The idea that utterances are formed of categorized prosodic constituents organized in a hierarchical structure has a long pedigree (Selkirk 1981; 1986; Beckman & Pierrehumbert 1986; Nespor & Vogel 1986; Pierrehumbert & Beckman 1988, among others). The prosodic categories assumed in this article are shown in Table 1, representing a version of the prosodic hierarchy recently argued for by Itô & Mester (2012; 2013).

Selkirk (1984) introduced the Strict Layering hypothesis (see also the references cited above), which holds that a prosodic node can dominate only nodes whose category is one step down on the prosodic hierarchy. Strict Layering rules out “level-skipping” structures like (1a) and recursive structures like (1b).

However, I follow recent developments in prosodic phonology arguing that both level-skipping and recursion are not only permitted but frequent. Recursion at the level of the prosodic word and above has been argued for by Ladd (1986); Inkelas (1989); Selkirk (1996); Wagner (2005; 2010); Itô & Mester (2009a; b; 2012) and Elfner (2012; 2015), among others. There may be constraints militating against these violations of Strict Layering (Selkirk 1996) (though see Kabak & Revithiadou 2009 for arguments against anti-recursion constraints) but they are not relevant for the analysis presented here.

Having introduced the prosodic hierarchy, we can now consider the organizing principles for how prosodic structures might correspond to syntactic structures.

Indirect reference theories, of which Match Theory is a recent iteration, hold that prosodic structure is the result of a negotiation between two competing pressures. On the one hand, there is pressure for the prosodic structure to correspond in particular ways to syntactic structure, and on the other hand there is pressure for prosodic structure to satisfy independent well-formedness conditions, which do not make reference to syntax. Sometimes these pressures come into competition, and this competition can be modelled in Optimality Theory (OT, Prince & Smolensky 1993). Note that employing OT to model syntax-prosody correspondence predates Match Theory—see Selkirk (1996; 2000) and Truckenbrodt (1995; 1999), among others.

To illustrate how OT allow us to model competing pressures at the syntax-prosody interface, consider a noun phrase consisting of a single word like the bare plural dogs. It may contain one or more phonologically empty functional heads, which project syntactic phrases, and thus have have a structure like that in (2).

Let’s assume that given the input syntactic structure in (2), there are just two candidate output structures available, shown in (3) (I assume that phonologically null syntactic categories like the null determiner in (2) are a priori excluded from mapping to prosodic constituents).1

In Match Theory, discussed in the next part of this section, syntactic phrases (XPs) are preferentially mapped to ɸs, and syntactic heads (X⁰s) are preferentially mapped to ωs. From the perspective of Match Theory, then, (3a) is the preferred candidate: in it, DP is mapped to a ɸ, whereas this is not the case for (3b).

However, there is reason to assume that single-word XPs in English are not necessarily mapped to ɸs. English ɸs are associated with particular phonetic properties—for instance, an H- or L- phrase accent at their right edge (Beckman & Pierrehumbert 1986; Selkirk 2000). There is no evidence to suggest that single-word DPs such as bare plurals or proper names behave as full ɸs, rather than as simple ωs.2

We may assume, then, that English single-word DPs do not by default map to ɸs, and that out of the two candidates in (3), the Match-violating structure in (3b) is in fact the winner. To account for this, I assume that the pressure for ɸs to be binary-branching outranks the pressure to map XPs to ɸs—see Ghini (1993); Inkelas & Zec (1995); Selkirk (2000); Elordieta (2007); Itô & Mester (2009a); Elfner (2012); Clemens (2014) and Bennett et al. (2015; 2016) for discussion of binarity in phrase-level prosody. In OT, we can embody each of these pressures in a constraint: on the one hand there is MATCH PHRASE, which enforces correspondence between XPs and ɸs, and on the other hand there is BINARITY(ɸ), which enforces binary-branching ɸs. They are defined informally in (4). In order for (3b) to beat (3a), BINARITY(ɸ) must outrank MATCH PHRASE, shown in (5).3

Having outlined the principles of indirect reference theories and constraint interaction, we can now flesh out some details of Match Theory. This sets us up for the discussion of function words in section 2.4.

Match Theory is a framework whose central tenet is that there is a pressure for certain syntactic categories in the input structure to correspond to certain prosodic categories in the output structure, and vice versa. Selkirk (2009; 2011) proposes that syntactic clauses correspond to intonational phrases (ɸs), syntactic phrases to phonological phrases (ɸs), and syntactic words to prosodic words (ωs). Following Itô & Mester (2013), I assume that a clause is a CP (or, if we are to only consider main clauses, perhaps a ForceP—see Selkirk 2009), that a phrase is an XP, and that a word is an X⁰. The correspondences assumed here are summarized in Table 2.

For each of the corresponding pairs in Table 2, there is a constraint (or pair of constraints) ensuring that a syntactic object in the input will have a counterpart prosodic object of the appropriate category in the output, and vice versa. These constraints are informally represented in (6).4

I refer the reader to Selkirk (2009) for detailed discussion of MATCH CLAUSE, and Elfner (2012) for MATCH PHRASE.

To see how these constraints might work in practice, we may assume that the NP hungry dog has the syntax in (7a), compliant with Bare Phrase Structure (Chomsky 1995). With this input structure, the maximally Match-compliant output prosodic structure would be (7b).

In (7), every X⁰ has a corresponding ω and every XP has a corresponding ɸ, and likewise every ω has a corresponding X⁰ and every ɸ has a corresponding XP. Therefore in the course of mapping (7a) to (7b), no violations of MATCH WORD or MATCH PHRASE are incurred.

However, not all X⁰s and XPs are mapped to ωs and ɸs. For instance, in the previous subsection we saw that a high-ranked BINARITY(ɸ) constraint may prevent XPs consisting of a single prosodic word from corresponding to ɸs. For the rest of this article, I focus on another case where a preferred correspondence in (6) breaks down: prosodically-reduced function words. These elements are syntactic X⁰s, so under the simplest imaginable form of MATCH WORD they should map to ωs, yet they generally map to prosodic clitics rather than independent ωs.

In the next and final part of this section, I discuss how prosodically-reduced function words have generally been approached in previous work, the dominant idea being that they are essentially “ignored” by syntax-prosody mapping principles like MATCH WORD and its precursors. Then in section 3, I propose an alternative account: the failure of an X⁰ to correspond to an output ω happens under essentially the same circumstances as when an XP fails to correspond to a ɸ: the relevant MATCH constraint is simply outranked. I propose that the relevant high-ranked constraint is SUBCAT, which encodes a functional element’s prosodic pre-specification.

Function words tend to have different prosodic properties from lexical words (Selkirk 1980; 1996; Kaisse 1985; Nespor & Vogel 1986; Inkelas 1989; Booij 1996, among many others). In English for instance, lexical words require at least one stressed syllable. Function words, by contrast, lack this requirement and their vowels are generally unstressed, often reduced to a schwa. (8) shows a preposition, an auxiliary and a determiner taking a reduced form.

I follow the analysis proposed by Itô & Mester (2009a; b) that English prepositions, auxiliaries and determiners have the prosodic category of “bare” syllables, and form recursive prosodic words with their complement.5 So under their analysis, each of the function words in (8) integrates into prosodic structure as follows:

Throughout this article, I refer to function words as “cliticizing” into an adjacent ω, but note that this is a purely phonological use of the term, and I make no claim about these forms having special syntactic behavior.

So it seems that function words are X⁰s in the syntax—P⁰s, Aux⁰s and D⁰s among others—and yet they consistently fail to map to ωs. How should we explain this? The consensus choice in the literature, which I argue against in this article, is that the syntax-prosody mapping principles simply “ignore” function words in some respect. To give an example from the pre-Match Theory literature, Truckenbrodt’s (1999) Lexical Category Condition, is stated in (10) (emphasis mine).

This idea has been carried over virtually wholesale into work using Match Theory. (11) provides three recent statements of MATCH WORD principles and constraints (emphases mine).

The following discussion from Selkirk (2011: 453) is also instructive (emphasis mine and bracket notation altered):

Similar claims can be found in Selkirk (1984; 1995; 2011); Hale & Selkirk (1987); Selkirk & Shen (1990); Chung (2003); Truckenbrodt (2007); Werle (2009); Selkirk & Lee (2015) and Guekguezian (2017), among others.

The common thread running through these works is that there is no impetus to parse function words as ωs. Yet the corollary of this—that the phrasal projections of functional categories should not be parsed as ɸs—has been challenged. For instance, Elfner (2012) shows that small clauses, TPs and possessed DPs in Irish, all of which are headed by a functional category, are preferentially mapped to ɸs. She attributes this to MATCH PHRASE, arguing that it does not distinguish between syntactic constituents headed by functional and lexical categories (Itô & Mester 2013 make the same claim). Furthermore, a large body of evidence has shown that coordinated phrases are generally parsed into a prosodic constituent to the exclusion of material outside of the coordination (Price et al. 1991; Fougeron & Keating 1997; Féry & Truckenbrodt 2005; Wagner 2005; 2010; Féry 2010; Kentner & Féry 2013). On the assumption that coordinations are headed by functional categories (Munn 1993), we have another case of a functional projection apparently governed by MATCH PHRASE. In this article, I take this kind of challenge to its conclusion, and argue that neither MATCH PHRASE nor MATCH WORD distinguish functional and lexical categories.

In the next section, I first offer an alternative to the “MATCH WORD ignores functional categories” analysis (henceforth the “lexical-only MATCH WORD” analysis), invoking the idea of violable prosodic subcategorization frames. Section 4 then provides several empirical advantages of this analysis. Following that, section 5 highlights some predictions of the lexical-only MATCH WORD analysis which can be shown to be false.

I propose that most English prepositions, auxiliaries and determiners come pre-equipped with the prosodic subcategorization frame in (14).

This should be read as “Fnc⁰ requires its mother node to be category ω, and it requires a sister node of any category on its right”.

Being associated with this frame forces Fnc⁰ to cliticize into whatever prosodic word shows up to its right. The mappings in (15) all show functional elements cliticizing into their complements.

This behavior is explained if SUBCAT, which enforces adherence to prosodic subcategorization frames, outranks both MATCH WORD and MATCH PHRASE. The three constraints are given formal definitions in (16), and the tableau deriving the prosodic structure of to Andy is shown in (17).8

Crucially, note that losing candidates (a-c) fare better than the winner when evaluated by MATCH WORD and MATCH PHRASE, yet because they each involve a violation of SUBCAT, they lose. To make this point as clear as possible, it is worth going through why each candidate, restated in (18), receives the violation marks that it does.

Candidate (a) is the most MATCH-adherent of the outputs, and were it not for the prosodic subcategorization frame associated with to, it would be the winner. Candidate (b) maps the PP node to a ɸ, just like candidate (a), but induces one more MATCH WORD violation than candidate (a) by failing to map the P⁰ head to to a ω. Candidate (c) earns its MATCH WORD violation mark by being guilty of different sin: it includes a ω that corresponds to no single X⁰. Furthermore, it receives its MATCH PHRASE violation by failing to map PP to a ɸ. Despite its failings, however, it still scores better on the MATCH constraints than the winner, candidate (e). Skipping to candidate (e), we see that it has all the combined sins of candidates (b) and (c): it fails to map P⁰ to a ω, it contains a “spurious” ω that doesn’t correspond to any X⁰, and it fails to map PP to a ɸ. Yet because it’s the only candidate to satisfy SUBCAT, it beats them. Finally, candidate (d) also manages to satisfy SUBCAT, yet it includes an extra MATCH WORD violation—by failing to map Andy to a ω—and so it is beaten by candidate (e).9

Before moving on, two points merit discussion. Firstly, there is the behavior of disyllabic function words. I follow Itô & Mester (2009a) and assume that (at least some) disyllabic prepositions and auxiliaries cliticize, as feet rather than syllables, into the ω to their left. These cases are discussed in more detail in section 4.2.

The second point is that there is variation in the behavior of auxiliaries. One class of auxiliaries is necessarily realized with, at minimum, one syllable. This list includes can, should, could, might, will and some forms of be (were, was, been). These are the auxiliaries to which the pattern described here most cleanly applies (as in (15b)). A second class of auxiliaries, however, may be optionally reduced to a non-syllabic consonant in certain environments. These include the forms of have and some forms of be, reducing to -’m, -’s, -’d, -’re and -’ve, as well as would, reducing to -’d. Regarding these “very reduced” auxiliaries, Kaisse (1985) and Anderson (2008) argue that they form a prosodic constituent with material to their left, and they are discussed in section 3.3.

The next section introduces the prosodic subcategorization frame associated with those English functional elements that cliticize to their left. I focus first on weak object pronouns, before moving on to the “very reduced” non-syllabic auxiliaries in section 3.3. It is argued that all left-cliticizing forms are associated with a prosodic subcategorization frame that is essentially the mirror image of the one we just saw.

I propose that weak object pronouns, contracted negation -n’t, and the “very reduced” auxiliaries are associated with the prosodic subcategorization frame in (20), which is essentially a mirrored version of (14).

Focusing for now on weak object pronouns, this frame accounts for their tendency to cliticize rightwards into the preceding prosodic word:10

The mapping is derived in the tableau in (22), again with all of the more MATCH-compliant candidates (a–c) losing out to the candidate that satisfies SUBCAT(them).

Note that here, I assume that English [verb+pronoun] sequences have the prosodic structure in (23), just as is proposed by Selkirk (1996). In the current proposal we have been able to simply specify the left-cliticizing behavior of object pronouns as a lexical idiosyncrasy, using the frame in (20). However, Selkirk is forced to posit a syntactic cliticization operation where object pronouns cliticize into the verb that selects them. This causes the [verb+pronoun] constituent to be parsed as a single lexical word, and, as a result, to be mapped to single prosodic word. For her, if this syntactic cliticization (essentially head-movement) did not happen then object pronouns would end up treated in the same way as stranded prepositions, on which see section 4.1.

The difficulty with Selkirk’s account is that the syntactic cliticization operation is not well-motivated for English. For one thing, it is hard to provide any evidence that the verb and pronoun form a complex syntactic head: verbs in English do not undergo head movement to T or C, so we can’t check to see whether the pronoun will move along with the verb as it undergoes head movement. For another thing, it is possible to provide evidence that object pronouns will phonologically cliticize into syntactic elements other than verbs, such as prepositions (24a–b) and the adjective worth (24c).11 Note that throughout this article, I provide descriptions and analyses of non-rhotic English.

If we were to maintain that the phonological reduction of English weak object pronouns results from syntactic head-movement into the X⁰ that selects them, we would need to claim that English pronouns syntactically incorporate into prepositions and adjectives too: another claim for which there is little syntactic evidence. I therefore suggest that the account presented here, in which the prosodic left-cliticizing property of object pronouns is a purely lexical property, and is not derivable from their syntax, is a better fit for the English data.12

Object pronouns are not, I propose, the only morphemes in the language to come pre-specified with a left-cliticizing prosodic subcategorization frame: in the final part of this section I discuss the “very reduced” auxiliaries such as -’d, as in we’d already left. In section 6, I discuss contracted negation -n’t, which I argue also has a left-cliticizing frame.

In section 3.1 it was argued that auxiliaries like can and should are associated with the right-cliticizing prosodic subcategorization frame in (14). However, not all auxiliaries fit this mold: in particular, there is a class of auxiliaries that may be reduced to a non-syllabic consonant, a sample of which are shown in (25).

These auxiliaries must be analyzed as cliticizing leftwards (Kaisse 1985; Anderson 2008). For one thing, to analyze them as cliticizing rightwards would mean claiming that (25b) and (25c) involve [zn] and [vm] syllable onsets respectively—onsets that are banned by English phonotactics. For another thing, even where it would be possible for these auxiliaries to cliticize onto the following word without creating an banned onset cluster, they do not do so. As shown in (26), although these auxiliaries could painlessly right-cliticize onto the following word, they instead left-cliticize onto the preceding word, triggering schwa-insertion.

In the system presented here, this behavior is expected if the “very reduced” auxiliaries are associated with the left-cliticizing prosodic subcategorization frame in (20). Note also that the behavior of these auxiliaries provides a crucial piece of evidence against a tempting generalization regarding the relationship between a language’s syntactic head-directionality and its direction of prosodic cliticization. Up until this point, it has seemed that all non-pronominal functional heads in English cliticize rightwards. Under a model in which prosodic constituency directly reflects syntactic constituency, this is exactly what we would expect.13 However, the left-cliticizing behavior of English’s very reduced auxiliaries provides the crucial evidence showing that prosodic behavior cannot be directly derived from head-directionality in the syntax.14

In the next section, I discuss two major empirical advantages that the model outlined here has over the lexical-only MATCH WORD model outlined in section 2.4.

Prepositions and auxiliaries in phrase-final position necessarily map to full prosodic words (Selkirk 1996). The evidence for this is that their vowel cannot be reduced to schwa:

This behavior can be derived from the analysis presented here: in these cases, where there is no material for the Fnc⁰ to cliticize into, SUBCAT is necessarily violated. The candidate that least violates the MATCH constraints is then picked as the winner, as shown in (28).

Note that more radical methods of satisfying SUBCAT, perhaps by altering the linear order of elements (Bennett et al. 2016) or epenthesizing material after the preposition, must be ruled out by other high-ranked constraints.

The non-reduction that we see with stranded prepositions and auxiliaries can be replicated with object pronouns—left-cliticizing elements—that occur at the beginning of a phonological phrase. As shown in (29), when object pronouns occur in phrase-initial position, they cannot be reduced. I believe this is a novel observation.

This behavior can be derived in the same way: left-cliticizing elements at the right edge of phonological phrases have nothing to cliticize onto, and so SUBCAT is necessarily violated. Consequently, the most MATCH-compliant candidate wins, as shown in (30).15

In this analysis, we have essentially reanalyzed the prosodic strengthening of function words in stranded positions as a TETU effect (“the emergence of the unmarked”, McCarthy & Prince 1994): the more marked form (the reduced function word) is blocked in the stranded environment, and so its complementary unmarked form (the unreduced function word) emerges.

I now briefly discuss how this account avoids running into a technical problem that befalls Selkirk’s (1996) analysis once it is placed in a theoretical landscape where prepositions, auxiliaries and determiners cliticize into recursive phonological words. Her analysis is as follows.

Selkirk argues that PPs like to Andy have the non-recursive structure in (31). Note that the category label “ɸ” is not important for the discussion here, what is important about Selkirk’s structure is that it is not recursive.

In her proposal, there is a high-ranked Alignment constraint operative in English, which ensures that the right edge of a ɸ always aligns with the right edge of a ω (ALIGN(ɸ,R;ω,R)). The structure in (31) satisfies this constraint. The preposition-stranding structure in (32a), however, would violate it, and so the alternative candidate (32b), in which the preposition is “promoted” to a ω, must be selected instead.

Yet once we assume that function words create recursive prosodic words such as (33), this explanation can no longer work (note that this assumption is taken wholesale from Itô & Mester 2009a; b—I refer the reader to their work for justification).

The reason why her account no longer works is that it is impossible to create an Alignment constraint that would penalize the structure in (34a), while allowing the structure in (34b)—structurally, they are the same.

This wasn’t a problem for Selkirk’s account, because the two syntactic constituents would form prosodic constituents of different categories, shown in (35), and so they could be distinguished on the basis of prosodic category alone. But in a contemporary landscape where both syntactic constituents map to prosodic constituents of the same category (ω), a discerning alignment constraint like Selkirk’s is no longer an option.

Fortunately, under the account here we can maintain the idea that both proclitics and enclitics form recursive prosodic words, while also accounting for their differing prosodic behavior: the structure in (34a) violates SUBCAT(to), while the structure in (34b) satisfies SUBCAT(them). We now move on to the second major empirical advantage of the proposal.

Not all function word can be phonologically reduced—some of them obligatorily form full ωs, with a stressed non-schwa vowel. One example of this is the demonstrative determiner that, which unlike the other determiners cannot have its vowel reduced to a schwa:16

Demonstrative determiner that stands in a clear contrast to complementizer that, which can be reduced:

The way that non-reducible function words are dealt with in the current analysis is simple: they just lack prosodic subcategorization frames. That is, at the syntax-prosody interface they are treated as regular “lexical” words like dogs. Therefore SUBCAT is inactive, and the most MATCH-compliant prosodic representation is picked instead. That representation is the one in which the DP node is mapped to a ɸ and both contentful syntactic heads are mapped to ωs, as shown in the tableau in (38).

I also propose that we can analyze certain “high-register” prepositions, such as via, in the same way. So the prosodic structure of via Andy’s would be as in (39), and it would result from via lacking a prosodic sucategorization frame.

Note that not all disyllabic function words have this prosodic behavior: Itô & Mester (2009a) propose that disyllabic prepositions like over and disyllabic auxiliaries like gonna have the structure in (40), repeated from (19). As mentioned in section 3.1, the prosodic behavior of these function words can be captured in the same way that we capture the behavior of their monoyllabic brethren, with a rightward ω-adjoining prosodic subcategorization frame.

So why should we think that via is different? My empirical justification comes from Itô & Mester’s own test for ω-adjunction vs. ɸ-adjunction in English. Essentially, on the basis of a similar analysis by McCarthy (1993), Itô & Mester (2009b) propose the following statement for the distribution of intrusive /r/ in non-rhotic English: intrusive /r/ is epenthesized in the onset of a maximal ω, but not in the onset of a non-maximal ω, where a maximal ω is a ω that is not dominated by any other ω.

We can illustrate this with the infamous “function word gap”, in which intrusive /r/ fails to appear at the juncture between a function word and a lexical word: Andy in (41a) constitutes a maximal ω, thus permitting an intrusive /r/ in its onset, while Andy in (41b) does not constitute a maximal ω, and so intrusive /r/ is blocked.

If we apply this test to via, we find that intrusive /r/ is indeed permitted between via and its complement.17 This stands in contrast with disyllabic auxiliaries like gonna, which do not license a following intrusive /r/—the expected result given the structures in (40).

If Itô & Mester’s test is valid, we are forced to assume that the complement of via is a maximal ω—an assumption that is compatible with the structure in (39), but not a structure like those in (40).^18,19

In this section, we have seen that the analysis presented here provides two empirical advantages over a lexical-only MATCH WORD analysis: it allows for a simple analysis of the phenomenon whereby “stranded” function words become full prosodic words, and it allows us to easily capture the behavior of certain function words that behave prosodically like lexical words.

At this point, however, it is important to address the counterintuitive nature of this analysis. I have argued that non-reducing functional elements are the unmarked case, since they are not associated with prosodic subcategorization frames. By contrast, the vast majority of function words, which do undergo phonological reduction, are treated as marked, since they are associated with prosodic subcategorization frames. This may seem a somewhat “backwards” way of looking at things—would it not be more intuitive to treat the exceptional non-reducing function words as the marked case, with reducible function words being unmarked?

I argue that the apparent counterintuitiveness of the analysis derives from the unmotivated assumption that function words form a uniform class whose default behavior is to reduce. Any failure to reduce would then have to be treated as exceptional. However, there is good reason to abandon this assumption: there is more than one way to reduce, and function words do not form a uniform class in terms of their prosodic behavior when reduced. The non-uniformity of prosodic reduction, both across languages and within a single language, is explicitly argued for in the next section, using evidence from English and Serbian.

The overall aim of the next section is to show that two key predictions of the lexical-only MATCH WORD analysis are incorrect. I first show that the lexical-only MATCH WORD analysis predicts that function words should form a prosodically uniform class within a language (relating to the last point mentioned), and that this does not hold empirically. Secondly, I show that function words can induce dramatic non-isomorphisms between syntactic and prosodic structures, which are not predicted under the lexical-only MATCH WORD analysis.

If MATCH WORD does not govern the prosodic behavior of functional items, and they are not pre-specified with any idiosyncratic prosodic information, we should expect that all functional items within a language should be treated in the same way. We have already seen one problem for this in English: prepositions, auxiliaries and determiners cliticize rightwards (section 3.1), while object pronouns cliticize leftwards (3.2). However, Selkirk (1996), anticipating this problem, proposes that object pronouns undergo syntactic incorporation into the verb, meaning that they are treated as a single morphosyntactic word at the syntax-prosody interface. Whatever the merits of this analysis (see section 3.2 for some arguments against it), the fact remains that across languages, different function words exhibit different, often idiosyncratic, prosodic behaviors.

To give an example from Serbian, Zec (2005) shows that function words come in two prosodic classes, which she terms “free” and “bound”. Free function words (when monosyllabic) adjoin at the ɸ level, as shown in (43).

Bound function words, on the other hand, adjoin at the ω level:

One of Zec’s pieces of evidence for this difference comes from the availability of 2nd-position clitics, whose distribution can be (at least partially) defined prosodically. The presence of a free function word in initial position, like mi in (45a), will block the placement of a 2nd-position clitic like = smo after the first ω. By contrast, a bound function word in the same position, like o in (45b), will not block the placement of a clitic after the first ω.

Note that my purpose here is not to discuss the conditions on 2nd-position clitic placement in Serbian: what’s important is that it is possible to diagnose at least two different prosodic behaviors for function words. Furthermore, recent work in the phonology of Bosnian-Serbian-Croatian clitics indicates that there may well be significantly more distinctions among functional elements in that language than those discussed here (Talić 2017). Prosodic differences between different classes of function words in various other languages are also examined in Nespor & Vogel (1986); Chung (2003); Bennett et al. (2018), among others. Ultimately, any theory that assumes that the prosodic behavior of function words can be derived from their being ignored by MATCH WORD will run into difficulty when trying to account for these mixed-behavior inventories of function words.

However, there is a tempting, weaker version of the present analysis that it is necessary to consider. Suppose that MATCH WORD ignores function words, just as in previous analyses, and the grammar makes use of just one “default” method to integrate them into prosodic structure—for English, this would be right-cliticization (as in Itô & Mester 2009a). The remaining exceptional function words, which either cliticize left or map to full ωs, are associated with subcategorization frames.

I believe this alternative is no simpler than the approach advocated in this article, and loses one of its key empirical payouts. Regarding the relative simplicity of the alternative analysis, it gives with the one hand and takes with the other: under the alternative analysis, it is no longer necessary to equip right-cliticizing function words with subcategorization frames—in this sense, it has an advantage over the main proposal advocated here. However, we would now need to stipulate that function words that map to full ωs have their own subcategorization frames, something that is unnecessary in the my proposal. Therefore the advantage in perspicuity we gain in one area is offset by what we lose in another. Secondly, a more serious charge against an alternative analysis relates to what we lose empirically. Under the alternative, my account of how stranded function words become prosodically strengthened (see section 4.1) no longer goes through. This is because my account relies on a MATCH WORD constraint that applies to all function words, including right-cliticizing ones. For these reasons, I propose that all versions of the lexical-only MATCH WORD model are incorrect, regardless of whether or not they admit prosodic subcategorization frames too.

In the next part of this section, I address a second false prediction made by lexical-only MATCH WORD accounts.

If function words are ignored by MATCH WORD, then we would expect that they are integrated into prosodic structure in whichever way is likely to create the fewest violations of MATCH PHRASE, MATCH WORD and any prosodic well-formedness constraints. In this subsection, I show that this is not borne out: function words can induce prosodic structures that are dramatically non-isomorphic to syntactic structure, creating structures that violate MATCH WORD and MATCH PHRASE in ways that cannot simply be the work of prosodic well-formedness constraints. In particular, I show that Selkirk’s (1996) EXHAUSTIVITY constraint, Itô & Mester’s (2009a) PARSE-INTO-ω constraint and Selkirk’s (2011) STRONG START constraint could not be responsible for the non-isomorphisms that we see. On the other hand, the non-isomorphisms that we do see can be nicely captured with the prosodic subcategorization model advanced here.

The relevant case of syntax-prosody non-isomorphism is what happens when right-cliticizing function words take complements composed of multiple prosodic words. An example is given in (46): a preposition takes a multi-ω complement.

Assume that there are two candidate output prosodic structures for this syntactic structure, shown in (47).

The prosodic structure in (47b) is more isomorphic to the syntactic structure than (47a): only in (47b) do Andy’s and house form a constituent to the exclusion of the preposition, just as in the syntactic structure. However, we can show that (47a)—the less isomorphic structure—is the correct one. Recall Itô & Mester’s (2009b) intrusive /r/ test: intrusive /r/ can be epenthesized in the onset of a maximal ω, but not in a non-maximal ω. If the structure in (47a) is the right one, we would predict that intrusive /r/ does not appear before Andy’s—this is because Andy’s does not constitute a maximal ω. By contrast if the structure in (47b) is the right one, we predict that intrusive /r/ should appear before Andy’s, since Andy’s is now a maximal ω.

Applying this test (48), we find that it is indeed impossible to epenthesize /r/ before a multi-ω complement, leading us to conclude that the non-isomorphic structure in (47a) is the correct one (also assumed by Itô & Mester 2009a). The same test is applied to the auxiliary gonna in (49), with the same result: gonna eat forms a prosodic constituent to the exclusion of cake.20

So why do we get the less-isomorphic structure over the more-isomorphic one? I propose that it is a consequence of the prosodic subcategorization frame associated with the functional element, being zealously enforced by its SUBCAT constraint. The tableau in (50) shows how the high-ranked SUBCAT(to) constraint overrules the objections of MATCH WORD and MATCH PHRASE to select the non-isomorphic structure, in the way we are used to by now.21

This analysis requires defending from a number of possible objections and alternatives. I first discuss possible alternative analyses that make use of prosodic well-formedness constraints which do not rely on prosodic pre-specification in the lexicon: Selkirk’s (1996) EXHAUSTIVITY constraint, Itô & Mester’s (2009a) PARSE-INTO-ω constraint and Selkirk’s (2011) STRONG START constraint. I then discuss the possibility of avoiding the problem entirely by using appropriately-defined MATCH constraints, which would truly “ignore” functional categories and projections, and show that this idea runs into the same problems.

EXHAUSTIVITY essentially punishes “level-skipping” in the prosodic hierarchy. (47b) runs afoul of it, since a ɸ directly dominates a σ, while (47a) does not. PARSE-INTO-ω punishes prosodic material that is not parsed into a ω. (47b) violates this constraint too, while (47a) does not. Finally, STRONG START (or at least the relevant version of it) punishes ɸs that start with a category that is lower on the prosodic hierarchy than a ω. (47b) violates this constraint since the preposition to is a bare σ that is not parsed into a ω, but (47a) does not violate it. Therefore for the input in (51), we see that each of these three alternative constraints have essentially the same effect as SUBCAT.

However, all three of these constraints are fatally incomplete as accounts of the behavior of English right-cliticizing function words. The problem only becomes apparent when (46), or some equivalently large FncP, is embedded inside a larger structure. What happens is that neither EXHAUSTIVITY nor PARSE-INTO-ω nor STRONG START are capable of forcing the function word to adjoin to its right, and they permit it to freely, and incorrectly, adjoin to its left. In the tableau in (52), candidate (b), in which the preposition left-adjoins into the preceding ω, receives the same number of MATCH WORD and MATCH PHRASE violations as the desired winner candidate (a). For clarity, the two candidates are shown as trees in (53).

SUBCAT does not run up against this problem: candidate (a) will not trigger a violation, since the subcategorization frame associated with to is satisfied, while candidate (b) will trigger a violation. Itô & Mester (2009a: 20) do make an oblique mention of this problem, stating that “[t]he general proclisis pattern of English means that fnc cannot cliticize to the left”, but this is not encoded in their constraint ranking. To rectify this situation, a tiebreaking constraint would be necessary—one which prefers right-cliticization to left-cliticization for (certain) English function words. This would essentially be equivalent to a SUBCAT constraint, but it would lack the flexibility of that constraint and would apply indiscriminately to all function words, including those which we do want to cliticize leftwards, such as weak object pronouns (on which see section 3.2). See the previous subsection (section 5.1) for discussion of why it would not be desirable to encode English’s general preference for right-cliticization into the interface constraints.

The reader might imagine that an alternative way of avoiding the problems caused by FncPs containing multiple ωs would involve redefining the MATCH constraints. If the MATCH constraints really do ignore function words, we could define them such that the ɸs in (54a) are viewed as the same ɸ, and the ωs in (54b) are viewed as the same ω—that is, adjoined functional items really would count as “invisible” to the MATCH constraints.

This would get us to a place where the two candidates in (47), repeated in (55), would be treated as equally valid by MATCH constraints: to the MATCH constraints, both structures would look like (56).

But once here, we end up with the same problem as we had before: what makes candidate (55a) beat (55b)? If we appeal to EXHAUSTIVITY, PARSE-INTO-ω or STRONG START, we end up with same problem that befell them when integrating the FncPs into larger prosodic structures, which is that structures in which proclitics procliticize fare just as well in the constraint ranking as structures in which proclitics encliticize (see the tableau in (52)). Ultimately, we are required to stipulate, somewhere, that function words must cliticize rightwards. That is, we are forced to simply re-state the effects of a general preference for proclisis, which, as before, causes problems when dealing with the prosodic behavior of English enclitics. In a language with a greater range of prosodic behaviors for function words (e.g. Serbian, as discussed in section 5.1), this approach would be a non-starter.

In this section, therefore, we have seen that two predictions of a “lexical-only MATCH WORD” model are incorrect. Firstly, such a model predicts that all function words within one language should be prosodically parsed in the same way. We saw in section 3.2 that this is not even true for English, and the previous subsection (5.1) presented some cross-linguistic evidence for its falsity. Secondly, the model would predict that syntax-prosody non-isomorphism should be minimized when integrating function words into prosodic structure, at least as far as is permitted by prosodic well-formedness constraints. Again, we saw that this is not the case. Furthermore, attempts to account for attested non-isomorphisms without using prosodic pre-specification end up “hardwiring” the prosodic behavior of particular classes of functional items into the grammar of that language, and essentially forcing all functional items to behave that way. This is undesirable, given the attested diversity in the behavior of function words within individual languages. In the next section, I pursue one further empirical consequence for the proposal advanced here, concerning the prosodic effects of contracted negation -n’t.