The negative concord illusion: an acceptability study with Czech neg-words

1 Introduction

In many languages, the type of phrase exemplified by the English no student requires the use of negation on the main verb of the local clause. Take the two following contrasting examples, from Czech, a language exhibiting this behaviour, and Standard English, where this construction leads to a different interpretation altogether:

In Czech, the interpretation of (1) is that there was not a single student who got an A. On the other hand, the same use of both the negative quantifier no and sentential negation in Standard English leads to the opposite reading, namely that it was not the case that no student received an A, meaning that there were at least some students who did (de Swart 2009; but for conflicting evidence, see Blanchette et al. 2018). Languages that behave in the way Czech does have become known as “negative concord languages”.

Cases of negative concord appear with a class of words known as “neg-words” (or “n-words” as per Laka 1990), such as the Czech žádný ‘no’ in (1). Giannakidou (2006) defines them as expressions that can be used in a structure with a sentential negation or another neg-word while yielding only one logical negation. There have been many theoretical attempts at analysing negative concord (for an overview, see Giannakidou and Zeijlstra 2017). One class of approaches sees the phenomenon as essentially a case of agreement (Haegeman and Zanuttini 1991; Zanuttini 1991; Zeijlstra 2004, 2008]). The proposal by Zeijlstra suggests that neg-words carry an uninterpretable negative feature that requires checking against an interpretable negative feature that is found on the verb. Under this view, negative concord is not dissimilar from other cases of agreement such as subject-verb agreement. Crucially for our purposes, this approach differentiates neg-words from negative polarity items (NPIs) and has them licensed in the syntax.

Another class of approaches sees neg-words as being subsumed under NPIs (Giannakidou and Quer 1995, 1997]; Ladusaw 1992). Under this view, neg-words in negative concord languages are similar to NPIs such as any. Some proposals see neg-words as universal quantifiers (Giannakidou 2000) that scope above negation; others see them as indefinites (Ladusaw 1992). This approach sees neg-words as in themselves non-negative (for accounts that treat neg-words as negative, see De Swart 2009; De Swart and Sag 2002). They are indefinites similar to phrases with NPIs such as any student, yet they come with a roofing requirement in Ladusaw’s (1992) parlance. What these accounts have in common is that they relate the conditions of the grammaticality of neg-words to the licensing conditions of NPIs.

A proper treatment of the theoretical landscape of negative concord is beyond the scope of this article. What is, however, crucial for the following investigations into the processing of the phenomenon is the understanding that negative concord is treated either as syntactic agreement or as NPI licensing. We return to the theoretical issues in the discussion and suggest how examining the processing of negative concord can inform theory.

While the theoretical literature disagrees on what the correct analysis of negative concord is, psycholinguistic investigations have revealed that both classic NPIs and syntactic agreement are susceptible to interference from grammatically (or semantically) irrelevant elements in online processing. One of the ways this is manifested is in grammaticality illusions. This is a phenomenon whereby under certain conditions, comprehenders mistakenly judge certain sentences as acceptable even though they are in fact ungrammatical (Phillips et al. 2011). Below, we review the literature on how these illusions arise in the general cases of agreement and NPI licensing. We then build a case for why the same effect ought to be seen with negative concord as well.

Let us begin by discussing agreement and its illusion phenomenon – agreement attraction:

While both (3) and (4) are ungrammatical due to the presence of the plural marking on the verb were, which mismatches the singular number of the subject head the key, comprehenders judge (3) in a speeded acceptability task as grammatical more often than (4) due to the former containing a noun (cabinets), which, while not the actual subject, is plural (Wagers et al. 2009).^[1]

These illusions have been well replicated in number agreement both in English (Hammerly et al. 2019; Patson and Husband 2016; Royer 2021) and in other languages, such as Greek (Paspali and Marinis 2020) and Spanish (Gonzalez Alonso et al. 2021). In Czech, Lacina and Dotlačil (2024) showed the presence of grammaticality illusions in number agreement, which were reliable yet of a small effect size. Additionally, there is a body of evidence showing that in reading tasks, the presence of a feature-matching attractor can cause relative speed-ups in the processing of sentences with ungrammatical number agreement (Franck et al. 2015; Jäger et al. 2020; Lago et al. 2015, 2019]; Parker and An 2018; Slioussar 2018; Tucker et al. 2015; Wagers et al. 2009). In general, agreement attraction manifested in comprehension either as illusions or in reading times has been shown to be present in many phenomena, including gender agreement (Gonzalez Alonso et al. 2021; Paspali and Marinis 2020; Slioussar and Malko 2016; Tucker et al. 2021) and honorifics (Kwon and Sturt 2016). The influences of various factors on the effect, for example that of case marking, have also been investigated (Avetisyan et al. 2020).

Let us now discuss the other processing phenomenon of interest here – npi illusions. Grammatically, NPIs are licensed in the scope of a downward-entailing operator (Ladusaw 1992), which is not the case in the following sentence:

It has been found that when exposed to sentences such as (5), comprehenders come under impression that the NPI ever is licensed even though it is not and that this effect is caused by the presence of the downward-entailing negative quantifier no in the relative clause (Drenhaus et al. 2005; Parker and Phillips 2016; Vasishth et al. 2008), which ought to be irrelevant. Just as with agreement attraction, the effect has been replicated repeatedly in both English and other languages (Muller 2022; Orth et al. 2021; Schwab 2022; Xiang et al. 2009; Yanilmaz and Drury 2018; Yun et al. 2018).

One recent crucial finding in the literature has been that only some licensors and only some NPIs can support illusions. Orth et al. (2021) find that while simple sentential negation grammatically licenses NPIs, its presence in a relative clause as an illusory licensor does not cause the effect in English. In Turkish, on the other hand, sentential negation did lead to illusions (Yanilmaz and Drury 2018). Schwab (2022) finds that in German, only strengthening NPIs such as jemals ‘ever’ can support illusions, while attenuating ones such as so recht ‘really’ do not. These results suggest a specificity to the effect, also dependent on the language in question.

Both classic NPIs and morphosyntactic agreement give rise to illusions and have even been proposed to be explainable by the same theories – cue-based retrieval (see Engelmann et al. 2019; Lewis and Vasishth 2005; Vasishth et al. 2008) and feature percolation accounts (see Eberhard et al. 2005; Orth et al. 2021). Cue-based retrieval theories explain attraction and NPI illusions by stipulating that the effects arise because comprehenders store incoming elements as bundles of features with dependency being established by means of memory retrieval, which depends on featural cues. In ungrammatical sentences where the attractor or illusory licensor matches one of the cues needed, processing speed-ups are predicted (Logačev and Vasishth 2016). Feature percolation accounts, on the other hand, suggest that features such as number, gender, or polarity in the case of NPIs percolate up the syntactic tree causing the illusion (Eberhard et al. 2005; Orth et al. 2021; Patson and Husband 2016). There are, however, many other competing separate theories of agreement attraction (see Dempsey et al. 2022; Hammerly et al. 2019; Häussler 2009; Keshev et al. 2024; Smith et al. 2018) and NPI illusions (see Giannakidou and Etxeberria 2018; Mendia et al. 2018; Schwab 2022; Xiang et al. 2013) that do not necessarily subsume the two phenomena under a single explanatory umbrella.

All of this points us to speculate whether the processing of negative concord is also susceptible to grammaticality illusions. This is an empirical question that has so far not been addressed to our knowledge. In the current study, we do just that by conducting experiments with Czech neg-words and we show that there is indeed a grammaticality illusion.

2 Experiments 1 and 2

We ran two web-based acceptability judgement experiments with native speakers of Czech, a strict negative concord language. Experiment 1 used the speeded acceptability judgement task, while Experiment 2 tested the same sentences without a time limit for participants’ answers. In both experiments, our participants saw sentences with an unlicensed neg-word and an intervening relative clause that included a case of a combination of a licensed neg-word and sentential negation. We predicted illusion effects to arise in the speeded task (Experiment 1) and to be absent when enough time is given to participants (Experiment 2).

Let us now set out precisely what we hypothesize in the current study with regard to the acceptability of Czech sentences with negative concord within both the speeded task and the untimed one. The first hypothesis concerns both Experiment 1 and 2 and sets out a general effect of ungrammaticality:

Next, we propose the hypothesis that an illusion effect caused by the presence of negation and a neg-word in a relative clause ameliorates ungrammaticality in the speeded paradigm, but not in the untimed acceptability judgement task, as attraction and NPI illusion effects are seen in speeded tasks and early measures (Phillips et al. 2011):

2.3 Results

2.3.1 Experiment 1: speeded acceptability

The reader may consult the mean proportion of ‘yes’ responses by condition from Experiment 1 in Figure 1 and in Table 2 and the mean RTs of both types of responses in Figure 2 and in Table 3. The results of the models are summarized in Tables 4 and 5.

Figure 1:

Mean proportion of ‘yes’ responses by condition from Experiment 1, with standard errors.

Table 2:

Descriptive statistics of mean acceptance in Experiment 1 by condition.

Sentence (as per Table 1)	MCNeg	RCNeg	Mean acceptance	SD
a.	Negated	Affirmative	0.68	0.46
b.	Affirmative	Affirmative	0.08	0.27
c.	Negated	Negated	0.58	0.49
d.	Affirmative	Negated	0.14	0.35

Figure 2:

Mean response times (RTs) of ‘yes’ and ‘no’ responses by condition from Experiment 1, with standard errors.

Table 3:

Descriptive statistics of response times (RTs) in milliseconds by condition in Experiment 1.

Sentence (as per Table 1)	MCNeg	RCNeg	Mean RTs in ms	SD
a.	Negated	Affirmative	820	641
b.	Affirmative	Affirmative	631	448
c.	Negated	Negated	860	661
d.	Affirmative	Negated	780	591

Table 4:

The fixed effects of the nested responses model (Experiment 1).

Fixed effects	Estimate	SE	z value	p value
(Intercept)	−1.109	0.167	−6.646	<0.001***
MCNeg	1.828	0.101	18.07	<0.001***
MCNeg (ungrammatical): RCNeg	0.510	0.142	3.58	<0.001***
MCNeg (grammatical): RCNeg	−0.272	0.089	−3.05	0.0023**

Table 5:

The fixed effects of the nested log-RTs model (Experiment 1).

Fixed effects	Estimate	SE	df	t value	p value
(Intercept)	6.420	0.030	112.1	216.16	<0.001***
MCNeg	0.075	0.015	54.74	5.11	<0.001***
MCNeg (ungrammatical): RCNeg	0.082	0.013	3,184	6.21	<0.001***
MCNeg (grammatical): RCNeg	0.019	0.013	3,183	1.45	0.148

The response model revealed that there was a main effect of MCNeg – ungrammatical sentences were judged as acceptable less often compared to grammatical ones. Crucially, there was a significant effect of RCNeg within the ungrammatical condition, suggesting that the negation in the relative clause increased acceptance rates. The effect of RCNeg was in the opposite direction in the grammatical MCNeg condition, associated with lower rates of acceptance.

As for RTs, there was a main effect of MCNeg and within this factor, RCNeg was significant only when nested within the affirmative (ungrammatical) condition, where RTs were longer with the negated (illusory) relative clause.

2.3.2 Experiment 2: untimed judgements

For the untimed judgements of Experiment 2, we report the mean ratings per condition in Figure 3 and in Table 6. The results of the model are in Table 7.

The nested model showed a main effect of MCNeg. Again, ungrammatical sentences were judged by comprehenders as less acceptable. The effect of RCNeg was significant only in the grammatical condition and (marginally) insignificant in the ungrammatical one.

Figure 3:

Mean ratings by condition from Experiment 2, with standard errors.

Table 6:

Descriptive statistics of mean ratings in Experiment 2 by condition.

Sentence (as per Table 1)	MCNeg	RCNeg	Mean ratings	SD
a.	Negated	Affirmative	4.58	1.94
b.	Affirmative	Affirmative	2.03	1.11
c.	Negated	Negated	3.96	1.92
d.	Affirmative	Negated	2.24	1.35

Table 7:

The fixed effects of the nested ratings model (Experiment 2).

Fixed effects	Estimate	SE	z value	p value
MCNeg	1.767	0.201	8.79	<0.001***
MCNeg (ungrammatical): RCNeg	0.129	0.072	1.80	0.0725
MCNeg (grammatical): RCNeg	−0.454	0.069	−6.58	<0.001***

3 Discussion

Let us start with our hypotheses. We can see that Hypothesis A has been confirmed, since in both experiments, ungrammatical sentences with a missing negative marker in the matrix clause were judged as unacceptable more often compared to grammatical ones. Hypothesis B has also received support, since we saw that the presence of negation and a licensed neg-word in the relative clause increased the probability of acceptance for otherwise ungrammatical sentences in the speeded acceptability Experiment 1 but crucially not in the untimed judgement Experiment 2.

Turning now to interpretations, we can say that our results provide the first evidence for the existence of what might be termed the negative concord illusion effect. We ran two judgement experiments with native speakers of Czech where we exposed them to sentences with an unlicensed neg-word where an intervening relative clause was providing an illusory licensor in the form of a negated verb. We found that the inclusion of this syntactically irrelevant negation caused Czech comprehenders to come under a grammaticality illusion in a speeded acceptability task. They judged ungrammatical sentences as acceptable more often when a relative clause contained a negated verb compared to when it was non-negated. However, this effect disappeared when the task did not have a time limit, a pattern typical of linguistic illusions (Phillips et al. 2011).

The RT results also support the illusion interpretation, since we found that comprehenders took longer in the ungrammatical condition when the illusory licensor negation was in place. Presumably, they were experiencing interference from the syntactically irrelevant negation and had to suppress the tendency to answer incorrectly. On the other hand, the grammatical conditions took longer and there was no difference between them. These longer RTs could be evidence of comprehenders weighing in factors beyond syntactic well-formedness and evaluating the sentences’ plausibility and pragmatic felicity.

What requires discussion is our finding of an effect of the negation in the relative clause on the acceptability of grammatical sentences. Prima facie, this appears to contrast with cases of agreement attraction, where “illusions of ungrammaticality” have mostly not been observed (Hammerly et al. 2019; Wagers et al. 2009).^[3] However, we argue against this position. The stimuli in the affirmative relative clause conditions replaced the neg-word with an additional modifier (this was done to maintain the same sentence length across conditions), which added more semantic content. This could have increased the sentences’ plausibility. Overall, negated sentences are also more difficult to process in comparison to affirmative ones and they are less felicitous in “out of the blue” contexts (for an overview, see Kaup and Dudschig 2020). We believe that the lower acceptability of grammatical sentences with another negation in the relative clause was due to their decreased pragmatic felicity. That the effect was present even in the untimed Experiment 2 is evidence of this.

One argument against the illusion interpretation is that the two negations present in one of the conditions were causing general processing difficulties resulting in more uncertainty, leading to responses being drawn towards the chance level of 50 %, which would explain the opposite effect of the negation in the relative clause in the grammatical and ungrammatical conditions. This, however, cannot explain the asymmetric pattern observed in the untimed Experiment 2.

It is interesting to compare our negative concord results in Czech to the ones obtained in the domain of agreement attraction in the same language. In general, Czech comprehenders have been found to be mostly immune to the effect and to exhibit it only under the conditions of case syncretism, that is, the sameness of form between different combinations of case and other features (Caha 2019), with very small effect sizes (Chromý et al. 2023a, 2023b; Lacina and Chromý 2022; Lacina and Dotlačil 2024). In comparison, the illusion effect seen here appears to be robust. We speculate that this could be due to the negative marker ne- being invariant in its use – negation is signalled in the same way whether it is in the matrix or relative clause. Therefore, the situation is similar to the syncretic cases where attraction emerges in Czech in that the attractor (or illusory licensor) carries the same form as the correct verb, which allows the illusion to surface. While there has been no study on NPI illusions in Czech, experimental studies have contrasted neg-words and strong NPIs in the language with the focus on neg-raising (Dočekal 2014, 2020]; Dočekal and Dotlačil 2016, 2018]), suggesting that in Czech, neg-words and strong NPIs are distinct categories and are licensed in different environments.

Overall, the results bear the hallmarks of two other linguistic illusions which have been prolifically studied, namely agreement attraction (e.g., Wagers et al. 2009) and NPI illusions (e.g., Vasishth et al. 2008). Yet, under which umbrella the negative concord effect ought to be subsumed cannot be decided based on the current data. However, NPI illusions and agreement attraction have been found to be differently sensitive to specific manipulations, chief among them distance. Parker and Phillips (2016) and Muller (2022) showed that when the NPI is moved further away from the illusory licensor, the effect disappears. Yet the same was not the case for agreement attraction (Experiment 7 in Parker and Phillips 2016). Whether the negative concord illusion is susceptible to distance manipulation could tell us how to classify it.

It has recently been found that NPI illusions are sensitive to both the type of licensor used (Orth et al. 2021) and the type of NPI (Schwab 2022). In English, it was found that simple sentential negation does not give rise to illusions (Orth et al. 2021), but the use of negative quantifiers does. In Turkish, on the other hand, sentential negation has been found to elicit the effect (Yanilmaz and Drury 2018). In our study, we not only used sentential negation in the relative clause, but also another negative quantifier. Whether the illusion stays or disappears when the quantifier is removed and only sentential negation is kept would also be informative regarding the categorization status of the observed effect.

Additionally, further data examining this illusion could also be used as evidence in the theoretical debate on the nature of negative concord. Should it turn out that the processing profile of negative concord patterns more with the behaviour of agreement attraction, it would suggest that the phenomenon ought to be analysed as syntactic agreement. On the other hand, if the negative concord illusion ended up having the characteristics typical of NPI illusions, this would point in favour of subsuming negative concord under NPIs.

While this study provided the first evidence for the existence of an illusion in the processing of negative concord, these results need to be supplemented with further studies to ascertain its robustness. The use of different methods, such self-paced reading or corpus examinations of whether the phenomenon exists in speech errors, and cross-linguistic validation on other closely related Slavic languages such as Polish, and both strict negative concord languages (such as Greek or Hungarian) and non-strict ones, that only allow in certain sentential structures (such as many Romance languages; Giannakidou and Zeijlstra 2017), are in order.

4 Conclusions

The current study revealed that there is a type of grammaticality illusion that can be induced in languages exhibiting negative concord, that is, the requirement of neg-words for sentential negation in the same clause. This was shown in a speeded acceptability experiment with Czech comprehenders where an irrelevant negation in a relative clause caused the illusory acceptance of an unlicensed neg-word in the matrix clause. This effect disappeared in an untimed task. Further research is needed to determine whether this negative concord illusion is a case of NPI or agreement-related processes. This could be done by making use of the previously discovered differential sensitivity of the illusion effect to factors such as distance (Muller 2022).