Branigan & Pickering (B&P) rightly state that acceptability judgments access linguistic representations only indirectly via language comprehension and production processes. This makes it difficult to draw strong conclusions about the nature of representations, because “the data are compatible with particular grammar-processor pairings, not just with particular grammars” (sect. 1.1, para. 4–5).
This problem, however, applies to all methodologies, including priming. In developmental psycholinguistics, we generate and test theories that propose both descriptions of adult representations and explanations of how those representations develop (e.g., Goldberg Reference Goldberg2006; Pinker Reference Pinker1984). We have learnt that restricting ourselves to any method – even a well-studied method like priming – yields only incomplete data about the nature of linguistic representation. For example, in priming studies, we access children's linguistic representation through the lens of a still poorly understood effect of priming on children's sentence production. To interpret our data, we must make inferences about the mechanisms underlying priming and how these mechanisms use the child's emerging linguistic knowledge. If our inferences about those processes (and how they use linguistic representations) are flawed, the conclusions we draw about representations will be flawed. In other words, if we rely on priming only, we will generate an incomplete theory of linguistic representations, which is likely to fail once tested using different methodologies.
The solution is a multi-method approach. As we have argued in Monaghan and Rowland (Reference Monaghan and Rowland2017), by gathering evidence from different methods, we can converge on a more holistic understanding of the child's developing representations. Below, we illustrate our argument with two examples.
First, we examine how to determine the linguistic representations children hold at different ages. Structural priming studies have been informative here, showing that even young children's syntactic representations are abstract enough to support generalisation across verbs. For example, 3-year-old children produce more double-object datives (DODs) after a double-object dative prime than after a prepositional dative (PD) prime, even when the prime and target sentences share no content words (Peter et al. Reference Peter, Chang, Pine, Blything and Rowland2015; Rowland et al. Reference Rowland, Chang, Ambridge, Pine and Lieven2012; see Thothathiri & Snedeker (Reference Thothathiri and Snedeker2008b) for similar results in comprehension).
It is tempting to conclude from this that children's dative representations are not only abstract, but also adultlike – that “evidence from these studies suggests that, from a relatively young age, children's structural representations are similar to adults” (sect. 3.3, para. 2). Although priming studies tell us that children's dative representations are abstract, however, they are not necessarily adultlike. This would be to assume that the priming mechanism requires adultlike representations, which is yet to be ascertained.
In fact, findings from other methodologies reveal asymmetries in the pattern of PD and DOD acquisition, which suggest that the two are not equally adultlike early on. Although naturalistic studies show that children produce DODs earlier than PDs (Snyder & Stromswold Reference Snyder and Stromswold1997), early DOD use is restricted to a small set of high-frequency verbs (Campbell & Tomasello Reference Campbell and Tomasello2001). Children are more productive earlier with the PD, in the sense of being more willing to use PD structures in novel verb experiments. For example, Conwell & Demuth (Reference Conwell and Demuth2007) showed that 3-year-olds were more likely to generalise a novel verb heard in a DOD to a prepositional form (e.g., to produce he pilked the cup to Toby after hearing I pilked Toby the cup) than they were to generalise a novel verb heard in a PD to a double-object form. There is a similar asymmetry in novel verb comprehension (Rowland & Noble Reference Rowland and Noble2011).
A number of explanations might integrate these findings. Perhaps children's very early double-object datives are restricted to a few, frequent verb-specific patterns, which become so entrenched that it remains difficult to generalise the structure to novel verbs, even when representations become more abstract (Tomasello Reference Tomasello2000). Alternatively, PD representations may have a “head start on the process of becoming abstract” because of their structural similarity to the early acquired transitive structure (Campbell & Tomasello Reference Campbell and Tomasello2001, p. 266). More work is needed here. Our point is simply that, without a multi-method approach, we would not gain these insights into the nature of children's developing knowledge.
Our second example demonstrates how a multi-method approach combining computational modelling with experimental work enables us to test the complex interplay between representation and the processing of those representations. Distinguishing processing from representation is far from trivial, as defenders of acceptability judgments have indicated (sect. 1.2, paras. 1–2). Consequently, a theoretical model, as presented by B&P, provides only a first step as a description of representational features and the likely processes that operate over these representations. Computational modelling of experimental findings is needed to test the necessity and sufficiency of representation and processing in the language system, as well as the extent to which there is separability between representations and the processes operating over them.
Chang et al. (Reference Chang, Dell and Bock2006) showed that a computational model with distinct event semantics and syntactic knowledge was able to simulate a series of syntactic priming effects, but only when points of interaction between those representations was limited. Language, however, is acquired and processed in a rich, multimodal situation that goes far beyond the representations described in Chang et al. (Reference Chang, Dell and Bock2006). For instance, Smith et al. (Reference Smith, Monaghan and Huettig2017b) constructed a computational model of processing in visual-world paradigm tasks, determining how phonological, visual, and semantic representations are integrated during speech perception. They demonstrated that the behavioural data could be simulated most effectively only when these representations interpenetrated throughout processing, rather than assuming autonomy of these representations, cohering at the outcome of processing. An extension of this model to language development showed that differences in quantity of exposure to a rich, multimodal language environment was sufficient to simulate child and adult behavioural differences in visual world paradigm tasks (Smith et al. Reference Smith, Monaghan, Huettig, Twomey, Smith, Westermann and Monaghan2017a). In other words, combining insights from the rich interactivity of multimodal information with the possible advantages of modular processing of this richness requires computational implementations to distinguish alternative accounts.
In sum, the process of determining linguistic representations from empirical data is far from straightforward and requires a multi-methodological approach.
You have
Access
Branigan & Pickering (B&P) rightly state that acceptability judgments access linguistic representations only indirectly via language comprehension and production processes. This makes it difficult to draw strong conclusions about the nature of representations, because “the data are compatible with particular grammar-processor pairings, not just with particular grammars” (sect. 1.1, para. 4–5).
This problem, however, applies to all methodologies, including priming. In developmental psycholinguistics, we generate and test theories that propose both descriptions of adult representations and explanations of how those representations develop (e.g., Goldberg Reference Goldberg2006; Pinker Reference Pinker1984). We have learnt that restricting ourselves to any method – even a well-studied method like priming – yields only incomplete data about the nature of linguistic representation. For example, in priming studies, we access children's linguistic representation through the lens of a still poorly understood effect of priming on children's sentence production. To interpret our data, we must make inferences about the mechanisms underlying priming and how these mechanisms use the child's emerging linguistic knowledge. If our inferences about those processes (and how they use linguistic representations) are flawed, the conclusions we draw about representations will be flawed. In other words, if we rely on priming only, we will generate an incomplete theory of linguistic representations, which is likely to fail once tested using different methodologies.
The solution is a multi-method approach. As we have argued in Monaghan and Rowland (Reference Monaghan and Rowland2017), by gathering evidence from different methods, we can converge on a more holistic understanding of the child's developing representations. Below, we illustrate our argument with two examples.
First, we examine how to determine the linguistic representations children hold at different ages. Structural priming studies have been informative here, showing that even young children's syntactic representations are abstract enough to support generalisation across verbs. For example, 3-year-old children produce more double-object datives (DODs) after a double-object dative prime than after a prepositional dative (PD) prime, even when the prime and target sentences share no content words (Peter et al. Reference Peter, Chang, Pine, Blything and Rowland2015; Rowland et al. Reference Rowland, Chang, Ambridge, Pine and Lieven2012; see Thothathiri & Snedeker (Reference Thothathiri and Snedeker2008b) for similar results in comprehension).
It is tempting to conclude from this that children's dative representations are not only abstract, but also adultlike – that “evidence from these studies suggests that, from a relatively young age, children's structural representations are similar to adults” (sect. 3.3, para. 2). Although priming studies tell us that children's dative representations are abstract, however, they are not necessarily adultlike. This would be to assume that the priming mechanism requires adultlike representations, which is yet to be ascertained.
In fact, findings from other methodologies reveal asymmetries in the pattern of PD and DOD acquisition, which suggest that the two are not equally adultlike early on. Although naturalistic studies show that children produce DODs earlier than PDs (Snyder & Stromswold Reference Snyder and Stromswold1997), early DOD use is restricted to a small set of high-frequency verbs (Campbell & Tomasello Reference Campbell and Tomasello2001). Children are more productive earlier with the PD, in the sense of being more willing to use PD structures in novel verb experiments. For example, Conwell & Demuth (Reference Conwell and Demuth2007) showed that 3-year-olds were more likely to generalise a novel verb heard in a DOD to a prepositional form (e.g., to produce he pilked the cup to Toby after hearing I pilked Toby the cup) than they were to generalise a novel verb heard in a PD to a double-object form. There is a similar asymmetry in novel verb comprehension (Rowland & Noble Reference Rowland and Noble2011).
A number of explanations might integrate these findings. Perhaps children's very early double-object datives are restricted to a few, frequent verb-specific patterns, which become so entrenched that it remains difficult to generalise the structure to novel verbs, even when representations become more abstract (Tomasello Reference Tomasello2000). Alternatively, PD representations may have a “head start on the process of becoming abstract” because of their structural similarity to the early acquired transitive structure (Campbell & Tomasello Reference Campbell and Tomasello2001, p. 266). More work is needed here. Our point is simply that, without a multi-method approach, we would not gain these insights into the nature of children's developing knowledge.
Our second example demonstrates how a multi-method approach combining computational modelling with experimental work enables us to test the complex interplay between representation and the processing of those representations. Distinguishing processing from representation is far from trivial, as defenders of acceptability judgments have indicated (sect. 1.2, paras. 1–2). Consequently, a theoretical model, as presented by B&P, provides only a first step as a description of representational features and the likely processes that operate over these representations. Computational modelling of experimental findings is needed to test the necessity and sufficiency of representation and processing in the language system, as well as the extent to which there is separability between representations and the processes operating over them.
Chang et al. (Reference Chang, Dell and Bock2006) showed that a computational model with distinct event semantics and syntactic knowledge was able to simulate a series of syntactic priming effects, but only when points of interaction between those representations was limited. Language, however, is acquired and processed in a rich, multimodal situation that goes far beyond the representations described in Chang et al. (Reference Chang, Dell and Bock2006). For instance, Smith et al. (Reference Smith, Monaghan and Huettig2017b) constructed a computational model of processing in visual-world paradigm tasks, determining how phonological, visual, and semantic representations are integrated during speech perception. They demonstrated that the behavioural data could be simulated most effectively only when these representations interpenetrated throughout processing, rather than assuming autonomy of these representations, cohering at the outcome of processing. An extension of this model to language development showed that differences in quantity of exposure to a rich, multimodal language environment was sufficient to simulate child and adult behavioural differences in visual world paradigm tasks (Smith et al. Reference Smith, Monaghan, Huettig, Twomey, Smith, Westermann and Monaghan2017a). In other words, combining insights from the rich interactivity of multimodal information with the possible advantages of modular processing of this richness requires computational implementations to distinguish alternative accounts.
In sum, the process of determining linguistic representations from empirical data is far from straightforward and requires a multi-methodological approach.