1.1 The Distribution of NPI any in English

In English, any as an NPI requires a negative environment. Typically, this means occurring in the scope of sentential negation, as in (1). The occurrence of any outside the scope of negation is ungrammatical, as in (2).

Beyond this basic case, there are various subtleties regarding the precise characterisation of NPI licensing, and how semantic properties interact with the lexical requirements of NPIs, which are more or less stringent in the type of negative context required (see Giannakidou, 2011). To exemplify, any is licensed in questions, see (3). While questions do not involve overt negation, they can be analysed as related to nonveridicality. That is, involving the truth value of the content of the clause, or assertion of whether the event or state of affairs under consideration actually happened/will happen (see Giannakidou, 1998; Gil & Marsden, 2013).

Any also occurs without overt negation in contexts where a nonveridical or negative inference can be computed (see Xiang et al, 2015). It is therefore licensed in clauses embedded under a semantically negative verb or within the scope of semantically negative adverbs, as in (4) and (5) respectively, from Gil et al. (2019, p. 221).

In L2 English, acquisition involves two related issues. Firstly, the lexical representation of any must specify it as a negative polarity item, so that its distribution is restricted. In syntactic terms, one could propose that this means any bears an NPI feature (Gil et al, 2019), or an uninterpretable nonveridical feature (Gil & Marsden, 2013), which is checked in specific syntactic configurations. The second issue involves establishing what these configurations are, i.e. that an overt negative operator such as not, an interrogative operator, or an implicit negator introduced by the semantics of certain verbs and adverbs can create the requisite grammatical environments.

1.2 The Acquisition of any in L2 English

The use of any is often addressed in instruction in English as a foreign language (EFL, e.g., Berry, 2015). However, the information typically provided in instruction underdetermines the range of properties outlined above. This leads GMW to ask how L2 knowledge of a linguistic phenomenon develops when certain properties are taught, others are not taught but may be observable in incidental input, and still others are neither taught nor observable in input.

Based on studies of online EFL teaching materials (Marsden et al, 2018) and international EFL textbooks (Gil et al, 2019), it is shown that the typical pedagogical approach to any involves contrasting its use with some and characterising the choice as determined by clause type. Any occurs in questions and negative declaratives, while some occurs in affirmative declaratives.¹ Additional information may be provided as ‘exceptions’ to this basic rule, i.e., that certain adverbs collocate with any. Learners who have received such instruction would be able to correctly accept any with sentential negation and in questions (assuming such knowledge is accessible during performance). However, an overgeneralisation of this rule of thumb would complicate knowledge of more subtle properties. The absence of overt negation, as in (4a) and (5a) would lead learners to reject these sentences if they are relying on the ‘textbook-rule’ of using any with overt negation. Similarly, the presence of negation in a sentence like (2) could conceivably mislead learners into accepting this pattern due to the occurrence of a negator, even though the scope properties in this configuration rule it out.

Using paced acceptability judgement tasks, GMW investigated the acquisition of NPI properties of any by L1 Arabic and L1 Chinese-speaking learners. The aims of these studies were to explore whether learners’ judgements indicate an ability to reject the less observable properties of any’s distribution dependent on scope and licensing from implicit negation, and whether learners might potentially overgeneralise textbook rules. In Marsden et al. (2018), the judgement task elicited acceptability ratings for 8 sentence types, reflecting the NPI properties and pedagogical rules discussed above, with both grammatical and ungrammatical stimuli. In their study, 4 tokens of each type resulted in 32 experimental items, plus 32 fillers. Table 1 summarises result for the eight sentence types.

All groups had lower accuracy on ungrammatical sentence types, although the results of t-tests on the grammatical/ungrammatical pairs were not significant for the L1 English group, indicating that they were equally able to accept grammatical clauses and reject ungrammatical ones. They also generally had higher accuracy across the board compared to learners. Overall, both learner groups performed similarly. Accuracy was highest on those contexts which tested the typically taught pedagogical rules: the difference between questions and declaratives and the fact that any occurs with the sentential negator not. Judgements are less consistent on those contexts which are not covered in pedagogical materials. However, this does not demonstrate that the pedagogical rule is overgeneralised. Such overgeneralisation would be evidenced by rejection of sentences that do not involve overt negation. However, learners can distinguish grammaticality based on the semantics of the licensing verb or adverb, even if judgements are markedly less robust than the taught cases. This leads GMW to suggest that instruction may have a facilitative effect for EFL learners.

For L1 Arabic and L1 Chinese learners of English, a facilitative effect would also be expected from L1 transfer. The corresponding Najdi Arabic ʔayy and Chinese renhe/wh-existentials have a similar NPI distribution to any. The results provide limited grounds for assuming general positive transfer. There may be specific L1 effects for the L1 Chinese-speaking learners, who demonstrate a greater difference in acceptance of negative verbs versus negative adverbs as licensors (average accuracy 2.23 versus 3.36, respectively, on grammatical conditions). By contrast, the L1 Arabic group does not discriminate (2.88 grammatical negative verb versus 2.92 grammatical negative adverb). Gil et al. (2019, pp. 230–31) suggest that the form of Chinese negative adverbs may facilitate this distinction as the negator bu is often incorporated in these adverbs, e.g., jihu bu – almost not ‘hardly’. The implicit negative meaning of such English adverbs is rendered more transparent for L1 Chinese learners given the association with an overt negator in the corresponding L1 lexical items.

A final relevant finding is that at least some learners could discriminate consistently between the different conditions and so seemed to have completely acquired the relevant linguistic properties of NPI any and its licensing conditions. Individual accuracy was calculated on the basis of rejecting 3 out of 4 tokens in each of the ungrammatical conditions and accepting 3 out of 4 in the grammatical conditions. On this measure, 9 of 22 L1 Chinese participants could consistently and accurately distinguish appropriate versus inappropriate syntactic contexts for any, and 10 of 25 advanced-proficiency L1 Arabic speakers.

1.3 Properties of German

The similarities between English, Arabic and Chinese with respect to the distribution of existential NPI any may facilitate knowledge of NPI distribution. The overall patterns of judgements by L1 Arabic and L1 Chinese-speaking learners do not bear this out in detail. Learners demonstrate much less robust knowledge in environments involving lexical-semantic negation from verbs or adverbs licensing any. Given that Chinese and Arabic pattern similarly to English in pertinent respects, an interesting exploration of L1 effects would come from how learners acquire the distribution of any when the L1 provides no facilitative role. This is the case with German.

The closest equivalent to any is represented by forms with the indefinite marker irgend-. The irgend-series shares a number of functions with English any (see Haspelmath, 1997). This overlap motivates the assumption that the learner grammar will map irgend- onto any in a Feature Reassembly model (Lardiere, 2009). However, the syntax and semantics of the items differ in some subtle ways, potentially leading to learnability issues. Most pertinent as motivation for this study is the fact that the irgend-series is not subject to NPI distributional requirements. As we will see later in the discussion, clarifying the specific pattern of results requires a more detailed study of specific interpretive properties. We return to this below as it is necessarily somewhat speculative given that this was not included as a variable in the replication study.

The irgend-series is formed by prefixing irgend to existential or indefinite phrases (wh-words, pronouns such as jemand someone or determiners such as ein- [=a]). Where there is optionality between a non-prefixed existential and an irgend-compound, irgend- introduces an implication of ignorance or indifference on the part of the speaker as in (6) (see Haspelmath, 1997, p. 45; Kratzer & Shimoyama, 2002).

While irgend-forms are not NPIs (as evidenced by 6), syntactic context does affect their interpretation as free choice any(thing) versus non-specific some(thing) (see Penka, 2020, p. 653). Sentences 7–10, which translate the ungrammatical English sentences in Table 1, illustrate these interpretations. In each case, an irgend-form is possible, but will be interpreted as non-specific something/one rather than as an NPI.

The irgend-series also occurs in the grammatical contexts for NPI any, illustrated in (11)–(15). A slight complication comes from the fact that German has a negative quantifier kein- which expresses not-any. An irgend- compound under the sentential negator is not possible without a specific pragmatic intention. For the sake of completeness, example (13) illustrates this pattern. The implication in this sentence is that the homework assignment is something special, i.e. “not just any old homework”.

Let’s return to questions of learnability that this comparison raises. As already noted, Feature Reassembly (Lardiere, 2009) would assume that L1 German-L2 English learners map the lexical item any onto the features associated with irgend-, due to the shared functions and distribution. However, one element of the subsequent learning task is to restrict the distribution of any as an NPI.² In all contexts where any occurs, a representation which implicates the existential indefinite properties mapped from irgend- would grammatically license the occurrence, but potentially assign a different interpretation (see above). Crucially from a learnability perspective, ungrammatical usages would also be licensed by this representation, and on the basis of positive evidence, restructuring poses a problem. The presence of the some/any distinction in English might aid in arriving at a target representation. Nevertheless, without negative evidence, it is difficult to rule out ungrammatical occurrences of any in L1 German-L2 English. In many contexts, there is optionality between some and any, complicating the task of arriving at a target NPI representation which specifically disallows any in certain contexts. In a judgement task, where ungrammatical occurrences are presented, learners who have not reassembled features so as to restrict the distribution will syntactically license the usage (though potentially with different meanings).

Of course, as noted in the studies of teaching materials by GMW, negative evidence is available from instruction to the effect that any is restricted to overtly negative contexts or questions as compared to some (see Tesch, 1990 for similar findings from EFL materials for German-speaking learners). In common with GMW’s predictions, the signature of reliance on a learned pedagogical rule would be a rejection of grammatical sentences where licensing of any is due to semantic features of verbs or adverbs, and acceptance of ungrammatical sentences with any outside the scope of negation.

Based on the discussion so far, and the insights furnished by GMW’s experiments, we explore potential effects of pedagogical rules in explicit input, versus observability in incidental input. The hypothesised interplay between these factors and the eight sentence types is outlined in Table 2.

As discussed in section 1.2, pedagogical rules focus strongly on occurrences in questions and negative declaratives, facilitating knowledge of sentence types 1, 2, and – as a common pedagogical counter example – sentence type 5. Sentences involving any licensed by semantically negative verbs and adverbs occur in incidental input. Knowledge of these is therefore in principle derivable from positive evidence, even though these features are not typically taught. Ungrammatical occurrences of types 6–8 in Table 2 are not taught. They are by definition not available from incidental input as ungrammatical sentences do not occur, and they would not be ruled out by mapping from L1 German.

Key comparisons are therefore (i) between the taught grammatical questions and negative declaratives versus untaught types with negative verb and adverbs. A difference in the acceptability ratings would index potential effects of instruction. Such instructional effects can be further explored by (ii) comparing the rating accuracy of grammatical questions (1) versus parallel ungrammatical affirmative declaratives (5), and also by contrasting negative declaratives (2) with sentences containing scope violations (6). Because the ungrammatical sentences are by definition absent from positive input, teaching effects would be indexed by more accurate ratings of sentence type 1 versus 5 and 2 versus 6.

Finally, input and L1 effects would be identified by comparisons of (iii) grammatical, observable but untaught sentence types with negative verbs (3) and adverbs (4) versus ungrammatical ones with non-factive verbs (7) and possibility adverbs (8), which are neither taught nor observable. A tendency to reject ungrammatical sentences altogether would indicate an overgeneralisation of the pedagogical rule, while a tendency to accept all of them would indicate L1 effects. More accurate ratings in sentence types 3 and 4 against 7 and 8 would furthermore point towards developing grammatical L2 knowledge despite the absence of an overt licensor.

2.1 Participants

72 learners of English studying in an English Language Teaching degree programme participated in the study. This sample size would permit detection of effects as small as η_p² = 0.13 (Kumle, Võ & Draschkow, 2021; Schütze, 2016). Participants’ mean age was 22.48 years, (SD = 3.29), and German was their L1, either mono- or multilingually. 53 identified as female, 18 as male, and one as non-binary. All participants had had eight years of regular, instructed EFL school teaching prior to their tertiary studies; in addition to that, around 11% reported a stay abroad between one and six months, while around 14% had stayed for a longer period of time in an English-speaking country. The great majority, though, had never had an extended stay in an English-speaking environment (75%).

All 72 participants took part voluntarily, anonymously, and with explicit consent, but without any financial renumeration. At the time of testing, participants had been studying English at tertiary level for at least two years and were approaching C1 proficiency level (CoE, 2001), i.e. “advanced” as defined by the Common European Framework of reference for language. This was measured by proficiency tests at the beginning of the degree and after two years in the study programme. The data of all 72 participants were used for reliability and bias analyses of the acceptability judgement ratings.

2.2 Materials and Method

As in the GMW studies, participants rated the naturalness of sentences containing any in a paced acceptability judgement task. 12 lexicalisations for each of the sentence types were created, resulting in 96 tokens, which were distributed in pseudo-randomised experimental lists following a Latin Square design. Participants saw 24 randomised test items (3 items per experimental condition) as well as eight fillers, either systematic distractors or items piloting a separate study on negative inversion. Test items, raw data and R-script are accessible via https://osf.io/9qt46/.

The acceptability judgement task was administered online through SoSciSurvey (Leiner, 2021). Participants were informed that the experiment was part of a study about learners’ grammatical knowledge, the procedure was explained briefly, and two untimed items for practice were provided.

The sentences were rated on a four-point Likert scale, labelled I’m sure this sounds natural, I think this sounds natural, I think this does not sound natural, and I’m sure this does not sound natural. Since the acceptability judgement scores were intended to elicit participants’ intuition about the naturalness of a given sentence, a gradable rating option was chosen instead of a binary right-wrong choice. Contrary to Marsden et al. (2018), an even point-scale was chosen so that participants were forced into one or the other direction, thereby avoiding clusters in the classic middle-ground option cannot decide. For the same reason, the option I do not know was omitted. Thus, it was assumed that even mild and subtle intuitions about a sentence’s acceptability could be captured.

2.3 Data Coding and Analysis

Signal Detection Theory (SDT) was used to analyse how participants discriminated between grammatical and ungrammatical sentence types and whether they showed any general bias towards accepting or rejecting (Huang & Ferreira, 2020). Huang & Ferreira (2020) propose using SDT for judgement data as it delivers insights into how well participants differentiate between grammatical and ungrammatical sentences and whether they have any bias in their preferences. To this end, acceptability ratings were coded as follows. Detecting a grammatical sentence’s grammaticality resulted in a hit; rejecting a grammatical sentence resulted in a miss. If participants correctly detected the ungrammaticality of a sentence, they scored a correct rejection, or true negative. Where they rated ungrammatical sentences as grammatical, they scored a false alarm. Based on this 2×2 matrix, the SDT indices d’ and c were calculated; d’, a measure of sensitivity, captures participants’ reliability when distinguishing between grammatical and ungrammatical sentences. Index c, the criterion location index, expresses conservative or liberal biases.

Mixed ordinal regression modelled potential effects of sentence type on the acceptability ratings. The four Likert points were treatment-coded as ordered categorical data, with the levels “1”, “2”, “3”, and “4”, “1” reflecting the least appropriate, and “4” reflecting the most appropriate rating. Thus, a rating such as I’m sure this sounds natural for a grammatical sentence would be coded as “4”, most appropriate, while the same rating for an ungrammatical sentence would score a “1”, least appropriate. The sentences were sum-contrast-coded, so that the predictor levels represent the difference between these levels and the grand mean, thus effectively ‘centring’ the effects of the variable’s levels at this value (Schad et al., 2020). This coding was chosen since none of the eight sentence types could reasonably be thought of as a baseline level against which the other estimates would then be calculated.

For the mixed modelling, we used cumulative link mixed regression with Laplace approximation from the ordinal package (Christensen, 2019, version 2019.12.10) in R (R Core Team, 2021). Main regression effects from the cumulative link mixed model are reported based on Type-II Wald χ2-tests, as implemented in the RVAideMemoire package (Hervé, 2022). Partial effects statistics (Wald) were derived from the model summaries produced within ordinal.

(16)	Maria bedauert, dass sie irgendwas bei der Feier gegessen hat. Maria regrets that she irgend.what at the party eaten has. ‘Maria regrets that she ate anything/something at the party.’

(17)	Maria glaubt, dass sie irgendwas bei der Feier gegessen hat. Maria thinks that she irgend.what at the party eaten has. ‘Maria thinks that she ate something at the party.’

(18)	Maria hat kaum irgendwas bei der Feier gegessen. Maria has hardly irgend.what at the party eaten ‘Maria hardly ate anything at the party.’

(19)	Maria hat wahrscheinlich irgendwas bei der Feier gegessen. Maria has probably irgend.what at the party eaten. ‘Maria probably ate something or other at the party.’