3.1 Research questions and hypotheses

As seen in the previous sections, most studies on embedded gapping in English and in Spanish have taken a binary perspective in terms of grammaticality: grammatical non-embedded gapping vs. ungrammatical embedded gapping in English, and grammatical embedded gapping with non-factive verbs vs. ungrammatical embedded gapping with factive verbs in Spanish.

It seems that by confining oneself to the grammatical/ungrammatical contrast, one does not manage to have a complete perspective on the functioning of the embedding of fragments in general and of gapping in particular. Adopting an approach based on acceptability makes it possible to address certain limitations of the traditional binary classification of sentences. In light of this, the following research questions guided our study:

RQ1. Is there an interaction between gapping and embedding (in other words, is embedded gapping considered acceptable)?

RQ2. Is there an interaction between gapping and factivity (in other words, is gapping considered acceptable if it is embedded under a factive verb)?

With regard to RQ1, we hypothesize that constraints on gapping constructions are less strict than what the literature on ellipsis has traditionally claimed, and, thus, embedded gapping should be acceptable in Spanish (at least with some verbs). With regard to RQ2, we hypothesize that the acceptability of embedded gapping is sensitive to the semantic class of the embedding predicate and that there is an interaction between gapping and factivity; in particular, we expect that gapping constructions embedded under a factive verb would be less acceptable than their full counterparts. Overall, we expect a gradience in acceptability rates, far from the categorical distinctions in terms of grammaticality that we see in previous studies.

In order to test these two hypotheses, we ran a first acceptability judgment task, which was followed by a follow-up experiment (which addressed the issue of mood alternation with emotion verbs in Spanish). We present these experiments in the following lines.

3.2 Experiment 1

3.2.4 Analyses and results

Acceptability judgments (1–7) were entered into a mixed-effect linear regression analysis using the lme4 package (Bates et al. 2015) in R (R Development Core Team 2008) with the predictors Gapping and Embedding Complexity and random intercepts and slopes for Gapping and Embedding Complexity for Items and Subjects (Barr et al. 2013). A more detailed explanation of the models used for the data analysis, as well as tables summarizing the models’ fixed effects are given in the Appendix.

The results of Experiment 1 revealed no main effect of Gapping, as participants judged equally acceptable the Embedded Gapping conditions compared to their non-elliptical counterparts. Overall, these results go against what has been reported in the literature for English: based on two experimental studies (a written questionnaire and an auditory comprehension study), Carlson (2001) shows that in English there is a preference for non-gapping over gapping structures. The same preference for non-gapping structures in English is confirmed by Bîlbîie et al. (2019), who show that languages may differ with respect to the preference between an elliptical construction and its non-elliptical counterpart. Coming back to our Spanish data, Table 1 summarizes the mean acceptability judgments in raw scores and z-scores. These results are shown in Figure 1.⁸

Table 1

Mean acceptability judgements for Experiment 1 (SD in parentheses).

	Gapping		No Gapping
	raw scores	z-scores	raw scores	z-scores
No Embedding	6.55 (1.32)	0.44 (0.67)	6.61 (0.98)	0.45 (0.57)
Emb. No Factive	6.28 (1.28)	0.25 (0.75)	6.21 (1.56)	0.23 (0.89)
Emb. Factive	4.66 (1.98)	–0.67 (1.09)	4.61(1.94)	–0.70 (0.99)

Figure 1

Mean acceptability judgments in z-scores for Experiment 1.

The linear mixed model revealed, however, a main effect of Embedding (p < .01), as embedded configurations were judged less acceptable than non-embedded ones (mean rates: 5.44 vs. 6.58). As Figure 2 shows, this penalty for embedding is not related to gapping. In addition to the main effect of Embedding, there was a significant main effect of Factivity (p < .001), as embedded clauses under a factive verb are less acceptable than embedded clauses under a non-factive verb (mean rates: 4.63 vs. 6.24). Once again, as Figure 3 illustrates, this was regardless of whether there is gapping or not. We explored further this sensitivity to the semantic type of embedding predicates by distinguishing between (emotion) true factive verbs and (knowledge) semi-factive verbs. As shown in Figure 4, embedding predicates under true factive verbs were rated as less acceptable than predicates under semi-factive verbs (mean rates: 3.9 vs. 5.6).

Figure 2

Effect of Embedding.

Figure 3

Effect of Factivity.

Figure 4

Zoom into the different kinds of factive verbs.

The interaction between Gapping and Embedding was not significant, and neither was the interaction between Gapping and Factivity.

3.2.5 Discussion

Coming back to our first research question, we hypothesized that embedded gapping should be acceptable in Spanish. Our results show that, although items with embedding are rated lower than items with no embedding, embedded gapping is indeed acceptable in Spanish (the Embedded Gapping conditions are significantly different from the ungrammatical controls p < .001),⁹ and this is so regardless of gapping. Concerning our second research question, we hypothesized a gradience in acceptability rates with respect to the semantic class of the embedding predicate and an interaction with gapping. In particular, we expected that sentences that contain an embedded clause introduced by a factive verb would be judged less acceptable with gapping than with full sentences. This hypothesis was not borne out by our results, as both conditions [Gapping, Embedding Factive] and [No Gapping, Embedding Factive] were judged equally (un)acceptable by our participants.

There are two reasons which may have disfavored factive verbs in our experiment. First, this finding may be related to our choice of mood for the verb in the embedded construction in the [No Gapping, Embedding Factive] condition. Recall that, for consistency reasons, we decided to employ the indicative mood in this condition. While the indicative form is grammatical in contexts with emotion factive predicates, this remains the more marked form, compared to the subjunctive (Lope Blanch 1958; Borrego et al. 1987; Bosque 1990; Porto Dapena 1991; Real Academia Española 2010). This probably translated into the lower acceptability judgments in this condition. A second possibility is that our results may have been influenced by the context wh-question we used before the target experimental sentences. Note that the second conjunct introduced by a factive verb does not answer that question; rather it addresses a different QUD (= Question Under Discussion), e.g. no longer what someone did, but what effect it had on the speaker, leading to a discourse incoherence (see Section 4.2 below). These two aspects (i.e. mood selection and context wh-question) were addressed in Experiment 2.

3.3 Experiment 2

3.3.2 Materials

The same 24 experimental items, in the same 6 experimental conditions, employed in Experiment 1 were used in Experiment 2. There were, however, two important modifications. First, in the [No Gapping, Embedding Factive] condition, we took into account the mood preferences of the embedding predicate: the verb in the embedded construction was presented in its subjunctive form if the embedding predicate was a true factive (= emotion verb) as in (18), but in its indicative form if the embedding predicate was a semi-factive (= knowledge verb) as in (19).

1. (18)

1. En
2. in
1. el
2. the
1. bar,
2. bar
1. Pablo
2. Pablo
1. pidió
2. ordered.IND
1. una
2. a
1. cerveza
2. beer
1. y
2. and
1. me
2. me
1. molesta
2. bother.3SG
1. que
2. that
1. Juan
2. Juan
1. pidiera
2. ordered.SBJV
1. un
2. a
1. whisky.
2. whisky
1. ‘At the bar, Pablo ordered a beer and I am bothered that Juan ordered a whisky.’

1. (19)

1. Para
2. for
1. la
2. the
1. fiesta,
2. party
1. Juan
2. Juan
1. trajo
2. brought.IND
1. zumos
2. juices
1. de
2. of
1. frutas
2. fruits
1. y
2. and
1. veo
2. see.1SG
1. que
2. that
1. César
2. César
1. trajo
2. brought.IND
1. bebidas
2. drinks
1. alcohólicas.
2. alcoholic
1. ‘To the party, Juan brought fruit juices and I see that César brought alcoholic drinks.’

Second, in order to avoid discourse incoherence, the contextual wh-question was replaced by an initial locative/temporal adjunct that served the same function of contextual anchor (i.e. circumstantial frame setter). These changes are shown in the sample item in (20).

(20)	a.	[Gapping, No Embedding]
		En el bar, Pablo pidió una cerveza y Juan un whisky.
		‘At the bar, Pablo ordered a beer and Juan a whisky.’
	b.	[Gapping, Embedding Non-factive]
		En el bar, Pablo pidió una cerveza y sospecho que Juan un whisky.
		‘At the bar, Pablo ordered a beer and I suspect that Juan a whisky.’
	c.	[Gapping, Embedding Factive]
		En el bar, Pablo pidió una cerveza y me molesta que Juan un whisky.
		‘At the bar, Pablo ordered a beer and I am bothered that Juan a whisky.’
	d.	[No Gapping, No Embedding]
		En el bar, Pablo pidió una cerveza y Juan pidió un whisky.
		‘At the bar, Pablo ordered a beer and Juan ordered a whisky.’
	e.	[No Gapping, Embedding Non-factive]
		En el bar, Pablo pidió una cerveza y sospecho que Juan pidió un whisky.
		‘At the bar, Pablo ordered a beer and I suspect that Juan ordered a whisky.’
	f.	[No Gapping, Embedding Factive]
		En el bar, Pablo pidió una cerveza y me molesta que Juan pidiera un whisky.
		‘At the bar, Pablo ordered a beer and I am bothered that Juan ordered a whisky.’

Some further minor adjustments were made to the lists of embedding predicates in the second conjunct: we used 8 non-factive verbs (repeated in three different items: creer ‘believe, imaginarse ‘imagine’, parecer ‘seem’, sospechar ‘suspect’, suponer ‘suppose’, contar ‘tell’, decir ‘say’, rumorearse ‘be rumored’) and 10 factive verbs: 6 true factive (emotion) verbs (each repeated 2 times: encantar ‘like’, gustar ‘like’, horrorizarse ‘be horrified’, molestarse ‘be bothered’, preocuparse ‘be worried’, sorprenderse ‘be surprised’) and 4 semi-factive (knowledge) verbs (each repeated 3 times: comprobar ‘confirm’, observar ‘observe’, saber ‘know’, ver ‘see’).

In addition to the 24 experimental items, there were 24 filler items from an unrelated experiment, different from those used in Experiment 1. The new filler items tested the acceptability of possessive relative clauses introduced by the pronoun cuyo/a ‘of which’, as in (21a).¹⁰ The condition shown in (21b), where the relative clause was introduced by the pronoun que instead of cuya, was taken as an ungrammatical control in order to make sure that participants were doing the task carefully, and not just providing random judgments.

(21)	a.	Llegamos a la base de la pirámide cuya altura impresiona a los turistas durante la visita.
		‘We arrive at the base of the pyramid the height of which amazes the tourists during their visit.’
	b.	Llegamos a la base de la pirámide que la altura impresiona a los turistas durante la visita.
		‘We arrive at the base of the pyramid the height which amazes the tourists during their visit.’

3.3.3 Procedure

Experiment 2 was also administered via IbexFarm. Participants were instructed to read the sentences carefully and to judge their acceptability by using a 10-point Likert scale,¹¹ where 1 means completely unacceptable and 10 means completely acceptable. Participants did not have the option to go back to change a previous judgment. A yes/no comprehension question was shown after each sentence, which was used as a second measure to control that participants were performing the task carefully and thoroughly. Answers to the comprehension questions were monitored to eventually exclude unfit participants from further analyses. Prior to the experiment, participants completed a short language background questionnaire and 4 practice items. Completing the experiment took 15–20 minutes.

3.3.4 Analyses and results

The average percentage of correct answers to comprehension questions was over 75% for all participants and, therefore, none of them was excluded from the analyses.

Acceptability judgments (1–10) were entered into the same kind of mixed-effect linear regression analysis used in Experiment 1. The summary of the model’s fixed effects is given in the Appendix.

The linear mixed model revealed this time a significant main effect of Gapping (p < .05), as participants rated the Gapping conditions as less acceptable than the No Gapping conditions (mean rates: 8.62 vs. 8.85). Table 2 summarizes the mean acceptability judgments in raw scores and z-scores. These results are shown in Figure 5.

Table 2

Mean acceptability judgements for Experiment 2 (SD in parentheses).

	Gapping		No Gapping
	raw scores	z-scores	raw scores	z-scores
No Embedding	9.46 (1.21)	0.61 (0.46)	9.16 (1.79)	0.51 (0.56)
Emb. No Factive	8.98 (1.64)	0.38 (0.58)	9.10 (1.49)	0.44 (0.49)
Emb. Factive	7.44 (2.54)	–0.25 (0.91)	8.30 (2.17)	0.12 (0.70)

Figure 5

Mean acceptability judgments in z-scores for Experiment 2.

In addition to the main effect of Gapping, there was a main effect of Embedding (p < .01), as embedded configurations were judged less acceptable than non-embedded ones (mean rates: 8.45 vs. 9.31). As Figure 6 shows, this penalty for embedding varies as a function of Gapping, with embedded configurations being less acceptable with gapping than with no gapping (mean rates: Gapping Embedding 8.21 vs. Gapping No Embedding 9.46; No Gapping Embedding 8.70 vs. No Gapping No Embedding 9.16). The interaction between Gapping and Embedding was significant (p < .05).

Figure 6

Effect of Embedding.

Moreover, there was a significant main effect of Factivity (p < .001), as embedded clauses under a factive verb are less acceptable than embedded clauses under a non-factive verb (mean rates: 7.87 vs. 9.04). Crucially, as Figure 7 shows and as opposed to Experiment 1, this effect varies as a function of Gapping, with embedded constructions under factive verbs being judged less acceptable with gapping than with no gapping (mean rates: 7.44 vs. 8.30). The interaction between Gapping and Factivity was significant, too (p < .001). The interactions are plotted in Figure 9. We explored further this sensitivity to the semantic type of embedding predicates by distinguishing between (emotion) true factive verbs and (knowledge) semi-factive verbs. As Figure 8 illustrates, embedding predicates under true factive verbs were rated as less acceptable than predicates under semi-factive verbs, especially in gapping constructions (mean rates: True factive Gapping 6.86 vs. No Gapping 8.19; Semi-factive Gapping 8.05 vs. No Gapping 8.41).

Figure 7

Effect of Factivity.

Figure 8

Zoom into different kinds of factive verbs.

Figure 9

Interactions Experiment 2.

3.3.5 Discussion

Judging by the results of Experiment 2, it seems reasonable to assume that the potential confound in our Experiment 1 (i.e. mood selection) did have an effect on the results. The mood selection created a floor effect in our Experiment 1, as the indicative choice (the marked form compared to the more natural subjunctive) cancelled the effect of our experimental factors. In particular, this choice seems to have been responsible for the lack of interaction between Gapping and Embedding, and between Gapping and Factivity in Experiment 1. These interactions, however, were statistically significant in Experiment 2, where this confound was addressed.

Coming back to our two research questions, we hypothesized that embedded gapping should be acceptable in Spanish (at least with some verbs). This hypothesis was borne out by the results of Experiment 2, since we observed that embedded gapping is as acceptable as embedded non-gapping, especially under non-factive verbs. Although there is an interaction between Gapping and Embedding, embedded gapping (with both factives and non-factives combined) is more acceptable than the ungrammatical controls (this difference turned out significant, p < .001).¹² With regard to our second research question, we hypothesized that embedded gapping acceptability varies depending on the semantic class of the embedding predicate and it decreases in particular with factive verbs. Our experimental findings show that there is a clear gradience in the participants’ acceptability judgments. Factivity plays indeed a major role: embedded clauses under a factive verb are less acceptable than under a non-factive verb. In addition to this, our results show that there is an interaction between Gapping and Factivity within factive verbs, in that embedded clauses under a factive verb are more acceptable in non-gapping contexts compared to gapping contexts. Moreover, factive predicates do not have the same effect on participants’ judgments, as embedded clauses under a true factive (emotion) verb are less acceptable than embedded clauses under a semi-factive (knowledge) verb.

These results have important implications with regard to previous literature. First, they are in line with the dichotomy proposed by Hooper (1975) between semi-factive (e.g. find out, know, see, etc.) and true factive verbs (e.g. regret, forget, amuse, etc.): semi-factive verbs are much more similar to non-factive than to true factive verbs. Second, they provide an explanation to the inconsistency in the use of the diacritic marks in previous studies on Spanish embedded fragments, as shown in (9b) and (10b) above. Given the fact that in our experiments there is a clear contrast between embedded gapping under a factive (emotion) verb and ungrammatical control items, we consider that these cases are better defined in terms of semantic and discursive infelicity rather than a syntactic constraint violation.

From the results of Experiments 1 and 2 combined, we conclude that the No Embedding Constraint on gapping, postulated by Hankamer (1979), Neijt (1979), Johnson (2009; 2014) a.o., and considered to be a strong syntactic constraint specific to gapping constructions must be reconsidered. Moreover, a more general semantic constraint seems to be at work: non-factive verbs embed more easily than factive ones (Karttunen 1971; Kiparsky & Kiparsky 1971); and, within factive predicates, semi-factive verbs embed better than true factive ones. Embedded gapping and embedded coordinated clauses in general are thus sensitive to the semantic class of the embedding predicate. An approach based on acceptability, rather than one based on categorical grammaticality, seems to be a better fit to capture these effects.

4.1 Between a Small Conjunct Gapping and a Large Conjunct Gapping

In the literature on ellipsis, there are two main syntactic accounts of gapping: (i) Small Conjunct Gapping (SCG), involving a subclausal (low) coordination, and (ii) Large Conjunct Gapping (LCG), appealing to a clausal (high) coordination.

Under the SCG approach, two kinds of analyses have been explored: (ia) the first, without ellipsis, appealing to some leftward movement that combines across-the-board (ATB) movement of the shared head verb out of each conjunct and asymmetric extraction of non-shared constituents preceding the head verb out of the first conjunct (Johnson 2004; 2009), as illustrated in (22a); (ib) the second, involving ellipsis (deletion) of an approximately vP-sized constituent (Coppock 2001; Lin 2002; López & Winkler 2003 a.o.), as illustrated in (22b).

Under the LCG approach, the various analyses which have been proposed fall under two main directions: (iia) the first, involving ellipsis (deletion) of a phrasal projection of the verb, such as TP or CP (Ross 1967; 1970; Jackendoff 1971; Sag 1976; Hankamer 1979; Jayaseelan 1990; Hartmann 2000 a.o.), as illustrated in (23a); (iib) the second, proposing a construction-based analysis that appeals to a dedicated meaning-form rule, i.e. a construction, mapping a headless structure to a clausal meaning, as illustrated in (23b): Categorial Grammar (Dowty 1988; Steedman 1990; 2000; Kubota & Levine 2015; 2016), Construction Grammar (Goldberg & Perek 2018), LFG (Patejuk & Przepiórkowski 2017), HPSG (Abeillé et al. 2014; Bîlbîie 2017; Park to appear; Park et al. 2019).

The contrast in (24) below, with ungrammatical embedded gapping in English (24b), is considered as strong evidence for an SCG analysis, i.e. a low (subclausal) coordination. In SCG approaches, it is assumed that a TP from a matrix clause may not dominate a vP from an embedded one; thus, this account automatically rules out embedded gapping.

However, as shown by the two experiments presented in Section 3, the ban on embedded gapping is not as strict (and universal) as traditionally assumed. The possibility to embed gapping (as revealed by the Spanish data) is unexpected under a low analysis of gapping. The SCG account is thus not equipped to handle embedding facts.

If one wants to take into account cross-linguistic variation with respect to embedded gapping (e.g. acceptability in Spanish vs. grammatical ban in English), one possibility would be to consider a “hybrid” analysis of gapping (Erschler 2016; Fernández-Sánchez 2017), by taking into account both SCG (22) and LCG (23) structures: an SCG analysis for non-embedded gapping, and an LCG analysis for embedded gapping. It is worth mentioning that the same “hybrid” (“two-source hypothesis”) approach has recently been proposed by Potter et al. (2017), based on scope ambiguities in gapping, a semantic aspect which has long presented a challenge to theories of the syntax-semantics interface. According to Potter et al. (2017), scope ambiguities in gapping systematically rely on a structural ambiguity: scope-taking elements (e.g. modal auxiliaries, sentential negation, modal and frequency adverbs) that occupy a position above the vP domain coordinate structure scope over this vP domain, yielding a wide scope above the coordination, whereas scopal elements occurring within a CP domain coordinate structure have a distributive scope under the coordination. The two interpretations which are available for the example (25) in English, displaying the modal auxiliary can and the sentential negation not, are derived by Potter et al. (2017) from two different syntactic configurations: in (25a), the wide scope is linked to an SCG analysis, whereas in (25b), the distributive scope comes with an LCG analysis.

Therefore, gapping is considered to be, in general, syntactically ambiguous between two structures: vP-domain sized coordinate structure (which handle the wide scope of semantic operators) and CP-domain (clause-sized) coordinate structures (which account for the distributive scope of semantic operators). From this perspective, homogeneous-SCG and homogeneous-LCG analyses face difficulties in accounting for this scope ambiguity in gapping: the SCG approach cannot handle the distributive scope, while the LCG approach cannot deal with the wide scope. Our embedding facts (in particular, the (un)availability of embedded gapping across languages) seems at first glance to favor this kind of “hybrid” syntactic approach. This is the position adopted by Erschler (2016) and Fernández-Sánchez (2017), according to whom cross-linguistic differences (e.g. between English and Spanish in our case) with respect to embedded gapping would reflect a syntactic ambiguity in the size of coordinated categories. An SCG approach would account for non-embedded gapping and wide scope of semantic operators, while an LCG approach would account for embedded gapping and distributive scope of semantic operators.

Coming back to our Spanish experimental data, as we have already mentioned, the SCG analysis cannot be the source of gapping in Spanish, as the embedding of a gapped structure under a clause-embedding predicate and complementizer is incompatible with such an analysis. The only option available for these embedding cases remains the LCG. A semantic correlation (in line with Potter et al. 2017)¹³ would be the unavailability of wide-scope reading in these contexts, which seems to be the case. If in the non-embedded gapping in (26), both distributive-scope (26a) and wide-scope (26b) readings are available, in the embedded gapping in (27a), only a distributive-scope reading is available (27b).¹⁴

This would suggest that at least embedded gapping invokes an LCG in Spanish, but does not necessarily rule out the possibility to have a second source of gapping (i.e. SCG) outside embedding contexts (in Spanish or English). However, there is no empirical motivation for an SCG analysis even outside embedded gapping configurations. Despite its semantic motivation (i.e. wide scope reading), the SCG account still faces many difficulties. Below, we give some empirical evidence against SCG both in Spanish and English.

In SCG approaches, verb movement is accompanied by an asymmetric extraction of the non-shared constituents (generally, the subject) preceding the head verb out of the first conjunct only. As Abeillé et al. (2014) and Bîlbîie (2017) show, this assumption makes wrong predictions with respect to the distribution of initial conjunctions in Romance languages. The SCG approach wrongly predicts that initial conjunctions, such as o ‘either’ in Spanish, which mark the left edge of the first conjunct in Romance (Bîlbîie 2008), should occur after the alleged moved material (28b), and not before (28a).

Moreover, Kubota & Levine (2016) signal an empirical problem related to the distribution of the adverb merely in English, which is a strictly vP adjunct (compare (29a) and (29b)). In gapping contexts, it should precede the putative vP in the gapped clause in the case of SCG (30a), contrary to facts (30b). This behavior is quite unexpected under a vP-coordination approach and, thus, constitutes evidence of a clausal size for the gapped conjunct.

Another difficulty for the SCG approach is its incapacity to account for cases that involve elements that are moved to or base-generated in the left periphery of CP, such as topicalized elements or fronted wh-words (Repp 2009; Kubota & Levine 2016). As the gapped clause in SCG is stipulated to contain only an untensed low vP projection, there are no landing sites for these fronted elements,¹⁶ traditionally analyzed as being above TP projection. This is illustrated in (31) for English and in (32) for Spanish.

The “hybrid” analysis of gapping (à la Potter et al. 2017) predicts a correlation between the coordination size and the interpretation of scope-taking elements. In this approach, some syntactic contexts condition wide scope only, such as contexts where there is a constituent at the left periphery of the left conjunct which is shared by both conjuncts (in other words, it does not have any correlate in the gapped conjunct). Potter et al. (2017) argue that, in such a context as in (33a), an LCG analysis is unavailable, gapping applying to a structure where the remnant topicalization is unavailable (since the CP-Top position, receiving topicalized elements, is already occupied by the left peripheral PP); consequently, only a wide-scope reading is available. However, as Park et al. (2019) show, there are similar contexts where this prediction is not borne out. A context such as (33b), where there is a wh-extracted element at the left periphery and a negation preceding the putative vP coordination, easily gives rise to a distributive-scope reading.

Furthermore, Park (to appear) discusses the example in (34a) from López & Winkler (2003), where the negation has wide scope over both coordinate structures (cf. the topicalized NPI at any time); nevertheless, a vP-coordination structure is implausible on the general assumption that English topicalization targets the left edge of CP. So, if one has to admit an LCG approach, we observe that the putative full counterpart (with syntactic reconstruction of the verb in the gapped clause) is ungrammatical (34b). Therefore, if one has an LCG structure, the gapped clause has a fragmentary structure, without verbal reconstruction.

Another empirical problem for the SCG approach (signaled by Johnson 2009 himself) concerns cases such as (35) in English and (36) in Spanish with negation having scope over the first conjunct only (narrow scope of negation, cf. Repp 2009). The SCG approach which appeals to an ATB verb movement (Johnson 2004; 2009) does not stipulate that the negation could move from the first conjunct exclusively, the second conjunct being interpreted as positive.

An additional theory-internal problem for the SCG approaches which appeal to a move-and-elide operation (Coppock 2001; Lin 2002; López & Winkler 2003, etc.) is related to cases of gapping with multiple remnants. As Kubota & Levine (2016) show, in an example such as (37), three complements of the verb bet must move out of the VP in order to survive ellipsis. However, a common assumption on the Heavy NP Shift movement is that shift movement cannot apply iteratively in a single VP; thus, there are no rightward adjunction sites for all of these remnants.

All the evidence presented here suggests that an SCG analysis creates more problems than it solves, irrespective of the cross-linguistic variation (between Spanish and English) with respect to embedded gapping. As illustrated above, such an analysis is problematic not only for embedded gapping contexts, but also for non-embedded gapping cases. The “hybrid” analysis of gapping is thus not on the right track.

Therefore, an LCG analysis of gapping seems to be a better fit to account for all these facts. The widespread LCG analysis (as well as SCG) assumes movement of the remnants to the left periphery, prior to ellipsis (PF-deletion): Farudi 2013¹⁷; Boone 2014; Weir 2014, etc. However, as discussed at length by Culicover & Jackendoff (2005), Abeillé et al. (2014) and Bîlbîie (2017), extraction of remnants in the left periphery is not empirically supported: in particular, contrary to what has been claimed since Ross (1967), remnants do not obey island constraints in English (Culicover & Jackendoff 2005; Kubota & Levine 2016) or Romance languages (Abeillé et al. 2014; Bîlbîie 2017). When the gap contains more material than the head, remnants can occur in what would be an island for extraction. We illustrate these syntactic island violations with “propositional” islands (that contain an embedded clausal domain). In English, remnants can appear in sentential adjuncts (38a–b) or relative clauses (38c–d).¹⁸ The same propositional island violations apply to gapping in Spanish: sentential adjuncts (39a), relative clauses (39b–c), and also sentential subjects (39d–e). In addition, remnants extraction to the left periphery has to be multiple in the case of gapping (which requires at least two remnants in the gapped clause); yet, this specific multiple extraction cannot apply to the ungapped counterparts (Culicover 2009; Park to appear). We thus observe that there is no parallelism between the distribution of gapping remnants and extraction dependencies.¹⁹ Consequently, LCG analysis appealing to a move-and-elide approach, where the remnants escape ellipsis via (multiple) leftward movement is not empirically supported.

Furthermore, a major argument which is usually mentioned in favor of a syntactic reconstruction mechanism in gapping constructions is supported by the so-called connectivity effects, i.e. a structural parallelism between the source and the gapped clause, with respect to morpho-syntactic properties of remnants (e.g. Hartmann 2000). A challenging example for this kind of approach is the attested occurrence in (40), where the syntactically marked remnant a él has as correlate a preverbal clitic in Spanish. Any LCG approach based on syntactic reconstruction faces difficulties with respect to syntactic mismatches.

Lastly, any LCG account appealing to a syntactic reconstruction mechanism is challenged by cases where gapped clauses do not have the same distribution as their complete counterparts: a gapped clause may be introduced by functors such as constituent negation adverbs (e.g. and not in (41a)) or lexicalized comparative connectives (e.g. as well as in (41b)), which do not allow verb reconstruction (Culicover & Jackendoff 2005; Abeillé et al. 2014; Bîlbîie 2017). Similarly, in Spanish, if the gapped conjunct contains a constituent negation (42a) or if it is introduced by the corrective conjunction sino ‘but’ (42b), verb reconstruction is impossible.²⁰

Such empirical data stand as a serious challenge for any structural variant of LCG (namely, move-and-elide approaches, based on syntactic reconstruction).

On the other hand, we can still have a uniform treatment for both non-embedded and embedded gapping, if we adopt a fragment-based analysis (with semantic reconstruction of ellipsis), as proposed by Abeillé et al. (2014) and Bîlbîie (2017) for gapping in Romance, and by Ginzburg & Sag (2000) for fragments in general. They propose a construction-based analysis which does not derive the unusual meaning/form mapping in the gapped clause from hidden syntactic structure, i.e. a “what you see is what you get” syntactic structure, as schematized in (23b) above.

The main syntactic constraint in this kind of approach formalizes Hankamer (1971)’s “Major Constituent Condition”, namely each remnant in the gapped clause must be paired with some “major” correlate in the source, and hence must match a possible subcategorization of the verbal predicate in the source.²¹ This syntactic constraint accounts for the contrast in (43): the ungrammaticality of (43b) is explained by the fact that the PP in good spirits cannot be subcategorized by the verb become (seel also the contrast in (43c)). Otherwise, remnants can have a different category (and even a different word order) from that of their correlates in the source clause, as illustrated in (43a), where the second contrastive pair contains the AP crazy and the NP an incredible bore.

The syntactic structure of a gapped clause is a fragment containing a cluster of (at least two) remnants. If there is no head verb in the gapped clause, the category of the fragment is thus appropriate for combination with functors selecting some non-finite constituent. Remnants are constrained to unify their HEAD features with the HEAD features of some contextual correlates, using the context SALIENT-UTTERANCE introduced by Ginzburg & Sag (2000).

At the semantic level, each remnant must be in a contrastive relation with a correlate in the source (in each contrastive pair, remnant and correlate belong to the same set of alternatives and at the same time they display a semantic opposition, neither of them subsuming the other, cf. a.o. Sag 1976; Hartmann 2000; Repp 2009). The semantic reconstruction is operated by building the content of the fragment from the meaning of the source clause, the remnants and their correlates. An exhaustive proposal (accounting for scope ambiguities in gapping) is given in Park et al. (2019) and Park (to appear), in terms of a semantic underspecification-based analysis (by using Lexical Resource Semantics) linked to a single, uniform syntactic structure. Unlike Potter et al. (2017) who assume that scope ambiguity (wide vs. distributive scope) has a syntactic source and reflects an ambiguity in the size of coordinated categories, Park et al. (2019) and Park (to appear) convincingly show that one can derive both wide and distributive interpretations from exclusively large conjunct structures (i.e. a homogeneous LCG approach), if one adopts a framework with a flexible syntax-semantics mapping. The latter approach attributes to gapping (and to coordination, in general) an underspecified meaning, and the scope ambiguity arises as the result of different ways of specifying it: the conjuncts denote either tensed propositions (and in this case, only the distributive-scope reading arises) or eventuality descriptions (and in this case, only the wide-scope reading is available).

Lastly, at the discourse level, some symmetric discourse relation must hold between conjuncts (Levin & Prince 1986; Kehler 2002). Therefore, gapping seems to be perfectly compatible with resemblance relations (where events are interpreted as independent from one another, e.g. parallelism and contrast), but infelicitous with asymmetric discourse relations, such as cause-effect relations. Furthermore, as proposed by Steedman (1990), Reich (2007), etc., gapping can be reduced to the question/answer relation: the discursive pattern in a gapping construction would be a pair-list answer to an implicit (multiple) wh-question. Moreover, both source and gapped conjuncts involve the same Question Under Discussion (QUD).

The overall gapping construction is a particular type of asymmetric coordination with the main clausal conjunct being non-elliptical and verbal, and the gapped clause fragmentary and non-verbal (Abeillé et al. 2014; Bîlbîie 2017).

Furthermore, unlike the SCG approach, our fragment-based analysis correctly predicts the distribution of double conjunctions, such as o…o… ‘either…or…’ in Spanish, which necessarily occur at the left edge of the conjunct (cf. (28a) above).

Coming back to our specific topic, namely the availability of embedded gapping, as we mentioned above, we do not exclude cross-linguistic differences (e.g. Spanish vs. English), but we strongly assume that the No Embedding Constraint is not universal. Our construction-based approach could easily handle this potential cross-linguistic variation by using the syntactic feature IC (Independent Clause): therefore, for a language such as English, which disallows embedded gapping, syntactic properties of gapping will include the feature [IC +], indicating that the gapped clause must be an independent clause; on the other hand, for a language such as Spanish, which allows embedded gapping, the formal analysis will not include such a syntactic constraint, the gapped clause behaving either as an independent clause or as a complement of an embedding predicate.

Nevertheless, much work remains to be done in order to refine these cross-linguistic differences. Though the classical literature on gapping in English assumes a grammatical ban with respect to embedded gapping, experimental evidence shows a more nuanced picture, confirming Weir (2014)’s intuitions. Bîlbîie et al. (2019) show, based on acceptability judgment tasks, that embedded gapping in English is affected by (i) the semantic class of the embedding predicate (though embedded gapping is indeed dispreferred even with non-factive verbs, there is a general sensitivity to factivity: namely, embedded clauses under a factive verb are less acceptable than under a non-factive verb), and by (ii) the presence/absence of the complementizer (the absence of complementizer renders embedded gapping more acceptable). These results seem to show that, from a cross-linguistic perspective, a more robust and stronger syntactic constraint would be related to the presence (in Spanish) and absence (in English) of the complementizer in embedded fragments rather than postulating a very rigid syntactic No Embedding Constraint. An exhaustive cross-linguistic analysis of embedded gapping and fragments in general is beyond the scope of this paper.

4.2 Explanations for the semantic (non-factive vs. factive) asymmetry

We now come back to the contrast we observed in acceptability judgments between non-factive and factive verbs in embedded gapping. Previous accounts on embedded fragments (Weir 2014; Fernández-Sánchez 2017) derive the contrast between the non-factive and factive verbs from their different syntactic structures, following ideas which are widespread in the generative literature, namely that non-factive verbs would display a more complex syntactic structure than factive ones (Haegeman 2006; de Cuba 2007; de Cuba & Ürögdi 2010; Haegeman & Ürögdi 2010, etc.).²² Complements of non-factive verbs are therefore supposed to be syntactically bigger than complements of factive ones: there are two complementizers (a “higher” and a “lower”) in the clausal complement of non-factive predicates, as illustrated in (44a), but only one (a “lower”) in the clausal complement of factive ones, as observed in (44b).²³

All previous accounts base their data on syntactic explanations. Besides its popularity, the size-of-complement approach faces several challenges raised by our experimental results. First, it does not make a distinction within the class of factive predicates (between true factive and semi-factive ones); there are only two syntactic structures for non-factives and factives respectively. In order to account for the continuum we observed along three different semantic classes (namely, non-factive, semi-factive and true factive), we would need an additional third syntactic structure for complements embedded under a semi-factive predicate. The syntactic approach is thus not well equipped to account for in between the factive and non-factive complementation types. Second, the size-of-complement approach does not account for the cross-linguistic variation within the same semantic class; we have seen that embedded gapping under non-factive verbs presents a cross-linguistic difference: in English, embedded gapping is dispreferred even with non-factive verbs, whereas in Spanish embedded gapping under a non-factive verb is quite acceptable. If a syntactic approach is involved, the structures proposed to deal with non-factive predicates should differ from one language to another, in order to account for this asymmetry within non-factives. Third, a more general problem of the syntactic approach modelling semantic distinctions is the fact that it predicts a categorical contrast (grammatical vs. ungrammatical) in participants’ judgments, whereas a non-syntactic approach allows more flexibility/gradience in acceptability judgments, as reflected by our experimental data. Lastly, if one has to assume a syntactic approach, our experimental results seem to go against the classical view (cf. Haegeman 2006) that factives are less complex than non-factives and would rather conform to the opposite view (cf. Kiparsky & Kiparsky 1971), according to which non-factives are less complex than factives.

In this paper, we have shown that there is no universal syntactic No Embedding Constraint (at least in some languages, such as Spanish). Unlike Fernández-Sánchez (2017), who adopts Haegeman (2006)’s analysis, there is no need to postulate two different syntactic analyses for the two semantic classes. We can have a uniform syntactic analysis for both non-factives and factives, as the account we summarized above in Section 4.1, which does not postulate an isomorphism between syntax and semantics. The effects we observe in syntax across semantic classes could come from other linguistic levels. Therefore, the differences we observed with respect to embedding in general and with respect to embedded gapping in particular could not come from their syntax, but rather from the semantic and discursive properties they have by themselves or in interaction with coordination and ellipsis.

The general preference for non-embedded items is quite expected, since complex sentences generally get lower acceptability ratings than simpler ones. The difficulty with coordinating a simple clause and a complex clause may result from a more general parallelism constraint on coordination (Frazier et al. 1984; Frazier et al. 2000). According to Frazier et al. (2000) and Carlson (2001), processing of coordinated clauses is facilitated by syntactic parallelism between the conjuncts, in particular by the parallelism of the internal structure of the conjuncts: the second conjunct in a coordination is processed faster when that clause is structurally parallel to the first clause than when it is not. More generally, these studies show that there is a facilitation bias associated with less complex structures.

Despite cross-linguistic variation with respect to embedded gapping (languages which allow vs. languages which strongly disprefer embedded gapping), a notable data point concerns a general decreasing acceptability across three semantic classes of predicates, which is observed not only for Spanish, but also for other languages (Romanian, French, English, cf. Bîlbîie et al. 2019). Both in gapping and non-gapping structures, embedding under a non-factive verb is more acceptable than under a semi-factive verb and embedding under a semi-factive verb is more acceptable than under a true factive verb. In order to account for this general tendency, we have to consider some discursive explanations.

The penalty on factive verbs may have two sources. Firstly, it may come from the non-assertive nature of factive predicates (Hooper & Thompson 1973; Hooper 1975). By definition, the complements of factive verbs are presupposed and are therefore backgrounded. On the other hand, the complements of non-factive verbs contain the foregrounded information, they are asserted (i.e. they are not presupposed nor backgrounded). The discourse function of the embedded clause would thus play a role in our embedding phenomena.²⁴ The generalization is that embedded assertions (under non-factive verbs) are more acceptable than embedded presupposed clauses (under factive verbs). According to Hooper & Thompson (1973), this is because it is inappropriate to emphasize elements of a sentence whose proposition is already known, whose truth is presupposed. On the other hand, this is due to the abstractness of non-factive verbs: in the literature (Deane 1992; Van Valin 1995; Goldberg 2006; Ambridge & Goldberg 2008), non-factive verbs such as think are semantically light or neutral, making little contribution to the pragmatics of the assertion (i.e. they have a parenthetical status).²⁵

However, this is not a satisfactory explanation for the behavior of semi-factive verbs, which are presupposition triggers, as any other factive predicates. We propose here a revised explanation, by adopting the distinction between assertive and non-assertive context change-potential (CCP), proposed by Farkas (2003) to explain the mood alternation in Romance embedded clauses. Non-factives, as well as semi-factives, are strong intensional predicates (in terms of Farkas 2003), which are analyzed as assertives: their complements have assertive CCP. Common to main assertions and the complements of these two kinds of predicates is that they are assertively added to an epistemic context. Unlike non-factives and semi-factives, true factive (emotion) predicates are non-assertives: their complements have evaluative CCP. A formal characterization of the three semantic classes of predicates (non-factives, semi-factives and true factives) is given in Table 3.

In the case at stake, the difference in acceptability could be explained by a mismatch in terms of the Assert feature. An embedded conjunct under a non-factive or a semi-factive verb will be [Assert +]; therefore, it will involve an assertive component, exactly as the non-embedded first conjunct. On the other hand, an embedded conjunct under a true factive verb will be [Assert –]; thus, it will involve an evaluative component, unlike the first conjunct, which is [Assert +]. The generalization – based on the notion of assertion – will be: the more asserted the complement is, the easier it is embedded.

Farkas (2003: 18) notes that “there are various other areas where [the distinction between assertive and non-assertive CCP] might prove useful”. Our experimental data show that this distinction is indeed useful not only in explaining the mood alternation in embedded clauses, but also in explaining the gradience in acceptability judgments with respect to coordination and ellipsis phenomena. If we combine Farkas (2003)’s and previous insights in the literature, we can easily account for the decreased acceptability across these three semantic classes of predicates embedding the second conjunct in coordination contexts.

Secondly, the penalty on factive verbs may come from a discourse incongruence, namely a mismatch between the conjuncts with respect to the Question Under Discussion (QUD) each of them answers. Discourse proceeds by raising and answering implicit questions (Ginzburg & Sag 2000; Roberts 2002; Ginzburg 2012 a.o.). For our purposes, we adopt the distinction between the main point of an utterance and its backgrounded material (Simons 2007). The main point of an utterance refers to propositional content that is intended to answer the QUD, whereas backgrounded material refers to propositional content that is not directly relevant to the QUD. There is a striking difference between non-factive and factive predicates with respect to this dichotomy (cf. Anand & Hacquard 2014). Non-factive predicates can have a parenthetical status (Hooper 1975); consequently, the main point of the utterance is the content of the embedded complement, while the main clause containing the embedding predicate is backgrounded. By contrast, factive predicates (and in particular emotive predicates) systematically forefront the evaluative component (e.g. the emotive component); in these latter cases, the main point of the utterance would be the main clause containing the embedding predicate. Interestingly, Anand & Hacquard (2014) observe that occasionally semi-factives (such as know) can behave like non-factives, i.e. the content of its embedded complement becomes main point.

As for the distinction within factive predicates (between true factives – emotive verbs, and semi-factives – cognitive verbs), it is assumed that emotive factives are semantically more complex than cognitive factives (Anand & Hacquard 2014, and references therein). While this latter subtype of factives presupposes their propositional complement p, the former subtype of factives presupposes both p and the attitude holder’s belief in p. According to the classical analysis of emotive factives, these true factive predicates entail the truth of their complement, an emotional attitude of the subject towards the truth of their complement, and the subject’s belief that it is true. Thus, emotive predicates presuppose that the subject necessarily has prior knowledge of the situation described in the embedded complement, with respect to which she experiences an emotive reaction. The main point is thus about the content of a private state, while the veracity of the embedded complement is backgrounded. This semantic and pragmatic complexity, coupled with the lack of assertion, would explain why the emotive factives are rated lower than the other semantic classes of predicates. The different discursive behavior of emotive verbs compared to semi-factive or non-factive verbs may be observed in (45) below, adapted from the example (18) above. Negating the content of the embedded complement in the second conjunct is possible for a semi-factive verb such as saber ‘know’ (45b) or a non-factive verb like sospechar ‘suspect’ (45c), but not for a true factive verb such as molestarse ‘be bothered’ (45a). This contrast can be explained if we assume that the embedded complement under an evaluative (true factive) predicate cannot be the main point of the utterance.

As mentioned above, decreasing acceptability across three semantic classes of predicates is generally observed in our experiments. However, the effect worsens in embedded gapping with factive predicates, in particular emotive factives. In order to account for this worsened effect only in gapping cases, we have to look at the specific semantic and discursive constraints applying in gapping constructions. From the perspective of discourse structure, gapping is traditionally seen as related to the question/answer relation (Kuno 1976; Steedman 1990; 2000; Reich 2007; Repp 2009, etc.): a gapping construction thus is reduced to some congruent sentential (pair-lists) answers to an implicit (multiple) wh-question, which constitutes the QUD. As Steedman (1990: 248) observes, “even the most basic gapped sentence, like Fred ate bread, and Harry, bananas, is only really felicitous in contexts which support (or can accommodate) the presupposition that the topic under discussion is Who ate what?.” Therefore, in a gapping construction, the most salient QUD is reconstructed on the basis of the information structure of the source clause, and this very question has to be answered by the gapped clause too. Repp (2009) argues that the conjunction and imposes a strict discursive and semantic parallelism in gapping, which she defines in terms of balanced contrast: “In gapping, both conjuncts must make the same kind of contribution to a common discourse topic.” Repp (2009: 83).

Based on all these assumptions, we can say that in a regular gapping construction, as well as in embedded gapping with non-factive verbs, both conjuncts address a same QUD. However, in the case of embedded gapping under factive predicates, the conjuncts address different QUDs. In the Spanish example given in (18) above and repeated below in (46a), the second conjunct embedded under the factive predicate molestarse ‘be bothered’ answers a different QUD, namely no longer ‘what drink did Juan order in the bar’, but ‘what effect did it have on the speaker’, while this is not the case with non-factive verbs such as sospechar ‘suspect’ in (46b), where the second conjunct still answers the same QUD as the first conjunct, the embedding predicate having in this case a parenthetical content.

The remnants of the gapped clause are positioned in discourse-prominent slots (they are contrastive foci, cf. Kuno 1976; Hartmann 2000; Johnson 2014 a.o.). The decreased acceptability with embedding under a factive verb could be explained by the presence of conflicting constraints: it is pragmatically anomalous to treat the gapped clause as simultaneously backgrounded and discourse-prominent. Therefore, a discursive account could be a good predictor of the acceptability of embedded gapping. We recall here the cross-linguistic generalization made by Ambridge & Goldberg (2008: 375) to account for the variation with respect to island constraints: “languages appear to select different cut-off points in how backgrounded a constituent may be while containing a gap […]. Languages differ as to the location of the cut-off point […].” This generalization seems to apply to embedding gapping too. While we expect to find cross-linguistic variation with respect to acceptability across the three semantic classes presented above, we do not expect to find any language in which embedded gapping under a true factive verb is more acceptable than embedded gapping under a non-factive verb.

In Section 4.1, we argued that parallelism constraints in gapping prove stronger at the discourse and semantic level than at the syntactic level: gapping is felicitous only with symmetric discourse relations, and it involves (at least) two contrastive pairs. Though it was long believed that there is a syntactic constraint disallowing embedding in gapping constructions, we can now attribute these dispreferences to semantic and discourse factors. The advantage of postulating a non-syntactic explanation for these dispreferences is the fact that it predicts the acceptability of several embedded gapping cases, which would be otherwise ruled out by a purely syntactic account.

If one takes a closer look at the attested corpus data on Spanish embedded gapping (see the examples given in (11) above and repeated below in (47) from García-Marchena 2015, 2018), one observes that in these embedded gapping cases, either the two contrastive pairs which are involved are alternating participants (with switching roles), or there is a contrast on polarity, and not just on contrastive elements. The reinforced contrast in these complex cases shows the importance of the semantic contrast requirement on gapping. If this contrast requirement is met, we predict that embedding in general could be possible, and indeed it is, as shown by the English attested data (drawn by a Google search) in (48), from Park (2016). Though gapping is traditionally considered to be restricted to coordination, we observe in (48) that gapping can involve subordination (the gapped clause can be embedded under complementizers such as before in (48a) or because in (48b)) provided that the above semantic and discourse requirements are met.²⁶

Therefore, specific constraints on gapping, corroborated with the general constraints on coordination and discourse, could explain why the acceptability of embedded gapping with factive verbs (in particular, emotive verbs) is highly decreased compared to non-gapping contexts.

In conclusion, the decrease in acceptability judgments in our experimental work should be attributed, on the one hand, to some general constraints on discourse and coordination, and, on the other hand, to specific semantic and discourse conditions on gapping constructions. Acceptability is the most decreased in embedded gapping with true factive verbs, because of the cascade effect made by the superposition of all of these constraints.

In light of the observations made in this section, we can conclude that there is no need to postulate complex (and different) syntactic mechanisms in order to derive the effects we observe with embedded gapping and/or with the semantic distinction between non-factive and factive predicates. A uniform syntactic analysis can account for both regular (non-embedded) and embedded gapping. As for the semantic asymmetry between non-factive and factive predicates, this could be explained in non-syntactic terms, by appealing to (more general) semantic and discursive constraints.