1.1 Revisiting previous conclusions about antecedent-ellipsis mismatches

There are two reasons to revisit previous conclusions about the ellipsis acceptability cline. First, it is difficult to distinguish the accounts proposed by Arregui et al. (2006) and Kim et al. (2011). Both accounts appeal to parser operations to capture the gradient acceptability, and without any strongly divergent predictions, they cover much of the same empirical ground (see Kim & Runner 2018 for discussion).

Second, the experimental findings for antecedent-ellipsis mismatches resemble findings from previous studies on other syntactic dependencies involving anaphora and agreement, where different conclusions are drawn about the nature of the underlying representations, mental processes, and processing architecture. For instance, many studies have shown that ungrammatical subject-verb agreement relations in configurations like (5) are often overlooked during real-time language processing, due to the presence of a non-subject plural lure, e.g., “the cabinets”, leading to increased acceptability relative to sentences that lack a plural item (e.g., Pearlmutter et al. 1999; Wagers et al. 2009). This effect is described in the literature as “agreement attraction”, and constitutes a kind of “acceptable ungrammaticality”, in the same sense that antecedent-ellipsis mismatches constitute a case of acceptable ungrammaticality (Arregui et al. 2006).

Many researchers have argued that the mistaken acceptability of sentences like (4) arises not because of the application of parser-specific heuristics, extra-grammatical processes, or separate structure-building systems for the grammar and parser, but rather because the grammatical constraints on syntactic dependency formation are implemented in real time using a noisy memory retrieval system that creates the opportunity for errors and gradient acceptability (Wagers et al. 2009; Dillon et al. 2013; Tanner et al. 2014; Lago et al. 2015; Tucker et al. 2015; Parker & Phillips 2016; 2017). On this view, long-distance syntactic dependencies such as subject-verb agreement are implemented in real time by retrieving a licensor from memory using a cue-based retrieval mechanism (McElree 2000; 2006; McElree et al. 2003; Lewis & Vasishth 2005; Lewis et al. 2006). This mechanism works by probing previously processed material for a constituent that matches a particular set of feature-based retrieval cues, such as +subject and +plural for subject-verb agreement, to satisfy the structural and featural requirements on the dependency. In sentences like (5), a partial match to the non-subject noun “cabinets” based on the +plural feature can give the impression that agreement is licensed, boosting acceptability of an otherwise ungrammatical sentence (Wagers et al. 2009).

This retrieval-based account of acceptable but ungrammatical subject-verb agreement relations has been extended to a wide range of acceptable ungrammaticalities involving case licensing (Sloggett 2013), reflexive licensing (Patil et al. 2016; Parker & Phillips 2017), negative polarity item licensing (Vasishth et al. 2008), and even agreement in noun phrase (NP) ellipsis (Martin et al. 2012; 2014), but not yet cases of acceptable ungrammaticalities involving VP-ellipsis. Crucially, in each of these studies, the acceptable ungrammaticalities are not taken as evidence for a special class of parser-specific rules or distinct structure-building systems for the grammar and parser. Rather, they all assume that parsing relies on a single structure-building system (i.e., the grammar) implemented in a noisy general cognitive architecture. Based on these findings, taken together with the growing evidence that ellipsis is resolved in real-time using the same error-prone retrieval mechanism proposed for these other dependencies (Martin & McElree 2008; 2009; 2011; Martin, et al. 2012; 2014), the question arises whether acceptable ungrammaticalities involving ellipsis can be captured in a similar fashion.

1.2 The present study

The present study tests whether the VP ellipsis acceptability cline can be captured under the same retrieval-based account that has been proposed for the wide range of acceptable ungrammaticalities observed for other family-related dependencies. On this view, the acceptability cline observed for antecedent-ellipsis mismatches does not reflect the application of parser-specific heuristics or a parser-grammar misalignment. Rather, such effects are a natural consequence of a single structure-building system that relies on noisy memory retrieval mechanisms to recover an antecedent. According to this account, when a comprehender encounters an ellipsis site, they engage a retrieval process to find an antecedent that matches the structural and morphological retrieval cues (i.e., the search criteria) compiled at the ellipsis site. When an antecedent does not match the search criteria in some way, as in the case of mismatching antecedent-ellipsis sentences, a processing disruption is observed, which can be detected in acceptability ratings. Based on previous findings on the relation between memory retrieval and acceptability (e.g., Dillon et al. 2015), the size of that disruption is determined by the degree of match between the retrieval cues and the corresponding features of the antecedent (i.e., “probe-to-target” similarity), such that antecedents with a greater mismatch are more difficult to interpret and consequently rated as less acceptable than their matching counterparts.

This proposal is developed in two experiments in the current study. To preview, Experiment 1 uses untimed acceptability judgments to confirm the acceptability cline reported in previous studies (Arregui et al. 2006; Kim et al. 2011). Then, as proof-of-concept, Experiment 2 uses an established computational model of memory retrieval (ACT-R; Lewis & Vasishth 2005) to show that the observed acceptability profiles follow from independently motivated principles of working memory, without invoking multiple structure-building systems. The model accurately predicts the behavioral profiles observed in Experiment 1 as a consequence of probe-to-target similarity, without adjusting the model’s default parameters.

2.1 Participants

Participants were 36 native speakers of English who were recruited using Amazon’s Mechanical Turk web service (https://www.mturk.com). This sample size was determined by a statistical power test generated using previously published results. Given the baseline .73-point difference between voice mismatches reported in Arregui et al. (2006), the power analysis suggested that 35 participants would be needed to achieve power of 80%. One additional participant for a total of 36 was recruited to balance list presentation. Participation required an IP address in the United States, and each participant was screened for native speaker abilities. The screening probed participants’ knowledge of the constraints on English tense, modality, morphology, and syntactic islands. Participants were required to meet the following qualifications: location was restricted to the United States, HIT approval rate for all requesters’ HITs was greater than or equal to 98, and the number of HITs approved for each participant was greater than or equal to 5000. Participants were assigned a qualification after they completed the study that would prevent them from participating in the experiment more than once. The experiment lasted approximately 20 minutes, and each participant received $3 for participating in the experiment.

2.2 Materials

Materials consisted of 12 item sets of VP ellipsis sentences with matching voice (active-active and passive-passive), 12 item sets of sentences with mismatching voice (active-passive and passive-active), and 12 item sets of sentences with gerundive antecedents (verbal and nominal gerunds), as shown in Table 1. These item sets were created in a way to intersect all previous experimental studies on ellipsis mismatches (Arregui et al. 2006; Kim et al. 2011; Kertz 2013 Kim & Runner 2018). All materials were taken directly from the Appendices of Arregui et al. (2006). The full materials list is provided in the Supplementary Materials. The sentences from each item set were distributed in a Latin square design across 2 lists, such that each participant read 6 sentences for each (mis)match type.

2.3 Procedure

Sentences were presented using Ibex (http://spellout.net/ibexfarm/). Participants were instructed to rate the acceptability of each sentence using a 7-point scale (7 = most acceptable, 1 = least acceptable). Participants were required to complete the experiment in one hour, which gave them adequate time to rate each sentence. All participants completed the task within the time limit. Each sentence was displayed in its entirety on the screen along with the rating scale. Participants could click boxes or use the numerical keypad to enter their ratings. The order of presentation was randomized for each participant.

2.4 Data analysis

Data were analyzed using linear mixed-effects models, with fixed factors for experimental manipulations and a fully specified random effects structure, which included random intercepts and slopes for all fixed effects by participants and by items (Baayen et al. 2008; Barr et al. 2013). Each model included simple effect-coded fixed effects (+0.5/–0.5 for each level within each mismatch type). To ensure that all comparisons were made, three additional models were coded to test for differences across mismatch types (voice match vs. voice mismatch, voice match vs. gerundive antecedents, voice mismatch vs. gerundive antecedents). Models were built using the lme4 package (Bates et al. 2018) in the R software environment (R Development Core Team 2014). If there was a convergence failure or if the model converged but the correlation estimates were high, the random effects structure was simplified following Baayen et al. (2008) until convergence obtained. An effect was considered significant if the absolute t-value was greater than 2 (Gelman & Hill 2007).

2.5 Results

The mean acceptability ratings by condition for Experiment 1 are presented in Figure 1. Figure 2 provides the histograms of ratings for each condition. Overall, matching voice sentences were rated higher than mismatching voice sentences and gerundive antecedent sentences (matching voice vs. mismatching voice: $\widehat{\beta }\text{= 1}.49,SE\text{= 0}.12,t\text{= 12}.24$ ; matching voice vs. gerunds: $\widehat{\beta }\text{= 0}.91,SE\text{= 0}.23,t\text{= 3}.90$ ), and sentences with a gerundive antecedent were rated higher than mismatching voice sentences $(\widehat{\beta }=0.58,SE=0.25,t=2.27)$ . Within the voice match sentences, active-active forms were rated higher than passive-passive forms $(\widehat{\beta }=0.56,SE=0.13,t=4.30)$ . Within the voice mismatch sentences, passive-active forms were rated higher than active-passive forms $(\widehat{\beta }=1.11,SE=0.18,t=6.11)$ . There was also a significant interaction between voice match and voice $(\widehat{\beta }=-0.55,SE=0.17,t=-3.17)$ . Lastly, within the gerundive antecedent sentences, verbal gerunds were rated higher than nominal gerunds $(\widehat{\beta }=0.82,SE=0.20,t=3.92)$ .

Figure 1

Mean acceptability ratings and standard errors by participants for Experiment 1.

Figure 2

Histograms of ratings for Experiment 1.

2.6 Discussion

The results of Experiment 1 replicated each portion of the acceptability cline reported in previous studies using a single, uniform acceptability measure across (mis)match types. Previous studies based on Likert scale judgments (Arregui et al. 2006) and magnitude estimation judgments (Kim et al. 2011) showed that voice matches are more acceptable than voice mismatches and sentences with gerundive antecedents, and that sentences with gerundive antecedents are more acceptable than voice mismatches (i.e., matching voice > gerundive antecedents > mismatching voice). Furthermore, active-active forms were shown to be more acceptable than passive-passive forms, sentences with a verbal gerund antecedent were shown to be more acceptable than sentences with a nominal gerund antecedent, and passive-active forms were shown to be more acceptable than active-passive form. Experiment 1 confirmed each of these contrasts. The results of Experiment 1 provide a single, uniform measure of the acceptability cline that will serve as the basis for the computational model of ellipsis resolution described in the next section.

A concern with the results of Experiment 1 is whether the observed effects are specific to elliptical structures. For instance, the observed contrasts may simply reflect an identity contrast for coordinated structures, or a general requirement on information-structural coherence (e.g., Kertz 2013), with VP-ellipsis showing one instance of these constraints in action. However, recent work by Kim and colleagues (Kim et al. 2011; Kim & Runner 2018) has shown that the contrasts are specific to ellipsis. Kim and colleagues reasoned that if the contrasts observed for voice and nominal mismatches reflect general well-formedness constraints that are not specific to ellipsis, then the same effects should be equally present for the non-ellipitical counterparts. Across several experiments, Kim and colleagues consistently found that the degradation in acceptability for voice and nominal mismatches was limited to VP-ellipsis structures. These findings suggest that the observed contrasts reflect an ellipsis-specific structural identity condition, which sets the stage for the memory-based account proposed in the next section.

3.1 Model details

A computationally complete model of retrieval for ellipsis resolution must specify (i) the fixed mechanisms and principles of the memory architecture, (ii) the cues that drive retrieval, (iii) the antecedent structures encoded in memory that those cues target, and (iv) the linking hypothesis that relates the principles and mechanisms of retrieval to behavioral measures. These components are described in detail in the next four subsections.

3.1.3 Antecedent structures

The model assumes the same antecedent structures proposed in Kim et al. (2011) and Arregui et al. (2006), as shown in Figures 3 and 4. Following Kim et al. (2011), the voice head distinguishes active and passive antecedents. On their analysis, the passive voice head contributes an overt morpheme “-en”, whereas the active voice head is phonologically null. This distinction is adopted in the current study, such that marking is not used as a cue for active structures (represented in Table 3 as “null”). Following Arregui et al. (2006), the critical distinction between the gerund antecedents is that a VP is included in the verbal gerund, but not in the nominal gerund. Lastly, based on recent experimental evidence on ellipsis processing, it will be assumed that the ellipsis site includes a pointer that links the ellipsis clause to a syntactic representation of the antecedent in memory (Martin & McElree 2008; 2009; 2011). But the findings reported here are equally compatible with alternative accounts that assume structure sharing (Frazier & Clifton 2005) or a cost-free process that involves copying antecedent information into the ellipsis site (Frazier & Clifton 2001).

Figure 3

Assumed syntactic structures of active and passive antecedents (adapted from Kim et al. 2011).

Figure 4

Assumed syntactic structures of verbal and nominal gerund antecedents (adapted from Arregui et al. 2006).

3.2 Simulations

5,000 Monte Carlo simulations were run for each condition in Table 1 (i.e., Active-Active, Passive-Passive, Active-Passive, Passive-Active, Verbal Gerund, and Nominal Gerund). All parameters were set to the default values described in previous work (Lewis & Vasishth 2005; Vasishth et al. 2008), with the exception of the Latency Factor (F), which was chosen to set the predictions on the appropriate time scale for sentence processing. The same parameter setting was applied to each condition. A comprehensive list of the parameter values used in the current study is shown in Table 4. Each simulation included the full series of hypothesized retrievals, based on Equations 1–7. I report the average predicted latency by condition, which provides a measure of how long on average antecedent retrieval took in each condition.

3.3 Model results

Figure 5 shows the model’s predicted mean retrieval latencies for each condition in Table 1, mapped to the corresponding acceptability judgments from Experiment 1. Across conditions, there is a tightly coupled inverse relation between predicted retrieval latencies and acceptability, such that as retrieval latencies increase, acceptability decreases. Overall, the model accurately predicted the three key components of the acceptability profile: matching voice > gerundive antecedents > mismatching voice (where > indicates “more acceptable than”). For voice match sentences (Active-Active vs. Passive-Passive), the model predicted increased difficulty for passive-passive sentences, which maps to decreased acceptability, due to the increased processing time associated with passive clause structures, i.e., retrieval of the subject noun phrase at the object gap following the verb. For the voice mismatch sentences (Passive-Active vs. Active-Passive), the model predicted increased difficulty for Active-Passive forms. Both mismatches involve additional routine retrievals to resolve passivization, and hence increased processing time, which maps to decreased acceptability relative to voice matches. But Active-Passive mismatches are predicted to be more difficult than the Passive-Active mismatches due to the combination of mismatches on voice and morphological cues, e.g., passive + -en, as shown in Table 3. For gerundive antecedents, the model predicted increased difficulty for nominal gerunds. In the case of the verbal gerund mismatch, a VP antecedent is available, but in an unexpected structural position, yielding a mismatch on the clause cue. By contrast, no suitable VP antecedent exists for nominal gerunds, as noted by Arregui et al. (2006), violating position and category expectations, as shown in Table 3. Taken together, these results can be considered a success for the model, as there is a direct correspondence between predicted processing time and acceptability (adjusted R² = 0.88), without adjusting the model parameters.

Figure 5

The antecedent-ellipsis acceptability cline presented as predicted retrieval latency for the antecedent-ellipsis (mis)matches in Table 1, mapped to the corresponding judgments from Experiment 1.

4.1 Summary of results

The present study used computational modeling to better understand why antecedent-ellipsis mismatches vary in acceptability. Previously, antecedent-ellipsis mismatches motivated the proposal that ellipsis resolution relies on special processing rules that either (i) repair and restructure an ungrammatical antecedent to satisfy syntactic identity constraints at the ellipsis site (Arregui et al. 2006), or (ii) over-generate a set of potential antecedent representations that are ranked and searched based on grammar-independent processing heuristics (Kim et al. 2011). These accounts reinforce the standard view in (psycho)linguistics that the parser and grammar reflect distinct structure-building systems that operate on different time scales using a distinct set of rules and representations (Fodor et al. 1974; Levelt 1974; Townsend & Bever 2001; Ferreira et al. 2002; Ferreira & Patson 2007).

The alternative view pursued in the current study is that antecedent-ellipsis mismatches are the product of a single structure-building system (i.e., the grammar) implemented in a noisy cognitive architecture. This hypothesis was motivated by a series of recent studies on a related class of acceptable ungrammaticalities involving subject-verb agreement, reflexive licensing, case licensing, and negative polarity item licensing, which claim that spurious acceptability arises not because of the application of parser-specific heuristics, repair rules, or a parser-grammar misalignment, but rather because the constraints on linguistic dependency formation are implemented using a noisy memory retrieval system that creates the opportunity for errors (Vasishth et al. 2008; Wagers et al. 2009; Phillips, et al. 2011; Dillon et al. 2013; Sloggett 2013; Tanner et al. 2014; Lewis & Phillips 2015; Tucker et al. 2015; Parker & Phillips 2016; 2017; Parker & Lantz 2017). Based on these findings, the present study tested whether antecedent-ellipsis mismatches could be accounted for in a similar fashion, leading to a more unified account of acceptable ungrammaticalities.

To this end, Experiment 1 used untimed acceptability judgments to confirm the acceptability cline reported in previous studies (Arregui et al. 2006; Kim et al. 2011). Results replicated each portion of the acceptability cline reported in the literature, based on a within-participant comparison of the key mismatches using a single, uniform acceptability measure across ellipsis forms. Then, as proof-of-concept, Experiment 2 used an established computational model of memory retrieval (ACT-R; Lewis & Vasishth 2005) to show how the observed acceptability profile follows directly from independently motivated principles of working memory without invoking multiple representational systems. Specifically, the model predicted differences in processing time as a function of the degree of match between the retrieval cues at the ellipsis site and the features of the antecedent (i.e., the independently motivated principle of probe-to-target similarity), with a tightly coupled inverse relationship between processing time and acceptability.

4.2 What distinguishes the retrieval-based account?

All accounts of the VP ellipsis acceptability cline, including the current account, accurately predict certain aspects of the acceptability profile, and all accounts link acceptability to the extent of the syntactic mismatch. So, what distinguishes the current retrieval-based account from previous accounts (e.g., Arregui et al. 2006; Kim et al. 2011; Kertz 2013)? I suggest that there are three distinguishing factors.

First, the current retrieval-based account provides an explicit description of how ellipsis is resolved during moment-by-moment processing and how online processes map to acceptability judgments. Previous accounts (i.e., Arregui et al. 2006; Kim et al. 2011; Kertz, 2013) are underspecified about the precise moment-by-moment parser operations, and do not provide sufficient details about the cost metrics associated with the proposed processing operations (i.e., repair, information structure analysis, search) or how they are applied in real time. For instance, Kim and colleagues reference parsing strategies involving a serial search procedure, but they do not describe the costs associated with a serial search. Furthermore, the proposal for a serial search predicts response times that are incompatible with the empirical data on retrieval times in sentence comprehension (e.g., Murdock 1971; McElree 2000; 2006; McElree et al. 2003). The current retrieval account, by contrast, explicitly defines each step of the computation, and provides a detailed description of the associated costs and how they map to behavioral measures, based on independently motivated principles of working memory. However, it is important to emphasize that there are limitations on the inferences that we can draw from the processing of ungrammatical strings. An important task for future work is to explore the predictions of the current retrieval-based account under a broader range of grammatical configurations. For instance, one prediction of the current account is that introducing multiple VPs that match the proposed retrieval cues should trigger similarity-based interference, in both grammatical and ungrammatical configurations. This prediction will be tested in future research.

Second, all accounts must consider that real-time ellipsis resolution relies on memory retrieval mechanisms to access the antecedent (Martin & McElree 2008; 2009; 2011; Martin et al. 2012), and that retrieval processes directly impact acceptability judgments (Bever 1970; Dillon et al. 2015; Kim & Runner 2018). If key portions of the acceptability cline can be explained based solely on the routine operations of the memory retrieval system, as shown in the current study, then the motivation for additional extra-grammatical operations is reduced. Importantly, the current retrieval-based account does not deny that additional post-retrieval factors (e.g., information structure constraints, antecedent reanalysis) can further impact acceptability, but rather claims that post-retrieval operations cannot be the source of the observed effects. To elaborate, retrieval is necessary, but not sufficient for interpretation, and ellipsis resolution does not culminate upon completion of memory access. As such, there are many additional constraints on interpretation, such as those involving information-structural constraints (e.g., Kertz 2013) and reanalysis (e.g., Arregui et al. 2006) that may be applied as well-formedness checks on the output of retrieval, further shaping acceptability. The proposal that certain constraints are applied as post-retrieval checks on dependency resolution is not new (e.g., Dillon et al. 2013 for a similar proposal for reflexive licensing), and the application of those constraints will feed interpretation and acceptability as comprehension unfolds over time. However, these additional processes should not disrupt the monotonic relationship between the retrieval operations that guide initial processing and acceptability. It is clear from the good model fits obtained in the current study that the memory system places a strong constraint on the acceptability profile.

The third factor that distinguishes the current retrieval-based account from previous accounts is that the current account improves empirical coverage. Specifically, the retrieval-based account captures four facts that existing accounts fail to explain. These differences are summarized below.

1. Active-Active > Passive-Passive: Existing accounts do not explain why Active-Active matches are more acceptable than Passive-Passive matches. More specifically, existing accounts do not predict a difference between these forms. According to previous accounts, acceptability for active-active and passive-passive matches should be identical, since both involve a fully-matched antecedent. Under the retrieval-based account, the degraded acceptability of passive-passive matches reflects increased processing cost due to additional retrievals to resolve passivization (see § 3.1.3). The current account provides an explicit description of these processes.

2. Passive-Active > Active-Passive: Existing accounts predict no difference between voice mismatches. Under the repair account proposed by Arregui et al. (2006), restructuring a passive antecedent to create an active match and vice versa should involve the same number of repair steps. Arregui and colleagues argue that there should be a difference between these mismatches, based on early findings that comprehenders are more likely mis-recall an active as a passive than the other way around (Mehler 1963). But in the previous studies on ellipsis mismatches, participants were not asked for an explicit recall judgment or required to recall the sentence verbatim, as in Mehler’s study, and there is little reason to believe that comprehenders incorrectly encoded or misremembered the passive structures used in the current materials (which were obtained from Arregui et al. 2006). The search-based account proposed by Kim et al. (2011) assumes that ellipsis for both mismatches targets the same projection, in which case both mismatches should involve the same amount of search work to recover a matching antecedent. Under the retrieval-based account, the differences in acceptability between these mismatches reflect more mismatching cues required by the passive ellipsis.

3. Voice matches > Verbal gerund mismatches: Kim et al. (2011: 349) noted that it is unclear under existing accounts why verbal gerunds are less acceptable than normal VPs, since a VP antecedent is available in the syntactic representation of the sentence. The current account offers an explanation for this difference. Under the retrieval-based account, verbal gerunds are less acceptable because they violate the expectation for an antecedent which occurs as a matrix, as opposed to embedded, verb phrase. This effect may be grounded in Frazier and Clifton’s “main assertion principle”, which favors antecedents in the main clause (Frazier & Clifton 2005; Frazier 2013). This preference is implemented in the cue structure with the clause cue.

4. Gerundive antecedent mismatches > Voice mismatches: Existing accounts do not offer an explanation for why voice mismatches are worse than gerundive mismatches. Under the current retrieval-based account, voice mismatches involve mismatching cues and additional retrievals (and hence increased processing costs) to resolve passivization. Likewise, the gerund constructions also contain a main clause VP that matches some of the retrieval cues, e.g., [+VP, +main clause], which helps boost their acceptability. But this VP does not bear the appropriate E-feature required to license ellipsis (Merchant 2011).

4.3 Questions for future research

The present study showed that key aspects of the ellipsis acceptability cline can be captured with a single structure-building system implemented in a noisy memory retrieval system. These findings are consistent with recent arguments that the acceptable ungrammaticalities observed for other linguistic dependencies are the product of noisy memory retrieval mechanisms (Vasishth et al. 2008; Wagers et al. 2009; Phillips et al. 2011; Dillon et al. 2013; Sloggett 2013; Tanner et al. 2014; Lewis & Phillips 2015; Tucker et al. 2015; Parker & Lantz 2017; Parker & Phillips 2016; 2017). The current results thus contribute to a more unified account of acceptable ungrammaticalities across linguistic dependencies and provide a theoretical foundation to investigate other types of antecedent-ellipsis mismatches discussed in the literature.

For instance, an important task for future research is to determine how to capture the “comet” mismatches described in Arregui et al. (2006), e.g., Seeing the comet was nearly impossible, but John did see the comet, and the category mismatches discussed in Kim et al. (2011), e.g., An admission of guilt was needed, but the suspect wouldn’t admit guilt. It is possible that the comet mismatches could be explained based on an extension of the current proposal for the gerundive mismatches by exploiting the expectation for a matrix, rather than an embedded, VP (Frazier & Clifton 2005; Frazier 2013). Likewise, category mismatches like those tested by Kim et al. (2011) could be explained by exploiting the expectation for a VP antecedent using the corresponding category cues. I leave further investigation of this suggestion to future work.

Another type of mismatch worth investigating under the current proposal includes the so-called “tolerable” vs. “intolerable” mismatches, e.g., John likes this movie and Bill might like this movie too, vs. *John is fond of this movie and Bill might be fond of this movie too (Lipták 2015). It is unclear at present how the acceptability profile of tolerable vs. intolerable mismatches might be captured under the current retrieval-based framework, without appealing to post-retrieval semantic or discourse-level processes. A similar problem arises for the contrast between voice mismatches in sentences like *This problem was looked into by John, and Bob did look into this problem too, vs. This problem was to have been looked into, but obviously nobody did look into this problem (Kehler 1993b). These sentences are problematic for existing accounts because both involve a passive-active mismatch but differ in acceptability. One possibility is that this contrast might involve an interaction of memory retrieval effects, syntactic processing, and discourse-level processing, as discussed in recent accounts of ellipsis resolution (Kim & Runner 2018). An important task for future work is to tease apart the contributions of these different factors.

Another question is how the current retrieval-based account applies to other types of ellipsis, such as sluicing (IP ellipsis). Both VP ellipsis and sluicing rely on the same memory retrieval system (Martin & McElree 2011), but unlike VP ellipsis, sluicing does not tolerate voice mismatches (Merchant 2001; 2013b). One potential explanation for this difference is that sluicing and VP ellipsis rely on different retrieval cues. For instance, VP-ellipsis relies on a VP category cue with corresponding verbal cues (e.g., voice, morphology) to guide retrieval, but sluicing relies on cues at the IP level, and thus does not deploy the verbal cues that give rise to gradient acceptability for mismatching VPs. At this level of analysis, the different behaviors for sluicing vs. VP ellipsis with respect to voice mismatch effects might reflect differences in what cues the grammar makes available to the retrieval system.

Lastly, it is worth discussing how the current retrieval-based approach fits with existing formal theories of VP-ellipsis. Formal theories of ellipsis typically fall into one of two categories. Syntactic theories assume that the content of the ellipsis site involves detailed structure, whereas referential theories assume that the ellipsis site involves a null proform/pointer, akin to other types of referential expressions, such as pronouns (e.g., Tanenhaus & Carlson 1990; Hardt 1993; Ginzburg & Sag 2000; Culicover & Jackendoff 2005; Martin & McElree 2008). Some syntactic theories assume that a full copy of the antecedent is reconstructed at the ellipsis site, whereas referential theories assume that the ellipsis site does not include a copy of the antecedent, but rather a direct link to the antecedent (see Phillips & Parker 2014 for a review). The current retrieval-based account fits most naturally with a pointer-style analysis of ellipsis resolution, which is consistent with recent experimental work showing that real-time ellipsis resolution is mediated by a pointer mechanism (e.g., Martin & McElree 2008). However, the current account does not rule out the possibility that a representation of the antecedent is reconstructed at the ellipsis site, possibly as a post-retrieval operation.