Work in syntactic satiation has typically focused on ‘island’ structures (ex. 3–4), wh-questions which are ungrammatical in English because they are said to violate constraints governing the movement of wh-phrases in English.

More specifically, well-formed English questions (ex. 1–2) involve the creation of a ‘filler-gap dependency’ between the pronounced (the filler) and interpreted (the gap) wh-phrases. Though this dependency can span across multiple clauses, there are nevertheless conditions that govern the formation of the filler-gap dependency. When these conditions are violated, movement of the wh-filler to the front of the sentence is disallowed. In example (3), for instance, introducing a noun phrase (‘the claim’) between the filler and the gap embeds the wh-gap within a noun phrase from which wh-movement is not possible. Likewise, when the wh-gap appears within a subject phrase (‘a bottle of’), as in (4), the resulting sentence is ungrammatical. Because these phrases – namely, complex noun phrases and subjects, respectively – block the formation of wh-dependencies, they are considered ‘islands’ to extraction (here represented using brackets).

In the first experimental investigation of satiation, Snyder (2000) asked native English speakers to rate the grammaticality of several types of island structures.¹ Participants rated each sentence type a total of five times. To determine whether there had been any improvement in ratings, the number of ‘grammatical/acceptable’ responses in the first two vs. the last two exposures was compared. Sentences were said to improve, or ‘satiate,’ if there were more ‘grammatical/acceptable’ responses in the second half than in the first half of the study.

Notably, Snyder (2000) found that while some ungrammatical structures satiated, others did not.² However, more recent work has been unable to replicate some of these original findings. For instance, the satiation effects initially observed for CNPC islands have been replicated by some (e.g., Sag et al., 2007; Hofmeister and Sag, 2010; Goodall, 2011; Snyder, 2017 using acceptability ratings), but not by others (Hiramatsu, 2000 using Likert scale ratings; Sprouse, 2009 using magnitude estimation). In addition, related work by Sag et al. (2007) and Hofmeister and Sag (2010) investigated CNPC islands using self-paced reading where participants were asked to read two types of CNPC islands word-by-word: In the first type, wh-fillers were bare wh-phrases (e.g., ‘who’ or ‘what’), whereas in the second type, the wh-fillers were more informative which-NP phrases (e.g., ‘which convict’), which have been shown to be more acceptable (Karttunen, 1977; Maling and Zaenen, 1982; Pesetsky, 1987, 2000; etc.). Both Sag et al. (2007) and Hofmeister and Sag (2010) reported a similar result. Participants rated which-NP CNPC islands more acceptable than CNPC islands with bare wh-phrases. Additionally, reading times for CNPC islands with which-NPs did not differ from their grammatical, non-island counterparts. Results from both these studies were taken as evidence that under some circumstances, processing costs for CNPC islands could be drastically attenuated strictly by manipulating a single processing-related factor [(namely, the informativeness of the wh-element; but see Goodall (2015) for evidence of residual island effects even with highly informative filler phrases)]. We return to this point in the discussion.

Subject islands have been under similar debate. Although Snyder (2000) only showed a marginally significant effect of satiation, Hiramatsu (2000), Francom (2009), and Chaves and Dery (2014) have found significant satiation effects for Subject islands. Work by others, however, either replicated Snyder’s (2000) marginal effects (e.g., Snyder, 2017) or failed to detect satiation effects in these island types (e.g., Sprouse, 2009; Goodall, 2011; Crawford, 2012; etc.).

In sum, at issue is not only the question of (i) what mechanisms underlie satiation, but also the more fundamental question of (ii) whether what has been termed ‘satiation’ in CNPC and Subject islands is even the same phenomenon. In part because the basic facts of satiation remain unclear (e.g., there is no consensus regarding which structures do and do not satiate), it has been difficult to interpret what satiation as a phenomenon means both for experimental and for theoretical linguistics.

At a minimum, investigations into the phenomenon of satiation represent a methodological question for the design of acceptability judgment studies. For instance, a better understanding of the factors underlying satiation may have consequences for understanding individual variation in judgments, the number of times target items may be repeated, proximity of individual target items to each other, etc. Beyond that, satiation potentially implicates the interaction between grammatical constraints and how those constraints are mentally represented. This is particularly true in the case of grammatical violations, like CNPC and Subject islands, whose status in both the experimental and theoretical literature is still under debate.

Unlike satiation, syntactic priming – where exposure to a syntactic structure can facilitate subsequent processing of that same structure (Bock, 1986) – is a well-known and well-attested phenomenon. A large body of work (e.g., Bock, 1986; Branigan et al., 1995; Pickering and Branigan, 1998; Bock and Griffin, 2000) in priming has shown that speakers are better able to access structures (e.g., passive sentences) that they’ve previously been exposed to. And, though most of the research in priming focuses on production, similar priming effects have also been found in studies of comprehension. In general, the ability to facilitate access to recently exposed structures has been attributed to two complementary mechanisms that are not mutually exclusive (Hartsuiker et al., 2008): (1) residual activation of combinatorial nodes in a syntactic structure (often lexically based), resulting in a short-lived priming effect (e.g., Pickering and Branigan, 1998; Branigan et al., 1999) and (2) Implicit learning of mappings between message-level representations and syntactic structures, resulting in a longer-term priming effect (Bock and Griffin, 2000; Chang et al., 2006; inter alia).

Residual activation accounts typically locate priming in the lexical units which connect to the larger syntactic structure (e.g., Pickering and Branigan, 1998; Branigan et al., 1999; Pickering et al., 2000; though see Scheepers, 2003). Since recent exposure momentarily increases the activation level of syntactic structures, priming occurs when the parser selects structures which are more active in memory, e.g., structures with higher residual activation levels. Because these accounts attribute priming to the moment-by-moment activation levels of particular lexicon-to-structure combinations, they also predict a short-term time course for priming (e.g., Roelofs, 1992; Pickering and Branigan, 1998). In particular, because the activation of lexical units is believed to decay quickly and automatically, priming effects are short-lived. Further, because residual activation accounts take priming to involve the links between lexical units and their larger syntactic structure, this account also predicts a stronger priming effect when prime and target sentences share lexical items (e.g., Pickering and Branigan, 1998; Cleland and Pickering, 2003). Indeed, this ‘lexical boost’ effect has been replicated in a number of production studies (e.g., Pickering and Branigan, 1998; Cleland and Pickering, 2003; Bernolet et al., 2013) and in nearly all comprehension studies (see Tooley and Traxler, 2010 for review).³ But, other work has shown that priming can still occur absent lexical repetition in production (e.g., Pickering and Branigan, 1998; Scheepers, 2003; Kaschak and Glenberg, 2004; Hartsuiker et al., 2004) and comprehension (e.g., Luka and Barsalou, 2005; Thothathiri and Snedeker, 2008a,b; Traxler, 2008; Ivanova et al., 2012a,b).

A second mechanism contributing to structural priming – implicit learning – attributes priming to changes that occur independent of the lexicon; so, lexical repetition between prime and target sentences is not predicted to influence the strength of priming (Bock and Griffin, 2000; Chang et al., 2000, 2006; Bock et al., 2007). Rather, priming occurs as the result of cumulative, lasting learning from experience: Encountering a given message with a given structure reinforces learning of that meaning-to-message mapping. Consequently, the structure becomes more accessible the next time the processing system encounters the same type of message. Because priming under this account is the by-product of cumulative changes at the abstract structural level, priming is predicted to be relatively long-lasting (e.g., Hartsuiker and Kolk, 1998; Bock and Griffin, 2000; Bock et al., 2007; Hartsuiker et al., 2008). Work by Bock and Griffin (2000) measured the proportion of prepositional datives that participants produced after hearing a prepositional dative prime (e.g., “A boy is giving an apple to a teacher.”) or a double-object prime (e.g., “A boy is giving a teacher an apple.”). To test the longevity of priming, they varied the number of unrelated sentences intervening between the prime and target structures. Consistent with prior work hinting at the persistence of priming, they found that effects could persist through as many as 10 intervening sentences.

The role of ungrammatical structures, though, is unclear. Most work in priming has focused on structural facilitation in the context of fully grammatical sentences – sentences whose structures can be mentally represented by the comprehender. Some researchers argue against the possibility of priming in ungrammatical sentences. For example, Sprouse (2007) suggests that priming “is predicated upon the existence of a licit representation. Given that ungrammatical structures have no licit representation… there should be no syntactic priming effect for ungrammatical structures” (Sprouse, 2007, p. 128). In contrast, other work (Kaschak and Glenberg, 2004; Luka and Barsalou, 2005; Ivanova et al., 2012a,b, 2017; etc.) has suggested that priming need not be limited to fully grammatical sentences.

At the lexical level, a series of experiments by Ivanova et al. (2012a,b, 2017) investigated if and how comprehenders build syntactic representations for anomalous ditransitive sentences (ex. 5a–b), when the verb is (a) a nonce word void of any semantic meaning, (b) a grammatically unacceptable verb, or (c) missing altogether. These anomalous sentences were compared against a fully grammatical counterpart (d).

Crucially, Ivanova et al. (2012a, 2017) used the presence/absence of syntactic priming effects (assessed via the proportion of participant-produced sentences matching the structure of the prime) to diagnose whether comprehenders had built syntactic representations for anomalous sentences.⁴ They found evidence of structural priming – and thus the presence of abstract syntactic structure – with nonce-verb primes (5a), with illicit verb primes (5b) and even when the prime contained no verb (5c). Thus, work by Ivanova et al. (2012a, 2017) suggests that even when comprehenders encounter incomplete and/or ungrammatical sentences, they do not “abandon” the syntactic route altogether. In addition to using other available information, comprehenders do attempt to construct a representation for the sentence via syntax.

An open question, though, is whether findings from Ivanova et al. (2012a,b, 2017) can be straight-forwardly extended to account for structures as degraded as island structures (ex. 3–4). Anomalies in those works were largely localized to a single, albeit structurally important, lexical item – namely, the verb. Indeed, Ivanova et al. (2012b) themselves raise the question of whether their results may generalize to sentences where the locus of ungrammaticality extends beyond the level of individual lexical items – e.g., as in island structures (Ivanova et al., 2012b, p. 367).

Earlier work by Kaschak and Glenberg (2004) and Luka and Barsalou (2005) provide insights into what happens on the sentence level, although they did not test island structures. Specifically, Kaschak and Glenberg (2004) found priming-like effects in structures like ‘These vegetables need cooked.’, which are acceptable in some dialects, but ungrammatical in standard American English. In their experiment, half of the participants were exposed to the ‘needs’ structure during an initial training phase while the other half did not undergo training. Afterward, all participants were asked to read structurally similar sentences, such as ‘The valiant hero wants recognized for his courageous actions.’ Kaschak and Glenberg (2004) found faster word-by-word reading times for the novel ‘wants’ structures only for participants who had participated in the training session. This, they argued, provided evidence that participants were “learning to comprehend” the novel structure via a new meaning-to-message mapping (e.g., through implicit learning). Similar work by Luka and Barsalou (2005) investigated priming in a variety of moderately ungrammatical structures (e.g., ‘I miss having any time to do anything.’, ‘Who did you hire because he said would work hard?’). Participants first read sentences that were structurally similar to the target sentences, and after a 5-min break, rate the acceptability of the target sentences. Luka and Barsalou (2005) found acceptability improvements in as little as one prior exposure to a structurally similar sentence.

Taken together, these results indicate that priming may, indeed, be possible even with structures that initially seem unacceptable. Nevertheless, because work examining priming with ungrammatical sentences is relatively new, the limits of this priming effect are still unclear and the mechanisms and/or processes that underlie priming in ungrammatical sentences are not yet well-understood. Moreover, prior work has tended to either look at only one specific kind of anomaly, or has grouped together various types of ungrammatical sentences without comparing them systematically. Thus, it is not yet known how generalizable prior findings are, or whether different kinds of ungrammaticality may pattern differently with regard to the possibility of priming.

The current work uses methods established in priming research to guide investigations into satiation, and in so doing, aims to shed light on broader issues related to the representation of ungrammatical sentences. Given the parallels between syntactic satiation and syntactic priming – namely, that both are linked to increased exposure – it may be possible for the underlying mechanism(s) responsible for satiation to be related to those in priming. The current work aims to contribute to our understanding of satiation and priming in three ways:

If the same factors known to influence priming – namely, the proximity between individual (prime-to-target) exposures and the type of lexical overlap between structures – produce higher acceptability ratings for target sentences than for primes, this suggests that acceptability ratings may be sensitive to the same factors that affect processing. Such a finding would provide reason to suspect that priming and satiation can be linked to the same underlying mechanisms. Alternatively, if we observe no rating improvements between primes and targets, this would not rule out the possibility of a relationship between satiation and priming, but would make any such relationship indirect.

In priming, the proximity of exposure between prime-target pairs has been used to distinguish between effects arising from short-term residual activation decay and/or longer-term effects arising from implicit learning. We use this same logic to investigate whether rating improvements (satiation) may be short- or long-term. If acceptability ratings from prime to target sentences improve (i.e., satiate) when primes and targets are close together (Lag1; one intervening sentence), but show small improvements or no improvements when they are far apart (Lag5: five intervening sentences), this may point to satiation being a short-lived effect that decays over time. But, if both lags show comparable rating improvements, this could point to satiation as a long-term effect analogous to implicit structure-learning.

Finally, lexical repetition often elicits a (short-lived) strengthening of the priming effect. According to residual activation accounts, this is because lexical repetition facilitates access to previously built syntactic structures. If acceptability is also sensitive to lexical repetition, we might find an analogous acceptability-rating ‘boost’ in Lag1 (primes and targets are close together) when prime-target pairs share lexical items. In particular, we may see stronger effects when the head of the syntactic island is repeated – given the significance of the head noun in the island structure – than when phrases unrelated to the island are repeated.

Experiment 1 investigated acceptability rating improvements for CNPC and Subject islands in prime-target pairs. While prior work in satiation has compared rating improvements over the course of an entire study, our priming-style (prime-target) design allowed us to test whether factors known to affect priming might also affect satiation similarly. If so, this might provide reason to suspect that priming and satiation share underlying mechanisms. We tested two factors: (1) lexical repetition and (2) proximity of exposure between the prime and target sentences. We varied lexical repetition such that primes and targets shared either the head of the island phrase or a phrase unrelated to the island. We predicted that repetition of the head of island phrases might produce a priming ‘boost’ akin to ‘lexical boost’ effects that have been observed in priming work. In addition, we varied proximity of exposure by manipulating the number of unrelated sentences (one vs. five) between primes and targets, to probe whether potential acceptability improvements are short-term (e.g., from activation decay of structural representations) or long-term (e.g., as a result of implicit structural learning).

Experiment 2 used an online measure – self-paced reading times – to investigate whether the acceptability rating improvements in Experiment 1 were related to on-line island processing effects. We tested for the presence of reading time improvements, indicative of processing facilitation, for CNPC and Subject islands when primes and targets were close together (Lag1) and when they were further apart (Lag5). Based on results from Experiment 1, we predicted that if the acceptability rating improvements found in CNPC islands (but not Subject islands) reflected online processing facilitation, we should find corresponding prime-to-target reading time facilitation in CNPC islands (but not Subject islands) in Experiment 2. We also investigate whether online facilitation effects in Experiment 2 were short- or long-term. If target sentences are read faster than prime sentences in Lag1, but not in Lag5, this would point toward a short-lived priming effect. But if reading times for targets in both Lag1 and Lag5 are faster than their primes, this would suggest a long-lasting effect.

Unlike in Experiment 1, which found no rating improvements for Subject islands regardless of proximity between prime and target sentences, Experiment 2 found faster reading times for target sentences when Subject islands were separated by only one intervening sentence (Lag1). This effect lasted through several words in our region of interest. When sentences were further apart (Lag5), we found a prime-to-target facilitation localized to only one word in the region of interest. The finding that reading times for target sentences are facilitated by a preceding prime suggests that comprehenders are able to build representations of ungrammatical Subject islands and then draw on those representations to facilitate later processing of that same structure. In other words, Experiment 2 suggests that priming is possible in Subject islands. Moreover, the pattern of differences between Lag1 and Lag5 suggests that the type of priming observed for Subject islands may be attributed to rapid decay of lingering structural activation. This is similar to what has been proposed to account for short-term priming in grammatical sentences.

Conversely, reading times between prime-target pairs in CNPC islands did not appear to differ in Lag1. Despite results from the stop-being-grammatical task (see Differences between CNPC and Subject Islands: The Stop Being Grammatical Task), we find no reading time slowdowns associated with either the word signaling the filled-gap (w5) or the point where the processor recognizes the illicit extraction (w7) when sentences were close together. Surprisingly, we did observe a localized one-word priming effect (w5) for CNPC islands when primes and targets were far apart (i.e., in Lag5).

The reading time pattern presented by CNPC islands is difficult to interpret because no prior work has predicted a structural priming effect that only surfaces at longer intervals (Lag5) between prime and target. Even implicit learning accounts of priming, which predict a long-lasting effect, do not do so in the absence of short-term ones. Moreover, reading times for CNPC islands did not behave as one might have expected based on the stop-being-grammatical task. Results from the stop-being-grammatical task (see Differences between CNPC and Subject Islands: The Stop Being Grammatical Task) showed that comprehenders begin perceiving CNPC islands to be ungrammatical as early as the fifth word in the sentence (with most comprehenders concurring by the seventh word). Thus, comprehenders seem aware of the ungrammaticality of CNPC islands relatively early in the sentence. Yet, we do not detect processing difficulty (reading time slowdowns) at any point in CNPC sentences when prime and target are close together (Lag1).

It is worth noting that the reading time patterns we found for CNPC islands do resemble those reported for this same island type by Sag et al. (2007) and Hofmeister and Sag (2010). They investigated different issues, but used the same self-paced reading paradigm and found that reading times for CNPC islands did not differ from those in fully grammatical sentences. Crucially, their results showed that manipulating a single processing-related factor (bare wh-phrases vs. which-phrases, see Syntactic Satiation) was sufficient to effectively produce a reading-time ‘floor effect’ in CNPC islands. Though it may be possible that reading times for CNPC islands in Experiment 2 also exhibited a similar floor effect, this account provides little explanation for why reading times slowdowns were not detected for CNPC primes, which are not facilitated by prior exposure. At the moment, we leave the question of why CNPC islands did not show expected reading time slowdowns as a question for future work.

In sum, Experiment 2 leads us to conclude the following: First, reading time facilitation effects from primes to targets in Subject islands suggest that comprehenders are able to build a syntactic structure for this purportedly ungrammatical island-violation structure in real time, and that this structure can facilitate subsequent processing. Second, the results for CNPC islands suggest that structure-building for island sentences may be limited: If, following Ivanova et al. (2012a,b, 2017), we treat processing facilitation as a diagnostic for structure-building, our results indicate that comprehenders only build structures for some ungrammatical sentences. Thus, the different patterns of priming observed for Subject vs. CNPC islands reinforce the idea that the mechanisms involved in facilitating comprehension of ungrammatical sentences may not be a uniform, across-the-board phenomenon. Third, our results suggest that the proximity between prime and target sentences can affect online processing of Subject and CNPC islands, though the effect manifests itself differently for the two island types.¹⁶

Prior work has sought to directly address which factors might contribute to the different patterns of satiation across island types (cf. Hiramatsu, 2000; Kluender, 2004; Sag et al., 2007; Crawford, 2012; Chaves and Dery, 2014; inter alia). That issue is not the main focus of the experiments reported in this paper. However, both Experiments 1 and 2 suggest that Subject islands and CNPC islands behave differently. Therefore, it may be reasonable to suggest that what has been grouped under the same ‘satiation’ umbrella may actually be two different underlying mechanisms, targeting different kinds of island violations, that happen to yield superficially similar consequences.

Prior work has attempted to classify island constraints under different syntactic (e.g., Ross, 1967; Huang, 1982; Chomsky, 1986; Rizzi, 1990) or semantic (e.g., Szabolcsi and Zwarts, 1993) mechanisms. To date, though, these typologies (e.g., “strong” vs. “weak” island effects) are neither very straight-forward nor fully agreed-upon (Szabolcsi and den Dikken, 2003; Szabolcsi, 2006; etc.). However, theories (Ross, 1967; Kluender, 1998, 2004; Hiramatsu, 2000; etc.) that suggest a typological distinction between CNPC and Subject islands may be able to capture the pattern of results presented here. For instance, some accounts consider CNPC islands to be “weak” and Subject islands to be “strong” by virtue of the severity of the violation (quantified in terms of subjacency violations).¹⁷ Though our work cannot speak to the validity of these classifications, it is worth noting that our results do provide evidence against grouping CNPC and Subject islands as a natural class. Clearly, further work is required to pinpoint what precisely defines the asymmetric satiation and priming effects that we observe.

The different pattern of behaviors for CNPC and Subject islands may also speak to an ongoing debate concerning the status of island violations in general. On one hand, with CNPC islands we (unexpectedly) found reading time differences between primes and targets when primes and targets were far apart but not when they were close together. This could be argued to lend support to accounts that primarily attribute island effects to processing effects (e.g., Kluender and Kutas, 1993; Kluender, 1998, 2004; Sag et al., 2007; Hofmeister and Sag, 2010; Pearl and Sprouse, 2012, 2015; but see Phillips, 2013). On the other hand, online facilitation effects for Subject islands were not strong enough to ‘spill over’ to acceptability improvements. This suggests that while the acceptability of island sentences may be affected by processing-related factors, attempts to locate island effects wholly outside the grammar are insufficient (Ross, 1967; Chomsky, 1986; Sprouse et al., 2012a,b; Phillips, 2013; Yoshida et al., 2014). As in the case of satiation, it may be that the role of processing-related factors may affect these two island types differently.

Traditional measures of satiation have relied on acceptability judgments, which is a consequence of how satiation as a phenomenon has been defined. However, our results show that there is a benefit to looking at satiation using multiple methods. Ratings from the acceptability judgment task (Experiment 1) provide a ‘first look’ into the potential link between satiation and priming. Strikingly, once we adapted the task to an online measure (Experiment 2), it became apparent that acceptability ratings alone did not allow us to fully differentiate between the mechanisms targeting the two different sentence types. The emerging picture is admittedly complex, but adds new empirical evidence to a subfield of linguistics – satiation research – that has been characterized by a lack of consensus from the outset.

Finally, while prime-target proximity effects have been thoroughly investigated in the priming literature, our work is the first (to our knowledge) to take some initial steps toward investigating proximity in studies of acceptability ratings. Therefore, an independent contribution of our work is to highlight the need to control for distance between targets in acceptability judgment tasks.