R1. How can linguistic representation be investigated?

We begin by clarifying the nature of our proposal. Although commentators agree about the importance of experimental methods for investigating linguistic representation, some of them seem to assume that we advocate entirely renouncing the use of acceptability judgments. But as we make clear throughout the article, our argument is that researchers interested in linguistic representation should not be restricted to using only acceptability judgments. Specifically, we contend that “the representations underlying language use need not and, in fact, should not be investigated only via [acceptability] judgments” (sect. 1, para. 2). We further argue that “[our] goal is to consider alternative (experimental) methods to acceptability judgments that potentially address the linguistic representations implicated in language processing” (sect. 1.1, para. 5), that “acceptability judgments are not enough,”(sect. 1.2) and that “researchers concerned with linguistic representations should not rely solely on such judgments, and should call on additional methodologies” (sect. 1.2, para. 10). Moreover, we argue that “acceptability judgments can be used (with appropriate controls) alongside structural priming (and perhaps other experimental methods; see sect. 1.3) as a means of developing representational hypotheses” (sect. 1.5, para. 2). (Emphasis added throughout.)

We therefore do not propose abandoning acceptability judgments as a linguistic method (Adger; de Ruiter & de Ruiter; Hall, Mayberry, & Ferreira [Hall et al.]), nor do we argue that structural priming is the only tool that can be used to study linguistic representation (González-Márquez, Feist, & Ströbel [González-Márquez et al.]; Hartsuiker & Bernolet; Ziegler, Snedeker, & Wittenberg [Ziegler et al.]). (We suspect that some commentators may have misinterpreted the phrase “end the current reliance on acceptability judgments” in the abstract as an exhortation to “end the current use of acceptability judgments.”)

We do argue, however, that priming has a privileged status. It is an implicit method that provides access to linguistic structure (i.e., strong generative capacity), and it is unaffected by an intervening theory “P” (Chomsky Reference Chomsky1981, p.283, footnote 39); see section 1.1, paragraph 7 of the target article. Acceptability judgments (whether obtained informally or formally) provide extensive and valuable, but suggestive, data (in the “context of discovery”; Hagoort). In this respect, they stand alongside the use of large-scale corpora to identify the types of utterances that occur and the contexts in which they occur (as Lester, Du Bois, Gries, & Martin [Lester et al.] note). Both of these methods provide evidence about the classes of acceptable sentences but need further assumptions to help determine linguistic structure – and here is where they encounter the problem of constituency tests. Other behavioral and neuroscientific methods (i.e., apart from priming) may provide evidence about linguistic structure without the need for constituency tests, but it is much harder to rule out explanations relating to the processing theory “P.”

Adger argues that linguists use experimental tests of a theory “as an injunction to think about what the implications of the result mean.” We agree that they should do this, but we do not believe that they typically do so (or else, papers in theoretical linguistics would make far more reference to experimental findings than they do). In sum, we wholeheartedly agree with Hagoort that there is no excuse for failing to use the available experimental and quantitative methods.

R2. How useful is priming?

A number of commentators express concerns about the versatility of structure priming in practice and the extent to which structural priming studies are able to provide support for specific representational proposals (de Ruiter & de Ruiter; Feldman & Milin; Francis; Hall et al.; Hartsuiker & Bernolet; MacDonald & Weiss; Mahowald, Futrell, & Gibson [Mahowald et al.]; Maldonado, Spector, & Chemla [Maldonado et al.]; Martin, Huettig, & Nieuwland [Martin et al.]; Meyer; Rees & Bott; Ziegler et al.).

As Mahowald et al. note, the meta-analysis carried out by Mahowald et al. (Reference Mahowald, James, Futrell and Gibson2016b) shows that structural priming is robust and highly replicable. Therefore, criticisms of other priming paradigms (e.g., in social psychology) are not relevant (MacDonald & Weiss). The meta-analysis also highlights limitations in studies that compare the magnitude of priming under different circumstances. We stress that a large number of converging findings may be necessary to produce confidence in theoretical claims.

With respect to our specific account, our claims concerning the monostratal nature of syntactic representation are based on many individual studies that, together, argue against some level of underlying structure. Our intention in proposing the account in section 2 is to encourage researchers to address both specific and general representational questions, in order to broaden the evidence base for theory development. Some of our specific claims, of course, may prove incorrect or incomplete in light of further research.

Hence, we acknowledge a need for reliable, replicable results, and for caution about underpowered studies, the interpretation of null results, and so on (e.g., de Ruiter & de Ruiter; Mahowald et al.; Maldonado et al.; Meyer; Rees & Bott; Ziegler et al.). These concerns apply to any scientific method. Clearly, priming studies require careful design, and some studies may be difficult to carry out in practice. Structural priming, however, is not inherently more limited than other experimental (or non-experimental) approaches.

Other commentators are concerned that the value of priming is limited to cases involving pairs of roughly equivalent structures (e.g., de Ruiter & de Ruiter; Meyer). As we note, most current evidence comes from choices in language production, when a speaker can use one structure (e.g., passive) or another (e.g., active) to describe the same situation. In such cases, it is clearly necessary for alternatives to exist (and to be sufficiently frequent that, in practice, participants can be induced to produce both alternatives). But there are other methods of inducing priming, such as priming of alternatives that differ in meaning (Scheepers Reference Scheepers2003), priming choices in comprehension (e.g., Branigan et al. Reference Branigan, Pickering and McLean2005), priming of eye movements during comprehension (e.g., Arai et al. Reference Arai, van Gompel and Scheepers2007), priming response times in production (Smith & Wheeldon Reference Smith and Wheeldon2001; Segaert et al. Reference Segaert, Wheeldon and Hagoort2016), priming response times in comprehension (Knoeferle Reference Knoeferle2014), or repetition suppression of the Blood Oxygen Level Dependent (BOLD) signal in functional Magnetic Resonance Imaging (fMRI) adaptation (Segaert et al. Reference Segaert, Menenti, Weber, Petersson and Hagoort2012). These methods do not always require semantically similar alternatives, and they should allow us to use priming more extensively in the future.

A particular set of priming effects might appear compatible with more than one linguistic account (e.g., Francis; Hartsuiker & Bernolet). In some cases, there are ways to rule out one alternative (see sect. R10 for some specific examples). In other cases, the data may fail to distinguish between accounts – and this is why priming should be established as a general method in which conclusions are drawn from large numbers of studies considered together.

Some commentators raise concerns about the fact that priming effects reduce with increasing exposure (Lester et al.; Ryskin & Brown-Schmidt). Experimenters can take this into account (especially when small effect sizes are anticipated) through careful decisions about the size and composition of the stimulus set, testing adequate number of participants and items, or appropriate statistical analyses. Ziegler et al. observe that magnitudes of priming may vary across tasks. This simply means that, just as for other experimental tasks (e.g., sentence recall versus picture description tasks), researchers should be careful about comparing results across tasks.

Finally, we do not share Meyer's pessimism concerning the relevance of priming for studying children and patients. For example, priming studies have successfully addressed longstanding questions about the extent to which typically developing children and children with a specific language impairment have abstract syntactic representations (Bencini & Valian Reference Bencini and Valian2008; Foltz et al. Reference Foltz, Thiele, Kahsnitz and Stenneken2015; Garraffa et al. Reference Garraffa, Coco and Branigan2015; Messenger et al. Reference Messenger, Branigan and McLean2011; Rowland et al. Reference Rowland, Chang, Ambridge, Pine and Lieven2012).

R3. The directness of the relationship between grammar and processing

MacDonald & Weiss, Martin et al., and Meyer raise questions about the directness of the relationship between grammar and processing. Comprehension-to-production priming cannot be simply procedural because comprehension and production are different processes. Therefore, we suggest it provides access to representation (as argued in Branigan et al. Reference Branigan, Pickering and Cleland2000). Of course, it could reflect a shared sub-process – for example, accessing a linguistic constraint. But if this sub-process is the same across production and comprehension, it is hard to see what it could reflect except for the representation itself. In fact, any form of structural priming (including production-to-production) is informative about linguistic representation, and researchers do not need to conduct comprehension-to-production priming studies to constrain linguistic theory. Note that a linguistic theory could treat the frequency of a syntactic constraint (e.g., V → V NP PP) as part of its linguistic representation (see MacDonald & Weiss), but our concern is with the representation of the constraints themselves rather than their frequency.

We do not claim that the apparent automaticity of priming is necessary for making claims about representation rather than processes. Instead, it suggests that priming is not affected by strategies or metalinguistic processes – concerns that may affect acceptability judgments. In other words, apparent automaticity gives us more confidence about linguistic representations.

R4. Does priming access all representations?

Some commentators suggest that priming may be insensitive to some linguistic representations (Koring & Reuland; Maldonado et al.; Rees & Bott). In other words, there may be levels of linguistic representation that are unaffected by priming. Some representations may be inaccessible to any method, whether because they have been compiled into different representations online (Berwick & Weinberg Reference Berwick and Weinberg1984) or because they have undergone attrition (see Kootstra & Rossi). The commentators, however, are more concerned with the possibility that representations can be accessed but cannot be affected by prior exposure – a possibility that would reduce somewhat the effectiveness of priming. We propose that the use of a representation (whether in production or comprehension) will affect subsequent use of that representation; this is the logic that all accounts of priming depend on. But this is an assumption; if it is too strong, then priming would be somewhat less sensitive than we assume. Such a limitation, of course, would not affect the conclusions that we can draw from (positive) demonstrations of priming.

Of course, failure to find priming does not demonstrate that it does not occur; a study may have been insensitive or underpowered. For example, priming without lexical repetition in comprehension sometimes has occurred (e.g., Pickering et al. Reference Pickering, McLean and Branigan2013; Thothathiri & Snedeker (Reference Thothathiri and Snedeker2008a)) and sometimes has not (e.g., Branigan et al. Reference Branigan, Pickering and McLean2005; Traxler et al. Reference Traxler, Tooley and Pickering2014). Similarly, evidence for intonational priming is limited (de Ruiter & de Ruiter), but to demonstrate that intonational structure is represented (in roughly the form assumed by many linguists) and that it cannot be primed would involve extensive experimentation. It is too early to conclude that there is a level of linguistic representation that cannot be primed.

R5. What can acceptability judgments tell us?

Adger; Ambridge; Gaston, Huang, & Phillips [Gaston et al.]; Hall et al.; and Sprouse & Almeida contend that we unfairly criticise a method that has been used fruitfully (and frugally) for decades – although other commentators (e.g., González-Márquez et al.; Lester et al.) argue that we, in fact, overemphasize its significance. We acknowledge that acceptability judgments have been a fertile method for generating hypotheses historically, and we do not advocate abandoning them. Indeed, their use is appropriate for field linguistics (Hall et al.) when identifying the possible sentences of a language and developing representational hypotheses (i.e., in the “context of discovery”), but even in these situations, researchers should carry out priming experiments when possible.

Our main concern with acceptability judgments does not lie with their informality (Sprouse & Almeida), though we do advocate avoiding informality where possible. We agree that careful experimentation can determine which judgments can be explained by processing limitations (e.g., Sprouse et al. Reference Sprouse, Wagers and Phillips2012). Presumably, those judgments that cannot be explained by processing limitations are the judgments that are of interest to linguists, in the sense that they determine the set of sentences the grammar must be able to account for (i.e., its weak generative capacity). Sprouse and Almeida say that establishing these data is not in doubt, and the concern is with what they call interpreting them. In other words, they argue that determining the set of sentences generated by the grammar is straightforward (i.e., the weak generative capacity) and that the difficult issue is identifying the structural representations that underlie those sentences (i.e., the strong generative capacity).

There may not be major concerns in practice about identifying the set of sentences that should be generated by the linguistic theory (though there remain concerns about consistency across types of speaker and task; Dąbrowska Reference Dąbrowska2010; Sprouse Reference Sprouse2009). Acceptability judgments, however, do not allow us to determine the underlying structural representations, because they cannot distinguish among competing analyses of the same data. In contrast, priming allows us to distinguish among such analyses – in Sprouse and Almeida's terms, to interpret the data. People may use representations that intrinsically include lexical content, or they may not. They may use representations that include elements that are not phonologically realised, or they may not. The different representations proposed for DO sentences (e.g., a ternary-branching structure involving no “moved” elements versus a binary-branching structure involving “moved” elements; Pollard & Sag Reference Pollard and Sag1994 versus Larson Reference Larson1988) are different interpretations of the same data. But they cannot both correspond to a speaker's mental representation for these sentences.

Some commentators defend constituency tests. For example, Gaston et al. propose that the apparent contradictions that they yield can be resolved by further enquiry. This is not correct; their contradictions are basic and widespread. For example, almost anything can be coordinated, so it would appear that almost anything can be a constituent. But this claim is possible only if we assume massively flexible constituency (Pickering & Barry Reference Pickering and Barry1993; Steedman Reference Steedman2000) – a proposal contested by the great majority of linguistic theories. And in practice, linguistic theories disagree about the most fundamental points regarding structure (not merely the structure of ellipsis or Wh-questions). Our point is that traditional linguistic approaches assume that strong generative capacity can be determined by combining the sets of sentences compatible with acceptability judgments (i.e., weak generative capacity) with constituency tests. But if the tests are as flawed as they appear to be, this is impossible, even if Sprouse and Almeida are correct about the consistency of acceptability judgments.

In response to the query of Gaston et al. about the nature of structural priming, the example of lexical priming that we use to introduce priming as a phenomenon (sect. 1.4) involves words that are related but not identical (i.e., associative/semantic priming). It may be more appropriate to compare structural priming with repetition priming of words (e.g., Scarborough et al. Reference Scarborough, Cortese and Scarborough1977), which is long lasting, like structural priming (Bock & Griffin Reference Bock and Griffin2000). Structural priming appears to be relatively unaffected by contextual changes – for example, between writing and speech (Cleland & Pickering Reference Cleland and Pickering2006) – but a more detailed exploration of such issues is necessary to fully understanding priming itself. Understanding a method is, of course, very important for drawing conclusions using that method, but priming (as a cognitive phenomenon) is far better understood than the complex combination of language processing, metalinguistic reflection, and decision-making that underlies acceptability judgments. Similarly, there is no well-grounded theoretical motivation for constituency tests (see sect. 1.2).

R6. The stability of linguistic representations

Several commentators are concerned with the stability of the linguistic representations that are investigated in structural priming experiments (Cai & Liu; Francis; Günther; Hall et al.; Kootstra & Rossi; Martin et al.; MacDonald & Weiss; Meyer; Ryskin & Brown-Schmidt; Ziegler et al.). Some commentators suggest that their apparent instability and the fact that priming itself can bring about representational changes call into question the use of structural priming to investigate representation.

We agree that exposure to particular linguistic structures can affect speakers' linguistic representations, and any linguistic theory must be able to account for such effects. In practice, some researchers have argued that priming may offer an intervention to enhance foreign-language learning (McDonough & Mackey Reference McDonough and Mackey2006) or recovery from aphasia (Cho-Reyes & Thompson Reference Cho-Reyes and Thompson2012). Moreover, Ivanova et al. (Reference Ivanova, Pickering, McLean, Costa and Branigan2012b) found that participants occasionally produced a DO structure with a verb such as donate after encountering another DO sentence with donate, even though such sentences are usually regarded as unacceptable (Ambridge). In other words, priming can change the speaker's representation (at least temporarily). Other priming effects involve repeated exposure (e.g., developing the needs + ed construction; Kaschak & Glenberg Reference Kaschak and Glenberg2004). Importantly, similar effects occur in studies in which participants repeatedly make acceptability judgments, thereby showing that such judgments are affected also by exposure (Luka & Barsalou Reference Luka and Barsalou2005; Snyder Reference Snyder2000; cf. Sprouse Reference Sprouse2009).

So, the method of investigation (priming) sometimes changes the phenomenon that it is investigating. But this observation is not problematic for our proposals. Our concern is with the nature of the representations that underlie language use at any given time, and such representations need not be static. Priming allows us to take a snapshot of linguistic representations. Moreover, we can track how those representations change across time and what causes them to change (e.g., Cai & Liu; Kootstra & Rossi; Rowland & Monaghan). For example, structural priming supports a developmental trajectory during L2 learning from item-specific representations in less proficient L2 speakers to abstract representations in more proficient L2 speakers (Bernolet et al. Reference Bernolet, Hartsuiker and Pickering2013). Other studies have shown how children's syntactic representations at different ages do or do not differ from those of adults in specific ways – for example, with respect to their independence from lexical or semantic content (Messenger et al. (Reference Messenger, Branigan, McLean and Sorace2012b)). Of course, we do not claim that they are the same in all respects (Rowland & Monaghan).

R7. Priming, individuals, and groups

Priming experiments are typically conducted over fairly large groups of participants, and some commentators (Hall et al.; MacDonald & Weiss) therefore assume that anything priming tells us must be at the group level – that is, holding over groups of participants with the same linguistic representations as each other. We do not make this assumption. Researchers typically assume a common core of structural representations: that people who speak a particular language have similar linguistic representations. Of course, the acceptability judgment method also tends to make this assumption. But speakers of a language may have some different representations. Again, structural priming allows us to investigate variations across individual speakers (idiolectal variations) or populations (e.g., L1 versus L2 speakers of a language).

As with any other languages, priming should be used to understand the representation of sign languages (Hall et al.), and Hall et al. (Reference Hall, Ferreira and Mayberry2015) paves the way for such research. L1 and L2 ASL signers may have differences in syntactic processing, but the priming data suggest that – as with L1 and (proficient) L2 speakers of spoken languages (Schoonbaert et al. Reference Schoonbaert, Hartsuiker and Pickering2007) – the syntactic representations they use during processing may have the same (relevant) characteristics. In these studies, syntactic priming does not tell us how these representations were constructed during processing, their interfaces with other representations, how those representations are instantiated in the brain, or how they developed (although see Slevc & Ivanova for relevant discussion).

Contrary to MacDonald & Weiss's contention, it is possible to study priming in individual participants. Such studies might require extensive data from an individual, just as in any single-case study. Moreover, if a structure can be represented variably within an individual, we can use priming to determine which representation is used more often (Günther).

Finally, Francis's observation that structural representations are not necessarily diachronically stable highlights how priming can address diachronic changes in structural representation (Pickering & Garrod Reference Pickering, Garrod, Hundt, Mollin and Pfenninger2017). Her example shows how priming could distinguish between competing theoretical proposals that acceptability judgments have not been able to discriminate, in the same way as priming has been used to examine historically motivated accounts of lexical representation (Marslen-Wilson et al. Reference Marslen-Wilson, Tyler, Waksler and Older1994). In principle, corpus-based studies of structural priming (Gries Reference Gries2005; Szmrecsanyi Reference Szmrecsanyi2006) also could be used to track diachronic changes in structural representations.

R8. Structural priming and language processing (in monologue and dialogue)

A number of commentators discuss what priming may or may not tell us about language processing. Our focus in this paper is strictly on representation, so we consider processing only to the extent that it is informative about representation (thus, we do not address questions such as the timecourse of priming [Martin et al.], or how priming may change people's parsing preferences [Ryskin & Brown-Schmidt]).

For example, Slevc & Ivanova point out that priming can reflect not only people's final parses, but also their intermediate parses (e.g., Van Gompel et al. Reference Van Gompel, Pickering, Pearson and Jacob2006). Therefore, we need to be careful about whether any priming effect depends on a temporary but abandoned representation rather than on the final representation. Investigations of abandoned analyses are, of course, also informative about the process of language comprehension (which analyses people consider, whether they hold on to abandoned analyses, and so on). Such issues, however, are not directly relevant to linguistic representation.

Slevc & Ivanova also discuss how priming might interact with other aspects of cognition, such as depth of processing and working memory resources (see also Hartsuiker & Bernolet; Martin et al.). These issues may be important to an understanding of processing – for example, with respect to the relationship between individual differences in priming and other aspects of cognition. Differences in method (and associated processing demands) also can affect priming magnitudes (e.g., Mahowald et al. Reference Mahowald, James, Futrell and Gibson2016b; Ziegler et al.), so care must be taken when comparing studies. But if such factors are controlled within an experiment, they do not undermine the informativeness of priming for understanding linguistic representation, as Slevc & Ivanova acknowledge.

We agree with Kootstra & Rossi that naturalistic studies of priming are beneficial, but for current purposes (i.e., linguistic representation), it does not matter whether any effects occur under laboratory conditions or not. We are fascinated by the priming of code-switching, but this programme of research is relevant primarily to processing (i.e., language selection) – assuming, of course that the code-switcher selects between two independent representations of different languages. In sum, priming is versatile, and its implications go well beyond linguistic representation (see Pickering & Ferreira Reference Pickering and Ferreira2008).

MacDonald & Weiss argue that priming may reflect a general (nonlinguistic) tendency to repeat prior actions, as demonstrated by parallel priming effects in motor actions and syntax (Allen et al. Reference Allen, Ibara, Seymour, Cordova and Botvinick2010) and between domains such as music and language (van der Cavey & Hartsuiker Reference Van der Cavey and Hartsuiker2016) (see also Kempson & Gregoromichelaki). But if priming occurs between language and non-language, it does not prevent researchers using priming to discriminate linguistic theories; it is just that some representations are shared with other domains. To take a more formal example, Steedman (Reference Steedman2002) presented an account of action based on principles (functional composition and type raising) that are used also in Combinatorial Categorial Grammar (Steedman Reference Steedman2000). Extending these principles to action in general does not invalidate his account of linguistic representation; it just means that the same principles are used in quite different domains.

R9. Alternative approaches to linguistic representation

Alongside structural priming, some commentators suggest further methods for studying linguistic representation. Ziegler et al. argue that priming is not special. If any method appears to relate to linguistic representation (e.g., multivoxel pattern analysis relating to thematic roles), then they assume that it can be used to constrain linguistic theory. This is reasonable in principle but, in practice, is likely to always face the problem that any effect could be due to the “parsing procedure P.” For example, it might appear that the N400 in Event-related Potential (ERP) research could be diagnostic of semantic anomaly and, hence, semantic representation, whereas the P600 could be diagnostic of syntactic anomaly and, hence, syntactic representation. Whether an anomalous utterance in an ERP experiment produced a N400 or a P600 would then be informative about linguistic representation (e.g., Osterhout & Mobley Reference Osterhout and Mobley1995 on anaphora). It turned out, however, that some semantically anomalous utterances produced a P600 (e.g., Kim & Osterhout Reference Kim and Osterhout2005). This finding means that these ERP components are not directly informative about linguistic representation. Rather than focus on one method, researchers might be able to combine findings from diverse methods and draw general conclusions about linguistic representation. But there is little sign of this happening, presumably because theoretical linguistics would ascribe any such findings to “P.” Priming does not face this concern, in part because there is no need for an indirect intervening theory, but also because it has an established methodology that is largely consistently applied across studies (a property that it shares with acceptability judgments, of course).

We agree with Rowland & Monaghan that multiple methods, including corpus research and novel word studies, are desirable in developmental research. Such methods, however, have important limitations for studying representation. It is unclear whether novel word studies, for example, are informative about underlying representations or about processing. We cannot tell whether children's reluctance to produce DO structures with novel verbs that they have previously encountered in PO structures is because they do not have an abstract (generalized) DO representation, or because they find it difficult to produce unusual word orders. In contrast, priming between DO sentences involving different verbs supports an abstract DO representation underlying these sentences (Rowland et al. Reference Rowland, Chang, Ambridge, Pine and Lieven2012). As we note in R6, priming studies can address how representations change (e.g., whether a structure is lexically specific at time 1 but not at time 2).

With respect to nonexperimental methods, Gonzalez-Marquez et al., Lester et al., and Rowland & Monaghan emphasize the potential offered by corpus and distributional analyses for both developmental and adult research (see also Rees & Bott). These approaches offer valuable data about weak generative capacity and factors that may influence structural choices in language use (e.g., Bresnan et al. Reference Bresnan, Cueni, Nikita, Baayen, Bouma, Krämer and Zwarts2007) (see also sect. R10). Recent advances in statistical analysis help researchers to exclude confounding variables. However, straightforward distributional analyses are less informative than the substantial corpus-based evidence of structural priming (see Gries & Kootstra Reference Gries and Kootstra2017) – an approach that also may address some of the concerns about statistical power highlighted in R2. Moreover, experimental and corpus-based evidence for priming complement each other. (Note that Kempson & Gregoromichelaki and De Ruiter & De Ruiter query whether syntactic priming occurs in conversation on the basis of the negative results of Healey et al. [Reference Healey, Purver and Howes2014]). But their conclusions contrast with many other demonstrations of priming in conversation (e.g., Gries Reference Gries2005; Gries Reference Gries and Schönefeld2011; Myslín & Levy Reference Myslín and Levy2016; Reitter & Moore Reference Reitter and Moore2014; Szmrecsanyi Reference Szmrecsanyi2006).

Hagoort, Martin et al., and Ziegler et al. propose that neuroimaging techniques can address many representational questions. As Hagoort notes, fMRI adaptation (based on variations in the BOLD response dependent on previous experience) is a form of priming and supports the segregation of semantic, syntactic, and lexical processing in ways that are shared across comprehension and production (Menenti et al. Reference Menenti, Gierhan, Segaert and Hagoort2011; Segaert et al. Reference Segaert, Menenti, Weber, Petersson and Hagoort2012). But neuroimaging research not based on priming has so far been less informative about representation.

Cai & Liu suggest that, although priming offers deeper insights than intuition-based data into the microscopic aspect of linguistic representation, it is less well suited to evaluating the macroscopic characteristics of language systems and, specifically, the interconnectivity of elements at each level. Complex network approaches may offer an interesting way to study system-level representational similarities (e.g., between native and non-native speakers) and changes in these systems (e.g., during L2 learning), in a way that can complement the insights from structural priming into the development of linguistic representations in particular populations (such as L2 learners), or in principle in individual speakers.

R10. Details of our theory

Our linguistic account assumes an architecture that distinguishes distinct semantic, syntactic and phonological representations. We agree with Hagoort that it is largely compatible with the Memory Unification and Control model (Hagoort Reference Hagoort2014; Hagoort & Indefrey Reference Hagoort and Indefrey2014), a neurobiological model of language based on data from imaging studies that identifies discrete neuronal network contributions for unifying semantic, syntactic, and phonological information. However, various commentators dispute specific aspects of our account.

Ziegler et al. are right that there is less structural priming evidence relating to semantic structure than to syntactic structure. This fact reflects the greater interest historically in the latter than the former (e.g., Bock Reference Bock1986). But it is not contentious that structural priming of semantic structure occurs and that it is informative for theories of semantic representation (Maldonado et al.; see also Bunger, Papafragou, and Trueswell Reference Bunger, Papafragou and Trueswell2013, for priming of event structure, and Rees & Bott for extension to pragmatic representation). Furthermore, studies converge in supporting a lexically independent level of semantic representation that includes quantificational information. We anticipate that the growing interest in the use of structural priming to investigate semantic representation will give rise to more detailed theoretical proposals based on converging evidence from multiple studies, as has occurred for syntactic representation.

Ambridge, González-Márquez et al., Günther, Lester et al., and Ziegler et al. express some scepticism about a level of syntactic representation that is independent of semantic and lexical information. The finding that priming of syntactic structure is not enhanced when thematic mappings or semantic features are repeated (Bock & Loebell Reference Bock and Loebell1990; Huang et al. Reference Huang, Pickering, Yang, Wang and Branigan2016; Messenger et al. (Reference Messenger, Branigan, McLean and Sorace2012b)) suggests that semantic information is not incorporated into syntactic structure. It specifically fits badly with the construction-based approach advocated by Günther, who suggests that priming effects might hold between entire constructions (which could be related but of different types), or just their syntactic (i.e., form) or their semantic (i.e., meaning) components. The straightforward interpretation of a construction grammar is that it involves form-meaning pairings and that these pairings can be primed. This is compatible with integrated syntax and semantics. Günther suggests that priming can occur between the form or the meaning side of a form-meaning pairing. But if so, this means there is autonomy of syntax. Such autonomy can explain priming between sentences with the same syntax but very different meanings (e.g., from locatives to passives; Bock & Loebell Reference Bock and Loebell1990).

In our account, thematic relations and other semantic information are represented separately from syntactic structure and are mapped to syntactic representations (e.g., Chang et al. Reference Chang, Bock and Goldberg2003; Köhne et al. Reference Köhne, Pickering and Branigan2014; Cai et al. Reference Cai, Pickering and Branigan2012). Not surprisingly, some mappings are preferred over others (e.g., Bresnan et al. Reference Bresnan, Cueni, Nikita, Baayen, Bouma, Krämer and Zwarts2007; Ferreira Reference Ferreira1994; Kempen & Harbusch Reference Kempen, Harbusch, Pechmann and Habel2004). Such preferences, however, do not imply that semantic (or indeed, lexical) content is part of the syntactic representation. Moreover, acceptability judgments for sentences reflect their likelihood of occurrence (Bresnan Reference Bresnan, Featherston and Sternefeld2007), so it is unsurprising that sentences involving preferred mappings are rated as more acceptable than sentences involving less preferred mappings (Ambridge). In fact, Ambridge's results suggest that acceptability judgments are affected by semantic factors, a point that reinforces our conclusion that acceptability judgments do not straightforwardly reflect syntactic representation.

With respect to the relationship between lexical content and syntactic structure, Ziegler et al. argue that function words can be a locus of priming, based on Ferreira's (Reference Ferreira2003) study. In fact, Ferreira argued for an autonomous, lexically independent syntactic component, consistent with our account: Participants did not produce structures that included complementizer that following primes that included complementizer that unless the primes also involved the same syntactic structure. In other words, repetition of the function word alone did not determine choice of structure (see also Bock Reference Bock1989).

Similarly, our claim that syntactic representations are independent of closed class features is not undermined by Feldman & Milin's finding of differences in priming of single word production (measured by response latencies) depending on the modality of the prime (drawing vs. word), and differences in the specific verb form that was produced. Any such effects do not appear to relate to structural priming (as far as is known; Pickering & Branigan Reference Pickering and Branigan1998).

Regarding open-class lexical content, priming evidence provides no support for redundant representations in which the same utterance (or part of an utterance) is represented both as a single lexicalized chunk and as the integration of several more schematic constructions (Günther). If this were the case, we might expect stronger priming from frequently encountered phrases (which should be more likely to be redundantly represented), as there would be two potential sources of priming (Rowland et al. Reference Rowland, Chang, Ambridge, Pine and Lieven2012). Konopka and Bock (Reference Konopka and Bock2009), however, found that idiomatic expressions and nonidiomatic expressions yielded equivalent priming.

Hartsuiker & Bernolet suggest that the lexical boost to priming (which appears to be short lived) need not reflect a link between syntactic representations and the lexicon but may instead arise from an explicit memory of the prime that rapidly decays, with the verb acting as a retrieval cue. Malhotra et al. (Reference Malhotra, Pickering, Branigan, Bednar, Love, McRae and Sloutsky2008) demonstrated (using a dynamical systems model) that our assumption of a link between syntactic representations and the lexicon is compatible with a short-lived lexical boost. The longevity of the lexical boost, therefore, does not discriminate between our account and alternative accounts. Furthermore, it does not matter whether explicit memory processes are involved in priming: Any such memories must still make reference to abstract structure (e.g., that the verb give appeared with an NP and an NP) to be able to explain priming between nonidentical sentences (Branigan & McLean Reference Branigan and McLean2016). Note also that these memories cannot be based on specific wordforms (because the lexical boost is unaffected by repetition of specific features such as tense, aspect and number; Pickering & Branigan Reference Pickering and Branigan1998).

Other commentators discuss alternative ways of conceptualizing an abstract level of syntactic representation. Hudson supports our overall approach but suggests that priming evidence suggests that we should abandon the standard assumption that constituent-structure is basic, and with dependencies (e.g., subject-verb agreement) being defined with respect to constituent-structure. Instead, he treats dependency as basic and constituency as derivative at best. It seems most straightforward to us to define priming (e.g., of Prepositional Object vs. Double Object) in constituent-structure terms (e.g., priming VP → V NP PP rather than V → V NP NP), but it presumably would be possible to define it also in terms of dependencies, for instance from V to N (i.e., the head noun of the NP) and V to P (i.e., the head preposition of the PP) rather than dependencies from V to N and V to N. Of course, priming of subject-verb agreement itself (Haskell et al. Reference Haskell, Thornton and MacDonald2010) fits particularly well with a dependency-based analysis. An analysis of priming in terms of a dependency-based grammar such as Word Grammar (Hudson Reference Hudson2010) would be very interesting.

Kempson & Gregoromichelaki propose that our data do not need to be interpreted in terms of hierarchical representations such as those derived from context-free constraints (or similar), and interpret it in terms of sequential patterns of processing actions (see also O'Grady). If those patterns discriminate between linguistic structures that prime differently (e.g., the sequential pattern for POs differs from the pattern for DOs), then their approach may be able to capture priming data, and an analysis in terms of dynamic syntax also would be very interesting (Cann et al. Reference Cann, Kempson and Marten2005).

Adger and Koring & Reuland (see also Gaston et al.) challenge our use of priming between unergatives and unaccusatives to propose that they have similar syntactic representations. In contrast, their favored accounts assume that unaccusatives have a very different structure from unergatives – a difference that would almost certainly preclude priming, given the sensitivity of priming to comparable syntactic differences (e.g., priming does not occur between constructions with different numbers of arguments). Koring & Reuland simply assume an analysis involving NP-movement, but we query the assumptions that underlie analyses such as theirs. We do not dispute syntactic distinctions between types of subjects (though at least some relevant distinctions between unergatives versus unaccusatives may be semantic rather than syntactic; e.g., Zaenen Reference Zaenen1988), but argue that these distinctions do not require different constituent-structure representations. Similarly, Adger assumes that a distinction between Lily froze solid and *Lily danced tired means that Lily froze and Lily danced must have different constituent structures. Instead, it shows that danced and froze have syntactic properties that allow them both to combine with just a subject but which have different effects on combination with a subject and another constituent. We cannot conclude that Lily froze and Lily danced have different constituent structures. Approaches such as those of Koring & Reuland and Adger would lead to extremely complex constituent structures and are challenged by theories assuming flat structures (e.g., Culicover & Jackendoff Reference Culicover and Jackendoff2005; Pollard & Sag Reference Pollard and Sag1994).

Our concern with many other experimental methods is that their results can be due to processing (i.e., processing theory “P” in Chomsky Reference Chomsky1981). Therefore, we did not review all results (indeed, we did not have the space to do so). As an example, Bever and Sanz's (Reference Bever and Sanz1997) finding of faster responses to a probe word associated with a subject noun that preceded an unaccusative compared to an unergative verb could be due to semantic differences between unaccusatives and unergatives (rather than reactivating an antecedent at an NP-trace) (see also Agnew et al. Reference Agnew, van de Koot, McGettigan and Scott2014). Both Koring & Reuland and Adger point to many experimental studies that contrast unaccusative and unergative sentences, but none of them demonstrate that the distinctions are due to constituent structure differences. Many experimental methods such as studies using reactivation or visual world paradigms provide results relating to interpretation; there may be an underlying syntactic cause of the effects or there may not, but we typically cannot tell. In contrast, priming is informative about syntactic representation.

We also note that our description of “mainstream generative grammar” (Koring & Reuland) is our best attempt to capture the key issues of a family of views and to ignore factional distinctions. It is not helpful to argue that we are using concepts that are obsolete within the authors' favored branch of linguistic theory. We are simply trying to describe a complex state of affairs and identify the core issues that we wish to dispute.

Kwon & Lee argue that priming evidence provides only limited benefit compared with acceptability judgments. But criticising a new method on the grounds that it has not discovered as much as an old method would simply lead to stagnation. Their specific point about distinguishing linear and hierarchical relations is more insightful. We believe that priming data support a single level of structure that is specified for linear and hierarchical relations together, rather than one in which hierarchical structure is independently represented. For example, Bernolet et al. (Reference Bernolet, Hartsuiker and Pickering2007) found priming between German and Dutch noun phrases that both have adjective-verb order (the X that red is) but not between English and Dutch, presumably because English has verb-adjective order. In other words, the priming is not localized at an unordered hierarchical level (relative clause, unspecified for order) and, therefore, we propose that such an unordered level does not exist. In response to Hartsuiker & Bernolet, the absence of priming for PO structures following shifted PO prime sentences is compatible with an account in which there is no priming of independent hierarchical relations, and thematic-function and thematic-order priming effects counteract each other (Cai et al. Reference Cai, Pickering and Branigan2012).

Of course, we do not assume that syntax can be reduced to simple sequential structure (see also Gaston et al.). Otherwise, we would not assume constituent-structure representations such as those in Figure 1. Priming also cannot be reduced to linear order. For example, it occurs when the internal structure of prime and target constituents are different (e.g., complex vs. simple noun phrases in dative priming; see Fox Tree & Meijer Reference Fox Tree and Meijer1999; Pickering & Branigan Reference Pickering and Branigan1998), or when the critical alternation occurs at a different location in the prime and target – for example, datives in main versus subordinate clauses (Branigan et al. Reference Branigan, Pickering, McLean and Stewart2006) or noun phrases sentence-initially versus -finally (Melinger & Cleland Reference Melinger and Cleland2011).

Hartsuiker & Bernolet and Francis question how we interpret the representation of coerced sentences. Hartsuiker & Bernolet argue that the control (event) sentences could be given a coerced reading (i.e., both sentences could have a missing predicate represented). But the picture presented with the prime sentence forced it to have an event reading. Additionally, although coerced and control sentences did not differentially prime the production of full VP targets (relative to VP-primes such as began drinking the champagne), they did differentially prime the production of coerced targets, with coerced primes inducing more coerced targets than control primes. This finding further suggests that participants did not postulate a predicate for control primes. Francis proposes that the data of Raffray et al. (Reference Raffray, Pickering, Cai and Branigan2014) might be due entirely to syntactic priming and so might indicate that the three sentence types (began the speech, began the champagne, began drinking the champagne) all have different syntactic structures. The problem with this explanation is that a purely syntactic account might argue that began the champagne had a covert V (corresponding to drinking) but began the speech could not have a covert V – and if this is the case, began the champagne should behave more similarly than began the speech to began drinking the champagne, contrary to the findings of Raffray et al..

R11. Further directions

We argue in the target article that structural priming can be used to address outstanding linguistic controversies, and highlight some specific examples (sect. 3.1). Commentators identify interesting further cases that exemplify the potential of this method.

Francis discusses how priming might be used to clarify the syntactic representation of quantificational NPs and specifically whether they have the same representation as collective NPs. Assuming that an appropriate experiment can be designed (e.g., an alternative form for a lot of papers), then any finding that A bundle of sticks primes A lot of papers would suggest that quantificational NPs and collective NPs involve the same syntactic representation, in which the first noun acts as the head. We suspect that such priming would occur, because priming appears to be sensitive to categorical distinctions rather than featural distinctions (e.g., tense or number; Pickering & Branigan Reference Pickering and Branigan1998).

O'Grady highlights how priming could be used to test specific representational questions across typologically diverse languages. As we note in section 3.2, such studies can address fundamental questions about the diversity of linguistic representations and potential language universals. With respect to subject-verb sequences, it may well be possible to prime neither is versus are (for example) and demonstrate that such priming constitutes priming of subject-verb agreement, rather than simply priming of the verb form (see Haskell et al. Reference Haskell, Thornton and MacDonald2010). Any such priming could be due to semantic rather than syntactic relations, as both types of dependency affect choice of verb form (cf. Bock et al. Reference Bock, Nicol and Cutting1999). And even if the effects are syntactic, they could relate to a unit consisting of the whole sentence rather than one consisting of the subject and verb. For example, the locus of priming for neither are satisfactory could be NP_+ pl V_+ pl AP rather than NP_+ pl V_+ pl. We also suggest that priming different interpretations of Everyone didn't finish the project takes place at the semantic rather than syntactic level.

Rees & Bott point to Bott and Chemla's (Reference Bott and Chemla2016) finding that scalar implicatures can be primed (see also Maldonado et al.). We see this as a fascinating example of how priming can be extended to further linguistic domains, specifically diverse aspects of structure relating to interpretation. Priming of pragmatic representations is, of course, compatible with it being sensitive to all levels of linguistic representation.

Cai & Liu and Kootstra & Rossi suggest ways in which structural priming can be used to investigate bilingual representation. In particular, Kootstra & Rossi note that priming can be used to study first language attrition. We agree: It could be used to determine whether an L1 had become completely inaccessible or not. Of course, it would be possible for a language to be temporarily inaccessible but to become accessible again later in life, and it might be possible to use priming to render an inaccessible language accessible again.

Finally, we support Hagoort's suggestion that fMRI adaptation offers promising directions for future research that may allow us to identify more precisely the type of priming that occurs, as different linguistic representations lead to adaptation in different brain areas. This approach may be particularly appropriate for studying structures that do not involve alternations. It is also likely to yield important findings about the neural basis of linguistic representation. It shows one way in which our proposals about the use of priming to understand linguistic representation may be extended and therefore underlines their value in the development of the cognitive sciences of language.