Sage Journals: Discover world-class research

Abstract

Children with developmental language disorder (DLD) present with unexplained difficulties with both expressive and receptive language. A key conundrum is why their prognoses are often poorer than children with evidenced brain lesions. How can children with early focal lesions acquire language so well compared with children with DLD who do not present with gross neurological abnormalities? A possibility raised by Dell and Chang in 2013 is that learning in DLD is fundamentally derailed by an exemplar-based learning style. Such a process has been robustly demonstrated in studies of artificial neural networks. This theoretical article discusses exemplar theory, the role exemplars may play in language development, and surveys evidence suggesting an exemplar-based learning style in DLD. It concludes that this provides a plausible new approach to both explaining the condition and designing new therapies.

Keywords

developmental language disorder syntax exemplar theory artificial intelligence

Introduction

Children with developmental language disorder (DLD) have severe difficulties acquiring, comprehending and producing language. Difficulties are evidenced across a wide range of domains, including word-learning, vocabulary, morphology (in particular verb morphology), use of complex sentences, pragmatics and phonology. A broad overview of linguistic difficulties in DLD is provided by Leonard (2014), while bearing in mind the shift in terminology from Specific Language Impairment (SLI) to DLD (see below).

In this review paper, I propose a new account of DLD, arguing that this condition is due to an exemplar-based style of learning and processing. The primary motivation for this argument is the ‘tough-but-fragile’ paradox. This is my term to describe an observation by Dell and Chang (2013). According to this paradox, the language system, in most individuals, is robust, being able to overcome severe obstacles such as early neurological damage. However, in individuals with DLD, it is extremely fragile, with severe long-term implications for language development. I then propose, based on Dell and Chang (2013), that exemplar-based learning enables us to explain this paradox. A novel aspect of this account is the fact that it is based on the principles of machine learning/artificial intelligence (AI). I will argue that, given substantial improvements in the power of generative AI systems, these may act as good models of human learning and cognition and may also be used to identify ways in which learning may break down.

A Note on Developmental Language Disorder

DLD is a recent label which revises the criteria used to diagnose SLI. A key change is the lowering of IQ cut-offs, with many studies of DLD using a cut-off of −2 standard deviations, as opposed to −1 standard deviations in the SLI literature. Consequently, the populations of children with SLI and children with DLD do not wholly overlap. This raises the question of whether research into SLI may inform our understanding of DLD. An argument in favour of comparing across research literatures is that linguistic profiles in children with language difficulties are not substantially influenced by IQ. For example, the linguistic phenotype in children with SLI closely resembles that of children who have a combination of language difficulties and moderate learning difficulties (e.g. Bishop, 1994; Rice et al., 2004). In fact, the finding that language profiles are largely unaffected by IQ constitutes an important motivation for lowering IQ cut-offs in the CATALISE consensus definition of DLD (Bishop et al., 2016, Criteria and Terminology Applied to Language Impairments: Synthesising the Evidence, statement 23, pg.15). In the light of these findings, Bishop (2020) argues that the research literature on SLI remains relevant to a discussion of DLD.

Given the above arguments, I will use the term ‘DLD’ when discussing research conducted using the recruitment criteria for SLI. The reader should bear in mind that most studies published prior to 2018 will use the ‘SLI’ label and should be aware that recruitment criteria (particularly with respect to IQ) differ slightly.

The Tough-But-Fragile Paradox

To my knowledge, Dell and Chang (2013) were the first authors to explicitly formulate the tough-but-fragile hypothesis. Regarding the ‘fragility’ of language in DLD, there is, indeed, evidence that children with DLD experience severe and protracted language difficulties. Botting (2020) found that 40% of a group of 84 24-year-old adults with DLD scored below <2 standard deviations on the Clinical Evaluation of Language Fundamentals (CELF) core language index (Semel et al., 2006). Given that the CELF is standardised up to 21;11, this study most likely overestimated their language ability. Unfortunately, due to the practicalities of running longitudinal studies, there are few studies of DLD in adulthood, and Botting’s (2020) study is unique in combining a large sample with standardised language assessments. Nonetheless, poor prognosis is widely recognised by clinicians and is emphasised by the CATALISE consensus definition of DLD, which specifies ‘language problems enduring into middle childhood and beyond’ (Statement 2: Bishop et al., 2017).

This poor prognosis contrasts with the better outcome of children with early focal lesions whose language systems are surprisingly ‘tough’. In a study of 38 children, aged 5 to 8 who experienced early focal lesions, Bates et al. (2001) observed normal range performance on measures derived from spoken language samples; word type frequency, number of propositions per utterance, Mean Length of Utterance (MLU) in morphemes and counts of lexical, morphological and grammatical errors. A more recent study by Newport et al. (2022) investigated 15 adolescents who suffered a left-lateralised perinatal ischaemic stroke. They did not significantly differ from age-matched peers on CELF 5 Word Structure (Semel et al., 2017), Test of Reception of Grammar (TROG) overall accuracy (Bishop, 2005), TROG complex sentences and a test assessing comprehension of actives and passives. This may reflect the reorganisation of language functions in the brain, with the right hemisphere taking over language processing.

While the research literature on prognosis in both DLD and early focal lesions is small and fraught with methodological issues (e.g. the lack of a standardised language test for adults), existing data support the tough-but-fragile hypothesis. Current evidence suggests that many children with early focal lesions will have a better language prognosis than many children with a diagnosis of DLD.

Exemplars as a Solution to The Tough-But-Fragile Problem

Dell and Chang (2013) propose that a dependence on exemplar-based representations may explain fragility in an otherwise tough language system. These representations are detailed, concrete, tied to specific usage events and most likely depend on episodic memory (Ambridge, 2020a). Many accounts of exemplar-based representations focus on word meaning (e.g. Ross & Makin, 1999). A child may initially underextend a word, for example assume that dog only applies to the family’s pet dog (e.g. Dromi, 1987). This is an exemplar-based representation as it is detailed and tied to specific usage events, for example experiences with a family pet. Subsequently, the child may develop a more abstract representation applicable to dogs outside the home. A key characteristic of exemplars is that they contain information which is irrelevant from the perspective of the adult system. For example, a semantic representation of the family’s pet dog will contain information which is not relevant to the abstract concept of DOG, for example its name, size and temperament. Key papers outlining exemplar-based approaches to linguistic representation and language acquisition are Ambridge (2020a), Bybee (2013) and Pierrehumbert (2001).

Exemplars are often contrasted with prototypes. While a prototype is also derived from multiple usage events, it is a stored representation, for example a representation of the ‘ideal’ dog. By contrast, according to exemplar theory, categories are constructed spontaneously based on the simultaneous activation of multiple exemplars.

Exemplar theorists highlight the fact that we retain detailed information about our linguistic experiences. For example, phonological representations may contain subphonemic information. In frequent every, the second syllable is rarely produced (/e.və.riː/ changes to /ev.riː/). By contrast, it is realised as syllabic /r/ in mid-frequency words, for example memory, and is fully-articulated in low-frequency words (Hooper, 1976; cited by Pierrehumbert, 2001)¹. Similar processes occur in phrases. For example, the /I/ in give is more centralised in frequent phrases, for example give a chance versus give a pen (Hay & Bresnan, 2006; cited by Ambridge, 2020a). These phenomena indicate that ‘detailed phonetic memories are associated with individual words’ (Pierrehumbert, 2001, p. 139). Hay and Bresnan’s (2006) data also suggest that phonetic detail is also stored within phrasal units.

Syntactic representations may also contain fine-grained contextual information. For example, constructional idioms exhibit detailed context-dependent meanings which go beyond or even contradict their literal interpretation. For example, the What’s X doing? construction, e.g. what’s that jacket doing in my wardrobe?, is used to highlight incongruity. Bybee (2013) argues that the only way to explain this is via exemplar theory. Because pragmatic interpretations form part of our detailed linguistic traces, we are able to extrapolate across these and thereby arrive at the relevant form-function pairing.

Exemplar Representations in Child Language

Children exhibit a dependence on exemplars which may result in linguistic generalisations which are ‘faulty’ in the sense that they do not occur in the adult system. One example is children’s conflation of tense and aspect. Because achievement and accomplishment verbs (e.g. jump, arrive) imply a natural end-state, they are more likely to occur in the past tense. Some studies indicate that children’s tense marking is more accurate for these verbs, indicating a sensitivity to the probabilistic alignment of tense and aspect in the input (Leonard et al., 2007; Owen, 2010).

Other studies find that children’s inflection of past tense is strongly influenced by the phonological properties of the stem. Young children often omit the regular past tense affix when the stem already ends in an alveolar plosive (/t/ or /d/), for example Yesterday she mend it (e.g. Owen, 2010). This, most likely, reflects a process of comparing the stem mend with other stored forms already inflected for past tense, for example laughed, robbed, hurt, put. Because the stem shares characteristics with stored inflected forms, it ends with an alveolar plosive, the affixation rule is not applied. Bybee and Slobin (1982) citing Zager (1982) describe this as a product-oriented schema because the morphological system is focused on the product of the affixation, that is, a form which ends with /t/ or /d/, rather than on the process of adding an affix. This phenomenon is exemplar-based as it involves activation of detailed representations of inflected forms and also involves generalisation based on irrelevant information: the stem-final consonant, which is not a categorical determinant of the inflected form in the adult system.

Children’s use of the passive also demonstrates exemplar-based generalisation. In competent adult speakers, this involves the movement of a post-verbal argument into subject position. Almost any active transitive sentence may be passivised irrespective of verb semantics. Often a key motivation is to move an argument into a position where it is topicalised. However, early passives exhibit semantic constraints which do not play such a strong role in the adult system. A key property is that the subject must be highly affected (Pinker, 1989). Exactly when the passive becomes ‘abstract’, and consequently less influenced by semantic constraints, has been rigorously debated. Even as old as five or six, children struggle to comprehend passive sentences where the subject is not highly affected (Ambridge, 2020a).

How Important Are Exemplars to Language Acquisition?

There is strong evidence that at least some processes evident in early child language exhibit exemplar-like properties. However, this does not entail a central role for exemplars in language development. Critics of exemplar theory question whether we can store the kind of detailed representations required by the theory. Moreover, it is not clear how exemplar-based analogy would work without some kind of a priori specification of which dimensions are most relevant (Demuth & Johnson, 2020; Lieven et al., 2020; Schuler et al., 2020). In response to these criticisms, and others, Ambridge (2020b) acknowledges that exemplar-based representations must become more abstract over time. Given the many criticisms of exemplar theory, it might be best to regard exemplar-based phenomena as a characteristic of an early-stage of language acquisition, which is not a necessary precursor to the development of more abstract representations.

However, there is evidence that exemplar-based language use persists into adulthood. Though competent adult speakers rarely make past tense errors, product-oriented schemas may exist in the adult system. For example, Bybee and Slobin (1982) found that adult suffix-omission errors tended to occur on verbs where the stem form ended in a /t/ or /d/. Albright and Hayes (2003) also found an effect of stem phonology with respect to grammaticality judgements of past tense forms. Moreover, past tense forms which are in free variation in many varieties of English, for example fit_PAST versus fitted_PAST, exhibit the influence of product-oriented schemas.

Turning to the passive, the concept of affectedness influences adults’ judgements of the grammaticality of passive sentences long after it has ceased to impact on performance in controlled tests of comprehension (Ambridge et al., 2016). The authors conclude that ‘a semantic constraint on the passive must be incorporated into accounts of the adult grammar’ (p. 1435). In addition, alongside the much-studied be passive, adult speakers also use the more formulaic get passive, which has strong connotations of affectedness (Thompson et al., 2013). Finally, Street and Dąbrowska (2014) identified lexical semantic constraints on adults’ abilities to process passive sentences, for example the frequency with which verbs co-occur with active and passive frames.

Data from adult usage of the past tense and the passive indicate that exemplars do not suddenly give way to abstract representations, but exist, albeit in a limited fashion, within the adult system. Consequently, the shift from exemplar-based to abstract representations can be conceived as a gradual and graded process. This is an important principle when discussing the possibility of exemplar-based representations in DLD.

How Dell and Chang Evoke Exemplars to Explain Tough-But-Fragile Systems

Dell and Chang (2013) argue that exemplar-based learning can severely affect generalisation within a system which is ostensibly ‘tough’. This is demonstrated via an artificial neural network model. Such networks display excellent learning as they have a strong tendency to minimise prediction error. Once a model has reached a certain level of complexity, in terms of the number of nodes and training epochs, the removal of nodes has relatively little impact on the model’s capacity to learn and generalise. This ‘toughness’ mimics children’s recovery from focal legions.

However, neural networks are also fragile in the sense that they are prone to overfitting. This involves modelling the noise in the training data, which, in turn, impacts on generalisation. Here, the term ‘noise’ is used to describe information which is not relevant to the learning task and does not support generalisation. For example, using lexical aspect or stem phonology to determine past tense inflection misses the broader generalisation that regularity is an abstract feature in the verb lemma. Likewise, using affectedness to determine the suitability of the passive misses the broader generalisation that virtually any post-verbal argument may be passivised for discourse reasons.

Though Dell and Chang do not use the term overfitting in their 2013 paper, they refer to a paper by Chang (2002) which discusses overfitting in the context of neural network models which exhibit poor generalisation (see below). They also describe a process of generalisation based on irrelevant information. Focusing on the past tense, they suggest that children might decide to add the regular past tense suffix based on stem phonology. They propose that a child might decide that a verb beginning /pe/ is a regular verb. Existing data support the possibility that children attend to the phonological properties of the stem when inflecting regular verbs, but with the proviso that the stem-final properties are most relevant. We have already seen that some children exhibit product-oriented schemas, such that verbs which end with an alveolar plosive are less likely to be regularly inflected (e.g. Owen, 2010).

Chang’s (2002) study artificially induced overfitting in an artificial neural network model by adding redundant connections to create, a ‘linked path’ model. The additional connections linked the message/lexical system with the sequencing system. Such a manipulation caused the model to learn correspondences between noun lemmas and their thematic roles, for example assuming that dog is always an AGENT. This would lead the model to successfully process the sequence the dog chased the cat, but not the cat chased the dog. This can be described as overfitting as it involves the acquisition of irrelevant information, namely the correspondence between the dog and the AGENT thematic role. While Chang artificially manipulates his model, in most neural networks overfitting happens naturally when the number of data points approaches the number of parameters in the model.

It must be acknowledged that overfitting is not a critical problem in machine learning. Engineers can employ various techniques to minimise overfitting, for example adding noise to the training data. Moreover, with substantial data, large networks recover from initial overfitting, a phenomenon called ‘double descent’ (Belkin et al., 2020). As our brains are more complex than the most sophisticated AI model to date, it seems unlikely that they are going to be worse at recovering from overfitting. Nonetheless, given that overfitting is an emergent property of computational systems directly modelled on human neural architectures, it provides a potentially important explanation for language learning difficulties.

Dell and Chang (2013) note that exemplar-based learning strategies may be relatively accurate and are therefore reinforced. In machine learning terms, the model is trapped in a ‘local minimum’, a term used to describe the entrenchment of a suboptimal generalisation. Product-oriented schemas for the past tense may be regarded as an example of this. This generalisation is at least partially successful as many verbs ending in an alveolar plosive do not take the past tense suffix. Albright and Hayes (2003) identify another suboptimal generalisation, noting that all verbs in English ending in a voiceless fricative (/f/, /θ/, /s/, /ʃ/) inflect for past tense via addition of the /t/ allomorph. Though this generalisation is completely reliable, it may foster an overdependence on stem phonology, which may prevent the child from attaining the broader generalisation, that regularity is encoded in the lemma.

Passive use may likewise exhibit the fossilisation of suboptimal learning strategies. Be passive constructions containing unaffected subjects are more likely in formal contexts and in print (Hundt et al., 2024). Consequently, most passives we hear on a daily basis will contain affected subjects. For this reason, a passive schema based on affectedness may have sufficient reliability to become entrenched in adults with a reduced exposure to formal and print contexts (Street & Dąbrowska, 2014).

While, to my knowledge, Dell and Chang (2013) are the only researchers to propose an overfitting account of DLD, there have been numerous overfitting accounts of autism. These focus on sensory processing abnormalities, which are characterised as attempting to learn from the noise in the input (Cruys et al., 2013; Idei et al., 2020; Pellicano, 2010). In addition, autistic individuals often perform poorly on a visual categorisation task, which involves learning from noisy input. Categories consist of an arrangement of dots in two-dimensional space. Exemplars of these categories are generated by randomly displacing dots, but with noise carefully calibrated to make the category learnable. In comparison with non-autistic controls, adult autistic participants find it difficult to learn the categories under these noisy conditions (Gastgeb et al., 2012). Again, this learning style can be explained via a learning system which models noise rather than signal.

The research on autism described above suggests that individual differences can arise in the degree to which learning is based on exemplars. This is an important precondition of an exemplar-based account of DLD. It should be noted, however, that this does not presume a shared genetic basis for DLD and autism, a claim which is not supported by recent research (Nudel et al., 2024).

Examples of Exemplar-Based Learning in DLD

I will now explore whether widely observed linguistic difficulties in DLD may be explained via exemplar-based learning. For brevity’s sake, I will focus on two key linguistic characteristics of DLD, difficulties with tense-marking and complex sentences. In addition, I will investigate evidence for a chunk-based learning and processing style.

Tense-Marking

Tense-marking difficulties in DLD are well-evidenced and may constitute an important linguistic marker of DLD (Conti-Ramsden et al., 2001). A typical error is the Optional Infinitive (OI) whereby the infinitive is used instead of an inflected form (Rice et al., 1998). OI errors occur in both present and past tense contexts, affecting both regular and irregular verbs. Numerous accounts of tense-marking difficulties have attributed errors to the analysis of stored forms, and the development of linguistic schemas which are based on factors which do not guide generalisation in the adult system, for example phonological properties of the stem. These depict exemplar-like processes, though it is important to note that none, except Owen Van Horne et al. (2017), have explicitly used the term ‘exemplar’.

A frequent finding is that zero-affixation errors, for example she go, he laugh, are influenced by the phonological properties of the stem. Marchman et al. (1999) found that children with DLD (n = 31) were less likely to add the regular past tense suffix to verbs where the stem already ends in an alveolar plosive for example *he mend_PAST. The children were therefore using a product-oriented schema (though it is important to note that the authors did not use this term). Importantly, this pattern did not reflect a general avoidance of a syllabic affix, as it was also evident for irregular forms ending in -t or -d, for example *he build_PAST. Further evidence for product-oriented schemas is provided by Calder et al.’s (2022) study of 30 children with DLD. Product-oriented schemas were evident for both regular past tense verbs, and third person singular present tense forms, for example *he surprise_PRESENT. Marshall and van der Lely (2006) likewise found a strong impact of stem phonology on regular past tense suffixation in children with DLD, this time focusing on phonotactic probability/legality rather than stem-final segments.

A second finding supporting an exemplar-based account is that, for regular past tense verbs, the input frequency of the inflected form impacts on children with DLD, but not language-matched typically-developing peers. Van der Lely and Ullman (2001) found that the input frequency of the inflected form was significantly negatively associated with zero affixation errors in the DLD group, but not the language-matched control group. In Dutch, Rispens and De Bree (2014) found that, for novel verbs, the DLD group were likely to make errors involving the erroneous inclusion of a -te suffix for stems ending in a voiced consonant. They hypothesise that this may reflect the high frequency of the stem plus -te pattern in the input. These findings suggest a role for exemplars as, according to exemplar-theory, linguistic generalisation involves the activation of multiple prior experiences. Consequently, we would expect strong frequency effects.

If past tense difficulties reflect exemplar-based processes, it follows that we can improve past tense inflections by deliberately exposing children to stimuli which disrupt these processes. This was observed by Owen Van Horne et al. (2017). Using a sentence imitation task, they trained 18 children with DLD to produce regular past tense forms. Verbs were divided into easy and difficult sets based on a norming study, which investigated zero-affixation rates for a wide range of verbs (Owen Van Horne & Green Fager, 2015). This manipulated (a) their phonological properties, for example whether they lent themselves to a product-oriented schema, (b) their lexical aspect; in particular whether they implied a natural end state, and (c) input frequencies for inflected forms. Those children who were initially trained on verbs which deliberately disrupted their exemplar-based learning, for example atelic verbs ending in alveolar plosives, demonstrated significantly greater learning. By contrast, those children who were initially trained on ‘exemplar-like’ verbs demonstrated weaker learning. The fact that ‘exemplar-busting’ stimuli promote more rapid learning provides further proof for the exemplar-based nature of language learning in DLD.

Complex Syntax

Children with DLD frequently struggle to produce and comprehend sentences with non-canonical word order such as passives (e.g. Van der Lely, 1996), object questions (e.g. Deevy & Leonard, 2004) and relative clauses (e.g. Novogrodsky & Friedmann, 2006). Most accounts of this phenomenon posit difficulties with syntactic movement processes (Novogrodsky & Friedmann, 2006), or short-term/working memory (e.g. Deevy & Leonard, 2004). To my knowledge, no one has yet proposed an exemplar-based account of such difficulties. However, the data are consistent with an exemplar-based interpretation which highlights the role of discourse.

While psycholinguists focus on the structure of complex sentences, they often overlook their discourse properties. Relative clauses ground the head noun in the discourse, making it identifiable by both speaker and hearer (Fox & Thompson, 1990). Consequently, relative clause internal Noun Phrases (NPs) are usually pronominal. Questions, which, in English, have a similar syntactic structure, exhibit similar discourse properties. For example, in object questions, subjects are usually pronominal (Riches & Garraffa, 2017). Passives likewise have an important discourse function. By enabling us to place an underlying non-subject in a topicalised position, they facilitate topic maintenance (Ferreira, 2021). Experimental support for this view is provided by Thompson et al. (2013) who found that competent adult speakers produce frequent passives when patients are topicalised. For this reason, passive subjects are usually pronominal.

Discourse fundamentally shapes our use of complex sentences. Contemori and Belletti (2014) identified an intriguing phenomenon whereby speakers strongly avoid producing object relatives with full NPs, for example I would rather be the child that the father is combing, and instead, shift the subject of the relative clause into the by-phrase of a passive, for example I would rather be the child that is being combed by the father. 97% of adult participants exhibited this behaviour. Though Contemori and Belletti propose a syntactic movement account of their findings, discourse may be an important factor. The subject slot within an object relative, occupied by the father in the first example above, is discourse-given. This is because subjects are typically discourse-given, and, in addition, the relative clause has a backgrounding function. Consequently, the father, clashes with the discourse properties of this slot. This clash may be resolved by moving the father into a by-phrase. Because by-phrases are optional when discourse-given, their very presence signals discourse-new information.

Note that the preferred structure, I would rather be the child that is being combed by the father, is highly complex, containing both a relative clause and a passive. But it seems that, for the participants in the study, the drive to generate sentences which are discourse appropriate overrode formal constraints related to complexity. Interestingly, children with DLD exhibit this kind of behaviour, frequently passivising relative clauses when they contain two lexically restricted NPs, though at rates far below adults (Novogrodsky & Friedmann, 2006; Riches, 2012).

Exemplar accounts are consistent with such discourse-related phenomena because they presuppose that stored grammatical constructions retain detailed contextual information. Goldberg (2014) demonstrates this phenomenon, arguing that in ditransitive sentences, for example she gave him the ball, there is a strong bias for the indirect object to be pronominal. By contrast, the indirect object in prepositional datives frequently contains new information, for example she gave it to Prakash . The respective discourse properties of these grammatical constructions affect how they combine with complex constructions. For example, we rarely question or relativise the indirect object slot in a ditransitive construction consisting of a verb and two consecutive NPs, for example who did she give _ the ball? and the man who she gave _ the ball. . . These sentences sound more felicitous if we add the preposition to at the end, thereby ensuring that extraction is from the indirect object slot within a prepositional dative, for example who did she give the ball to _ ?, the man who she gave the ball to _. Because questions and relative clauses topicalise the wh-phrase / head noun, wh-movement must take place from a position which can also be topicalised, such as the indirect object position of prepositional datives.

If grammatical constructions strongly specify discourse properties, it follows that complex sentences will be hard to interpret when their discourse properties are atypical. For example, object relative clauses which contain lexically-restricted NPs are more difficult to process than those containing pronouns (Gordon et al., 2004). Even choice of pronoun matters. When discourse-accessible me and you are used in object relative clauses, they facilitate processing in comparison to third person pronouns whose referent is contextually-derived (e.g. Haendler et al., 2015).

The degree to which processing is impacted by discourse is likely to reflect the degree to which representations are exemplar-based. When an individual has representations which are sufficiently abstract, they may have only minor difficulties processing complex sentences with atypical discourse properties. However, an individual whose representations are more exemplar-based will be strongly affected when discourse properties are atypical. It is therefore important, in controlled studies of syntactic processing, to create sentences which are naturalistic from a discourse perspective. However, studies of complex syntax in DLD employ stimuli which, in discourse terms, are highly atypical. One example is object relatives containing discourse-new NPs, which are frequent in studies of DLD (e.g. Novogrodsky & Friedmann, 2006). Likewise, many studies of passive comprehension which present passive sentences in isolation (e.g. Van der Lely, 1996) do not present them in a context where they have a topicalisation function. Such stimuli, though lacking in ecological validity, are chosen for practical reasons. Presenting these constructions with pronouns would involve the creation of a rich context, which is difficult within a controlled experimental setting. Moreover, the inclusion of pronouns would involve the introduction of case-marking cues, which add an additional and potentially confounding variable to already complex experimental designs. Consequently, many controlled studies of complex syntax in DLD assess an individual’s ability to adapt their syntactic representations to an atypical discourse context rather than their ability to process complex sentences in discourse contexts where they would naturally occur.

While it is difficult to run a controlled experiment on complex syntax with naturalistic stimuli, it is possible to create or facilitate a more naturalistic context in studies which elicit spoken language within a conversational setting. According to the logic of the above argument, children with DLD will demonstrate better abilities with complex sentences in such naturalistic contexts. Admittedly some naturalistic studies do suggest severe difficulties with complex sentences. For example, Marinellie (2004) sampled the language of 15 children with DLD in a conversational context, and found that rates of complex sentences, for example relative and adverbial clauses, were approximately half those of language-matched controls. Paradis et al. (2022), also using a conversational task, found a similar pattern for bilingual children with DLD.

However, open-ended conversational tasks may not be ideal for eliciting complex syntax, as sentence complexity may be influenced by factors unrelated to structural language, for example pragmatic abilities, conversational style, motivation, and rapport. Expository discourse tasks, where participants must create a monologue describing an inherently complex topic, provide a better means of eliciting syntactically-complex utterances. Using this paradigm, Nippold et al. (2008) identified very similar rates of complex sentences in adolescents with DLD (n = 106, mean age 13;11) compared to age-matched language-typical peers (n = 252). 26.5% of their sentences contained adverbial clauses compared to 27.6% in the control group. Regarding relative clauses, differences were again, small (7.5% vs. 8.9%). Group differences were non-significant. These findings suggest that difficulties with complex sentences in DLD are less evident when an appropriate discourse context is created.

Some studies of passive production in DLD have also employed naturalistic elicitation paradigms. These elicit passives via a prompt which topicalises the patient, for example what happened to X? Findings are quite mixed, but again point to the possibility that supposed difficulties with complex sentences in DLD may reflect the atypical nature of the stimuli used in more controlled studies. Both Leonard et al. (2006) and Stanford and Delage (2023) elicited passives via topicalisation prompts. Neither study identified an overall effect of group (DLD vs. age- and language-matched controls) when the dependent variable comprised long and short passives. It should, however, be noted that Leonard et al. (2006) found reduced production of long passives in DLD, and in Stanford and Delage (2023), there was a pronounced effect of prompt type, with the DLD group producing fewer passives after a linguistic prompt rather than a visual prompt which highlighted the patient. Overall, despite the mixed findings, these studies suggest that difficulties producing the passive are relatively minor when an appropriate discourse context is constructed.

More serious difficulties with the passive were identified by Du et al. (2024) in a study of the bei passive in Mandarin. Like the above studies, this study carefully manipulated the discourse context, employing prompts which topicalised the passive subject. The authors found substantially lower rates of passive use in children with DLD than both age- and language-matched controls. Nonetheless, it is debatable whether the bei passive has identical discourse properties to the English be passive. Some researchers have argued that it signals adversativity, that is an event with negative implications for the subject, and this factor outweighs its use as a topicalisation structure (Liu, 2011). If this is the case, the key question is whether the stimuli employed in the study involved the relevant degree of adversity.

To summarise, while complex sentences have salient discourse characteristics, experimental studies of complex syntax in DLD tend to employ stimuli which lack appropriate discourse contexts, and may even have atypical discourse characteristics. Consequently, assuming that the discourse properties of complex sentences are encoded in their grammatical representations, such studies assess children’s flexibility in adapting their exemplars to atypical discourse contexts. By contrast, numerous studies which create or facilitate discourse contexts which support the use of complex sentences (Leonard et al., 2006; Nippold et al., 2008; Stanford & Delage, 2023) find only minor difficulties in DLD. Such a pattern can be easily explained via an exemplar-based account which argues that fine-grained contextual information, including discourse properties, is stored within grammatical representations.

The Use of Chunks During Language Processing

Some researchers have suggested that children with DLD have a dependence on high-frequency chunks during language production and comprehension. Such multi-word units are exemplar-like as they lack the abstraction which characterises adult language use. Bybee (2013) has characterised chunk-dependence as a consequence of exemplar-based learning. Moreover, an overdependence on input frequency is consistent with Chang’s (2002) linked-path model which predicts input-dependence regarding word order, though it should be noted that, in this model, such effects arose at the message level, for example making links between the lemma of dog and the AGENT role, rather than reflecting the storage of multi-word units.

Some studies of DLD have presented novel items in specific syntactic frames, and observed an unwillingness to ‘go beyond the input’ by using that item in an unattested item-frame combination. Skipp et al. (2002) modelled novel verbs in a variety of frames, for example V, SV, VO, SVO. They found that children with DLD (n = 28, mean age 6;0) rarely spontaneously produced the novel verb in a different frame to the attested frame, though the degree of input-dependence was similar to language-matched controls. Riches et al. (2006) found that children with DLD who heard novel verbs in frames containing pronouns, for example he’s meeking it, seemed to acquire a lexically-specific VERB + it schema. This was evidenced in a task designed to elicit transitive sentences with novel verbs. The children with DLD were more likely to replicate the pronominal frame used during the training stage.

A further exploration of chunk dependence was conducted by Kueser and Leonard (2020). They performed an elicited imitation task using sentences which controlled the frequency of chunks, for example corn on the cob versus corn on the tray. The study also modelled the predictability of the final word in the chunk. The effect of chunk frequency in the DLD group and language matched controls was stronger than in the age-matched controls. When chunk frequency was reduced, accuracy in the DLD group (percentage of words correctly recalled) fell by 25.5% (71.3–45.8). This contrasted with a fall of only 15.3% in the language-matched group (98.0–82.7), and 19.8% in the age-matched group (54.2–34.4).²

This finding would suggest heightened chunk frequency in the DLD group. Unfortunately, for the purposes of the current discussion, the key interaction between group (DLD, age-matched and language-matched) and condition (presence vs. absence of a high-frequency chunk) did not reach significance (p = .099). This led the authors to conclude that there were no between-group differences in chunk-dependency. However, the study reported only the interaction between Group (3 levels) and Frequency (2 levels). This is an omnibus test as it simultaneously models all possible sources of variation, including those related to group comparisons which are of minimal theoretical interest, for example the age-matched controls versus the language matched controls. Such an approach is ‘unnecessarily conservative’ according to the American Psychological Association Board of Scientific Affairs (Wilkinson, 1999), a claim which is supported by statistical simulations (Schad et al., 2020). Given the conservative nature of the analysis, and the substantially different group profiles there is, I would argue, a case for further studies investigating this phenomenon.

Caveats

While there is, I believe, converging evidence for an exemplar-based account of DLD, there are some potential difficulties with this account.

First, the definition of what constitutes an exemplar, and the range of properties which may be used to characterise exemplars, could be more firmly operationalised. The key criterion is that exemplars exhibit constraints, or principles of generalisation which are ‘noisy’ in the sense of not existing within the adult system. Examples are semantic constraints on affixation in regular verbs, strong discourse constraints on the production and comprehension of complex sentences, and the possible existence of a chunk-based learning style. Further empirical and theoretical work is clearly needed to facilitate identification of exemplars.

Another issue is whether an exemplar-based learning style can explain co-occurring difficulties, for example motor difficulties, or dyslexia. Though the theory focuses primarily on language difficulties, it is possible that an exemplar-based learning style may be found across domains. For example, researchers have argued that performance on a visual pattern-learning task by autistic individuals may indicate atypical domain-general learning mechanisms (Gastgeb et al., 2012). According to this logic, the hypothesised exemplar-based learning style in DLD may have ramifications beyond language. However, a detailed discussion of how this might manifest is beyond the scope of this paper.

A further limitation of this account is that it does not readily explain inconsistent language use, for example the within-child co-occurrence of inflected and infinitival forms of the same verb. Such inconsistency would not be predicted by an exemplar account. One possibility is that inconsistent performance may reflect the application of a metalinguistic strategy. A child who is dependent on a product-oriented schema may become aware that most verbs in the past tense need to be inflected either via suffixation or a stem change. They may therefore develop a metalinguistic strategy which adds the suffix whenever their product-oriented schema results in no change. Because of the cognitive effort involved in such a process, they may only apply this intermittently. Though this goes beyond the exemplar account by proposing an alternative mechanism, exemplar-based processes remain a key underlying factor.

Many readers will question whether the study of artificial neural networks can serve as an empirical basis for making claims regarding language impairments. The emergence of powerful large language models (LLMs) has generated both excitement and controversy in the field of linguistics. Some have proclaimed that the incredible power of these models and the extremely naturalistic language they produce effectively rings the death knell for theories of Universal Grammar (Piantadosi, 2023). Others have argued that these models are nothing more than technically-sophisticated versions of the horse Clever Hans, and that they produce and comprehend language in a very superficial manner (Chomsky et al., 2023). For example, these models lack a genuine understanding of intentionality, and their semantic representations are, in effect, floating signifiers which are not linked to real-world objects or situations. Moreover, their training data greatly exceed the children’s linguistic input, making them poor models of acquisition.

While many of the criticisms of LLMs as models of human learning and cognition are valid, there are a number of reasons why we might wish to adopt them as psychologically plausible models. First, they mimic neural architectures and learning mechanisms. In terms of architectures, they involve neurons with multiple inputs, corresponding to dendrites, and a single output, corresponding to axons, and these neurons are arranged in layers as they are in the human brain. They also implement prediction-based learning, a process which many also assume to underly human learning (Friston, 2010). Finally, their linguistic behaviours closely resemble those of humans. For example, Cai et al. (2024) ran 12 classic psycholinguistic studies using ChatGPT and identified human-like performance in 10 of these. Paradigms included lexical and structural priming, and implicit causality (Gary scared/feared Anna because. . .).

A further issue is whether overfitting can genuinely happen in the human brain. Though this phenomenon occurs naturally in artificial neural networks, it is relatively easy to overcome. Therefore, is it plausible that this process also happens in the human brain, which is vastly more complex than the most complex artificial neural network? Two possible candidate mechanisms which are specific to the human brain are myelination and precision. Myelination, a neurological process which prunes weak connections and thereby fortifies strong connections plays an important role in brain development. In the context of overfitting, it reduces noise within neural networks. With regard to myelination, Chang’s (2002) linked-path model could be described as ‘under-myelinated’ as it exhibits a surfeit of connections due to the addition of paths between the message-level and sequential system. There is, moreover, evidence for reduced myelination (i.e. over-connectivity) in DLD (Krishnan et al., 2022), which could result in more noisy networks.

Precision is, in simple terms, an estimate of the learning potential of any unpredicted stimulus (Parr & Friston, 2017). Recent research has identified neurons in the auditory cortex which spike when an unpredicted stimulus is encountered, and consequently, it is likely that they encode estimates of precision (Pérez-González et al., 2024). Furthermore, recent data gathered by Blake et al. (2025) indicate that children with DLD find it difficult to identify collocations when presented with pairs consisting of collocations and a non-collocational distractor (hat and scarf vs. scarf and hat, and hard worker vs. hard swimmer). This suggests a diminished sensitivity to the input frequency of word combinations. Such a weakness would most likely impact on precision-based neural mechanisms which, by definition, model input frequencies. The responsiveness of precision-related neurons appears to be influenced by acetylcholine (Pérez-González et al., 2024) raising the possibility that DLD may be linked to low levels of this neuromodulator.

Possible Future Directions

As mentioned above, definitions of exemplars and descriptions of the range of characteristics that they display are relatively underspecified, and further work is needed in this area. Further behavioural studies could be conducted to explore the putative exemplar-related effects described above, for example the relationship between verb telicity and successful regular past tense affixation in DLD, the extent to which the production of complex sentences by children with DLD is influenced by discourse phenomena, and the degree to which children with DLD exhibit a dependence on high-frequency chunks. Neurobehavioural/neuroimaging studies may also be conducted in an attempt to identify neural signatures of exemplar-based learning. The kind of visual pattern-learning tasks which have been used with autistic children to explore exemplar-based process (e.g. Gastgeb et al., 2012) could also be administered to children with DLD to ascertain, first, whether they exhibit poorer generalisation, and second whether performance on this task correlates with linguistic measures.

A particularly exciting possibility is that exemplar-based theory will be applied in intervention. I have already discussed an intervention for the past tense which is closely informed by exemplar theory (Owen Van Horne et al., 2017). This employs what could be described as an ‘exemplar-busting’ approach, by initially introducing stimuli which are designed to deliberately disrupt exemplar-based representations. A similar technique could be applied in various grammatical domains. As mentioned above, when complex sentences are introduced into a clinical context, they typically contain full NPs and consequently are already potential ‘exemplar busters.’ What may be lacking is the application of a learning regime which manipulates item difficulty. Owen Van Horne et al. (2017) specifically found that a difficult-to-easy gradient was most effective. In the context of complex sentences, this would involve introducing more naturalistic complex sentences, containing pronouns rather than full NPs, at a later stage.

The Tough-But-Fragile Paradox Revisited

The evidence for an exemplar-based account of DLD can best be described as emerging. Nonetheless, a key attraction of this approach is that it provides a solution to the tough-but-fragile paradox, whereby the language system recovers well in cases of early neurological damage, but shows poor prognosis in children with DLD. This is a key paradox which all accounts of DLD must explain. The idea that powerful neural systems can be severely affected by overfitting has been consistently demonstrated in artificial neural networks, raising the possibility that this may also happen in the human brain. It remains to be seen whether existing accounts of language difficulties in DLD can provide such a persuasive explanation. One possibility is that subtle weaknesses in specific cognitive systems, for example phonological working memory, or auditory processing, may, via a series of cascading effects, lead to severe long-term consequences. But given what we know about neural plasticity, and the impressive language development of children with early focal lesions, these accounts are questionable. By contrast, an overfitting account offers an attractive alternative which is theoretically well-formulated, empirically grounded, and has important implications for intervention.

Supplemental Material

sj-docx-1-fla-10.1177_01427237251371762 – Supplemental material for Developmental Language Disorder: A Consequence of Exemplar-Based Learning?

Supplemental material, sj-docx-1-fla-10.1177_01427237251371762 for Developmental Language Disorder: A Consequence of Exemplar-Based Learning? by Nick Riches in First Language

Footnotes

ORCID iD

Nick Riches

Author Contributions

Nick Riches: Conceptualisation; Investigation; Writing—original draft; Writing—review & editing.

Funding

The author received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Supplemental Material

Supplemental material for this article is available online.

Notes

References

Albright

Hayes

(2003). Rules vs. analogy in English past tenses: A computational/experimental study. Cognition, 90(2), 119–161.

Ambridge

(2020a). Against stored abstractions: A radical exemplar model of language acquisition. First Language, 40(5–6), 509–559. https://doi.org/10.1177/0142723719869731

Ambridge

(2020b). Abstractions made of exemplars or “You’re all right, and I’ve changed my mind”: Response to commentators. First Language, 40(5–6), 640–659. https://doi.org/10.1177/0142723720949723

Ambridge

Bidgood

Pine

J. M.

Rowland

C. F.

Freudenthal

(2016). Is passive syntax semantically constrained? Evidence from adult grammaticality judgment and comprehension studies. Cognitive Science, 40(6), 1435–1459. https://doi.org/10.1111/cogs.12277

Bates

Reilly

Wulfeck

Dronkers

Opie

Fenson

Kriz

Jeffries

Miller

Herbst

(2001). Differential effects of unilateral lesions on language production in children and adults. Brain and Language, 79(2), 223–265.

Belkin

Hsu

(2020). Two models of double descent for weak features. SIAM Journal on Mathematics of Data Science, 2(4), 1167–1180. https://doi.org/10.1137/20M1336072

Bishop

D. V. M.

(1994). Is specific language impairment a valid diagnostic category? Genetic and psycholinguistic evidence. Philosophical Transactions of the Royal Society of London: B. Biological Sciences, 346, 106–111.

Bishop

D. V. M.

(2005). Test of reception of grammar—Electronic. Harcourt Assessment.

Bishop

D. V. M.

(2020, February 23). BishopBlog: Changing terminology for children’s language disorders: Reflections on special issue of perspectives of the ASHA Special Interest Groups. BishopBlog. https://deevybee.blogspot.com/2020/02/changing-terminology-for-childrens.html

10.

Bishop

D. V. M.

Snowling

M. J.

Thompson

P. A.

Greenhalgh

, & CATALISE consortium. (2016). CATALISE: A Multinational and Multidisciplinary Delphi Consensus Study. Identifying language impairments in children. PLoS One, 11(7), e0158753. https://doi.org/10.1371/journal.pone.0158753

11.

Bishop

D. V. M.

Snowling

M. J.

Thompson

P. A.

Greenhalgh

., & the CATALISE-2 consortium. (2017). Phase 2 of CATALISE: A multinational and multidisciplinary Delphi consensus study of problems with language development: Terminology. Journal of Child Psychology and Psychiatry, 58(10), 1068–1080. https://doi.org/10.1111/jcpp.12721

12.

Blake

Wright

Dąbrowska

Riches

Winckel

. (2025). Non-verbal cognition as a predictor of linguistic outcomes in children with and without developmental language disorder [Manuscript submitted for publication].

13.

Botting

(2020). Language, literacy and cognitive skills of young adults with developmental language disorder (DLD). International Journal of Language & Communication Disorders, 55(2), 255–265. https://doi.org/10.1111/1460-6984.12518

14.

Bybee

J. L.

(2013). Usage-based theory and exemplar representations of constructions. The Oxford Handbook of Construction Grammar. Oxford University Press. https://doi.org/10.1093/oxfordhb/9780195396683.013.0004

15.

Bybee

J. L.

Slobin

D. I.

(1982). Rules and schemas in the development and use of the English past tense. Language, 58(2), 265–289.

16.

Cai

Duan

Haslett

Wang

Pickering

(2024). Do large language models resemble humans in language use? Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics, Bangkok, Thailand 37–56. https://doi.org/10.18653/v1/2024.cmcl-1.4

17.

Calder

S. D.

Claessen

Leitão

Ebbels

(2022). A profile of expressive inflectional morphology in early school-age children with developmental language disorder. Clinical Linguistics & Phonetics, 36(4–5), 341–358. https://doi.org/10.1080/02699206.2021.1931454

18.

Chang

(2002). Symbolically speaking: A connectionist model of sentence production. Cognitive Science, 26(5), 609–651.

19.

Chomsky

Roberts

Watumull

(2023, March 20). Noam Chomsky: The false promise of ChatGPT. The New York Times.

20.

Contemori

Belletti

(2014). Relatives and passive object relatives in Italian-speaking children and adults: Intervention in production and comprehension. Applied Psycholinguistics, 35, 1021–1053.

21.

Conti-Ramsden

Botting

Faragher

(2001). Psycholinguistic markers for specific language impairment (SLI). The Journal of Child Psychology and Psychiatry and Allied Disciplines, 42(6), 741–748.

22.

Cruys

S. V.

de, de-Wit

Evers

Boets

Wagemans

(2013). Weak priors versus overfitting of predictions in autism: Reply to Pellicano and Burr (TICS, 2012). I-Perception, 4(2), 95. https://doi.org/10.1068/i0580ic

23.

Deevy

Leonard

L. B.

(2004). The comprehension of wh-questions in children with specific language impairment. Journal of Speech Language and Hearing Research, 47(4), 802–815.

24.

Dell

G. S.

Chang

(2013). The P-chain: Relating sentence production and its disorders to comprehension and acquisition. Philosophical Transactions of the Royal Society B: Biological Sciences, 369(1634). https://doi.org/10.1098/rstb.2012.0394

25.

Demuth

Johnson

(2020). Exemplar-based learning probably requires learning abstractions: A commentary on Ambridge (2020). First Language, 40(5–6), 573–575. https://doi.org/10.1177/0142723720903609

26.

Dromi

(1987). Early lexical development. Cambridge University Press.

27.

(2024). The production of passives in Mandarin children with and without developmental language disorder. First Language, 1–13. https://doi.org/10.1177/01427237241298796

28.

Ferreira

(2021). In defense of the passive voice. American Psychologist, 76(1), 145–153. https://doi.org/10.1037/amp0000620

29.

Fox

B. A.

Thompson

S. A.

(1990). A discourse explanation of the grammar of relative clauses in English conversation. Language, 297–316.

30.

Friston

(2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

31.

Gastgeb

H. Z.

Dundas

E. M.

Minshew

N. J.

Strauss

M. S.

(2012). Category formation in autism: Can individuals with autism form categories and prototypes of dot patterns? Journal of Autism and Developmental Disorders, 42(8), 1694–1704. https://doi.org/10.1007/s10803-011-1411-x

32.

Goldberg

A. E.

(2014). The information structure of ditransitives: Informing scope properties and long-distance dependency constraints. In S. K. Bourns & L. L. Myers (Eds.), Perspectives on linguistic structure and context (pp. 3–16). John Benjamins. https://www.jbe-platform.com/content/books/9789027270597-pbns.244.01gol

33.

Gordon

P. C.

Hendrick

Johnson

(2004). Effects of noun phrase type on sentence complexity. Journal of Memory and Language, 51(1), 97–114.

34.

Haendler

Kliegl

Adani

(2015). Discourse accessibility constraints in children’s processing of object relative clauses. Frontiers in Psychology, 6, 1–17. https://doi.org/10.3389/fpsyg.2015.00860

35.

Hay

Bresnan

(2006). Spoken syntax: The phonetics of giving a hand in New Zealand English. The Linguistic Review, 23(3), 321–349. https://doi.org/10.1515/TLR.2006.013

36.

Hooper

J. B.

(1976). Word frequency in lexical diffusion and the source of morphophonological change. In W. M. Christie (ed.), Current progress in historical linguistics (pp. 95 –105). North Holland.

37.

Hundt

Dallas

Nakanishi

(2024). The be- versus get-passive alternation in world Englishes. World Englishes, 43(1), 86–108. https://doi.org/10.1111/weng.12633

38.

Idei

Murata

Yamashita

Ogata

(2020). Homogeneous intrinsic neuronal excitability induces overfitting to sensory noise: A robot model of neurodevelopmental disorder. Frontiers in Psychiatry, 11, 762. https://doi.org/10.3389/fpsyt.2020.00762

39.

Krishnan

Cler

G. J.

Smith

H. J.

Willis

H. E.

Asaridou

S. S.

Healy

M. P.

Papp

Watkins

K. E.

(2022). Quantitative MRI reveals differences in striatal myelin in children with DLD. eLife, 11, e74242. https://doi.org/10.7554/eLife.74242

40.

Kueser

J. B.

Leonard

L. B.

(2020). The effects of frequency and predictability on repetition in children with developmental language disorder. Journal of Speech, Language, and Hearing Research, 63(4), 1165–1180. https://doi.org/10.1044/2019_JSLHR-19-00155

41.

Leonard

L. B.

(2014). Children with specific language impairment. MIT Press.

42.

Leonard

L. B.

Deevy

Kurtz

Krantz Chorev

Owen

Polite

Elam

Finneran

(2007). Lexical aspect and the use of verb morphology by children with specific language impairment. Journal of Speech, Language, and Hearing Research, 50(3), 759.

43.

Leonard

L. B.

Wong

A. M. Y.

Deevy

Stokes

S. F.

Fletcher

(2006). The production of passives by children with specific language impairment: Acquiring English or Cantonese. Applied Psycholinguistics, 27, 267–299.

44.

Lieven

Ferry

Theakston

Twomey

K. E.

(2020). Similarity, analogy and development in radical exemplar theory: A commentary on Ambridge (2020). First Language, 40(5–6), 600–603. https://doi.org/10.1177/0142723720909554

45.

Liu

(2011). The bei passive and its discourse motivations. Chinese Language and Discourse, 2(2), 198–231. https://doi.org/10.1075/cld.2.2.03liu

46.

Marchman

Wulfeck

Ellis Weismer

(1999). Morphological productivity in children with normal language and SLI: A study of the English past tense. Journal of Speech Language and Hearing Research, 42, 206–219.

47.

Marinellie

S. A.

(2004). Complex syntax used by school-age children with specific language impairment (SLI) in child–adult conversation. Journal of Communication Disorders, 37(6), 517–533. https://doi.org/10.1016/j.jcomdis.2004.03.005

48.

Marshall

C. R.

van der Lely

H. K. J.

(2006). A challenge to current models of past tense inflection: The impact of phonotactics. Cognition, 100(2), 302–320.

49.

Newport

E. L.

Seydell-Greenwald

Landau

Turkeltaub

P. E.

Chambers

C. E.

Martin

K. C.

Rennert

Giannetti

Dromerick

A. W.

Ichord

R. N.

Carpenter

J. L.

Berl

M. M.

Gaillard

W. D.

(2022). Language and developmental plasticity after perinatal stroke. Proceedings of the National Academy of Sciences, 119(42), e2207293119. https://doi.org/10.1073/pnas.2207293119

50.

Nippold

M. A.

Mansfield

T. C.

Billow

J. L.

Tomblin

J. B.

(2008). Expository discourse in adolescents with language impairments: Examining syntactic development. American Journal of Speech-Language Pathology, 17(4), 356.

51.

Novogrodsky

Friedmann

(2006). The production of relative clauses in syntactic SLI: A window to the nature of the impairment. International Journal of Speech-Language Pathology, 8(4), 364–375. https://doi.org/10.1080/14417040600919496

52.

Nudel

Chrsitensen

R. V.

Kalnak

Lundberg

Schwinn

Sørensen

Mikkelsen

Nissen

Christoffersen

L. A. N.

Kjerulff

B. D.

Hansen

T. F.

Burgdorf

K. S.

Banasik

Bay

Kjærgaard Boldsen

Brodersen

Brunak

Demur

A. B.

Nordahl Christoffersen

L. A.

, . . . Werge

(2024). Developmental language disorder—heritability and genetic correlations with other disorders affecting language. Psychiatry Research, 342, 116212. https://doi.org/10.1016/j.psychres.2024.116212

53.

Owen

A. J.

(2010). Factors affecting accuracy of past tense production in children with specific language impairment and their typically developing peers: The influence of verb transitivity, clause location, and sentence type. Journal of Speech, Language, and Hearing Research, 53(4), 993–1014. https://doi.org/10.1044/1092-4388(2009/09-0039)

54.

Owen Van Horne

A. J.

Fey

Curran

. (2017). Do the hard things first: A randomized controlled trial testing the effects of exemplar selection on generalization following therapy for grammatical morphology. Journal of Speech Language and Hearing Research, 60(9), 2569. https://doi.org/10.1044/2017_JSLHR-L-17-0001

55.

Owen Van Horne

A. J.

Green Fager

. (2015). Quantifying the relative contributions of lexical and phonological factors to regular past tense accuracy. International Journal of Speech-Language Pathology, 17(6), 605–616. https://doi.org/10.3109/17549507.2015.1034174

56.

Paradis

Sorenson Duncan

Thomlinson

Rusk

(2022). Does the use of complex sentences differentiate between bilinguals with and without DLD? Evidence from conversation and narrative tasks. Frontiers in Education, 6, 1–18. https://doi.org/10.3389/feduc.2021.804088

57.

Parr

Friston

K. J.

(2017). Uncertainty, epistemics and active inference. Journal of the Royal Society Interface, 14(136), 20170376. https://doi.org/10.1098/rsif.2017.0376

58.

Pellicano

(2010). Individual differences in executive function and central coherence predict developmental changes in theory of mind in autism. Developmental Psychology, 46(2), 530.

59.

Pérez-González

Lao-Rodríguez

A. B.

Aedo-Sánchez

Malmierca

M. S.

(2024). Acetylcholine modulates the precision of prediction error in the auditory cortex. eLife, 12, RP91475. https://doi.org/10.7554/eLife.91475

60.

Piantadosi

S. T.

(2024). Modern language models refute Chomsky’s approach to language. In Gibson

Poliak

(Eds.), From fieldwork to linguistic theory: A tribute to Dan Everett (pp. 353–414). Language Science Press.

61.

Pierrehumbert

J. B.

(2001). Exemplar dynamics: Word frequency, lenition and contrast. In Bybee

J. L.

Hopper

P. J.

(Eds.), Typological studies in language (Vol. 45, p. 137). John Benjamins Publishing Company. https://doi.org/10.1075/tsl.45.08pie

62.

Pinker

(1989). Learnability and cognition. MIT Press.

63.

Rice

M. L.

Tomblin

J. B.

Hoffman

Richman

W. A.

Marquis

(2004). Grammatical tense deficits in children with SLI and nonspecific language impairment. Journal of Speech Language and Hearing Research, 47(4), 816.

64.

Rice

M. L.

Wexler

Hershberger

(1998). Tense over time: The longitudinal course of tense acquisition in children with specific language impairment. Journal of Speech, Language, and Hearing Research, 41(6), 1412–1431.

65.

Riches

N. G.

(2012). Sentence repetition in children with specific language impairment: An investigation of underlying mechanisms. International Journal of Language & Communication Disorders, 47(5), 499–510. https://doi.org/10.1111/j.1460-6984.2012.00158.x

66.

Riches

N. G.

Faragher

Conti-Ramsden

(2006). Verb schema use and input dependence in 5-year-old children with specific language impairment (SLI). International Journal of Language and Communication Disorders, 41(2), 117–135. https://doi.org/DOI:10.1080/13682820500216501

67.

Riches

N. G.

Garraffa

(2017). A discourse account of intervention phenomena: An investigation of interrogatives. Glossa: A Journal of General Linguistics, 2(1), 74. https://doi.org/10.5334/gjgl.100

68.

Rispens

J. E.

De Bree

E. H.

(2014). Past tense productivity in Dutch children with and without SLI: The role of morphophonology and frequency. Journal of Child Language, 41(1), 200–225. https://doi.org/10.1017/S0305000912000542

69.

Ross

B. H.

Makin

V. S.

(1999). Prototype versus exemplar models in cognition. In R. J. Sternberg (Ed.), The nature of cognition (pp. 205–241). MIT Press.

70.

Schad

D. J.

Vasishth

Hohenstein

Kliegl

(2020). How to capitalize on a priori contrasts in linear (mixed) models: A tutorial. Journal of Memory and Language, 110, 104038. https://doi.org/10.1016/j.jml.2019.104038

71.

Schuler

K. D.

Kodner

Caplan

(2020). Abstractions are good for brains and machines: A commentary on Ambridge (2020). First Language, 40(5–6), 631–635. https://doi.org/10.1177/0142723720906233

72.

Semel

Wiig

E. H.

Secord

(2006). Clinical evaluation of language fundamentals—Fourth edition UK (CELF-4UK). Harcourt Assessment.

73.

Semel

Wiig

E. H.

Secord

(2017). Clinical evaluation of language fundamentals—Fifth edition UK (CELF-5UK). Pearson Assessment.

74.

Skipp

Windfuhr

Conti-Ramsden

(2002). Children’s grammatical categories of verb and noun: A comparative look at children with specific language impairment (SLI) and normal language (NL). International Journal of Language and Communication Disorders, 37(3), 253–271.

75.

Stanford

Delage

(2023). The contribution of visual and linguistic cues to the production of passives in ADHD and DLD: Evidence from thematic priming. Clinical Linguistics & Phonetics, 37(1), 17–51. https://doi.org/10.1080/02699206.2021.2006789

76.

Street

J. A.

Dąbrowska

(2014). Lexically specific knowledge and individual differences in adult native speakers’ processing of the English passive. Applied Psycholinguistics, 35(1), 97–118. https://doi.org/10.1017/S0142716412000367

77.

Thompson

Ling

S. P.

Myachykov

Ferreira

Scheepers

(2013). Patient-related constraints on get- and be-passive uses in English: Evidence from paraphrasing. Frontiers in Psychology, 4, 1–24. https://doi.org/10.3389/fpsyg.2013.00848

78.

Van der Lely

H. K. J

. (1996). Specifically language impaired and normally developing children: Verbal passive vs. adjectival passive sentence interpretation. Lingua, 98, 243–272.

79.

Van der Lely

H. K. J.

Ullman

. (2001). Past tense morphology in specifically language impaired and normally developing children. Language and Cognitive Processes, 16(2–3), 177–217. https://doi.org/10.1080/01690960042000076

80.

Wilkinson

(1999). Statistical methods in psychology journals: Guidelines and explanations. American Psychologist, 54(8), 594.

81.

Zager

(1982). A real-time process of morphological change. SUNY Buffalo. https://elibrary.ru/item.asp?id=7321500

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.89 MB