The distribution of the null finite complementizer (C) in English has been a longstanding puzzle. A small selection of the basic facts are as follows:

Numerous attempts have been made to capture why both overt C (that) and null C (Ø) are allowed in (1a) but null C is disallowed in (1b) and (1c). These range from purely syntactic accounts (Stowell 1981; Webelhuth 1992; Pesetsky 1995; Pesetsky & Torrego 2001) to morphosyntactic and linearization-based accounts (Bošković & Lasnik 2003; Bošković 2005; Lohndal & Samuels 2013) to prosody-based accounts (An 2007a; b), as well as an array of variationist, usage-based, and frequency-based accounts (see overviews in Shank et al. 2016; Liang et al. 2021). These many different approaches show varying degrees of success in capturing the full range of data and in demonstrating sufficient explanatory power. For our purposes here, it suffices to say that despite significant advancement in our understanding, the phenomenon of null C distribution remains an open question.

In this paper, I will pursue a line of thinking not yet represented in the literature cited above: null C is a prosodically conditioned allomorph of the finite complementizer morpheme. Prosodically conditioned allomorphy is a subcase of phonologically conditioned allomorphy (Carstairs 1988; 1990; Paster 2006; Inkelas 2014) where a given allomorph is licensed by its local prosodic environment. In short, the proposal of this paper is that the null allomorph can only be selected when C occupies a medial position within a prosodic constituent, i.e. non-adjacent to a prosodic boundary. This idea builds upon a set of observations about the range of environments which prohibit null C, going back at least to Bošković & Lasnik (2003), later taken up by An (2007a; b). The observation is that contexts in which null C is prohibited all show the obligatory presence of an Intonational Phrase (IP) boundary immediately preceding C. In other words, null C is prohibited when it would otherwise occupy an initial position within a prosodic constituent. This can be conveniently seen in (1b) above, where C heads a Clausal Subject (That the teacher was lying) and is adjacent to the initial boundary of the sentence itself, and in (1c), where a Parenthetical breaks up the prosodic phrasing of the sentence, represented by comma punctuation. The full range of contexts which prohibit null C that will be addressed in this paper to support the prosodic generalization is as follows: Clausal Subjects, Topicalized Clauses, Extraposition, Parentheticals, Non-Bridge Verbs, Noun Complement Clauses, Right-Node-Raising, Gapping, and Pseudoclefts.

The proposal outlined so far is a synchronic account of the facts, but it is crucial to consider diachrony as well. To that end, I will also address how the prosodically conditioned allomorph Ø can be understood as deriving from progressively weakened forms of overt C in the history of English (e.g., /ðæt/ > /ðət/, /(ə)t/ > /Ø/)—forms which are themselves conditioned by prosodic boundaries. This will require a broader discussion of prosodic constraints on the phonological reduction of English function words (prepositions, auxiliaries, determiners, etc.). In particular, I rely on facts about the application of vowel reduction and deletion of word-initial segments. (2a) below shows that deletion of word-initial /h/ on forms of the auxiliary verb have is allowed in medial position, while (2b-c) show deletion prohibited when preceded by a prosodic boundary, matching the distribution of null C in (1a-c) above.

The segmental strengthening in (2b) and (2c) is referred to as “Domain-Initial Strengthening (DIS)”, a well-established crosslinguistic and diachronic effect whereby stronger articulation of segments is maintained in the initial position of prosodic constituents (Lass & Anderson 1975; Escure 1977; Pierrehumbert & Talkin 1992; Fougeron & Keating 1997; Beckman 1998; Cho & Keating 2001; Fougeron 2001; Keating et al. 2004; White & Turk 2010). For convenience, I will refer to the inverse weakening effect in (2a) as “Domain-Internal Weakening (DIW)”.

The further claim of this paper is that, in cases where a given morpheme develops allomorphs over time that correlate with prosodic position (i.e. the emergence of prosodically conditioned allomorphy), “strong” forms of allomorphs will be associated with initial prosodic positions, while “weak” forms will be associated with medial prosodic positions. In the case of the finite complementizer, over time, DIW causes C to undergo phonological reduction, eventually to zero in the late Middle English period, and the zero-form emerges as a distinct null allomorph correlating with medial prosodic positions. In contrast, DIS prevents reduction in domain-initial position, and so the “strong” form of the complementizer (i.e. that) is preserved in initial position. Note that selection of null C is a case of optional allomorphy, since overt C can also still be selected in medial position. DIW simply yields an extra optional choice for the form of the morpheme, one that has been diachronically conditioned (and subsequently synchronically acquired by learners) to correlate with medial prosodic positions.

The paper is organized as follows: In Section 2, I present the full dataset on null C distribution along with the current state of our understanding of the prosodic properties of the relevant sentences. In Section 3, I briefly address and critique prior accounts of this data, focusing specifically on the prosodic account of An (2007a; b). Section 4 lays out the main proposals of the paper (briefly sketched above), split into a synchronic component, whereby null C is characterized as a prosodically conditioned allomorph, and a diachronic component explaining the historical development of null C and the origins of prosodic conditioning. Section 5 discusses some interesting consequences of the account, while Section 6 concludes.

This section presents the facts on null C distribution, starting with sentences that allow null C and then moving to cases where null C is prohibited. Anticipating the main proposal of the paper (that null C is a prosodically conditioned allomorph), I will also include information on the prosodic form of these sentences, focusing on aspects of prosodic phrasing and the assignment of Intonational Phrases in particular.

The English finite complementizer may optionally be phonologically overt (that) or phonologically null (Ø) in the contexts shown below, adapting data and judgements from Aoun et al. (1987) and An (2007b), which trace back to work by Rosenbaum (1965), Perlmutter (1968), Postal (1974), Stowell (1981), Hornstein & Lightfoot (1991), and Pesetsky (1995).

Each of these cases shows an embedded finite clause immediately following a lexical word, such as a verb (3a), adjective (3b), or noun (3c). In (3a-b), the syntactic structures for these sentences are typically characterized as involving local complementation or head-argument structures, while (3c) is a restrictive object relative clause, the precise syntax of which is beyond the scope of this paper. These facts and related ones are routinely cited in syntactic literature and are easy to confirm (Perlmutter 1968: 215; Stowell 1981: 409–412; Quirk et al. 1985: 1222ff; Aoun et al. 1987: 540–545; see also Mindt 2011: 149 et seq. for many examples with adjectives). Going forward, I will primarily focus on (3a) as the most commonly cited case where null C is licensed.

Turning to the prosodic features of these sentences, an important first step is to outline properties of intonational phrasing in English. Briefly, an Intonation Phrase (IP) is defined as a prosodic unit which is flanked by pauses and/or boundary tones and which contains a number of pitch accents. It is commonly assumed that the syntactic level of the “clause” is obligatorily parsed as an IP, along with a number of other syntactic items, including parentheticals, extraposed and dislocated constituents, tag questions, and vocatives (Selkirk 1978; 1984; 1996; Nespor & Vogel 1986). For the sentences in (3) above, however, the prosodic feature of note is that both the main clause and the embedded clause can optionally be parsed into separate IPs or together into a single IP (Emonds 1970; Downing 1970; Selkirk 1978; 1984; 2005; 2011; Pierrehumbert 1980; Nespor and Vogel 1986; Ladd 1986; 1996; Truckenbrodt 2005; Downing 2011). The sentences in (4) illustrate these two options (from Nespor & Vogel 1986: 199):

Note that the pronunciation of the first IP in (4a) will show a “continuation rise” on believes (a pitch accent on the first syllable with a drop-off in the second syllable, followed by a pitch-rise and a short pause), signaling a juncture between IPs.1 The most important idea going forward is that all three cases in (3) are known to allow the “integrated” IP-parsing shown in (4b), whereas the bulk of evidence suggests that the rest of the sentences discussed below—many of which contain diverse syntactic constructions with varying rules and derivations—do not allow integration between the main clause and embedded clause. As a result, the common thread amongst all of the sentence-types below is that they involve a configuration where the complementizer is obligatorily adjacent to an initial IP-boundary.

Let’s now turn to cases where overt C is allowed, but null C is disallowed. As will be seen, the number of environments in this category far outnumber the cases where null C is allowed.

(5) shows two cases where C is positioned at the beginning of the main sentence, a Clausal Subject in (5a) and a Topicalized Clause in (5b), and null C is strongly prohibited. In fact, the prohibition on null C in sentence-initial position is likely one of the more consistent judgements in the literature. There is not much to say here in terms of prosody, other than to note that (5a) can be straightforwardly parsed into at least two IPs: (_IP That the teacher was lying) (_IP was hardly obvious). It is unclear if an “integrated” IP-phrasing is available for (5a), although this is immaterial for our purposes here. The same is true for (5b), although it is known that topicalized constituents are typically obligatorily parsed into separate IPs, maintaining a Topic-Comment information structure (Downing 1970: 41–45), so maintenance of two IPs is to be expected. The upshot is that, in both cases, C is always obligatorily preceded by an IP-boundary.

(6) shows two cases where adjacency between an embedded clause and a preceding lexical word (the verb believe, in this case) is disrupted: Extraposition in (6a), a process which shifts the clause rightward (Aoun et al. 1987: 540–541), and insertion of a Parenthetical between verb and embedded clause in (6b). In both cases, null C is judged to be disallowed.

Prosodically, both of these cases are fairly well-described. It is routinely reported that extraposed objects are obligatorily parsed into a separate IP from the IP of the main clause, with a clear prosodic break between them (Nespor & Vogel 1986: 188; Hale & Selkirk 1987; Zec & Inkelas 1990: 376–377; Truckenbrodt 1995; 1999; Bošković 2001; 2005; An 2007a; b). Thus, (6a) must be parsed as (_IP Sue said yesterday) (_IP that Kurt had left), with no discernable option for integrated IP-parsing. The case of parentheticals is similar: it is well-established that parenthetical phrases “break up” the prosodic and syntactic material into which they are inserted, and as noted above, parentheticals are a part of the class of syntactic objects which are known to be obligatorily parsed as separate IPs, with attendant pauses flanking the parenthetical and corresponding continuation rise (see Dehé 2009: 571 and citations therein; see also Nespor & Vogel 1986: 188–190; Selkirk 1984: 295–296; and discussion in Gussenhoven 2004; Astruc 2005; Dehé & Kavalova 2007; and Kaltenböck 2007). (6b) therefore must be parsed as (_IP I believe) (_IP based on intuition) (_IP that Sue likes linguistics). Note that both of these cases result in configurations where C is obligatorily preceded by an IP-boundary.

Given the dataset so far, it might be tempting to explain licensing of null C by appealing roughly to the pattern of adjacency between the C of the embedded clause and the preceding lexical word (an idea that is indeed explored by Pesetsky 1995 and Bošković & Lasnik 2003). Unfortunately, however, not all lexical words license null C in a following embedded clause. To start, (7) shows null C disallowed after “non-bridge verbs”, a class containing “manner of speaking” verbs (quip, whisper), subject experiencer verbs (love, hate), and “factive” verbs (regret, resent). The identification of these types of verbs goes back to Ross (1967: 252), Zwicky (1971), and Erteschik-Shir (1973), see also Stowell (1981: 397).

The intonational phrasing of sentences with verbs like quip is addressed by Franks (2000; 2005), citing the judgements in (8)-(9) below. (8a-b) shows that an IP-boundary may precede, but not follow, the C of an embedded clause after a non-bridge verb like quip, while, interestingly, (9a-b) shows that a boundary may optionally precede or follow C after a verb like say.

Franks also judges that, for a sentence with quip, “having a break before that is highly preferred to pronouncing these sentences as a single Utterance” (2005: 19). In other words, there is a preference to maintain separate IPs for both main clause and embedded clause in sentences with non-bridge verbs. It is also relevant, according to my judgement, that sentences with quip vs. say differ in the appearance of a phrasal pitch accent on the verb. Accent is obligatory on quip, likely representing a boundary tone, but only optional on say, correlating with the integrated IP-phrasing discussed in (4b) above. The presence of a pitch accent has also been independently noted for factive verbs like resent and regret (also non-bridge verbs), which always receive accent when followed by an embedded clause (Kallulli 2006: 214; de Cuba & Urogdi 2010: 47–48).

Another instance of a lexical word not licensing null C is shown in (10), a “Noun Complement Clause (NCC)” construction where an embedded clause expresses the informational content of a preceding noun (claim, story, etc.):

Intonational phrasing of NCCs is analyzed by An (2007b), who first notes that NCCs have been traditionally analyzed as appositives—modifiers which are standardly assumed to parse into separate IPs, similar to parentheticals (Stowell 1981). However, there is little overt discussion in prior literature of the intonational phrasing of English NCCs. Instead, An (2007b) presents evidence from NCCs in Serbo-Croatian, Tagalog, and Brazilian Portuguese, showing that NCCs are standardly parsed into separate IPs in these languages, and by assumption, across languages. The interested reader is directed to An (2007b: 39–42) for overall discussion, as well as Bošković (2001) for Serbo-Croatian, Richards (1999) for Tagalog, and Guimarães (1998) for Brazilian Portuguese. Adding to this, it is my judgement that the same pattern identified by Franks (2000; 2005) for non-bridge verbs also applies straightforwardly to NCCs. In other words, an IP-boundary may precede (=11a), but not follow (=11b), the complementizer of a NCC, and a prosodic break after the head noun is more preferred to parsing the entire sequence as a single unit (=11c).

Taken together, the facts on Non-Bridge Verbs and NCCs strongly suggest that a common feature of these constructions is the obligatory presence of an IP-boundary preceding C, a consequence of the apparent prohibition on integrated IP-parsing for these sentence-types.

The final set of sentences which do not license null C also mostly involve adjacency between C and a preceding lexical word. However, in each case, some additional operation has applied to the sentence, such as displacement, coordination, and/or ellipsis, depending on the preferred analysis. (12a) contains an example of Right-Node-Raising (RNR), consisting of an embedded clause placed in a coordination structure with two clause-taking verbs (Postal 1974; Aoun et al. 1987). (12b) shows (Verb-)Gapping, where two embedded clauses are coordinated and the verb of the second conjunct has undergone ellipsis under identification with the verb of the first conjunct (Kuno 1976; Sag 1980; Aoun et al. 1987; Johnson 1997). (12c) shows a Psuedocleft, a construction consisting of a copula joining together an embedded wh-clause in the preverbal subject position and (for our purposes) an embedded clause in postverbal position (Postal 1974; Pesetsky 1995; Den Dikken et al. 2000; Den Dikken 2006; Herment & Leonarduzzi 2012).

Starting with RNR, the prosodic facts are described by Abbot (1976) and Swingle (1993), who observe that each component of an RNR construction (the two conjuncts and the rightward element) must be parsed into separate IPs, indicated by the typical comma punctuation in (12a) above (see also An 2007b: 47–48 for additional discussion): (_IP They suspected) (_IP and we believed) (_IP that Peter would visit the hospital). Selkirk (2002) also finds a clear preference amongst speakers for a prosodic break after each of the coordinated verbs under RNR, which are themselves pronounced with pitch accents indicating narrow focus (see also Hartmann 2000; Féry & Hartmann 2005; Ha 2008 for similar observations). For Gapping, Selkirk (2002; 2005) again identifies a clear prosodic break after the gapped word, although the boundary may not be as strong as an IP-boundary (Selkirk defines it as a “Major Phrase” boundary): (_IP Mary believed Peter finished school) (_IP and Bill) (_?IP that Peter got a job).

Intonational properties of pseudoclefts have also been identified in past work. Both Collins (2006: 1713) and Herment & Leonarduzzi (2012: 714) find that speakers parse pseudoclefts into two or more IPs in the majority of cases. Higgins (1979: 8) and Declerck (1984: 253; 1988: 5) observe a consistent IP-boundary after the copula of a pseudocleft, referred to as “colon intonation”, shown in (13a) below (see Frascarelli & Hinterhölzl 2007; Frascarelli & Ramaglia 2009 for similar effects in Italian and German). This intonational pattern is associated with a “specificational” interpretation of pseudoclefts, where the postverbal constituent acts as a “value” for the “variable” expressed in the pre-verbal wh-clause (Higgins 1976: 95, Declerck 1988: 6–7; Hartmann 2006: 8). However, findings by Collins (1991: 81), Herment & Leonarduzzi (2012: 716), and Van Luven (2018: 62–65) complicate this picture somewhat, showing that pseudoclefts can sometimes manifest with an IP-boundary before the copula as well (=13b).

My impression is that the IP-parsing in (13b) is indeed acceptable, and as such, pseudoclefts present the only (partial) counterexample to the otherwise consistent prosodic generalization outlined so far: that null C is prohibited in sentences which show an obligatory IP-boundary before C. Note, however, that Kearns (2007) actually finds evidence that null C is indeed allowable in cleft sentences, contrary to the overall judgements cited in other literature. It is possible that the variation in IP-phrasing discussed above may be responsible for these inconsistencies, see below for fuller discussion of data from Kearns (2007).

This concludes the overview of environments which allow and prohibit null C. To recap, null C is allowed in sentences with local complementation or a head-argument relation between an embedded clause and a preceding lexical word (=3). In contrast, null C is prohibited with Clausal Subjects and Topicalized Clauses (5), with Extraposition and Parentheticals (6), with Non-Bridge Verbs (7), Noun Complement Clauses (10), and with Right-Node-Raising, Gapping, and Pseudoclefts (12).

The data presented here so far relies on the wealth of acceptability judgements reported in previous literature. However, it should be noted that the characterization of different cases as either purely licensing or purely prohibiting null C is likely an oversimplification. Kearns (2007: 312) presents a corpus study of written English showing significant variation in the distribution of null C in many of the above constructions, and these results can provide quantitative grounding for acceptability judgements. Kearns finds, for example, that null C appears 48–55% of the time with clausal complements to verbs and adjectives, largely as expected given the judgements in (3) above. More unexpected, however, is the finding that null C also appears in “prohibiting” environments, such as with Extraposition (31% of sentences), environments with it-subjects (20%) (a variant of Extraposition), and some cleft constructions (~9%). Even so, note that frequencies of null C in these environments are significantly reduced, compared to the standard licensing environment. This is important because it suggests that there is something about the environments in (5)–(7), (10), and (12) which creates a dispreference for null C, even if null C is not fully banned in every case. Any model which attempts to explain the distribution of null C should, therefore, be flexible enough to allow for variation while also explaining which constructions show a dispreference for null C and why.

I believe the account based on prosodically conditioned allomorphy articulated in Section 4 will fulfill this aspiration, since it does not attribute the distribution of null C to categorical syntactic, morphological, or semantic factors, which might predict more robust splits in the acceptability of sentences. An account of null C based in prosody has the advantage of locating the licensing principles in a module of language (prosody) which is at the intersection of various systems, including morphosyntax and semantics to be sure, but also allowing room for factors like variable prosodic phrasing, which is influenced by discourse context, information structure, and the size and length of syntactic units. Before continuing to the main account, Section 3 briefly addresses some prior proposals for null C distribution, focusing specifically on the syntax-prosody mapping account by An (2007a; b).

Numerous attempts have been made to capture why both overt C and null C are allowed in the contexts represented in (3) above, but null C is disallowed elsewhere. Many prior analyses account for null C distribution in purely syntactic or morphosyntactic terms:

See An (2007a; b) for additional discussion of these and other accounts, such as Pesetsky & Torrego’s (2001) feature-checking analysis.

Importantly, none of these approaches are able to fully capture the range of facts on null C distribution. They are all either forced to set some datapoints aside entirely or are obliged to introduce extra stipulations in order to cover errant data. In brief: Stowell’s and Pesestky’s accounts make incorrect predictions for null C licensing in the context of Right-Node-Raising. Bošković & Lasnik’s null affix account runs into problems with Non-Bridge Verbs (see Bošković & Lasnik 2003: fn.1) and must also introduce a number of separate and distinct kinds of null complementizer to account for different configurations. Finally, Lohndal & Samuels’ account cannot capture the facts for Clausal Subjects and Topicalized Clauses (see Lohndal & Samuels 2013: fn.6). Space does not permit a full overview of critiques of all prior theoretical accounts, but instead see Ormazabal (1995), Bošković & Lasnik (2003), and especially An (2007a; b) for relevant overviews and comprehensive evaluations.

For the purposes of this paper, I will choose to address the subsequent proposal by An (2007a; b), which outlines a distinct approach to null C distribution appealing to prosodic factors. An introduces the important insight that while the various constructions with that and Ø may have different syntactic derivations, they must all be assigned prosodic structure, and therefore prosody offers a potential means of identifying a common through-line for the distribution of overt and null variants of C. An’s account is based on a proposed principle of syntax-prosody mapping called the Intonational Phrase Edge Generalization:

This states that a phonologically null item cannot occupy the edge of an Intonational Phrase. Importantly, as discussed in the preceding section, An observes that many of the contexts in which null C is banned clearly involve obligatory parsing of the embedded clause into an IP, and therefore postulates that a principle like IPEG could ban alignment of null C at the left edge of this IP, which would create a misalignment between prosodic and syntactic constituents. An’s account is attractive because it successfully covers many, if not all, of the facts from the preceding section. However, it nevertheless runs into some problems which make it undesirable as an account of null C distribution, as follows:

First, the IPEG is essentially a stipulation. It does not appear to follow from independent principles of grammar, aside from the notion that syntax and prosody should “match up”. Lohndal & Samuels (2013: 5–6) present this criticism by asking what the motivation could be for a principle like the IPEG to exist. While IPs do generally line up with the edges of clauses (CPs), it is unclear why there should be an explicit requirement that they do so built into the grammar, especially in light of the fact that variation in IP-parsing is indeed observable, as with the cases of “integrated” IP-parsing in the previous section. See Lohndal & Samuels (2013) for additional discussion along these lines, as well as further critiques on conceptual grounds.

This leads to a second objection: the IPEG crucially requires strict isomorphy between syntactic and prosodic constituents at the level of the clause. This is a strong theoretical commitment which is not broadly accepted and does not clearly conform to our understanding of variable relations between syntax and prosody. Many models of syntax-prosody mapping assume some level of isomorphy in the initial syntax-to-prosody mapping (see, e.g., Match Theory, Selkirk 2009; 2011; Elfner 2012; 2015), but also allow for many adjustments to prosodic structure in response to external factors like discourse and information structure, length of constituent, rate of speech, and other elements of parsing and production (Nespor & Vogel 1986; Selkirk 1996). These adjustments yield many instances of non-isomorphy between syntactic and prosodic structure. Importantly, in order to maintain the IPEG, no such contextual adjustments are allowable. If null C is aligned to an IP-boundary in a resultant prosodic structure, the derivation is “stuck”, without the possibility of, for example, selecting a different IP-parsing configuration to avoid or repair an aberrant null edge. The flexibility of prosodic phrasing is a longstanding observation in much prior work; see Selkirk (1978; 1984; 1996) and Nespor & Vogel (1986) for examples of variation in IP-parsing in English. As a result, we would expect that variable phrasing should be available in many cases to resolve an itinerant null edge if IPEG is indeed a constraint on the mapping from syntax to prosody. In order for An’s account to work, however, this option cannot be countenanced.

Taken altogether, I believe these issues speak against the legitimacy of IPEG, and so I will not adopt it here. However, as already noted, I will take as a starting point An’s basic observations about the relevance of IP-boundaries and incorporate this into an account of null C distribution that still appeals to prosodic structure; namely, via the concept of prosodically conditioned allomorphy. This account, consisting of both a synchronic and diachronic component, is presented in the next section.

The data on prosodic phrasing presented in Section 2 shows the following generalization, which I will call the IP-Boundary Generalization:

This basically encapsulates An’s (2007a; b) observations about the correlation between the obligatory presence of an IP-boundary and prohibition on selection of null C. We have already seen, however, that An’s account invoking syntax-prosody mapping and the IPEG is not ideal to account for this correlation.

I will now make an alternative proposal characterizing null C distribution as a case of prosodically conditioned allomorphy, an approach to the phenomenon which has not been explored thus far. The proposal consists of a synchronic component presenting how to understand the relation between the overt and null forms of the finite complementizer, followed by a diachronic component explaining the historical origin of the null allomorph, as well as the origins of the prosodic conditioning factors associated with it.

In this section, I will discuss some consequences of the synchronic and diachronic proposals and also address specific aspects of the data in Section 2 and how it justifies a prosodic approach to null C distribution.

First, note that the synchronic account as presented focuses on the level of IP only. However, DIS has been shown to occur in domain-initial position at all prosodic levels—not just at IP—and is stronger at “higher” levels, see (27b) above. This is relevant to models of syntax-prosody because it leads to predictions about where prosodic boundaries should be postulated. For example, in cases where integrated IP-parsing occurs, illustrated in (4b) (repeated below), we know that both Word Initial Deletion and null C are both available in this context, and therefore we must additionally assume that C is not only medial to IP (non-adjacent to an IP-boundary) in (4b) but also medial to lower-level prosodic constituents, such as a Phonological Phrase (PhP).

Although I will not develop a full account here, there are generally two options for the prosodification of function words at levels below IP: (i) C might be integrated into a lower prosodic constituent on its left, having the status of a prosodic enclitic, as in (28a) below, or (ii) C may be integrated into a lower prosodic constituent on its right, having the status of a prosodic proclitic, as in (28b). In the latter, case, C would be aligned to the initial prosodic boundary of a PhP, and therefore both Word Initial Deletion and null C itself are predicted to be disallowed. This cannot be the case, however, if we understand the availability of deletion and null C to specifically correlate with medial position and therefore the absence of a preceding prosodic boundary. Thus, for the account of null C presented here, I must assume that lower-level prosodic organization of C in sentences with integrated IP-parsing looks something like (28a), rather than (28b). Both of these sentences show PhP-parsing for English, which is generally known to at least parse subject Noun Phrases and Verb Phrases into separate PhPs (Nespor & Vogel 1986).

Notably, the concept of enclitic complementizers has precedent in the syntactic and prosodic literature; see, for example, Schütze (1994), Pesetsky (1995), Bošković & Lasnik (2003), and Bošković (2005). Each of these works posit that when an embedded clause is complement to a lexical word (believes in (28)), the embedded C bears a closer structural relation to that lexical word than to the embedded clause itself. Schütze (1994: 463–464) suggests on the basis of data from Serbo-Croatian that an embedded C can be prosodically organized into the same IP as the selecting matrix verb it is adjacent to. Pesetsky (1995), Bošković & Lasnik (2003) and Bošković (2005) make these ideas more explicit for English data by proposing that null C in particular is either a suffix or a syntactic enclitic, attaching to a lexical category to its left and, in the case of Bošković (2005), subject to various restrictions imposed by intervening prosodic boundaries. More broadly, a number of papers have explored evidence for more widespread enclisis in the prosodification of English function words, including Wheeldon & Lahiri (1997), Lahiri & Plank (2010), Ito & Mester (2019), and Tyler (2019). The reader is directed to these works for further information. For our purposes here, the important thing to reiterate is that an account which appeals to the differential effects of DIS and DIW must at least allow the option of enclisis for C in order to maintain correct predictions for the occurrence of Word Initial Deletion and the prosodic conditioning of null C.

I will now move on to address some specific aspects of the data in Section 2 which highlight the advantages of a prosody-based account such as the one presented here. First, perhaps the most intuitive aspect of the account is the way that it captures null C licensing with Clausal Subjects and Topicalized Clauses, repeated below, both of which position C in sentence-initial position.

It has already been mentioned that selection of null C in these constructions is strongly unacceptable. In fact, of the many constructions discussed so far, selection of null C in these sentences may be the most consistently cited example of null C prohibition in the literature, and I am unaware of any example where null C is actually attested in such an environment, including in corpus work by Kearns (2007) which does otherwise show variability for many other constructions. This is, in fact, predicted by the account of this paper, given the effects of DIS. Since in most prosodic frameworks, the beginning of a sentence will be parsed as an IP-boundary at least, and possibly as an Utterance-boundary as well (a higher prosodic constituent than IP), sentence-initial position is predicted to be the strongest prosodic position possible, and therefore the most resistant to phonological reduction. Consequently, this position would also be expected to yield the highest level of unacceptability should the prosodically conditioned null allomorph be inserted there.

This account has substantial advantages compared to other proposals, some of which were already briefly noted in Section 3. For example, proposals by Lohndal & Samuels (2013), which attempt to derive null C licensing via linearization mappings from syntax, simply leave clause-initial null C aside as an unexplained phenomenon within their account; almost certainly a mistake given the consistent prosodic patterns pointed out here. In contrast, Bošković & Lasnik (2003) propose a morphosyntactic account characterizing null C as an affix which must attach to an adjacent lexical word. This successfully derives many of the adjacency effects on null C licensing seen in the data, including prohibition of null C in clause-initial position, where there is no preceding lexical word for null C to affix to. However, the authors do note that their account provides no explanation for why affixal null C cannot simply attach to an adjacent nominal subject immediately to the right of C, e.g., teacher or John in (5) above (Bošković & Lasnik 2003: fn.31). While not fatal to their account, it does require more bespoke complexity in the subcategorization requirements of the supposed null C affix and again fails in large part to identify the common pattern tying clause-initial null C to other cases of null C prohibition.

Next, let’s address null C prohibition under Extraposition and Parenthetical insertion as in (6), repeated below.

It should be noted that Extraposition is particularly difficult to capture in prior accounts that appeal to the adjacency relation between a lexical word and embedded C (e.g., Bošković & Lasnik 2003; Lohndal & Samuels 2013) because even under Extraposition, embedded C can remain adjacent to some lexical word, such as an adverb (as with strongly in (6a)) or the NP complement of a preposition (I said to Kay that Sue likes linguistics). Accounts which tie the licensing of null C to some aspect of underlying syntax, such as Stowell’s (1981) appeal to Proper Government or Pesetsky’s (1995) appeal to syntactic head-raising, actually fare better in this respect because a clear connection can be drawn between the syntactic rule of Extraposition (however it is characterized) and subsequent licensing or non-licensing of null C. At the same time, a prosodic account is also able to draw clear connections, given the prosodic effects of Extraposition, which prevent IP-integration.

Turning now to Parenthetical insertion, it should be noted that parentheticals are rarely addressed in prior accounts of null C distribution, but may, in fact, be the strongest argument in favor of a prosodic approach. This is because one of the defining criteria for parentheticals is that they may be freely inserted at a number of points in the structure of a sentence and do not “disrupt” the structure around them. Parentheticals are generally assumed to attach externally to the unit that hosts them, either via adjunction or some other operation (see Emonds 1973; 1976; 1979; McCawley 1982; Haegeman 1991; Potts 2004 for a variety of approaches to parenthetical attachment). This is significant because, if null C is allowed in one sentence but disallowed in the same sentence with a parenthetical inserted, this suggests that the licensing factor for null C in these sentences is not strictly syntactic or semantic—not strictly a part of the “underlying” form of a sentence. In other words, we cannot generally assume that a parenthetical disrupts or destroys an underlying structural configuration, because parentheticals are understood to be external to such configurations. At the same time, it is well known that parentheticals do indeed disrupt the prosodic structure of a sentence, as discussed above, breaking up an otherwise-integrated IP structure. If parentheticals affect prosodic structure, but not syntactic or semantic structure, and null C is prohibited when a parenthetical is inserted (as in (6b)), this strongly suggests that the licensing of null C should be located in the prosodic form of a sentence, rather than elsewhere.

Looking at the whole of the data presented in Section 2, it should also be emphasized that the constructions which disallow null C are heterogenous in nature. They include constructions where an embedded clause occupies a special syntactic position (Clausal Subjects) or a “derived” position related to discourse/pragmatic phenomena (Topicalized Clauses, Extraposition), the insertion of other phrases in relation to the embedded clause (Parentheticals), different underlying argument structures and structural relations between the embedded clause and other categories (Non-Bridge Verbs, Noun Complement Clauses), the application of rules of coordination and ellipsis, and the assignment of focus (Right-Node-Raising, Gapping, Wh-clefts). Looking at the problem from a high altitude, however, it is worth asking what linguistic system could tie together all of these phenomena and end up yielding the same effect (prohibition of null C). This highlights the explanatory power of the prosodic account here: despite their diverse syntactic, semantic, and discourse/pragmatic properties, all of the relevant environments must be assigned prosodic structure, and it is this shared structure which determines null C distribution through the medium of prosodically conditioned allomorphy.

Having established some reasons to believe the prosodic account is preferable, I believe that the proposals of this paper can also serve as the starting point for a broader program of research on the boundary between prosody, syntax, and morphology, since it sets up a series of testable predictions for future work to address. First, note that null C is only one manifestation of the larger phenomenon of null allomorphy in English. The account in this paper predicts that null allomorphy may (and even should) also emerge in other prosodic environments. This may allow for extensions of the account to reduced and null counterparts of other function words in varieties of English, such as reductions of verbal auxiliaries. In the sentences below, both reduced and null forms of the auxiliaries are licensed except when a prosodic boundary occurs before the auxiliary (illustrated via Parenthetical insertion, with comma punctuation indicating prosodic breaks):3

A reviewer asks what might constrain rampant reduction to zero in the case of other function words, if it is indeed available and conditioned by prosodic context. To answer this, note that reduction to zero has practical consequences due to loss of information from the surface form. It is possible that null C is licensed over time due to the “redundancy” of C for marking the clause-edge in specific contexts. When a clause is embedded under high-frequency clause-taking verbs like say or think, the existence of the embedded clause becomes highly predictable, and this readily allows reduction to zero over time. The same may be true for verbal auxiliaries have and be referenced above, whose presence is also signaled by participial forms of the verbs they select for. Other function words like the prepositions to and for do not have this functional redundancy. They convey semantic information that cannot be recovered purely via the predictability of the context around them, and so they show phonological reduction, but have so far resisted reduction to zero.

An additional important aspect of any prosodic account is that prosodic phrasing in general (and IP-parsing in particular) is variable, subject to a number of factors like the length of a phrase or clause, the prosodic prominence of elements, speech style, and rate of speech (Nespor & Vogel 1986: 193). This creates predictions for the manifestation of null allomorphs in different environments. For example, a longer or more complex embedded clause is more likely to induce IP-boundaries compared to a shorter clause, and this should potentially have consequences for the selection of a null complementizer: the shorter the clause, the higher the likelihood of prosodic integration, and therefore of null C licensing). Consider also the fact that focus prominence (shown by Small Caps below) has been observed to induce a prosodic juncture after a focused word (Selkirk 2002):4

This predicts that null C should be dispreferred after a focused word. My judgements on this are not entirely clear. Null C seems to be available, but I do not judge the other reduced forms of the complementizer to be as acceptable. At the same time, this is complicated by the nature of focus-assignment, which also implies that post-focal material is presuppositional and/or “given”. Interactions of givenness with null C licensing would also need to be explored in a future study, along with the impact of downstep and post-focal pitch compression for the determination of prosodic boundaries. Lastly, the nature of the prosodification of function words via proclisis or enclisis (as discussed at the start of this section) will also have to be more firmly established.

To conclude this section, any account of the English finite complementizer should mention the classic puzzles of the that-trace effect (obligatory null C with long-distance subject extraction) and the so-called anti-that-trace effect (obligatory overt C with short-distance subject extraction).

In the case of the that-trace effect, the account in this paper predicts that null C should be available at least, but the prosodic account does not provide a mechanism that would enforce (rather than simply allow) the selection of null C. The anti-that-trace effect is more noteworthy, since it appears to prohibit null C in a context where prosodic integration is otherwise available (non-restrictive relative clauses). It is notable, however, that reduced forms of C can be selected in (33b), as expected in a prosodically weak environment. The prohibition is simply that C cannot be fully null. Given the application of short-distance subject extraction in these sentences, I believe other external factors (antilocality, avoidance of parsing ambiguity, etc.) may explain this peculiarity (see Bickerton 2014 and Douglass 2017 for proposals).

In this paper, I have argued for an account of null C distribution based on the observation that the contexts which prohibit selection of null C also show an obligatory IP-boundary preceding the position of the complementizer, while contexts which allow null C do not show an obligatory boundary. This is formalized in a synchronic proposal that null C is a prosodically conditioned allomorph of the finite complementizer, licensed to appear in domain-medial position within a prosodic constituent. Overt C, on the other hand, is licensed to appear in either domain-initial or domain-medial position. This account is complemented by an additional diachronic proposal explaining how such a null allomorph gradually arose in the history of English through phonological reductions of the full form that, and how this particular allomorph ultimately came to be associated with specific prosodic environments as a result of the principles of Domain-Initial Strengthening and Domain-Internal Weakening, which govern the distribution of phonological reduction within prosodic constituents.

This account builds upon a line of work from Pesetsky (1995) through Bošković (2001), Bošković & Lasnik (2003), and An (2007a; b), emphasizing the relevance of prosody for explaining the distribution of nul C, rather than, e.g., purely syntactic properties of sentences. Each of these works identify important factors that have led to the core proposal here, from adjacency between the complementizer and other lexical words to the relevance of prosodic patterns which ultimately tie together the different constructions which disallow null C. This paper takes the final step to argue that the entirety of null C distribution can be captured by prosodic factors, once understood both diachronically and synchronically in the context of phonological weakening processes, principles of intonational phrasing in English, and the possibility of prosodically conditioned allomorphy.