Previous ERP studies of negation in the visual modality have shown a relatively late integration of negation in the comprehension process. This means that ERP effects of negation processing have often shown up in later time-windows, typically 600 ms and onward (Lüdtke et al., 2008) as compared to non-negated information where a typical N400 is observed in earlier time-windows (Fischler et al., 1983). This delay in integration has been argued to be caused by two steps involved in the processing of negated information through the two-step simulation model (Kaup et al., 2006; Lüdtke et al., 2008). The first step involves the simulation of what is expressed in the scope of negation and the second step involves the integration of negation and the simulation of the negated concept.

Other studies emphasizing a dynamic view of negation have highlighted the role of context in processing (Nieuwland and Kuperberg, 2008; Tian et al., 2010, 2016; Orenes et al., 2014, 2016). For instance, contextual cues provided by sentence structure (e.g., simple or cleft) shift the focus of the information and hence affect the representation of the affirmative concept in the integration of negation (Tian et al., 2016). Similarly, Nieuwland and Kuperberg (2008) found that the processing of negation is less effortful in pragmatically plausible contexts than in implausible contexts. More specifically, incongruities in plausible, negated contexts elicited an N400 effect comparable to that elicited by incongruities in affirmative sentences, while this was not the case in implausible contexts.

Revisiting the processing of incongruities in negated and affirmative contexts in a previous ERP study (Farshchi et al., 2021), we investigated the integration of negation in visually presented affirmative and negated sentences such as the ones shown in example (1). The first part of the sentence contained either an affirmative adjective (e.g., authorized), a sententially negated adjective (e.g., not authorized) or a prefixally negated adjective (e.g., unauthorized). Prefixal negation was included because very little is known about the processing of such negated items (Tottie, 1980). The second part of the sentence contained a critical word in the form of one of the members of an antonym pair (e.g., correct/wrong), which rendered the whole sentence either congruent or incongruent.

(1) The White House announced that the new Obama biography was authorized/not authorized/unauthorized and the details in the book were correct/wrong in actual fact.

In Farshchi et al. (2021), incongruities in affirmative sentences elicited an N400 reflecting difficulties with the integration of the incongruent word (cf. Kutas and Hillyard, 1980; Fischler et al., 1983) followed by a P600 suggesting a re-analysis of the sentence (cf. Kolk et al., 2003). In sententially negated sentences, however, both congruent and incongruent conditions elicited an equally large N400, suggesting processing costs for the negated forms also in the congruent condition. A difference between incongruent and congruent sententially negated sentences showed up in the form of a later negativity (450–600 ms). This negativity suggests a delayed detection or processing of incongruities due to the complex interpretation of the negated meanings, e.g., not authorized, and possibly a memory search for the negated information presented earlier in the sentence. Finally, in prefixally negated sentences, incongruities elicited a sustained anterior negativity, which was interpreted as reflecting taxed working memory processes (cf. Müller et al., 1997; Martín-Loeches et al., 2005) in the retrieval of the prefixally negated meanings, e.g., unauthorized. These findings reveal that the processing of sententially negated information was more difficult than that of affirmative information, and the processing of prefixally negated information was even more difficult than the other two forms. The results of Farshchi et al. (2021) do not reflect the two processing steps described by the two-step simulation model. If negation had been processed in two steps, the negator not should have been ignored initially, and consequently, the congruent negated condition (not authorized with wrong) should have elicited a larger N400 compared to the incongruent negated condition (not authorized with correct). This was, however, not the case. Instead, the results suggest that factors such as the multifunctional meaning potential of not and memory retrieval processes have an impact on ease of processing, which highlights the importance of context for the processing of negated meanings.

As far as we know, the investigation of negated sentences using ERPs in the auditory modality is restricted to two studies (Herbert and Kübler, 2011; Herbert and Kissler, 2014). In Herbert and Kübler (2011), participants were asked to evaluate the truth value of false statements such as Dogs don’t bark and true statements such as Dogs don’t fly. All sentences were spoken by a female German speaker and an inter-stimulus interval (ISI) of 1500 ms was used between the negation word and the target word, corresponding to a stimulus onset asynchrony of 1810 ms. The ERPs to the target words (bark/fly) were recorded. The authors did not find any modulation of the N400 by truth value (e.g., Dogs don’t fly vs. Dogs don’t bark) (cf. Nieuwland and Kuperberg, 2008) or by semantic relatedness (e.g., semantic relatedness between dog and bark and semantic unrelatedness between dog and fly in the statements, respectively) (cf. Fischler et al., 1983; Lüdtke et al., 2008). Instead, false statements elicited a larger frontal negativity (300–600 ms) followed by a larger parietal positivity (600–1000 ms). We argue that the lack of an N400 effect in response to semantic relatedness may have been caused by the relatively long ISI of 1500 ms between the negator (n’t) and the target word (bark/fly), which may have allowed for processing and integration of negation. Note, however, that if negation had indeed been integrated, the N400 ought to have been modulated in the congruent condition (i.e., increased N400 for bark as a dog indeed barks), which it did not. The length of the ISI could have affected the pattern of ERP effects as this ISI is significantly longer than the typical ISIs used in ERP paradigms.

In the second study, Herbert and Kissler (2014) used the same sentences as Herbert and Kübler (2011) but with a shorter ISI between the negator and the target word (approximately 500 ms). The experiment consisted of a passive listening task followed by an active truth evaluation task with repeated presentations of the sentences. The question was whether the first task would facilitate the processing of negated sentences in the second task (resulting in enhanced N400 and late positive potential), or would lead to the same effects attenuated due to the repetition of the sentences. In the second task, half of the participants performed the truth evaluation mentally and the other half did the evaluation through button presses. The overall effects were larger in the active task than in the passive task, suggesting that active evaluation of sentences requires more resources than passive evaluation. Additionally, and in line with the results of the first study, Herbert and Kübler (2011) found a larger late positive potential for incongruent negated sentences compared to congruent negated sentences in the passive task.

Despite the valuable findings of these two studies, we argue that the presentations were somewhat unnatural: either the negated forms were presented with long ISIs or stimulus sentences were repeated, which could have affected processing. Also, no affirmative counterpart was included in the design of these studies that would allow for a comparison between the processing of negated and non-negated information. This calls for further exploration of processing negation in the auditory modality, which is what we do in the present study.

Thirty-two right-handed native speakers of English (21 females; mean age = 24.8 years, range = 20–33 years) participated in the study with the following countries of origin: US (12), UK (10), Canada (3), Australia (3), South Africa (2), Ireland (1) and New Zealand (1). All participants listed English as their native language. Six participants listed an additional language as their second native language. All participants provided written consent prior to the experiment and received a cinema voucher as compensation. None reported any neurological disorders or hearing problems.

The stimulus sentences from Farshchi et al. (2021) were read aloud by a female native speaker of British English and recorded specifically for the experiment. One item from the previous stimulus set was deemed unnatural and removed. The total number of sentences was 306 sentences (17 adjective sets × 3 sentence frames × 6 conditions) presented in three lists of 102 sentences each so that each participant listened to only two conditions of the same sentence frame. Sentence frame refers to one of the three sentential contexts created for each adjective set. Condition refers to the combination of negation (affirmative/prefixal negation/sentential negation) and congruency (congruent/incongruent). See Supplementary Appendix A for a complete list of stimulus sentences. The final lists were pseudo-randomized in that the sentences from the same frame were separated by at least two or more other frames.

To ensure that the ERP effects were elicited by the manipulations only, identical primes and critical words were used. The stimulus sentences were spliced into four different parts (Table 1) with the use of Audacity (Audacity team, 2019) and Praat speech editing software (Boersma and Weenink, 2019). The sentence parts were combined into one audio file with breaks of 9 and 11 ms, respectively before the critical words (parts 2 and 4 in Table 1). These intervals were chosen based on the average pause length between the offset of the preceding word and the onset of the target word in the original recordings. The average duration of the sentences was 7.13 s (SD = 0.96). The EEG was time-locked to the onset of the critical word (part 4 in Table 1), which was identified using Praat. PsychoPy software program (Peirce et al., 2019) was used for the presentation of the experiment.

Participants were informed about the procedure, signed a consent form, and were seated at a 110 cm distance from a monitor with an electrode cap on their head. Participants were instructed to focus on a cross in the middle of the screen while listening to sentences presented through loudspeakers placed on either side of the screen. After each trial, three question marks appeared on the screen prompting participants to answer “yes” or “no” to the question “Did the sentence make sense, logically?” by pressing a green or red button on a keyboard. The placement of the buttons was counter-balanced across participants. Response accuracy, response time, and continuous EEG were recorded. There was no official pause or break although participants were told they could take a short break if needed. The task lasted approximately 40 min.

EEG recording and processing criteria were identical to those presented in Farshchi et al. (2021). Continuous EEG data was recorded at a 500-Hz sampling rate from 30 scalp electrodes (NeuroScan Easycap), 2 mastoid electrodes and 4 facial electrodes for recordings of EOG. From a total of 3,264 trials (102 items by 32 participants), 3% of the data was rejected with similar number of trials in each condition (519–534).

The procedure for building statistical models was identical to that adopted in Farshchi et al. (2021). In the analyses of both behavioral and ERP data, sentence type (affirmative, sentential negation, and prefixal negation) and congruency (congruent and incongruent) were used as predictors. In all the analyses, affirmative and congruent conditions were coded as the reference level. Subject and item information were included as random effects. Item information was only available for behavioral data.

For ERPs, the mean amplitude in three time-windows was analyzed in order to capture three responses observed in previous auditory studies (Herbert and Kübler, 2011; Herbert and Kissler, 2014): the N400 (450–650 ms), the P600 (650–800 ms), and a late positive potential (LPP, 800–1000 ms). Six separate analyses were performed for the three time-windows over the anterior region (electrode sites F3/4, FZ, FC3/4, FCZ, C3/4, and CZ) and the posterior region (CP3/4, CPZ, P3/4, PZ, O1/2, and OZ) as differences in topographical distributions have previously been observed between the visual and auditory modalities (e.g., McCallum et al., 1984; Holcomb and Neville, 1990, 1991). Post hoc pairwise comparisons were performed, and z-scores were adjusted accordingly. Supplementary Appendix B contains more details about the analyses.

The proportions of accurate responses and mean response times within the congruent and incongruent conditions for each sentence type are listed in Table 2. In the analysis of the accuracy rates, the model with the strongest predictive accuracy included a significant interaction between sentence type and congruency. The output of the model revealed lower accuracy rates for both prefixal negation (β = –0.88, SE = 0.18, z = –4.67, p < 0.001) and sentential negation (β = –0.70, SE = 0.19, z = –3.66, p < 0.001) compared to affirmative forms within the congruent condition. Lower accuracy rates were found for the incongruent than the congruent condition in affirmative sentences (β = –0.84, SE = 0.20, z = –4.08, p < 0.001) but a significant interaction effect between both negation types and congruency revealed that this direction of effects was not observed for prefixal negation (β = 1.19, SE = 0.25, z = 4.68, p < 0.001) or sentential negation (β = 0.80, SE = 0.25, z = 3.16, p < 0.01). No other congruency effects were found for the two negation types, nor were there any significant differences between the two negation types.

For the response time analysis, there were neither significant interaction effects nor significant effects of sentence type. However, the main effect of congruency was significant suggesting that the responses in the incongruent condition were slower than those in the congruent condition (β = 0.13, SE = 0.04, t = 2.17, p < 0.05).

As seen in Figure 1, incongruities in affirmative sentences elicited an N400 effect over the posterior region. This N400 effect was followed by a P600 effect over the same region and continued through the third time-window (1000 ms) over both regions. For sententially negated sentences, no N400 effect was observed for the incongruities but a larger positivity for incongruities in the P600 time-window over the anterior region. For prefixally negated sentences, the effect of incongruity was restricted to a larger positivity in the LPP time-window, and this effect approached significance over the posterior region.

In all six analyses of the ERP data, a significant interaction between the two predictors, sentence type and congruency, was observed. We pursued the interactions from each analysis with three contrasts between the congruent and incongruent conditions within each sentence type. The results of these contrasts are shown in Table 3 and discussed below.

When comparing the findings from the auditory processing with those from the visual processing of the same sentences (Farshchi et al., 2021), the N400–P600 response for the incongruities in affirmative sentences in the auditory modality was replicated, which signals that the affirmative sentences were easiest to process and rightly served as the baseline.

For sententially negated sentences, the ERP patterns differed between the two modalities. In the visual modality, incongruities elicited a larger centro-parietal negativity in the P600 time-window, which was taken to reflect delayed and costly processing of the negated sentences. In the auditory modality, incongruities in the sententially negated sentences instead elicited a larger positivity (650–800 ms) over the anterior region. We argued earlier that this positivity could indicate the weak nature of the incongruities in negated contexts where the incongruent endings might have been taken as felicitous sentence completions.

In both modalities, prefixal negation is the most difficult condition to process based on behavioral and ERP patterns. In the visual processing of these forms (Farshchi et al., 2021), incongruities elicited a sustained anterior negativity expanding over both time-windows (300–600 ms), which was taken to reflect higher cognitive effort and working memory load (Müller et al., 1997; Martín-Loeches et al., 2005). In the auditory processing of these sentences, instead, a larger posterior positivity (600–1000 ms) was found for incongruities. This points to a delayed and cognitively more demanding process in sentence comprehension and integration (Kolk et al., 2003; van Herten et al., 2005; Bornkessel-Schlesewsky and Schlesewsky, 2008). Together, these ERP effects reflect a greater difficulty in processing prefixally negated contexts compared to the other conditions.

Overall, the results indicate that there are many similarities between the two modalities, at least in terms of the overall difficulty in processing between the conditions. It may be argued that the effects found in the auditory processing of negated sentences (late positivities) are easier to interpret and align with previous findings as they are more well-established and more commonly found in studies of violations of semantic expectancy (Kolk et al., 2003; van Herten et al., 2005; Bornkessel-Schlesewsky and Schlesewsky, 2008). Accordingly, it can be argued that the auditory processing of negated sentences is easier than their visual word-by-word processing as the ERP effects in the visual modality (sustained anterior and centro-parietal negativities) reflected a higher working memory and cognitive load, which may be factors that have masked the actual cognitive processes behind the comprehension of negated meanings.