2. Possible explanations for the small-antecedent effect

This section outlines alternative analyses of the small-antecedent effect, both those that have appeared in the literature (the Antecedent Containment Hypothesis, or ACH, Matsuo Reference Matsuo, Megerdoomian and Bar-el2001; Sailor Reference Sailor2014) and another viable candidate hypothesis related to parallelism, a factor known to influence antecedent selection.

3. Experiments 1a and 1b

The first two experiments establish the basic effect: that in ACD, but not ellipses in a coordinated clause, a small antecedent is more likely to be chosen. The only difference between the two experiments is the determiner heading the DP containing the ACD: a definite or a universal. The experiments were conducted separately so they are reported separately here.

3.1. Experiment 1a

Methods

Two types of clauses containing ellipsis (C_Es) were compared: a coordinated clause and a relative clause modifying a definite description. The clause containing the antecedent (C_A) contained a pre-verbal manner adverb.

1. (9)

   1. a. Eddy quietly sang the song and Fred did too. C_E = coordinated.

   2. b. Eddy quietly sang the song that Fred did. C_E = relative clause (ACD)

Two versions of nine sentences were created following the pattern in (9). Materials appear in Appendix A. These items were included among 61 other one- or two-sentence discourses from an unrelated experiment. Each of the sentences was followed by a forced-choice question like (10). (The order of answer options was counterbalanced.)

1. (10) What did Fred do?

   1. a. quietly sing a song

   2. b. sing a song

Participants were instructed to choose the answer that corresponded to their “understanding of what the sentence said”. They were instructed to read at a normal speed and answer with their first impression. In principle, a response of the unmodified verb phrase is compatible with participants resolving the ellipsis with a modified VP. That is, because of the entailment relation, (10b) will be a true albeit under-informative response even if a participant resolves the ellipsis with (10a). The experiment thus relies on the assumption that even if participants are reporting under-informative responses, they would do so equally across embedded and coordinated conditions. Twenty-four native English speakers completed the questionnaire via Amazon Mechanical Turk using Turktools (Erlewine and Kotek Reference Erlewine and Kotek2016) and were paid $1.50 for their participation. Two counterbalanced lists were created, so that no participant saw any one item in more than one condition. Participants were instructed to read each sentence or two-sentence discourse and click on the radio button next to their answer.

Results

The results are presented in Figure 1, which shows the mean proportion of large (i.e. modified) antecedent responses for each condition (71% vs. 17%). A logistic mixed-effects model (Baayen et al. Reference Baayen, Davidson and Bates2008; Jaeger Reference Jaeger2008) was fit using the lme4 package (Bates Reference Bates2005) in R (R Development Core Team 2012). The model included C_E type as a fixed factor, and participant and item as random factors. The models also included random intercepts and random slopes for both these factors and their interaction (a maximal random effects structure, as recommended by Barr et al. Reference Barr, Levy, Scheepers and Tily2013). There was a significant effect of C_E type (Est.=−5.1210, SE = 1.1609, z = −4.411, p < 0.001).

Figure 1: Mean proportion of large-antecedent responses in Experiment 1a

3.2. Experiment 1b

Methods

Experiment 1b tested whether the small-antecedent effect extends to ellipses within restrictors of universally quantified nouns. Looking ahead to the Parsimony Hypothesis, note that quantifiers such as every and all presuppose that the set forming their restrictors is non-empty (Lappin and Reinhart Reference Lappin and Reinhart1988; Diesing Reference Diesing1992). This is what makes the sentences in (11) pragmatically anomalous, since they presuppose the existence of American kings and unicorns.

1. (11)

   1. a. All/every American king(s) lived in New York.

   2. b. All unicorns have accounts at the Chase Manhattan Bank.

Experiment 1b tested the same materials as Experiment 1a, but with a universal every instead of a definite heading the relevant noun. All other aspects of the experiment were the same, except that participants (N = 9) were undergraduate students who completed a pen-and-paper written questionnaire for course credit. All were native English speakers.

Results

The results are presented in Figure 2, which shows the mean proportion of large (i.e. modified) antecedent responses for each condition (84% vs. 6%). A logistic mixed-effects model was fit with C_E type as a fixed factor and participant and item as random factors.Footnote ¹⁰ The analysis showed a significant effect of C_E type, such that coordinated C_Es differed from relative clause C_Es (Est. = −8.560, SE = 2.086, z = −4.104, p < 0.001).

Figure 2: Mean proportion of large-antecedent responses in Experiment 1b

Discussion

In out-of-the-blue contexts like those presented, participants chose a small antecedent in relative clause conditions (ACD) but not in coordinated conditions, both under definites (Experiment 1a) and universals (Experiment 1b).

These experiments establish the basic small-antecedent effect. I now turn to the predictions of the proposed hypotheses, the Parsimony Hypothesis (PH) and the antecedent-containment hypothesis (ACH).

The accounts in Matsuo (Reference Matsuo, Megerdoomian and Bar-el2001) and Sailor (Reference Sailor2014) – which I have grouped together as the non-categorical ACH – attribute the small-antecedent effect to the syntactic height at which the C_E is attached: economy conditions on QR preferentially place the site of C_E below the adverb, and this precludes a large antecedent as that would be an instance of antecedent containment. In syntactic configurations where there is no containment problem precluding a large antecedent, even for an embedded C_E, the ACH does not predict the kind of small-antecedent preference found for the C_Es that must invoke QR to allow a large antecedent. The competing Parsimony Hypothesis (PH), on the other hand, makes a positive prediction: even if the overt syntax makes available a large antecedent, a small antecedent would be chosen according to the PH if the C_E is presupposed as part of a definite description or strong quantifier.

These two approaches can be tested with (12a), where a post-verbal adverb is followed by an extraposed relative clause C_E. The landing site of QR can be diagnosed by the position of the extraposed relative clause (Williams Reference Williams1974; Fox and Nissenbaum Reference Fox, Nissenbaum, Bird, Carnie, Haugen and Norquest1999). Because the relative clause is outside the VP constituent containing the adverb, the processor must give (12a) the parse in (12b) where QR has applied; in this case, economy considerations on QR do not militate against a high attachment, for the simple reason that the only possible parse for (12a) is one in which the containment problem is avoided.

1. (12)

   1. 

The PH and the ACH make divergent predictions about sentences like (12a) with post-verbal adverbs, and those tested in Experiments 1a and 1b with pre-verbal adverbs as in (13).Footnote ¹¹ These predictions are tested in Experiment 2.

1. (13) John quietly sang the song that Fred did.

While the ACH predicts a preference for small antecedents in (13) (against a coordinated baseline) it does not make the same prediction about the extraposed C_E in (12a), since economy constraints cannot rule out a small antecedent. The ACH predicts either that the choice of antecedent will be at chance, or perhaps that it will pattern with coordinated cases.Footnote ¹² Either way, according to the ACH, we have no reason to expect that (13) and (12a) would pattern alike in terms of antecedent selection.The Parsimony Hypothesis (PH) on the other hand does make the strong, positive prediction that both (12a) and (13) will pattern alike and exhibit a small-antecedent effect, and to the same extent. That is because in both cases, the ellipsis is part of the restrictor of a definite description and hearers, according to the PH, are less willing to accommodate the specification of the manner of the event.

4. Experiment 2

Methods

Experiment 2 manipulated two factors, each with two levels. C_E Type: main (coordinated) or embedded (relative) and Adverb Position: pre-verbal or post-verbal. These two factors were crossed to create the four conditions shown in (14).

1. (14)

   1. a. Edna quietly entered every room that David did. embedded/pre

   2. b. Edna entered every room quietly that David did. embedded/post

   3. c. Edna quietly entered the room and then David did. main/pre

   4. d. Edna entered the room quietly and then David did. main/post

The PH predicts that the extraposed C_E in (14b) will pattern with the non-extraposed C_E in (14a) in preferring a small antecedent compared to the main clause conditions in (14c) and (14d). The ACH predicts an interaction between C_E Type and Adverb Position such that (14a) will show a small-antecedent effect compared to (14b), which is predicted to either allow large antecedents at the rate of coordination or at least at a higher rate than (14a).Footnote ¹³

Sixty-four sentences were created in the four conditions, following the pattern in (14). Ten items used every as the determiner; six used the. In Experiments 1a and 1b, relative clauses were compared to highly parallel coordinated clauses that contained the particle too. This baseline might overestimate the expectation for large antecedents for coordinated clauses, because such particles demand parallelism of de-accented material (and a fortiori elided material, see section 2.2). In an attempt to control for this confound, half of the items in Experiment 2 did not use such a particle in the coordinated case and instead used then as in (14). The other half of the items included also in both the coordinated and embedded ellipses as in (15).Footnote ¹⁴

1. (15)

   1. a. Larry energetically played every sport that Fred also did. embedded/pre

   2. b. Larry played every sport energetically that Fred also did. embedded/post

   3. c. Larry energetically played every sport and Fred also did. main/pre

   4. d. Larry played every sport energetically and Fred also did. main/post

If differences in antecedent selection are simply due to the presence or absence of particles such as too or also, we expect to find no small-antecedent effect in Experiment 2.

Four counterbalanced lists were created in which each item appeared once in only one condition. Each item was followed by a forced-choice question as in Experiments 1a,b. The placement of the adverb in the answers was counterbalanced. The 16 items were included among six unrelated experiments, all followed by a comprehension question or a naturalness rating. Materials are provided in Appendix B. Forty-eight native English speaking undergraduates participated in a computerized questionnaire. The items werepresented in random order with the items from the other experiments.

Results

Figure 3 shows the mean proportion of large-antecedent responses for each condition for 47 participants.Footnote ¹⁵ Main clause (coordinated) conditions showed agreater proportion of large-antecedent responses than the embedded (relative) conditions. In both cases, the large antecedent was chosen more often when the adverb was post-verbal.

Figure 3: Mean proportion of large-antecedent responses in Experiment 2

The results were analyzed with a logistic mixed-effects model. The model reported here included as fixed factors C_E Type and Adverb Position, and participant and item as random factors. The models also included random intercepts and random slopes for both these factors and their interaction. There was a significant effect of C_E Type (Est. = 1.8514, SE = 0.3364, z = 5.504, p < 0.001) and a significant effect of Adverb Position (Est. = −0.6889, SE = 0.2372, z = −2.905, p < 0.01). There was no interaction.

Table 1 shows the means broken down by the presence of also. Items with also showed a slight numerical increase in large-antecedent responses in the post-verbal condition, for both the embedded and coordinated ellipses. There was no significant effect of the presence of also (Est. = 0.04668, SE = 0.32706, z = .143, p > .5).Footnote ¹⁶

Table 1: Experiment 2, mean proportion of large antecedents (standard deviations)

Discussion

The results of Experiment 2 were not those predicted by the ACH. Regardless of the position of the adverb, embedded C_Es were more likely to take small antecedents than main C_Es. If the small-antecedent preference were due to the low attachment of the C_E when the adverb is pre-verbal, then we should have observed an interaction. The results are consistent with the Parsimony Hypothesis, which predicts that it is the nature of the C_E that is at stake: more semantically specified (but contextually unmotivated) material is less preferred in presupposed clauses. An asserted C_E imposes no such requirements and a large antecedent is significantly more available. It is important to understand that while the rates of large-antecedent responses are higher for the extraposed embedded C_E than the non-extraposed embedded C_E – which might give the impression that the two conditions do not pattern alike, as the ACH predicts – this difference is a part of a second main effect of Adverb Position that holds across C_E-type. This means that the difference between the two embedded conditions can be factored out: adverb position is an additional factor that promotes small antecedents across clause types, which I will now address.Footnote ¹⁷

The main effect of Adverb Position was not predicted by either the ACH or PH. Nonetheless, I would like to briefly outline a possible explanation for this effect. It is well known that there are semantic differences between pre-verbal and post-verbal adverbs (Jackendoff Reference Jackendoff1972; Peterson Reference Peterson1997; Shaer Reference Shaer, Fabricius-Hansen, Lang and Maienborn2000). Morzycki (Reference Morzycki, Kennedy and McNally2008) argues that pre-verbal adverbs may be interpreted non-restrictively whereas post-verbal adverbs must always restrict the VP. If Morzycki makes the bet in (16a) and it turns out that Floyd performed the nose job but not easily, he loses. In (16b), where the adverb is pre-verbal, it is not clear whether the bet is won or lost.

1. (16)

   1. a. I'll bet you $80 that Floyd could perform a successful nose job in a moving taxi easily.

   2. b. I'll bet you $80 that Floyd could easily perform a successful nose job in a moving taxi. (Morzycki Reference Morzycki, Kennedy and McNally2008: 105–106(17,18))

This is because pre-verbal adverbs have both a restrictive and non-restrictive reading and a non-restrictive modifier can ‘escape’ the propositional content of the complement clause. (See Potts Reference Potts2005 on how this dimension of ‘not-at-issue’ content is integrated.) Relevant to the present discussion is the fact that there is existing experimental evidence to support the idea that the search for antecedents for VPE need not target non-restrictive material. Frazier and Clifton (Reference Frazier, Clifton, Harris and Grant2011) show that, rather unexpectedly, the subject of the second clause in (17a) can be interpreted as though modified by the restrictive relative in the subject of the first clause. However, the same effect was not found with non-restrictive relatives (17b).

1. (17)

   1. a. A man who had climbed Mt. Greylock called and a woman did too.

   2. b. A man, who had climbed Mt. Greylock, called and a woman did too.

They argue that parallelism requirements are responsible for the enrichment to the meaning of the subject of the second clause in (17a). They suggest that not-at-issue material like the non-restrictive relative can escape the kinds of parallelism requirements imposed on (17a). It is possible, then, that a somewhat similar explanation applies to the main effect of Adverb Position in Experiment 2: a non-restrictive pre-verbal adverb similarly escapes parallelism requirements and need not be part of the ellipsis. Post-verbal adverbs, which are only restrictive, necessarily form part of the main assertion of the antecedent clause. As established by Hardt and Romero (Reference Hardt and Romero2004) and Frazier and Clifton (Reference Frazier and Clifton2005), content from a main assertion is more likely to be resolved as part of an ellipsis. While this proposal for the effect of Adverb Position requires further testing, it bears repeating that in Experiment 2, this effect held regardless of C_E type: it will not by itself explain the small-antecedent effect when C_E is part of a presupposed restrictor clause.

Another informative result of Experiment 2 concerns the role of also, a parallelism-promoting item. Numerically, the presence of also does appear to increase large-antecedent responses in the post-verbal conditions, but the effect was not significant. So while such items may boost large-antecedent responses, their presence cannot explain the contrast between main and embedded C_Es.Footnote ¹⁸

One potential drawback is that in Experiments 1a, 1b, and 2, coordinated clauses are used as a baseline. Even without discourse particles like too and also, coordinated clauses may impose parallelism constraints that demand not just matching antecedent and ellipsis, but overall parallelism between the C_A and C_E as discussed in section 2 (Frazier et al. Reference Frazier, Clifton and Munn2000; Frazier and Clifton Reference Frazier and Clifton2001; Carlson Reference Carlson2003). To ensure that the small-antecedent effect is not a by-product of comparing embedded C_Es to parallelism-seeking coordinated clauses, Experiment 3 was designed to hold the relation between C_E and C_A constant. To do this, VP ellipsis was tested in temporal modifiers headed by after, when and before as in (18a).

1. (18)

   1. a. Sally quickly read the book after/when/before Fred did.

   2. b. After/when/before Sally quickly read the book, Fred did.

These were compared to minimally different sentences in which the ellipsis was in a main clause (18b). Neither involves coordination, and so we do not expect overall parallelism constraints to enforce maximal similarity between the two clauses. The only difference is that the ellipsis is in a matrix clause in (18b), but in a temporal embedded clause in (18a).

These stimuli offer an additional way to test the PH against the ACH. Sailor (Reference Sailor2014) concentrates almost exclusively on adverbial subordinate C_Es like these. Investigating these provides a further test of the family of approaches the ACH encapsulates. Furthermore, the PH makes predictions about ellipsis in temporal adjuncts because they are presuppositional, containing definite descriptions of times. A number of other authors argue that the prepositions after and before embed expressions that refer to unique times (Beaver and Condoravdi Reference Beaver, Condoravdi, Young and Zhou2003; von Stechow Reference von Stechow2002, Reference von Stechow, Klein and Li2009; Condoravdi Reference Condoravdi2010; Sharvit Reference Sharvit2014). Heim (Reference Heim, Reuland and ter Meulen1987), for instance, gives a paraphrase with a definite description:

1. (19) Mary left before/after John arrived.

   ↝ Mary left before/after the time at which John arrived.

As von Stechow (Reference von Stechow, Klein and Li2009) and others point out, this paraphrase incorrectly presupposes that John arrived just once. Beaver and Condoravdi (Reference Beaver, Condoravdi, Young and Zhou2003) propose that the before-clause contains an operator equivalent to earliest, which does not require one event but does still presuppose there is a unique earliest time:Footnote ¹⁹

1. (20) “A before B” = 1 iff (∃t ∈ A) t < earliest(B)

   “A after B” = 1 iff (∃t ∈ A) t > earliest(B)

2. (21) Sally quickly read the book before Fred did.

   ↝ Sally quickly read the book before “the earliest time” that Fred read the book.

When-clauses have likewise been argued to semantically contain a definite description of times. Caponigro (Reference Caponigro and Young2004), Caponigro and Pearl (Reference Caponigro and Pearl2009) and Hall and Caponigro (Reference Hall, Caponigro, Li and Lutz2010) argue that when-clauses are free relatives, and they may contain a maximality operator (in the sense of Link Reference Link, Baeuerle, Schwarze and von Stechow1983): the when-clause denotes a set of time intervals and the maximality operator shifts this set to the maximal element in that set. In (22) the maximal element is the largest time interval in which Sally read the book.Footnote ²⁰

1. (22) John quickly read the book when Sally did.

   ↝ John quickly read the book at/during the time interval in which Sally read it.

Since (22) has this “hidden” definite description, we expect its restrictor (the set of times denoted by the C_E) to be subject to constraints on parsimony, just as relative clauses that form the restrictor of the overt definite article are.Footnote ²¹ Resolving to a large antecedent requires hearers to accommodate in (22) not just the particular time at which Sally read the book, but also the manner of that event. The Parsimony Hypothesis predicts readers will choose the unmodified antecedent more often in the embedded clause ((18a), (22)) than the main clause (18b).

5. Experiment 3

Methods

Experiment 3 tested two factors each with two levels. C_E Type: main vs. embedded (using after, before, or when); and Adverb Position: pre-verbal vs. tpost-verbal. The two factors were crossed to create the four conditions in (23).

1. (23)

   1. a. Peter quietly walked into the house after Lesley did. embed/pre

   2. b. Peter walked into the house quietly after Lesley did. embed/post

   3. c. After Peter quietly walked into the house, Lesley did. main/pre

   4. d. After Peter walked into the house quietly, Lesley did. main/post

Twenty items were created in four conditions following the pattern in (23), using the three kinds of temporal subordinators (when, after, before). These items were included among twenty other items, which consisted of twelve three- or four-sentence discourses testing an unrelated phenomenon. Eight filler items were included. Each sentence was followed by a forced-choice question with the modified and unmodified antecedents as options. Four counterbalanced lists were created, so that each participant saw each item once in only one condition. Materials appear in Appendix C. Forty native English speaking undergraduates participated in a written comprehension questionnaire. As with Experiment 2, the ACH predicts an interaction between C_E Type and Adverb Position whereas the PH predicts a main effect of C_E type.

Results

Figure 4 shows the mean proportion of large-antecedent responses. When the ellipsis was presented in a main clause, participants chose the large antecedent more often than if the ellipsis was found in a temporal embedded clause. In both cases, more large-antecedent responses were observed if the adverb was post-verbal in the antecedent.

Figure 4: Mean proportion of large-antecedent responses in Experiment 3

A mixed model was fit, which had as fixed factors C_E Type and Adverb Position, and participant and item as random factors with a maximal random effects structure. The analysis showed a significant effect of C_E Type (Est. = 1.1405, SE = 0.3703, z = 4.135, p < .001) and a significant effect of Adverb Position (Est.=−0.7754, SE = 0.2383, z= −3.254 , p < .001). There was no interaction.

Discussion

The results of Experiment 3 are not compatible with the hypothesis that the small-antecedent effect is due only to a countervailing preference for large antecedents in coordination due to parallelism. The embedded clauses differed from non-coordinated main clauses in the likelihood of large-antecedent selection. Additionally, the experiment confirmed that the small-antecedent effect is not due to a containment problem, as the ACH predicts: for the condition where the C_E is not contained in the modified VP (the post-verbal embedded condition), readers were as likely to chose an unmodified antecedent compared to main clauses as they were for the non-contained pre-verbal embedded conditions. The separate main effect of adverb position replicates that found in Experiment 2 and bears out Frazier and Clifton's (Reference Frazier, Clifton, Harris and Grant2011) proposal that non-restrictive material is less likely to be targeted as part of an antecedent.

Summarizing, Experiment 3 shows once again that the small-antecedent effect arises in embedded clauses generally, and it cannot be attributed to a containment problem alone as per the ACH. Instead, the predictions of the PH are borne out. Like ACD constructions headed by the and every which showed the small-antecedent effect, temporal adverbial clauses, which are likewise presuppositional, exhibit a small-antecedent effect compared to asserted main clauses. In the next section, I provide a more explicit formulation of the PH and show how it is part of a more general constraint on accommodation in processing as argued by Crain and Steedman (Reference Crain, Steedman, Dowty, Karttunen and Zwicky1985) and Altmann and Steedman (Reference Altmann and Steedman1988).

6. Parsimony and accommodation in processing

Crain and Steedman (Reference Crain, Steedman, Dowty, Karttunen and Zwicky1985) claim that the famous garden path in is a result of pragmatic considerations, not parsing constraints (see Frazier Reference Frazier1978).

1. (24) The horse raced past the barn fell. (Bever Reference Bever and Hayes1970)

The target, but not initially pursued, restrictive relative (RR) parse requires the listener to not only accommodate a (unique) horse that was raced past the barn, but also a contrast set of horses which do not have the property ascribed by the relative.Footnote ²² The more-readily pursued main verb (MV) analysis only requires the accommodation of the horse, not a contrast set. In this sense, the MV parse requires a more parsimonious or simpler discourse model. Crain and Steedman (Reference Crain, Steedman, Dowty, Karttunen and Zwicky1985) propose a Principle of Parsimony: the reading that carries fewer unsupported presuppositionsFootnote ²³ will be favored over one that carries more. However this is implemented formally, the intuition is clear that the RR parse requires the reader to imagine a more elaborate discourse model. Grodner et al. (Reference Grodner, Gibson and Watson2005) found that even without ambiguous input, readers find restriction difficult to process in a context that does not motivate restriction. They compared (unambiguous) restrictive and (unambiguous) non-restrictive relatives and found that in null contexts, the restrictive relatives were read more slowly. (When supporting contexts were provided, the effect reversed direction.) They argue that the results support the Principle of Parsimony, in that the restrictive relative requires a more elaborate discourse model (evoking a contrast set) which the non-restrictive does not.Footnote ²⁴ Additionally, a number of studies have shown that individuals prefer unmodified NPs in a number of ambiguous constructions and that the Principle of Parsimony is responsible (Ni Reference Ni1996; Sedivy Reference Sedivy2002; Spivey-Knowlton and Sedivy Reference Spivey-Knowlton and Sedivy1995).

The small-antecedent effect has a similar, albeit not identical flavour. In the case of ellipses in presupposed clauses, choosing a large, modified antecedent, as in (25b), leads to a richer discourse model than the small antecedent (25a): the large resolution ascribes not only the property of being read by Mary to the book in question, but the manner as well.

1. (25)

   1. a. John quickly read the book that Mary did read.

   2. b. John quickly read the book that Mary did quickly read.

If the ellipsis is part of the restrictor of a presuppositional element (a definite description, a universal), it requires accommodation in a null context (Karttunen Reference Karttunen1974; Stalnaker Reference Stalnaker, Munitz and Unger1974; Lewis Reference Lewis1979; see von Fintel Reference von Fintel, Reimer and Bezuidenhout2004 for a recent overview). When readers are given a choice between parses that require accommodation, they choose the parse that requires the ‘least’ accommodation, in the sense that the relative clause with an unmodified (small) VP denotes a superset of the entities picked out by a relative with a modified (large) VP.

1. (26)

   $$\underbrace{{\lcub {{\rm x}:{\rm Mary}\;{\rm quickly}\;{\rm read}\;{\rm x}} \rcub }}_{{{\rm large}\;{\rm antecedent}}}\subseteq \underbrace{{\lcub {{\rm x}:{\rm Mary}\;{\rm read}\;{\rm x}\;} \rcub }}_{{{\rm small}\;{\rm antecedent}}}$$

Put another way, the unmodified relative is less informationally rich and, in the null context, more parsimonious.Footnote ²⁵

In this way, Parsimony not only militates against introducing more individuals into a discourse model (as in Crain and Steedman Reference Crain, Steedman, Dowty, Karttunen and Zwicky1985) but it evaluates between different “amounts” of accommodation, where the subset relation ranks the information content of the restrictors.

1. (27) Parsimonious Accommodation

   All else being equal, in a null context, given two parses [Det A] and [Det B], where Det is a presuppositional determiner and A and B are properties such that A ⊂ B, choose parse [Det B].

Parsimony only regulates restrictor material that requires accommodation. When the elided phrase is part of asserted content, the pressures of parsimony are called off. This is consistent with Experiments 1–3, where main clauses did not show a small-antecedent preference. In the last section, we saw that temporal adjunct clauses could be profitably analyzed as involving definite descriptions of times. If their content is subject to accommodation, then here too a less informationally rich (i.e. unmodified) ellipsis is chosen.

It is important to stress that Parsimonious Accommodation is intended here only as a processing heuristic. There is certainly no barrier to accommodating “more” than is strictly necessary in, for instance, bridging scenarios:

1. (28) John read a book about Schubert and wrote to the author. (Heim Reference Heim1982: 371)

As Heim (Reference Heim1982, p. 371ff.) points out, the author we accommodate is not just any author (that would certainly be more parsimonious) but the author of the book mentioned in the first clause. In my terms, this amounts to accommodating a richer set of properties of the individual. But here the pressures of discourse coherence motivate the bridging.Footnote ²⁶ My claim is that Parsimonious Accommodation governs parsing in impoverished contexts: when faced with a syntactic ambiguity, readers appeal to it. This, of course, was also the intended application of the Principle of Parsimony in Crain and Steedman (Reference Crain, Steedman, Dowty, Karttunen and Zwicky1985).

A reviewer asks how the processor enforces (27): does the processor compare the available options in parallel and evaluate which one requires fewer commitments? Another option, as pointed out by the reviewer, is that the reader might pursue the small antecedent first without entertaining the large antecedent. We leave these interesting questions about time-course to future research.

The natural prediction is that non-presuppositional determiners will lift the small-antecedent effect. The simplest direct test of this prediction would be to choose a relative clause C_E that is headed by a non-presuppositional determiner. This is not so straightforward, however, since only strong, presuppositional determiners license ACD (Carlson Reference Carlson1977; Diesing Reference Diesing1992). For instance, the weak cardinal reading of few in (29a) does not license ACD, whereas the partitive version (which is always presuppositional) does (29b):Footnote ²⁷

1. (29)

   1. a. ??I read few books that you did.

   2. b. I read few of the books that you did. (Diesing Reference Diesing1992: 71)

Since non-presuppositional determiners and quantifiers cannot head ACD, the most direct test of the Parsimony Hypothesis is not possible. Instead, I test non-restrictive relatives, in which the content of the relative is asserted and therefore not subject to Parsimonious Accommodation.

7. Experiment 4: Restrictive vs. Non-restrictive

Unlike restrictive relatives, the content of non-restrictive relatives does not serve as the restrictor of the head noun, and so is not subject to accommodation. Instead, non-restrictive relatives form a type of secondary assertion (see Potts Reference Potts2005).Footnote ²⁸

1. (30)

   1. a. John read the book that is on the table.

      assertion: John read the book

      presupposition: there is a (unique) book on the table

   2. b. John read Moby Dick, which is on the table.

      assertion: John read Moby Dick

      secondary assertion: Moby Dick is on the table

Parsimonious Accommodation is not expected to constrain ellipsis choice in non-restrictive relatives because their content is not presupposed. Testing non-restrictives, however, is not without problems, as non-restrictive relatives often resist ACD:

1. (31) *Dulles suspected Philby, who Angleton did. (May Reference May1985: 12(23))

Vanden Wyngærd and Zwart (Reference Vanden Wyngærd and Zwart1991) report that non-restrictive ACD is ameliorated with as well or not. (See also Pesetsky Reference Pesetsky2000; Lasnik Reference Lasnik1999.)

1. (32) ?Dulles suspected Philby, who Angleton did not/as well.  (Vanden Wyngærd and Zwart Reference Vanden Wyngærd and Zwart1991: 154(19))

To the extent that non-restrictive ACDs are acceptable, my judgement is that a large antecedent is available.

1. (33)

   1. a. Dulles wrongly suspected Philby, who Angleton did wrongly suspected as well.

   2. b. Maria quickly spoke to Regina, who Bill did not quickly speak to.

However, we saw in Experiment 2 that additive modifiers like as well could boost the proportion of large antecedents (due to the parallelism constraints it imposes). To counteract this effect, Experiment 4 tested restrictive and non-restrictive relatives both with as well. This may increase the overall number of large antecedents, but I assume this would apply equally across C_E-type.

1. (34)

   1. a. Rob secretly suspected Mary, who Peter did as well. non-restrictive

   2. b. Rob secretly suspected the woman that Peter did as well. restrictive

      What did Peter do?

      Secretly suspect Mary

      Suspect Mary

The restrictive relative forms the descriptive content of the definite determiner, and so requires accommodation. I have argued that accommodation in processing is subject to parsimony, and readers will choose the less informative NP meaning (the unmodified VP). The non-restrictive relative is not presupposed, and instead serves as a type of secondary assertion (Potts Reference Potts2005). If the small-antecedent effect is due to pressures on accommodation, as hypothesized, we predict a difference between the two relatives in antecedent choice such that small antecedents will be chosen more often in restrictive relatives.

Methods

Twelve item sets like (34) were created with the two-level factor C_E-type: non-restrictive and restrictive. The order of answer options (modified, unmodified) was counterbalanced. Twenty participants completed one of two lists using Amazon Mechanical Turk and were paid $1.25 for their participation. All participants identified themselves as native English speakers.

Results

The mean proportion of large antecedents is reported in Figure 5. Non-restrictive relatives received 39.1% large-antecedent responses; restrictives, 27.5%. A mixed model was fit with C_E type as a fixed factor, and participant and item as random factors. The model also included random intercepts and random slopes for participants and items. There was a significant effect (Est.=−0.8087, SE = 0.3995, z = −2.025, p < 0.05).

Figure 5: Mean proportion of large-antecedent responses in Experiment 4

Discussion

The results point in favour of the hypothesis that a pressure for small antecedents is conditioned by the presuppositional versus assertoric nature of the clause containing the ellipsis. Relative clauses that form the restrictor of a definite must be accommodated, and this accommodation is subject to Parsimony, giving rise to a greater preference for an unmodified VP resolution. Non-restrictive relatives do not form the restrictor of the presuppositional determiner, but are instead secondary assertions that do not require accommodation. Hence, there is not the same pressure to choose a small antecedent.

The raw scores for large antecedents in non-restrictive relatives are not overwhelmingly high, as they are in main clauses in Experiments 1a, 1b or Experiment 2. In fact, they do not reach chance. A reviewer points out, however, that there may be several reasons for “small antecedents”, Parsimonious Accommodation being only one of them. In this particular case, there may be several reasons why non-restrictives do not prefer large antecedents. First, the baseline grammaticality of ACD in non-restrictive relatives may not be high, and this could have the effect of creating more noise in the results if less acceptable sentences are harder to judge. Second, this experiment only employed pre-verbal adverbs, and we have already seen that pre-verbal adverbs in the antecedent are less likely to correlate with modified ellipsis interpretations. Finally, it is important to keep in mind that raw numbers are hard to interpret, especially without a baseline of coordination. The important result is that the non-restrictives are different from restrictive relatives, even when both include the parallelism-inducing as well, which supports the claim that Parsimonious Accommodation plays a role, although further research may find that there are additional factors that militate against large antecedents.

8. General discussion

I began with the intuition, reported by Matsuo (Reference Matsuo, Megerdoomian and Bar-el2001), that in ACD constructions manner adverbs are not targeted as part of the ellipsis. I showed that while that is not an absolute, grammatical, prohibition, it is a replicable trend in speakers’ interpretation of elided sentences. I set out to determine the reason for the small-antecedent effect, ruling out along the way a very plausible syntactic explanation (the ACH). Ultimately, the explanation that fits the experimental results best is the one that attributes the small-antecedent effect to a very general pragmatic principle that militates against accommodating a richer context than needed (Crain and Steedman Reference Crain, Steedman, Dowty, Karttunen and Zwicky1985). When readers are given a choice between parses that require accommodation, in a null context they choose the parse that requires the ‘least’ amount of modification in the sense of Parsimonious Accommodation. Evidence in favour of the Parsimony Hypothesis—and Parsimonious Accommodation in general—involved comparing restrictive to non-restrictive relatives; in the latter case the content of the relative does not need to be accommodated because it is a (secondary) assertion.

It becomes an interesting question whether adverbial modification is dis-preferred in embedded contexts generally, even outside of ellipsis, as in (35):

1. (35) John read the book that Fred quietly read.

Of course, we are forced in (35) to accommodate the adverb because it is overtly present. This is unavoidable. What the small-antecedent effect has shown us is that when faced with an ambiguity, readers will choose a syntactic parse that does not force “more” accommodation than necessary. This demonstrates that ellipsis resolution is readily influenced by pragmatic principles governing general expectations about information exchange and complexity in a discourse. This is generally recognized in the literature (Hardt and Romero Reference Hardt and Romero2004; Frazier and Clifton Reference Frazier and Clifton2005) but we are only beginning to understand how to dissociate these pressures from syntacic-semantic principles governing the identity conditions on ellipsis resolution. I should point out, though, that Grodner et al. (Reference Grodner, Gibson and Watson2005) found that in null contexts, even unambiguous restrictive relative clauses were read more slowly than non-restrictives. They suggest that the Principle of Parsimony governs routine syntactic parsing, not just in cases of syntactic ambiguity. This raises the possibility that “excess” modification within a restrictive relative, perhaps as in (35), will pose processing costs. I leave it to future research to investigate this.

As noted, the simplest direct test of the PH would have used relative clause C_Es headed by a non-presuppositional determiner. This is not possible, however, as only strong, presuppositional determiners license ACD (Carlson Reference Carlson1977; Diesing Reference Diesing1992; Fiengo and May Reference Fiengo and May1994). Weak determiners can, however, contain relative clauses—especially in presentational contexts like (36a) (sm is the reduced version of some, and it is unambiguously weak):

1. (36)

   1. a. Here are sm books that John bought.

   2. b. Here are some books that John bought.

With sm, (36a) does not presuppose that the restrictor is non-empty; it introduces to the common ground that there is a set of books John bought. The non-reduced some in (36b) is ambiguous. While weak determiners do not allow ACD, they can have elided VPs as long as the VP is not contained in the antecedent, for example if the antecedent is sentence-external as in (37):

1. (37)

   1. a. Speaker A: The candlesticks that were carefully cleaned aren't here.

   2. b. Speaker B: Here are sm that were carefully cleaned.

A large antecedent is certainly possible in (37b), as the proposal predicts, but the fact that the manner-modified antecedent VP itself requires accommodation might prime the reader to further accommodate the same kind of VP event in the elided clause.

Since Parsimonious Accommodation does not tie the small-antecedent effect to anything particular about ACD, it is predicted that any ellipsis in a presuppositional DP will be constrained by parsimony and show a small-antecedent effect. In (38a), the elision is part of a subject relative under a definite; in (38b), the elision is in a presentational (non-presupposed) relative. (I have put the antecedent VP in an asserted clause to avoid priming accommodation.)

1. (38) The man quietly sang the song and left.

   1. a. The woman that did sing the song stayed.

   2. b. There's a woman that did quietly sing the song and she stayed.

While testing is required, to my ears (38a) does not commit to the manner of the elided VP in the same way that (38b) does. Nonetheless, the effect may be attenuated because in the creation of such examples, the context becomes already quite a bit richer. I leave this for further research.

Aside from pragmatic processing questions, the small-antecedent effect also bears on prominent issues in the syntax of ellipsis. One of these issues is the claim that ellipses target the largest constituent available (MaxElide). Another is the representation of manner adverbs in the VP. I turn to these in the concluding subsections.