Do structural priming and inverse preference effect demand cognitive resources? Evidence from structural priming in production

ABSTRACT Implicit learning theories assume that structural priming is based on an error-based prediction mechanism (e.g. Chang et al. [2006]. Becoming syntactic. Psychological Review, 113(2), 234–272.), which predicts stronger priming when the bias of the verb in the prime sentence towards a syntactic structure mismatches the actual sentence’s structure (inverse preference priming). We investigated whether structural priming and inverse preference priming are modulated by cognitive resources such as demand on memory. Experiments 1 and 2 showed inverse preference priming in a priming task that exerted a relatively low cognitive load (sentence reading followed by picture description), but Experiments 3 and 4 found no such effect in a more demanding task (i.e. sentence reading, sentence recognition judgment, picture description, and picture recognition judgment). In the less demanding experiments, structural priming was always stronger and inverse preference priming was marginally stronger. These findings suggest an important role of cognitive resources in error-based learning.


Introduction
Structural priming is the tendency for speakers to reuse the structure of a previous prime sentence in a subsequent target sentence (Pickering & Ferreira, 2008). This tendency of repeating structures is enhanced when the bias of the verb in the prime sentence towards a particular sentence structure mismatches that sentence's actual structure (i.e. inverse preference priming). For instance, Bernolet and Hartsuiker (2010) found that double object (DO) sentences (e.g. the Dutch equivalent of "The clown gave the swimmer a ball") exerted a stronger priming effect when the verb was more strongly biased towards the prepositional object (PO) dative (e.g. "The clown gave a ball to the swimmer"). The enhancement of priming when the structure mismatches the verb's structure bias, suggests an important role for prediction error in syntactic processing (Chang, 2000;Chang et al., 2006). To minimise prediction error for upcoming sentence structures, speakers would rapidly adapt to the experienced prediction error and continuously renew the weights of the abstract syntactic representations that contribute to structure prediction. This error-based learning procedure is argued to be automatic, and therefore should not be influenced by explicit memory (see Chang et al., 2006, p. 256). However, in order to be surprised and then learn from the prediction errors, speakers first need to detect the discrepancy with the expectations. They might not be able to construct these strong expectations without efficient memory storage and retrieval of the abstract representations in language production (Reitter et al., 2011). Therefore, we tested the structural priming and inverse preference priming effects in tasks that differ in their demand on cognitive resources.
There is considerable evidence for robust structural priming effects in many languages (Mahowald et al., 2016). Importantly, structural priming is an implicit method to tap into representations at the syntactic processing level (Branigan & Pickering, 2017). Therefore, we use this paradigm to investigate syntactic processing in production. There are three theories of structural priming, that differ in whether they predict that abstract structural priming is automatic and whether there is an inverse preference effect. We discuss these theories below.
First, the residual activation account proposes that the lemma of a verb is connected to syntactic information (Roelofs, 1992), including combinatorial nodes of the structures that the verb allows (Pickering & Branigan, 1998). For example, the lemma of a dative verb "give" is connected to the combinatorial nodes of both DO and PO structures. In this model, structural priming is the result of the residual activation of the combinatorial nodes of a specific structure, which contributes to the repeated choice of prime structure when formulating the target sentence. Moreover, structural priming is strengthened when the prime and target share the verb (i.e. lexical boost). In this case, the repeated activation of the syntactic head (e.g. the lemma of the verb "give") would also strengthen the activation of combinatorial nodes and boost structural priming via the temporarily strengthened link between lemma and combinatorial node. Such spreading activation between lemma and syntactic nodes is usually considered an automatic process that should not be modulated by cognitive resources (Levelt, 1989(Levelt, , 1995Roelofs, 1992). Therefore, this account does not predict inverse preference priming and cognitive load effects on structural priming. Instead, it might predict a preference priming. This is because the more preferent structure might be more frequent and more strongly activates the structure representation in sentence processing, resulting in higher resting activation (Reitter et al., 2011).
Second, the implicit learning account suggests a nonlexicalist abstract representation for syntax (Bock & Griffin, 2000;Chang, 2000;Chang et al., 2006). It suggests that syntactic acquisition is based on an automatic learning mechanism driven by prediction errors. This learning effect is long-lived and contributes to a speaker's language network. Adaptations to the syntactic representations would be especially strong when an unexpected sentence structure is encountered. For instance, "give" is a DO-biased verb in English (Gries & Stefanowitsch, 2004): It is followed more often by a DO structure than a PO structure. However, when "give" is followed by the unexpected PO structure (i.e. the prediction of a DO based on the verb "give" mismatches the actual input, a PO), the prediction error triggers the weight changes of the units that represent the experienced structure and this would lead to more use of the unexpected PO structure in subsequent sentences. Therefore, this system predicts a stronger priming effect for unexpected structures (i.e. inverse preference priming). A study in Dutch (Bernolet & Hartsuiker, 2010) indeed confirmed these predictions (also see Jaeger & Snider, 2013).
Except for the automatic error-based learning mechanism, this account also proposes a non-automatic explicit memory mechanism to interpret the lexical boost effect, the phenomenon that structural priming is stronger when lexical items, such as the verb, are repeated between prime and target (see Chang et al., 2006, p. 256). This hypothesis is sometimes referred to as a "multifactorial account" of structural priming (Bernolet et al., 2016;Chang et al., 2012;Hartsuiker et al., 2008). It suggests that the repetition of the verb between prime and target serves as a memory retrieval cue. This cue to the short-term memory of the prime structure would make it particularly likely that the structure of the prime sentence is repeated in the syntactic planning of the target sentence. Moreover, this memorybased boost is argued to be short-lived and to decline rapidly if there are several fillers between prime and target (Hartsuiker et al., 2008). Note that the original implicit learning theory from Chang et al. does not predict a cognitive load effect on long-lived priming, because it proposes that the error-based learning system which is responsible for long-lived priming is automatically driven and does not interact with memory resources. Instead, the multifactorial implicit learning account proposes the important role of memory resources on both long-lived priming and short-lived priming, predicting an influence of memory resources on both structural priming and inverse preference priming.
Third, a cognitive model that assumes a non-automatic view of syntactic processing is the ACT-R model of syntactic priming (Reitter et al., 2011). This model explains both structural priming and inverse preference priming in terms of procedures for learning and memory retrieval in production. In this model, declarative memory is responsible for the storage and retrieval of syntactic chunks. The short-term priming (e.g. lexical boost) is a result of spreading activation from a retrieval buffer (i.e. working memory) that holds temporary chunks (e.g. lexical form chunks) to the associated syntactic chunks in long-term memory. In particular, the spreading activation decays over time. Long-term priming effects (i.e. inverse preference priming and syntactic persistence) result from "base-level" learning, in which base-level activation of the syntactic chunks changes with repeated retrieval from memory in language processing (e.g. preferred/frequent structures have higher base-level activation). Similar to spreading activation, the base-level activation also decays over time, but more slowly. Therefore, long-term priming also involves some activation decay, similar to shortterm priming. In contrast to the residual activation model and implicit learning theory, this memory-based ACT-R model predicts an influence of cognitive load on both structural priming and inverse preference priming, because both effects require memory resources for the structure retrieval and base-level learning procedures.
Is structural priming automatic or constrained by resources? Note that implicit learning accounts argue that structural priming is driven by prediction, and there is evidence that prediction is sometimes resource-constrained. Specifically, several studies, both in production and comprehension, suggest that linguistic predictions require cognitive resources like attention and working memory (for a review, see Pickering & Gambi, 2018). For example, Huettig and Janse (2016) found that Dutch comprehenders with larger memory capacity and faster processing speed showed a stronger prediction effect of target nouns based on article gender (neuter vs. common) in visual-world comprehension (e.g. "het paard (the_neuter horse)" or "de piano (the_common piano)"). In another visual world comprehension study, Ito et al. (2018) found that prediction effects on the basis of the verb's lexical semantics (e.g. "the lady will fold/find the scarf") was delayed when the comprehension task was accompanied by a concurrent memory task. Similar results have been found in production. For instance, Zhang et al. (2020) found a weaker structural priming when speakers performed a difficult mathematical task after the prime sentence compared to an easy task. Together, these findings suggest a non-automatic view of linguistic prediction and of structural priming, in line with the claim that sentence processing more generally is constrained by the capacity of cognitive resources rather than fully automatic (Allen et al., 2015;Frank, 2013;Gordon et al., 2001;Hartsuiker & Barkhuysen, 2006;Hartsuiker & Moors, 2017;Levelt & Kelter, 1982;Lewis, 1996;Lewis & Vasishth, 2005;Stowe et al., 1998).
Several studies showed that lexically-mediated priming decays or suffers from cognitive load. For instance, in a question-answering situation, speakers tended to reuse the prepositional form of questions (e.g. "At what time do you close") in answers (e.g. "At five o'clock") during a real-life conversation (Levelt & Kelter, 1982). Importantly, this tendency disappeared when there was an additional clause between questions and answers. This finding suggested that speakers may temporarily retain the recent form in the short-term memory which serves a function that contributes to the syntactic processing of following sentences (Levelt & Kelter, 1982;Slevc, 2011). Moreover, Hartsuiker et al. (2008) found persistence of structural priming even when they presented a number of fillers (i.e. 0/2/6 lags) between prime and target, replicating the long-lived priming observed in previous studies (Bock et al., 2007;Bock & Griffin, 2000;cf. Bernolet, et al., 2016). Importantly, they also found that the lexical boost effect rapidly decayed over lags. This short-lived lexical boost suggests that speakers used a lexical cue to retrieve the structure from explicit short-term memory and used this structure in the target sentences. Moreover, Man et al. (2019) found a lexical boost in healthy adults but not in patients with aphasia who had a short-term memory impairment, again suggesting an important role of memory for the lexical boost in syntactic processing. Therefore, these findings are compatible with both the ACT-R model that assumes spreading activation from lexical chunks to syntactic chunks, and the multifactorial account that suggested two systems: an error-based implicit learning system that is responsible for the long-lived structural priming and an explicitmemory system responsible for the short-lived lexical boost effect. On the latter two-system account, structural priming is at least partially non-automatic.
Abstract (i.e. non-lexically boosted) structural priming also decays or suffers from cognitive load (Bernolet et al., 2016;Zhang et al., 2020). Bernolet et al. (2016) further investigated the role of explicit memory for structural priming when there was no lexical overlap between prime and target. Participants were instructed to carry out either a regular structural priming task or a structure-memory task. In the latter task, speakers memorised the prime structure and then used it in the following target picture description. In both experiments, the lag (the number of fillers separating the prime and target sentences) was varied. Bernolet et al. found that both structural priming and the memory of the structure decreased with lag. These findings suggest that explicit memory not only contributes to the short-lived lexical boost effect, but also to the longer-lived abstract structural priming effect. Further evidence for a role of memory in structural priming was obtained by Zhang et al. (2020). These authors manipulated the cognitive load (i.e. high vs. low) that was exerted by an additional arithmetic problem between prime and target. They found a similar decay of both structural priming and memory of structure when the cognitive load was increased. This cognitive load effect occurred regardless of the lexical overlap between prime and target. They argued that there was a cue-dependent memory retrieval process that serves both the procedures of structural priming and structure memory retrieval. Together, these findings suggest that the component of cognitive resources (e.g. memory) does not only play an important role in the procedure that underlies the lexical boost but also the more abstract, learning-based component of structural priming, supporting a non-automatic view (e.g. ACT-R model and multifactorial account). Moreover, this decreasing priming effect seems to be a short-lived rather than a long-lived persistence effect. For example, in Bernolet et al.'s study, priming of datives was rather small at Lag 2; it was absent at Lag 6. This short-lived structural priming can also be interpreted as a result of fast decay of the activation of the combinatorial nodes in the residual activation account.
On the other hand, structural priming for patients showed a different pattern. Patients with amnesia (n = 4) or short-term memory impairment in aphasia (n = 12) showed similar structural priming as healthy adults, suggesting that explicit memory might not be necessary for syntactic processing Yan et al., 2018). Ferreira et al. (2008) found a comparable magnitude of structural priming in amnesic patients as in healthy adults and this priming effect persisted over lags (0/1/6/10). This finding suggested an automatic, long-lived learning effect that is independent of the memory system. Furthermore, Yan et al. (2018) not only found comparable structural priming but also a comparable lexical boost effect in patients and controls. Importantly, the degree of short-term memory deficits for patients with aphasia was not correlated to the size of the lexical boost effect, suggesting that the lexical boost effect is independent from cognitive resources (e.g. memory). These findings support an automatic view of syntactic processing.
In order to tap into the mechanisms of syntactic processing, we investigate the influence of memory on both structural priming and inverse preference priming. There are three important reasons to conduct this study. First, it is unclear whether structural priming is constrained by memory resources. As reviewed above, previous studies investigated the memory effect on the lexical boost and syntactic persistence for healthy adults (Bernolet et al., 2016;Hartsuiker et al., 2008;Zhang et al., 2020) and on patients with aphasia or amnesia Hartsuiker & Kolk, 1998a;G. Man et al., 2019;Yan et al., 2018), but showed mixed results. Some studies showed memory effects on syntactic choice and sometimes explicit recall of structure (Zhang et al., 2020), but others did not (e.g. Yan et al., 2018). Second, there are several possible interpretations of the decay of abstract structural priming over lags observed in previous studies. It is possible that these effects reflect the influence of memory on structure retrieval in sentence processing, but it is also possible that they reflect a fast decay of the activation of syntactic representation (e.g. residual activation model, Pickering & Branigan, 1998). For instance, in Bernolet et al. (2016)'s study, priming of datives was rather small at Lag 2; it was absent at Lag 6. This short-lived structural priming can also be interpreted as a result of fast decay of the activation of the combinatorial nodes in the residual activation account. Therefore, a manipulation of cognitive load between prime and target sentence can tap into the influence of cognitive resources on abstract structural priming.
Third, the inverse preference priming effect might be more resource intensive than structural priming. For instance, in a visual world comprehension study Chen et al. (2022) found that inverse preference priming started in a later time window than abstract structural priming, suggesting different loci of structural priming and inverse preference priming. This does not seem to be consistent with implicit learning accounts, as these predict a concurrent effect of both structural priming and its modulation by inverse preference. This is because both of these effects would be driven by adaptations to the weights during prime processing as a function of prediction error. Previous studies only tested for a memory effect or a lag effect on structural priming, but it is important to test for the memory effect on both structural priming and inverse preference priming at the same time. If both effects involve the same mechanisms of (memory-demanding) prediction of upcoming structure, they should appear or disappear concurrently across tasks that vary in memory demands.
In the current study, we report four structural priming experiments varying in demand for cognitive resources.
In the experiments reported below, we manipulated prime structure type (i.e. DO vs. PO) and prime sentence bias (i.e. whether the structure bias of a prime verb matches the prime structure or not). Furthermore, we varied the tasks across experiments to assess any influence of cognitive load in both the structural priming and inverse preference priming effects. In the priming task of Experiments 1 and 2, speakers were instructed to read the prime sentence and then use a preamble (Experiment 1) or a given verb (Experiment 2) to describe the following target picture. There was no interval with fillers or other tasks between prime and target. In contrast, in the recognition-memory priming task of Experiments 3 and 4, speakers needed to execute not only the priming task but also an additional recognition memory task (Bock & Loebell, 1990). After reading the prime sentence or describing the target picture, they decided whether this sentence or picture had been presented in the previous trials. Speakers needed to retain these sentences and pictures in memory so that they can execute the recognition task immediately after processing the present prime sentence or target picture. We, therefore, expected cognitive load to increase with increasing exposure to trials throughout the experiment, leading to a high demand of cognitive resources. This recognition task has been used in many production studies, which descriptively seemed to show rather weak structural priming (e.g. Bock et al., 2007;Bock & Griffin, 2000), compared to tasks such as simply hearing or repeating the prime sentence and picture matching (e.g. Bernolet & Hartsuiker, 2010;Branigan et al., 2000;Buckle et al., 2017;Hartsuiker et al., 2008). However, a formal comparison of the tasks that keep materials constant has not been carried out so far.
If syntactic processing is based on a residual activation mechanism, structural priming should be automatic and independent of cognitive resources or prediction errors. Therefore, the residual activation account predicts that structural priming is unaffected by the resource demands of the task and unaffected by the structural bias of the prime. If instead syntactic processing is based on a error-driven learning mechanism, structural priming should be strengthened when prediction errors increase, leading to the prediction of an inverse preference priming effect. Importantly, if there is an external memory system that interacts with the error-based learning system, as assumed by the multifactorial account (Bernolet et al., 2016), structural priming should become smaller when the cognitive load increases with the recognition-memory task in Experiments 3 and 4. Similarly, the ACT-R model also predicts a cognitive-load effect on structural priming, because it assumes that structural priming is supplemented by a memory retrieval process.

Participants
We recruited 60 native Mandarin speakers (50 females and 10 males with an average age of 21). The online experiment was shared via a social platform (i.e. Wechat) to access the target group of university students from China. None of the participants has participated in the other experiments of this study. 1 Before the experiment, participants were instructed to read the consent form; they could only enter the test if they agreed to the form. The study was approved by the ethics committee of the Faculty of Psychology and Educational Sciences, Ghent University. Participants were paid 35 yuan (approximately US$ 5.50).

Materials
We used the same 48 sets of materials as Chen et al. (2022) except that the target sentences were adapted for a production experiment. Note that Chen et al. normed a set of Mandarin verbs for bias in a large sample (n = 367), and selected items that were possible with either a DO or a PO, but displayed a clear bias. Each set involved four prime sentences (Bias-matched DO, Bias-matched PO, Bias-mismatched DO, Bias-mismatched PO; Table 1) and one target picture with a sentence preamble involving a different verb from the primes ( Figure 1). The structure bias of verbs was calculated as the log-odds for a DO or PO response. There were four DO-biased verbs (M = .92, SD = .40), four PObiased verbs (M = -2.01, SD = .34) in prime sentences, and four neutral-biased (slightly PO-biased; M = -.22, SD = .12 2 ) verbs in target pictures. No lexical content was shared between prime and target. We also included 96 filler pairs. These fillers consisted of transitive sentences involving two entities (e.g. "The waitress kicked the cowboy", see Chen et al., 2022;Huang et al., 2016) and target pictures that were accompanied by a sentence preamble. There were 20 verbs in the filler sentences.
In the critical items, the target picture always depicted three entities corresponding to an agent, recipient, and theme. With the sentence preamble (e.g. "The grandpa passes"), the target sentences were constrained to a ditransitive structure. We counterbalanced the spatial position of the three entities on the target pictures to balance out any effects of spatial expectations. For example, in Figure 1 the agent "grandpa" was on the left, the theme "iron" was in the middle, and the  recipient was on the right; we, therefore, denote the order from left to right as "A-T-R". The other three orders were "A-R-T", "R-T-A", and "T-R-A". Additionally, the preambles for the filler pictures constrained the responses to an active transitive structure. Half of the filler pictures showed the agent on the left and recipient on the right; the other half had the agent on the right and recipient on the left. We constructed 4 lists of experimental items in a Latin Square design, with 48 experimental trials for each list and 12 trials for each condition. We presented all 144 trials in a pseudo-random order. The first three trials were fillers and each experimental trial was preceded by at least one filler trial. Participants were randomly assigned to one list.

Procedure
Participants completed the experiment on the online platform "LimeSurvey v3.15" with their own computer. They were redirected to LimeSurvey via a link that they received from the social platform. Participants were instructed to familiarise themselves with all the entities, which were shown with their names underneath the pictures before the experiment. During the experimental phase, on each trial participants read a sentence (e.g. "Wujing song Tegong yiba shouqiang (The policeman gives the spy a gun)") aloud and then clicked the "Next(下一页)" button to trigger the corresponding picture (e.g. Figure 1); they were instructed to use the preamble to describe this picture. All the responses were recorded by participants' smartphones. There were two practice trials before the experimental phase.

Scoring
We coded a response as DO if the target verb was followed by a noun phrase indicating the recipient and then a noun phrase indicating the theme; as PO if the target verb was followed by a noun phrase indicating the theme and then by a prepositional phrase (PP) including a preposition and a noun phrase indicating the recipient; or as "other" responses. We excluded the "other" responses from analysis.

Data analysis
For the data analysis, we used Generalized logistic mixed models (glmer) with the "lme4" package in R (Bates & Maechler, 2009). Because a model with a full maximal random effect structure did not converge, our analyses involved the random effect model with the random intercept and random slopes of prime structure and prime bias for subjects and with the random intercept and random slopes of prime structure and the interaction between prime structure and prime bias for items (Barr et al., 2013). The predictor "Prime structure" indicated the main effect of structural priming. The interaction between prime structure and prime bias indicated the difference of structural priming between the bias-matched condition and bias-mismatched condition (Table 3). Based on this model, we then used the "emmeans" package to compute the contrast between bias-matched and bias-mismatched conditions for both the DO and PO primes. We used deviation coding for both prime structure and prime bias (Scheepers et al., 2017). Data and scripts are available online https://osf. io/49wzu/. Table 2 shows the frequency of target responses and the priming effect (expressed as a proportion) for each condition. We excluded 0.4% other responses. The table shows that descriptively, there is a priming effect of 0.29 in the match condition and of 0.34 in the mismatch condition. Table 3 shows the results of the GLMM fixed effects. The predictor of prime structure was significant (β = 2.63, SE = 0.37, z = 7.06, p < .001), indicating a main effect of structural priming.

Results
Importantly, the interaction between the predictor of prime structure and the prime bias was significant (β = 0.53, SE = 0.27, z = 1.98, p = .048), indicating stronger priming in the bias-mismatch (0.34) than the biasmatch (0.29) prime conditions ( Table 2): Participants showed more priming when the prime structure mismatched the structure bias of the corresponding prime verb than when it matched the structure bias of the prime verb (inverse preference priming). Furthermore, the contrast between the bias-match and bias-mismatch conditions for the DO primes was not significant (i.e. 0.73

Discussion
Experiment 1 demonstrated a clear priming effect of dative structures via an online platform. Participants tended to reuse the structure of the previous prime sentences when they were required to construct a sentence to describe the target picture. Importantly, there was stronger structural priming after the bias-mismatch primes than after bias-match primes, demonstrating an inverse preference priming effect. This finding is consistent with the implicit learning theory, which proposes that structural priming is driven by prediction errors, and the ACT-R model, which assumes that inverse preference priming is driven by base-level learning. In particular, when predicted structure (based, for instance, on verb bias) mismatches the actual sentence structure, the increased prediction error triggers weight changes in the network of syntactic representations, which leads to the increased use of the unexpected structure in the following sentences. This inverse preference priming effect in production is consistent with the findings of a visual-world comprehension experiment in Mandarin that used the same materials (Chen et al., 2022). However, two of the present findings were inconsistent with Chen et al. (2022). First, our speakers showed a slight preference for DO responses in the target sentences (i.e. 57% DO responses), whereas Chen et al. showed a clear preference for the PO structure in the norming data. Second, the inverse preference priming effect did not emerge in an analysis on DO primes separately and was only marginally significant in such an analysis on PO primes, whereas Chen et al. found an inverse preference priming in the DO primes instead of PO primes. Except for the processing modality (production vs. comprehension), an important difference between Experiment 1 and Chen et al.'s Mandarin comprehension study is the predictability of the target structure across the experiment. For example, Experiment 1 only included dative structures (i.e. involving three entities for target trials) and active transitive structures (i.e. involving two entities for filler trials), which clearly indicated the difference between target and filler trials. Moreover, in the pictures, speakers were given a preamble with an agent and a verb of the corresponding event. For example, "Yeye di (The grandpa passes)" for the target picture or "Fuwuyuan ti (The waitress kicked)" for the filler picture, strongly constrained the speaker's responses (i.e. the active structure only for filler pictures and datives only for target pictures). Because of this high degree of constraint (responses could only be actives, DO, or PO sentences), speakers might not require too much effort to retrieve the structure from memory in the syntactic construction of the target sentences. In contrast, Chen et al.'s comprehension study not only included the materials we used as experimental trials in Experiment 1, but also additional filler trials with various structures (i.e. active, passive, shifted-PO, Ba-construction, locative/attributive clause etc.), some of which required reference to three entities, like the target items. Importantly, most of those filler structures can also be used with the target dative verbs, due to the flexible word order of Mandarin (Chen et al., 2022;Li, 1996;Li et al., 1993;Z. Man, 2003). Therefore, the target structure was much less constrained than in our Experiment 1. One possibility is that when the target structure is less constrained, so that selecting a structure is more difficult, speakers might pay more attention to the prime structure, resulting in stronger structural priming than that in Experiment 1.
In order to detect whether the constraint of the target structure influences error-based structure prediction in production, Experiment 2 used similar fillers as our earlier comprehension experiment. Moreover, we used a verb rather than a preamble in the target sentences, which reduced the constraint on the responses even further. If the error-based prediction effect is indeed influenced by the predictability of target structure across the experimental procedure, we should observe a stronger structure prediction effect (both structural priming and inverse preference priming). Additionally, the inverse preference priming should be significant for the DO primes, similar to the previous comprehension studies. Experiment 2: priming task with structure variability

Participants
We tested 85 further participants from the same population as Experiment 1. Twenty-five participants were excluded because they produced Other responses for more than half of the target sentences, leaving 60 participants in the data analysis (12 males and 48 females with an average age of 21). They were paid 35 yuan for their participation.

Materials
The target items (both prime sentences and target pictures) were the same as Experiment 1, except that we provided a verb underneath the pictures rather than a sentence preamble (Figure 2) and that across the board the target structure was less constrained than that of Experiment 1. In particular, we constructed a more diverse set of filler items based on the materials of Chen et al. (2022), which involved a range of structures (i.e. active, passive, shifted-PO, Ba-construction, locative/attributive clause, etc.). Given that the presence of fillers with a PO-shifted or BA-construction (i.e. with emphasis or topicalization of the theme or recipient for ditransitive events, Li, 1996;Li et al., 1993) may influence the structure choice in the target sentences (Cai et al., 2012), we replaced such filler sentences in Chen et al.'s materials with active or passive structures (e.g. "The thief took the professor's magazine", Appendix A). All 144 trials of sentences and pictures involved three entities (two animate and one inanimate).

Procedure and scoring
The procedure was the same as in Experiment 1.

Data analysis
Similar to Experiment 1, we used deviation coding for the predictors of prime structure and prime bias. We analysed their main effects and interaction (see Table 5). Because of model convergence issues (Barr et al., 2013), the random effect structure involved random intercepts, random slopes of prime bias and the interaction between prime structure and prime bias for both subjects and items, and a random slope of prime structure for subjects (Table 4). Table 4 shows the frequency of target responses and the priming effect for each condition. We excluded 15% of the responses. The table shows a priming effect of 0.25 in the match condition and of 0.31 in the mismatch condition. As in Experiment 1, there was a preference for the DO structure in the target responses. The predictor of prime structure was significant (β = 2.36, SE = 0.32, z = 7.28, p < .001), demonstrating a clear structural priming effect. Furthermore, the interaction between prime structure and prime bias was significant (β = 0.89, SE = 0.45, z = 1.99, p = .047), indicating an inverse preference priming effect: structural priming was stronger when the prime structure mismatched the prime verb bias (0.31) than when it matched (0.25). The contrast between bias-matched and bias-mismatch conditions was not significant for PO primes (p = .68), but significant for DO primes (p < .01).

Discussion
Experiment 2 again demonstrated robust structural priming and inverse preference priming in a paradigm without an additional memory task between prime sentences and target pictures. Moreover, the descriptive effect sizes of inverse preference priming in Experiment 2 was similar to that in Experiment 1(0.06 vs. 0.05).
Presumably because of the manipulation of structure variability across the experimental list and the larger range of structural options with a single verb rather than a preamble on the target stimuli, speakers produced more other responses in Experiment 2 than in Experiment 1. Interestingly, speakers showed inverse preference priming for DO primes (i.e. 0.87 vs. 0.80) but not for PO primes (i.e. 0.56 vs. 0.56) in Experiment 2. This time, the pattern of inverse preference priming resembled that in our comprehension study in Mandarin (Chen et al., 2022), consistent with a similar error-based prediction system of implicit learning in production and comprehension. However, even though the target dative structures were less constrained in Experiment 2 than in Experiment 1, the speakers seemed to show descriptively somewhat weaker structural priming than that in   Experiment 1 (0.28 vs. 0.32). In order to test whether the structure constraint influenced structural priming, we implemented a combined analysis of all experiments below that we will discuss after reporting Experiments 3 and 4.

Experiment 3: priming + recognition memory task
In order to investigate whether demand on cognitive load modulates the structural priming that we observed in Experiments 1 and 2, we now added a recognition memory task (Bock & Loebell, 1990) after the processing of prime sentences and target pictures. Importantly, the recognition task arguably increases speakers' cognitive load in sentence production, because speakers need to continuously remember all the sentences and pictures of the preceding trials, in order to carry out the recognition task with each presented sentence or picture. Thus, the materials encountered before in the experiment exert a continuous and increasing load on memory. In order to increase the difficulty of memorising the materials, we manipulated the similarity of the filler sentences (half of the filler sentences differed by only one argument (i.e. agent, patient, or verb) from the other half of the filler sentences). If memory plays a role during structural priming (e.g. if there is retrieval and reuse of previously used syntactic representations), then an additional recognition memory task should interfere with priming. Therefore, speakers should show weaker structural priming in Experiment 3 than in Experiment 1. If not, they should show a similar priming effect as in Experiment 1.

Participants
We tested 60 further participants (15 males and 45 females with an average age of 22) with the same constraints on participation as in Experiment 1. Participants were again recruited on Wechat platforms and then finished the experiment via Limesurvey. They were paid 35 yuan for participation.

Materials
The experimental items were the same as in Experiment 1, except that the list of fillers (i.e. active transitives) was reconstructed for the purpose of a recognition memory task (Bock & Loebell, 1990). The 48 experimental trials, consisting of prime sentences and target pictures, only occurred once in this experiment. Then, we selected 24 filler sentences and their corresponding filler pictures from Experiment 1 as unrepeated items for the recognition task. In order to increase the difficulty of this task, half of these filler sentences (i.e. 12 sentences) were replaced with new sentences that only had one different argument (i.e. agent, patient, or verb) from the other half of filler sentences. After the first occurrence of the sentence-picture fillers, they were repeated three times in the rest of the experiment, resulting in 72 repeated filler trials. Therefore, the ratio of repeated and unrepeated trials was 1:1.

Procedure
Following Bock and Loebell (1990), participants were instructed to decide whether the given sentence or picture has been shown in the previous trials of the experiment. Therefore, in order to recognise the given stimuli, they needed to memorise all the sentences and pictures of the preceding trials. In addition to the recognition memory task, participants were also instructed to finish the production task, as in Experiment 1. For each trial, they repeated the sentence aloud and then clicked the "Next" button to trigger the recognition task; they decided whether this sentence has occurred in previous trials by clicking the button for "是(Yes)" or "否 (No)" on the screen. After clicking the "Next" button, they saw the corresponding picture and used the preamble to describe this picture; then again, they clicked the button for "是(Yes)" or "否(No)" to decide whether this picture has occurred in previous trials of the experiment. In contrast to the fully random fillers (separate sentences and pictures) in Bock and Loebell's study, we presented the fillers in sentence-picture pairs, which allows for more direct comparison with Experiment 1. Again, participants used their smartphone to record all their responses. Two practice trials preceded the experimental phase.

Scoring
The scoring was the same as Experiment 1.

Data analysis
Again, we used GLMM to analyse the main effect of prime structure, prime bias, and their interaction (see Table 7). Then we accessed the results of the contrast between match and mismatch condition for both DO and PO primes. Due to model convergence issues (Barr et al., 2013), the random effect structure involved random intercepts and random slopes of Prime structure for subjects and items, and random slopes of prime bias and the interaction between prime structure and prime bias for subjects. In order to test whether structural priming decreases with increasing memory load over the course of the experiment, we included the interaction between prime structure and trial order as a predictor into the LME model. Then, we compared it with the model that excluded this predictor. The model comparison was significant (p = .025). Hence, we included the predictor of trial order in our data analysis. 3 Table 6 shows the frequency of target responses and the priming effect for each condition. We excluded 1.3% of the responses. The accuracy of the recognition memory task was 93.83% (SD = 0.05). The table shows a priming effect of 0.08 in the match condition and of 0.11 in the mismatch condition. There was again a preference for DO responses in the target sentences (i.e. 71%). The predictor of prime structure was significant (β = 1.12, SE = 0.22, z = 5.09, p < .001) ( Table 8). Consistent with Experiments 1 and 2, participants tended to reuse the prime structure in the description of the target pictures, although this effect was descriptively much smaller than in the earlier experiments (we report a cross-experiment comparison after Experiment 4). The interaction between prime structure and trial order was significant (β = −0.30, SE = 0.12, z = −2.57, p = .010), suggesting that priming decreased with increasing exposure to the trials throughout the experiment. Importantly, the interaction between prime structure and prime bias was not significant (β = −0.23, SE = 0.36, z = −0.63, p = .53). The difference in the number of DO responses in the bias-matched and bias-mismatched conditions was not significant for either DO (p = .89) or PO primes (p = .23) ( Table 7).

Discussion
In Experiment 3, we again found structural priming when there was an additional recognition memory for prime sentences and target pictures, demonstrating a robust structural priming effect in production. Importantly, structural priming was considerably smaller in Experiment 3 than in Experiment 1 (which did not include a memory task). Thus, it seems structural priming is much smaller with the recognition memory task than without it. Consistently, structural priming effect decreased with increasing exposure of trials while the demand of cognitive resources was also increasing (i.e. because participants need to memorise more materials throughout the experiment). This finding suggests a decreasing priming effect with increasing cognitive load over the course of the experiment.
Although the priming effect was 3% larger in the mismatch than in the match conditions (i.e. 0.11 vs. 0.08), this inverse preference priming effect was not significant. In contrast, the inverse preference effects of about 5% in Experiments 1-2 were significant. It is possible of course that the null effect in Experiment 3 was a false negative, possibly related to the smaller priming effect overall. Alternatively, the inverse priming effect might be directly affected by task demands. We discuss the theoretical implications of such a state of affairs after reporting Experiment 4 and the cross-experiment comparison.
Experiment 4: priming + recognition memory task with structure variability Experiment 3 showed a descriptively small and statistically non-significant inverse preference priming effect. It is unclear whether the experiment failed to detect an existing effect, or whether the inverse preference effect varies with task demands. Experiment 4, therefore, provided a further test of inverse preference priming. Additionally, Experiment 3 only involved active and dative sentences, which might not make a strong demand on cognitive resources. The repeated materials   might be easily recognised in the additional memory task. In order to increase the cognitive load exerted by the recognition task, we now used the fillers of Experiment 2 and presented the recognition memory task of Experiment 3.

Participants
We tested 74 further participants from the same population as Experiment 1 and excluded fourteen participants who produced Other responses for more than half of the target sentences. Thus, we had 60 participants in the data analysis (15 males and 45 females with an average age of 21). They were paid 35 yuan for their participation.

Materials
The experimental materials were the same as Experiment 1. Again, forty-eight sets of experimental materials only occurred once for each participant. Following Experiment 3, we selected 24 unrepeated filler sentences and their corresponding filler pictures of Experiment 2 for the recognition memory task. Half of the filler sentences were replaced with new filler sentences with one argument that was different from the other half of the original filler sentences (i.e. either the agent, patient, verb, or theme). After the first occurrence of the sentence-picture fillers, they were repeated three times in the experiment, so that there were 72 repeated filler trials. Similar to the experimental trials, all filler trials involved three entities (two animates, one inanimate).

Procedure and scoring
The procedure was the same as in Experiment 2, except that the participants were required to use the given verb to describe the pictures rather than a sentence preamble and finish the recognition memory task after they read a prime sentence or described a target picture.

Data analysis
We used deviation coding for the predictors of prime structure and prime bias, and then analysed their main effects and interaction (see Table 9). Due to model convergence issues (Barr et al., 2013), the random effect structure involved a random intercept and a random slope of prime structure for both subjects and items, a random slope of prime bias, and the interaction between prime structure and prime bias for subjects. Similar to Experiment 3, in order to test the influence of trial order on structural priming, we included the interaction between prime structure and trial order as a predictor into the LME model and then we compared it with the model that excluded this predictor. The model comparison was not significant (p = .57). Therefore, we did not include this predictor in our data analysis. Table 8 shows the frequency of target responses and the priming effect for each condition. We excluded 18% of the responses. The accuracy of the recognition memory task was 94.05% (SD = 0.05). The table shows a priming effect of 0.12 in the match condition and of 0.15 in the mismatch condition. Again, there was a preference for DO responses in the target sentences (i.e. 73%). The main effect of prime structure was significant (β = 1.59, SE = 0.26, z = 6.07, p < .001), suggesting a stable structural priming effect. The interaction between prime structure and prime bias was not significant (β = −0.01, SE = 0.44, z = −0.03, p = .98). The contrast between bias-matched and bias-mismatch conditions was not significant for either PO primes (p = .68) or DO primes (p = .78).

Discussion
Similar to Experiment 3, Experiment 4 demonstrated robust structural priming but no statistically significant inverse preference priming effect. However, unlike Experiment 3, we did not observe a decreasing priming with the increasing exposure of trials in Experiment 4. The possible explanation is that, Experiment 4 involved fillers with various structures leading to a relatively high cognitive load in the early trials. Therefore, the changes of cognitive load over trials in Experiment 4 is much smaller than that in Experiment 3 and thus the decreasing priming effect was absent in Experiment 4. Descriptively, both structural priming and inverse preference priming seemed to be much weaker in Experiments 3-4, which made a high demand on memory, compared to Experiments 1-2, which made a lower demand. We defer discussion of this pattern of findings until we have reported a cross-experiment comparison.

Combined analysis of all experiments
In order to investigate the influence of the cognitive load imposed by the task and the structure variability on both structural priming and inverse preference priming cross experiments, we analysed the data of Experiment 1 (i.e. priming task, involving the active structure only for fillers), Experiment 2 (i.e. priming task, involving various structures for fillers), Experiment 3 (i.e. priming + recognition-memory tasks, involving the active structure only for fillers) and Experiment 4 (i.e. priming + recognition-memory tasks, involving various structures for fillers). Then we treated the task (priming vs. memory), type of fillers (active-only vs. variable), prime structure, prime bias, their twoways interactions, three-ways interactions, and fourway interaction as predictors in a linear mixed model (Table 10). The best fit random effect model involved an intercept and slope of the interaction between prime structure and prime bias for subjects, and an intercept, slopes of the four-way interaction between prime structure, prime bias, task and filler, the threeway interaction between prime structure, task and filler, the three-way interaction between prime structure, prime bias and task, and the two-way interaction between prime structure and filler for items.

Results and discussion
There was a significant main effect of prime structure (p < .001), suggesting an abstract structural priming overall the experiments. Importantly, the interaction between task and prime structure was significant (p < .001), demonstrating weaker priming with the additional recognition-memory task. This finding suggests an influence of cognitive resources on structural priming.
The three-way interaction between task, prime structure, and prime bias was only marginally significant (p = .068); thus, it is not clear whether inverse preference priming is reduced with cognitive load. Additionally, there is no compelling evidence that priming was affected by structure constraint (p = .092), the effect size seems rather small (i.e. β = 0.24; descriptively 0.207 vs. 0.210 for Experiments with structure constraint comparing to that without structure constraint). We further discuss the influence of cognitive load and structure constraint in General Discussion Figure 3.

General discussion
In four experiments in Mandarin, we investigated the role of cognitive resources in syntactic processing. In particular, we manipulated prime structure (DO vs. PO) and prime bias (i.e. whether the prime verb bias matches or mismatches prime structure) in two different tasks (priming vs. priming + recognitionmemory task) in sentence production. We reported four critical findings ( Figure 3). First, all experiments clearly showed structural priming. Second, structural priming was weaker in experiments with an additional task (both a priming and recognition-memory task) than in experiments with a priming task only (i.e. Experiment 1 vs. 3; Experiment 2 vs. 4). Third, structural priming was stronger when the prime verb bias mismatched prime structure, suggesting inverse preference priming. However, we only observed inverse preference priming in Experiments 1 and 2, which used a priming task only, but not in Experiments 3 and 4, where speakers' cognitive load increased with an additional memory task. Fourth, neither structural priming nor inverse preference priming effects seem to be affected by the structural diversity of the filler items and thus the degree to which the target sentence structure is constrained. We discuss these findings below. First, we found robust structural priming when the verb was different between prime and target in four web-based experiments of Mandarin production. This finding of abstract structural priming replicates the results from previous lab-based studies of Mandarin production (Cai et al., 2011(Cai et al., , 2012(Cai et al., , 2015Chen et al., 2020;Huang et al., 2016). Such an abstract structural priming was also observed in our previous comprehension study with the same materials (Chen et al., 2022), suggesting syntactic representations operate in a similar manner for both production and comprehension.
Second, abstract structural priming was weaker when cognitive load is increased. This result is consistent with earlier findings. Zhang et al. (2020) found weaker structural priming when speakers' cognitive load increased Note. Experiment 3 and 4 with an "#" marker in the title label are the experiments with an additional recognition-memory task. "Priming" in the y-axis label refers to the numerical priming effect size (i.e. the difference of DO proportion between DO and PO primes). The comparison of priming effects between match and mismatch condition refers to the interaction between prime bias and prime structure. "*" means p < .05 and "n.s." means nonsignificant. Error bars reflect standard errors from a by-participant analysis.
(i.e. a difficult vs. easy arithmetic secondary task). Bernolet et al. (2016) observed a reduction of priming with an increasing lag between prime and target. Additionally, Man et al. (2019) found weaker structural priming in a patient with aphasia (with a memory deficit) compared to healthy adults, which is also indicative of an important role of cognitive resources in syntactic processing. In short, our finding provides evidence for a non-automatic view assuming that syntactic processing demands cognitive resources for grammatical coding (e.g. multifactorial accounts, see Bernolet et al., 2016, Chang et al., 2012, Zhang et al., 2020; ACT-R model, see Reitter et al., 2011), but not for an automatic view that considers syntactic planning as an automatic procedure that does not demand many cognitive resources (e.g. residual activation account, Pickering & Branigan, 1998). Third, we used the same materials as the visual world structural priming study in comprehension reported by Chen et al. (2022) and like our previous studies we observed an inverse preference priming in Experiments 1 and 2 (i.e. larger prediction errors induced a stronger priming effect), this time in production. Importantly, the inverse preference priming in Experiments 3 and 4, in which cognitive load was increased, was much weaker. However, there was only a trend of the cognitive load effect on inverse preference priming in the combined analysis across experiments. These findings suggest the possibility of cognitive-load effect on error-based learning. We further discuss inverse preference priming below.
On the one hand, the inverse preference priming effect that we found in the priming task when speakers read the written prime sentence aloud, replicated the findings in a task when speakers comprehend the audio prime sentence (Bernolet & Hartsuiker, 2010;Jaeger & Snider, 2013). These findings support a shared error-based learning system between production and comprehension, in line with the learning-as-processing assumption of the implicit learning account (Chang et al., 2006;Peter et al., 2015): the learning procedure is based on prediction errors, leading to the weight changes of representation, resulting in abstract syntactic generalisation in both production and comprehension. Importantly, such an error-based learning effect can also be explained by the assumption of base-level learning in the ACT-R model (Reitter et al., 2011): the weight changes of the association links between the lexical forms and syntactic chunks during the learning procedure results in inverse preference priming. However, inverse preference priming cannot straightforwardly be explained by accounts that consider structural priming a result of residual activation of abstract syntactic Figure 3. Priming effects in all experiments. Note. Experiment 3 and 4 with an '#' marker in the title label are the experiments with an additional recognition-memory task. 'Priming' in the y-axis label refers to the numerical priming effect size (i.e., the difference of DO proportion between DO and PO primes). The comparison of priming effects between match and mismatch condition refers to the interaction between prime bias and prime structure. '*' means p < .05 and 'n.s.' means nonsignificant. Error bars reflect standard errors from a by-participant analysis.
On the other hand, inverse preference priming was much weaker (and statistically not significant) in Experiments 3-4, which we argue exert a stronger cognitive load than the earlier experiments. Future research needs to test whether there is indeed a modulation of inverse preference priming by load, which would plead for an important role of cognitive resources in errorbased learning. If this finding is robust, it would support a non-automatic view of syntactic processing (e.g. Hartsuiker & Moors, 2017). For instance, it might require cognitive resources (e.g. memory and attention) to make a prediction of upcoming structure, to match it to the actual input, or to make the appropriate changes. Error-based learning might therefore fail if the participants have insufficient cognitive resources to perform this processing. The absence of inverse preference priming with high cognitive load seems consistent with some of the findings in Bernolet et al. (2016), where structural priming disappeared with a lag of six fillers between the prime and target. This finding can be explained by the non-automatic view of implicit learning (e.g. the multifactorial account (Bernolet et al., 2016;Chang et al., 2012) or the ACT-R model (Reitter et al., 2011)).
Fourth, both structural priming and inverse preference priming effects seem to not be affected by the structural constraint of the target sentences. In particular, there was no clearly weaker priming for the experiments with structure constraint (Experiments 1 and 3) than the experiments without structure constraint (Experiments 2 and 4) (i.e. only marginally significant with small effect size). This finding is inconsistent with the task-demand hypothesis, which predicts stronger priming effect in the experiments with less structure constraint. This hypothesis assumes that speakers pay more attention or memory resources on the structure cue of prime sentence when selecting a target structure in sentence production is difficult. However, our recent study of Dutch comprehension supports the taskdemand hypothesis in comprehension, suggesting that comprehenders make prediction when it's needed with unpredictable target structure (Chen & Hartsuiker, n.d). We found an abstract structural priming when filler sentences involved dative verbs with shifted-PO (e.g."De non geeft aan de soldaat een boek (The nun gives to the soldier a book)") and verb-final (e.g. PO, "Men heeft de schurk aan de politie overgeleverd (One has the scoundrel to the police handed over)") structures, leading to low predictability of the target structure, but not when that is predictable (i.e, without shifted-PO and verb-final fillers). One possibility to interpret the inconsistent effect of structure constraint is that its influence on structural priming differs between production and comprehension. It will be interesting to further investigate this possibility in the future studies.
There are some additional findings in our study. The preference of DO structure in the target responses overall in this study (e.g. 0.57% DO responses in Experiment 1) was somewhat unexpected given that Chen et al.'s (2022) norming data of Mandarin dative verbs suggested a PO-structure bias in our target verbs (Mean = −0.22, SD = 0.12). It is possible that the exposure to bias-mismatch primes and DO structure contributed to the increased use of DO, because DO is less frequent than PO. Therefore, we performed a baseline experiment that was based on Experiment 1 but with intransitive prime sentences. 4 The proportion of DO responses in this baseline condition was 0.59, which is comparable to the DO proportion in Experiment 1 (i.e. 0.57). This finding argues against the assumption that exposure of bias-mismatched primes or dative structure contributed to the slight preference for DO responses. There are two further possibilities to interpret this difference in structure preferences between this experiment and the norming study. First, the overall structure bias of experimental dative verbs in the norming study was based on the evaluation of fortyeight dative verbs with overall strong PO bias (Mean = −1.47, SD = 1.34, only 7 slightly DO-biased verbs, 34 strongly PO-biased verbs, and 7 fully PO-biased verbs). The preceding trials with strong PO-biased verbs might have reduced the DO responses for the norming data of our target verbs, because there were no fillers in the norming study. Second, structure bias of dative verbs in written production (i.e. PO biases in norming experiment) might be stronger than that in speech production (i.e. baseline experiment) (van de Velde et al., 2015). However, given that previous studies found clear inverse preference priming of prime verb bias even though they manipulated various structure bias in target verbs at the same time (Bernolet & Hartsuiker, 2010;Fine & Jaeger, 2013;Jaeger & Snider, 2013), the small preference of DO responses in target verbs may not influence the critical pattern of inverse preference priming for prime verbs in our experiment.
One limitation of the cognitive-load effect in our study is that, the influence of the recognition memory task on both priming effects (i.e. abstract structural priming and inverse preference priming) could be related to processing of other relevant features rather than just memory. This is because, in the recognition memory task, speakers might not only require memory resources to memorise the sentences and pictures, but also unfold linguistic processing (e.g. semantic or syntactic feature that is related to sentence processing) to access the representation of the sentences which might influence structural priming.
Additionally, the cognitive load effect was tested with a between-subject design (i.e. the recognition memory task was tested throughout the experiment in Bock and Loebell (1990)), which might also have contributed to the difference in the priming effects. In a further experiment, we therefore recruited 96 participants and tested the cognitive load effect with a further within-subject experiment. This experiment presented two blocks with either priming task or dual tasks with SVO fillers. However, we observed a strong carry-over effect in experimental sessions where the dual task occurred in the first block and the priming-only task in the second block. This carry-over effect suggested that participants persisted in memorising the sentences during the second block (even though they were no longer required to), leading to a rather weak priming effect (e.g. 0.03 or 0.06 in the priming task of the second block). 5 Future studies could therefore use a working memory task with non-linguistic materials (e.g. memory span of numbers) to control the use of memory resources for an additional task between prime and target with a within-subject design and further investigate the role of memory (e.g. working memory) in syntactic processing. There is one possible confound between the cognitive load and time lag in our study. In particular, the experiments with additional memory tasks (Experiments 3 and 4) induced a longer time lag than the experiments with priming task only (Experiments 1 and 2), because participants need to finish an extra task which costs more time. However, the passing of time per se might has little substantial effect on priming effects. For instance, Hartsuiker and Kolk (1998b) presented conditions with or without a 1000 ms delay between prime and target stimuli and observed no difference in the priming. It is therefore much more likely that the effects observed in our experiments are due to cognitive load differences rather than just the passing of time.
The cognitive-load effect on structural priming that we found in production is consistent with a non-automatic view of prediction. This non-automatic view is also supported by several comprehension studies, which found that both syntactic and semantic processing require cognitive resources like attention and working memory (Huettig & Janse, 2016;Ito et al., 2018; for a review, Pickering & Gambi, 2018). Together, these findings suggest a modality-independent prediction mechanism that demands cognitive resources to predict the upcoming information, suggesting the possibility of interference between production and comprehension processing (e.g. prediction-by-production mechanism, see Pickering & Gambi, 2018). According to the cross-modality prediction mechanism, we expect the cognitive-load effect on structure prediction in comprehension. For instance, abstract structural priming, the lexical boost effect, and inverse preference priming in comprehension should also be constrained by cognitive load as in production. Therefore, it would be interesting for the future studies to investigate these hypotheses.
In conclusion, we investigated the effect of cognitive load in structural priming in four experiments. We found that: (1) abstract structural priming became much weaker in experiments with dual tasks when speaker's cognitive load increased than in experiments with a priming task only; (2) inverse preference priming seemed to be weaker when cognitive load increased. These findings indicate that there is an important role of cognitive resources in syntactic processing and therefore support a non-automatic view of structure prediction. Notes 1. Participants were only allowed to participate in one of the experiments (including the baseline experiment in the general discussion). In the online payment procedure, their personal information (name and account) was checked. 2. Given that it was impossible to find verbs that were exactly equally biased, the verbs that we chose here were slightly PO-biased. 3. We also analysed the predictor of trial order in Experiments 1 and 2. There was no significant interaction between trial order and prime structure. 4. We replaced the dative prime sentences with intransitive sentences in Experiment 1 (e.g., "The baby cries", as baseline) and then tested the distribution of structure responses with the same materials. We recruited 24 native Mandarin speakers (19 females and 5 males with an average age of 22) via the same social platform (i.e., Wechat). 5. Participants in the within-subject experiment were recruited online from the similar population of the experiments in our study. None of them has participated in the previous experiments. Given that the withinsubject design for the priming task and dual tasks might induce the carry-over effect between blocks with different tasks, we did not include this experiment in the current study.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Funding
This work was supported by the China Scholarship Council (CSC) and The Development Project for Young Researchers of South China Normal University (22KJ06).