An exemplar model should be able to explain all syntactic priming phenomena: A commentary on Ambridge (2020)

Abstract

The authors argue that Ambridge’s radical exemplar account of language cannot clearly explain all syntactic priming evidence, such as inverse preference effects (greater priming for less frequent structures), and the contrast between short-lived lexical boost and long-lived abstract priming. Moreover, without recourse to a level of abstract syntactic structure, Ambridge’s account cannot explain abstract priming in amnesia patients or cross-linguistic priming. Instead, the authors argue that abstract representations remain the more parsimonious account for the wide variety of syntactic priming phenomena.

Though there is currently no consensus over the exact mechanisms underlying syntactic priming – the phenomenon whereby people’s sentence processing is influenced by previously experienced syntactic structures – these effects have historically been viewed as key evidence that speakers store abstract representations of syntactic structure. This conclusion stems from evidence that priming occurs between language comprehension and production (Bock et al., 2007; Branigan et al., 1995) and in the absence of lexical or thematic overlap (Bock, 1989; Messenger et al., 2012), suggesting a shared store of representations based on abstract syntax. Notably, sentences with similar surface features but different underlying syntax (e.g. prepositional datives vs infinitives: Susan brought a book to Stella/Susan brought a book to study) do not show priming (Bock & Loebell, 1990). Therefore, whether the purported mechanism of priming is residual activation of stored representations (Branigan & Pickering, 2017; Malhotra et al., 2008), error-based learning from syntactic prediction (Chang et al., 2006) or a combination of the two (Reitter et al., 2011; Segaert et al., 2016), all existing models of syntactic priming posit a layer of abstract syntactic representation in language processing.

By contrast, Ambridge (2020) argues that a radical exemplar model of language can explain syntactic priming effects without appealing to abstract syntactic representations. Within his account, processing a prime sentence activates concrete exemplars of similar sentences stored in a person’s memory. Priming occurs because, when the speaker analogises over exemplars to form a subsequent sentence, it is the prime exemplars that are more readily available for retrieval, having been recently activated. This model of priming might account for some effects, such as the increase in priming in the presence of lexical overlap (i.e. lexical boost), however, we argue, it cannot easily explain the full range of syntactic priming phenomena, many of which implicate a role for abstract syntactic representations in language processing.

Priming effects are typically stronger for infrequently-experienced syntactic structures (Hartsuiker et al., 1999; Jaeger & Snider, 2013). Error-based learning accounts most easily explain this inverse frequency effect: processing a prime sentence with an unexpected (low frequency) syntactic structure leads to a larger prediction error signal, increasing the weighting of that representation and consequently its likelihood of re-use (Chang et al., 2006). For frequently-experienced structures, the smaller prediction error signal makes priming less likely to occur. The exemplar model makes the opposite prediction: higher frequency structures would generate a greater number of stored exemplars to analogise over, increasing the likelihood of priming. It is not clear how low frequency structures, for which a speaker would have fewer stored exemplars, would lead to greater priming in this model.

Accounts of syntactic priming must also be able to explain observed differences in the timespan of syntactic priming effects. Lexically-mediated priming (lexical boost) is typically large but short-lived, whereas ‘abstract’ priming (where the source of priming is lexically-unrelated syntactic forms) is typically smaller but long-lasting (Bock & Griffin, 2000; Branigan & McLean, 2016; Hartsuiker et al., 2008). The exemplar model provides a simple explanation for lexical boost effects, but does not address the different timespans of lexically-mediated and abstract priming. Whilst Ambridge discusses possible decay mechanisms for exemplar models, he does not apply this to the priming data. In fact, he rejects the notion of ‘abstraction-as-forgetting’ (p. 541) suggesting instead that exemplars are stored intact, though details may become inaccessible due to memory decay or interference. In particular he claims that it is implausible that speakers retain an abstract representation of sentences following decay, but this is precisely what the priming evidence indicates: priming results specifically in long-lasting effects on abstract syntactic forms. To explain these differences, both residual activation and error-based learning accounts of syntactic priming suggest different memory systems serve different aspects of processing (Chang et al., 2012; Malhotra et al., 2008). Short-term memory may support more immediate and explicit priming effects as in the case of lexical overlap, however abstract priming effects persist because they entail changes to the underlying representations stored in long-term memory.

Without recourse to a layer of abstract syntactic representations supporting language processing, an exemplar account cannot easily explain priming phenomena within particular groups of speakers. Patients with amnesia, such as Korsakoff’s and anterograde amnesia, who have such profound impairments in declarative memory that they cannot formulate new concrete memories (i.e. exemplars), nonetheless show abstract syntactic priming (Ferreira et al., 2008; Heyselaar et al., 2017). Since amnesia patients have preserved non-declarative memory, such findings are typically explained by an account in which syntactic priming effects are predominantly supported by non-declarative memory systems (Chang et al., 2012). By contrast, these findings are not well-explained by an exemplar model in which syntax generation is primarily influenced by the retrieval of exemplars, stored within declarative memory.

Perhaps the most striking evidence that speakers represent syntactic information in abstract form comes from evidence of cross-linguistic syntactic priming in bilingual adults and children (van Gompel & Arai, 2018). That is, processing a syntactic structure in one language can prime a bilingual speaker to use the same structure in their other language (Kantola & van Gompel, 2011; Vasilyeva et al., 2010). Notably, this effect persists with very little surface similarity between the two languages, such as between Korean and English (Shin & Christianson, 2012) and Chinese and English (Chen et al., 2013), or for structures that are grammatically possible in both languages, but rarely produced in one (such as prenominal adjectives in Spanish; Hsin et al., 2013). Without experience of sentences that share some surface similarity, these cross-linguistic syntactic priming effects are hard to explain in an exemplar model. Syntactic structure is the only aspect that is shared between cross-linguistic primes and targets, providing the strongest evidence that speakers store abstract representations of syntax.

In sum, the radical exemplar model of syntactic processing cannot clearly explain the wide range of observed priming effects. Of course, work is still needed to clarify the exact nature of the mechanisms and representations that support language processing and explain priming effects, but the exemplar model is no more immune to these issues than other existing models. Furthermore, it must address the range of findings presented here in order to provide a viable explanation of syntactic priming. Based on the evidence discussed, we suggest that to discount accounts in which speakers store abstract syntactic representations in long-term memory would be to discount a more parsimonious way of explaining human sentence processing.