Negation is far from being a uniform phenomenon, as it exhibits subtle nuances that do not always align with the straightforward logical interpretation commonly associated with its use. In fact, there are non-standard uses of the negative marker, for which its presence fails to affect the truth conditions of the proposition it occurs in. This phenomenon is usually referred to as Expletive Negation (EN). The contrast between Standard Negation (SN) and EN is visible in the following Italian example:

The negative marker non affects the proposition in which it appears in (1a), as its propositional content is that Gianni will not come to the party. However, it does not fulfil the same function in (1b), where the propositional content that Gianni will come to the party is not negated, in spite of the presence of non. In fact, the negative marker can be omitted altogether, as shown in (1c), without affecting the grammaticality or the propositional meaning of the sentence. Such expletive use of the negative marker has been described in earlier literature using terms such as pleonastic, redundant, or non-truth-functional negation (Jespersen 1917; Horn 1989), reflecting its lack of contribution to the semantics of the sentence, and aside from examples such as (1a–c) in Italian. This type of negation is not limited to examples such as (1a–c) in Italian; it is reported in other sentence types within Italian (Greco 2019a) and in a number of other languages as well (Jin & König 2020).

Interestingly, languages exhibiting EN can be divided into two classes: those with two distinct negative lexical items – one for SN and one for EN – and those with a single negative lexical item used for both functions. Examples of the former include Latin (non vs. ne) and Greek (min and dhen), while examples of the latter include Italian (non) and German (nicht) (see Greco 2020 and references therein).

While the standard use of negation affects the truth-conditions of a given sentence, the expletive use of a negative marker is accounted for in the literature in different ways, mostly analysing it as either a non-negative or a negative element. On the one hand, in the non-negative approach, EN has been traditionally associated with semantic emptiness and vacuity. Approaches that endorse this view consider EN as a vacuous morpheme that has lost its negative value (Haegeman and Zanuttini 1991; Espinal 1992; Fischer 1992; van der Wouden 1994a; 1994b; Van der Wurff 1999, among others). Espinal (1992), for instance, accounts for the expletive interpretation of negation in terms of Logical Absorption, a structural mechanism that takes place at the Logical Form. Logical Absorption is triggered by the presence of specific lexical items – such as comparative quantifiers of inequality, as well as certain prepositions and verbs (Espinal 1992: 347–348) – that in certain languages license EN incorporation in their representation. Through this operation, EN results as semantically empty. Espinal’s approach ultimately anchors the expletiveness of the negative marker in EN to a structural operation at the Logical Form, through which another element incorporates the logical content of negation.

On the other hand, according to the negative approaches proposed by Abels (2005), Greco (2017; 2019b; 2020), Delfitto et al. (2019), the negative marker of EN always realises a negative operator, but it does not act on the propositional layer and, therefore, it does not affect the truth-value conditions of a given sentence. Specifically, Greco (2017; 2019b; 2020) argues that the twofold interpretation of negation (standard or expletive) is encoded in syntax. The standard reading of negation, which reverses the truth-condition of a sentence, arises when the Negation Phrase (NegP) is merged in the TP-area (Zanuttini 1997) during the vP phase (Chomsky, 2001). Thus, the negative particle has access to the predicative core of the sentence in the case of SN. The expletive reading is instead yielded when NegP is merged in the CP Domain (Left Periphery, c.f. Rizzi 1997; Rizzi & Bocci 2017) when the vP is already closed, meaning that NegP cannot access the predicative domain.1 In this view, the reading of negation is strictly determined by its structural position, and EN and SN are seen as two manifestations of the same element.

Five Areas of Interest (AoI) were taken into consideration for data analysis:

More specifically, the verb of the matrix clause (AoI 1) is considered for the relation it entails with EN according to semantic approaches, such as Cépeda’s (2017) explanation of temporal sentences described above. The conjunction (AoI 2) is taken into account to observe any possible effect stemming from the difference in sentence type. Finally, AoIs 3 and 4 account for the (expletive) negative clause, while AoI 5 accounts for any possible end-of-sentence effects. As for the metrics, in line with our predictions, for AoIs 3 and 4 we extracted both early (Gaze duration, i.e., the sum of the durations of all fixations made on an AoI till the eye goes past it to the right) and late measures (Total reading time, i.e., sum of the durations of fixations made on an AoI), along with Regression in (the number of regressive fixation on an AoI) and Visit (the number of times the eye enters the AoI) count to observe possible revisiting of critical areas. Additionally, for AoI 3 (the negative marker), we extracted the Skipping rate, which accounts for the expected skipping of short and frequent (function) words and on AoI 4 we extracted Regression path duration (the sum of the durations of all fixations made from the entry to an AoI till its exit from the right, including any regressive fixations to previous material and excluding any later re-visit to the AoI after following its exit) as a measure of sentence processing difficulty. Given that they were presented before the critical area introducing negation, for AoIs 1 and 2, we extracted only late measures (Total reading time, Regression in count and Visit count). For AoI 5, we focused only on Total reading time, Regression path duration, and Regression out (the number of regressive launches from an AoI) count to examine potential wrap-up effects. A schematic summary of the dependent variables extracted for each AoI, together with the experimental predictions made for each, can be found in the Supplementary Materials. These predictions are grounded in previous empirical findings in the eye-tracking literature, which show that each measure indexes different processing stages, and in the theoretical assumptions motivating the present study.

Regarding our data analysis approach, generalised linear mixed models (GLMMs) with a Poisson distribution were used for discrete count variables, including Visit count, and Regression in and out count. For the Skipping rate in AoI 3, we applied a GLMM with a binomial distribution. Reading time (RT) measures, which included Gaze duration, Total reading time, and Regression path duration, were analysed using linear mixed-effects models (LMMs) after log-transforming the RTs. The fixed factors for each model were sentence type (causal vs. temporal) and negation status (affirmative vs. negative). Since EN and SN appeared in two distinct sentence types (temporal and causal, respectively), a direct comparison was not considered appropriate. Any observed differences in reaction times could stem from varying processing demands associated with each sentence type rather than being attributable to the type of negation. Therefore, the effect of negation status (sum-coded as –1 negative and +1 positive) was examined separately for each sentence type by coding sentence type as a factorial variable and treating first temporal and then causal sentences as a reference level in separate analyses. The interaction between sentence type and negation status was also included in the fixed structure. The only exception to this line of reasoning was the glmms fitted for AoI 3, which, since this area was not presented in affirmative sentences, only included sentence type as a fixed factor, with causal sentences as a reference level. The random structure of each model was then established using a data-driven approach, starting with the maximal random effects model (following Barr et al. 2013), which included random intercepts and slopes for participant and item variability. As these models resulted in singular fits, we followed the parsimonious model-selection approach advocated by Bates and colleagues (2015), guided by the results of a randomised principal component analysis (rePCA).

For RT data trimming, we employed a model criticism approach following an initial mild trimming step (Baayen & Milin 2010). In the first stage, RTs below 80 ms and above 1000 ms were excluded to filter out anticipatory and excessively delayed responses. After log-transforming the reaction times and fitting the model, standardised residuals were computed by dividing each residual by the standard deviation of the residuals. These standardised residuals were then examined to identify potential outliers. Specifically, any data points with standardised residuals exceeding an absolute value of 2.5 were flagged. Since none of the models produced more than 3% of observations with residuals surpassing this threshold, no further data trimming was deemed necessary.

For the sake of conciseness, only measures that produced statistically significant effects will be discussed thoroughly in the results section. The full analysis script with results can, however, be found in our online repository on OSF (osf.io/abwx3).

AoI 1 and 2. None of the extracted measures produced significant effects in these areas.

AoI 3. For RT measures, we found no significant effects on the negative particle non in the models analysing RT measures, with only a marginally significant effect for Gaze duration (β = 0.06, p = 0.089), signalling a trend of longer RTs on the negative marker in temporal sentences (i.e., EN). As for discrete variables, however, non in temporal sentences had significantly fewer visits (β = –0.190, p = 0.025), although no effect is seen for Skipping rate (β = –0.131, p = 0.121) or Regression in count (β = –0.096, p = 0.505). The differences in Visit count, Skipping rate and Regression in count between the two conditions under examination can be appreciated by looking at Figure 1.

AoI 4. This interest area consisted of the full verb, which could be under the scope of negation, composed of two separate words, an auxiliary and a past participle. The analyses of all RT measures for this area all produced significant effects. For Gaze duration, an effect of negation is seen: negative sentences had longer RTs on the verb. This effect is observed for both causal sentences presenting SN (β = –0.075, p < 0.001) and temporal sentences presenting EN (β = –0.054, p < 0.001). The same effect is seen for Regression path duration (for causal as reference: β = –0.075, p < 0.001; for temporal as reference: β = –0.056, p < 0.001). However, the effect found in Regression path duration seems to be driven by Gaze duration itself, as the difference between Regression path duration and Gaze duration did not reach 60 milliseconds in any of the four conditions, and can therefore be considered negligible. When it comes to Total reading time, an effect of negation is only seen for causal sentences (β = –0.041, p = 0.013), but not for temporal ones (β = –0.002, p = 0.919). In addition, in both models an effect of sentence type (for causal as reference: β = –0.070, p = 0.003; for temporal as reference: β = 0.007, p = 0.022) and a marginally significant interaction between negation status and sentence type models (for causal as reference: β = 0.040, p = 0.090; for temporal as reference: β = –0.004, p = 0.077) can be found. The differences between the four conditions on Gaze duration, Regression path duration and Total reading time are shown in Figure 2.

The results for discrete variables provide additional information, enriching the processing picture provided by RTs. In fact, for Visit Count, the only significant effect found is an effect of negation status in the model taking temporal sentences (those presenting EN in the negative condition) as reference (β = 0.048, p = 0.044), with the verb in affirmative sentences being visited more, while no effect is seen in the causal model (that is, the one pertaining to sentences containing SN in their negative condition). The same effect of negation status in the temporal sentences is replicated for Regression in count (β = 0.101, p = 0.016).

AoI 5. For the last word, no significant effects were found in the models analysing Total reading time or Regression out count. As for Regression path duration, an effect of sentence type was detected, which was mirrored in both models, showing that the duration of regression paths starting from the last word was longer for causal sentences than for temporal ones (β = |0.068|, p = 0.021).

The eye-tracking results reveal different processing patterns for negation across the two sentence structures analysed, namely causal and temporal clauses, licensing respectively either standard (SN) or expletive negation (EN). For SN, the pattern that emerges is much in line with previous research by other scholars (Wason 1965; Trabasso et al., 1971; Clark and Chase 1972; Carpenter and Just 1975; Carpenter et al., 1999; Hasegawa et al., 2002; Kaup et al., 2007) when a coherent pragmatic licensing context is missing (see Nieuwland & Kuperberg 2008; Orenes et al. 2014) as well as with the previous SPR study using the same stimuli (Porrini et al. 2025): on the verb (AoI 4) in the scope of negation gaze duration, total reading time and regression path duration were all longer in negative sentences compared to affirmative ones. For EN, only some of the effects are replicated: the same effects are seen for gaze and regression path duration, but not for total reading time, where no effect of expletive negation is seen. In addition to this, for temporal sentences, the number of visits and regressions in on the verb from the left were higher in the affirmative condition as opposed to the negative one, where EN was presented. No similar effect is found in causal sentences (presenting SN), suggesting that the difference between the two types of negation can be seen in re-reading behaviour.

Compared to previous studies utilising a SPR methodology, the advantage of implementing an eye-tracking paradigm is that it captures different stages and nuances of sentence processing. Namely, the detection of eye movements during online reading enables a distinction between early measures – typically associated with the initial lexical integration – and late measures – reflecting comprehension difficulty, including meaning integration, which was crucial to our initial predictions. Furthermore, it allows us to assess the effects of both types of negation based on reading patterns observed through multiple measures, rather than relying on single p-values in our interpretation of the results. Going back to the patterns outlined above, for SN, the present eye-tracking findings not only replicate previous results but also provide a more fine-grained picture of its processing. In particular, the predicted increase in both early and late RT measures on the verb in the scope of negation (AoI 4), despite no differences in the number of visits and regressions, suggests that processing the (standard) negation of the predicate demands more cognitive effort than affirmative sentences.

A comparable initial effort is observed in early measures for EN for the same AoI (the verb in the subordinate clause); however, such an effect is not maintained in late measures. An increase of the RTs for gaze duration – immediate lexical access and early word processing – was predicted by our initial hypotheses, and it is observed on the verb (AoI 4) for EN compared to its affirmative counterpart. The similarity between SN and EN in terms of early measures appears to rule out the semantic emptiness of EN, thus agreeing with the so-called negative approaches (Abels 2005; Greco 2017; 2019b; 2020; Delfitto et al. 2019). These approaches postulate that while on the one hand EN and SN both represent a negative entry, on the other, the negative operator affects the predicative core of the sentence only in the latter case. Again, in line with these theories and with our experimental prediction, we found that EN and SN exhibit distinct processing patterns at later stages of processing, with EN not showing extra processing cost compared to affirmative sentences in late RT measures such as total reading time. This might suggest that the presence of the negative marker in its expletive use causes no extra processing cost when dealing with entire sentence comprehension and meaning integration. It must be noted that, contrary to our prediction, for regression path duration, EN and SN also showed similar processing patterns. As outlined in the results section, this result seems to be mainly driven by the high correlation between gaze duration (an early processing measure) and regression path duration, and is thus not interpreted as a late-stage processing effect.

In addition to the effects found in the critical area presented after introducing negation, namely AoI 4, no significant effects were found on the two initial portions of our experimental sentences: the verb in the main clause (AoI 1) or the finché and perché conjunctions (AoI 2). This is of particular interest for our temporal sentences, as it suggests that no reanalysis of the temporal relation between main and subordinate clause occurred after reaching the negation, and that the semantic contribution of EN did not need to be reassessed after reaching the negative marker and the verb in the subordinate clause, as might have been the case if the negation was initially assumed to be standard, and only later the negative marker was reanalysed as a semantically empty expletive element, in line with non-negative accounts of EN.

Concerning the effects on the negative marker itself (AoI 3), interestingly, no significant difference in skipping rate between the two sentence types was attested. This suggests that despite the inherent negative semantics of the temporal conjunction finché (see 1.1) and the naturalness of EN in such constructions (as attested in Porrini et al. 2025), the presence of non following finché was not more expected than after perché to a degree that could be captured through this measure. The unexpected trend found for gaze duration on the same AoI, with longer reading times on non in the temporal sentences than in the causal one, further suggests that the negative marker was not strongly expected after finché.5 Our analysis of AoI 3, however, revealed a lower visit count in the EN condition. This finding potentially suggests that negation requires more cognitive resources when it is standard, especially at later stages of processing. Nevertheless, as it was not paired with a significance on the measure counting regressions, it must be taken cautiously. Additionally, it must be noted that comparing the negative particle directly in two sentence types is not without its issues, as any significant effects found might be due to differences in processing effort required by the sentence types, and not necessarily by the type of negation used. The result must be considered also in combination with the effects of sentence type detected on total reading times on AoI 4 and for regression path duration on the last word (AoI 5), both suggesting that causal sentences require more effort than temporal ones. While no immediate effort is observed on the conjunction itself (AoI 2), our results highlight a general difference in processing effort between causal and temporal clauses later in the sentence, independent of the presence of negation (standard or expletive). This finding does not affect our conclusions with respect to EN or SN, as the experimental design allowed us to assess the cost of these two types of negation relative to their affirmative counterparts. Thus, any differences between causal and temporal sentences should not influence our interpretation of the effects of negation. Nonetheless, it warrants a brief discussion. While the difference may be attributed to differences in the processing effort necessary to correctly interpret causal and temporal relations between clauses, we would refrain from claiming that causal relationships are inherently more difficult to interpret than temporal ones. In fact, in a previous study using the same stimuli (Porrini et al. 2025), native speakers rated the temporal versions of the sentences as more natural than the causal ones. This was likely due to the need to construct sentences that were coherent in both their affirmative and negative forms. As a result, the causal sentences did not always describe events where the matrix clause followed directly from the subordinate clause, which may have made the causal link less salient in some cases, making the causal sentences less natural and, consequently, potentially harder to process.

The acceptability ratings collected in the study by Porrini and colleagues (2025) might also provide an explanation for another reading pattern detected in the present study. Namely, in temporal sentences (i.e., those presenting EN in their negative condition), participants visited the verb (AoI 4) significantly more in the affirmative condition. This might be explained by the fact that the results of the acceptability judgment task showed that Italian native speakers find temporal sentences containing the negative marker (with EN) more natural than their affirmative counterparts. This might have led to increased re-reading behaviour on the verb when the negative marker is absent.

Overall, our findings provide support for negative accounts of EN. Furthermore, the results are compatible with the view of EN as an instance of a complement function that does not alter the truth-condition of the sentence in which it appears, but rather, acts at a different level of meaning (a.o., Greco 2017; 2019b; 2020; Delfitto et al. 2019; for EN in temporal clauses: Tovena 1996; Español-Echevarría & Vegnaduzzo 2000; Cépeda 2017). Although further research needs to be conducted to address this issue, it could be argued that the early and late processing profiles we observed for the two types of negation reflect their different roles within the syntactic structure (Greco 2020). As mentioned in the introduction, both EN and SN involve the integration of a NegP into the sentence structure, with SN involved in the TP domain, and thus affecting the predicative core of the sentence directly, while EN is involved in the CP domain, when the predicative relations have already been established. The early measures, where the presence of both EN and SN caused an effect compared to affirmative sentences, may reflect this shared need to integrate the negation structurally, regardless of whether NegP operates on TP or CP. In contrast, the difference observed in late measures – where only SN incurred a greater processing cost compared to its affirmative counterpart, while EN did not – may be attributed to the differences between these structural positions: only when negation operates on the TP layer, thus having access to the predication, it contributes to the propositional content reversing the truth conditions of the clause, which may entail greater integration effort at later stages of interpretation. On the other hand, when negation operates on the CP domain, without affecting the truth conditions of the proposition, no effort at later stages is expected, reducing processing effort for EN compared to SN. As we have seen in the Introduction, some semantic approaches to EN view it as a constraint on the temporal sequence of the two events described in the sentence (Cépeda 2017). While it is not possible to draw predictions about the processing effort associated with such a difference, our results show no significant effect on the verb of the matrix clause (AoI 1), signalling that in our stimuli the relation established by the presence of EN in the finché-clause does not entail a differential processing cost associated with the main event.