1.1 Models of Sentence Production

A psycholinguistic theory (or model) of sentence production makes explicit what happens in a speaker’s mind when they turn a prelinguistic intention to communicate something into a linguistic expression that can be articulated as either spoken, written, or signed language. Sentence production is conceptually driven: speakers start out with a meaning (encoded in a preverbal message) and end up with a form (a phonological representation). At the same time, there is consensus among psycholinguists that production is not a shallow or direct mapping process from thinking to speaking: production proceeds through separable representational steps, corresponding broadly to the divisions between semantics, syntax, and phonology, whereby linguistic units of different kinds are selected, retrieved, and assembled online.

Since the seminal work of Garrett (Reference Garrett and Bower1975), psycholinguists have worked within a consensus model of production with a two-stage architecture in which lexical semantic representations are retrieved independently of form-based, phonological representations (Bock, Reference Bock, Miller and Eimas1995; Bock & Ferreira, Reference Bock, Ferreira, Goldrick, Ferreira and Miozzo2014; Bock and Levelt, Reference Bock, Levelt and Gernsbacher1994; Dell, Reference Dell1986; Ferreira, Morgan, & Slevc, Reference 73Ferreira, Morgan, Slevc, Rueschemeyer and Gaskell2018; Ferreira & Slevc, Reference Ferreira, Slevc and Gaskell2007; Garrett, Reference Garrett and Butterworth1980a; Reference Garrett and Fromkin1980b; Reference Bock1982; Reference Baker1988; Levelt, Reference Levelt1989; Roelofs & Ferreira, Reference Roelofs, Ferreira and Hagoort2019). Following Ferreira and colleagues (Reference 73Ferreira, Morgan, Slevc, Rueschemeyer and Gaskell2018), we will call this the two-stage consensus model for sentence production.

To illustrate the main components of the consensus model, we will consider the grammatical encoding stages of a simple active English sentence (see Figure 2), suitable as a description of the transitive event of a “ball hitting a boy” (see Figure 3). One type of elicitation task used in sentence production studies, including structural priming studies, involves the use of pictures of events with two or three characters engaged in some kind of action, as in Figure 3. Participants are instructed to describe “what’s happening in the picture” using one sentence. The combination of providing speakers with a picture and giving them instructions constrains the production task in important ways. First, pictures constrain the types of ideas or messages speakers are likely to form. This is an important condition in priming studies, where the interest is mostly frequently in seeing what the effect of a prime sentence is on production, while controlling for aspects of the message. Second, the instruction to use one sentence also constrains speakers to focus on describing an event in isolation, and not, for example, use the image as a prompt for a longer story or narrative.

Figure 2 Grammatical encoding model for the generation of the sentence “The ball hit the boy.”

Figure 3 Picture used to elicit a transitive sentence.

2.1 The Use of Structural Priming to Address Representational Questions

Early priming studies were designed to inform models of grammatical encoding. Two important questions addressed via priming are: (1) the division of labor between lexical representations and structural representations; (2) the relationship between semantic representations and syntactic representations. We will review these studies and discuss whether the results are consistent or with the consensus model outlined in Section 1, or whether they suggest modifying the model.

More recently, a general awareness across psycholinguistics of the benefits of having well-powered replications of earlier studies, along with advances in statistical methods, have led researchers to attempt to replicate and extend several of the original priming studies. Recent studies have also extended the use of structural priming paradigms to investigate a broader range of representational questions than they were originally developed for, such as the nature of sentence-level semantic representations. The potential of structural priming as an experimental method to investigate syntax and semantics is increasingly being recognized outside of the psycholinguistics of sentence production.

2.1.1 Priming and the Separation between Lexical and Structural Processing in Sentence Production

The seminal study by Bock (Reference Bock1986b) examined abstract (i.e., lexically independent) priming. Recall that all nouns and verbs differed across prime and target trials: if the prime sentence had the verb punch and the nouns were fans and referee (as in 1.1) target pictures depicted a different kind of action (strike) and nominal referents (e.g., lighting, church).

Additional evidence came from Bock (Reference Bock1989), who also found that primes with the same surface structure primed one another irrespective of whether they shared the same function words (i.e., prepositions): prepositional dative primes (2.1a) containing the preposition (P) for were equally good as dative primes (2.1b) containing to in eliciting prepositional dative descriptions with to (2.1c). This result is consistent with the view that what matters for priming is shared surface syntactic phrase structure (NP-VP-NP-PP) and that there is no additional contribution from function word identity (P = to). We will revisit this finding in Section 2.1.2.

1. (2.1)

   Primes
   a.	The secretary baked a cake for her boss (prepositional dative with for)
   b.	The secretary took a cake to his boss (prepositional dative with to)
   Target
   c.	The girl is giving a flower to her teacher (prepositional dative with to)

To more carefully examine the separable contributions of lexical and structural repetition in structural priming, however, one needs to orthogonally manipulate both factors. Priming paradigms are well suited for this. In a series of experiments using structural priming in the written modality, Pickering and Branigan (Reference Pickering and Branigan1998) manipulated sentence structure (DO dative, PO dative) and verb repetition (same verb, different verb) across primes and targets. Participants were given a printed booklet with a list of written sentence fragments, and were asked to complete each fragment in writing with the first thing that came to mind, provided that the fragment and the completion resulted in grammatical sentence in English. In order to make the task as similar as possible on both prime and target trials, primes were also presented as fragments that participants were asked to complete. Prime fragments were constructed so as to bias the structure (DO, PO) with which they could be completed. For example, a prime fragment like 2.2a, where the verb show is followed by a direct object makes it more likely that people will complete the sentence with a DO structure. Target fragments ended at the verb (e.g., showed …).

1. (2.2)

   Written prime fragments
   a.	The racing driver showed the torn overall … (po-inducing, same verb)
   b.	The racing driver showed the helpful mechanic … (do-inducing, same verb)
   c.	The racing driver gave the torn overall … (po-inducing, same verb)
   d.	The racing driver gave the helpful mechanic … (do-inducing, different verb)
   Target
   e.	The patient showed …

The results showed abstract structural priming – that is, priming when the verbs in the prime were different from the verbs in the targets (e.g., give in the prime, show in the target). Fragment primes that induced POs (like 2.2c) led to more PO target completions (e.g., the patient showed the prescription to the doctor) than did DO-inducing prime fragments. Importantly, Pickering and Branigan (Reference Pickering and Branigan1998) also found an enhanced priming effect when the prime and target used the same verb, (e.g., show in the prime and show in the target). This enhanced priming effect has since been referred to as the lexical boost in structural priming. Before we continue, a terminological observation is in order. Although the original Pickering and Branigan study correctly used the term “verb repetition” to refer to the condition with the same verb in the prime and the target, much subsequent priming literature, including reviews and meta-analyses (e.g., Pickering & Ferreira, Reference Pickering and Ferreira2008; Mahowald et al., Reference Mahowald, James, Futrell and Gibson2016) uses the term “overlap” instead of lexical repetition. Following the convention in the later priming literature, we will on occasion use the term overlap to refer to conditions where some dimension of the prime (e.g., lexical, semantic, syntactic) is repeated, or shared, between the prime and the target.

In order to test that the locus of priming is indeed at the more abstract grammatical level of the verb lemma, not the verb form or lexeme, Pickering and Branigan investigated whether repetition of morphological features, specifically tense (2.3a), aspect (2.3b), and number (2.3c) would increase priming magnitude. Recall that in the consensus model the different verb forms in 2.3 (i.e., show/shows/showed/showing) are subsumed under one abstract lemma [show]_v. None of these additional form manipulations changed the magnitude of the lexically boosted priming effect.

1. (2.3)

   Written prime fragments
   a.	The racing driver shows/the mechanic/the torn overall (do/po prime, different tense)
   b.	The racing driver was showing/the mechanic/the torn overall (do/po prime, different aspect)
   c.	The racing drivers show/the mechanic/the torn overall (do/po prime, different number)
   Target
   d.	The patient showed …

On the basis of the finding that priming is lexically (verb) independent, priming is boosted when verbs are repeated, but feature repetition does not matter, Pickering and Branigan (Reference Pickering and Branigan1998) argued that the results provide experimental evidence for the lemma/lexeme distinction. They also proposed a modification and extension of the lemma level for the representation of verbs in the consensus model, following up on the lemma model of Roelofs (Reference Roelofs1992; see Figure 4).

Figure 4 Lexicalist verb lemma model for sentence production.

With respect to priming, two important features of Pickering and Branigan’s account are that the mechanism assumed for both abstract and lexically boosted priming is activation of lexical-syntactic representations at the lemma level. This account is known as the residual-activation account of structural priming. Each verb lemma is linked to its argument structure options in configurations that Pickering and Branigan call combinatorial nodes, as part of a speaker’s long-term linguistic memory. So, for example, when a dative verb like give is used in a double object structure (e.g., “Bert gives Ernie a feather”), the NP_NP combinatorial nodes are activated. When give is used in the prepositional object (“Bert gives a feather to Ernie”), the NP_PP nodes are activated. Abstract priming is modeled via residual activation of the combinatorial nodes linked to the verb. Lexically boosted priming results from the combined residual activation of the verb lemma and of the combinatorial nodes.

The Pickering and Branigan model integrates a linguistic account of verb argument structure with a psycholinguistic processing model. There is a long tradition in linguistic theory that considers the main verb as the key element in a sentence (Chomsky, Reference Chomsky1965; Grimshaw, Reference Grimshaw1990; Levin, Reference Levin1993; Levin & Rappaport Hovav, Reference Levin and Rappaport Hovav1995; Pinker, Reference Pinker1989). Because of the central role of the verb in determining its argument structure, this model is lexicalist: it assumes that the meaning and syntax of a sentence are projected from the properties of the main verb.

2.2 Why Does Structural Priming Occur?

Thus far, we have looked structural priming studies that were aimed at representational questions. These studies have contributed greatly to our understanding of the architecture of the sentence production system, specifically the stage called grammatical encoding. We also saw that structural priming is useful for the experimental investigation of linguistic representations in adults. In Sections 3 and 4, we will also see how priming applies to investigating representational questions in children and speakers of more than one language.

In this sub-section, we will follow another important line of inquiry into the functions and the cognitive mechanisms of structural priming in adult speakers. Two different perspectives on priming have contributed to this topic: one comes from the study of priming in dialogue settings, the other one comes from connecting priming to life-long language learning. We will examine each perspective in turn.

2.2.1 Priming as a Mechanism for Alignment in Dialogue

Pickering and Garrod (Reference Pickering and Garrod2004, see also Pickering & Garrod, Reference Pickering and Garrod2013) were the first to propose that the functions of priming, both lexical and structural, are best understood within the context of a psycholinguistic theory of dialogue.

The psycholinguistic model of dialogue is called the interactive alignment model. Pickering and Garrod propose that during communicative activities speakers align representations at multiple levels: phonetic-phonological, lexical, semantic-pragmatic, and syntactic. On this view, structural priming is an abstract and implicit form of alignment. Because the mechanism is implicit and automatic, it does not require deliberate allocation of resources, thus serving the function of making an incredibly complex task (comprehending and producing language) easier for both producers and comprehenders.

Pickering and Garrod’s alignment model is massively interactive: they assume interactivity between all levels of representation, both within interlocutors and between them. Alignment at one level (e.g., lexical) leads to alignment at other levels (e.g., syntactic), which is the explanation for the lexical boost effect. Pickering and Garrod argue that production in monologue settings (the classic approach to the study of sentence production) is only a special case of dialogue. What distinguishes monologue from dialogue is the amount of alignment between interlocutors. Alignment (via priming) is the reason why speakers can often complete their interlocutor’s sentences and why the units of analysis in dialogue may not be the classic model of a sentence produced from scratch. The prediction of this account, is therefore, that priming effects in dialogue settings should be greater than in monologue settings. This prediction is borne out.

Branigan, Pickering, and Cleland (Reference Branigan, Pickering and Cleland2000) developed a priming paradigm suitable for the experimental study of priming in dialogue. This paradigm has been hugely successful in eliciting priming effects in adults. The “priming in dialogue paradigm”, called confederate scripting, involves the presence of an interlocutor who, unbeknownst to the participants, is a confederate of the experimenter and provides the primes. In this classic dialogue paradigm, the confederate and the participant sit at table across from each other, and play a dialogue game involving cards to be described, placed in front of them. A screen between them prevents them from seeing each other’s cards (see Figure 5).

Figure 5 Typical setup for a spoken sentence dialogue production priming study with a live confederate.

Instructions to the participants specify that the goal of the study is to understand how people communicate when they cannot see each other. To make the game setup of the experiment work, at the beginning of the experiment, participant and confederate are presented with two ordered sets of cards each: a set of cards to be described placed face down in a box. This is the description set. The other set of cards is already displayed in an array, face up in front of the participant. This is the selection set. When the confederate describes a picture, the participant has to verify whether the description matches any of the pictures in their selection set. If it does, the card is then placed into a discard box.

The confederate follows a predetermined script in describing the cards, and the form of the confederate’s description (e.g., active/passive, DO/PO) is systematically manipulated. This is the equivalent of a priming manipulation in monologue. Using the confederate-scripting technique, it has been possible to show that structural priming effects are much larger in dialogue settings than in monologue settings. Confederate scripting techniques can be adapted to examine how sentence production is influenced by many interactional, social, and sociolinguistic variables, which are harder to study in monologue settings.

Hartsuiker and colleagues (Reference Hartsuiker, Bernolet, Schoonbaert, Speybroeck and Vanderelst2008) developed a computerized written dialogue variant of the of the priming paradigm with a confederate. Instead of sitting in front of a real person like in the face-to-face dialogue version, participants were told they are chatting live via the computer with a real person in another room. The task was to describe pictures to each other using the chat line, and to verify each others’ descriptions. The scripted set of primes was predetermined and typed out by the computer, instead of a real interlocutor. Hartsuiker and colleagues (Reference Hartsuiker, Bernolet, Schoonbaert, Speybroeck and Vanderelst2008) primed datives (PO/DO) with and without verb overlap, and found large priming and lexical repetition (verb repetition) effects in the written dialogue paradigm. Importantly, these effects were comparable in magnitude to the priming and lexical repetition effects observed in spoken dialogue and much larger than those found in written monologue (Pickering and Branigan, Reference Pickering and Branigan1998) and spoken monologue (Branigan et al., Reference Branigan, Pickering and Cleland2000) versions. In two additional experiments, Hartsuiker and colleagues also examined the time-course of priming effects, which leads to the next important set of questions developed in the priming literature, concerning priming mechanisms.

2.2.2 Abstract Structural Priming as Implicit Learning

In all of the priming studies reviewed this far, prime sentences immediately preceded targets (i.e., there was no additional intervening material between prime and target). One important question addressed in this section concerns the duration of priming effects, as it turns out that answers to this question provide information about the mechanisms that support priming.

Studies that examine the duration of structural priming often use the term “long-term” or long-lasting” to refer to priming effects that persist when the prime and the target are separated either by intervening distractor trials, or a short time interval, in the order of minutes. This differs from other areas of psychology, where long-term priming is defined on the scale of days or weeks (Moutsopoulou et al., Reference Moutsopoulou, Pfeuffer, Kiesel, Yang and Waszak2019). We will follow the convention adopted in the structural priming literature and talk about long-term priming even when it occurs on this shorter time-scale for simplicity and consistency with the primary literature. Although there are important reasons to examine the effects of priming over longer time-scales, especially in studies that use priming procedures as an intervention, the question of whether priming decays immediately, or whether it persists over distractor trials, has provided crucial information to understand how priming works and how it is connected to learning.

Starting from the 2000s, a number of studies examined whether priming decays rapidly or whether it persists over intervening filler trials. Bock and Griffin (Reference Bock and Griffin2000) examined the time-course of abstract priming with active/passive, and DO/PO primes and targets in a paradigm in which a variable number of filler trials (up to 10) was inserted between the prime and the target (See Figure 6). Fillers were constructed so as not to compete directly with either active/passive or DO/PO priming, so they included a variety of sentence structures but did not include transitive or dative sentences. Likewise, pictures were chosen so as to not lead to transitive or dative descriptions. Priming effects were found at all lags (from lag 0 to lag 10), and interestingly, the magnitude of these effects was undiminished over lags.

Figure 6 Typical sequence of events in a structural priming experiment with a lag manipulation.

Recall that the Residual Activation Model of structural priming (Pickering & Branigan,Reference Pickering and Branigan1998; see also Cleland & Pickering, Reference Cleland and Pickering2003) was proposed to account for both abstract and lexically boosted priming. Abstract priming was due to residual activation of the combinatorial nodes connected to the verb lemma, the lexical boost effect came from the combined residual activation of a repeated verb lemma node and the residual activation of the link between the lemma node and the combinatorial nodes. Because activation decays rapidly, a purely activation-based account predicts that priming should decay rapidly with time or interfering trials. Bock and Griffin took their results to argue against an account of priming that relies only on Residual Activation, and suggested an account that views priming as the reflection of the implicit ability of the production system to adapt or change in response to linguistic experience. In other words, they proposed that priming was a form of implicit learning. Additional studies have compared the magnitude and time course of priming effects in versions of the task where participants repeat the prime (see Bock & Griffin, Reference Bock and Griffin2000) to those observed in variants of the task where participants don’t repeat the prime (Bock et al., Reference Bock, Dell, Chang and Onishi2007) finding similar effects (for relative clauses, see Tooley & Bock, Reference Tooley and Bock2014; for actives and passives, see Litcofsky & van Hell, Reference Litcofsky and Van Hell2019). In an important study comparing the time-course of abstract priming and lexically boosted priming over lags (0, 2, 4, and 6) in both the spoken and the written modality, Hartsuiker and colleagues (Reference Hartsuiker, Bernolet, Schoonbaert, Speybroeck and Vanderelst2008) reported a large lexical boost effect at lag 0, but crucially, the boost disappeared after lag 0. Abstract priming persisted and was still there at lag 6, independent of modality.

The connection between priming and learning has since been implemented in many different computational models (Chang, Reference Chang2002; Chang, Reference Chang2009; Chang et al., Reference Chang, Dell, Bock and Griffin2000; Chang, Dell, & Bock, Reference Chang, Dell and Bock2006; Chang, Janciauskas, & Fitz, Reference Chang, Janciauskas and Fitz2012; Dell & Chang, Reference Dell and Chang2014; see also Jaeger & Snider, Reference Jaeger and Snider2013; Reitter, Keller, & Moore, Reference Reitter, Keller and Moore2011, among others).

Chang and colleagues were the first to model structural priming using a type of neural network model called a simple recurrent network (Chang et al., Reference Chang, Dell, Bock and Griffin2000; Reference Chang, Dell and Bock2006; for a useful overview tying together comprehension, production, priming and learning, see Dell & Chang, Reference Dell and Chang2014). The distinctive property of recurrent networks is that they can be trained to learn sequential patterns, hence the idea that they could be suited to learn how to produce sequences of words. In neural networks, information is represented in a distributed fashion over series of interconnected nodes arranged into layers (e.g., an input layer, an output layer, and one or more intermediate or “hidden” layers). Each node (the digital analog of a biological neuron) is capable of receiving and sending numerical signals or activation according to a mathematical function, and activation is said to spread throughout the network. The input activation to each node is a mathematical combination of the activation levels of the nodes connected to it from the previous layer. The output activation of a node is computed based on the combination of its input activations and the type of activation function. How the network behaves depends on the numerical strength of the connections or weights between nodes. Recurrent neural networks that are trained to learn sequential patterns use information about their prior state to make predictions about each upcoming element. The comparison between what the model predicts and what it observes (either in the initial training set, or in the prime) generates a prediction error signal that the model then uses to adjust its connection weights so that its future predictions are more accurate. This is known as backpropagation, or error-driven learning. Learning via weight changes, however, does not happen instantaneously; therefore, error driven learning cannot account for immediate priming effects, but it can account for overall priming during the course of an entire experiment.

Chang and colleagues (Reference Chang, Dell, Bock and Griffin2000)’s recurrent neural network model was built and trained, using error-driven learning, to produce novel English sentences corresponding to transitive, dative, and locative messages. The model had separate components for lexical and for structural processing, so it is referred to as a dual-path model. The message layer represented the meanings of transitive, dative, and locative structures, using event semantic representations. The output layer (called the sequencing layer) specified both hierarchical and linear constituent orders. The model was trained to map messages onto sequences of words, one word at a time, according to the rules of English, so the model output was a sequence of words with the correct word order for the message. The same model and the same mechanism were then employed to simulate all known behavioral priming effects at the time, including priming when the prime and the target were separated by filler trials, as in the study by Bock and Griffin (Reference Bock and Griffin2000). To simulate priming, the model was set to produce a prime structure (or sequence, in the model) with error-based learning “on.” Over trials, this led to weight changes throughout the network, biasing the system to produce a similar structure to the one presented in the prime. Because weight-changes happen gradually, immediate effects such as priming at lag 0 and the lexical boost cannot be accounted for via changes in the connection weights between message-level representations and the sequencing level, so another mechanism is required to account for immediate priming effects. In another version of the model, Chang and colleagues (Reference Chang, Dell and Bock2006) proposed that immediate priming effects are due to an explicit memory trace of the prime sentence. The lexical boost effect was explained with a cue-based retrieval mechanism whereby the repeated lexical element (e.g., the verb) acts as cue for the retrieval of the verbatim prime sentence (Bock & Griffin, Reference Bock and Griffin2000; Chang et al., Reference Chang, Dell and Bock2006; see also Hartsuiker et al., Reference Hartsuiker, Bernolet, Schoonbaert, Speybroeck and Vanderelst2008).

A consequence of connecting priming to learning is the prediction that we should observe inverse preference effects. The reason is that less expected primes will generate a larger error signal, and hence a larger weight adjustment (and greater learning). This could explain why across many studies less frequent structures, such as passives, show larger priming effects than actives, and why encountering a biased-verb in a dis-preferred argument structure also primes more (Bernolet & Hartsuiker, Reference Bernolet and Hartsuiker2010; Jaeger & Snider, 2008; Reference Jaeger and Snider2013).

A more recent and updated computational model (Chang et al., Reference Chang, Janciauskas and Fitz2012) provides a learning account for both short-term and long-term priming, but with a different time-course, hypothetically supported by different memory systems: one fast system relying on explicit (declarative) learning and one slow system relying on implicit (nondeclarative) learning.

The activation-based model introduced by Pickering and Branigan (Reference Pickering and Branigan1998) to account for the lexical boost could not account for long-lasting priming because it had no memory component; however, it is possible to add memory and learning components. Malhotra and colleagues (Reference Malhotra, Pickering, Branigan, Bednar, Love, McRae and Sloutsky2008) did this by implementing an activation-model with additional learning and short- and long-term memory components. This dynamic-activation model captures immediate and long-term abstract priming, as well as lexically boosted immediate priming. Along similar lines, Reitter and colleagues (Reference Reitter, Keller and Moore2011) proposed a neural network model to account for both immediate abstract priming and the lexical boost via spreading activation. Lexical boost effects are accounted for via normal spreading activation. Structural representations (e.g., the passive) in this model are themselves represented in long-term memory with a base level of activation. Retrieval of a structure from memory results in a long-term change to its base level of activation, resulting in persistent and cumulative effects.

In addition to predicting relatively long-lived and cumulative effects during the course of an experimental setting, an interesting and important question is whether these effects persist outside of the priming study itself. In a series of experiments aimed at testing the predictions that priming is learning, Kaschak and colleagues (Coyle & Kaschak, Reference Coyle and Kaschak2008; Kaschak, Reference Kaschak2007; Kaschak & Borregine, Reference Kaschak and Borreggine2008; Kaschak, Kutta, & Schatschneider, Reference Kaschak, Kutta and Schatschneider2011, Kaschak, Kutta, & Coyle, Reference Kaschak, Kutta and Coyle2014; Kaschak, Loney, & Borreggine, Reference Kaschak, Loney and Borreggine2006) tested whether priming effects would also persist and generalize outside of a priming intervention study and whether inverse preference effects would be observed, with greater priming for less frequent structures, as assessed during a baseline phase.

In Kashak’s (Reference Kaschak2007) and Kashak and colleagues’ (Reference Kaschak, Kutta and Schatschneider2011) studies, priming was cumulative in that concentrated input was given for one structure only. The priming phase was preceded and followed by baseline and post-priming phases to examine changes immediately after priming (Kaschak, Reference Kaschak2007) and one week after the experiment (Kaschak et al., Reference Kaschak, Kutta and Schatschneider2011). In the baseline phase, participants provided base rates for PO and DO datives in a fragment completion task similar to Pickering and Branigan’s (Reference Pickering and Branigan1998). The baseline task served to confirm that DO datives for English speakers are more frequent than POs. During the priming phase, participants received cumulative blocked priming with either one or the other structure: they either completed a block of sentence fragments that either induced POs or DOs. Priming effects were found to be larger for the less frequent structure (PO) confirming the inverse frequency effect. Lasting effects were found immediately post-priming (Kaschak, Reference Kaschak2007) and up to one week after the priming intervention phase (Kaschak et al., Reference Kaschak, Kutta and Schatschneider2011).

Another line of inquiry into whether there is an implicit learning component to priming examines priming in clinical populations with deficits in explicit memory, but intact implicit memory abilities, as assessed via neuropsychological tests (Ferreira et al., Reference Ferreira, Bock, Wilson and Cohen2008; Heyselaar et al., Reference Heyselaar, Segaert, Walvoort, Kessels and Hagoort2017). Ferreira and colleagues (Reference Ferreira, Bock, Wilson and Cohen2008), and more recently Heyselar and colleagues (Reference Heyselaar, Wheeldon and Segaert2021), tested participants with anterograde amnesia. People with this condition are impaired in their ability to form (and consequently, to recall) new memories. At the same time, they are still capable of learning some skills implicitly, such as mirror-reading (Cohen & Squire, Reference Cohen and Squire1980). In both Ferreira and colleagues’ (Reference Ferreira, Bock, Wilson and Cohen2008), and Heyselar and colleagues’ (Reference Heyselaar, Wheeldon and Segaert2021) studies, participants with anterograde amnesia showed intact abstract priming effects, in the face of an impaired ability in recognizing previously encountered filler items.

Priming has also been investigated in older adults whose different memory systems are susceptible to differential decline (Heyselar, Wheeldon, & Segaert, Reference Heyselaar, Wheeldon and Segaert2021). In a cross-sectional study with participants aged between twenty and eighty-five years, Heyselaar and colleagues (Reference Heyselaar, Wheeldon and Segaert2021) found that short-term abstract priming did not change with age, but that long-term, cumulative priming (i.e., the tendency to increasingly produce a primed structure over the course of the experiment, irrespective of whether it is immediately primed) did. Older adults were more likely to produce primed structures if they had already produced them during the experiment.

Heyselaar and colleagues (Reference Heyselaar, Wheeldon and Segaert2021) propose that different components of nondeclarative memory underly both short-term and long-term structural priming. They account for short-term priming via residual activation, along the lines of Malholtra and colleagues (Reference Malhotra, Pickering, Branigan, Bednar, Love, McRae and Sloutsky2008) and Reitter and colleagues (Reference Reitter, Keller and Moore2011), but with the difference that the representations are part of implicit, not explicit memory for the prime.

In conclusion, there is general consensus that a multi-faceted account of structural priming is needed in which a short-term component, responsible for the lexical boost, operates along with a longer-term lexically independent mechanism. On many accounts, abstract priming is understood as a form of implicit learning. The view that abstract priming reflects implicit learning has consequences for how we view language use throughout the lifespan: speakers constantly adapt, or tune their language to the input they receive, in some sense, they never stop learning language (Chang et al., Reference Chang, Dell, Bock and Griffin2000; Reference Chang, Dell and Bock2006; Dell & Chang, Reference Dell and Chang2014).

3.1 Priming to Investigate Children’s Early Sentence Representations

Priming has contributed to an important debate in first language acquisition concerning the abstractness of children’s early sentence representations. One view is that children start out combining words with item-specific representations – their early multi-word utterances are assumed to be organized around specific words (e.g., “eat”, but not the category “verb”) or stored sequences of words (e.g., “eat it”, but not “Verb-Noun”) (Tomasello, Reference Tomasello1992; Reference Tomasello2000a; Reference Tomasello2000b; Reference Tomasello2003). Syntactic abstractions are assumed to emerge gradually, drawing on individual experiences with particular lexical items. On other accounts (e.g., Fisher, Reference Fisher2002; Fisher et al., Reference Fisher, Gertner, Scott and Yuan2010; Gertner, Fisher, & Eisengart, Reference Gertner, Fisher and Eisengart2006; Valian, Reference Valian1986), children’s early lexical and syntactic representations are abstract from the beginning of combinatorial language. Children are assumed to have access to lemmas with all of the relevant grammatical information. This is where abstract structural priming paradigms comes in.

Applying the priming logic (see Section 1), if abstract priming is found in a child’s production, then the inference is supported that the child has sufficiently abstract phrasal or sentence-level representations to support this kind of priming. Abstract priming is therefore a stringent test for representational abstraction.

Using priming with young children requires taking into consideration children’s developing cognitive abilities (in particular working memory, attention, inhibitory control) and their developing language production system (McKee, McDaniel, & Garrett, Reference McKee, McDaniel, Garrett, Fernández and Cairns2018). Young children have less practice coordinating all of the different processes required to comprehend and speak (for a discussion, see Valian, Reference Valian, Barner and Barron2016). Working memory (WM) capacity, attention, and inhibitory control develop with age, and these functions are required to select lemmas among competitors, keep lemmas active at the right stage of sentence formulation, and provide precision timing when inserting lemmas into their appropriate positions (Dell, Oppenheim, & Kittredge, Reference Dell, Oppenheim and Kittredge2008). Whereas there is evidence that children’s overall sentence production system is organized qualitatively like the adult system (for an overview, see McKee McDaniel, & Garrett, Reference McKee, McDaniel, Garrett, Fernández and Cairns2018), lexical access is more challenging for young children (McKee, McDaniel, & Garrett, Reference McKee, McDaniel, Garrett, Fernández and Cairns2018). Accordingly, priming studies with children often include a familiarization phase to facilitate lexical access, priming sentences are presented with accompanying pictures, the number of fillers is reduced, and in some cases between-subjects designs are employed.

The first study to use a structural priming paradigm with young children, was conducted by Savage and colleagues (Reference Savage, Lieven, Theakston and Tomasello2003). They examined abstract and lexically boosted priming for transitives (actives and passives) in English-speaking three- to six-year-olds. The study was presented as a turn-taking game between a live experimenter and the child. Agents and themes in both primes (3.1a–3.1d) and targets (3.1e–3.1h) were always inanimate. In addition to primes with no lexical overlap with the target (3.1a, 3.1b) to test abstract priming, there were primes that contained the pronoun it used both for the agent and the patient (3.1c, 3.1d). Under the hypothesis that children’s early sentences first develop around high-frequency templates, or “islands” such as “it’s verb-ing”, “it’verb-ed”, the idea was that prior to developing a more abstract representation, children would use the template from the prime “it is verb-ing it”. Savage and colleagues (Reference Savage, Lieven, Theakston and Tomasello2003) found abstract priming only in six-year-olds. Younger children showed priming only in the conditions with pronouns in the prime (3.1c, 3.1d), which resulted in targets produced with pronouns (3.1e, 3.1f).

1. (3.1)

   Primes
   a.	The digger pushed the bricks
   b.	The bricks got pushed by the digger
   c.	It is pushing it
   d.	It got pushed by it
   Targets
   e.	It is breaking it
   f.	It got broken by it
   g.	The hammer broke the vase
   h.	The vase got broken by the hammer

On the basis of these results Savage and colleagues proposed that three- to four-year-olds lack abstract sentence representations. They also proposed an influential developmental model, according to which children’s early sentences are either lexically specific or organized around high-frequency templates (“islands”) and learning occurs through a gradual process of abstraction over these templates.

Another study, however, found earlier abstract priming for nominal phrases with a prenominal modifier (3.2a) and post-nominal relative clause structure (RC) (3.2b). Branigan, McLean, and Jones (Reference Branigan, Mclean, Jones, Brugos, Clark-Cotton and Ha2005) developed a game-version of the priming paradigm adapting the popular British children’s game “Snap.” The child and the experimenter each have a stack of cards in front of them, and take turns turning over the cards and describing them, looking for matching pairs of cards. When cards match, the first player to shout out “Snap” wins the cards in play. The game continues till one player has won all of the cards. Using this technique with three- to four-year-olds, Branigan and colleagues (Reference Branigan, Mclean, Jones, Brugos, Clark-Cotton and Ha2005) found both abstract priming (3.2c and 3.2d primed 3.2e and 3.2f) and a lexical boost effect, meaning that priming was larger when the head noun in the prime (cat) appeared again in the target card (3.2a and 3.2b primed 3.2e and 3.2f, more than 3.2c and 3.2d did).

1. (3.2)

   Primes
   a.	The blue cat
   b.	The cat that is blue
   c.	The red dog
   d.	The dog that is red
   Targets
   e.	The green cat
   f.	The cat that is green

Huttenlocher, Vasilyeva, and Shimpi (Reference Huttenlocher, Vasilyeva and Shimpi2004) also used structural priming to address the question of abstractness in children’s early sentences. They tested four- to five-year-olds using transitive and dative primes, with a task and materials similar to those in Savage and colleagues (Reference Savage, Lieven, Theakston and Tomasello2003). Huttenlocher and colleagues (Reference Hartsuiker, Pickering and Veltkamp2004), additionally, compared production-to-production priming (Experiment 1: children repeated the prime) to comprehension-to-production priming (Experiment 2: children listened to, but did not repeat the prime). In Experiment 1, children aged 4;5–4;8 (mean 4;8) showed significant priming effects for both transitive and dative trials. In Experiment 2, children aged 4;2–5,7 (mean 4;5) also showed abstract priming effects. Cross-experiment comparisons showed that prime repetition did not make a difference for priming in this age group. In a third experiment with four- to five-year-olds (4;1–5;7, mean 5;3), all of the priming sentences were presented consecutively, in a priming block, followed by a later target block where it was the child’s turn to describe all of the targets without any additional input from the experimenter. Again, they found significant priming for both transitive and dative structures, suggesting a longer lasting effect of the prime beyond immediate (0 lag) priming in children. Shimpi and colleagues (Reference Shimpi, Gámez, Huttenlocher and Vasilyeva2007) tested priming of transitive and dative sentences in a younger group of English-speaking children. They tested three- to four-year-olds, using different variations of the priming procedure: blocked (as in Huttenlocher et al., Reference Huttenlocher, Vasilyeva and Shimpi2004, Experiment 3) and alternating prime-target trials with repetition of the prime (as in Huttenlocher et al., Reference Huttenlocher, Vasilyeva and Shimpi2004, Experiment 1). For the blocked version of the experiment, where the target block was presented after the priming block, they found significant structural priming only for the four-year-olds, but not for the three-year-olds. For the version where primes and targets alternate, three-year-olds (2;6–3;5, mean 3;2) showed abstract priming for both datives and transitives.

Taken together, Huttenlocher and colleagues (Reference Hartsuiker, Pickering and Veltkamp2004) and Shimpi and colleagues (Reference Shimpi, Gámez, Huttenlocher and Vasilyeva2007) provide evidence that abstract priming for transitives and datives is robust in four-year-olds and occurs with or without sentence repetition. In younger children, specifically three-year-olds, abstract priming occurs only when children also repeat the prime. One aspect that leads to qualifying these results, somewhat, concerns the scoring criteria used: Huttenlocher and colleagues (Reference Hartsuiker, Pickering and Veltkamp2004) and Shimpi and colleagues (Reference Shimpi, Gámez, Huttenlocher and Vasilyeva2007) used lax scoring criteria and counted as actives sentences with unexpressed objects (e.g., “the truck is dumping mud”) and counted as passives truncated passives (e.g., “the truck got hit”).

Bencini and Valian (Reference Bencini and Valian2008) tested abstract priming with transitives in three-year-olds (range 2;11–3;6, mean 3;2) using both strict and lax scoring criteria. Bencini and Valian (Reference Bencini and Valian2008) developed different scoring schemes as to what counted as an active or a passive. Lax scoring for passives included truncated passives, passives with uninflected verbs, and passives with incorrect inflections. For example, “the ice cream is melt by the sun” and “the ice cream is melting from the sun” were scored as other in the strict scoring scheme, but passive according to lax scoring. The results showed significant priming effects for both actives and passives, using both strict and lax scoring criteria. When considering lax scoring (but not strict), priming accumulated over trials, with greater priming in the second than the first half of the experiment, consistent with learning over trials.

From the point of view of error-driven learning, it is interesting to note that lax scoring captured those gradient descriptions on a similarity cline with the model presented in the prime. For example, both sentences “the icecream is melt by the sun” and “the icream is melting from the sun” have many of the structural elements of the passive (they are correct at the level of functional processing). What is missing or is incorrect are elements required at the positional level.

Foltz and colleagues (2014) was the first priming study to attempt to relate priming effects to WM capacity. They primed nominal phrases with prenominal modifiers and RC structures, with and without lexical overlap (c.f. Branigan et al., Reference Branigan, Mclean, Jones, Brugos, Clark-Cotton and Ha2005, reviewed earlier) in a group of German-speaking typically developing children (range 4;2–5;7, mean 4;10) and in a group of children with language impairment (range 4;0–5;9, mean 5;10). They also included both lax and strict scoring criteria to also include non-adult attempts at producing primed structures. WM was assessed by having children repeat sequences of numbers of increasing length. They found large priming effects for both groups of children, but neither group exhibited a lexical boost (contrary to Branigan et al., Reference Branigan, Mclean, Jones, Brugos, Clark-Cotton and Ha2005). Importantly, they found a positive relation between working memory and the production of more complex RC structures, and interpreted this result as evidence that sequence memory is particularly relevant to structural priming.

In another study relating structural priming in young children to their WM, Foltz and colleagues (Reference Foltz, Knopf, Jonas, Jaecks and Stenneken2020) used priming to test the abstractness of sentence representations in children before the age of three (range 2;7–2;11). This study provides the strongest evidence for abstract priming before age three, to date. Foltz and colleagues tested three groups of German speaking two-year-olds. Targets were embedded in a question-answering game with a repeated lead-in (What is Emma doing? Emma is.). Children did not repeat primes. Children also provided phonological working memory measures (assessed with a nonword repetition task) and sentence production measures from a developmental language assessment battery. The priming task used infinitive transitives (3.3a) and infinitive intransitives (3.3b). One group of “older” two-year olds (range 2;7–2;11, mean 2;9) participated in a priming study where primes were immediately followed by targets (alternating prime-targets). Two additional groups of “younger” two-year-olds (range for both groups: 2;0–2;6, mean 2;3) were primed with either alternating or blocked primes. Children started with a baseline phase where they provided answers to the lead-in question without a prime. Results showed reliable priming from infinitive transitive primes to infinitive transitive targets compared with infinitive intransitive primes, only for the older group of two-year-olds. For this group, higher sentence production scores were positively related to increased production of infinitive transitive sentences when primed with infinitive transitives. In addition, WM scores were positively related to the primed production of infinitive transitives, but the strength of this effect decreased with age.

1. 1. a. Baby kitzeln (infinitive transitive)

      [baby to tickle/baby tickling]

   2. b. Laufen (infinitive intransitive)

      [to run/running]

In summary, structural priming paradigms and the priming logic have been successfully applied to address highly debated representational questions in first language acquisition. Finding abstract priming – that is, when primes and targets do not share lexical content – at a given age, argues against representations being lexically specific at that age.

3.2 Priming to Investigate Developmental Trajectories in Child Syntax: The Case of Passives

Priming studies in children have contributed empirical data to address the question of why certain structures, such as the passive, are acquired later than related structures, such as actives. It has been proposed (e.g., Borer & Wexler, Reference Borer, Wexler, Roeper and Williams1987; Reference Borer and Wexler1992; Fox & Grodzinsky, Reference Fox and Grodzinsky1998; for a review, see Deen, Reference Deen, de Villiers and Roeper2011) that late mastery of the passive derives from maturational constraints on grammatical representations and computations. As we saw in Section 3.1, however, in structural priming studies children produce passives much earlier than in spontaneous speech. If under priming conditions children produce passives, this suggests that a lack of passives in spontaneous speech cannot be imputed to immature grammars. This also shows that priming paradigms are well suited to disentangle whether the late appearance of certain forms in children’s speech are due to maturational constraints in their grammars, or other constrains such as developing processing abilities, attention, working memory, and so on.

Several studies have used priming to track the developmental trajectories for passives in three- to four-year-olds and six- to nine-year-olds, focusing on representational questions with respect to the semantics of the passive (Messenger, Branigan, & McLean, Reference Messenger, Branigan and McLean2011; Messenger et al., Reference Messenger, Branigan and McLean2012) and on the question of whether the formal aspects of passives are acquired earlier than the form-meaning mappings (Messenger, Branigan, & McLean, Reference Messenger, Branigan and McLean2012). Messenger and colleagues (Reference Messenger, Branigan and McLean2011) found that short (truncated) passive primes (3.4a) primed full passive targets (3.4c) in three- to four-year-olds (range 3;4–4;10, mean 4;1). These results suggest that three- to four-year-old English-speaking children already have a generalized representation for the passive that includes both its full and truncated versions.

1. (3.4)

   Primes
   a.	The girls are being shocked (short passive)
   b.	The sheep shocking the girl (active)
   Target
   c.	The king is getting scratched by the tiger (full passive)

Messenger and colleagues (Reference Messenger, Branigan, McLean and Sorace2012) manipulated the semantic roles in the primes so as to either be identical to those in the target, or to be different. Targets were always pictures of actional verbs (e.g., scratch, kick. push). In the first experiment, they used prime sentences and pictures with actional verbs and agent/patient thematic roles (3.5a, 3.5b) and primes with non-actional verbs (e.g., shock, scare, surprise) and theme/experiencer roles (3.5c, 3.5d). They found that English three- to four-year-olds (3;1–4;11, mean 4;2) were equally primed and produced passives both when thematic roles strictly matched across prime and target and when they did not match.

1. (3.5)

   Primes
   a.	A cow is chasing a witch (active, agent-subject, patient-object)
   b.	A witch is being chased by a cow (passive, patient-subject, agent-oblique)
   c.	A sheep is shocking a girl (active, theme-subject, experiencer-object)
   d.	A girl is being shocked by a sheep (passive, experiencer-subject, theme-oblique)
   Target
   e.	A tiger is scratching a king (active, agent-subject, patient-object)

In a second experiment with three- to four-year-olds (3;4–4;11, mean 4;2), they compared priming from theme/experiencer (3.6a, 3.6b) with priming from experiencer/theme primes (3.6c, 3.6d) and found that they equally primed passives in agent/patient targets.

1. (3.6)

   Primes
   a.	A sheep is shocking a girl (active, theme-subject, experiencer-object)
   b.	A girl is being shocked by a sheep (passive, experiencer-subject, theme-oblique)
   c.	A sheep is loving a girl (active, experiencer-subject, theme-object)
   d.	A girl is being loved by a sheep (passive, theme-subject, experiencer-oblique)
   Targets
   e.	A tiger is scratching a king (active, agent-subject, patient-object)
   f.	A king is being scratched by a tiger (passive, patient-subject, agent-oblique)

These findings show that priming children’s passives is not restricted to an overlap in narrow semantic roles and argue against the view that children’s early passives are limited agent/patient (actional) verbs.

Children produce passives under priming conditions, but why do they fail to do so in spontaneous speech? Why do they produce more errors than adults, and why do they seem to get the syntax of passives right before passive semantics? Priming studies can help eliminate explanations that appeal to children’s immature grammatical systems and point towards processing accounts (Bencini & Valian, Reference Bencini and Valian2008).

Messenger, Branigan, and McLean (Reference Messenger, Branigan and McLean2012) examined the priming of passives in a group of six-year-olds, (range 6;2–6;11, mean 6;7) and a group of nine-year-olds (range 8;8–10;0, mean 9;6) and found that one common developmental error with passives, reverse role errors (saying: The dog is chased by the cat, instead of The cat is chased by the dog), occurred frequently in six-year-olds and decreased with age. Reverse role errors were also reported in Bencini and Valian’s (Reference Bencini and Valian2008) study in three-year-olds (these errors were excluded from strict scoring, but included in lax scoring). The reason why children may produce more reverse role errors in priming studies (i.e., “getting the meaning wrong while getting the syntax right”) has a likely explanation in the development of working memory and inhibitory control mechanisms.

Tracking the development of the utterances that children produce under priming conditions from earlier ages might be revealing of developmental trajectories themselves (what elements of a complex construction are harder or easier to acquire) and inform theories of priming as learning.

3.3 The Time-Course of Lexically Boosted and Abstract Priming in L1 Acquisition

Priming as learning makes a number of predictions concerning the time-course of the lexical boost and abstract priming in both children and adults. Like some of the studies with adults (e.g., Kaschak, Reference Kaschak2007; Kaschak, 2011), longer-lasting priming effects have been found with children aged four-to-five.

Huttenlocher and colleagues (Reference Hartsuiker, Pickering and Veltkamp2004; see Section 3.1) found priming effects for datives and passives following a priming block with either passive or dative primes, and Savage and colleagues (Reference Savage, Lieven, Theakston and Tomasello2006) found priming for passives following a block of passive primes. These effects persisted one week after priming and one month after priming for the group of children who had also been tested again at one week after the priming intervention. In a pretest posttest blocked design with passives, Kidd (Reference Kidd2012) found sustained production of passives at higher rates compared to the pretest.

A few studies have compared the time-course of the lexical boost and abstract priming employing a lag manipulation and also testing both children and adults with the same materials and tasks. These studies are generally consistent with priming as learning: lexically boosted priming is either absent at the youngest ages (Chang et al., Reference Chang, Janciauskas and Fitz2012) or decays rapidly. Abstract priming effects last over a lag 2 manipulation (Branigan & McLean, Reference Branigan and Mclean2016) and extend beyond the immediate experiment itself (Messenger, Reference Messenger2021). The prediction that priming should be larger in young children than in adults, however, has not yet been confirmed (Messenger, Reference Messenger2021).

Chang and colleagues (Reference Chang, Janciauskas and Fitz2012) tested three- to four-year-olds, five- to six-year-olds, and adults in a computerized description task, in which participants heard prime sentences describing short video animations of transitive events and produced descriptions of similar target events. They included a lexical manipulation (same/different verb). They found a large lexical boost effect in adults, a smaller lexical boost in the five- to six-year-old group, and no lexical boost in the youngest group. Importantly, they found equivalent (albeit small) abstract priming effects across all the age groups. These results appear to be consistent with an account of the lexical boost that relies on an explicit memory trace for the prime sentence, along with the assumption that explicit memory for sentences develops with age.

In the interactive snap-card game version of priming, Branigan and McLean (Reference Branigan and Mclean2016) tested English-speaking three-and-a-half to five-year-olds and adults. In addition to a lexical overlap manipulation, there was also a lag manipulation (lag 0 and lag 2) to measure the time course of priming effects. Priming effects in children were large (25%), as was the lexical boost effect (an additional 20%) for immediate priming (lag 0). There were no age effects in children. The lexical boost decayed rapidly at lag 2 (a 3% boost), but the abstract priming effect persisted (15%). This overall pattern of effects did not differ significantly from that found in adults. Consistent with the view that priming is learning, children also showed cumulative priming over trials for passives when considering lax scoring criteria, but not with strict scoring (similar to Bencini & Valian, Reference Bencini and Valian2008). There was no evidence of cumulative effects for adults.

Messenger (Reference Messenger2021) examined short-term and longer-term priming outside of the priming task itself, in children (age range 3;3–4;10, mean age 4;2) and in adults. The study targeted transitive priming, specifically focusing on passives. The study also included a control group, both for the children and the adults. Control groups provided baseline measures for passive production. The priming phase employed the snap-game-technique introduced in Section 3.1, with half of the experimenter’s cards on the prime trials presented in the active and half in the passive. The experiment also included immediate and delayed test conditions, assigned between groups. For the immediate group, the experimenter and the participant simply continued to play the snap-game. The participant’s cards were still transitive events, but crucially, the experimenter’s cards and descriptions were not, so, no additional priming with passives (or actives) was provided. The delayed test condition used the same strategy (i.e., participants were no longer primed with passives). The delayed test condition simply followed another activity (a lexical repetition task). As expected, there were priming effects during the priming phase. In addition, children who had participated in the priming session with actives and passives produced more passives post-priming than children assigned to the baseline condition in which they received no priming. Although Messenger (Reference Messenger2021) used both strict and lax scoring schemes, neither the strict nor the lax scoring scheme showed evidence for larger priming in children than adults. Passives were, in fact, extremely rare in children’s baseline productions; at baseline and at all phases, adults produced a greater number of passives than children.

Priming magnitude in the absence of lexical repetition has been found to be very variable across studies and across structures, so it remains to be determined whether the predicted larger priming effect will show up in children. One possibility is that during the test phase where children listened to intransitive sentences, these sentences in fact acted as primes for simpler sentence structures than passives. The developmental studies reviewed in this section add to the priming studies reviewed in Section 3.1 that report abstract priming effects at the youngest ages.

3.4 Structural Priming in Bilingual Children

The priming logic can be applied to address the question of how early bilingual children represent sentence structures in their two languages, in a variant of the paradigm known as crosslinguistic priming, in which primes are presented in one language and targets are elicited in the other language. Applying the priming logic crosslinguistically, if one or more structures prime across languages, they are recognized as similar by the bilingual mind. The very first crosslinguistic priming study was conducted with adults (Bock & Loebell, Reference Loebell and Bock2003) and most crosslinguistic priming studies to date concentrate on adults; there are only a handful of bilingual priming studies focus on children. Notably, priming studies with adult bilinguals overwhelmingly look at adults who have acquired a second language after acquiring their first one (sequential, or late bilinguals).

In the first crosslinguistic priming study with bilingual children Vasilyeva and colleagues (Reference Vasilyeva, Waterfall, Gámez and Gómez2010) examined crosslinguistic priming bi-directionally (from English to Spanish and Spanish to English) in bilingual five-to-six-year-olds (range 5;2–6;5, mean 5;11). Although the children were exposed to both languages, they were unbalanced, early bilinguals, with Spanish as their dominant language. They used a between-subjects design both for priming direction (English–Spanish, Spanish–English) and prime condition (active, passive), so that for a given child, priming was concentrated on one structure.

1. (3.6)

   Primes (Spanish-to-English)
   a.	La pelota rompió la ventana
   	[The ball broke the window]
   b.	La ventana fue rota par la pelota
   	[The window was broken by the ball]
   Targets
   a.	The dog was washed by the cat
   b.	The cat washed the dog

They found reliable priming from Spanish passive primes (3.6b) to English passives (3.6d) but not from English to Spanish. The authors explained the lack of a priming effect from English to Spanish by appealing to differences in frequency and usage between passives in the two languages.

In a more recent follow-up to this initial study, Gámez and Vasilyeva (Reference Gámez and Vasilyeva2020) tested balanced early bilingual six-year-old children, and found bi-directional priming for passives (Spanish to English and English to Spanish). The study also included a verb repetition condition and interestingly, they failed to find a lexical-boost effect.

In the context of the debate concerning the abstraction of children’s sentence representations, the crosslinguistic priming data in early bilinguals provide supporting evidence for abstraction, as crucially, primes and targets in different languages do not share words.

3.5 Section Summary

Structural priming studies with young children have contributed crucial empirical data to our current understanding of first language development. In general, priming paradigms provide a useful technique to examine children’s grammatical competence while controlling for processing demands associated with lexical retrieval. Priming studies across development have also contributed to our understanding of priming itself. Priming as learning predicts larger effects in less proficient speakers.

Although it is not always possible to compare child priming and adult priming directly because of differences in scoring criteria across labs, some studies show that children exhibit abstract priming before they show lexically boosted priming. In a few studies that adopt lax scoring schemes in addition to strict ones, lax scoring has allowed to detect cumulative effects of priming in children.

Another promising, but still methodologically challenging area of investigation, concerns the relationship between priming and other measures of linguistic and socio-cognitive development.

4.1 Priming to Investigate Sentence Representations in Late Bilinguals

Two main hypotheses concerning how bilingual (or multilingual) speakers represent the syntax of a later acquired language are typically contrasted in the bilingual priming literature: a Separate-Syntax Separate Production System account (De Bot, Reference De Bot1992), according to which highly proficient bilinguals represent and processes lexical and syntactic structures separately, and a Shared-Syntax/Shared Grammatical Encoding System account (Hartsuiker, Pickering, & Veltkamp, Reference Hartsuiker, Pickering and Veltkamp2004), according to which late bilinguals eventually integrate representations for similar syntactic constructions into one shared language-independent representation. For dissimilar constructions, representations are kept language-specific, and separate.

Experiments that employ crosslinguistic priming in both directions (L1 to L2 and L2 to L1) and compare the strength of priming effects across languages to the strength of priming within-languages offer one way to distinguish between different theoretical proposals concerning how multilinguals represent and process syntax in their different languages. If highly proficient multilinguals rely on shared language-independent representations, the prediction is that priming should be equally strong irrespective of priming direction. Additional linguistic variables of interest concern the degree of structural similarity between the two languages, the effects of lexico-semantic similarity in primes and targets (the bilingual equivalent of the lexical boost effect), and the influence of verb bias – that is, whether both languages have the same grammatical encoding options to describe similar events.

The first study to develop the crosslinguistic priming paradigm was Loebell and Bock (Reference Loebell and Bock2003) who adapted the priming with running recognition memory paradigm to examine crosslinguistic priming of transitives (active, passive) and datives (PO, DO) from English to German. The choice of German and English and active/passive versus DO/PO offers an interesting comparison. DO and PO datives have the same constituent order in both English and German, but actives and passives do not: in German passives, the lexical verb occurs in final position (4.2b). Participants were native speakers of German (L1-German) who had lived in an English-speaking country for at least two years, and who were highly proficient in English (their L2). Loebell and Bock found crosslinguistic priming effects for English and German datives (4.1a, 4.1b) in both directions: DO/PO datives in German primed DO/PO datives in English and vice-versa. For transitives (4.2a, 4.2b), the results were different: there was no crosslinguistic priming for passives. Loebell and Bock explained their results in terms of the need for constituent order to be shared for crosslinguistic priming to occur. They also claim that crosslinguistic priming is explained by implicit learning at the level of the linguistic processes that build utterances. In a far-reaching discussion of the implications of these results they also suggest that crosslinguistic structural priming provides a unified psycholinguistic account of bilingual language use (e.g., code switching, when a linguistic element from language A is inserted into an unfolding utterance in language B) and contact-induced language change.

1. (4.1)

   Dative primes (German-to-English/English-to-German)
   a.	Der Musiker verkaufte etwas Kokain an den Agenten/The musician sold some cocaine to the undercover agent/ (po)
   b.	Der Musiker verkaufte dem Agenten etwas Kokain/The musician sold the undercover agent some cocaine/ (do)
   Target
   c.	girl, paintbrush, boy, hand (dative picture)

1. (4.2)

   Transitive primes (German-to-English/English-to-German)
   a.	Eine Gruppe Jugendlicher überfiel den Verkäufer/A gang of teenagers mugged the salesman/ (active)
   b.	Den Verkäufer wurde von einer Gruppe Jugendlicher überfallen/ The salesman was mugged by a gang of teenagers/ (passive)
   Target
   c.	truck, car, tow (transitive picture)

Since the seminal study by Loebell and Bock (Reference Loebell and Bock2003), crosslinguistic structural priming has been found across several pairs of languages. Notably, these language pairs are predominantly Indo-European, and with English as the L2 (English–Spanish, English–Swedish, English–Dutch, English–Chinese, and English–Polish). Although there are studies with other structures, such as genitives and relative clauses, the most frequently primed structures in crosslinguistic research are transitives and datives (Bernolet, Hartsuiker & Pickering, Reference Bernolet, Hartsuiker and Pickering2007; Reference Bernolet, Hartsuiker and Pickering2009; 2012; Reference Bernolet, Hartsuiker and Pickering2013; Cai, Pickering, Yan, & Branigan, Reference Cai, Pickering, Yan and Branigan2011; Desmet & Declerp, Reference Desmet and Declercq2006; Hartsuiker, Pickering, & Veltkamp, Reference Hartsuiker, Pickering and Veltkamp2004; Kantola & van-Gompel, Reference Kantola and van Gompel2011; Salamoura & Williams, Reference Salamoura and Williams2006; 2007; Schoonbaert, Hartsuiker, & Pickering, Reference Schoonbaert, Hartsuiker and Pickering2007; for a reviews, see Hartsuiker & Pickering, Reference Hartsuiker and Pickering2008; van-Gompel & Arai, Reference van Gompel and Arai2017).

An influential study by Hartsuiker and colleagues (Reference Hartsuiker, Pickering and Veltkamp2004), used crosslinguistic priming to propose a bilingual extension of the lexicalist model of grammatical encoding in language production. Using a dialogue-based priming paradigm with a confederate they tested Spanish–English bilinguals (L1-Spanish, L2-English). Passive sentence primes were presented in Spanish (4.3a) and were found to prime English passive descriptions to targets pictures (4.3e) compared to active Spanish primes (4.3b). The study also included an intransitive baseline (4.3c) and a constituent fronting construction (left dislocation) (4.3d) in which the object appeared sentence initially (left dislocated object). The main discourse-function of left dislocation is to topicalize the dislocated constituent. Although, interestingly, left dislocated object primes resulted in numerically more passives than active primes and intransitives, these contrasts were not significant.

1. (4.3)

   Primes
   a.	El camión es perseguido por el taxi (passive)
   	[The truck is chased by the taxi]
   b.	El taxi persigue el camión (active)
   	[The taxi chases the truck]
   c.	El taxi acelera (intransitive) [The taxi accelerates]
   d.	El camión lo persigue el taxi (left dislocated object)
   	[The truck obj (it) chases the taxi subj]
   Target
   e.	missile, boat, hit (transitive picture)

The bilingual production model is known as the shared syntax model. As mentioned at the beginning of Section 4.1, the model resembles (and is an extension of) the monolingual lexicalist model of sentence production proposed by Pickering and Branigan (Reference Pickering and Branigan1998; see Section 2.1.1). Bilingual structural priming is captured at an abstract level, depending on the structures primed. For priming at the level of argument structure, the model is lexicalist, in that it assumes that it is always mediated by verb lemmas (as in Pickering and Branigan, Reference Pickering and Branigan1998). Conceptual nodes representing verb-based relational information (e.g., “chase” concept) are connected to language-specific lemmas (e.g., chase [X, Y] in English, perseguir [X, Y] in Spanish). Lemmas are connected to shared category nodes (e.g., verb) and to shared (language independent) combinatorial nodes (active/passive). Lemmas are also connected to general language nodes (e.g., Spanish, English) that act like a switch, controlling what language is being produced.

The prediction of the shared syntax model is that priming effects should be of equal strength within and between the languages of a highly proficient bilingual speaker: L1 to L1, L2 to L1, L2 to L2, and L1 to L2. Some crosslinguistic structural priming studies confirm these predictions, finding equivalent priming from L1-to-L2 and L2-to-L1 (Schoonbaert et al., Reference Schoonbaert, Hartsuiker and Pickering2007; Kantola & van Gompel, Reference Kantola and van Gompel2011), and even equivalent priming between additional later acquired languages (L2 and L3) (Hartsuiker et al., Reference Hartsuiker, Beerts, Loncke, Desmet and Bernolet2016), while others do not (Cai et al., Reference Cai, Pickering, Yan and Branigan2011). We will first review studies that have reported equivalent priming.

Schoonbaert and colleagues (Reference Schoonbaert, Hartsuiker and Pickering2007) reported comparable within-language and between-language priming for datives in L1-Dutch L2-English speakers (4.4). Although there was a numerically greater priming effect within-language, this was not statistically significant. In a subsequent reanalysis of the data from this experiment, Hartsuiker and Bernolet (Reference Hartsuiker, Beerts, Loncke, Desmet and Bernolet2016) showed that the difference in priming magnitude found in the earlier study was due to speakers’ L2 (English) proficiency. Consistent with the shared syntax model, highly proficient L2-English speakers showed similar abstract priming patterns within their L1 and in their L2, and bi-directionally between the two.

1. (4.4)

   Primes
   a.	The cowboy throws the clown a book/De cowboy gooit de clown een boek (do)
   b.	The cowboy throws a book to the clown/De cowboy gooit een boek naar de clown (po)
   Target picture
   c.	Prisoner, apple, pirate, sell

Schoonbaert and colleagues also included a lexical manipulation – called meaning-equivalent condition – where the verb in the prime and in the target are identical in within-language priming, and are translation equivalents in between-language priming. This is shown in 4.5, where primes (4.5a, 4.5b) use translation equivalent verbs – that is, verbs with the same meaning (English to throw and Dutch gooien) – and in targets 4.5c and 4.5d, where throw is used when the target is English, and gooien is used when the target is Dutch. When priming from L1-Dutch to L2-English, a meaning-equivalent boost to priming was detected – that is, priming was larger when verbs were translation equivalents (gooien, throw) than when they were not (gooien, sell, as in 4.4). This semantic-equivalent boost effect was, however, smaller than the well documented within-language verb repetition effect (the classic lexical boost discovered by Pickering and Branigan, Reference Pickering and Branigan1998, see Section 2.1.2). The results were also asymmetrical. The translation equivalent boost was found going from L1-Dutch to L2-English, but not the other way around. Schoonbaert and colleagues accounted for the asymmetry by assuming that the links between conceptual representations and lemmas in an L2 are weaker than in an L1. Their overall proposal for how bilinguals represent lexical and syntactic information in structures that are similar in their two languages is that lemmas are language-specific, but combinatorial nodes are shared. According to this model, shared conceptual nodes are differentially connected to L1 and L2 lemmas, with stronger concept-to-lemma connections in an L1 than in an L2.

1. (4.5)

   Primes
   a.	The cowboy throws the clown a book/De cowboy gooit de clown een boek (do same verb/ meaning equivalent verb)
   b.	The cowboy throws a book to the clown/De cowboy gooit een boek naar de clown (do same verb/ meaning equivalent verb)
   Target picture
   c.	Teacher, hat, monk, throw

Consistent with the shared syntax model, equivalent within- (L1 to L1 and L2 to L2) and between-language priming has been found for datives in L1-Swedish L2-English speakers in a written sentence fragment completion priming task (Kantola & van-Gompel, Reference Kantola and van Gompel2011).

In experiments that primed relative clause attachment, Hartsuiker and colleagues (Reference Hartsuiker, Beerts, Loncke, Desmet and Bernolet2016) compared within-language priming (L1 to L1) to between-language priming (L1 to L2 and L2 to L1) and to priming between the two later acquired languages (L2 and L3). Note that priming relative clause attachment involves priming at the message level, because primes with different attachment sites have different meanings. So the meaning differences between relative clauses with different attachment sites are greater than the differences between the typical structures used in priming. The first experiment used a fragment-completion priming task and primed relative clause attachments from L1-Dutch, L2-English, and L3-French to L1-Dutch. Two additional experiments crosslinguistically primed relative clause attachment from L1-Dutch to L2-English, and L1-Dutch to L3-French. A fourth experiment used a confederate priming picture description task testing crosslinguistic priming of datives from L1-Dutch, L2-English, and L3-German to L2-English. Across all experiments, priming effects within-language and between-languages were found to be comparable. Finding priming between later acquired languages (an L2 to an L3) suggests that at high levels of proficiency, representations can also be shared across multiple languages. Consistent with this, Facipieri, Vann, and Bencini (Reference Facipieri, Vann and Bencini2022) found bidirectional crosslinguistic priming (L2-to-L3 and L3-to-L2) for passives with the same task and materials between English L2 and Spanish L3.

Differential priming effects for within- versus between-language priming were, however, reported by Cai and colleagues (Reference Cai, Pickering, Yan and Branigan2011) with early bilingual adult speakers of Cantonese and Mandarin. Cai and colleagues (Reference Cai, Pickering, Yan and Branigan2011) also had a cognate verb condition with verbs that are both translation equivalents (as in Schoonbaert et al., Reference Schoonbaert, Hartsuiker and Pickering2007) and, additionally, are etymologically related. Although the study by Schoonbaert and colleagues examined priming between languages (English and Dutch), for which many cognates exist, the items chosen for the experiments were overwhelmingly noncognates. Cognates may have a privileged status in a bilingual’s mind. For example, Costa, Caramazza and Sebastián-Gálles (Reference 72Costa, Caramazza and Sebastian-Galles2000) found that Catalan–Spanish bilinguals were faster in naming pictures that had cognate names in Catalan and Spanish than pictures with noncognate names. Cai and colleagues (Reference Cai, Pickering, Yan and Branigan2011) reported a lexical boost effect in within-language priming, and a much smaller between-language cognate boost effect. To account for the within-language versus between-language priming asymmetry, Cai and colleagues (Reference Cai, Pickering, Yan and Branigan2011) do not reject the overall architecture of shared syntax model, but they propose a more active role for the global language nodes, by assuming that they act exactly like the other nodes in the model. The idea is that the language nodes activate when a specific language is spoken, and this activation spreads to all lemmas of that language, resulting in a within-language boost even in the absence of lexical overlap. Moreover, they speculate that there may also be an inhibitory mechanism at play in accessing shared representations, such that when one language is turned on, the other language may be inhibited. As for the status of cognates, they argue that cognates are represented as separate lemmas, against positions that view cognates as sharing representations.

The shared syntax model, however, is not the only way to capture crosslinguistic priming effects. An alternative model is that speakers have separate but connected representations for similar structures, as proposed recently by Ahn and Ferreira (Reference Ahn and Ferreira2024). On this “Separate-but-Connected” account (Ahn & Ferreira, Reference Ahn and Ferreira2024), crosslinguistic structural priming occurs because the structure in one language activates a related, but separate, structure in the other language. The Separate-but-Connected account also predicts smaller between-language priming than within-language priming. The reason is that between-language priming results from the transmission of activation between two separate combinatorial nodes (one for each language), whereas within-language priming results from the residual activation of one combinatorial node.

Another question addressed via priming is whether priming in an L2 depends on whether the bilingual speaker’s L1 has that structure or not. At high proficiency levels, L1 experience with a structure in the L1 does not seem to influence within-language priming in an L2. In a within-language priming study for DOs in proficient L2-English speakers, L1-Spanish and L1-German showed equal amounts of priming (Flett, Branigan & Pickering, Reference Flett, Branigan and Pickering2013). But we will address the question of how priming, L1 and L2 characteristics, and prior linguistic experience with a structure interact with proficiency in the following section.

In Section 2.1.4, we addressed the question of whether priming can occur between constructions that are similar at the functional level but differ in constituent order. As was the case for monolingual priming, the evidence is mixed, with some crosslinguistic studies failing to find priming when constituent order differs (Loebell & Bock, Reference Loebell and Bock2003; Bernolet et al., Reference Bernolet, Hartsuiker and Pickering2007) and some finding priming irrespective of constituent order differences (Desmet & Declerq, Reference Desmet and Declercq2006; Chen, Jia, Wang, Dunlap & Sin, Reference Chen, Jia, Wang, Dunlap and Shin2013; Shin & Christianson, Reference Shin and Christianson2009). Desmet and Declerq (Reference Desmet and Declercq2006) found crosslinguistic priming of relative-clause attachment from Dutch to English, despite differences in constituent order. Chen and colleagues (Reference Chen, Jia, Wang, Dunlap and Shin2013) found bi-directional priming of passives between English and Chinese, again, despite differences in constituent order. In a recent computational model of crosslinguistic priming as implicit learning, Khoe and colleagues (Reference Khoe, Tsoukala, Kootstra and Frank2021) successfully modeled crosslinguistic priming between structures in English and Dutch with different word orders.

Many of the crosslinguistic effects reviewed above can be accounted for nicely by the shared syntax model. This model, however, deals with adult multilinguals who have achieved high levels of proficiency in their L2. Two questions arise with respect to priming and learning a second language: whether priming can in fact also function as an intervention to stimulate language learning, and whether priming effects change across different developmental stages in an L2. We will address each of these issues in the following sections.

4.2 Priming and Second Language Learning

The proposal that priming is an adaptive language-learning mechanism that continues into adulthood (see Section 2.2.2) has sparked an interest in examining whether priming supports L2 learning and extends beyond the priming task itself (for a review, see Jackson, Reference Jackson2018). Again, as noted in Section 2.2.2, use of the terms long-term or long-lasting priming should be qualified with the modifier “relatively.” In monolingual adults and children, long-term priming studies typically examine priming effects over varying intervening distractor trials. These studies have provided data in support of the view that priming is a form of implicit learning, whereby the production system learns to express messages in certain ways in response to input, so it makes sense to ask the same type of questions when adults learn an additional language. Additionally, some priming studies with adult L2 learners have tested whether priming can function as a learning intervention, by also including pre- and post-priming assessments.

The first priming study to test priming as a learning intervention in an L2 context was McDonough (Reference McDonough2006), who used the confederate priming paradigm with datives. This study was designed to include a baseline and a post-priming phase, using a picture description task to elicit datives. The priming phase targeted PO and DO datives. Participants were low-proficiency L2 speakers of English. In the priming phase, priming was successful for POs, but not for DOs. PO production carried over into the post-priming phase, with more PO descriptions produced than in the baseline phase. Another experiment primed DOs only, but again, priming was not significant, and participants produced more POs despite having being primed with DOs. There was, however, an indication that priming itself led to a greater production of DOs. L2 learners who had produced at least one DO at baseline were more likely to produce DOs during priming. McDonough’s interpretation of the lack of priming for DOs takes inspiration from L1 priming work that takes lack of evidence for priming as evidence for item-specific representations. McDonough, in fact, suggests that at the early stages of L2 development for DO structures in English, these structures may be tied to specific items, specifically pronouns, because DOs occur more frequently with pronouns in English. Based on these priming results, McDonough proposed an item-specific developmental trajectory for L2 syntax, along the lines of similar proposals in L1 development (e.g., Savage et al., Reference Savage, Lieven, Theakston and Tomasello2003).

In the first within-language L2 priming study that directly considered proficiency as a moderator variable, Kim and McDonough (Reference Kim and McDonough2008) examined active/passive priming in English L2 with Korean L1 speakers who were grouped into three proficiency groups based on performance on a cloze test. They included a verb overlap condition (same/different verb) and included verbs that varied in the frequency of occurrence in the passive in English, as determined by corpora. Kim and McDonough reported abstract priming and a lexical boost for all three proficiency levels, but the effect size of the lexical boost was larger for the lower-proficiency group.

Another test of priming as implicit learning in an L2 context was conducted by Shin and Christianson (Reference Shin and Christianson2012). This study is the only study, to my knowledge, that compares the effects of priming with additional explicit grammatical rule instruction embedded in the same experiment, and includes immediate and post-priming assessment phases. Pretest and posttest assessments consisted of a pictured-based sentence production task, with the target structures stimulated in the priming task (POs/DOs and phrasal verb constructions), and a grammaticality judgment task. Pre-tests took place immediately before the priming phase, and the immediate posttest occurred right after. For the long-term test, participants came back to the lab the next day. In the priming phase, L1-Korean L2-English speakers who were proficient in English (and had lived in an English-speaking country for at least three years) were primed with PO/DO datives and separated phrasal verb constructions (The man is putting the fire out/The man is putting out the fire). The priming phase included a lag manipulation (lag 0 vs. lag 4–5). At lag 0, there was also an instruction manipulation (explicit instruction, no instruction). In the explicit instruction condition, participants were provided with a grammatical explanation of the rules for DO/POs and particle verb placement in English. All manipulations were between subjects. Shin and Christianson found different effects for DO/POs versus phrasal verbs within the priming phase itself, immediately post-priming, and one day later. All priming conditions (lag 0, with explicit instruction, 0 lag without explicit instruction, and 4–5 lag – which was always without explicit instruction) yielded priming effects within the priming task itself. For DO/POs, explicit instruction at lag 0 resulted in greater priming than priming at lag 0 without explicit instruction, and greater than priming at lag 4–5. For separated phrasal verb constructions, lag 0 priming resulted in equal amounts of priming irrespective of explicit instruction, and these effects were larger than priming over lag 4–5. For DOs, the advantage of explicit instruction persisted in the immediate post-priming production task. For the long-term phase one day later, however, only participants who had received priming over lag 4–5 demonstrated sustained production of DOs, compared to those who had received 0 lag priming and 0 lag priming with explicit instruction. For phrasal verb constructions, effects were similar, independent of priming manipulation and instruction. Shin and Christianson imputed the differential effects of priming manipulation on longer-term maintenance and generalization of target structures to novel items to an interaction between implicit learning and structural complexity. They argued that for more complex structures such as DO datives, compared to phrasal verbs, implicit learning (induced via the lag manipulation) may be optimal for long-term learning.

Priming studies in an L2 setting demonstrate the potential of using priming translationally, as an intervention to support second language learning. One additional attractive feature of the L2 priming studies conducted in the context of second-language learning is that they employ pretest-posttest designs to determine whether priming effects generalize and persist outside of the priming task. Many L2 priming studies confirm that the effects of priming interventions generalize and persist (e.g., McDonough & Mackey, Reference McDonough and Mackey2008; McDonough & Chaikitmongkol, Reference McDonough and Chaikitmongkol2010; McDonough, Trofimovich, & Newman, Reference McDonough, Trofimovich and Neumann2015, Shin & Christianson, Reference Shin and Christianson2012), although the question remains as to how exactly these effects are modulated by proficiency, first language properties (e.g., structural similarity), and learner characteristics (for discussion and review of within-language priming in an L2, see Jackson, Reference Jackson2018).

4.3 Priming and Developmental Trajectories in an L2

Priming has increasingly been used to track how late bilinguals, and second language learners’ representations change developmentally. This line of inquiry parallels the interest in applying priming to address development of an L1. Whereas priming studies in L1 development track priming over age, the bulk of the studies that address L2 learning consider the relationship between known priming effects, such as the lexical boost and abstract priming and L2 language proficiency.

A study that directly sets out to use priming to examine developmental trajectories in an L2, including proficiency as a variable, was conducted by Bernolet, Hartsuiker, and Pickering (Reference Bernolet, Hartsuiker and Pickering2013), with L1-Dutch L2-English speakers from the Dutch-speaking region of Belgium (Flanders, adjective Flemish). They aimed to test a developmental model of L2 syntax acquisition by tracking lexically dependent and abstract priming effects over proficiency levels. Across two experiments, they primed genitive constructions from Dutch to English and within English (L2). The English Saxon genitive (‘s genitive) is a difficult structure for Flemish learners of English. Although Dutch has a construction that resembles the Saxon genitive in constituent order (possessor/possessed), this structure is not marked morphologically, and it occurs less frequently than the prepositional genitive. In both the between-language experiment (Experiment 1) and the within-language experiment (Experiment 2), possessed objects (e.g., a skateboard) were the same across prime and target for the repeated meaning conditions (4.6a, 4.6b) and different for the different meaning conditions (4.6c, 4.6d). In the same meaning condition, head nouns used to refer to the possessed objects were therefore translation equivalents in the between-language priming experiment and identical in the within-language priming experiment.

1. (4.6)

   Primes
   a.	The nun’s skateboard is blue/De non haar skatebard is blauw (’s genitive, same)
   b.	The skateboard of the nun is blue/Het skatebord van de non is blauw (of genitive, same)
   c.	The nun’s bottle is blue/ De non haar papfles is blauw (’s genitive, different)
   d.	The bottle of the nun is blue/Het papfles van de non is blauw (of genitive, different)
   Target picture
   e.	pirate, blue skateboard

In the between-language priming experiment there was a lexical boost effect (i.e., a translation equivalent boost) and increased abstract priming of ’s genitives (the dis-preferred structure) as proficiency increased. At the lowest proficiency levels, bilinguals showed no between-language priming. In the within-L2 priming experiment, there was priming for ’s genitives and a large lexical boost. Importantly, proficiency interacted with priming magnitude and the lexical boost, with different effects for high- versus low-proficiency speakers. Less proficient speakers showed large priming effects in the lexical overlap condition and smaller abstract priming effects. There was a trend towards stronger abstract priming in more proficient speakers. In a combined analysis with both experiments, within-language priming turned out to be stronger than between-language priming. Bernolet and colleagues (Reference Bernolet, Hartsuiker and Pickering2013) proposed a developmental trajectory according to which at lower proficiency levels L2 speakers start out with lexically specific representations and gradually abstract over items. At high levels of proficiency, multilinguals merge their representations for similar structures into one shared representation. On this account, the shared syntax model represents an “end state” for language learning.

A more fleshed out version of developmental model for L2 can be found in Hartsuiker and Bernolet’s (Reference Hartsuiker and Bernolet2017) study. We can illustrate the developmental stages using the acquisition of the English DO dative as an example. An L2 learner would start out with lemma-specific syntactic representations. For example, an English-L2 learner’s initial representation for the verb give would be item specific in that it would be connected to the DO argument structure on a verb-specific basis, with no generalization over English verbs that also occur in the DO. Because at the beginning stages of L2 learning there is no abstract representation of the DO, priming relies on L2 speaker’s explicit memory for the prime. Lexical overlap acts as a retrieval cue for the prime, thus explaining why at very low proficiency levels L2 speakers exhibit greater lexically dependent priming. With increased proficiency, syntactic representations abstract over verbs that are observed in the DO (e.g., the nodes for various English ditransitive verbs, such as give and show become connected to a single generalized DO node). At higher proficiency levels bilinguals will also abstract away from language-specific constructions and merge representations on the basis of both structural and semantic-pragmatic similarity.

Although there is some priming evidence that experience with a structure in the L1 does not affect priming magnitude in an L2 in highly proficient speakers (Flett et al., Reference Flett, Branigan and Pickering2013), other priming studies have found that the priming magnitude and longer-term effects of priming depend on interaction between proficiency and a combination of the properties of the L1 and the L2. For example, Kaan and Chun (Reference Kaan and Chun2018) claim to have found an effect of the frequency of a structure in the L1 when comparing priming of English DOs and POs in monolingual English and L2-English speakers who had Korean as their L1. Both Korean and English have POs and DOs, but whereas DO is more frequent in English, PO is more frequent in Korean. They observed cumulative priming during the experiment (as a function of the number of DOs/POs experienced prior to a given trial) and an inverse preference effect consistent with language biases: native English speakers showed more priming for POs, whereas L2-English speakers showed greater priming for DO. This suggests an influence of L1 experience. Future studies can exploit the priming paradigm to investigate how priming effects are modulated by proficiency, first language properties (e.g., structural similarity, frequency of the structures), and learner characteristics.

5.1 A Revised Model of Grammatical Encoding

Models of grammatical encoding have tacitly imported an assumption from a long tradition in linguistics that views the main verb as being central to encoding those aspects of meaning that matter to syntax, or how semantics is mapped onto syntax. For example, in Section 2.1.1, we noted that the Pickering and Branigan (Reference Pickering and Branigan1998) lemma model of grammatical encoding is verb-centered and lexicalist: it assumes that grammatical encoding necessarily requires activating a verb lemma, which then activates its combinatorial nodes.

There are reasons to question whether this entirely verb-centered view of sentence representation and sentence production is correct. Linguists working in very different theoretical frameworks (Borer, Reference Borer2005; Goldberg, Reference Goldberg1995; Levin & Rappaport Hovav, Reference Levin and Rappaport Hovav2005; Ramchand, Reference Ramchand2008; for a review and discussion, see Marantz, Reference Marantz2013) have observed that verbs appear in a larger set of argument structures (Pickering and Branigan’s combinatorial nodes) than typically assumed. Consider a prototypically transitive verb in English: kick. ‘Kick’ in English can occur in at least eight grammatical frames (5.1a–5.1e; examples from Bencini and Goldberg, Reference Bencini and Goldberg2000):

1. 1. a. Pat kicked the wall

   2. b. Pat kicked Bob the football

   3. c. Pat kicked Bob black and blue

   4. d. Pat kicked the football into the stadium

   5. e. Pat kicked at the wall

   6. f. Pat kicked her foot against the chair

   7. g. Horses kick

   8. h. Pat kicked her way out of the operating room

Each one of these occurrences of the verb kick in a sentence is an instance of a more general type of event. For example, 5.1b entails transfer: the sentence can be paraphrased as “Pat caused Bob to have the ball by kicking it.” On its own, however, the verb kick does not entail transfer (cf. 5.1a, 5.1c, 5.1d, etc.). So where is this component of meaning to be located? This question is at the heart of the debate between lexicalist and constructionist treatments of argument structure. The classic way of dealing with this is to assume different verb senses for each argument structure (e.g., Levin and Rappaport Hovav, Reference Levin and Rappaport Hovav2005; Rappaport Hovav and Levin, Reference Rappaport Hovav, Levin, Butt and Geuder1998; Pinker, Reference Pinker1989), so, on this view, there would, in fact, be multiple senses of the verb kick, including a sense that does not entail transfer and one that does. From the point of view of the speaker, this means that there are multiple long-term linguistic representations, each one corresponding to a different sense, and that the speaker selects the correct sense based on the message she/he intended to express. Goldberg (Reference Goldberg1995) pointed out that the multiple sense view also results in having to stipulate implausible verb senses for many verbs, to account for sentences like the ones in 5.2a–5.2c.

1. 1. a. The woman sneezed the foam off the cappuccino [sneeze sense: X cause Y move Z by sneezing]

   2. b. The train screeched into the station [screech sense: Y moves while screeching]

   3. c. She saw him into the room [see sense: X accompanies Y while moving]

An alternative view is to recognize the existence of verb-independent semantic generalizations directly linked to sentence structures; in other words, to recognize that sentence structures themselves contribute to the interpretation of a sentence. This level of generalization above the verb is captured in Argument Structure Constructions (Goldberg, Reference Goldberg1995, Reference Goldberg2006, for an introduction and application of Construction Grammar to the analysis of English, see Hilpert, Reference Hilpert2014; Hoffman, Reference Hoffmann2022). Constructions are, more broadly, pairings of “form and function” that exist at all levels of generalization (Goldberg, Reference Goldberg1995, Reference Goldberg2006). Semantically, argument structure constructions encode meanings that are central to human experience and cognition. Goldberg (Reference Goldberg1995) calls this the “scene-encoding hypothesis” (p. 39). Across languages, it is common to find sentence patterns that encode basic meanings such as experiencing a stimulus, undergoing a change of state, causing a change of state in another entity, transferring an object, moving along a path, and so on. The meaning side of an argument structure construction is expressed using the semantics of event structures (see Section 2.1.5.3). On a constructional approach, a semantic division of labor is posited between the more abstract relational meanings carried by constructions and the more specific meanings carried by verbs (Goldberg, Reference Goldberg1995; Reference Goldberg2006). For example, in 5.1b, the meaning [X cause Y to have Z] is associated directly with the DO construction, the semantic contribution of the main verb is, in turn, to modify the event by providing the specifics of the means by which the transfer is achieved (i.e., the act of kicking). In the case of the two intransitive verbs sneeze and screech in 5.2, sneeze modifies the X cause Y to move meaning expressed directly by the English caused motion construction, and sneeze expresses the means. Similarly, screech expresses the manner. The “function side” of a construction encompasses semantics and information structure. On a constructional account, although passives express the same semantic relations as actives, they differ in their information structure, as passives typically serve the discourse function of making the patient more prominent than the agent, and placing it in a topical position (for a constructional analysis of English passives, see Hilpert, Reference Hilpert2014).

With this additional representational machinery in the grammar, we are ready to propose a revision of the consensus model of grammatical encoding that captures all of the priming findings in monolingual adults, including the more recent studies that examined the nature of semantic representations, reviewed in Section 2.1.5. The constructional (as opposed to lexicalist) model recognizes the existence of verb-independent mappings from messages to grammatical encoding (Bencini, Reference Bencini, Hoffmann and Trousdale2013; Reference Bencini2017; Reference Bencini, Cardinaletti, Branchini, Giusti and Volpato2020). Figure 7 shows the generation of the sentence The girl gives her teacher a flower. In addition to verb-specific participant roles, associated with the core semantics of the verb give (“giver”, “givee”, “given”), the model recognizes the semantic contribution of the event structure X causes Y to have Z associated with the DO construction in English. In this case, the integration of verb and constructional semantics is seamless, as the number of verb-specific roles matches up with the number of argument slots (variables) in the construction. The constructional model also accounts for those cases of “implausible verb senses” in 5.2, when the number of verb participant roles is a subset of the number of argument roles (see Figure 8). To account for the production of The woman sneezed the foam off the cappuccino, the intransitive, monovalent predicate sneeze has only a participant role (we can call it “the sneezer”). It is the abstract English caused motion construction that contributes the two additional argument roles required and the overall more abstract cause-to-move interpretation, while the predicate contributes the means component of the event (Goldberg, Reference Goldberg1995). The combination of verbs and constructions is by no means an unregulated relationship. I refer the reader to the relevant linguistics literature (e.g., Goldberg, Reference Goldberg1995, Reference Goldberg2006).

Figure 7 Constructional sentence production model, same number of argument roles in argument structure construction and verb.

Figure 8 Constructional sentence production model, number of argument roles in argument structure construction is greater than in the verb.

Although constructional approaches vary (see Hoffman and Trousdale Reference Hoffmann and Trousdale2013; Ungerer and Hartmann, Reference Ungerer and Hartmann2023), one key finding is that constructions have the right level of abstraction to account for both for abstract (lexically independent) and for lexically boosted priming and the full range of priming effects reviewed in this Element. By recognizing sentence level form-function mappings, the constructional model means the finding that the more dimensions align (lexical, syntactic, semantic, pragmatic), the greater the priming can be accounted for. Because constructional approaches allow for the existence of intermediate-level generalizations, they make it also possible to account for the results of Ziegler and colleagues (Reference Ziegler, Bencini, Goldberg and Snedeker2019), which is a challenge for the notion of purely abstract syntactic priming. The constructional model is consistent with implicit learning models with a dual architecture and event semantic representations for the message (along the lines of Chang et al., Reference Chang, Dell and Bock2006) and with more structurally guided models of sentence production (Bock and Ferreira, Reference Bock, Ferreira, Goldrick, Ferreira and Miozzo2014; Konopka and Bock, Reference Konopka and Bock2009; Konopka and Meyer, Reference Konopka and Meyer2014).

To conclude, there are advantages to adopting a broadly constructional approach to linguistic representation over theoretical frameworks in linguistics that consider constructions to be epiphenomenal. The fact that constructions can be primed provides behavioral evidence that they are active in real-time language use. There is a welcome convergence between linguistic and psycholinguistic approaches that recognize priming as a suitable method to tap into linguistic representations (Bencini, Reference Bencini2002; Bencini, Reference Bencini, Hoffmann and Trousdale2013; Bencini, Reference Bencini, Cardinaletti, Branchini, Giusti and Volpato2020; Branigan, Pickering, Liversedge, Stewart, & Urbach, Reference Branigan, Pickering, Liversedge, Stewart and Urbach1995; Cai & Zhao, Reference Cai and Zhao2024; Branigan & Pickering, Reference Branigan and Pickering2017).