Atypical language development matters: A commentary on Ambridge (2020)

Ambridge (2020) proposes an exemplar model for children’s acquisition of (among other things) words and sentence-level constructions. In this model, language users store exemplars of every utterance they have heard or produced, each of which contains its surface form/phonetic detail, the child’s understanding of their own and/or the speaker’s meaning, and the context/situation/task at hand. Apparent abstractions in the model actually derive from analogies that children have computed ‘on the fly’, which are weighted by degrees of similarity between stored exemplars and the target. A key claim of this model is that it provides a unifying account of language acquisition, in that the processes of learning across all domains (phonology, words, sentences) are the same; moreover, such identical learning processes explain attested inter-domain correlations. Thus, children who are better at storing – or making analogies based on – sounds, meanings, and/or situations at the producing/understanding words level would also be the ones proficient at doing the same storage and/or analogies at the producing/understanding sentences level. According to Ambridge’s model, then, abstracting syntactic constructions boils down to the same storage and analogical processes as generalizing word meanings.

My commentary is based on findings from children with atypical language development, because the more protracted period of language development frequently observed in children with atypical language can afford a developmental separation of variables relevant for language acquisition (Levy, 1994; Tager-Flusberg, 1994b); moreover, cases of atypical children offer ‘natural experiments’ in which specific variables deemed critical for language acquisition to proceed are changed, enabling their roles to be scrutinized independently (Gleitman, 1984; Tager-Flusberg, 1994a). The case of atypical language that I will focus on here is that of children with Autism Spectrum Disorder (ASD), which is a neurodevelopmental disorder characterized by deficits in social communication and interaction, and by excessive repetitive/stereotypical behaviors (American Psychiatric Association, 2013). The findings I will discuss come from the Longitudinal Study of Early Language (LSEL; Naigles & Fein, 2017), in which my colleagues and I have tracked the developmental course of the production and comprehension of language in a sample of 30+ children with ASD, and compared them with a control sample of typically developing (TD) children who were matched on expressive language with the children with ASD at the beginning of the study. Using the Intermodal Preferential Looking paradigm (Naigles & Tovar, 2012), across a number of visits the children were tested on their comprehension of two grammatical constructions: SVO word order (e.g., distinguishing ‘The girl tickles the boy’ from ‘The boy tickles the girl’) and wh-questions (e.g., seeing an apple hit a flower pot and then distinguishing ‘What did the apple hit?’ and ‘What hit the flower?’). They were also tested on two word learning principles: the shape bias (e.g., extending ‘dax’ to a new instance of the same shape over same color) and syntactic bootstrapping (e.g., mapping ‘gorping’ in ‘The duck is gorping the bunny’ onto a novel causative action over a novel noncausative action). The relationships among the children’s performance with these tasks will demonstrate the separability of their acquisition – and abstraction – of grammar from either their word learning or pragmatics abilities. Thus, they tell against Ambridge’s ‘exemplar-based’ model, especially his proposal that these levels are governed by the same processes and so can be explained by a single (exemplar) account of language representation.

Three sets of findings are particularly relevant here. First, the children with ASD, as a group, engaged in syntactic bootstrapping (Naigles et al., 2011), which most researchers would agree demonstrates that they have abstracted an SVO or transitive sentence frame (Fisher et al., 1991; Gleitman, 1990; Noble et al., 2011). Ambridge might argue that the children’s behavior is based instead on analogies between ‘The duck is gorping the bunny’ and stored exemplars of, say, ‘Mom is pushing the baby,’ and that those children with larger vocabularies might be expected to perform these analogies better because they might have stored more utterances including various animate agents, actions, and animate patients. Indeed, these children with ASD’s successful performance correlated with their vocabulary size; however, their performance also was predicted – independently – by their speed of processing very different SVO sentences at an earlier point in development (Naigles et al., 2011). Thus, syntactic proficiency and lexical proficiency contributed jointly but independently to the children’s use of sentence frames in determining verb meaning; one was not reducible to the other. Postulating an abstract SVO structure can account for these data whereas Ambridge’s radical exemplar model cannot.

Second, the children with ASD demonstrated successful comprehension of wh-questions, even with visual and auditory stimuli that were quite removed from the children’s usual situations (where apples don’t usually hit flower pots; Goodwin et al., 2012; Jyotishi et al., 2017). As might be predicted from both Ambridge’s model and theories of ASD, the children’s wh-question performance correlated positively with their lexical and with their social/pragmatic development (Durrleman et al., 2016; Tager-Flusberg, 1994b). However, their performance also was predicted – independently – by their degree of understanding SVO sentences earlier in development (Jyotishi et al., 2017). That is, variation in wh-question comprehension was not reducible to variation in pragmatics, nor was it reducible to variation in vocabulary. Variation in syntax also played an independent contributory role, and in so doing, demonstrates that syntax is a separable component during language development (Jyotishi et al., 2017). In Ambridge’s model, each instance of language use is stored as an exemplar, so atypical language should simply be reducible to being better or worse at storing or retrieving exemplars, in general. Whereas what our data show is that children who are better at representing SVO sentences early are better at understanding wh-questions later, independently of their ease of representing pragmatics/speaker intentions. Postulating separable levels of grammar and pragmatics can handle these findings whereas Ambridge’s radical exemplar model cannot.

Third, our findings demonstrate that these children with ASD are not using the same processes when acquiring different domains of language. While our TD group demonstrated a shape bias in novel word learning across multiple visits, our ASD group never did (Potrzeba et al., 2015; Tek et al., 2008). Moreover, while developing a shape bias has been postulated to depend on the number of shape-oriented words in a child’s vocabulary (Perry & Samuelson, 2011), most of the children with ASD were rated as knowing just as many shape-oriented words as their TD peers, and yet did not show a shape bias (Abdel-Aziz et al., 2018). Finally, most of the children with ASD who did not show a shape bias nonetheless demonstrated abstraction abilities at varied levels by (a) understanding sentences with familiar verbs in SVO word order, (b) using transitive syntax to map novel verbs onto causative actions (i.e., doing syntactic bootstrapping), and (c) demonstrating wh-question comprehension (Naigles & Tek, 2017). Thus, the children were not lacking in the ability to abstract (or analogize) in general, nor did they lack in number of relevant stored exemplars. One must conclude from these data that the process of making generalizations or abstractions with at least some lexical semantic constructs is impaired in these children, while the process of making generalizations or abstractions with at least some grammatical constructs is functioning in typical fashion. In which case there are at least two independent processes of abstraction or generalization going on, which is just what Ambridge’s radical exemplar model cannot account for.