Branigan & Pickering (B&P) tell a straightforward tale. Linguists rely on grammaticality judgments to uncover representations. Judgments have limitations, but no other psycholinguistic methods systematically reveal linguistic structure. Fortunately, priming offers a direct window onto representation, providing evidence for two distinct levels: a surface syntactic form, independent of meaning and void of lexical content, and a semantic form that includes information about thematic roles, quantifier scope, and information structure.

We are fond of priming ourselves, but this elegant story is misleading in several ways. First, the priming literature does not strongly support the theory that B&P propose. As they dive deeper, the loose ends and contradictions emerge, but their final conclusions bypass this complexity. If we rearrange the evidence a bit, the theoretical ambiguity becomes clearer.

The primary evidence for syntactic representations comes from studies of argument alternations (dative or active-passive) that perfectly confound surface syntax with thematic mappings. B&P note that a few foundational studies demonstrated that syntax can be primed independent of thematic mappings (sect. 2.1). Thus, they privilege syntax in their theory. But there is now an equally robust literature demonstrating that thematic mappings can be primed independent of syntax (e.g., Cai et al. Reference Cai, Pickering and Branigan2012; Chang et al. Reference Chang, Bock and Goldberg2003; Cho-Reyes et al. Reference Cho-Reyes, Mack and Thompson2016; Hare & Goldberg Reference Hare, Goldberg, Hahn and Stones1999; Salamoura & Williams Reference Salamoura and Williams2007; Ziegler & Snedeker Reference Ziegler and Snedeker2016b). B&P acknowledge this work (sect. 2.4) but treat it as a secondary, interface phenomenon: Thematic information remains separate from syntax (Fig. 1).

Similarly, the observation that priming can occur in the absence of lexical overlap motivates a theory in which the syntactic skeleton is separate from the lexical content. To account for the lexical boost, B&P must complicate their story, by linking lemmas to structures (sect. 2.3). But perhaps we should revisit the claim that the syntactic structure lacks lexical nodes. Indeed, function words can be a locus of priming (Bencini et al. Reference Bencini, Bock and Goldberg2002; Ferreira Reference Ferreira2003). We know that only partial overlap in the syntactic skeleton is needed for structural priming (sect. 2.1), but we don't assume that the unnecessary pieces are removed from the syntactic representation. Lexical content may be similar: always present and sometimes contributing to priming via overlap.

The evidence for their semantic level is also sparse. We know: (1) Quantifier scope can be primed, (2) this priming is isolated to the particular quantifier used (e.g., each does not prime every), and (3) it abstracts away from the nouns and verbs in a sentence (Feiman & Snedeker Reference Feiman and Snedeker2016; Raffray & Pickering Reference Raffray and Pickering2010). However, B&P's claim that scopal priming is bound to thematic roles and cannot be captured by an LF representation is controversial (Chemla & Bott Reference Chemla and Bott2015). It rests on a single null result with prime stimuli (A boy climbed every tree) that have not been shown to produce priming when thematic roles are the same. Furthermore, the manipulation used confounds verb-specific roles, thematic roles, and the notion of deep subject/object. It's just too early to conclude that scope and thematic roles are tightly coupled, or that LF isn't the locus of scopal priming.

It seems that, under the right conditions, almost any linguistic representation, mapping, or process can be primed. Consequently, evidence for priming is always interpretable to some degree (it demonstrates a commonality between prime and target). But the absence (or magnitude) of an effect is often less constraining, because there is so much variability across tasks and stimuli. In some comprehension tasks, there is no priming in the absence of verb overlap (Arai et al. Reference Arai, van Gompel and Scheepers2007), while in others, abstract priming is robust (Thothathiri & Snedeker Reference Thothathiri and Snedeker2008a; Reference Thothathiri and Snedeker2008b). This problem isn't unique to comprehension. The pattern of effects in production can depend on how the sentences are elicited (stem completion vs. full sentence generation; Ziegler & Snedeker Reference Ziegler and Snedeker2016a).

Understanding this instability is critical; we suspect that the answer lies in thinking through the processes involved in each task and how they engage both stored representations and representations that are constructed on the fly. To do this, we will have to move beyond the notion of priming as a static, atemporal phenomenon that targets stable representations independent of the process of production (or comprehension).

While priming is not the transparent window that B&P promise, psycholinguists do have a much wider range of tools than the paper suggests. Some are behavioral. Novel word generalization studies going back to the 1950s have revealed structural regularities in the representation of linguistic form and meaning (Berko Reference Berko1958; Fisher Reference Fisher2002; Gropen et al. Reference Gropen, Pinker, Hollander, Goldberg and Wilson1989; Naigles Reference Naigles1990; Prasada & Pinker Reference Prasada and Pinker1993). Similarly, artificial language learning sheds light on the representations that learners extract from linguistic data and use to guide generalization (Pothos Reference Pothos2007; Reber Reference Reber1967; Saffran et al. Reference Saffran, Hauser, Seibel, Kapfhamer, Tsao and Cushman2008; for review, see Erickson & Thiessen Reference Erickson and Thiessen2015).

New methods for analyzing imaging data also provide greater constraint on representational theories. For example, multi-voxel pattern analysis, a class of machine-learning algorithms that examine patterns of neural activity (Haxby et al. Reference Haxby, Gobbini, Furey, Ishai, Schouten and Pietrini2001), has revealed regions of the left temporal cortex that appear to bind arguments to something roughly like thematic roles (Frankland & Greene Reference Frankland and Greene2015). Structural priming is therefore only one useful tool of many.

Finally, we think that B&P are too pessimistic in their assessment of linguists and their tools. There are longstanding traditions of experimental work in phonetics, phonology, and language acquisition. In recent years, experimental work has also become common in syntax, semantics, and pragmatics (e.g., Arunachalam Reference Arunachalam2013; Cowart Reference Cowart1997; Myers Reference Myers2009; Sorace & Keller Reference Sorace and Keller2005). Indeed, the question of how armchair judgments translate into generalizable conclusions has received considerable attention (Sprouse & Almeida Reference Sprouse and Almeida2012; Sprouse et al. Reference Sprouse, Schütze and Almeida2013). From our perspective, the remaining disputes do not reflect an over-reliance on grammaticality judgments or a dearth of appropriate methodologies; they stem from: (1) the close parallels between the theories that are still standing (similar operations assigned to different theoretical levels), (2) the lack of falsifiability for the contrasting features, and (3) the danger we all face of letting our “affection for [our] intellectual child[ren]” (Chamberlin Reference Chamberlin1897) guide our interpretation of the data.