Exemplars,Generalisability and the Broader DLD Profile: A Commentary on Riches (2026)

Abstract

Riches’ article proposes that developmental language disorder is a consequence of exemplar-based learning. While there is a lot of merit to the proposal, there are also areas that need further clarity – primarily whether the difficulty lies in the exemplars, the generalisation/abstraction processes that operate over exemplars, or both.

Commentary

Riches (2026) puts forward two hypotheses. First, that language acquisition is characterised by exemplar-based learning. Second, that this exemplar-based learning is somehow faulty in children with developmental language disorder (DLD). Although his focus is on the second of these, since it is predicated on the first, we also need to consider what exemplar-based learning is.

The general idea of exemplar-based learning is that what children hear is stored as specific instances or events – not just what is spoken and how it is spoken (e.g. phonetic features), but the overall context of the communication exchange, such as speaker features, surrounding objects and so forth (e.g. a dog’s size and temperament, to use Riches’ example). As such, these exemplars contain information that is potentially useful (e.g. the dog’s name) and information that is potentially less useful or misleading (e.g. a car passed in the background while the spoken words unfolded). The congruence of previously stored exemplars is important since it pinpoints such things as what the likely past-tense is for a particular verb or what the function of a novel object may be. So things like categorisation occur on the fly – a situation involving a previously unseen dog generates an assortment of exemplar features that match with a multitude of exemplars that exist from prior knowledge for the child to conclude that this new animal is more than likely a dog. In essence, previous exemplars involving dogs are likely to have similar ‘dog’ features to them, while at the same time having other information that likely varies across exemplars (e.g. background objects and who the speaker was) and is therefore seen as superfluous. This is perhaps a simplistic description of the process, but it serves for considering how it might apply to DLD, while Riches cites excellent reading material such as Ambridge (2020) and Bybee (2013) should readers wish to explore exemplar-based learning further. I am very much in line with this approach, particularly as my work could be classed as an exemplar approach that focuses on the speech children hear from their caregivers (using computational models that form increasingly large chunks of sub-lexical and lexical knowledge as they progress through caregiver speech input, simulating the learning trajectories of the children).

We can now turn to the key part of the article – can exemplar-based learning account for the language profiles seen in DLD? Here, Riches gives two examples. First, tense marking. Children with DLD have difficulty in using the correct inflected form (e.g. using he go shops instead of he’s going to the shops). On the one hand, since some of these errors are influenced by properties of the stem, Riches suggests children with DLD may be operating a product-oriented schema, which presumably means the process that operates across exemplars is focused on the product rather than something else. On the other hand, it is perhaps the information within the exemplars that is the problem: Riches points to a study where performance was improved in children with DLD when training them on verbs that broadened the range of exemplars so that they were less entrenched (Owen Van Horne et al., 2017). Here, we get our first question: Does the difficulty lie in the exemplar system or the process that operates over exemplars to infer some form of meaning or production? I suppose one could be faulty in storing exemplars such that product-schemas were more likely to be invoked, so perhaps Riches’ idea is along these lines, but it is a little unclear. Interestingly, tense-marking errors also occur in children without DLD and have been explained by the process operating over exemplars. Since children hear forms of the resulting errors – he go or she laugh – in their natural language discourse (e.g. where did he go and why did she laugh), they can be straightforwardly generated by a system that allows substitutions of words or phrases based on shared distributional regularities (see Freudenthal et al., 2015). In a nutshell, then, where does the problem lie – in the exemplars, in the abstraction process, or both?

The second example is complex syntax. Details on the difficulties relating to complex syntax in children with DLD can be found in the paper itself. The take-home message is that Riches suggests the difficulty in complex syntax seen in children with DLD may be because these children are exemplar-based, whereas children without DLD are more abstract-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’, p. 16). So, a further question is, are we talking an exemplar-based system that generates abstraction on the fly (as per Ambridge, 2020) or one where abstraction is inherent in the creation of exemplars?

Finally, let us consider aspects of DLD that are not covered within Riches’ account. I will highlight just two – smaller vocabularies (Ansari et al., 2025) and greater difficulty repeating nonsense words as syllabic length increases (Marton & Schwartz, 2003) – since both aspects suggest that sub-lexical knowledge is either lacking in these individuals or something causes difficulty in properly accessing or interpreting the information. Certainly, what seems to be the case in nonsense word repetition is that children with DLD more easily repeat nonsense words when they share characteristics with existing words than when they do not (Jones et al., 2010). To me, this seems to point more towards a problem in the process that operates over existing exemplars than in the exemplars themselves. This is because existing knowledge clearly helps, suggesting that there are appropriate exemplars that can be used. But perhaps there is some difficulty in combining this information, particularly when it is not well embedded as exemplars (as is the case for nonsense words that share little information with existing words). Again though, we have the alternative possibility that it is the exemplars themselves – perhaps not being stored, not stored in a way that easily permits their access, or some other reason – because as word (or nonsense word) length increases, so does the difficulty in forming coherent representations. For example, if there was difficulty in forming exemplars, this would be less impactful when words are short than when they are long.

As a summary, then, the exemplar-based view of children with DLD as set out by Riches (2026) seems to hold some promise, but the devil is in the details. Of course, there are alternative accounts as to what may be the underlying difficulties for these children, although there is no cover-all account (see Marshall, 2024). Properly assessing different explanations may well require different accounts to be implemented as computational instantiations and testing them against the developmental profiles of DLD. Although not straightforward – it requires faithful interpretations of each account – it would allow us to discover those explanations that hold the most promise. Whether exemplars are the answer, time will tell.