Accounting for non-constituents in hybrid quotations: Why unquotation is not the answer

Abstract

This paper addresses a serious challenge to some recent semantic accounts of quotation: the existence of ‘non-constituent hybrid quotations’, as in Vera said she was ‘very happy and incredibly relieved’ by the supreme court’s decision. These pose a threat to theories that have to make the assumption that hybrid quotations must be co-extensive with syntactic constituents. Responses to the challenge have been proposed, first a quote-breaking procedure, and subsequently unquotation. I argue that these responses fall short of providing empirically satisfactory accounts of the phenomena. Other theories of quotation are not under threat of non-constituent hybrid quotations. I single out a particular family of theories, depiction theories, which have the added advantage of doing justice to the core mechanisms at the heart of quoting.

1. Introduction

This paper addresses what may strike some as only a marginal phenomenon in the great scheme of things: the existence of hybrid quotations that are not mapped onto syntactic constituents. Yet, hybrid quotations are important when it comes to understanding the broader phenomenon that is quotation, and they can often ‘make or break’ theories that seem empirically adequate as long as theorists restrict themselves to (at first sight) more straightforward quotational phenomena, such as direct discourse reports (DD) and pure quotation (PQ). By way of illustration, so-called ‘Name’ theories of quotation as were initially championed by e.g. Quine (1940) do a good job of explaining PQ but prove quite incapable of handling hybrid cases. Here, I will argue that non-constituent hybrid quotations invite the conclusion that even the best and most sophisticated semantic theories of quotation must be found wanting. Thus, their existence provides indirect support for pragmatic theories.

The cases I wish to discuss are aptly illustrated by examples like the following.

The quotations in (1) and (2) do not coincide with strings capable of occurring as syntactic constituents. This would not be an issue if they were instances of embedded DD ³ or PQ. But the fact that they are hybrid quotations poses a challenge to a family of important theories. Section 2 will be devoted to a broad-brush characterisation of the main kinds of theories of quotation, semantic, and pragmatic, with a view to showing the relevance of the present study in spite of its narrow focus. The very notion of hybrid quotation is explained in Section 3. In Section 4, I introduce the theories that have to endorse the assumption that hybrid quotations are co-extensive with syntactic constituents. I also show that this assumption sits uneasily with some of the empirical data. In Section 5, I briefly touch on a first solution initially defended by Maier, the quote-breaking procedure. In Section 6, I review a second solution, originally devised by Shan and subsequently adopted by Maier, namely unquotation. In Section 7, I show why, in spite of its strengths, unquotation still falls short. In Section 8, I return to a family of pragmatic theories to show that they do not (have to) endorse the constituency constraint on hybrid quotation and face none of the issues discussed earlier. A conclusion follows.

My focus in this paper is solely on written quotations. That is not because I think that they require a different theory than spoken quotations. Quite the contrary: theories of quotation, especially those devised by philosophers of language – and they are in the majority – tend to focus exclusively on the written medium, and that is often to the detriment of adequate theorising (cf. Clark & Gerrig 1990: 800; Saka 2011b; De Brabanter 2013, 2017). Yet, in the context of the present paper, I use only written data because those are the data that are most advantageous to the very theories that I criticise, since those theories assign a key role to quotation marks. The upshot is that if they are found inadequate to account for written data, it is unlikely that they could fare any better with respect to spoken data, where the existence of something like quotation marks is quite doubtful. By contrast, pragmatic theories – especially of the depictive kind – were designed with spoken quotations in mind. If they can handle written data, they can readily do so with spoken data as well.

2. Two main families of theories

Theories of quotation can be broadly separated into two main families: semantic and pragmatic. Semantic theories regard quotation as a phenomenon to be explained in terms of the conventional meaning of morphosyntactic or lexical elements. Chief among those are quotation marks, which are widely assumed to be referential devices that are necessary for quotation to take place (e.g. Davidson 1979; Cappelen & Lepore 1997, 2005; Benbaji 2004, 2005), and whose conventional meaning is commonly assumed to encode a verbatim constraint: a quotation must faithfully reproduce what it stands for (e.g. Cappelen & Lepore 1997; Maier 2014 for a qualified notion of verbatimness). This constraint is reflected in the truth-conditions that semanticists have proposed for quotation. When quotation nonetheless seems to have different effects on truth-conditions, semantic theories distinguish different ‘varieties’ of quotation and provide different logical forms for these. It is standard nowadays to make a distinction between pure quotation, direct discourse, mixed quotation (MQ), and scare quoting (SQ). Those are usually considered to require distinct semantic and syntactic explanations (Cappelen & Lepore 1997; Gómez-Torrente 2005; Maier 2014); sometimes one of them is regarded as falling outside the bounds of semantics (cf. SQ in Cappelen & Lepore 2005).

Pragmatic theories, in contrast, insist that quotations are produced by quoters rather than generated by a morpheme or lexical item (Saka 2011a; Salkie 2016). Though convenient, marks of quotation are regarded as non-essential disambiguators (Washington 1992; Saka 1998; Recanati 2001; De Brabanter, 2022). The more radical approaches regard quotation as an iconic communicative act, a demonstration on a par with certain kinds of gestures or paralinguistic features (Clark & Gerrig 1990; Clark 1996, 2016; Recanati 2000, 2001; De Brabanter 2017). On that view, quotations are crucially different from ordinary linguistic acts that prototypically rely on conventional form–meaning pairings. They are what Peirce (1998) called icons, viz. signs which signify through selective resemblance, not convention. As for the ‘varieties’ of quotation, they are not distinct quotational phenomena; they are explained in terms of how the iconic act that is quotation interacts with convention-based linguistic acts: is the quotation embedded in a conventional act or concurrent with one or free-standing?

Although the semantic accounts have been dominant among the specialists most concerned with theorising quotation – philosophers of language and formal semanticists – they have come under attack from the pragmatic camp. Criticism has centred on the central role accorded to quotation marks, the attendant inattention to spoken quotations, the widespread assumption that all quotations are referential, the verbatim constraint, the comparative neglect of the central depictive/pictorial nature of quotation. I cannot rehearse any of these criticisms in detail (see Saka 1998, 2011a; Recanati 2001; De Brabanter 2013, 2017, 2022; Salkie 2016), but their existence is crucial to make sense of the present study. Although it is legitimate to think that the failure of the unquotation account of hybrid quotation would not in itself be a big deal, that failure becomes significant against the background of the other flaws of semantic accounts. If unquotation proves inadequate in fixing a problem that plagues some of the best accounts of hybrid quotation on the semantic marketplace, then the viability of semantic accounts is open to question. This, I take it, is enough to justify a study with as narrow a scope as mine: it is one among an array of studies that have pointed out some serious issues with semantic accounts.

3. Hybrid quotation

Hybrid quotation is especially common in certain written genres in English, news writing paramount among them. In the literature, it also goes by other names, notably mixed (e.g. Geurts & Maier 2005), double-duty (García-Carpintero 2005), and subclausal quotations (Potts 2007). The examples below illustrate what are often regarded as two separate varieties of hybrid quotation, namely mixed quotation (in Cappelen & Lepore’s 1997 narrower sense than Geurts & Maier’s 2005) and scare quoting (e.g. Predelli 2003). I have argued elsewhere that the distinction between mixed and scare quoting rests on shaky foundations (De Brabanter 2017: 238) and will therefore mostly use the umbrella term hybrid quotation, which I borrow from Recanati (2001). ⁴

Example (3) exemplifies mixed quotation as originally defined by Cappelen & Lepore (1997), i.e. a hybrid quotation that is under the scope of a reporting verb (told, said) in a construction that appears to mix indirect discourse (ID) with DD, while (4) illustrates scare quoting, by and large the use of a quotation as a way of providing a metalinguistic comment, e.g. that ‘the enclosed expression is used non-standardly, or is unsuitable for the tone of the discourse; [or, very differently] that the enclosed material is borrowed from an allegedly identifiable source’ (Predelli 2003: 4).

To really grasp what hybrid quotation is, it is useful to contrast it with more widely studied quotational phenomena, namely DD and PQ. Prototypically, these quotations function as NPs in clause structure – they are replaceable by, e.g. that in examples like (5) and (6).

Importantly for present purposes, the internal syntactic structure of the quoted strings in (5) and (6) is segregated from their surroundings: never mind that the quoted string in (5) is a sequence of an AdjP and an NP. The complement of said is not generated via a phrase-structure rule like VP ⇒ V AdjP NP. Semantically, the quotation is about a speech act by Kim, not (directly) about some object that is remarkably red. Likewise with (6), which is about a certain grammatical structure, not a particular time interval, and is not generated by a rule that states that PPs can be subjects. Note further that strings instantiating just any syntactic structure (or no structure at all, e.g. She the with me six) can be placed between quotation marks in embedded DD and PQ, without altering the grammaticality of the host sentence:

Just the opposite can be observed with hybrid quotations, and that is because they involve strings of words whose internal structure is part and parcel of their host structure. Consider example (9), in which the nominal (similar to an N’ in X-bar theory) billion Muslims who reject their ideology does double duty, as García-Carpintero (2005) would put it: it is quoted and it functions as the head of the NP the billion Muslims who reject their ideology. You could not substitute, say, a PP, NP, AdvP, VP or a finite clause for this string without making the host structure ungrammatical. Nor could you substitute word salad as was done successfully in (7) and (8). Example (10) shows the ungrammaticality that results from substituting a PP for the quoted nominal in (9):

Likewise with (3), in which the clause consider running for the Presidency does double duty as well, as it is both quoted and the complement of the auxiliary would: again, nothing but a bare infinitival clause fits in this slot. Similarly with (4), where nothing but a past participle of a transitive verb (like seen) will yield an acceptable sentence.

4. The constituency constraint on hybrid quotations

Over the past twenty years, quotation, notably its hybrid variant, has attracted the attention of an increasing number of formal semanticists, who have proposed the most articulate and explicit accounts of the phenomenon. One feature that their proposals often have in common is a principle of congruence between syntax and semantics: only syntactic constituents (only nodes in a syntactic tree) receive a semantic interpretation. As a consequence, for a hybrid quotation to be ascribed a semantic value it must map onto a node in a tree, i.e. it must be a constituent.

It is worth emphasising the stringency of this constraint. In particular, on Maier’s (2014) analysis, the tree diagram that represents a sentence with a hybrid quotation must contain a node that dominates exactly the hybrid string because it is there in the structure that a ‘mixed quotation rule’, which transforms an ordinary constituent into a ‘mixed-quoted’ constituent of the same category, is inserted. Therefore, hybrid quotations must be co-extensive not just with potential constituents – strings capable of occurring as a constituent in some syntactic structure – but with actual constituents in the structure in which they occur. Consider (11). The string every effort is a potential constituent because it can occur as an NP in a clause like Every effort deserves a reward. But in (11), every effort alone is not a constituent. It only forms one in combination with to stop its adverts appearing next to inappropriate content.

The various semanticists who have endorsed the syntax–semantic congruence thesis have also accepted the constituency constraint on hybrid quotations. Most of them are formal semanticists – Geurts & Maier (2005), Maier (2007, 2008, 2014, 2015), Potts (2007), Shan (2011) – and one of them a philosopher, Benbaji (2004, 2005).

I concentrate on the proposal in Geurts & Maier (2005) because it has gone through several developments and improvements at the hands of Maier, the most relevant of which will be examined in detail below. Geurts & Maier (2005) capture the semantic value of a scare quotation ‘α’ as ‘what (contextually salient) speaker x means by α’. Thus, the quoted predicate in example (12) is interpreted as in (12′):

For the hybrid quotation to have a semantic value, it must form a constituent. In (12), we have an NP, an atheist, which expresses the property that the Nice attacker’s brother attributes to the word. This guarantees the correct contribution to the truth-conditions of (12), whether the echoed speaker used the phrase in its usual sense or idiosyncratically.

Around the time that Geurts & Maier set out their views on hybrid quotations, several writers pointed out that actual examples could be found that flouted the constituency constraint. To my knowledge, Abbott (2005: 20), Cumming (2005: 80–81), and Tsohatzidis (2005: 228) were the first to do so. Example (1), repeated here as (13) is Abbott’s, and (14) is Cumming’s: ⁵

Both of these examples show hybrid quotations that are ‘too long’ to form constituents. In (1), the quotation spans two constituents which, together, do not form a larger constituent: largish sums is the direct object, and several benign institutions a prepositional complement, of donated. Whatever syntactic analysis you defend (binary, with a VP donated largish sums, or ternary, with donated, largish sums and to several benign institutions the immediate constituents of donated largish sums, to several benign institutions) largish sums and to several benign institutions together are never mapped to a node in a syntactic tree. ⁷ In (14), weight … air is the nominal that heads the NP that complements the preposition by, whereas because … fluids is an adjunct that most likely modifies the VP that begins with suspected. Again, at no point do these two elements together form a node in a syntactic derivation.

There is a second category of relevant examples in which we have something like the reverse problem, i.e. where the quotation stops short of forming a constituent. Thus, in (2), repeated here as (15), constituency would require his behaviour, the direct object of the conjunction of verbs enabled and encouraged, to be quoted too.

Maier (2014) will later refer to the two categories of examples as involving superconstituent and subconstituent quotation, respectively, and I will adopt this terminology in what follows. It is worth noting that instances of both categories are by no means infrequent. Examples (16) and (17) are two more superconstituent examples like (13) and (14). Both illustrate the not unusual case in which the second element in the quotation is a separate main clause (even a distinct sentence in (15)). Similar to subconstituent (15) are (18) and (19).

Informally, it is clear that no proper meaning can be attributed to the quoted strings in the examples in this section of the paper. Here, for illustration, are a couple of failed attempts at doing just that, using the plain-English gloss of the meanings of hybrids à la Geurts & Maier:

More specifically, the syntactic analyses proposed by Potts, Maier, or Shan use variants of a grammar in which each node in a tree is associated with a semantic type, e.g. e for (individual) entity, or t for truth-value. ⁸ On those formalisms, it is impossible to ascribe a semantic type, and therefore an interpretation, to any of the quotations in the examples (11) to (19).

Were the phenomenon to be marginal, one could argue that it does not need to be addressed. But it is not. ⁹ Therefore, any proper theory of quotation ought to be able to account for it. But how can the theories just mentioned provide semantic values for strings that are not constituents?

5. A first solution: quote-breaking

To be fair, Emar Maier was immediately aware of the issue raised by Cumming. In Maier (2007, 2008), he proposed supplementing the theory devised jointly with Geurts with a ‘quote-breaking procedure’ that made constituents out of quoted strings that are not constituents. With respect to an example like (13), where a non-constituent made up of two constituents is quoted, the quote-breaking procedure yields the output in (20). This could be regarded as the underlying form of (1), in which two bona fide constituents, the direct object largish sums and the prepositional complement to several benign institutions, occur as separate hybrid quotations:

De Brabanter (2010: 117–118) and Shan (2011: 433–434) pointed out that the procedure fails to account for the fact that all the quoted words were pronounced together as part of a single uninterrupted utterance. Taking his cue from Shan’s criticism, Maier quickly gave up on the quote-breaking procedure as an empirically adequate account of both subconstituent and superconstituent cases (Maier 2014: 7:52). In its place, he adopted a version of Shan’s mechanism of ‘unquotation’. It is to this solution that I now turn.

6. A second solution: Unquotation

Shan initially defines unquotation as ‘including non-quoted material inside a quote’ and ‘typically punctuated using square brackets’ (2011: 432). He makes a distinction between syntactic and semantic unquotation. The former is illustrated in (21), where an expression that is part of a quotation is replaced by a metalinguistic description of that expression. The latter is at work in (22), where an expression that is part of a quotation is replaced by another expression with a similar semantic value:

With respect to Abbott’s original examples (with superconstituent quotations), Shan suggests that semantic unquotation can be enlisted to meet the constituency requirement: ‘[w]e can analyze these examples by postulating semantic unquotes at the edge’ (2011: 433) of the problematic quotes. The unquoted version of (13) is given informally in (23).

The idea is that in the original (1), donated represented a gloss by the reporter, i.e. an unquotation. In other words, some words outside the quotation marks in the (surface) structure are actually part of the quotation in the (underlying) syntactic structure.

It is easy to see that a similar analysis can be extended to cases of subconstituent quotations. Thus, his behaviour in (15) is analysed as having been unquoted. Similarly with with it in (18) and with Mr Trump in (19):

However, some examples of superconstituent hybrid quotations do not readily lend themselves to this sort of treatment. Consider (17) again. What is the right unquotation analysis of such an example? Is it (27), where the quotation now corresponds to a juxtaposition of two sentences? But that is not a constituent! Is it (28), where the quotation now comes to resemble an instance of free direct discourse? In other words, it is no longer a hybrid quotation, and therefore no longer has to satisfy the constituency constraint.

Leaving those problems aside and focusing on the examples where the unquotation strategy can be applied, the verdict is clear: on the proposed analyses, all quotations are constituents: [donated] largish sums, to several benign institutions is a VP; a staff of dozens who enabled … and encouraged [his behaviour] is an NP; couldn’t go through [with it] is a VP; a lot of disagreements [with Mr. Trump] is an NP.

So there is no doubt that unquotation offers a solution to a large range of cases that would otherwise refuse to be handled by the theories proposed by Potts, Shan, and Maier. But is it a good enough solution?

7. Problems with unquotation

Shan borrows the idea of unquotation, together with the term, from computer science. However, there is a difference between showing that a mechanism useful in computer science can be enlisted to account for a natural-language phenomenon and showing that that mechanism is actually at work in the phenomenon under scrutiny. After all, Shan talked of postulating semantic unquotations in superconstituent hybrid quotations. Is unquotation intended as a mere notational ‘fix’?

To be fair, something like unquotation is practised widely in journalism and academic writing, where square-bracketing a part of a quotation signals a modification, by the quoter, of the words uttered in the source context. The term unquotation, however, is not in common currency, with typesetters and editors seemingly preferring interpolation (e.g. Bringhurst 2004: 317; Chicago Manual of Style, 2010). Typically, words can be replaced or added to a quotation. ¹⁰ In (29), his is bracketed to indicate a change in the letter of the reported utterance, which must have contained a my. In (30), the Sunni ruling family is added between square brackets to help readers make sense of the name al-Khalifas.

The fact that unquotation is grounded in actual natural-language use is definitely a plus. And there is more. Maier (2015: 365) insists rightly that ‘unquotation is more than a superficial typographical invention of modern day editors. … it marks a genuine semantic operator that occurs, covertly, in many other registers of written and spoken communication as well’. To back up this claim, Maier provides illustrations similar to (31), in which the them is most probably the implicitly unquoted counterpart of an us in the source utterance, a change no doubt motivated stylistically.

All of what precedes definitely makes unquotation very attractive. But there are enough reasons to resist this attraction. I go through them below.

7.1 The empirical grounding of unquotation is shakier than it initially looks

I wrote above that one major asset of unquotation (as opposed to quote-breaking) is that the mechanism is independently motivated: something like unquotation does happen in actual language use. Though true, this requires some qualification: there are also some striking dissimilarities between Shan and Maier’s semantic unquotation and real-life interpolation, to which I now turn.

7.1.1 Location of unquotation in the quotation

Remember that Shan points out that his unquotations occur ‘at the edge’ of hybrid quotations. The few examples of non-constituent hybrid quotations that Maier analyses in his (2014) are of the same ilk, inevitably, since unquotation is always meant to extend the quotation on the left or the right until it spans a constituent. ¹¹

That is very much unlike the interpolations discussed in style manuals. These have no prescribed location in the quotation. Compare (29) and (30). In connection with the APA style principles, Perrin writes that ‘most often, the words you add are specific nouns to substitute for pronouns that are vague outside the context of the original work. However, you may substitute a different tense of the same verb (for example, used for use)’ (2010: 74). This is a far cry from the semantic unquotation that is enlisted in an account of HQ.

Consider (32), an expanded version of (11), a quite extreme example, as the constituency constraint demands unquotation at both edges of the second quotation: the second complement of categorised – the PP as sensitive – has to be unquoted, and so does had, because the negator attaches to it. This is shown in (33).

The above is by no means a knock-down argument against unquotation in hybrid quotations, but it is enough to weaken the view that the same mechanism exists in natural-language use. The Shan and Maier type of unquotation is definitely idiosyncratic.

7.1.2 Explicit interpolation versus implicit unquotation

Another criticism that can be levelled at semantic unquotation is that it is implicit, whereas empirically observed interpolation is customarily explicit. Were it not for the square brackets, one would usually not know that this or that word or phrase has been interpolated into a quotation, witness (29) and (30). Maier writes that in his grammar of hybrid quotation, ‘we have the same unquotation in the syntax [as in ordinary interpolation] but it does not get expressed phonologically’ (2014: 7:57). Now, as Maier (2015) pointed out too, covert interpolation does occasionally crop up in ordinary writing. The quotation in (31) was a case in point. So is (34), which is a newspaper headline:

In the body of the Guardian article, the quotation is “hate your guts” with generic your, as shown in (35). Replacement of your by his may suggest that there exists a stylistic prescription in hybrid quotations against deixis that is not adjusted to the reporter’s perspective (see Maier 2014: 7:57). This prescription is far from absolute, as examples like (35) and (36) are frequent in the English-language press (See De Brabanter 2020):

Still, real-life cases like (31) and (27), in which I am ready to agree that there is covert interpolation/unquotation, are rather different from non-constituent cases like (1), (2), (18), and (19). In (31) and (34), some recognition of unquotation by the reader is necessary to arrive at a correct interpretation of the sentence. This is probably not a cognitively demanding task for readers; after all, them in (31) and his in (34) are the pronouns they would expect to occur in plain indirect discourse. Still, they must refrain from interpreting the pronoun from the perspective of the reported context, and that is unlike what happens in an example like (36), where my refers to the utterer in the reported context (i.e. Trump), not to the reporter.

By contrast, Shan’s (2011) and Maier’s (2017) semantic unquotation in hybrid quotation is entirely clandestine. ¹² That, notwithstanding Maier’s assurances to the contrary (2017: 261–262), makes this brand of unquotation quite unlike even the covert interpolation in (31) and (34). In the end, Shan’s and Maier’s only evidence for unquotation is the fact that without it the hybrid quotation is not a constituent. This is suspicious. Besides, it is quite unclear how readers would work out that utterances like (1), (2), (18), and (19) involve unquotation. So, semantic unquotation seems a phenomenon that fulfils no clear function in language use; none, that is, other than enabling a theory to account for examples the existence of which it cannot otherwise justify. Note in passing that, as far as I have been able to find out, none of the theories that require hybrid quotations to be constituents ever provides empirical grounds for this requirement. Hence, it may look as if the requirement is entirely generated by the needs of the theory the semanticist is developing.

7.2 Unquotation ascribes too many words to the reported speaker

The next objection depends on the fine detail of how we understand semantic unquotation. So far, I have interpreted it as the substitution of a reporter’s word or phrase for an expression that is part of the quotation. All the examples provided by Shan suggest that that is the mechanism at play. Besides, that is the way Maier understands unquotation. ¹³ In fairness, Shan’s text may provide a reason for a slight hesitation. Initially, in connection with example (37), Shan offers this comment: ‘In [(37)], Bush could have used neither the word eclectic nor eckullectic’ (2011: 432).

Does this formulation leave open the possibility that Bush might have produced no word at all between my and reading? In other words, is Shan implying that semantic unquotation allows insertion of a reporter’s words in a position where there were no words in the source utterance? The answer seems to be ‘no’. With respect to the examples that immediately follow (37) – reproduced here as (38) and (39) – Shan writes that ‘each boy could have referred to his uncle using a different expression, and the politician could have said not my job but this despicable position of deception’ (2011: 432).

In those cases it is quite clear that the words plugged into the quotation replace words uttered by the reported speaker.

If the above is accepted, then I can show that both Shan’s and Maier’s accounts of unquotation make incorrect predictions about the kinds of words that must have been used by the reported speaker.

We start with a complex example that involves a hybrid quotation in a language different from the language of reporting. The reason that I wish to discuss (40) is because of the ellipsis it contains. On the most likely reading, this ellipsis is an addition by the writer, a Times Literary Supplement journalist, not one that occurred in the reported speaker’s original text. ¹⁴

Presumably, the source text had libérée des scrupules conventionnels qui le ligotent (‘free from the conventional scruples that tie him down’), with le referring to Tom Brangwen. In order to restore a constituent hybrid quotation, the underlying structure in (41) has to be postulated.

On my understanding of semantic unquotation, the unquotation Tom Brangwen replaces words (namely, a different phrase that refers to that character) in that position. But the ellipsis points very strongly suggest that this other phrase, the cliticised pronoun le, occurred between qui and ligotent. In other words, the putative unquotation clashes with the ellipsis: in Delavenay’s text there was simply no phrase referring to Tom Brangwen after ligotent.

Although I see nothing in Shan (2011) that can address this problem, Maier’s (2014) proposal may have a ready response. After all, Maier hypothesises that when there is unquotation it is a construction that is quoted, namely a grammatical object that may have holes in it. In this particular case, Maier could perhaps argue that the construction has a hole for the second argument of ligoter, and that what gets filled by the unquotation is not a position in linear order, but in the argument structure of ligoter.

So, maybe there is a fix for the problem just brought up. But other examples resist fixing. Consider (42). The relative clause that follows beat-up modifies it, hence that clause is part of the NP headed by beat-up. As a consequence, the constituency constraint requires that which is clearly targeting Wada is unquoted, as shown in (43).

This means that something like that string (though not that string) was uttered by former Anti-Doping Agency chief, Ings, in the source context after he said Russia. Consider, however, that if Ings had not said anything after Russia, if that piece of information had been provided by the journalist strictly to help readers, the quotation in (42) would still be totally acceptable. When I first read (42), that is precisely how I interpreted the relative clause. Yet, if my understanding of semantic unquotation is to the point, this interpretation is simply not available to Shan or Maier. The unquotation account puts words in the mouths of quoters that they did not utter.

Should some readers want less conjectural evidence, that can be supplied too. Consider (44), an amended version of (14). ¹⁵

In this example, it is most likely that the because PP modifies the VP suspected that the mercury was really supported by the weight and pressure of the air. Thus, in order for the quotation to span a constituent – a very large VP – unquotation must include everything that precedes the quotation except Pascal, yielding the representation of the structure in (45).

In other words, Pascal ¹⁶ must originally have written words to the same effect as the unquoted ones, and in a position immediately before the quotation as it appears in (44), for instance think that quicksilver ¹⁷ must actually be sustained by the. That seems possible. However, as with (42), one feels that (44) is an acceptable utterance even without these two conditions (regarding the kinds of words and their location in the sentence) being fulfilled. The meaning of the unquoted words could have been expressed elsewhere, or it could have been reconstructed by the quoter, Rupert Hall, on the basis of his reading of Pascal. Again, the demands of the unquotation account appear to be too high.

For the final word on this, we can turn to the actual text, the English translation of Pascal’s letter to ‘Monsieur Périer’. The original text has (46):

Pascal’s all these effects is anaphoric to his previous the effects ascribed to the abhorrence of a vacuum. But there is no earlier reference in this letter to mercury being supported by the weight and pressure of the air. There are several later references to mercury (as quicksilver), and these warrant Hall’s conclusion that Pascal suspected that, among other effects, mercury was supported by the weight and pressure of the air. But the unquotation analysis requires that something like suspected that the mercury was really supported by the occurred in Pascal’s text right next to the quotation in (44). I maintain that that is not the case. This way, the unquotation analysis precludes the interpretation which, based on the philological evidence, is the only correct one.

7.3 Drawing some lessons

The discussion in Sections 7.1 and 7.2 does not engage with the details of Shan’s and Maier’s formal frameworks. I trust, however, that it correctly reflects the assumptions and commitments that those frameworks rest on. And it is those that matter for present purposes and that ultimately prove unsound.

Maier (2014) had mounted a forceful and skilful defence of unquotation, even arguing ‘that it is an essential ingredient for an empirically adequate analysis of mixed quotation in natural language’ (2014: 7:1; emphasis added). Although a brilliant instance of formal creativity and astuteness, unquotation as a repair mechanism fails empirically.

The prescribed peripheral location of unquotation, its essential covertness, and the unfounded attributions of words that it entails – taken together those issues should be enough to persuade us that it is unsuccessful.

If unquotation as a repair mechanism falls short, I venture, it must be because, ultimately, there is nothing to repair. The underlying assumptions of Shan’s and Maier’s theories are ill-founded. In particular, the verbatim and the constituency constraints to which Maier is committed simply do not reflect what quoters do. The idea that quoters will resort to unquotation only when it turns out that producing a verbatim constituent proves problematic presupposes too much deliberate monitoring of grammatical structure on their part. ¹⁹ At the same time, it is remarkable that so much controlled effort would be expended on a process to which readers remain totally oblivious (see Section 7.1).

I conclude that the most sensible move is to renounce the constituency constraint, together with the verbatim assumption in Maier’s case.

8. Theories that work

There are at least two types of theories that have no trouble handling non-constituent hybrid quotations. They divide into (i) a subset of the semantic theories that take quotation marks to be essential ingredients in generating hybrid quotations – these are essentially theories of quotation marks (e.g. Cappelen & Lepore 1997, 2005; Predelli 2003; García-Carpintero 2005, 2011; Gómez-Torrente, 2017) – and (ii) pragmatic theories that treat quotation as first and foremost an iconic communicative act (Clark & Gerrig 1990; Clark 1996, 2016; Saka 1998; Recanati 2001; De Brabanter 2017) and therefore regard quotation marks as useful but optional disambiguators. ²⁰

I will not go into the theories of the first type, which, as hinted in Section 2, I think are ultimately inadequate to the task of explaining quotation. On the depictive theory that I endorse, which is profoundly shaped by Clark & Gerrig (1990) and Recanati (2001), quotations, like other iconic communicative acts (as realised e.g. through gesturing or prosodic features), can be concurrent or not with an ongoing linguistic act, one that is convention-based. When they are not, as in embedded DD and PQ, they occupy the place of a convention-based communicative act, or of part of such an act. In (5) and (6), they are made to function as arguments of a V or VP, and could be replaced by non-quotational NPs (underlined), as in (47) and (48). In (49), the quotation, an instance of free direct speech (Quirk et al. 1985: 1033; Huddleston & Pullum 2002: 1029), stands for a whole clause, which could have taken the non-quotational form in (50).

A quotation that is concurrent with an ongoing convention-based act is not a distinct ‘variety of quotation’. In reality, hybrid quotation is not sui generis, and the labels used to designate it are misnomers. In so-called hybrid (mixed, double-duty, subclausal) quotation, what happens is just this: a string that is part of a convention-based communicative act is simultaneously involved in the performance of an iconic act. The latter act is an ordinary quotation, albeit superimposed upon part of a convention-based act. The quoted string is therefore hybrid (in the sense of functioning both conventionally and iconically), but it is not a ‘hybrid quotation’ in the full sense of instantiating a distinct type of communicative act.

In hybrid cases, the quotation, being merely superimposed, does not segregate the internal syntactic structure of the quoted string from that of the structure that embeds it, as happens in embedded DD and PQ (cf. (7) and (8)). Neither does it segregate the internal semantics of the quoted string from that of the embedding structure. This kind of segregation only occurs when a quotation or other iconic act is ‘recruited’ (Recanati 2001) to occupy a syntactic slot on its own, as in embedded DD and PQ. Thus, hybrid quotation is more basic than embedded DD and PQ in that it quotes without additional syntactic recruitment.

What the quotation contributes in hybrid cases is an extra layer of meaning, to the effect that the string occurring in the quotation instantiates some properties of words uttered elsewhere and thereby suitably resembles those words. Two main ‘quotational points’ (Recanati 2001) can be distinguished, corresponding typically (but far from systematically) to the distinction between quotation affecting a string under the scope of a reporting verb (cf. Cappelen & Lepore’s (1997) mixed quotation), and quotation affecting a string outside the scope of a reporting verb (cf. scare quoting). Thus, the quotation in (51) is very likely to be produced with the intention of ascribing the quoted words to the justice ministers designated by the subject of said, while that in (52) is most likely intended to show to the reader that there is something about the word bought that the journalist has reservations about.

Keeping separate the contributions of the convention-based act and of the iconic act has the virtuous consequence that no a priori restriction is placed on the boundaries of the superimposed quotation. The depiction theory thus predicts the occurrence of non-constituent hybrid quotations. Let’s return to (42). We saw that the unquotation account was too demanding in terms of the words produced by the reported speaker, Ings. In particular, Ings was required to have said something to the same effect as which is clearly targeting Wada in the source situation. On the depiction account, it does not matter to the quoter whether the quoted words match with a constituent, and that is because the quotation does not interact with the syntax of the sentence. Hence, whether Ings uttered additional words with the right meaning is immaterial.

For a second illustration, we return one last time to (17). We saw that neither quote-breaking, because it does not preserve textual connectedness, nor unquotation, whose two most plausible outcomes were inadequate, offered a way out. But does the depiction theory have something more convincing to say?

At first sight, it proposes a deceptively similar type of analysis to Maier’s original one: according to depiction theory, in (17), only the first quoted part, a slave to reality, is hybrid. The second, after the full stop, is a non-hybrid instance of free direct speech. Thus, as on the quote-breaking analysis, two different objects are distinguished, with the first quoted string a hybrid while the second is not. However, in contrast to the quote-breaking account, there is no question of the quotation being split in two: it is the same iconic act that spans across the end of the first sentence and the whole of the second, with nothing to disrupt connectedness. It just so happens that initially the quotation is concurrent with part of a convention-based communicative act whereas it subsequently occurs on its own. I take this analysis to be the correct one.

9. Conclusion

In this paper, I have offered a description of non-constituent hybrid quotations. I have shown how they present a challenge to a family of theories which, for theory-internal reasons, have to assume that hybrid quotations must be co-extensive with syntactic constituents. I have briefly discussed a first response to the challenge, the quote-breaking procedure, and dismissed it. I then devoted two long sections to a second response, unquotation, and to the problems which I think it raises. ²¹

Originally introduced into quotation theory by Shan, unquotation has subsequently been adopted by Maier, who has offered very detailed treatments of various quotational phenomena, notably hybrid quotation, that appeal to unquotation. Although I have been quite critical of Maier in this paper, I must make it emphatically clear that his contributions meet the highest standards of inventiveness, of explicitness, and of sheer dedication to taking up all the challenges that empirical data throw his way. So, while I side with e.g. McCullagh (2017: 31) or Gregoromichelaki (2017: 247) in their rejection of the theory – because it is just superfluous, or too ad hoc – I thought the right thing to do was to provide an in-depth critique, one that is sufficiently explicit that Maier or other upholders of unquotation could in turn reply.

In the end, my main grievance is that, when used to repair surface non-constituency, the unquotation mechanism seems to fulfil no role in utterance interpretation and is most likely to go unnoticed by addressees of communicative acts containing a non-constituent hybrid quotation. In other words, it seems to have been devised with the sole purpose of handling empirical data that contradict the constituency constraint. In spite of Maier’s plea for the empirical validity, indeed indispensability, of unquotation, the mechanism seems to correspond to nothing tangible in quoters’ actual practices. ²²

By contrast, the depictive family of theories directly addresses the core feature of quotation – it being an iconic communicative act. Accordingly, in their account of hybrid quotation, they separate the contribution of that iconic act from that of the convention-based linguistic act. The upshot is that they face none of the challenges that plague the theories which have to endorse the constituency constraint.