors into account (e.g., Chow et al., 2016; Pickering & Gambi, 2018, among others). Yet such associative mechanisms can also be noisy and potentially lead to erroneous predictions (e.g., Kukona et al., 2011). Further, among L2 speakers, even such associative mechanisms may be slower and potentially more resource demanding, relative to L1 speakers, due to the more complex processes of lexical activation in the bilingual lexicon (Dijkgraaf et al., 2019).

Alongside prediction by associative priming, models of predictive processing also incorporate routes through which comprehenders can actively anticipate and pre-activate upcoming linguistic input, based on various types of bottom-up and top-down information that they encounter as a sentence unfolds. Such information can include grammatical or morphosyntactic information, like case marking (e.g., Hopp, 2015; Kamide et al., 2003; Schlenter & Felser, 2021), phonological or orthographic forms (e.g., Ito, Pickering, et al., 2018; Martin et al., 2013), contextual information, like the objects presented in a visual world display (e.g., Knoeferle & Crocker, 2007; Tanenhaus et al., 1995) or speaker voice (e.g., Corps et al., 2023; Kamide, 2012), as well as the event structure and higher-level meaning that people build up through the combined meaning of multiple words over time (e.g., Altmann & Kamide, 2007; Chambers & San Juan, 2008; Kim & Osterhout, 2005). However, making predictions via this route likely requires greater cognitive resources. Therefore, such higher-level information may contribute to predictive processing more selectively, namely, only when such information is maximally reliable and beneficial for comprehension (e.g., Brothers et al., 2017; Huettig, 2015; Kuperberg & Jaeger, 2016; Pickering & Gambi, 2018). The use of multiple and higher-level cues may be even more constrained among L2 speakers, precisely because L2 processing, in general, is more resource demanding, leaving fewer resources to devote to prediction (see, e.g., Ito & Pickering, 2021; Kaan, 2014, for further discussion).

There is evidence that both associative and more complex predictions occur in tandem. For instance, using the visual world paradigm, Kukona et al. (2011) found that when L1 English listeners heard a sentence fragment like Toby arrests the… they looked at an image of a plausible patient (e.g., crook), but also an image of a plausible agent (e.g., policeman) even though the subject role in the sentence had already been filled by Toby. Such competition effects were attenuated, but not eliminated entirely, when participants listened to passive sentences, like Toby was arrested by the… , where participants looked at the plausible agent (e.g., policeman), after by, but they still looked at the plausible patient (e.g., crook) more often than at unrelated distractors. Kukona and colleagues suggested that when there are strong associative connections between a verb and its arguments, this can lead to early activation of these arguments, even when such local thematic priming conflicts with sentence-level constraints, like word order and the distribution of grammatical and thematic roles. Further, the relative weight of such local – namely, lexical-level – versus sentence-level constraints can vary, based on the availability and strength of these different constraints in a sentence.

The relative weight of local- versus sentence-level cues during predictive processing can also vary based on listeners’ linguistic experience and skills (Borovsky et al., 2012, 2013). For instance, in a visual world experiment with a heterogeneous group of adult bilinguals fluent in English, Peters et al. (2018) tracked participants’ eye movements to four images while listening to sentences like The pirate chases the ship: the target (ship), a distractor semantically associated with the agent (treasure chest), an action-related distractor, i.e., a plausible object for the verb without considering the subject (cat), and an unrelated distractor (bone). Similar to the monolingual English participants in Borovsky et al. (2012), these bilingual participants’ looks to the action-related distractor (cat) temporarily increased after hearing the verb (chases), even though the referent was an unlikely continuation in the context of the preceding sentence fragment as a whole (The pirate chases…). Of note, these local semantic association effects were greater among participants who self-identified as non-native speakers of English and participants with lower English vocabulary skills. Peters and colleagues concluded that less proficient bilingual speakers are more likely to activate not only the most probable referent, given cumulative information from the preceding sentence fragment, but also referents that are only locally coherent, that is, semantically associated with the immediately preceding word, suggesting that simple associative prediction may play a greater role among less proficient language users.

Investigating how L1 and L2 German listeners use verbal semantics to predict post-verbal arguments across different word orders provides an ideal environment in which to probe the extent to which prediction proceeds via simple semantic association versus the use of higher-level information. The present study takes advantage of the verb-second (V2) property of German. Like Dutch, German exhibits V2 word order in main clauses, meaning that the finite verb always appears in second position in main clauses, regardless of whether the subject or another constituent occurs sentence initially. In instances where a non-subject constituent occurs sentence initially, the grammatical subject appears after the finite verb, as in (3).

This contrasts with languages with more rigid SVO word order, like English, where the subject always precedes the verb, even in instances where another constituent appears sentence initially (In summer, the frog jumps…).

At least at lower proficiency levels, L2 German speakers with L1 English, an SVO language, will continue to place the subject before the finite verb in adverb-first sentences (e.g., *Im Frühling der Frosch springt ins Wasser; Jackson & Ruf, 2017; Pienemann, 1998). Doing so maintains the canonical agent-verb-patient order of SVO sentences from their L1 English, even though it results in an ungrammatical verb-third sentence in German. Even at more advanced proficiency levels, L2 German speakers with L1 English exhibit continued difficulties with V2 word order and are less likely to produce non-subject initial sentences, compared to L1 German speakers (Carroll et al., 2000; O’Brien & Féry, 2015). In the context of L2 processing and prediction, less experience with the V2 property of German, and its consequence that the verb precedes the subject in non-subject first utterances, also means that L2 German speakers with English or another SVO language as their L1 have less experience anticipating post-verbal subjects. Thus, one primary question driving the present study is whether L2 speakers will predict a post-verbal subject when listening to precisely such AdvVS(O) sentences in German.

In two visual world experiments, the present study examines whether differences in word order modulate the use of semantic information for prediction among L1 and L2 German listeners. In Experiment 1, we presented participants with SVO and AdvVS sentences in order to investigate their use of semantically informative cues on the verb to predict the post-verbal argument. In Experiment 2, which focuses specifically on non-subject-initial AdvVS(O) sentences, we vary the argument structure of the verb. Participants were presented with transitive (AdvVSO) and intransitive (AdvVS) sentences in order to tease apart whether semantic cues on the verb lead to the anticipation of any compatible argument via simple semantic association or whether the preceding syntactic context constrains anticipation to the next grammatically possible referent, in this case, the subject. In so doing, the present study addresses the following research questions:

As seen in Table 2, the results from the AdvVS production task and the grammaticality judgment task revealed a range in productive and receptive knowledge of German AdvVS word order among the L2 participants, but most participants scored at or close to ceiling on these tasks.3 Results from the L2-to-L1 translation task demonstrated that the L2 participants knew, on average, 86.7% of the lexical verbs from the eye-tracking stimuli.4 Turning to the eye-tracking task, participants were highly accurate in clicking on the target at 98.6% across both groups (see Table 2), so data from all trials were included in the analyses reported below.

Both L1 and L2 German listeners were faster to look at the post-verbal object in SVO sentences with constraining verbs, as opposed to neutral verbs, and did so prior to target noun onset, paralleling results from previous research (e.g., Chambers & Cooke, 2009; Dijkgraaf et al., 2017; Ito, Corley, et al., 2018). The same pattern emerged in AdvVS sentences, in that both L1 and L2 German listeners were faster to look at the post-verbal subject in AdvVS sentences with constraining verbs, as opposed to neutral verbs, and did so prior to target noun onset, paralleling previous work with L1 listeners on semantic prediction in Dutch AdvVS sentences (Dahan & Tanenhaus, 2004). The L2 listeners were slower than the L1 listeners to anticipate the target referent for both SVO and AdvVS sentences, similar to other studies that directly compared the use of semantic cues for prediction by L1 vs. L2 speakers (e.g., Chun & Kaan, 2019; Dijkgraaf et al., 2019; Ito, Corley, et al., 2018). Critically, the mean difference in divergence points between the L1 and L2 listeners overlapped for each sentence type, suggesting that the L2 listeners exhibited no additional prediction delay with AdvVS sentences as opposed to SVO sentences, even though this required them to anticipate a post-verbal subject in a syntactic configuration not present in their L1. Further, in neither SVO nor AdvVS sentences did the proportion of looks to the target in constraining-verb versus neutral-verb sentences vary as a function of L2 proficiency. Although one must interpret such null effects with caution, this suggests that L2 proficiency did not modulate the magnitude of L2 semantically-cued prediction. In all, the findings suggest that L2 speakers – like L1 speakers – can use semantic cues for prediction across different word orders.

However, in AdvVS sentences, the target referent was also the only semantically-plausible referent associated with the lexical verb. Thus, the increased looks to the target referent after a semantically-constraining verb could stem from simple semantic association, rather than reflecting evidence that the participants specifically predicted a post-verbal subject (e.g., Chow et al., 2016; Kukona et al., 2011). To tease apart these two possibilities, we designed Experiment 2 to include transitive AdvVSO sentences, alongside visual scenes that depicted two possible post-verbal referents, namely, a plausible (agent) subject and a plausible (patient) object of the lexical verb. Given that Experiment 1 was also the first L2 study to our knowledge to examine the semantic prediction of post-verbal subjects, rather than post-verbal objects, we also included AdvVS sentences in Experiment 2, to test whether we could replicate these results from Experiment 1 with a new group of L2 speakers.

As seen in Table 4, results from the AdvVS production task and the grammaticality judgment task revealed a range in productive and receptive knowledge of German AdvVS word order among the L2 participants, but most participants scored at, or close to, ceiling on these tasks. The results from the L2-to-L1 translation task demonstrated that all L2 participants knew, on average, 81.1% of the lexical verbs from the eye-tracking stimuli. T-tests revealed that the L2 participants in Experiment 2 scored significantly higher than the L2 participants in Experiment 1 on the AdvVS production task, t(57) = 2.07, p = .043. No other differences were statistically significant (all ps > .15). Turning to the eye-tracking task, participants were highly accurate in clicking on the target noun at 95.2% across both groups (see Table 4), so data from all trials were included in the analyses reported below.6

Similar to Experiment 1, both L1 and L2 German listeners were faster to look at the post-verbal subject in AdvVS sentences with constraining verbs, as opposed to neutral verbs. Both groups were also faster to look at the post-verbal subject (agent) in AdvVSO sentences with constraining verbs, as opposed to neutral verbs, and did so prior to target noun onset, although this effect was less reliable among the L2 listeners. More importantly, a second set of divergence point analyses with AdvVSO sentences revealed that both the L1 and L2 listeners were more likely to look at the agent than the patient, and did so prior to noun onset only in AdvVSO sentences with a semantically-constraining verb. Looks to the agent did not reliably diverge from looks to the patient until after target noun onset in neutral-verb sentences in either listener group. This suggests that both the L1 and L2 listeners’ predictive eye movements were not guided by semantic association with the verb alone. Rather, they preferentially looked at the referent that was most likely to be mentioned next in the given syntactic configuration. Notably, this preference was found in both L1 and L2 listeners.

As in Experiment 1, the mean difference in divergence points between the L1 and L2 listeners was similar across all analyses, suggesting that the L1 and L2 listeners used qualitatively similar mechanisms for prediction and that any differences between the two groups were only quantitative in scope. As in Experiment 1, in neither AdvVS nor AdvVSO sentences did the proportion of looks to the target referent in constraining-verb versus neutral-verb sentences vary as a function of L2 proficiency. Although one must interpret such null effects with caution, this suggests that L2 proficiency did not modulate the magnitude of L2 semantic prediction – and, more specifically, the magnitude of predicting a post-verbal subject – in this experiment.

Experiment 1 showed that listeners use the semantics of a verb in V2 position to anticipate the post-verbal argument in German, regardless of whether that argument is a subject or an object. Further, the results with AdvVSO sentences in Experiment 2 suggest that the use of semantic information to predict argument roles cannot stem only from semantic association, in contrast to other research (e.g., Chow et al., 2016; Kukona et al., 2011). Critically, in constraining-verb AdvVSO sentences, both L1 and L2 listeners looked to the target (agent) subject more than the (patient) object prior to the onset of the post-verbal noun. In other words, both groups preferentially anticipated a plausible subject. This is more compatible with the hypothesis that, even in contexts in which strong associative connections exist between a verb and its arguments, speakers use sentence-level cues, like word order and argument structure, for prediction, at least when these are sufficiently strong and reliable.

This appears to contrast with the results for English active sentences from Kukona et al. (2011, Experiment 1), in which L1 English listeners looked at both a plausible agent (e.g., policeman) and a plausible patient (e.g., crook) when listening to sentence fragments like Toby arrests the… up until the onset of the target noun phrase. The present results are more in line with Kukona and colleagues’ results for English passive sentences, like Toby was arrested by… (Experiment 2), in which competition effects between plausible agent and plausible patient referents were attenuated by the time participants encountered the word by. Kukona and colleagues attributed this difference in active versus passive sentences to the fact that their passive sentences provided participants with additional syntactic cues (i.e., was and by) to constrain the target referent, as well as more time for these additional cues to influence prediction. As in Kukona and colleagues’ Experiment 2, sentences in the present study contained an additional word (a semantically neutral adjective, e.g., täglich, ‘daily’) between the verb and the post-verbal argument, thus providing more time for the integration of multiple cues and for predictive effects to emerge, compared to Kukona and colleagues’ Experiment 1. We also note that in both Kukona and colleagues’ passive sentences in their Experiment 2 and in our AdvVS(O) sentences, the post-verbal argument and target of prediction was the agent of the event. We hypothesize that agents may be privileged as targets of anticipation, as they are the most critical thematic role in establishing event representations (e.g., Bornkessel & Schleswesky, 2006).

Further support for this hypothesis comes from recent work with L1 Tagalog speakers (Bondoc & Schafer, 2022; Sauppe, 2016). Tagalog is a verb-initial language in which the order of the subsequent arguments is based on thematic roles (agent versus patient) and morphological markers on the verb that confer a more prominent status on one of the two arguments, which is referred to as the pivot (see Sauppe, 2016, and references cited therein for further detail). The pivot argument canonically and most frequently appears clause-finally, while the non-pivot argument appears immediately after the verb, leading to both VSO and VOS sentences. Sauppe (2016) showed that regardless of morphological pivot markers on the verb, and, thus, regardless of the most likely upcoming word order, participants were more likely to look at the most plausible agent after hearing the verb. Sauppe concluded that Tagalog speakers initially use verbal semantics to anticipate potential referents based on their thematic roles, rather than using morphological cues to identify the syntactic function (i.e., pivot role) and word order of the target referents (see also Bondoc & Schafer, 2022, for similar results). Together, these studies and our study highlight how agents may be more prominent than patients, leading listeners to identify the most plausible agent early on in sentence processing (e.g., Bornkessel & Schlesewsky, 2006). This line of research underscores the point that the mechanisms underlying semantically-cued prediction among both L1 and L2 listeners go beyond the use of simple semantic association to include, at the very least, the consideration of grammatical and thematic roles and their distribution across different word orders and argument structures.

To further disentangle the contribution of grammatical versus thematic roles in prediction, future studies should consider including sentences with verbs that require thematic roles other than agents and patients, such as implicit causality verbs associated with stimulus and experiencer roles, as in (9).

In both (9a) and (9b), the event-causing stimulus (i.e., Hexe ‘witch’) is considered more prominent than the experiencer (i.e., Frau ‘woman’), even though Hexe is the grammatical subject in (9a), but the grammatical object in (9b) (e.g., Hartshorne, 2014). If prediction is driven primarily by the prominence of thematic roles, then one would expect listeners to look at an event-causing stimulus, like Hexe, after hearing the verb in both (9a) and (9b). However, if prediction is driven primarily by the anticipation of the next grammatically possible referent, namely, the subject, then one would expect decreased looks to an event-causing stimulus, like Hexe, after hearing the verb in (9b) versus (9a). Such research could uncover additional evidence regarding the exact mechanisms that drive prediction, and which types of linguistic information are most crucial for prediction, during L1 and L2 processing.

Despite the delay in divergence points for the L2 listeners versus the L1 listeners, no qualitative differences in predictive processing emerged, paralleling other L2 studies that investigated the use of semantic cues for prediction (e.g., Chun & Kaan, 2019; Chun et al., 2021; Dijkgraaf et al., 2017, 2019; Ito, Corley, et al., 2018; Ito, Pickering, et al., 2018). Importantly, in contrast to previous L2 studies, which have all used sentences with SVO word order, the L2 listeners here engaged in semantically-cued prediction even with AdvVS(O) sentences – a word order that is not part of their L1 and that diverges from a more canonical agent-verb-patient word order. Further, like the L1 listeners, the L2 listeners always fixated on the most plausible post-verbal argument, even in AdvVS(O) sentences, when the syntactic configuration required listeners to use sentence-level cues to constrain their predictions to the subject (agent) of the target sentence, rather than using verbal semantics alone to anticipate any plausible post-verbal referent. Together, these results underscore the point that L2 speakers do not always privilege simple semantic association, instead using both syntactic context and argument structure to guide prediction.

Although the absence of any interaction between L2 proficiency and the use of verbal semantics in Experiments 1 and 2 must be interpreted with caution, given the relatively small sample size in both experiments, this result adds to the growing number of studies showing that L2 proficiency may not always modulate L2 prediction. In line with other recent L2 research, we tentatively hypothesize that L2 speakers will generate predictions based on those cues that are most reliable in the input, given their L2 knowledge and experience more broadly (see Kaan & Grüter, 2021, for further discussion). For the L2 listeners in this study, verbal semantics were a reliable cue for anticipating the most plausible and syntactically expected post-verbal argument across different word orders, regardless of L2 proficiency. This included the anticipation of post-verbal subjects in AdvVS(O) sentences – even in the absence of any experience using verbal semantics to anticipate post-verbal subjects from their L1.