On the history of NPIs and Negative Concord

Abstract

This article aims to better understand how Negative Polarity Items (NPIs) come into existence and how they change over time. It argues that an expression can become an NPI if its semantics makes it pragmatically useful in negative or downward entailing contexts, often because the meaning leads to pragmatic strength, but sometimes because its semantics leads to pragmatic attenuation. Special attention is given to two patterns involving pragmatic strength that can emerge historically: Negative Concord (NC) and what I call NPI Dualization. Both patterns, I argue, involve a pairing between an NPI that has an existential-like or low scalar semantics with a homophonous but semantically different expression with a freer distribution; the homophone is semantically negative in Negative Concord but semantically universal in NPI Dualization. The article argues that pragmatic strength plays an important role in the history of NPIs, both in their origin and in NPI Dualization, but is not directly relevant for their licensing synchronically. Instead, it argues for a return to the view that NPIs are lexically marked by a semantically meaningless distributional feature that needs to be valued syntactically. On a conceptual level, the article argues that historical shifts may be matters of likelihood.

Résumé

Cet article vise à mieux comprendre comment les items de polarité négative (IPN) voient le jour et comment ils évoluent dans le temps. Il soutient qu'une expression peut devenir un IPN si sa sémantique la rend pragmatiquement utile dans des contextes négatifs ou d'implication vers le bas, souvent parce que le sens conduit à une force pragmatique, mais parfois parce que sa sémantique conduit à une atténuation pragmatique. Une attention particulière est accordée à deux structures impliquant une force pragmatique pouvant émerger historiquement : la Concordance négative et ce que j'appelle la Dualisation des IPN. Je soutiens que ces deux modèles impliquent un jumelage entre un IPN ayant un sémantisme existentiel ou scalaire faible et une expression homophone mais sémantiquement différente avec une distribution plus libre ; l'homophone est sémantiquement négatif dans la Concordance négative mais sémantiquement universel dans la Dualisation des IPN. L'article propose que la force pragmatique joue un rôle important dans l'histoire des IPN, à la fois dans leur origine et dans la Dualisation des IPN, mais qu'elle n'est pas directement pertinente pour leur légitimation en synchronie. En revanche, il plaide pour un retour à l'idée que les IPN sont marqués lexicalement par une caractéristique distributionnelle sans contenu sémantique qui doit être évaluée syntaxiquement. D'un point de vue conceptuel, l'article soutient que les changements historiques peuvent être des questions de probabilité.

1. Background: NPIs bear [u-neg] and are syntactically licensed

I'd like to begin with a few basic facts about NPIs and then lay out the assumptions about how they are licensed that underpin the rest of the article. I hope to show elsewhere in more detail that these assumptions are well-founded.¹

It is uncontroversial that a great many languages have NPIs and that NPIs form a motley crew from a syntactic point of view; they include determiners (any), nominal and adverbial quantifiers (anyone, ever), verb phrases (budge an inch, sleep a wink), focus particles (even in the sense of ‘the least noteworthy’), prepositions (until), coordinating expressions (additive either), etc.

All NPIs are licensed at least by negation, their paradigmatic licensor.

1. (1)

   1. a. Elizabeth didn't let on anything about the secret plan.

   2. b. *Elizabeth let on anything about the secret plan.

In addition, many NPIs are also licensed by a host of other expressions, including the cross-linguistic equivalents of without, no one, before, if, only, emotive factive verbs like regret, negative propositional attitude verbs like doubt, negative verbs of saying like deny, etc. NPIs are also found in comparative and superlative constructions, as well as in direct and indirect questions. Not all NPIs appear in all of the same environments, and NPIs have been roughly grouped into so-called ‘strong’ and ‘weak’ NPIs, where the strong are choosier than the weak (e.g., van der Wouden 1997, Zwarts 1998, Gajewski 2011). Outside of their respective licensing contexts, NPIs result in ungrammaticality and are felt to merit a ‘*’rather than the ‘#’ that signals pragmatic infelicity. All things being equal, this suggests that NPI licensing should be given a grammatical account.

An early syntactic account of NPI licensing was proposed by Klima (1964). He analyzed NPIs as carrying a feature [+affective] that needs to be checked by one of the licensors just mentioned under what corresponds to c-command. This analysis runs into a problem with NPI licensing by quantificational determiners (see Ladusaw 1980). As (2) shows, both every and no license NPIs in their restriction or first argument, but only no also licenses NPIs in its second argument or scope. No matter what definition of c-command one takes, the contrast between no and every presents a problem for Klima's (1964) account, and any similar syntactic account.

1. (2)

   1. a. [Everyone/no one who ever had any dealings with Bob] thinks he is nice.

   2. b. [No one/*everyone] thinks Bob would ever say such a thing.

The every vs. no contrast is addressed in Ladusaw (1980). Following Ladusaw, NPIs are now widely thought to be restricted to semantically downward entailing (DE) contexts or contexts that reverse the usual direction of entailment (Fauconnier 1975). The negation marker not creates such a context: while we can infer that if a person named Amber works for the Department of Labor, she works for the federal government, the direction of entailment reverses from Department of Labor to federal government when the sentence is negated (i.e., Amber doesn't work for the federal government entails Amber doesn't work for the Department of Labor).² Though it is far from trivial to show that all NPI-licensing contexts are in fact DE (e.g., Heim 1984, von Fintel 1999, Horn 2002, Schein 2003), a great many clearly are, including those created by no and every.³ The Ladusaw/Fauconnier generalization, though ground-breaking descriptively, is, however, puzzling because it is not clear why it should hold.

Krifka (1995) and Chierchia (2013) aim to explain this generalization by exploiting the semantics of NPIs (see also Kadmon and Landman 1993, Lahiri 1998, Crnič 2014, a.o.). This type of account is based on the important observation that NPIs are often existentially-quantified expressions. For instance, the NPIs any and ever express existential quantifiers. As such, they occupy the bottom rung of the so-called Horn scale (Horn 1989), invoking alternatives expressed by the semantically stronger many, most and all and often, most of the time and always. NPIs like budge an inch and sleep a wink, which involve minimizers (an inch, a wink), behave similarly, semantically speaking. Existentially-quantified expressions make the weakest assertion, one which alternative sentences involving higher elements on the relevant Horn scale would entail, but in negative or more generally DE environments, matters are reversed and the existential quantifiers are the semantically strongest among the alternatives. Building on this observation, this type of account posits that NPIs come with a strength requirement that stipulates that sentences with NPIs must entail alternative sentences where an element higher on a Horn scale replaces the NPI. This restricts them to DE context. The grammaticality of John didn't eat any cookies is thus related to the asymmetric entailment between John didn't eat any cookies and John didn't eat many/most/all cookies. Conversely, the ungrammaticality of *John ate any cookies is related to its not entailing John ate many/most/all cookies.

Though this idea is appealing, it also raises some issues (e.g., Herburger and Mauck 2013), one being that existential NPIs often have synonymous or near-synonymous counterparts that do not have the distribution of NPIs. The NPI any, for example, expresses existential quantification just like some and a, as shown in (3).

1. (3)

   1. a. [[any]] = $\lambda f_{\left\langle {e, t} \right\rangle }.\lambda g_{\left\langle {e, g} \right\rangle }.[ { { \exists} x\colon \;f( x ) = 1} ] \;g( x ) = 1$

   2. b. [[some]] = $\lambda f_{\left\langle {e, t} \right\rangle }.\lambda g_{\left\langle {e, g} \right\rangle }.[ {{ \exists \hskip 1pt} x\colon \;f( x ) = 1} ] \;g( x ) = 1$

   3. c. [[a]] = $\lambda f_{\left\langle {e, t} \right\rangle }.\lambda g_{\left\langle {e, g} \right\rangle }.[ {{ \exists} x\colon \;f( x ) = 1} ] \;g( x ) = $1

Moreover, a and some activate the same stronger alternatives as any, namely those corresponding to many, most, and all. This is, after all, the original kind of Horn scale and is also what explains (on a neo-Gricean view) the fact that I ate some of the cookies gives rise to the inference ‘I didn't eat all of the cookies’ (e.g., Horn 1989). And yet, any is an NPI, some a Positive Polarity Item (and thus eschews being interpreted in the scope of a local negation), and a has a free distribution.⁴

1. (4)

   1. a. He found some photo/*any photo/a photo in the drawer.

   2. b. He didn't find *some photo/any photo/a photo in the drawer.

The strength requirement is crucial to distinguish NPIs like any from their synonymous non-NPI counterparts some and any on this type of account. Yet, as this requirement does not follow from the semantics of NPIs, it has to be stipulated.

A second issue is that not all NPIs have an existential semantics. A careful look at how NPIs are distributed across the lexicon reveals another, semantically and pragmatically rather different, species of NPIs. Examples include adverbial much (see Klima 1964), adverbial long, Catalan gaire (‘much’), French guère (‘much') and French grand-chose (‘big thing’) (see Israel 2011). Clearly, these NPIs are not low scalar in their meaning. For instance, much occupies a relatively high rung on a Horn scale, unlike typical low-scalar expressions (‘all’ > ‘much/many’ > ‘some’).⁵

Furthermore, the analyses in question do not straightforwardly predict that NPIs are ungrammatical outside of DE contexts and are not just contradictions. Although Chierchia (2013) explores ways of arguing that contradiction can lead to ungrammaticality (‘G-triviality’), all things being equal, an analysis that predicts the ungrammaticality of an NPI that fails to be licensed is preferable.

In the absence of an independent argument for the strengthening requirement on NPIs, and in light of the empirical and theoretical questions the analysis raises, we may consider taking another look at the view where what distinguishes the NPI any from some and a is a semantically meaningless, arbitrary syntactic feature that needs to be valued by expressions that create DE environments. Rather than calling it [+affective] (Klima 1964), we can update the terminology and call it [u-neg] and assume it is valued by the [i-neg] feature stemming from NPI licensors (e.g., not, without, no one, before, if, only, regret, doubt, etc.) In such an account, it would follow that NPIs are ungrammatical outside of DE-contexts. Moreover, it is not a problem that not all NPIs have an existential semantics, nor that many existential NPIs can have synonyms that are not NPIs. But what about the every vs. no issue?

As it turns out, the calculation of Local Polarity through Monotonicity Marking (Sánchez Valencia 1991, Icard and Moss 2013) offers the prospect of a syntactic account of NPI-licensing that, unlike Klima's, can also handle the contrast we see in no vs. every. Concrete proposals along these lines are worked out in Dowty (1994) and Ludlow (2002) and are briefly discussed in the Appendix. I therefore assume a syntactic analysis of NPI licensing, and ask: What happens historically with the NPI-feature [u-neg] and the expressions that bear it?

Exploring the origin of NPIs, section 2 argues that expressions that become NPIs have a semantics that makes them pragmatically useful and frequent in DE contexts. I hypothesize that this, over time, allows speakers to parse them as being marked [u-neg]. On this assumption, pragmatic usefulness provides the chance for an expression to develop historically into an NPI. Which semantically predisposed and pragmatically suitable expressions undergo this grammatical development is to some extent up to chance.

Section 3 explores the possible subsequent development of low-scalar NPIs, in particular their becoming Negative Concord expressions. It argues that Negative Concord is an epiphenomenon and really just involves the pairing of two homophones, one with a negative semantics and a free distribution, and the other with a low scalar semantics (in the wide sense) and a [u-neg] feature. Differences between Spanish-style and French-style Negative Concord are attributed to differences in the overt manifestation of the negation marker and what it tells the speaker about scope-taking. I also explore the loss of Negative Concord, a process that has already taken place in Standard English and German and is, I argue, under way in French.

While sections 2 and 3 discuss well-known patterns and developments, section 4 draws attention to a less widely-noticed generalization, namely that low-scalar NPIs can also have homophones that are semantically universal. I call this pattern ‘NPI Dualization’. NPI Dualization is taken to indicate that pragmatic strength is not only a factor in the genesis of NPIs but it may be playing an important role in NPI Dualization, namely as the one factor that stays constant when both quantificational force and distributional feature make-up change.

2. How NPIs come into existence

As we saw above, NPIs are, syntactically, highly heterogenous and can be of many syntactic categories. Semantically, however, they cluster in interesting ways. We can likely get a better understanding of why NPIs exist and what happens to them diachronically by looking at their semantics, and exploring their pragmatic effects.

2.1 The pragmatic usefulness of NPIs

Despite their syntactic heterogeneity, on a pragmatic level NPIs tend to fall into one of two different classes, which I argue is no accident. One class is larger and better studied than the other.

As has long been noted, many NPIs involve existential quantification expressing the smallest possible amount in some general sense (‘existential NPIs’). These can be basic existential quantifiers and determiners (e.g., any, anyone, ever), expressions of small quantity (a single), of a basic entity (a thing, a soul), or a small amount of a thing of little value (‘minimizers’, e.g., a hair, an iota). The latter also appear in collocational NPIs like give a damn, have a clue, budge an inch, sleep a wink. Some combine various of these properties. A related class of NPIs consist of or incorporate ‘even’ expressions in the sense of ‘the least noteworthy’. These ‘even’ NPIs include German auch nur, Italian anche solo. An existential and an ‘even’ expression appear jointly in entire paradigms of NPIs in various languages, including Hindi, Korean and Hebrew (see Lee and Horn 1994, Lahiri 1998). While existential and ‘even’-NPIs (and combinations thereof) seem to be the largest, and also most discussed class, we must also acknowledge NPIs like adverbial much (see Klima 1964), adverbial long, Catalan gaire (‘much’), French grand-chose (‘big thing’), certain uses of impressed, etc. These NPIs, discussed in detail in Israel (2011), occupy not the bottom rung on a Horn scale, but a (relatively) high rung.

Existential NPIs, ‘even’-NPIs, and relatively high scalar NPIs are pragmatically useful in DE contexts, but for different reasons. The first two, which I will subsume under the term ‘low scalar NPIs’, lend themselves to making semantically very strong, even emphatic claims in DE contexts: if Nico didn't find a single/even the slightest mistake in the calculation, he certainly didn't find three. If Mary won't budge an inch on her position, she certainly won't meet John half way. If Bill cannot even fry an egg, he certainly cannot prepare a three-course dinner. Disjunctive NPIs like additive either can also be included in this category.

Relatively high scalar NPIs have a different pragmatic function. As Israel (2011) notes, they generally attenuate matters in DE contexts. Attenuation is pragmatically useful, for instance when one wants to soften a negative comment or the force of a requirement; one can use Yolanda was not impressed by the design of the garden when one means ‘She disliked it.’ The fact that this type of NPI generally receives less attention likely has to do with the fact that it is less common. This makes a certain amount of sense. Strength and emphasis are communicatively important when expressing negation, attenuation arguably less so.⁶

2.2 Becoming an NPI is becoming [u-neg]

One way to conceptualize how an expression can become an NPI, or how it can come to carry [u-neg], is to note that its pragmatic usefulness in DE contexts can result in its frequent appearance there. We can then hypothesize that this allows the learner to parse such an expression as being formally restricted to such contexts, that is, as carrying the feature [u-neg]. This new parsing does not involve a change in meaning. Nor does it involve extra effort on part of the speaker, if we assume that NPI-marking and the valuation of the [u-neg] feature is provided for free by Universal Grammar, and is something that speakers already have at their disposal. On this view, expressions with a particular kind of semantics are susceptible to being parsed as bearing the NPI-feature [u-neg] because of their pragmatic usefulness in DE contexts, which follows from their semantics. But exactly which of the semantically and pragmatically suitable expressions actually end up being marked [u-neg] is to some degree up to chance (and/or perhaps other, yet to be determined factors). This probabilistic perspective suggests that in principle NPIs should be able to have synonyms that are not themselves NPIs. The coexistence of any, some and a discussed in section 1 can be taken to show just that.

One consequence of an expression becoming an NPI is the semantic bleaching that is often observed. While this is not relevant for NPIs like any, which express only existential quantification and are semantically as bare as can be, it characterizes many minimizing NPIs. Though inch literally denotes a measurement of distance (2.54 cm), the NPI budge an inch involves no real talk of measurement of distance. Semantic bleaching serves as a tell-tale sign to a learner that an existential noun phrase has become (part of) an NPI. Similarly, though ‘even’ NPIs are initially emphatic, their emphatic character diminishes as time goes by (see e.g., Kiparsky and Condoravdi 2006).

2.3 The possible loss of [u-neg]

Given the relative arbitrariness of which semantically predisposed expressions acquire [u-neg] – any did it, some and a did not – it would now not be surprising to find that expressions can lose their NPI-hood without change in meaning, reverting to a less restricted distribution. While the literature generally argues for a development from ‘less negative’ to ‘more negative’ (see e.g., Haspelmath 1997, Breitbarth et al. 2020), the loss of NPI-hood without attendant change in meaning is not unheard of (see e.g., Jäger 2010, Herburger and Mauck 2013).

One example illustrating the loss of [u-neg] is Dutch ooit (‘ever’). According to Hoeksema (1998), although until the 1960s ooit had the distribution of a weak NPI, as in (5), in present day Dutch it can also be used outside of DE contexts, as in (6):

The recent development of how speakers use ooit suggests that it has lost its [u-neg] feature, which limited its distribution earlier, while retaining its meaning as an existential adverbial quantifier.⁷ In other words, the relevant speakers used to have something like (7a) in their lexicon, now it is something like (7b):

1. (7)

   1. a. [[ooit _[u-neg]]] = $\lambda f_{\left\langle {ev, t} \right\rangle }.\; { \exists \hskip 1pt} e\;f( e ) = 1$.

   2. b. [[ooit]] = $\lambda f_{\left\langle {ev, t} \right\rangle }.\;{ \exists \hskip 1pt} e\;f( e ) = 1$

Further examples where a [u-neg] feature seems to have been lost without any change in meaning include German einig- and jemand. The determiner einig-, which shares a root with any (‘one’) is now used as a regular existential, without any distributional restriction to DE contexts. But it used to have the distribution of an NPI (Jäger 2010). Jäger similarly argues that Old High German ioman changed from existential, low scalar NPI to eventually becoming the regular indefinite jemand. Even NPIs that contain an overt instance of ‘even’ can shed their NPI-hood while keeping their existential meaning (Herburger and Mauck 2013). As these examples illustrate, just as an expression with a suitable semantics can come to bear the feature [u-neg], an expression that is lexically marked [u-neg] can also come to lose its marking without any shift in meaning. On a more general level it suggests that the history of NPIs is not unidirectional (see e.g., Haspelmath 1997, Breitbarth et al. 2020). I briefly return to this matter in section 3.7.3.

3. Negative Concord

NPI-hood cannot just be gained or lost; NPIs can also morph into Negative Concord terms (NC-terms). Looking mainly at data from Romance, in this section I show that the Medieval Romance languages featured a series of expressions that appeared in a wide variety of negative contexts but whose distribution has since shrunk, as far as these contexts are concerned. At the same time these expressions have started to appear elsewhere, having increasingly gained the ability to express negative meanings on their own in a class of environments that can be systematically defined. Building on these observations I argue that Negative Concord in Romance typically arises when a [u-neg] expression with a low scalar, existential-like interpretation comes to be paired with a semantically negative, [i-neg] homophonous counterpart. Independent factors related to scope are responsible for the two interpretations having an almost complementary distribution.⁸

3.1 NPIs becoming stronger

Medieval Romance developed a series of existential NPIs along the lines sketched initially in section 2. Examples include among many others Spanish nadie (< hominem natum ‘man born’), nada (< res nata, ‘thing born’), French rien (< res, ‘thing’), French personne, Catalan cap (< caput ‘head (of cattle)’), etc. The forebears of these expressions in Medieval Romance readily appeared in all sorts of NPI contexts, including if-clauses, questions and comparatives (see Martins 2000, Eckardt 2006, and Breitbarth et al. 2020, a.o.).

The acceptance of these expressions in NPI-contexts has decreased in Modern Romance. But there is considerable variation among the Romance languages in this regard (see Martins 2000). Within Iberian Romance, we find that in European Portuguese the relevant expressions now tend to be restricted to the scope of negation, ‘without’ or a negative quantifier. On the other hand, in Catalan they are still possible in interrogatives and in if-clauses:⁹

Unlike Portuguese, Spanish no longer features these expressions in if-clauses or questions (unless the questions are rhetorical and have a negative bias), but still allows them in the NPI-contexts created by ‘doubt’, ‘prefer, ‘before’, comparatives and even in the restriction of universal quantifiers in some instances (see also Bosque 1980, Laka 1994).

The gradual narrowing to more obviously negative contexts has constituted a long-standing philological puzzle. An important insight was that it can be thought of as a process of ‘weak’ NPIs turning into ‘strong’(er) ones, a point argued for in Martins (2000), Eckardt (2006) and others.¹⁰

Interestingly, these expressions in Old Spanish also appeared together with negation in elliptical contexts, both in coordination, as in (10), and in elliptical or fragment answers, as in (11). This is no longer possible in Modern Spanish.

In Catalan, which retains various Medieval properties, e.g., res and gens can still optionally appear with negation in elliptical answers, recalling in this respect the pure NPI gaire (‘much’), which requires a negation (e.g., Vallduví 1994).

In sum, it appears that in Medieval Romance, what are today considered NC-terms generally appeared in a wider series of negative contexts than they do today, including an extensive series of NPI contexts, and contexts involving verbal ellipsis.

3.2 Negative Concord: Important data

NC-terms in Modern Romance crucially differ from univocal NPIs, which Modern Romance languages also have (e.g., Spanish N algun- ‘any’), in that in addition to their having low-scalar readings in (a subset) of NPI contexts, they can also appear on their own with a negative interpretation. A paradigmatic context would be the case of elliptical answers, where pure NPIs are not possible, (see Who came? *Anybody.)^:

In addition, negative readings of NC-terms are also found in cases of verbal ellipsis in coordination, as shown for Spanish in (14) and Romanian in (15):¹¹

It is theoretically significant that even in postverbal position, NC-terms can have a negative interpretation on their own. They do so when they take narrow scope with respect to the event quantifier binding into the verbal predicate, which prevents them from expressing ‘sentence negation’ (e.g., Herburger 2001). Since such a narrow scope construal is not often feasible, examples of postverbal NC-terms without a higher negation or NC-term are not numerous. This, however, does not mean that they are not productive and thus should be set aside as idioms (see Penka 2011).

NC-terms with a negative interpretation can also be found postnominally when they take narrow scope, as in el viaje a ninguna parte (‘a trip to nowhere’), una serie sobre nada (‘a show about nothing’). Adding a higher negatively interpreted expression does not change the scope properties of the postnominal NC-term in such instances. We can see this in (17), which combines a negative, narrow scope postnominal NC-term (nada) with a preverbal negative NC-term with sentential scope (nunca). Though both NC-terms are negative, the result is not a double negation, because the scope of nada is below the nominal serie.

Finally, as is well-known and much discussed, Modern Romance languages differ with respect to the behaviour of preverbal NC-terms. In some languages and varieties, preverbal NC-terms co-occur with negation to express a single negation. This is obligatory in Romanian, as is illustrated in (18), and optional in Catalan.

And though in Medieval Spanish, Italian and Portuguese, the expressions in question often also appeared with negation preverbally, to express ‘sentence’ negation, already in the Medieval languages, they could at times appear without negation, alternating even within a single text, as pointed out in Martins (2000).

In Modern Italian, Portuguese and Spanish, this is the general pattern. In these asymmetric or non-strict Negative Concord languages, preverbal NC-terms routinely occur without negation, expressing sentence negation, as is illustrated in (20a). Adding negation results not in ungrammaticality, as is sometimes claimed, but in a double negation interpretation. Pragmatically, (20b) is felicitous as a denial of a previously uttered negative claim (see A: Nobody came. B: No, nobody DIDN'T come; A: Michelle didn't come. B: No, NOBODY didn't come) and is characterized intonationally by stress on the NC-term or the negation and a L+H*L! H% intonation contour (see e.g., Labov 1972, Espinal et al. 2016). In languages that have symmetric Negative Concord, double negations are distinguished only prosodically (see Espinal et al. 2016).

Based on this description of the facts, we can now draw the descriptive generalization in (21), which assumes that sentences are descriptions of event(ualitie)s, and that in the absence of an overt adverb of quantification (e.g., always, often) a tacit existential quantifier $\exists e\;$binds the event variable of the verb.

1. (21)

   1. a. In languages with asymmetric or non-strict Negative Concord like Spanish, NC-terms can have a negative interpretation when this does not require them to take inverse scope over the event operator $\exists e\;$that binds the event variable of a verb that is pronounced.

   2. b. In languages with symmetric or strict Negative Concord like Romanian, NC-terms can have a negative interpretation when this does not require them to take linear or inverse scope over the event operator $\exists e$ that binds the event variable of a verb that is pronounced.

This generalization means that in Spanish and Romanian, an NC-term that appears after an overtly realized verb can in principle be interpreted negatively but cannot take scope over ∃e and express sentence negation. However, when the verb is elided, as it is in elliptical answers and elliptical conjunction, a postverbal NC-term can be negative and express sentence negation, taking scope over ∃e that binds the event variable of the elided verb. It is, in other words, as if the overtly pronounced verb blocks non-linear scope. Whereas preverbal NC-terms can be interpreted negatively in a Spanish-type language taking linear scope over ∃e, this is not possible in a Romanian-type language. The difference between strict or symmetric and non-strict or asymmetric Negative Concord reduces to whether a preverbal NC-term can take scope over the event quantifier binding the variable of the pronounced verb on its own (non-strict) or not (strict). Languages like Catalan currently seem to allow for both options; preverbal NC-terms can, but need not, express sentence negation on their own.

3.3 Negative Concord as homophony¹²

Many of the facts described in (21) can be readily captured with a feature-based account. After showing how, I explore the historical development of NC-terms.

3.3.1 Synchronic facts

The feature-based account of NPI-hood argued for in section 1 fits organically with the view that Negative Concord is an epiphenomenon, merely a name for a systematic lexical ambiguity (see Herburger 2001). NC-terms, in this view, come in two phonologically indistinguishable versions, different in meaning and distributional features. One version has a low-scalar semantics (e.g., existential, ‘least noteworthy’, disjunctive) and bears the distributional feature [u-neg]. The other lacks [u-neg] and is semantically the negation of the low-scalar meaning; we can say it bears [i-neg]. The NC-terms nadie and nada, for example, thus have the following kinds of lexical entries:

1. (22)

   1. a. [[nadie _[u-neg]]] = $\lambda f_{\left\langle {e, t} \right\rangle }.\;[ {{\exists \hskip 1pt} x\colon \;Person\;( x ) } ] \;f( x ) = 1$

   2. b. [[nadie _[i-neg]]] = $\lambda f_{\left\langle {e, t} \right\rangle }.$ [∄${x\colon \;Person\;( x ) } ]\; f( x ) = 1$

2. (23)

   1. a. [[nada _[u-neg]]] = λf _〈e,t〉. [∃x: Thing(x)] f (x) = 1

   2. b. [[nada _[i-neg]]] = λf _〈e,t〉. [∄x: Thing (x)] f (x) = 1

Various facts observed earlier find a simple explanation under the homophony view. One is that just as NPIs are not restricted to elements in the nominal domain (‘indefinites’), neither are NCs; they can also be adverbials (nunca ‘never’ ‘ever’), scalar expressions (ni siquiera lit. ‘not even’ ‘even’), or conjunctions (ni…ni ‘neither nor’, ‘either or’), etc. This follows directly from the claim that NC-terms are, on one of their readings, NPIs, which are also not restricted to expressions of particular syntactic categories. The cross-categorial parallel that we see between NC-terms and NPIs is more difficult to explain in analyses that treat Negative Concord and NPI licensing as separate phenomena (e.g., Zeijlstra 2004, Penka 2011).

It also follows that NC-terms receive a negative interpretation when they appear on their own. As we saw in section 3.2, they do this in elliptical answers, conjunctions involving verbal ellipsis, postverbal or postnominal narrow scope, and in preverbal position in asymmetric or non-strict Negative Concord. In all these instances, the negative interpretation simply follows from the fact that in the absence of an NPI licensor, the negative reading of the NC-term is the sole grammatical option; employing the [u-neg] version would leave its feature unvalued.¹³

The homophony analysis captures the fact that, to varying degrees, NC-terms continue to appear in the scope of NPI licensors with non-negative meanings. Thus, in all Negative Concord languages, NC-terms can appear not only under ‘not’, another negatively read NC-term (‘Negative Spread’), but also under ‘without’. And as we noted, in various languages/varieties, they can also still appear under other NPI-licensors; in Spanish, for example, they can appear under ‘before’ and ‘doubt’.

Finally, the homophony analysis also predicts certain ambiguities. We expect, and indeed find, ambiguity when an NC-term appears preverbally in an embedded clause that is itself in the scope of an NPI-licensor. The ambiguity of (26) supports this. Its simple negation reading arises when nadie _[u-neg] is licensed by ‘doubt’, as in (27a). The double negation reading arises with nadie _[i-neg], as in (27b). (The characteristic prosody and context are needed for the double negation reading.)

3.3.2 Diachronic facts

Turning to the historical trajectory of NC-terms, we can now say that Negative Concord can arise when a univocal low scalar NPI bearing [u-neg] acquires a semantically-negative homophone with a free distribution: [u-neg] > [u-neg]/[i-neg]. This characterizes the historic origin of many NC-terms as low-scalar expressions (e.g., nada < res nata, cap < caput, res < res etc.). It is, however, also worth bearing in mind that quite a few NC-terms derive from expressions that were semantically negative in Classical Latin. These include Spanish nunca (< numquam ‘never’), Romanian neminem (< nemo ‘nobody’), and French nul and nulle part < (nullus ‘none’). In addition, many varieties of Modern Romance feature words that derive from Latin nec, which meant ‘(neither) nor’ and ‘not even’. These include Spanish ninguno (‘nobody’, ‘anybody’), ningun- (‘no’, ‘any’), Italian niente (‘nothing’, ‘anything’), Italian nessuno (‘nobody’, ‘anybody’), along with conjunctive expressions ni…ni (‘neither… nor’, ‘either… or’) and scalar particles like ni (‘not even’, ‘even’). The existence of NC-terms that historically derive from semantically negative expressions suggests that the homophony known as Negative Concord can arise from either side, either from a low scalar [u-neg] expression acquiring a negative [i-neg] homophone or from a negative [i-neg] expression acquiring a low scalar [u-neg] homophone.

In section 2.2 we noted that the choice of which semantically predisposed expression acquires the distributional NPI-feature [u-neg] is to some degree an arbitrary matter (e.g., any vs. some vs. a). It would now not be surprising to also find variation in whether or not a [u-neg]-bearing expression comes to be paired with a semantically-negative [i-neg] counterpart. The descendants of Latin aliquis (‘some (or other)’), discussed in detail in Gianollo (2018), provide an interesting example of this. Latin aliquis developed into the Spanish regular indefinite quantifiers algo (‘something’) and alguien (‘somebody’). It also developed into an epistemic non-NPI, prenominal indefinite determiner (algun-) (‘some (or other)’). In postnominal position, however, algun- in Spanish functions as an NPI. Going one step further, its Portuguese counterpart (N algum-) has the distribution of an NC-term. The same is true of the French descendant of aliquis, aucun.

This type of variation, which may seem puzzling, makes sense in the present perspective. What happens to aliquis in a particular Romance language is to some degree a random matter within a certain set of possibilities (stays existential, becomes NPI, or becomes also NC-term).¹⁴

3.4 On Negative Concord in French

Negative Concord in French shows an interesting development when compared to the languages considered so far. I will suggest that this difference is related to the history of sentential negation, in particular to the postverbal realization of negation in French as pas, and to the cost of inverse scope.

As in other Romance languages, French NC-terms typically functioned as NPIs in the medieval varieties (see Eckardt 2006), and continue to appear in the NPI environment provided by ‘without', and optionally (increasingly less) in the NPI environments provided by the comparative, ‘before', ‘not believe', rhetorical questions, etc. Their NPI-side is also visible when NC-terms are interpreted under a negation in a higher clause. The following examples are from Milner (1979), and are annotated for the relevant features.

In preverbal position, as in (31a), French NC-terms receive a negative interpretation, indicating asymmetric Negative Concord. This claim is based on the assumption that it is pas rather than ne that translates as semantic negation (see Jespersen 1917, Rowlett 1998). Support comes from the fact that in spoken French ne is readily omitted, and also from the fact that adding pas results in a double negation (with the typical prosodic and pragmatic characteristics) (e.g., Muller 1991):

Finally, as in other Romance languages, French NC-terms appear with negative interpretation on their own as elliptical answers (as in (32)–(33)), in elliptical coordination (as in (34)), and postverbally/postnominally with narrow scope (as in (35)). Interestingly, in examples like those in (35), ne has to be absent to express the desired reading where the NC-term takes scope below the event quantifier ∃e that binds the event variable of the verb. Ne, whose position is high, seems to be incompatible with this interpretation. Adding ne in (35a) would not be ungrammatical, but rather change the meaning, negating that a given event took place. This suggests that ne, when present, functions as a scope marker.

Other than how negation is realized, so far it might appear that French Negative Concord and Spanish/Italian/Portuguese Negative Concord are parallel. But, in striking contrast with the Spanish (36), which is interpreted as a single negation (‘Negative Spread’), its French counterpart in (37) is ambiguous between a single or double negation, that is between (38a) and (38b) (Corblin 1996, Muller 1991, de Swart and Sag 2002, a.o.). (As usual, the double negation reading has special prosodic properties; see Déprez 2018 and Yeaton 2018.)

The double negation reading of (37) indicates not just that both personne and rien are negative, but also that rien, despite being postverbal, takes scope over the event quantifier $\exists e$, something its Spanish counterpart nada crucially cannot do. If so, a sentence where a NC-term appears postverbally on its own (with optional ne) should be able to express sentence negation, as is in fact confirmed in (39). Also, adding pas results in a double negation, contrasting in this respect with univocal NPIs like qui que ce soit. (Examples are from Muller 1991:258)

When a postverbal NC-term occurs with pas, the availability of a negative interpretation seems to effectively block a [u-neg] interpretation of the NC-term, which would result in a simple negation reading. This may have to do with the fact that a simple negation reading could be achieved more economically by simply using the [i-neg] version of the NC-term, without bothering with sentential negation. The fact that in older varieties (e.g., 17^th century French), pas + NC-term was read as a simple negation – as is still possible in Québecois and in Haitian Creole (see Muller 1991, Déprez 1997, de Swart and Sag 2002) – indicates that the ability of postverbal negative NC-terms to take scope over the event quantifier is an innovation, relatively speaking.

To illustrate the properties of postverbal French NC-term we can consider (42), an example that I owe to an anonymous reviewer, who notes that it is four ways ambiguous. While in this sentence personne can only be taken to be negative, since there is no NPI licensor above it, jamais and rien can be interpreted either as NPI expressions licensed by the negative personne, or as negative expressions that outscope ${ \exists \hskip 1pt} e$. The readings are shown in (43):

In sum, while French NC-terms are like their Spanish and Romanian counterparts in that they are ambiguous between a [u-neg] and an [i-neg] interpretation, they are different in that the postverbal [i-neg] versions can take inverse scope over $\exists e.$

3.5 Postverbal NC-terms and Jespersen's cycle

At this point we may wonder why French but not Spanish NC-terms should have acquired the ability to take inverse scope on their negative reading. A look at micro-variation helps shed light on the possible reason for the difference. Modern European French is not alone in allowing postverbal NC-terms on their negative readings to outscope the event quantifier. Such readings have also been reported for ‘italiano populare’ of the 20^th century, and various Italian and Rhaeto-Romance varieties in northern Italy and Switzerland (Bernini and Ramat 1996, Zanuttini 1997, a.o.):

An interesting property shared by these varieties and French is that their negation marker (‘not’) is overtly realized below the tensed verb (see Bernini and Ramat 1996, Zanuttini 1997, a.o.) even though semantically it typically takes scope over it. Put differently, whereas Spanish and most Italian varieties are ‘NEG1’, the varieties mentioned here are ‘NEG3’ (and some even ‘NEG4’, Zanuttini 1997), and French is NEG2 or NEG3, depending on whether the register employs ne.

The low negation no/beka/pas reflects a step in what is called the Jespersen Cycle. According to Jespersen (1917), initially, the semantic negation was the preverbal ne, which was reinforced by the postverbal pas in French, which then itself became the negation without changing its postverbal position. The erstwhile negation ne became largely optional in spoken French in most instances, and, when pronounced, now marks the scope of the NC-term. This trajectory can now be represented as follows:¹⁵

The negation and the postverbal NC facts seem to be related (and are actually presented as one single fact, in Bernini and Ramat 1996). In particular, we can assume that once the semantically active negation marker is realized postverbally (NEG 2,3,4), speakers have evidence that a semantically negative expression (e.g., no, beka, pas) can take non-linear scope over the event quantifier and thereby express sentence negation. At some point, they can extend this to postverbal negative NC-terms, which thus acquire the capacity to outscope ${ \exists \hskip 1pt} e$ and express sentence negation. In varieties that pattern with Spanish, in contrast, the negation marker is robustly realized preverbally (NEG1), offering speakers little reason to assume that the surface position of a postverbal negative expression does not reflect its semantic scope. In order to express sentence negation with a postverbal NC-term, its [u-neg] version must be employed with a preverbal [i-neg] licensor.

Table 1 summarizes the development of Negative Concord in Romance as analyzed here.

Table 1: The development of Negative Concord in Romance

3.6 The demise of Negative Concord

Given the many [i-neg] uses of French NC-terms, at some point the [u-neg] uses may disappear, at which point French will cease to be a Negative Concord language and turn into a Double Negation language:

This has already happened in English, German and various other Germanic languages. For instance, the German negative determiner kein- (‘no’) functioned as an existential NPI determiner dehein(ig)/kein before it developed into an NC-term in Middle High German (Jäger 2010). It could then not only appear with existential meaning in NPI contexts, as in (48a), but also alone with negative meaning, as in (48b). While Negative Concord with kein- is still possible in some dialects (e.g., Bavarian, see Bayer 1990), in Standard Modern German kein- no longer shows Negative Concord but only has a negative reading. In other words, while Middle High German featured two versions of kein-, a [u-neg] one with existential meaning, and a negative [i-neg] one, modern standard German retains only the [i-neg] version. The trajectories of niemand (‘nobody’) and nichts (‘nothing’) are reported to be parallel (see Jäger 2010), and they too can be said to have developed from existential NPI to NC-term to negative quantifier.

The history of English also offers numerous examples of NC-terms becoming univocally negative expressions. Negative Concord was common in Old English and still possible in Middle English, as we can see in (49).

It then became less frequent in the late Middle English and Early Modern periods and is now no longer part of Standard varieties, where nobody is purely negative. Negative Concord continues, however, to be part of many traditional British dialects (e.g., Tubau 2016), of many dialects of white speakers in the US, and it is a systematic part of Black American English (e.g., Labov 1972). These dialects show considerable variation as to whether they employ asymmetric or symmetric Negative Concord.

Summarizing, we have seen how low scalar [u-neg] expressions can come to have a negative [i-neg] homophone and vice versa, giving rise to Negative Concord. The differences we observe in Negative Concord in Spanish-type vs. French-type languages stem from the inverse scope capabilities the negative readings of postverbal NC-terms have acquired in French-type languages. This in turn is tied to the emergence of a postverbal negation marker, which develops when a [u-neg] direct object morphs into an [i-neg] adverbial/sentential operator. It offers the learner evidence that postverbal negative expressions can take non-linear scope and express sentence negation; the learner can then extend the strategy to other postverbal [i-neg] expressions. Once all postverbal [i-neg] expressions can take scope over the event operator, there is pressure in the system to not employ [u-neg] versions for that, significantly decreasing their uses. This signals the beginning of the end of Negative Concord, paving the way to a Double Negation system.¹⁶

3.7 What drives the change?

As we have seen, the history of Negative Concord in Romance is one of change. This raises the question what drives this change?

3.7.1 Contagion

One way to conceptualize the genesis of Negative Concord is as involving contagion. It is as if a [u-neg] expression, which requires an [i-neg] expression to be licensed, comes to be ‘infected’ with negation and starts to carry the semantically active [i-neg] feature itself. Rather than saying ‘I need some [i-neg] to be grammatical’, it starts to say ‘I'm [i-neg] myself’. Bréal (1897: 222, quoted in Jesperson) expresses this idea of contagion, where speaking of French rien, personne, pas etc. he notes “These words, through their association with ne, have themselves become negative. They have done this to an extent that they do not need their companion. ‘Who is there? Nobody’ …” After looking up the meaning of these expressions in various dictionaries, he observes: “The two answers one obtains are contradictory, but, upon reflection, although they are opposite, they both have their reason to exist and their legitimacy.”

3.7.2 The progression of the [i-neg] reading

In light of the variation within Romance described above, we can now hypothesize that verbal ellipsis, elliptical answers and coordination structures are early targets for low scalar [u-neg] marked expressions starting to show negative, [i-neg] marked counterparts. As the data in (10)–(11) suggests, in Old Spanish, expressions like nadie etc. appeared together with negation in elliptical coordination structures and elliptical answers, something that is no longer possible. The contrast between symmetric and asymmetric Negative Concord furthermore suggests that [i-neg] readings next become possible in preverbal position (Spanish vs. Romanian). As I have described in (21), this difference has to do with whether a preverbal [i-neg] expression can outscope $\exists e$; it can in asymmetric NC, but not in symmetric Negative Concord. This does not seem to be a particularly difficult, deep or parametric matter for preverbal expressions. This may help explain why some languages, for example Catalan, can do both, and it may also help explain the considerable variation we find in this regard within dialects of English that feature Negative Concord (e.g., Labov 1972, Tubau 2016).¹⁷ The last place where [i-neg] versions come to take scope over $\exists e$ are postverbal positions (see French vs. Spanish). Finally, once [i-neg] versions enter the picture, they can also be employed with narrow scope reading relative to an event operator that binds into a verbal or nominal predicate.

1. (50) Progression of [i-neg] vis-à-vis [u-neg] :

   Verbal ellipsis > preverbal position > postverbal position

Why should the spread of the [i-neg] versions take place in this order and not the other way around? The progression suggests that there is considerable cost to taking scope over the event operator, especially when the verb is overtly realized. This cost keeps the [i-neg] version restricted to certain environments where this is relatively easy. The result is that the [u-neg] and [i-neg] readings of NC-terms appear in almost complementary distribution. The quasi-complementary distribution, in turn, makes Negative Concord eminently learnable, reinforced by the fact that (symmetric) Negative Concord is a common feature of Creole languages (e.g., Haspelmath 1997). We may surmise that what makes the ambiguity attractive to the learner is that, by and large, she can associate the presence of an NC-term with sentence negation, either because the NC-term is [u-neg] and licensed by an [i-neg] expression, or because it is a wide-scoping [i-neg] expression itself. One way that this balance represented by Negative Concord can tip is when sentence negation itself comes to be realized postverbally, as in French; the overtly postverbal negation then signals to the learner that it is possible for postverbal [i-neg] expressions to take scope over ${ \exists \hskip 1pt} e$ and express sentence negation that way. This can then be extended to NC-terms.

So far, we have considered how Negative Concord can arise from low-scalar expressions, but, as we noted earlier in section 3.3.2, it can also arise from negative expressions. A number of the negative expressions of Classical Latin (e.g., nihil ‘nothing’, nemo ‘nobody’) did not survive while others did, for instance, numquam ‘never’, nec ‘not even’, nec…nec ‘neither…nor’), nullus ‘none’ and they function as NC-terms in Modern Romance. How this development took place is studied in Gianollo (2018), who focuses on expressions deriving from nec. While the details are beyond the scope of this article, the Latin system began to change, according to Gianollo, when movement to a special pre-Infl position became impossible for negative quantifiers (including object negative quantifier). Thus, when the pre-Infl order became unavailable for negative quantifiers that were objects, it became impossible for them to take scope over ${ \exists \hskip 1pt} e$ and to express sentence negation. But losing this major function of a univocal negative quantifier makes it of little use overall, which in turn makes it fall into disuse. This appears to have happened with nihil and largely with nemo. Alternatively it can be repurposed as an NC-term, as seems to have happened with the descendants of nec, nullus and numquam.

3.7.3 Directionality and the Quantifier Cycle

Breitbarth et al. (2020) argue that a ‘Quantifier Cycle’ characterizes the historical development of existential determiners and quantifiers. Their claim, which builds on Haspelmath's (1997) indefinite map, is that the development generally proceeds from ‘less negative’ to ‘more negative’:

1. (51) Quantifier Cycle: (Breitbarth et al. 2020)

   existential quantifier > weak NPI > strong NPI > NC-term > negative quantifier

While there are numerous examples that fit this schema well, the negative expressions of Latin that survived into Romance can be taken to suggest that the historical development is more complex and less unidirectional. As far as Spanish ningun- (< nec unus) is concerned, not only did it change from negative expression to NC-term, the repurposing to the eventual NC-term seems to have involved a stage in Old Spanish where it was a pure NPI. As we saw in section 3.1, even in elliptical answers and conjunctions it did not appear on its own but was rather accompanied by non. The ‘counter-cyclic’ development of Spanish ninguno may perhaps be attributed to analogy to the development of NC-terms that arose from low-scalar expressions (e.g., Haspelmath 1997, Breitbarth et al. 2020). However, it is not clear that this type of argument extends to the loss of NPI-hood that we saw in connection with expressions like Dutch ooit. Historical development may at times indeed proceed from ‘more negative’ to ‘less negative’ contrary to what is stated in (51), see also Jäger 2010).

The quantifier cycle in (51) also suggests that an expression's becoming an NC-term is preceded by its becoming a strong NPI. Though there is clearly a gradual narrowing of NPI contexts in which NC-terms can appear in Modern Romance when compared to Medieval Romance, this narrowing, as we noted, varies considerably among Romance languages. It appears to be rather advanced in Portuguese but considerably less so in Catalan, for instance, with Italian, French and Spanish occupying various places in between. And yet, all five languages have Negative Concord. This suggests that narrowing need not precede an NPI's development into an NC-term, but may in fact follow it. Conceivably, once an existential/‘even’ NPI expression acquires a second, semantically negative [i-neg] interpretation, its association with negation becomes intuitively stronger for the learner, so much so that the contexts where the non-negative, NPI readings can appear in reduced number, to varying degrees across languages, to more evidently negative and more easily acquired contexts (e.g., under ‘not’, ‘without’, and [i-neg] NC-terms). This not only fits with Breitbarth et al's (2020) view that a gradual narrowing of NPI-contexts has to do with the relative difficulty of acquisition, but in fact motivates the narrowing: the existence of an [i-neg] version strengthens the association with the negation.

4. NPI Dualization

We have explored in considerable detail the relation that low scalar expressions marked [u-neg] may have with semantically negative homophones. Next, I show that NPIs with an existential semantics can also come to have a quite different homophonic partner, namely one with a universal-like (rather than negative) semantics. I will refer to this pattern, which is not part of the Quantifier Cycle in (51) and has received less attention, as NPI Dualization.¹⁸

4.1 Examples of the dualization of [u-neg] expressions

Ever provides an example of NPI Dualization. It is now generally considered an NPI, and appears in various NPI contexts such as those in (52), where it is interpreted existentially along the lines of ‘at any time/at some time’. Historically, however, ever also meant ‘always’ and appeared outside of DE contexts, as we can see in (53). The universal reading of ever is also still evident in ever after, forever, ever since, ever + adjective, and also a multitude of brand names (e.g., Everlast). And while the use of ever as ‘always’ outside of these locutions is no longer common, (54) illustrates that it can appear under certain conditions:

Both the low scalar [u-neg] reading and the universal readings of ever appear together to striking effect in this 17^th century poem, with the existential NPI reading in the first three lines, the universal, non-NPI reading in the last:

According to Jäger (2010), German immer developed from an existential NPI io mer to its current non-NPI version as a universal adverb of quantification (‘always’). This is possibly another example of NPI Dualization, only in this case the development proceeds in the other direction, from an existential [u-neg] marked adverb to an adverb of quantification with universal force.

Given what we have seen so far, it may appear that NPI Dualization is restricted to adverbial quantifiers. The expression at all, however, suggests otherwise. At all clearly functions as an NPI in Modern English, and can be paraphrased as ‘to any degree’, ‘in any way’, ‘in the least’, ‘whatsoever’. But, as the OED attests, at all had a universal reading in Middle English, meaning ‘in every way’, ‘altogether’, ‘wholly’. This latter reading can still be seen in (57). Related positive uses are also reported for Irish and Caribbean English, where at all seems to mean ‘of all’, and also for US regional varieties, where at all is read as ‘only’, as in (58).

Synchronically, the universal all in at all seems in conflict with the low scalar meaning of the NPI expression. NPI Dualization helps us make sense of the disconnect between form and meaning. In particular, we can say that in the past at all was semantically universal, but, as an instance of NPI Dualization it came to have an [u-neg] homophone with existential semantics. This second version is the one we encounter in most modern dialects of English.¹⁹

At this point one may also wonder about the two readings of any – the existential reading with the distribution of an NPI and the ‘free choice’ reading, which does not have the distribution of an NPI but is restricted to certain modal and ‘subtrigging’ contexts:

I am hesitant to include any as an example here, as the distribution of free choice any is limited to certain modal contexts, and it is not entirely clear if it has universal force (see Dayal 1998) or if it originates form an indefinite reading of any that has an additional indiscriminatory component (see Horn 2000 and numerous references cited therein). Horn (2000) notes that it is easy to see how an existential expression with an additional free choice component can come to express what often seems to amount to the equivalent of a universal meaning: if a randomly, freely chosen element satisfies a predicate, all elements that could have been chosen instead presumably would do too. The role of negation in this context is not obvious, however. I leave this as an issue for future study.

4.2 Analysis of NPI Dualization: The role of strength

How are we to make sense of NPI Dualization? The proposal that the homophony between [u-neg] and [i-neg] expressions can be thought of as change where [i-neg] is realized – in the environment or on the expression itself – clearly does not apply here; another explanation is needed.

One way to conceptualize NPI Dualization is to ask what existential [u-neg] and the universal-like homophones have in common, besides their phonology. Their interpretation offers a plausible answer: both interpretations lend themselves to making pragmatically strong statements. For universal-like expressions, like ever _universal, etc. this follows directly from the semantics, just as it does for every, all etc.; there is no stronger claim than one with universal force. Granted, not all the occurrences of a universal-like expression are pragmatically strong, since in DE-contexts they give rise to scalar implicatures (e.g., not all implies ‘some’), revealing the presence of stronger alternatives (e.g., ‘some’ in DE contexts). But nothing in the lexical entry of ever _universal, every, etc. forces them to appear in a DE context (they are not NPIs), and mostly they do not, as what is generally implicated by ‘not every’ is more readily expressed by ‘some’. It seems fair to say then that the vast majority of universal quantifiers are used to make semantically strong claims. This, together with the overall relative markedness of negative contexts (e.g., Horn 1989), makes it plausible to assume that speakers classify ever _universal as ‘strong’ in virtue of its meaning alone. What about the existential ever _[u-neg]? It is semantically weak (e.g., some normally implicates ‘not all’). But, because existential ever is lexically marked as [u-neg], it is restricted to locally DE contexts (which renders the relative markedness of negative environments irrelevant). Consequently, the lexical entry of ever _[u-neg] signals pragmatic strength as well.

Based on this reasoning, we might take NPI Dualization to show that pragmatic strength can become lexically associated with a particular lexical entry. The strength association may then persist while other meaning components, in particular quantificational force (existential vs. universal), and distribution ([u-neg] vs. free), change, creating a pairing of same-sounding but semantically-distinct expressions. The two expressions may co-exist, as the two kinds of ever in Ann Bradstreet's poem in (55) show, or one of the two may fall into disuse.

I have discussed only a few examples of NPI Dualization (ever, at all), and much remains to be explored. It would be interesting to see whether NPI Dualization interacts with Negative Concord. A possible candidate can again be found among adverbial quantifiers. Spanish jamás and French jamais and related expressions in Portuguese, Galician and Italian function as NC-terms now, showing an ambiguity between an existential [u-neg] interpretation and a negative or [i-neg] reading, corresponding to the NPI ever and the negative adverbial quantifier never, respectively. Yet historically jamás etc. derive from Latin iam magis (‘already more’), and would thus seem to have originally been closer in meaning to a universal. A universal use is in fact preserved in the locutions para siempre jamás (Spanish) and à (tout) jamais (French), which both translate as ‘forever and ever’.

This suggests that jamás/jamais exemplifies both Negative Concord and NPI Dualization. This sketch of an analysis requires more historical data to establish this more firmly.

5. Conclusion

Summarizing, I have explored how semantically predisposed expressions can become NPIs, and argued that this means they come to bear a purely distributional feature, [u-neg]. The low-scalar ones among the NPIs can develop into NC-terms. This happens when the [u-neg] version is joined by a homophonous, semantically negative [i-neg] version. The distribution of the [i-neg] version is limited not because it needs to have a feature valued, but because of the cost of taking scope over the event operator. Initially, [i-neg] interpretations are limited to instances where the verb is elliptical, before later extending (in some languages/varieties) to preverbal position. This results in a quasi-complementary distribution of the [u-neg] and [i-neg] versions, which makes the ambiguity learnable. The appearance of either version typically allows the learner to conclude that they are dealing with sentence negation. The balance can start to tip to pure [i-neg] paradigms, resulting in a Double Negation system, when the learner has independent evidence that postverbal expressions can outscope the event operator ∃e. This happens when, as part of the Jespersen Cycle, a former postverbal [u-neg] expression of a two-part negation becomes [i-neg].

In this article, I have advanced a view of NPI-licensing as a purely syntactic affair. I have argued that pragmatic strength is relevant not to the licensing of NPIs per se, but to the historical development of NPIs. One way in which it matters is that their pragmatic strength in DE contexts helps explain why some low scalar expressions come to be frequently used in, and then ultimately restricted to, such contexts. Importantly, however, the ability to express strength in DE-contexts makes an expression susceptible to occurring frequently in DE contexts, but is neither a necessary nor a sufficient condition for an expression becoming an NPI. That it is not sufficient is shown by the existence of numerous expressions with existential meaning that do not have the limited distribution of NPIs (e.g., some, a). That it is not a necessary condition is shown by the existence of attenuating NPIs. The other area in which pragmatic strength matters is in the history of NPI-Dualization. There an existential NPI quantifier or a universal one seem to have such a strong association with pragmatic strength that the association stays constant while quantificational force (existential or universal) and feature specification ([u-neg] or not) change. Obviously, many open issues remain. These include establishing the specifics of the syntactic licensing process (including its application to the weak/strong spectrum and the locality of NPI-licensing), and also studying in more detail the history of non-low scalar NPIs. I leave these and other issues for future study.