A prosodic constraint on wh-extraction from preverbal infinitival subjects

This paper introduces a series of mitigating circumstances improving the acceptability of wh-extraction from preverbal infinitival subjects in Rioplatense Spanish. It is argued that the factor behind these amelioration effects is encoded in prosodic structure, much in line with the hypothesis that certain island restrictions apply at PF. The linguistic principle accounting for the phenomenon is proposed to be a faithfulness constraint at the syntax-prosody interface stating that an extraction domain XP cannot be mapped as a prosodic word ω at PF. An alternative syntactic account based on freezing is shown to be unable to capture the relevant contrasts.


Introduction
famously noticed that ellipsis is capable of repairing island violations. Consider the contrast in (1), taken from Merchant (2008: 136). While the sentence in (1a) is ungrammatical due to a violation of a subject island, its counterpart under sluicing in (1b) is perfectly acceptable. 1 (1) a. * Which i a biography of t i is going to be published this year? b.
A biography of one of the Marx Brothers is going to be published this year. Guess which i ! a biography of t i is going to be published this year Restrictions on syntactic movement have been traditionally captured in (narrow) syntactic terms (see Newmeyer 2016, Szabolcsi and Lohndal 2017, Liu et al. 2022, and references therein). However, patterns like (1) have led to the hypothesis that certain movement restrictions are phonological in nature, e.g., Lasnik (2001), Merchant (2001Merchant ( , 2008, Fox and Lasnik (2003). The argument roughly goes as follows. Under the assumption that ellipsis involves deletion of phonological material at PF (e.g., Hankamer and Sag 1976, Sag 1976, Merchant 2001, among many others), the ameliorating effect in (1b) seems to show that the island restriction applies over the surface representation of the sentence, and not on the underlying syntactic structure. Therefore, the island violation must be computed at PF, and not during the syntactic cycle.
In this paper, we present evidence supporting the claim that some movement restrictions involve phonological factors and apply at PF. The data comes from violations of syntactic movement in Rioplatense Spanish 2 that can be seemingly 'repaired' under a heterogeneous set of grammatical factors. We claim that the common denominator connecting these scenarios is linked to prosodic constituency, i.e., the property determining the acceptability of the extraction can be tracked to the prosodic structure of the sentence. While we do not advance a fully-fledged theory accounting for the phenomenon, we do offer an explanatory conjecture. We contend that assigning the phonological properties of a word to a whole extraction domain XP prevents the possibility of interpreting a trace within XP.
The structure of the paper is as follows. In section 2, we introduce the basic data points illustrating the type of amelioration effect we are interested in. In section 3, we show that the syntactic accounts available at the moment are not capable of capturing the contrasts and that, as far as we can tell, there is no structural way to do it. In section 4, we sketch an analysis based on the observation that 'repaired' sentences lead to prosodic structures in which the extraction domain does not match a prosodic word ω in prosodic structure. Section 5 contains the conclusions.

The data
Our empirical domain consists of infinitival clauses functioning as subjects of completive clauses in Rioplatense Spanish. As the pairs in (2) and (3) illustrate, these subjects can appear both in preverbal or postverbal positions.
(2) a. Creo  However, an asymmetry arises if a constituent is wh-extracted from these subjects: the extraction is grammatical only if the infinitival TP is postverbal. This type of contrast has been well-attested in Spanish; see Haegeman et al. (2014: 102-118)  While wh-extraction yields ungrammatical results for the preverbal subjects in (4b) and (5b), we observe that, for many speakers, there is a series of mitigating circumstances improving the acceptability of these sentences. To begin with, as first noticed by Verdecchia (2018), the relevant extraction becomes possible if additional material, e.g., an adjunct PP, appears together with the infinitive within the subject TP. Compare (4b) to (6). The only difference between these sentences is that the latter has the locative adjunct PP en el secundario 'in high school' modifying the infinitive leer 'to read', while the former has a bare infinitive as the sole overt constituent within the clausal subject. This slight difference, however, introduces a stark contrast in acceptability for us and most (but, certainly, not all) native speakers consulted. The same amelioration effect obtains with example (5b) if an adjunct PP such as en las farmacias 'in the pharmacies' is added within the subject infinitival clause. This is shown in (7) A second amelioration factor we find in these contexts involves introducing a parenthetical adjunct separating the subject TP and the (embedded) finite verb. Thus, for instance, if a parenthetical like de acuerdo con tu experiencia 'according to your experience' is inserted between the infinitive comprar 'to buy' and the inflected verb causará 'will cause' in (4b), the acceptability of the sentence improves, as the vast majority of the informants who share the contrast in (4b) and (5b) report. Once again, the same amelioration effect can be obtained for (5b), as exemplified in (9). In this case, the parenthetical sin lugar a dudas 'without any doubts' separates the infinitive comprar 'to buy' and the inflected verb estaba 'was'.  The acceptability of the extractions in (4b) and (5b) also improves if the stranded infinitive is interpreted as contrastive focus and receives the corresponding intonation.
Consider first the dialogue in (10). In the answer in (10B), the infinitive leer 'to read' functions as contrastive focus and receives an emphatic nuclear accent. The example is acceptable to our ears, despite the fact that it exhibits the same type of wh-extraction that makes (4b) unacceptable. In contrast, this amelioration is not attested if the contrastive focus is on another constituent, such as problemas 'problems' in (11B). This contrast was confirmed by many of the speakers who also share the judgements regarding (4b) and (5b). Yet, it is worth noting that this effect is less systematic than the previous ones. A fourth and final amelioration effect is triggered by ellipsis; as discussed in the introduction, ellipsis is known to repair island violations in several contexts (Ross 1969, Merchant 2001. As shown in (14), ellipsis also repairs the unacceptability of (4b As mentioned in the introduction, the fact that ellipsis grants a 'repairing effect' in island violations has been used to argue that movement restrictions are phonological in nature. In section 4, we explore this hypothesis in order to offer a unified explanation of the four amelioration effects presented so far.

Structural accounts will not do
It is not clear for us how a purely structural account of islands could explain the amelioration patterns in section 2. Consider once again the sentences in (4b) and (5b). As mentioned, this movement restriction is reminiscent of more traditional subject island effects attested in Spanish. Starke (2001: 57) notices that preverbal subjects in Spanish are opaque for extraction, while postverbal subjects are rather transparent. Under this condition, the opacity of the preverbal subject is expected. That is, the subject of the sentence in (16a) occupies the [Spec,TP] position after moving there from [Spec,vP]. Since the subject already moved, wh-extraction of the PP becomes impossible. This is sketched in (18). When the subject is postverbal, as in (16b), it occupies its base position in [Spec,vP]. Thus, wh-extraction in this case is allowed, as shown in (19). There is a further argument against an explanation based on the Freezing Principle. Under this approach, the pattern discussed by Starke (2001) and Haegeman et al. (2014) should be sensitive to the four repair strategies discussed throughout section 2. However, the ungrammatical sentence in (16a) cannot be improved by adding an adjunct PP within the subject, e.g., (22), or by focusing the noun libros, e.g., (23). This shows that the data discussed in this paper do not form a natural class with more traditional patterns of subject islands. Another possible explanation for subject island phenomena stems from the assumption that specifiers are opaque for extraction, e.g., Huang (1982), Nunes and Uriagereka (2000), among others. This approach fails at capturing the relevant contrasts, as it predicts that all instances of subject extraction should be ungrammatical. 5 In conclusion, a purely structural approach cannot account for the amelioration effects attested in section 2. In the following section, we contend that prosodic structure provides a way of making sense of these patterns, at least in a preliminary fashion.

The conjecture: prosodic constituency and traces
We contend that the restriction attested in (4b) and (5b), and the ameliorating effects displayed by these patterns in (6) to (15) are related to the underlying prosodic structure of the sentence. As is known, the prosody of a sentence obtains from a series of hierarchically arranged constituents of phonological nature: syllables are organized into feet, feet are organized into prosodic words, prosodic words are organized into phonological phrases, and phonological phrases are organized into intonational phrases, e.g., Selkirk (1984), Nespor and Vogel (1986), Truckenbrodt (2007), Féry (2016). The resulting hierarchy is usually referred to as prosodic hierarchy. (24) We are concerned with two types of constituent in this hierarchy: phonological phrases ϕ and prosodic words ω. These units can be taken to be phonological counterparts of constituents in the syntactic representation. That is, phonological phrases relate to (lexical) syntactic phrases such as VPs, NPs or APs; prosodic words, on the other hand, relate to (lexical) syntactic heads: they typically realize lexical words together with their adjacent functional elements, which (in general) do not form prosodic words of their own. Following Selkirk (2011), we assume that syntactic and prosodic structures are, in principle, isomorphic. The constraints in (25) and (26)  These constraints are violable, i.e., they may be overridden due to a number of languagespecific preferences. This introduces mismatches in the mapping between syntax and prosody.
We believe that the generalization in (27) allows for the unification of the amelioration effects discussed in the previous section. We conceive (27) as a descriptive statement.
(27) Wh-extraction from a subject infinitival clause S inf in preverbal position leads to unacceptability if S inf has been mapped into a prosodic word ω that is immediately dominated by the phonological phrase φ containing the VP.
According to (27), the deviation of (4b) and (5b) is due to properties of their prosodic structures. Our claim is that in both cases the infinitive heading the infinitival subject is interpreted as a phonological word ω at PF. 6 As such, it is mapped into a single phonological phrase (S inf VO) φ together with the rest of the embedded clause, e.g., (28a) and (28b). Due to reasons we will discuss below, this underlying prosodic structure is incompatible with extraction of constituents from within the infinitival phrase.
What the four ameliorating scenarios introduced in the previous section have in common is that they prevent the generation of prosodic structures like (28a) and (28b). It goes without saying that ellipsis (i.e., phonological deletion) would prevent the violation of (27), so we will focus on the other three contexts. In these, the subject infinitival clause is mapped as a separate phonological phrase φ 1 from the phonological phrase φ 2 containing the inflected verb (and the rest of the predicate), i.e., a prosodic structure of the form (S inf ) φ 1 (VO) φ 2 obtains. As a consequence, extraction from S inf is permitted in these cases.
Let's consider first the cases in which an additional constituent within the infinitival clause rescues the wh-extraction violation. These are the cases presented in (6) and (7), repeated in schematic form together with their contrasting pairs in (29b) (27), for the structures in (29b) and (30b) to allow extraction from the infinitival TP, these representations must be mapped into the prosodic structures in (31), with the infinitive forming a phonological phrase φ 1 together with the adjunct PP.
We take that the relevant factor leading to these prosodic structures is syntactic branching within the infinitival subject, i.e., the fact that the subject consists of two prosodic words ω. Based on an acoustic analysis of reading data, Elordieta et al. (2003Elordieta et al. ( , 2005 6 An anonymous reviewer asked whether the effects hold with periphrastical verbs. As the example in (i) shows, the complexity of the verbal form does not make any difference.  This can be captured within our analysis: under the assumption that these predicates form a complex head in syntax, they are mapped into a single prosodic word ω. Thus, the generalization in (27) holds. and D' Imperio et al. (2005) observe that European Spanish displays the tendency to divide simple SVO declarative sentences into (S)(VO) prosodic phrasings (e.g. (32a)); this preference, however, becomes the rule if the subject has a branching structure. Gabriel et al. (2011) report, also based on reading data, that the overall proportion of (SVO) (e.g., (32b)) groupings is higher in Rioplatense Spanish. This result, however, follows from a higher proportion of non-branching subjects leading to (SVO) phrasings in Rioplatense; branching subjects, on the other hand, categorically lead to (S)(VO) (e.g., (32c)). Thus, syntactic branchingness in the subject must be considered a predictor of (S)(VO) prosodic structures in both European and Rioplatense Spanish. Gabriel et al. 2011 provide a syntactic motivation for the general tendency attested in (European) Spanish to separate subjects and predicates in distinct intonational units. They follow Ordóñez (1997) and Ordóñez and Treviño (1999) in postulating that preverbal subjects in Spanish are dislocated in the left periphery. The basic idea is that just like the DP el auto 'the car' is a dislocated DP doubled by a clitic within the clause in (33a), the preverbal subject Cosmo in (33b) is also dislocated and doubled by the verbal inflection.  Elordieta et al. (2005) propose that a phonological phrase boundary appears at the left of the extended projection of the VP, i.e., at the TP level. Since preverbal subjects in Spanish are dislocated constituents above this position, they are expected to form independent intonational units from the rest of the sentence. According to these authors, this explains the general preference for (S)(VO) phrasings in Spanish. They further argue that syntactic branching within the subject is an independent condition that applies on top of this preference, increasing even more the need to separate the subject from the rest of the clause.
Coming back to the representation in (31a) and (31b), we argue that syntactic branching is the only factor leading to these (S)(VO) groupings. This follows from the fact that clitic left dislocation does not apply to infinitival clauses in Spanish, e.g., (34 Since the infinitival clauses in (29) and (30) cannot be taken to participate in a configuration parallel to (33b), they must be taken to occupy the [Spec,TP] position when functioning as preverbal subjects. Thus, according to the analysis by Elordieta et al. (2005), they are predicted to be grouped together with the predicate unless they have a branching structure. This explains why the patterns in (29a) and (30a) are mapped as in (28a) and (28b), respectively, while (29b) and (30b) have the prosodic structures in (31a) and (31b).
Let's move to the repairing effect triggered by focus. As discussed, if the infinitive is interpreted as a contrastive focus and receives the emphatic nuclear accent, the resulting sentence is acceptable. These are the examples presented in (10) and (12), repeated in schematic form together with their contrasting pairs in (35b)  In these cases, we take that focus must be aligned with a φ-boundary (Truckenbrodt 1999, Féry 2013. This is attested in Rioplatense Spanish, where non-final foci display a boundary L-to their right (e.g., Gabriel 2010). Therefore, the abstract syntactic representations in (35b) and (36b) are expected to be mapped into the prosodic structures in (37a) and (37b), respectively.
Once again, these examples comply with the generalization in (27). That is, since in neither (35b) or (36b) the infinitival TP is mapped into a prosodic word ω within the same phonological phrase φ as the VP, wh-extraction from these syntactic domains is permitted.
The final amelioration effect to discuss is the one granted by introducing a parenthetical adjunct just after the infinitival subject. The relevant examples are those presented in (8) and (9), which we repeat in (38b) and (39b), respectively, in schematic form and together with their contrasting pairs. In this case, the analysis is very straightforward. As Dehe (2014: 31) puts it, "there is wide agreement that intonation is a defining feature of parentheticals, and that parentheticals are prosodically separate and independent from their host." This is also the case in (Rioplatense) Spanish, as the commas in the examples attempt to show. Thus, we take that (38b) and (39b) produce the prosodic representations in (40a) and (40b), respectively.
As the preceding discussion shows, the common denominator connecting all ameliorating conditions is that all of them produce a prosodic representation that prevents the scenario described by the generalization in (27); we include here the repairing effect granted by ellipsis in (14) and (15), since we assume that phonological deletion trivially leads to the satisfaction of (27). While this uniformity in the pattern strongly suggests that the generalization is on the right track and there are prosodic conditions affecting the acceptability of wh-extraction in these examples, it is far from obvious why this should be the case. That is, there is no theoretical principle providing a rationale for the restriction stated in (27) to be true. In what follows, we offer an explanatory conjecture attempting to derive this.
We may take the generalization in (27) as describing the effects of a faithfulness restriction in the mapping between syntax and prosody. That is, much in the spirit of Match Theory (Selkirk 2011), we posit a constraint requiring a strict correspondence between syntactic and prosodic units. Our constraint, however, suggests that the syntax-prosody interface involves matching requirements that are specific for certain grammatical processes. Syntactic movement is one of these processes.
As discussed, the relevant syntactic configuration involves an infinitival TP from which a constituent is extracted. If the extraction domain TP is mapped into a prosodic word ω, e.g., (28a) and (28b), movement produces an ungrammatical outcome. However, if the extraction domain TP is mapped into a phonological phrase φ, e.g., (31a), (31b), (37a), (37b), (40a) and (40b), the result becomes acceptable. Thus, we contend that the constraint at play in these configurations roughly looks like (41).
(41) An extraction domain XP cannot be mapped as a prosodic word ω.
In practice, the constraint in (41) forbids assigning the phonological properties of a word to a phrase XP consisting of a head X 0 and a trace. One could argue that the constraint in (41) is superfluous, as the result in (42) should follow from the constraint in (25), i.e., Match Phrase. That is, an extraction domain XP cannot be mapped into a prosodic word ω because it should rather be mapped into a phonological phrase φ.
However, the amelioration effects discussed throughout this section cannot be accounted for by appealing to (25). Match Phrase establishes a correspondence between an XP in syntax and a φ in prosody without taking into consideration whether XP is or is not an extraction domain. Thus, just as it could be used to derive the prohibition sketched in (42), it would also make the prediction in (43), in which an XP consisting only of its head cannot be mapped into a prosodic word.
The effect of this prohibition is not attested in the configurations we have been discussing. The sentences in (44) also have clausal subjects consisting of a bare infinitive. Arguably, these examples can be analyzed in terms of the prosodic representations in (28), in which the infinitival clauses headed by leer 'to read' and correr 'to run' are mapped into prosodic words ω at PF. 8 If the contrasts discussed throughout this section were to be explained as violations of Match Phrase, these sentences should be as ungrammatical as (4b)  In other words, a principle like Match Phrase misses the basic observation that the XP triggering the whole pattern contains a trace. Thus, the phenomenon requires an explanation connecting prosodic structure and syntactic movement, just as (41) does.
We believe that the constraint in (41) can be motivated in terms of linguistic performance. As is known, prosody and prosodic phrasing play an important part in language processing; see Speer and Blodgett (2006) and Pratt (2018) for discussion. This role is also attested in the on-line interpretation of movement dependencies, as prosodic information is actively used to decide the positions in which traces must be interpreted (Nagel et al. 1994, Straub et al. 2001 With this background in mind, consider the following abstract scenario. Take a DP that is an argument of a head X 0 ; X 0 and the DP are the only overt constituents within XP. At some point in the syntactic derivation, DP moves to a higher position and leaves behind a trace. The resulting structure is illustrated in (45).
Consider the interpretation process of the resulting sentence if the constituent XP is mapped into a phonological phrase φ in the prosodic representation, e.g., (46). Under the assumption that the principles of Match Theory (or any other equivalent faithfulness constraint) are independently active, the prosodic structure in (46) allows the hypothesis that the DP is an argument of X 0 . That is, if X 0 is assigned the phonological properties of a phrase, an interpreter can reasonably assume that it is a phrase in the underlying syntactic structure and, as such, they can attempt to locate a trace within its domain. Thus, the DP is likely to receive a proper interpretation as an argument of X 0 .
(46) ( X 0 ) φ EXPECTATION: can host traces The situation is quite different if XP is mapped into a prosodic word ω, e.g., (47). In this case, the interpreter once again will follow the faithful correspondences of Match Theory as guiding principles, but this time they will conclude that there must be no traces associated with X 0 , so the DP cannot be an argument of X 0 . That is, X 0 is assigned the phonology of a word, therefore the interpreter expects it to be a word at the syntactic level; since words cannot host traces (Baker 1988), the expectation is that there is no trace associated with X 0 . In sum, this sort of prosodic mapping makes it more difficult to retrieve the proper interpretation of a sentence like (45).
(47) X 0 ω EXPECTATION: cannot host traces This latter scenario is prevented if the constraint in (41) is active. Thus, there seems to be a quite straightforward functional motivation for it to hold.

Concluding remarks
In this paper, we have discussed a number of grammatical contexts that seemingly 'repair' an instance of wh-movement. The pattern involves movement from within an infinitival clause functioning as a preverbal subject. If wh-extraction leaves a bare infinitive behind, the sentence is deemed ungrammatical. However, the acceptability of the extraction improves dramatically under the following circumstances: (i) if the stranded infinitive has an argument or an adjunct, (ii) if a parenthetical adjunct separates the infinitival subject from its verb, (iii) if the bare infinitive functions as a contrastive focus, and (iv) if ellipsis erases the infinitival clause.
We have discussed that a more traditional approach based on the Freezing Principle or any other purely structural means does not seem to be able to capture the relevant data. But, as far as prosodic structure is considered, all amelioration effects seemingly form a natural class. In this line, we have proposed an analysis of these cases in terms of prosodic constituency. In all cases, what seems to improve the acceptability of the sentence is the avoidance of a particular type of prosodic phrasing: that in which the infinitival subject S inf forms a single phonological phrase φ together with the rest of the sentence, i.e., (S inf VO). We have argued that this follows from a specific faithfulness constraint at the syntax-prosody interface: an extraction domain cannot be mapped as a prosodic word ω. We have motivated this constraint on performance grounds.
If our analysis is on the right track, it offers support for the claim that certain aspects of movement restrictions, i.e., islands, are computed post-syntactically at PF. Of course, this does not mean that restrictions of this sort explain every island effect. That is, we are not proposing a general theory of islands, but only a preliminary analysis of the patterns we observe in Rioplatense Spanish in contexts of wh-movement from within infinitival clauses in preverbal subject position. The extent to which this phenomenon extends to other languages or to other island domains remains a question for further inquiry.