The target article is part of a welcome recent trend to take psycholinguistic results as able to adjudicate among competing theoretical proposals, rather than being treated as simply presupposing linguists' constructs. We wholeheartedly agree with this stance; in fact, we endorse it to a much greater extent than the authors advocate: From our point of view, the paper presents a rather conservative interpretation of the cited results in that it persists with the preoccupation of abstracting over behavioural/neuronal data to underlying abstract knowledge representations presumed to underlie their explanatory mechanisms (Gregoromichelaki & Kempson, Reference Gregoromichelaki, Kempson, Scott, Clark and Carstonforthcoming; cf. Ferreira et al. Reference Ferreira, Bock, Wilson and Cohen2008).
We support the claim that methods of structural priming comparisons can be informative about mechanisms underlying linguistic processing. However, the authors argue that structural priming results are explainable only by assuming separate linguistic representations encoding semantic/syntactic/phonological information. Here, perhaps surprisingly, the authors seem to adopt the standard linguistic stance that theoretical frameworks/explanations need to presuppose an abstract, static view of linguistic knowledge, separating models of competence from accounts of performance.
In contrast, we propose an alternative formal architecture based on Dynamic Syntax (DS) as the syntactic engine (Cann et al. Reference Cann, Kempson and Marten2005; Kempson et al. Reference Kempson, Meyer-Viol and Gabbay2001; Reference Kempson, Cann, Gregoromichelaki and Chatzikyriakidis2017) enhanced with incremental construction of Type-Theory-with-Records (TTR; Cooper Reference Cooper, Kempson, Asher and Fernando2012) conceptual representations (DS-TTR; Gregoromichelaki et al. Reference Gregoromichelaki, Kempson, Howes, Eshghi, Wachsmuth, de Ruiter, Jaecks and Kopp2013; Hough Reference Hough2015; Kempson et al. Reference Kempson, Cann, Gregoromichelaki and Chatzikyriakidis2016; Purver et al. Reference Purver, Gregoromichelaki, Meyer-Viol, Cann and Purver2010). While eschewing a level of syntactic representation and any competence/performance distinction, such a framework is able to account directly for the priming data as well as standard linguistic generalisations.
Concentrating on syntax, the main focus of the paper, the data presented provide no evidence for theoretical or implementational perspectives on syntactic knowledge that would necessarily assume string-level hierarchical representations or accessing of stored well-formedness constraints in some kind of context-free-grammar format. Instead, a formal grammar adopting a DS-TTR-style architecture consisting of routinised sequences of processing actions with no syntactic representations is much more compatible with the overall data. From this perspective, the appearance of abstract structural pattern-matching is epiphenomenal on the incrementality/predictivity of the processing of time-linearly unfolding stimuli. In contrast, the methodology that involves abstracting a level of syntactic representation over the actions impedes straightforward analyses of patterns of interlocutor coordination in dialogue (Gregoromichelaki et al. Reference Gregoromichelaki, Kempson, Purver, Mills and Cann2011).
Instead of syntactic hierarchical structure, according to DS-TTR, a small set of elementary domain-general processing actions underpins both parsing and generation: Cross-linguistically available sequences of such actions cluster into higher-order sequential patterns (macros) that can be learned online, activated long-/short-term, and stored as chunks triggered by specific word forms (Eshghi et al. Reference Eshghi, Hough, Purver, Demberg and Levy2013). It is the reuse, and potential for adjustment, of such sequences that accounts for the authors' findings of “syntactic-pattern” repetitions appearing as distinct and/or independent from semantic interpretations. These results can be explained more explicitly in DS-TTR because the framework addresses the pervasive local ambiguity problem of incremental parsing/generation by predictive activation of various potential probabilistically weighted processing paths (sequences of actions). These processing paths are taken to constitute part of the context and some of them lead to identical TTR conceptual structures (Hough Reference Hough2015; Hough & Purver Reference Hough, Purver, Chatzikyriakidis and Luo2017; Sato Reference Sato, Gregoromichelaki and Howes2011). For example, PO/DO or active/passive alternations in DS-TTR reflect the invocation of distinct sequences of actions to construct or linearise equivalent conceptual event frames (with distinguishing information-structure aspects reflected in the processing order). The parser/generator initially pursues the highest-ranked option, with the rest maintained in the context for conversational-repair purposes (Eshghi et al. Reference Eshghi, Howes, Gregoromichelaki, Hough and Purver2015; Hough Reference Hough2015). Success of one such path in achieving the intended conceptualisation will enhance the probability of perception/execution of the same action sequence subsequently if the word forms accessed make it available, while inhibiting the pursuance of alternatives.
Cumulative priming effects are predicted with additional repetition of the triggering word forms (the lexical boost effect), because phonological forms are stored in context for conversational purposes like clarification. Facilitation of retrieval is predicted even when repetition of the same word forms in conjunction with the same word order leads to distinct conceptual frames (Bock & Loebell Reference Bock and Loebell1990), a mechanism independently needed for ellipsis resolution, or in priming across languages (given that code-switching in DS-TTR does not involve shift of processing environment [Gregoromichelaki Reference Gregoromichelaki, Saka and Johnson2017]).
The TTR conceptual frames invoked in processing explain the observation that speakers may show behaviour indicating that they represent semantic elements they do not hear/utter. However, with sequences of actions modelling incremental conceptual integration of stimuli, there is no need for postulating movement or empty categories while it is also predicted that long-distance dependencies of the standard kind should trigger parallel sequential patterns subsequently even in the absence of semantic parallelism. Finally, given that the DS-TTR modelling of the grammar itself is driven by the generation of predictions of upcoming sequences of actions, any already pursued action paths will always be prioritised (Myslín & Levy Reference Myslín and Levy2016), tuning processing accordingly and thus explaining why comprehension is cross-primed by production and vice versa within and between speakers.
From this perspective, syntactic knowledge is not autonomous but derivative upon other forms of procedural knowledge, namely sequential action planning and comprehension with gradual elaboration of conceptual representations expressing stimuli categorisation as it occurs across cognition (Gregoromichelaki Reference Gregoromichelaki, Orwin, Howes and Kempson2013). Consequently, such knowledge needs to be modelled in an architecture that integrates simultaneous qualitatively related constraints from various sources, rather than separate modular components expressed in distinct vocabularies, as the authors advocate. For this reason, we believe the consequences of structural priming, while transparently operative when isolated in carefully controlled experimental designs, seem to disappear in investigations of corpora that reflect multiple other sources of constraints such as frequency, creativity, affective, and social effects (Healey et al. Reference Healey, Purver and Howes2014).
In conclusion, an explanation of the structural-priming results from a DS-TTR perspective dispenses with the heterogeneous multilevel representational nature of the grammar proposed by the authors. Yet, this more radical move we propose turns out to be much more supportive of the general conclusion the authors draw, namely, the relevance of psycholinguistic explorations in determining the nature of linguistic theories. It is also more compatible with recent neuro-physiological evidence (e.g., Covington & Duff Reference Covington and Duff2016). In fact, from our perspective, priming experiments provide valuable tools for guiding the formalisation/implementation of grammar models – for example, by providing measures estimating the temporal course of pattern memory decay, investigating the competition among alternatives resulting in inhibitory effects, and determining variable probability distributions of available sequences, all currently being theoretical and observation-based assumptions in need of further substantiation.
The target article is part of a welcome recent trend to take psycholinguistic results as able to adjudicate among competing theoretical proposals, rather than being treated as simply presupposing linguists' constructs. We wholeheartedly agree with this stance; in fact, we endorse it to a much greater extent than the authors advocate: From our point of view, the paper presents a rather conservative interpretation of the cited results in that it persists with the preoccupation of abstracting over behavioural/neuronal data to underlying abstract knowledge representations presumed to underlie their explanatory mechanisms (Gregoromichelaki & Kempson, Reference Gregoromichelaki, Kempson, Scott, Clark and Carstonforthcoming; cf. Ferreira et al. Reference Ferreira, Bock, Wilson and Cohen2008).
We support the claim that methods of structural priming comparisons can be informative about mechanisms underlying linguistic processing. However, the authors argue that structural priming results are explainable only by assuming separate linguistic representations encoding semantic/syntactic/phonological information. Here, perhaps surprisingly, the authors seem to adopt the standard linguistic stance that theoretical frameworks/explanations need to presuppose an abstract, static view of linguistic knowledge, separating models of competence from accounts of performance.
In contrast, we propose an alternative formal architecture based on Dynamic Syntax (DS) as the syntactic engine (Cann et al. Reference Cann, Kempson and Marten2005; Kempson et al. Reference Kempson, Meyer-Viol and Gabbay2001; Reference Kempson, Cann, Gregoromichelaki and Chatzikyriakidis2017) enhanced with incremental construction of Type-Theory-with-Records (TTR; Cooper Reference Cooper, Kempson, Asher and Fernando2012) conceptual representations (DS-TTR; Gregoromichelaki et al. Reference Gregoromichelaki, Kempson, Howes, Eshghi, Wachsmuth, de Ruiter, Jaecks and Kopp2013; Hough Reference Hough2015; Kempson et al. Reference Kempson, Cann, Gregoromichelaki and Chatzikyriakidis2016; Purver et al. Reference Purver, Gregoromichelaki, Meyer-Viol, Cann and Purver2010). While eschewing a level of syntactic representation and any competence/performance distinction, such a framework is able to account directly for the priming data as well as standard linguistic generalisations.
Concentrating on syntax, the main focus of the paper, the data presented provide no evidence for theoretical or implementational perspectives on syntactic knowledge that would necessarily assume string-level hierarchical representations or accessing of stored well-formedness constraints in some kind of context-free-grammar format. Instead, a formal grammar adopting a DS-TTR-style architecture consisting of routinised sequences of processing actions with no syntactic representations is much more compatible with the overall data. From this perspective, the appearance of abstract structural pattern-matching is epiphenomenal on the incrementality/predictivity of the processing of time-linearly unfolding stimuli. In contrast, the methodology that involves abstracting a level of syntactic representation over the actions impedes straightforward analyses of patterns of interlocutor coordination in dialogue (Gregoromichelaki et al. Reference Gregoromichelaki, Kempson, Purver, Mills and Cann2011).
Instead of syntactic hierarchical structure, according to DS-TTR, a small set of elementary domain-general processing actions underpins both parsing and generation: Cross-linguistically available sequences of such actions cluster into higher-order sequential patterns (macros) that can be learned online, activated long-/short-term, and stored as chunks triggered by specific word forms (Eshghi et al. Reference Eshghi, Hough, Purver, Demberg and Levy2013). It is the reuse, and potential for adjustment, of such sequences that accounts for the authors' findings of “syntactic-pattern” repetitions appearing as distinct and/or independent from semantic interpretations. These results can be explained more explicitly in DS-TTR because the framework addresses the pervasive local ambiguity problem of incremental parsing/generation by predictive activation of various potential probabilistically weighted processing paths (sequences of actions). These processing paths are taken to constitute part of the context and some of them lead to identical TTR conceptual structures (Hough Reference Hough2015; Hough & Purver Reference Hough, Purver, Chatzikyriakidis and Luo2017; Sato Reference Sato, Gregoromichelaki and Howes2011). For example, PO/DO or active/passive alternations in DS-TTR reflect the invocation of distinct sequences of actions to construct or linearise equivalent conceptual event frames (with distinguishing information-structure aspects reflected in the processing order). The parser/generator initially pursues the highest-ranked option, with the rest maintained in the context for conversational-repair purposes (Eshghi et al. Reference Eshghi, Howes, Gregoromichelaki, Hough and Purver2015; Hough Reference Hough2015). Success of one such path in achieving the intended conceptualisation will enhance the probability of perception/execution of the same action sequence subsequently if the word forms accessed make it available, while inhibiting the pursuance of alternatives.
Cumulative priming effects are predicted with additional repetition of the triggering word forms (the lexical boost effect), because phonological forms are stored in context for conversational purposes like clarification. Facilitation of retrieval is predicted even when repetition of the same word forms in conjunction with the same word order leads to distinct conceptual frames (Bock & Loebell Reference Bock and Loebell1990), a mechanism independently needed for ellipsis resolution, or in priming across languages (given that code-switching in DS-TTR does not involve shift of processing environment [Gregoromichelaki Reference Gregoromichelaki, Saka and Johnson2017]).
The TTR conceptual frames invoked in processing explain the observation that speakers may show behaviour indicating that they represent semantic elements they do not hear/utter. However, with sequences of actions modelling incremental conceptual integration of stimuli, there is no need for postulating movement or empty categories while it is also predicted that long-distance dependencies of the standard kind should trigger parallel sequential patterns subsequently even in the absence of semantic parallelism. Finally, given that the DS-TTR modelling of the grammar itself is driven by the generation of predictions of upcoming sequences of actions, any already pursued action paths will always be prioritised (Myslín & Levy Reference Myslín and Levy2016), tuning processing accordingly and thus explaining why comprehension is cross-primed by production and vice versa within and between speakers.
From this perspective, syntactic knowledge is not autonomous but derivative upon other forms of procedural knowledge, namely sequential action planning and comprehension with gradual elaboration of conceptual representations expressing stimuli categorisation as it occurs across cognition (Gregoromichelaki Reference Gregoromichelaki, Orwin, Howes and Kempson2013). Consequently, such knowledge needs to be modelled in an architecture that integrates simultaneous qualitatively related constraints from various sources, rather than separate modular components expressed in distinct vocabularies, as the authors advocate. For this reason, we believe the consequences of structural priming, while transparently operative when isolated in carefully controlled experimental designs, seem to disappear in investigations of corpora that reflect multiple other sources of constraints such as frequency, creativity, affective, and social effects (Healey et al. Reference Healey, Purver and Howes2014).
In conclusion, an explanation of the structural-priming results from a DS-TTR perspective dispenses with the heterogeneous multilevel representational nature of the grammar proposed by the authors. Yet, this more radical move we propose turns out to be much more supportive of the general conclusion the authors draw, namely, the relevance of psycholinguistic explorations in determining the nature of linguistic theories. It is also more compatible with recent neuro-physiological evidence (e.g., Covington & Duff Reference Covington and Duff2016). In fact, from our perspective, priming experiments provide valuable tools for guiding the formalisation/implementation of grammar models – for example, by providing measures estimating the temporal course of pattern memory decay, investigating the competition among alternatives resulting in inhibitory effects, and determining variable probability distributions of available sequences, all currently being theoretical and observation-based assumptions in need of further substantiation.