CHAPTER 26
NEGATIVE POLARITY ITEMS
26.1 INTRODUCTION: WHAT IS AN NPI?
NEGATIVE Polarity Items (NPIs) are a class of lexical items that need to be licensed by certain conditions. For instance, the English NPIs any and ever typically appear in the semantic and syntactic scope (i.e. the c-command domain) of an NPI licensor, such as negation (Klima, 1964; Ladusaw, 1979, 1980). As shown in (1), any and ever are grammatically licensed when they appear within the scope of negation (1a,b), but they are ungrammatical when there is no negation present (1c,d), or when a negation is present, but doesn’t c-command the NPI (1e,f).
(1) a. John didn’t talk to anybody.
b. John hasn’t ever talked to Bill.
*c. John has ever talked to Bill.
*d. John has talked to anybody.
*e. Anybody didn’t talk to John.
*f. The debate that nobody cared about will ever end.
Because of their apparent sensitivity to the presence of negation, any and ever are labelled ‘negative’ polarity items (NPIs).1 Although negation is a cross-linguistically attested licensor for NPIs, it must be noted that their distribution (within a single language or crosslinguistically) is quite broad and includes a vast range of negative and non-negative licensors, including negative quantifiers, conditionals, modal verbs, generic sentences, imperatives, questions, the scope of certain universal quantifiers and disjunctions (see Giannakidou, 2016, for a review). At the same time, many other lexical items, in addition to any and ever, can be NPIs, and the set of NPIs can vary from language to language. We give some English examples of both NPIs and licensors in (2) (examples taken from Linebarger, 1987), with each licensor and NPI underlined:
(2)a. John didn’tknow any French.
b. Few people have any interest in this.
c. Only John has a hope in hell of passing.
d. Everyone who knows a damn thing about English knows this word.
e. If you steal any food, they’ll arrest you.
f. He was taller than we ever thought he would be.
g. He refused to budge an inch.
Given the broad distribution of NPIs in English and cross-linguistically, a large body of the formal linguistic literature has focused on the licensing question. What is the essential grammatical property that can systematically account for the distribution of NPIs? Since NPIs do not seem to form a homogenous class, either within a single language or crosslinguistically (Zwarts, 1995, 1997; Haspelmath, 1997; Giannakidou, 1998), it is likely that there isn’t one single licensing condition that can explain everything, and an ultimate characterization of the licensing conditions cannot be derived without a close look at the variation in NPIs themselves (within and across languages). Indeed, various notions of licensing have been proposed in the literature (Baker, 1970; Fauconnier, 1975; Ladusaw, 1979, 1980; Linebarger, 1980, 1987; Kadmon & Landman, 1993; Hoeksema, 1994; Krifka, 1995a; Zwarts, 1995, 1996; Giannakidou, 1998, 2006; Lahiri, 1998; Von Fintel, 1999; Chierchia, 2006). The primary focus of this chapter is to present an overview of the experimental studies that have investigated the comprehension and acquisition of NPIs. But before we do that, we need to first understand the complexity of the linguistic properties of NPIs. In the sections later, we will first give a brief overview of the major linguistic notions that have been shown to be important for NPI licensing (section 26.2), and then we will present the recent findings and puzzles revealed by experimental studies in sentence processing and first language acquisition (sections 26.3 and 26.4).
26.2 LINGUISTIC PROPERTIES
In this section we briefly review three lines of research that have been important for our understanding of NPI licensing as a natural phenomenon.
26.2.1 Downward entailment
According to the influential Downward Entailing (DE) hypothesis (Fauconnier, 1975; Ladusaw, 1979, 1980; see also Hoeksema, 1994; Von Fintel, 1999; among others), NPIs are licensed in the scope of the DE operators.
(3) DE function:
A function f is DE iff for every arbitrary element X and Y, it
holds that X ⊆ Y → f(Y) ⊆ f(X)
DE functions reverse entailment relations. Negation is a typical DE operator. Consider (4) as an example. The set of red cars is a subset of all the cars. In an affirmative statement (4b), the entailment relation holds from the subset (the red cars) to the superset (the cars), but not vice versa (4a). The entailment relation reverses, however, in (4c) and (4d), which contain a sentential negation.
The DE hypothesis is extremely influential because it provides a unified semantic characterization of NPI licensing for the first time. It elegantly captures the fact that a set of seemingly unrelated items could all license NPIs because of their logical properties. Take the universal quantifier every as an example. It licenses NPIs in its restrictor, but not its scope. This is predicted because only the restrictor of every is DE, whereas its scope is not. The correlation between the NPI-licensing behaviour of every and its DEness is demonstrated in the following two examples.
(5) The restrictor of every is DE and licenses NPIs
a. Every student wrote a paper for this class. ⊨ Every student who enjoyed the class wrote a paper for this class.
b. Every student who did any homework for this class got an A on the final exam.
(6) The scope of every is not DE and doesn’t license NPIs
a. Every student wrote a paper for this class. ⊭ Every student wrote a paper on language change for this class.
*b. Every student who did their homework for this class ever got an A on the exams.
Despite its intuitive appeal, one empirical problem for the DE hypothesis is that NPIs can be licensed in non-DE contexts as well. The following is a (non-exhaustive) list of the NPI-licensing contexts that are non-monotonic (examples are taken from Giannakidou, 1999; Von Fintel, 1999; and Rothschild, 2006).
(7)a. Antecedents of conditionals
If John had walked on any weeds he would have tracked dirt into the house.
b. ‘most’
Most men with any revolutionary commitments were executed.
c. ‘only’
Only John did any work.
d. Exactly n
Exactly three people with any money showed up.
e. Adversative predicates
Sandy regrets that Robin bought any car.
f. Superlatives (and comparatives)
Emma is the tallest girl to ever win the dance contest.
g. yes-no non-rhetorical questions
Did Lucy see anyone?
One approach to solving this problem is to refine the original notion of DE. Von Fintel (1999) proposed to replace DE with a weakened condition, Strawson DE. A sentence S Strawson-downward-entails another sentence S’ if, when the presuppositions of both S and S’ are satisfied, S entails S’. For example, the sentence Only John ate vegetables doesn’t entail that Only John ate kale in the strict sense, because the presupposition of the second sentence John ate kale isn’t entailed by the first sentence. However, with the notion of Strawson DE, we would restrict the inference test to only those situations in which the presuppositions of both sentences are fulfilled. In a situation in which we know that John ate kale, then if Only John ate vegetables is true, Only John ate kale is also true. The Strawson DE criterion would correctly classify only as a NPI licensor, whereas the traditional DE criterion fails. Limited by space, we won’t go into a full discussion of the theoretical and empirical advantages of the Strawson DE condition discussed in Von Fintel (1999) (also see Gajewski, 2005; Guerzoni & Sharvit, 2007). But we note that it has also been argued that Strawson DE still doesn’t suffice to account for all the empirical problems that the traditional DE account encounters. We refer readers to discussion of the problems in Giannakidou (2006), Rothschild (2006), and Homer (2008).
26.2.2 Non-veridicality
Observing that DE operators capture a subset but not all of the NPI licensors, Giannakidou (1998, 1999) proposed (non-)veridicality sensitivity as the licensing condition for NPIs (see also Zwarts, 1995). Non-veridical contexts are those that do not entail or presuppose the truth of a proposition p. Sentential negation, being the most robust NPI licensor cross-linguistically, is not only non-veridical, but anti-veridical, since the proposition under negation is made necessarily false. The antecedent of a conditional, which also licenses NPIs, is non-veridical, because the truth of the proposition in the antecedent of a conditional is not warranted. The non-veridicality condition achieves broad empirical coverage. Giannakidou (1998, 1999) argued that many of the licensing environments that are not DE are non-veridical, including questions, conditionals, modal contexts, future tense, imperatives, etc.
The non-veridicality condition connects NPI licensing to a previous observation that NPI licensors appear to have graded strength depending on how negative they are (Zwarts, 1995, 1996, 1997). In Zwarts’s system, the strength of negation is a function of how many of De Morgan’s laws are satisfied by an expression. Some expressions, such as anti-additive and anti-morphic negations, are strong negations; mere DE quantifiers, such as few, are forms of the weaker negations. But the non-veridical licensors include also non-DE expressions (as discussed earlier), and they satisfy the fewest of De Morgan’s laws. From this perspective, there is a hierarchy among different classes of NPI licensors, and nonveridical licensors form a superset of the others.
26.2.3 Domain widening and exhaustifying the alternatives
Kadmon & Landman (1993) made the argument that the semantic/pragmatic contribution of any is to make a stronger statement than a regular indefinite noun phrase. They proposed that the semantics of any is the regular semantics of an indefinite noun phrase plus an additional domain widening function. Domain widening widens the interpretation of the regular noun phrase along a certain contextual dimension. Any is licensed only if the widened domain it introduces creates a stronger statement. This analysis essentially introduces a scalar component to the semantics of any. By uttering I don’t have any potatoes, the speaker could mean that he doesn’t have even a rotten potato (not to mention a normal potato), or even one single potato (not to mention a larger quantity). The exact scalar dimension is determined by context. The (implicit) scalar component in English any is explicitly introduced in some other languages. For instance, in Hindi, an overt morpheme even is part of the NPI form (Lahiri, 1998). Strengthening with domain widening can be naturally achieved with DE environments, since the assertion in which a DE operator quantifies over a larger/widened domain would entail the assertion with a smaller domain. This account therefore explains rather than stipulates why DE-operators are prominent NPI licensors (see also Krifka, 1995a). Building upon the insights of Kadmon & Landman (1993) and Krifka (1995a), Chierchia (2006, 2013) further implemented a compositional analysis. Without going into all the technical details, in a nutshell, the basic idea is this. The NPI any lexically activates alternatives. A covert ‘only’ is introduced such that during the semantic composition of a proposition, alternatives that are not entailed by the assertion need to be eliminated (i.e. exhaustifying the alternatives). It turns out that the exhaustification process leads to contradictions in Upward Entailing (UE) contexts, but not in DE contexts, and therefore NPIs (specifically the scalar type)2 are licensed in DE but not UE contexts.
26.2.4 Summary
The quick and brief overview in this section, although not doing full justice to the large literature, already reveals that NPI licensing involves satisfying constraints at multiple levels, including syntactic, semantic, and pragmatic constraints. Each of the proposals outlined earlier emphasized a particular dimension of NPI licensing, but it is plausible that we need more than one single condition to account for the full range of data. The multi-dimensionality of NPI licensing becomes even more apparent when we consider how NPIs are processed and comprehended in real time. The experimental studies we introduce in the next section will suggest that to successfully integrate an NPI into its sentential context, the comprehension system needs to recruit multiple different mechanisms, including both grammatical and extra-grammatical ones.
26.3 PROCESSING NEGATIVE POLARITY ITEMS
26.3.1 Basic sensitivity to the licensing condition
Given the complexity of the licensing condition for NPIs, a basic processing question is how quickly people can utilize the relevant grammatical constraints to integrate an NPI word into a sentence. To address this question, a commonly used experimental paradigm is to compare the processing profile of licensed and unlicensed NPIs, and examine how quickly people can detect the linguistic anomaly from the unlicensed NPI.
To understand the fine-grained time course of the NPI-licensing process in comprehension, a number of studies have adapted the Event-Related Potentials (ERP) technique, due to its excellent temporal resolution. The two most common ERP components that are relevant for sentence-level language comprehension are the so-called N400 and P600 effects. The N400 is a negative-going waveform that peaks at approximately 400 milliseconds (ms) from the onset of the stimulus presentation, with a primarily centro-posterior scalp distribution. Generally speaking, the amplitude of the N400 evoked by an incoming word indexes the degree to which that word’s semantic features match the semantic features that have been pre-activated by its context at the time it is encountered (Lau, Wagers, Stroud, & Phillips, 2008; Kutas & Federmeier, 2011; Kuperberg, 2013). For example, a highly predictable word sugar in the sentence I like my coffee with cream and sugar elicited a much smaller N400 response than a semantically incongruent word sock in the same sentential context I like my coffee with cream and sock (Kutas & Hillyard, 1984). Even though the term ‘pre-activation’ could be associated with active prediction of specific lexical items, we use the term here in a more neutral sense: it refers to the activation of relevant semantic features, regardless of whether active prediction or expectation of the upcoming word is at work, ahead of encountering the full linguistic input. In situations where the context is highly constrained, it is possible the comprehension system may actually anticipate a specific lexical item, especially if it is a content word. But predicting a specific lexical item is not a necessary condition to evoke reduced N400 amplitude. In the context of NPI licensing, before an NPI is encountered, it is not very plausible to assume that the comprehension system could have been anticipating the specific lexical form of an NPI (such as any or ever). But as long as, when an NPI is encountered, the semantic features/properties processed in the context match the relevant lexical semantic features of the NPI, we should expect to see a reduced N400 on the NPI.
In a series of studies on German (Saddy et al., 2004; Drenhaus et al., 2005, 2006; Drenhaus et al., 2007), a reduced N400 with a central maximum was found on the German NPI jemals (‘ever’) when it was licensed by negation, compared to the ungrammatical counterpart when the word jemals was not licensed. An example from Saddy et al. (2004) is given in (8):
N400 effects were also found for Dutch (Yurchenko et al., 2013), Turkish (Yanılmaz & Drury, 2018) and English NPIs (Shao & Neville, 1998), and in a magnetoencephalography (MEG) study by Tesan et al. (2012).
These findings provide strong evidence that comprehenders can very rapidly integrate the relevant linguistic information to license NPIs in a sentence—by 400ms after an NPI is presented, language users can already detect whether the NPI appears in the proper context. It should also be noted that the rapid incorporation of the NPI licensing constraints in online comprehension is equally available in languages in which an NPI can linearly precede its licensor. Pablos et al. (2012) found that when the Dutch NPI ook maar iets precedes a negation in a sentence, comprehenders actively anticipate a downstream negation after encountering the NPI, resulting in larger central anterior negativity on the negation if the distance between the NPI and the negation is long. Furthermore, such an effect is most salient when the negation is actually properly licensing the NPI by being in a c-commanding position, but the effect diminishes when the negation is in a non-c-commanding embedded position (see also Yanılmaz & Drury, 2018).
The particular ERP component associated with the NPI licensing process, namely the N400 response, also sheds important light on the specific mechanisms involved in NPI processing. Under the functional interpretation that N400 indexes the degree of semantic feature match between a word and its preceding context, the reduced N400 on a licensed NPI (and conversely the larger N400 amplitude on an unlicensed NPI) suggests that some semantic features/properties of the sentential context, pre-activated prior to the point at which an NPI is encountered, match certain lexical semantic features of the NPI. Since these studies all manipulated whether the sentential context contains negation or not, these results at least suggest that some property of negation is implicated in the lexical processing of NPIs. However, we do not have enough information to precisely identify the exact NPI licensing property that triggered the N400 effect. For example, is it the [+Neg] feature that is responsible for the early stage of the NPI licensing process, or is it the non-veridicality of the context, or is it the downward entailment property? All of these possibilities are consistent with the current findings. To tease them apart, one would need to examine a wide range of NPI licensors with non-overlapping properties, and see whether they all facilitate the lexical processing of an NPI during the N400 time window.
In this regard, it is interesting to note that not all the studies on NPI licensing have found increased N400 amplitudes on the unlicensed NPIs. Drenhaus et al. (2007) compared licensed and unlicensed NPI jemals (‘ever’) in German. Two kinds of licensors were included—negation and wh-questions. NPIs licensed under negation elicited smaller N400 than the unlicensed NPIs, but NPIs licensed under a wh-question did not lead to any N400 reduction. The authors attributed the difference between negation and wh-questions to their strength of licensing (see Zwarts, 1995, 1996). But it is worth noting that wh-questions are not negative, and it is possible that the N400 reduction on licensed NPIs is primarily triggered by negation. In another study, Steinhauer et al. (2010) examined the licensing of three different NPIs in English ever, any, and at all. Although there was an N400 difference between the licensed and the unlicensed NPI at all, they did not find an N400 difference between licensed and unlicensed ever and any. One important difference between this study and the others mentioned earlier is that Steinhauer et al. (2010) tested a larger set of licensors in their stimuli, including various negative licensors such as not, without, rarely/hardly, and also licensors that are not negative per se, but non-veridical, such as every, before, whether, and yes-no questions. It is possible that the N400 reduction on the NPI ever or any can only be triggered by pre-activated negative features and therefore, that the N400 effect in Steinhauer et al. (2010) might have been washed out by the use of both negative and non-negative licensors. This post-hoc explanation is obviously very tentative, but it is worth further exploration.
Results from both Drenhaus et al. (2007) and Steinhauer et al. (2010) suggest that the lexical processing of NPIs seems to be primarily sensitive to negation, but not non-veridicality in general. This issue is addressed in more detail in a later study in Xiang et al. (2016), which we will discuss later in this section. An additional finding from Steinhauer et al. (2010) is that not all NPIs are licensed in the same way either. In particular, although no N400 reduction was found for the licensed ever and any in this study (compared to their respective unlicensed controls), the morphologically complex NPI at all did induce an N400 effect. Steinhauer et al. (2010) suggested that the morphological complexity of the NPIs themselves may also matter.
The majority of the studies reviewed earlier also reported a posteriorly distributed P600-like late positivity effect, which was larger for unlicensed NPIs than for licensed ones.3 The P600 component is a positive-going waveform that peaks at about 600ms after the onset of a stimulus. This effect was originally associated with syntactic processing, since it is reliably elicited by syntactic errors (Hagoort et al., 1993; Osterhout & Holcomb, 1992) or grammatical but syntactically complex constructions (Osterhout et al., 1994; Kaan et al., 2000; Phillips et al., 2005; Gouvêa et al., 2010); on the other hand, there is a growing body of work on the ‘semantic P600’ effect (Kim & Osterhout, 2005; Kuperberg, 2007; Bornkessel-Schlesewsky & Schlesewsky, 2008; Van de Meerendonk et al., 2009; Brouwer et al., 2012; Paczynski & Kuperberg, 2012; Chow & Phillips, 2013), showing that words that are semantically implausible within their context can also elicit a large P600. Although the precise functional interpretation of the P600 is yet to be determined, a broad generalization that has emerged is that it reflects costs associated with a processing stage in which information from different sources is integrated into one coherent representation (Friederici & Weissenborn, 2007; Kuperberg, 2007; Bornkessel & Schlesewsky, 2008; Van Petten & Luka, 2012). Increased P600 amplitudes signal the detection of an integration error or integration difficulty, including costs associated with the process of reanalysis. In the particular context of NPI licensing, multiple streams of information—syntactic, semantic, and pragmatic—are recruited to construct a grammatical representation that can license NPIs. In an ungrammatical sentence that does not license NPIs, the comprehension system fails or has great difficulty integrating an NPI into the current grammatical representation, and therefore produces a large P600.
If the N400 and the P600 components index the workings of different processing mechanisms at different stages of the online comprehension process, combining the two together provides us with an opportunity to examine the finer-grained and more precise mechanisms that support the comprehension of NPIs. In particular, since it is likely that NPI licensing is not a uniform phenomenon, different licensors (or licensor-NPI pairs) may show different responses on one or both of the ERP components. Xiang et al. (2016) examined the English NPI ever under different licensors, and argued that there are different classes of licensors, grouped by the neurophysiological patterns they elicit. Their results suggest multiple different licensing mechanisms. Four kinds of NPI licensors were tested in this study, no, few, only, and emotive factives, such as surprised, glad, etc. An example of the stimuli is given in (9):
(9) It is hard to train a dog.
a. No dogs Andrew owns have ever responded to commands.
b. Few dogs Andrew owns have ever responded to commands.
c. Only dogs Andrew owns have ever responded to commands.
d. Andrew is surprised that the dogs he owns have ever responded to commands.
e. *The dogs Andrew owns have ever responded to commands.
ERP responses were measured from the onset of the critical word ever. The critical finding is that, when compared to the unlicensed ever in (9e), all the other conditions (a)–(d) showed a reduced N400 on ever, but only (a)–(c) showed a smaller P600. That is to say, in the N400 time window, all four licensors no, few, only, and the emotives patterned together and triggered a reduced N400 on ever; but in the later P600 time window, the emotive factive licensor deviated from other three kinds of licensors. In particular, in the P600 time window, the emotive factive condition elicited similar P600 amplitude to the ungrammatical condition (e), suggesting processing cost associated with NPIs licensed by emotive factive licensors during this time window.
These findings are difficult to explain under any account that assumes a single unified licensing mechanism for different licensors, be it DE, Strawson DE, or non-veridicality-based. Any single-route licensing account would predict no difference between the four licensors in (9). Xiang et al. (2016) interpreted their results as providing evidence for a qualitative distinction between a grammatical licensing and a pragmatic licensing mechanism, as represented by licensors no, few, and only on the one hand, and the emotive factives on the other. The relevant dimension of difference is how these licensors introduce negation into a proposition. No and few are negative determiners, and they introduce negation into the asserted meaning of a proposition. Only also introduces a negative meaning into the assertion by an exclusive component in its semantics. The example in (9c) means something like ‘only Andrew’s dog but not other dogs have ever responded to commands’. On the other hand, the emotive factive predicate surprised in (9d) introduces a negative implicature—that Andrew is surprised by something implicates that he did not expect it. Under this perspective, all four licensors activate some kind of negative feature. It is this pre-activated negative feature that triggered the reduced N400 amplitude on ever, since the lexical processing of ever is sensitive to negation. However, it is crucial to distinguish the lexical stage of processing from the full licensing of ever, since the latter involves a more combinatorial process of integrating the NPI at the sentential level. Xiang et al. (2016) argued that even though the comprehension system can extract the negative information from the four kinds of licensors equally quickly (as shown by the similar N400 effects), it nevertheless draws a line between a primary grammatical licensing mechanism and a secondary pragmatic licensing mechanism. The increased P600 amplitude triggered by the NPI word under an emotive factive licensor reflects the failure to license the NPI with a grammatical licensing mechanism, so that the comprehension system ultimately resorts to a pragmatic licensing mechanism, relying on the negative implicatures.
26.3.2 Illusory licensing effect
To understand the precise mechanisms that license NPIs, it is also useful to look at the errors people make in comprehension, because the errors that comprehenders make when they compute the proper licensing conditions potentially provide important information about the underlying processes.
It has been observed that sometimes comprehenders fall for ‘illusory licensing’ of NPIs. This often happens if a licensor is inserted into the preceding context—but crucially, it is not in the right structural (c-commanding) position (Drenhaus et al., 2005; Vasishth, Drenhaus, et al., 2005; Vasishth, Brussow, et al., 2008). An example is given in (10) (examples taken from Drenhaus et al. 2005).
(10)a. No man who had a beard was ever happy.
*b. A man who had a beard was ever happy.
*c. A man who had no beard was ever happy.
(10a) is grammatical, and (10b) is not, since there is no licensor in (10b). But the crucial sentence is (10c). In the (c) example, there is a negation in the sentence, but the NPI ever is not in the scope of the negation (i.e. the negation does not c-command the NPI). The NPI ever therefore remains unlicensed. In comprehension experiments, it was consistently observed that (10c) was more acceptable than (10b), and the unlicensed NPI word in (10c) elicited shorter reading times or smaller ERP amplitudes than the NPI word in (10b) as well (Drenhaus et al., 2005; Vasishth, Drenhaus, et al., 2005; Vasishth, Brüssow, et al., 2008; Xiang et al., 2009; Parker & Phillips, 2016). All of these findings suggest that the negation in (10c) triggered some sort of illusory licensing or interference effect.
On the surface, illusory NPI licensing looks very similar to the well-documented agreement attraction/illusion effect (Bock & Eberhard, 1993; Pearlmutter et al., 1999; Wagers et al., 2010) and interference in reference resolution (Badecker & Straub 2002; Gordon et al. 2006). For instance, a sentence like *the key to the cabinets are…, contains an agreement dependency between the singular subject the key and the plural predicate are, which is ungrammatical due to a number mismatch. However, the intervening noun cabinets creates an illusion/interference effect on various processing measures: such sentences are not uncommon in spontaneous production, they can be elicited in controlled laboratory experiments, they are judged to be relatively acceptable, and on-line reading times on the otherwise problematic verb are generally reduced compared to number mismatched verbs without interference (Bock & Miller, 1991; Bock & Eberhard, 1993; Pearlmutter et al., 1999; Wagers et al., 2009; Dillon et al., 2013).
The standard account of the agreement attraction effect has modelled it as an instance of memory interference. During incremental parsing of a long distance dependency, the tail of the dependency initiates the memory retrieval of the head. This retrieval is prone to interference due to overlapping features between the set of retrieval cues and other material stored in working memory (McElree et al., 2003; Lewis & Vasishth, 2005; Lewis et al., 2006; Wagers et al., 2009). Memory interference could be driven by partially matched morpho-syntactic features, as has been repeatedly shown by agreement attraction errors like the example earlier, and examples like ‘The new executive who oversaw the middle managers were dishonest’ (example from Dillon et al., 2013). In such cases, at the plural verb, for example, were, the memory retrieval mechanism sets out to retrieve a plural subject, guided by the retrieval cues [+plural] and [+subject]. A non-subject plural noun has some probability of being misretrieved due to its partial feature match with the retrieval cues.
At first glance, such a memory (mis)retrieval-based account could also be applied to NPI illusory licensing. For instance, Vasishth et al. (2008) argued that the parser uses semantic cues such as [+negative] and syntactic cues such as [+c-command] to retrieve a proper licensor for ever from previously processed material in memory. For (10b), no such match is found, and the sentence is determined to be unacceptable. For (10c), however, the quantifier no in the embedded subject position partially matches the search criteria: although it doesn’t match the syntactic cue [+c-command], it does satisfy the semantic cue [+negative]. During retrieval of a licensor, this partial feature match may boost the activation level of the memory representation of the embedded quantifier no, causing it to be more likely to be retrieved once its activation level goes beyond a certain threshold.
Although such an account is plausible, it seems to miss a crucial distinction between NPI licensing and other types of long distance dependencies such as subject-verb agreement. Specifically, while the latter dependency types involve mainly syntactic relations between lexical items (e.g. a subject and an agreeing verb, or a head noun and a verb in a relative clause), NPI licensing involves not only syntactic licensing (e.g. the c-command requirement on a proper licensor), but also logical-semantic and pragmatic licensing conditions. Xiang et al. (2009) argued that the NPI interference effect stems from over-application of an inference-based licensing mechanism that is already in place in the grammar (Baker, 1970; Linebarger, 1980, 1987, Giannakidou, 2006; Xiang et al., 2016). Some evidence for this comes from a study in Xiang et al. (2013), which showed that NPI interference is sensitive to speakers’ individual pragmatic reasoning abilities, as measured by the Communication Subscale from the Autism-spectrum Quotient (AQ; Baron-Cohen etal. 2001). The AQ test has been widely used to assess individuals’ Autistic traits, which were known to affect semantic and pragmatic processing. In this study, NPI interference was compared to agreement attraction on the same group of participants. It was found that participants with worse pragmatic skills were better at rejecting the illusory NPI licensing items, and people with better pragmatic skills accepted them more often. But there was no interaction between AQ scores and the acceptability judgments on the set of number agreement items.
All the studies reviewed earlier assumed that the illusory NPI licensing effect is a static phenomenon, but a set of experiments from Parker & Phillips (2016) clearly demonstrated that illusory licensing can be switched on or off systematically. In particular, the amount of processing time the comprehender has between encountering the licensor and the NPI appears to be an important factor in modulating the illusory licensing effect. Parker & Phillips showed that the standard illusory licensing effect in (11a) disappears if more material is inserted between the licensor and the NPI. This could be done by moving the NPI to a post-verbal position (11b), or embedding the NPI in a subordinate clause (11c), or inserting a parenthetical clause (11d). Although comprehenders accept (11a) to some degree, they fully recognize that (11b)–(11d) are as ungrammatical as sentences without any licensors.
(11)a. The authors that no critics recommended have ever received acknowledgement.
b. The authors that no critics recommended have received any acknowledgement.
c. The journalists that the editors recommended for the assignment thought that the readers would ever understand the complicated situation.
d. The authors that no critics recommended for the assignment have, as the editor mentioned, ever received a pay raise.
Very importantly, similar manipulations do not switch off the illusory agreement effect, as in The key to the cabinets, according to the janitor, probably were destroyed by the fire.
These interesting findings provide further evidence that NPI licensing, unlike morpho-syntactic agreement, involves semantic and pragmatic processing. More importantly, data like this also raises important questions about how semantic/pragmatic representations are dynamically consolidated over time in on-line comprehension. Parker & Phillips (in press) suggested that the memory encoding of the semantic/pragmatic representations may evolve over time such that the specific format of the encoding is different from the time the licensor is encountered to the time an NPI is encountered. It is this change in memory encoding that would affect the illusory licensing effect, although the details of the memory consolidation of semantic/pragmatic representations are yet to be worked out.
26.3.3 Summary
The experimental studies on NPI processing reviewed in this section form the initial basis for understanding how NPIs are comprehended in real time. The emerging picture is that incremental comprehension is very sensitive to the relevant information that contributes to NPI licensing. Current findings already reveal that small changes, such as switching from one licensor to another, or changing the position of the NPI word in the sentence, can produce different processing results. To model results like these, we need to look at both NPI-specific linguistic properties, including the syntactic, semantic, and pragmatic constraints on NPI licensing, and more domain-general comprehension mechanisms, such as how linguistic representations are encoded and retrieved over time.
26.4 ACQUIRING NEGATIVE POLARITY ITEMS
The complexity of NPI licensing and processing naturally raises the question of learn-ability. Children need to learn, from their input, the distributional constraints on NPIs. More importantly, the acquisition task is not simply for children to memorize a one-to-one relationship between a certain licensor-NPI pair. Instead, children need to learn the relevant syntactic, semantic, and pragmatic features that allow an NPI to be licensed. As we will see later, some licensors (and possibly some NPIs) are extremely rare in the early input of a child. Given the diversity of both the set of NPIs and the set of licensors, it is unlikely that children will be exposed to a sufficient number of exemplars of every possible kind of NPI licensing in early life. The challenge for a child is to make a leap from the limited observed input to a generalized set of semantic, pragmatic, and syntactic properties that are essential to licensing NPIs.
26.4.1 Children’s knowledge about NPI licensing
Studies that analysed children’s spontaneous speech from corpora have generally observed that young children make very few errors with NPI licensing. Tieu (2013, 2015) analysed the spontaneous production from forty children who speak either American English or British English, between 1 and 5 years old. Focusing on the production of one singular NPI any, these studies found that children produced very few unlicensed any—the error rate is only about 3 per cent. Furthermore, even though, for the majority of the cases, children’s use of any appeared with sentential negation, they also correctly produced instances of some other licensors, such as negative quantifiers, conditionals, comparatives, etc. Lin et al. (2014) and Lin (2015) analysed about forty monolingual Mandarin-speaking children, between 1 and 5 years old. The Chinese NPI shenme has overlapping but not identical distribution with the English NPI any. The most salient difference between the two is that Chinese shenme is lexically ambiguous between an NPI and a wh-question morpheme. For instance, the wh-word what in a wh-question like ‘what did you eat?’ would be translated into Mandarin as shenme. The dual function of shenme in Mandarin makes the learning task more complicated. Lin (2015) found that children’s early production of shenme predominantly represents its function as a question morpheme. But after 4 years of age, children start producing shenme in various NPI-licensing contexts. Like English-speaking children, Mandarin-speaking children made very few errors—they do not use the NPI shenme in contexts that do not license NPIs.
In addition to observations from children’s spontaneous production, experiments that examined children’s elicited comprehension and production also demonstrated that children have, at least some, non-trivial understanding of the licensing conditions of NPIs. A number of studies used the Truth-Value Judgment Task (TVJT) to probe children’s comprehension of NPIs (Thornton, 1995; Xiang et al., 2006; Lin, 2015, 2017; Tieu & Lidz, 2016). It was consistently found that young children have some understanding that the NPI any should be interpreted under the scope of sentential negation. Consider the following comparison between an NPI any and a regular indefinite:
(12)a. Donald and Daisy both can’t find any stars.
b. Donald and Daisy both can’t find a star.
The indefinite a star can take either narrow or wide scope relative to the sentential negation. But the NPI any can only take narrow scope under the negation. Results from TVJTs (Xiang et al., 2006; Tieu & Lidz, 2016) have found that, in contexts under which the scope interpretations of any and a regular indefinite need to be distinguished, children were very sensitive to their differences. For instance, Tieu & Lidz (2016) set up a scenario in which the two toy characters Donald and Daisy needed to find all the hidden stars. There were different types of hidden stars, including wooden stars, metal stars, fuzzy stars, etc. They succeeded in finding all the other stars except for the fuzzy one. Given this scenario, children were asked to judge whether the sentences in (12) was true or false. For the most part, they correctly rejected (12a), but accepted (12b) (see similar findings in Xiang et al. 2006). Findings like these suggest that children have some sophisticated understanding of what an NPI means. At its semantic core, any in (12a) is very similar to a regular indefinite, except that it introduces domain widening (Kadmon & Landman, 1993), which prevents any from taking wide scope over negation.
Lin (2017; see also Lin, 2015) carried out an elicited imitation task to investigate children’s knowledge of a range of different licensors in Mandarin Chinese. In this study, children were given a set of pre-recorded sentences, and they were instructed to repeat, as precisely as they could, every sentence they heard from the recording. A crucial assumption of the authors is that a repetition task is not simply a task of passive verbatim recall. Instead, while repeating the stimuli, children were also reconstructing the stimuli based on their internal grammatical rules and representations. If children failed to repeat certain constructions correctly, it may indicate that the particular stimulus presented to them was not totally compatible with their current grammar. Constructions involving four different kinds of NPI licensors were tested, including sentential negation, conditionals, epistemic uncertainty modals, and polar questions. Wh-questions were also included since the Chinese NPI shenme is also a wh-interrogative morpheme. In addition, there were also ungrammatical controls in which shenme is not licensed at all. The results suggested that the first use children acquire is the wh-use of shenme. Children as young as 35 months old were able to repeat back the wh-construction accurately. The NPI use of shenme under sentential negation and conditional and polar questions was acquired slightly later, reaching 74 per cent accuracy by 57 months old. Interestingly, epistemic uncertainty modals were the most difficult for children—the imitation accuracy for this construction is not reliably different from the ungrammatical controls even for the oldest children tested in this study (57 months). To our knowledge, this study is one of the very few that examined and revealed different developmental paths for acquiring different NPI licensing contexts.
26.4.2 What is in the input?
Our empirical understanding of children’s knowledge state regarding NPI licensing is still quite limited. There are only a handful of studies that have looked at the acquisition of NPIs. Most of these studies focused on a single NPI, or a single licensing environment like sentential negation. But even with the limited data, it is clear that children have robust knowledge of at least some aspects of NPI licensing. What is a plausible learning mechanism that could guide children to this state? A close look at the input data presented to children reveals some interesting but puzzling patterns. It is not obvious that there is sufficient evidence in the direct input that would help children make the full set of generalizations about the syntactic, semantic, and pragmatic properties of NPI licensing.
Tieu (2013, 2015) searched through parental speech samples for two children (based on the CHILDES corpora) and found that the NPI any appeared in about 1.02 per cent of all the utterances in one sample, and 0.54 per cent in the other sample. Tieu argued that the small number of tokens may not in itself cause learning problems, since the distribution of any in the parental speech did provide some positive evidence for children to make hypotheses about the syntactic or even semantic properties of any and its various licensors. For instance, any predominately appeared in the scope of negation in the parental speech, and some instances of other DE licensors were also observed. Lin (2017, 2015) found that, in a sample of 999 utterances of child-directed speech that contain the Mandarin NPI shenme, 976 (97.7%) of them are wh-questions, and the remaining handful of examples are almost evenly distributed among different licensing environments for the NPI shenme (see appendix A in Lin, 2017). Such imbalanced distribution in the input does not completely predict children’s performance in the elicited imitation task discussed earlier. On the one hand, wh-questions are indeed the first thing children understand about shenme, which is predicted given how frequent wh-questions are in the input; but at the same time children nevertheless showed some knowledge about the NPI any in the elicited imitation task, despite the extremely sparse distribution of NPI shenme in the input. Moreover, frequency in the input data alone could not explain why epistemic uncertainty modals are particularly difficult for children to acquire as NPI licensors, since epistemic modals do not seem to be less frequent than other licensors.
The distributional patterns of different licensing environments, although they may be imperfect, at least provide some positive evidence for children to start bootstrapping from the restricted distribution of NPIs. But as Tieu (2013) critically pointed out, the observable distributional patterns of NPIs do not automatically ensure that the evidence for acquiring the semantic representations of NPIs is immediately available from the input. For instance, the scalar use of English any introduces alternatives and widens the domain of the indefinite (as discussed in Kadmon & Landman, 1993; Chierchia, 2006, 2013), achieving a similar effect to focus. Tieu noted that the domain widening effect of any, however, is not immediately transparent by just observing that any co-occurs with negation or other DE operators. Additional contextual support is necessary to signal the domain widening effect. Tieu (2013) reported 577 instances of any through a search of twenty transcripts of conversations between parents and children, and found only two cases that presented clear contextual cues of domain widening or focus. Given the limited input, it remains a puzzle what features of the input provided the most informative cues to facilitate children’s development of the semantic and pragmatic representations.
26.4.3 Summary
The major research question on the acquisition of NPIs is about understanding how children acquire the full complexity of NPI licensing. Empirically speaking, we have very limited data to support any robust conclusions. But it already appears that children have some non-trivial understanding of NPIs at an early age. At the same time, it also appears that not all NPIs and licensors are acquired equally well or equally early. Current discussion of the input has mostly focused on finding out the distributional frequency patterns of NPIs. Given that distributional patterns alone do not seem to provide an adequate account of the developmental path, it is possible that there are other types of subtle cues in the input that correlate with learning, which the current research has yet to fully identify.
26.5 CONCLUSIONS AND FUTURE DIRECTIONS
The linguistic complexity of NPI licensing makes it a rich empirical domain for investigating the cognitive architecture of language processing and acquisition. An adequate processing model needs to explain how information from different sources, syntactic, semantic, and pragmatic, are incrementally combined, stored, accessed, and continuously consolidated over time to form an appropriate licensing context for NPIs. At the same time, what adults can do as language users has to be learned with a specific learning mechanism that can transform information from the external input to an internal grammar. Previous studies, as we discussed in this overview, have revealed exciting new empirical findings, but at the same time have also raised more questions. For future research, we suggest two possible directions that are worth exploration.
First, more varieties of NPIs and different licensing contexts should be experimentally investigated. Previous studies have largely focused on negation as the licensing context, and examined only the two NPIs ever and any. The empirical horizon needs to be widened to include other kinds of licensors and NPIs. Cross-linguistic investigation would be especially valuable since the distribution of NPIs and likely the precise licensing mechanisms vary from language to language. Second, the investigation of NPI licensing should engage with other research domains. For instance, downward entailment inferences are argued to be important for NPI licensing. But very little has been explored as to how people process logical inferences like downward entailment, and to what extent people’s ability to compute different downward entailment relations carry over to NPI processing. Similarly, for a subset of the NPIs, the scalar NPIs such as any, the proposed pragmatic strengthening function bears similarities to contrastive focus marking. Independent findings from research on focus processing/acquisition could shed light on questions such as what factors modulate the set of alternatives activated by any, or what cues in the discourse are salient to signal the presence of alternatives. It is possible that these related processes are supported by shared cognitive mechanisms.
1 Some NPIs, such as any, also obtain so-called free choice readings in modal environments and with imperatives, such as You may talk to any student, and Pick any card! We won’t discuss the free choice use in the current chapter.
2 There are also many non-scalar NPIs (see Lin, 1996; Giannakidou, 1998; Giannakidou & Yoon, 2016). Chatzikonstantinou (2016) provided experimental evidence to show that scalar and non-scalar NPIs could have different prosodic properties.
3 To our knowledge, the only two studies that did not find a P600 are Saddy etal. (2004) and Yurchenko etal. (2013). The original data from Saddy etal. (2004) was reanalysed in Drenhaus etal. (2006) using a symbolic resonance analysis and a hidden P600 was discovered. For Yurchenko et al. (2013), the authors acknowledged that the lack of a P600 may be due to insufficient power in the data, as well as, potentially, to task-specific effects.