Processing sentences with sentential and prefixal negation: an event-related potential study

ABSTRACT

This study is concerned with the integration of negation in relation to two models of the processing of negation: (i) the two-step model (Lüdtke et al., 2008), according to which negation involves two representations where negation is ignored in the first representation, and (ii) the pragmatic view (Nieuwland & Kuperberg, 2008), which posits that negation can be integrated without delay if it is used in a natural context. The processing of two negated forms (not authorised and unauthorised) and an affirmative form (authorised) was studied in complex congruent and incongruent contexts. Incongruities in affirmative sentences elicited a biphasic N400–P600. In both types of negated sentences, ERP patterns associated with higher processing difficulties (anterior and central negativities) were observed. The results did not support one or the other model, suggesting that the processing of negation cannot be fully captured by either of them.

KEYWORDS:

• Negated meanings
• ERP
• Sentential integration
• Processing cost
• Opposition
• Bounded adjectives

Introduction

Expressions of negation are powerful elements in human communication. They can be used for different purposes and give rise to different effects ranging from negating the existence of something to mitigating assertions as well as inducing irony and metaphor (Fein et al., 2015; Giora, 2006; Giora et al., 2010; Paradis & Willners, 2006). For a long time, the received view of language processing has been that expressions of negation delay processing and bring about comprehension difficulties (Dudschig et al., 2019; Ferguson et al., 2008; Fischler et al., 1983; Kaup et al., 2006; Lüdtke et al., 2008), but findings that this is not always the case have also been presented (Nieuwland & Kuperberg, 2008). Also, most previous research on the processing of negation has targeted negated forms such as not and no, while negation in the form of prefixes such as un, as in unauthorised, has been given very little attention. Such gaps and conflicting results raise questions about how negation is represented, processed and comprehended. Using event-related potentials (ERPs), we explore the time course of sentential integration of prefixal negation (e.g. unauthorised) and sentential negation (e.g. not authorised) as compared with (non-negated) affirmative forms (e.g. authorised). The aims are (i) to determine whether negation causes difficulty for comprehension (e.g. delay in integration) or whether it is processed similarly to affirmative forms, and if so, (ii) whether prefixal negation is processed similarly to sentential negation or to affirmative forms.

Most earlier behaviourial studies have provided evidence for the difficulty of the processing of negated information as reported by longer reading times and higher error rates (Clark, 1969; Just & Carpenter, 1976; Just & Clark, 1973; Slobin, 1966; Wason, 1959; Wason & Jones, 1963). Different explanations for the greater processing difficulty of negation have been proposed and tested, such as the extra word or syllable in the negator (Clark & Chase, 1972) and the negative semantics per se of negated expressions (Just & Clark, 1973; Wason, 1959; Wason & Jones, 1963). One of the proposals for the processing difficulty of negation is the two-step simulation model put forward by Kaup and colleagues (Kaup, 2001; Kaup et al., 2006; Kaup & Zwaan, 2003). According to this model, the processing of negation involves two steps or representations. For instance, in the first step in The door was not open, participants ignore the negator and process and represent what is in the scope of negation, that is “an open door”. This step is followed by an integration phase which involves the simulation of the opposite state, namely “a closed door”. The first step in this process creates a delay in the direct integration of negation and has been associated with the difficulty of processing negated meanings (Kaup et al., 2006).

Negation has also been studied using neurophysiological measurements such as ERPs. In an early ERP study, Fischler et al. (1983) investigated the comprehension of affirmative and negated sentences. In their study, participants were instructed to verify statements such as A robin is a bird (true–affirmative), A robin is a tree (false–affirmative), A robin is not a tree (true–negated) and A robin is not a bird (false–negated). In true–affirmative and false–negated statements (primed conditions), there was a semantic match between robin and bird, while in false–affirmative and true–negated statements (non-primed conditions), there was a mismatch between robin and tree. They measured the N400 response that previously had been shown to be larger in the case of more effortful processing of incongruent semantic information (Kutas & Federmeier, 2011; Kutas & Hillyard, 1980). Their results revealed a larger N400 effect for false compared to true affirmative sentences and the reverse pattern in negated sentences, where true–negated sentences elicited a larger N400 than false–negated sentences. In other words, a larger N400 effect was observed for the non-primed conditions compared to the primed conditions, irrespective of negation.

It was argued that when negated sentences are presented, participants first process the “inner proposition” and this proposition is then evaluated against their semantic memory. For example, in the case of A robin is not a tree, participants first process A robin is a tree before negating the proposition. In this first stage, the semantic mismatch between tree and robin leads to a larger N400 (Fischler et al., 1983). The findings by Fischler et al. (1983) were later explained by two different models of negation processing. The first model was the two-step simulation model discussed earlier (Kaup et al., 2006). According to this model, the processing of the “inner proposition”, suggested by Fischler et al. (1983), could be in line with the first simulation step such that the larger N400 elicited by tree, compared to bird in the sentence A robin is not a tree/bird is caused by the semantic mismatch in the processing of the concept without negation (A robin is a tree) that is, the semantic mismatch in the first simulation step.

Additional evidence for the two-step model has been provided by Lüdtke et al. (2008) who studied the time-course of integrating negation using ERPs. In their study, participants were presented with written sentences (e.g. In front of the tower there is a/no ghost) which were followed by a congruent picture (i.e. a ghost in front of a tower for the affirmative sentence and a lion in front of a tower for the negated sentence) or an incongruent picture (e.g. a lion in front of a tower for the affirmative sentence and a ghost in front of a tower for the negated sentence). Sentences were either affirmative or negated yielding four conditions similar to those in the experiment by Fischler et al. (1983). In order to test the time-course involved in the integration of negation, the pictures were presented either 250 or 1500 ms after the sentences. Just like in the study by Fischler et al. (1983), the N400 was always (irrespective of the time interval between the sentences and the pictures) larger in the non-primed conditions (i.e. the conditions where the picture that was shown was not mentioned in the previous sentence) compared to the primed conditions (i.e. the conditions where the picture that was subsequently shown was mentioned in the previous sentence). In later processing (550–1000 ms after the presentation of the image), the length of the delay (short vs. long) affected the ERPs in that a larger positivity was found for negated sentences compared to affirmative sentences after the long-delay condition. The authors conclude that these results suggest that in the short-delay condition, negation had not yet been fully integrated; hence, no effect of negation was found in the later time-window. However, in the long-delay condition, a main effect of negation showed up in the later time-window, which suggests that participants had time to integrate negation as reflected in the ERP effects.

The results by Lüdtke et al. (2008) also support the view that negation is processed in two steps. The N400 effect for primed conditions reflects processes associated with the first simulation step of the non-negated information, while the late positivity for negated sentences reflects the second step where negation is integrated. At the same time, it is important to note that the N400 was not larger for incongruent pictures (i.e. pictures that did not correctly depict the content of the sentences) than for congruent pictures in negated sentences. In other words, no larger N400 was observed for the incongruent negated sentences compared to the congruent negated sentences in either of the time delays. This suggests that the conclusion regarding the full integration of negation in the long time delay is in fact limited.

The delay in the integration of negation was further corroborated in a study where the comprehension of non-negated and negated information was tested in counterfactual statements (Ferguson et al., 2008). The stimuli consisted of sentences such as If cats were not carnivores, they would be cheaper for owners to look after. Families could feed their cat a bowl of carrots/fish and listen to it purr happily. In this example, the semantically primed word fish was expected to create an anomaly while the semantically non-primed word carrots would not. The results of this study showed that there was a larger N400 for the consistent but non-primed word (carrots) than that to the inconsistent primed word (fish) irrespective of negation. The authors therefore conclude that the meaning of the sentence was initially processed against real world knowledge and negation was ignored (Ferguson et al., 2008).

Furthermore, in two experiments, Dudschig et al. (2019) also investigated the processing of affirmative and negated sentences using semantic and world-knowledge violations. In the first experiment, negated sentences such as Zebras/Thoughts/Ladybirds are (not) stripy were presented while in the second experiment, a clause was added at the beginning of the sentence, with or without negation, as in It is (not) true that zebras/thoughts/ladybirds are stripy in order to give participants more time to process negation. In neither of the negated sentences did they find that congruency effects modulated the N400. Instead, the N400 response was modulated based on semantic incongruities regardless of negation. They conclude that simply providing participants with extra processing time does not make the integration of negation easier and that the N400 does not always reflect the processing of meaning at the sentence-level (Dudschig et al., 2019).

While evidence in favour of the two-step model has been presented using various paradigms and techniques, other findings suggest that the first step proposed by this model is not always present or mandatory in the processing of negated meanings (Orenes et al., 2014; Orenes et al., 2016; Tian et al., 2010; Tian et al., 2016). More specifically, Tian et al. (2016) argue that depending on the sentence structure in which negation is presented (simple vs. cleft), the focus of the content changes and therefore the presence of the first representation step is also affected. In an eye-tracking study using the visual world paradigm, they presented participants with simple sentences (e.g. John has/hasn’t ironed his brother’s shirt) and cleft sentences (e.g. It is John who has/hasn’t ironed his brother’s shirt), followed by two images, one consistent with the affirmative statement (an ironed shirt) and one consistent with the negated statement (a crumpled shirt). They argue that in simple sentences, the focus is on the verification of the affirmative statement, while in cleft sentences the focus is on “who has or hasn’t done it”. The results show that in negated simple sentences, participants’ attention was first on the image consistent with the affirmative statement followed by the image consistent with the negated statement. This indicates that the first representation step was available in the processing of negated simple sentences. However, in negated cleft sentences, participants’ attention was divided between the two images, suggesting that in negated cleft sentences, the focus was no longer on the verification of the affirmative statement (i.e. the first representation step). These results suggest that the contextual cues provided by the linguistic structure play a role for whether or not it is necessary to represent the affirmative statement in the processing of negated information (Tian et al., 2016).

In line with the importance of contextual factors, the other approach to the processing of negation proposes that negation is not necessarily more difficult if it is used in an appropriate and sufficient context. The results of a number of behavioural studies have shown that the processing cost for negation is only found if the context is inappropriate, insufficient, confusing or ambiguous (Arroyo, 1982; Glenberg et al., 1999; Wales & Grieve, 1969). In other words, it is possible to modulate the processing cost of negation if enough context for the comprehension of the negated forms is presented (Glenberg et al., 1999), or if negation fulfils its “natural function of distinguishing the exception from the norm” (Arroyo, 1982).

In a similar vein, a pragmatic account of negation processing was put forward by Nieuwland and Kuperberg (2008). According to their view, negation is not necessarily more difficult to process or it is not necessarily integrated with delay if it is used in an adequate context where it fulfils its natural function of rejecting a plausible statement. They criticised the use of unnatural and implausible statements such as A robin is not a tree (Fischler et al., 1983) and instead created two types of context based on naturalness and truth-value ratings: pragmatically licensed contexts (e.g. With proper equipment, scuba-diving isn’t very dangerous/safe and often good fun), and pragmatically unlicensed contexts (e.g. Bulletproof vests aren’t very dangerous/safe and used worldwide for security), in which negation is not informative and does not reject a plausible statement. They found that in the processing of licensed contexts, negation was readily integrated, and incongruities in these sentences elicited a larger N400, similar to those in affirmative sentences. In the processing of unlicensed contexts, however, this was not the case. Indeed, for the pragmatically unlicensed contexts, a larger N400 was observed for both congruent and incongruent conditions. Therefore, they conclude that negation is not necessarily more difficult to process if it is used in a natural context (Nieuwland & Kuperberg, 2008).

Convincing evidence has been provided in support of both views on the processing of negation (i.e. two-step model and the pragmatic account) . This suggests that while in some cases, the processing of negation involves two representations, in other situations, the two representation steps are not obligatory. Maybe it is not the case that one or the other account is wrong, but rather that the experiments tap into different cognitive mechanisms and processes. In the case of the two-step model, the experiments focus on category inclusion (a robin is a bird) or inherent category features of entities (stripy or not). In the pragmatic account, on the other hand, the focus is on slightly more complex but also more natural contexts, tapping into assessments about cause and effect, plausibility and relevance.

In the light of the discussion above, we revisit the integration of negation in context. Our study differs from the previous ones in essentially two ways. First, we added prefixally negated forms (un) to our experimental materials. To the best of our knowledge, prefixal negation has been investigated in three studies from the 1970s. The results from these studies are not in agreement with one another. While one study found that these forms were more difficult to process than their non-negated counterparts (Sherman, 1973), the other two did not (Hoosain, 1973; Sherman, 1976). Therefore, it remains unclear whether the y6 processing of prefixally negated forms is similar to that of sentential negation or that of affirmative forms. Second, instead of the relatively short sentences which have been used in previous studies, we use longer and more complicated sentences, which resemble sentence types that we encounter on a daily basis, for instance in news media outlets. Behavioural and ERP responses to these sentences are registered and analysed.

This study

Sentences such as The White House announced that the new Obama biography was authorised/unauthorised/not authorised and the details in the book were correct/wrong in actual fact are created and presented to participants. In each sentence, the first part contains the negated adjective (underlined), and the second part makes use of a pair of opposites (bold), according to which the meaning of the sentence becomes either congruent (i.e. the sentence makes logical sense) or incongruent (i.e. the sentence does not make logical sense), and to which the ERPs are time-locked. It is the participants’ task to decide the congruency by pressing one of two response keys. By investigating the ERP effects to the critical words, we can seek an answer to whether there are difficulties in the processing and integration of prefixally negated and sententially negated forms in the earlier context. In this study, processing refers to the patterns observed in both behavioural and ERP data. Integration is operationalised as the end result of successful processing. Successful processing or integration of the negated forms is measured relative to the patterns observed in the processing of the affirmative form that is taken as the baseline condition. Note that processing does not necessarily entail integration.

If the two-step model holds, negation (not authorised) should not have been integrated by the time participants reach the critical word (correct/wrong) but instead, the processing of negation should be in an intermediary stage of simulating the first representation step (i.e. the concept in the scope of the negation: authorised). This should be represented by a larger N400 for the congruent (e.g. not authorised with wrong) than for the incongruent (e.g. not authorised with correct) condition. This is because negation (not authorised) has not yet been integrated and the processing of the critical word and sentence congruency is assessed against the affirmative concept (authorised). However, if sentential negation has been integrated at the critical word, in line with the pragmatic view of negation, we expect to observe a larger N400 for the incongruent (e.g. not authorised with correct) compared to the congruent (e.g. not authorised with wrong) condition in both negated and affirmative sentences. Even though our stimuli are not directly comparable with those of Nieuwland and Kuperberg (2008) as they include naturalness and truth-value scores, we believe that we can compare the setup of this study with the pragmatic view of negation, as we also make use of longer and more complex contexts where different discursive factors may modulate the processing of negation.

With regard to prefixal negation, no previous ERP studies have been conducted on the processing of these forms and therefore, no clear predictions are made. If prefixally negated forms are processed similarly to affirmative forms, we should observe a larger N400 for the incongruent than the congruent sentences. However, if prefixal negation is in fact a negator in the sense that it takes scope over and operates on the meaning of another lexical element, ERPs should reveal patterns similar to the sentential negation condition, if those patterns are different from the non-negated sentences.

Method

Materials and design

In order to test the processing of negated and non-negated sentences, we created congruent and incongruent contexts. The combination of three “sentence types”: 1. affirmative (baseline condition): authorised, 2. prefixal negation: unauthorised, and 3. sentential negation: not authorised, and two “congruency” conditions: 1. congruent, 2. incongruent, resulted in a three by two design. To create the sentential contexts, 18 antonymous adjective pairs (e.g. authorised–unauthorised) were extracted from the Corpus of Contemporary American English (COCA) (Davies, 2008).

An important criterion in selecting these adjectives was the boundedness property. Previous work (Paradis & Willners, 2006, 2013) has shown that negation is sensitive to the configuration of boundedness of meanings (in this case adjectival meanings). Based on this property, adjectival meanings can be divided into bounded (non-scalar) and unbounded (scalar) meanings, where bounded non-scalar meanings are those that divide a conceptual domain into two distinct parts (e.g. dead-alive). Unbounded meanings, in contrast, are those that correspond to a range on a scale (e.g. small–large) (Paradis, 2001). When non-scalar adjectives such as dead-alive are negated as in not alive, the interpretation is typically the absolute opposite, that is, dead. However, when scalar adjectives such as small–large are negated as in not small, the interpretation is not necessarily the absolute opposite of the scale, but instead an attenuated reading of the opposite much like fairly in fairly large (Paradis & Willners, 2006, 2013). Therefore, in combination with scalar meanings, the negator functions in a similar way to a degree modifier (fairly). In other words, in combination with scalar meanings, negation may evoke a range of possible meanings on a scale with varying degrees of the property, which in this case is the property of size. Since we wanted to make sure that the negated adjectives in our study, (unauthorised and not authorised) were as similar in meaning as possible – that is they were interpreted as absolute opposites of the non-negated meaning (authorised) – we opted for bounded, non-scalar meanings instead of scalar meanings. This was because negation in the form of both prefixes and not, in combination with unbounded scalar meanings gives rise to a wider range of interpretations while that is not the case for bounded meanings (Farshchi et al., 2019).

The other criterion for the selection of these adjectives was frequency. The pairs were selected from three frequency categories where: (i) the prefixed form was more frequent than the non-prefixed form, (ii) the prefixed form was less frequent than the non-prefixed form, and (iii) the prefixed form was by and large as frequent as the non-prefixed form (see Appendix A). This was done in order to have a range of frequencies and to allow for generalisations to be made. For each adjective set, three sentence frames were created (see Table 1), where the first clause contained the negated or non-negated adjective and the second clause contained one item of an opposite pair (critical word), the role of which was to render the meaning of the sentence either congruent or incongruent. Since the critical words were pairs of opposites (e.g. correct/wrong), the critical word that rendered the negated sentences congruent rendered the affirmative sentence incongruent and vice versa. The average number of letters in the critical words was M = 6.97 (SD = 2.24). Both adjective items and critical words were kept constant across the three sentence frames so that each participant would see two versions of the same sentence frame but still see all possible combinations of conditions for each adjective set (i.e. six possible combinations of adjectives and critical words). Three lists were created, consisting of 108 sentences each (2 conditions by 3 sentence frames by 18 adjective sets). The lists were pseudo-randomized, that is, the same sentence frames were separated by at least two other types of sentence frames. PsychoPy software program (Peirce et al., 2019) was used for the presentation of the stimuli.

Table 1. Examples of sentences used in the study.

Adjective set: authorised/unauthorised/not authorised
Sentence frame 1	The new Obama biography was ADJ therefore the details in the book were correct/wrong in actual fact
Sentence frame 2	Her access to the food storage was ADJ therefore opening it for her friends to get food was correct/wrong according to her
Sentence frame 3	The information that she disclosed was ADJ therefore the board agreed that her decision was correct/wrong in taking

Notes: ADJ indicates one of the three adjectives in the adjective set. The underlined words are the critical words that were the target words for the ERP analysis and were kept constant across the three sentence frames.

Participants

Twenty-eight right-handed native speakers of English were recruited to participate in the study. Two were removed due to technical errors leaving a total of 26 participants to be included in the final analysis (18 females; mean age = 29.9 years, age range = 20–41). All participants provided written consent (approved by the Swedish Ethical Review Board) prior to the experiment and received a cinema ticket as compensation for participation. None of the participants reported any history of neurological disorders and all had normal or corrected-to-normal vision.

Procedure

First, participants were informed about the experimental procedures by the experimenter after which they filled in the consent form and the electrode cap was placed on their heads. Next, participants were seated at a 110 cm distance from a monitor and were instructed to read the sentences silently and to answer the question “Did the sentence make sense, logically?”. The experimenter then explained to the participant that the focus of the task was on the coherent and consistent link between the two parts of the sentence in order to help them identify whether the sentence was congruent or incongruent. Each trial started with a fixation cross in the middle of the screen, followed by the sentence upon a button press. Sentences were presented word-by-word at the centre of the screen, with a 300 ms word duration and a 200 ms inter-stimulus interval. After the final word, three question marks prompted participants to respond either yes or no by pressing one of the two right-left buttons on the keyboard. The responses were registered with index fingers of both hands and the right-left response keys were counter-balanced across participants. Both accuracy and response times were collected in addition to the continuous EEG. Prior to the experimental trials, participants performed two practice trials. The task lasted about 45 min.

EEG recording and processing

The continuous electroencephalogram was recorded at a 500 Hz sampling rate from 30 scalp electrodes (NeuroScan Easycap). In addition, two electrodes were placed on the mastoids, two electrodes were placed at the canthi of the eyes to monitor horizontal eye-movements, and two electrodes were placed above and below the left eye to monitor vertical eye-movements. Impedance levels were maintained below 5 kΩ for scalp and mastoid electrodes and below 15 kΩ for facial electrodes. EEG channels were online referenced to the left mastoid and later at the data processing stage, all scalp electrodes were re-referenced to the average of the two mastoids. For the processing of the data, EEGLab (Delorme & Makeig, 2004) was used. Offline filters of 0.01 and 40 Hz were used. Initial artefact rejection was performed on continuous data prior to the low-pass filter, during which portions of the data with bursts of excessive muscle noise (EMG) were removed. Next, Independent Component Analysis (ICA; runica routine of EEGLab) was run on continuous data and the ocular artefact components were identified in topographical maps and component series and were subsequently removed. The data was then segmented into 1100 ms epochs with a baseline of 100 ms prior to the critical word for each electrode site and each participant. Final artefact rejection was performed on the epoched data where trials with any remaining artefacts, not removed by the ICA, were removed (5%). From a total of 2808 trials (108 trials by 26 participants), the number of remaining trials after artefact rejection was 2665 with 440–448 trials per condition.

Statistical analyses

The analyses for both behavioural and ERP data were performed in the software R (version 3.4.3; R core team, 2016) using mixed-effects modelling and the R-package “lme4” (Bates et al., 2015). In a step-wise approach similar to that used in Pérez et al. (2018), we created multiple models of varying complexity in the analysis of behavioural and ERP data separately. By keeping the maximal fixed structure, various random structures were defined including a restricted maximum likelihood estimation. The best random structure was identified using likelihood ratio tests excluding models that did not converge. Next, keeping the random structure fixed, a stepwise approach was used comparing two models at a time starting with the highest maximal fixed structure using likelihood ratio tests. Subsequent comparisons were performed if the interaction or main effects were not significant. After determining the model with the highest predictive accuracy, we ran an ANOVA test on the model to make sure that all remaining predictors were significant contributors.

In the analysis of response accuracy and response time, sentence type with three levels (affirmative, prefixal negation and sentential negation) and congruency with two levels (congruent and incongruent) were used as predictors, and participant and item (sentence) as random effects. The best model for response accuracy included sentence type and congruency as main effects, by-participant random intercepts and by-item random slopes for sentence type. The same fixed structure was identified as the best structure in the response time analysis with by-participant and by-item random intercepts.

The same predictors, that is, sentence type and congruency were used in the analysis of the ERP data. Participant and channel were also included as random effects. The ERP analyses focused on two ERP responses and hence, two time-windows: 300–400 ms for the N400, and 450–600 ms for later effects such as the P600. These windows were selected after visual inspection of the individual data and based on previous research suggesting that the N400 effect typically occurs between 200-600 ms (Kutas & Federmeier, 2011) and the P600 effect onsets around 500 ms post-critical word onset (Bornkessel-Schlesewsky & Schlesewsky, 2008). For the purpose of ease of presentation and understanding, the analyses of these ERP responses were conducted in three separate regions of interest defined as follows: anterior (F3/4, FZ, FC3/4, and FCZ), central (C3/4, CZ, CP3/4, and CPZ), and posterior (P3/4, PZ, O1/2, and OZ). These are the regions in which N400 and P600 effects are typically reported in studies of semantic incongruity (e.g. Bornkessel-Schlesewsky & Schlesewsky, 2008; Kolk et al., 2003; Kutas & Federmeier, 2011; van Herten et al., 2005). In all reported time-windows, that is, the time-windows where there was a significant difference between the conditions, the model with the highest predictive accuracy included an interaction between sentence type and congruency. Additionally, the best random effect structure in all analyses included by-participant slopes for sentence type and congruency and by-channel random intercepts. More details of all models used in the analyses of both behavioural and ERP data are presented in Appendix B.

Results

In reporting the results of the regression output from each model, β represents the difference between the specified condition in relation to the specified intercept, SE represents the standard error and z-scores and p values indicate the significance levels of the effects.

Accuracy rates and response times

The behavioural analyses revealed a higher processing cost for prefixal negation compared to the other two conditions (see Table 2). The output of the model with the highest predictive accuracy revealed both sentence type [χ²(2) = 5.85, p = .053] and congruency [χ²(1) = 8.86, p < .01] as significant predictors. The individual predictors were further examined and a lower accuracy rate was revealed for the prefixal negation compared to the affirmative form (β = −0.54, SE = 0.26, z = −2.09, p < .05) and to the sentential negation form (β = −0.44, SE = 0.19, z = −3.37, p < .05). In addition, lower accuracy rates were found for the incongruent conditions compared to the congruent ones across the three sentence types (β = −0.34, SE = 0.11, z = −3.00, p < .001).

Table 2. Accuracy rates and average response times.

Negation type	Proportion of accurate responses		Average response time in seconds
	Congruent	Incongruent	Congruent	Incongruent
Affirmative	0.87 (0.09)	0.80 (0.11)	1.69 (0.58)	1.94 (0.85)
Sentential negation	0.83 (0.11)	0.80 (0.09)	1.74 (0.78)	1.97 (0.86)
Prefixal negation	0.78 (0.08)	0.71 (0.12)	1.86 (0.81)	1.72 (0.75)

Note: Numbers in parentheses are standard deviations.

The response time analysis revealed a main effect of congruency [F(1, 2107) = 4.32, p < .05]. No significant differences were observed between any of the three sentence types. The congruency effect showed longer response times for incongruent than for congruent responses (β = 0.34, SE = 0.16, t = 2.07, p < .05).

ERP data

The ERP analyses revealed a biphasic N400–P600 response for the incongruities in affirmative sentences. For both negated conditions, incongruities resulted in larger negativities over two regions of interest, the central and anterior. In sententially negated sentences, a negativity was observed for the incongruent conditions, in the central region while in prefixally negated sentences, incongruities elicited a larger negativity in the anterior region (see Figure 1). The more detailed results are presented below, and three post-hoc pair-wise contrasts (general linear hypothesis tests) between the congruent and incongruent conditions in each sentence type over the three regions of interest and adjusted z-scores for each model are presented in Table 3.

Figure 1. ERP effects of incongruities in affirmative and negated sentences. The left panel indicates the grand averages for the congruent and incongruent conditions and the right panel indicates the topographic maps of the difference for the incongruent minus congruent conditions for affirmative sentences, sententially negated and prefixally negated sentences.

Table 3. Post-hoc contrasts of congruency in the three sentence-types.

	Anterior	Central	Posterior
N400 (300–400 ms)
Affirmative	NS	NS	β = –0.87 (0.30) z = –2.89**
Sentential Negation	NS	NS	NS
Prefixal Negation	β = –1.13 (0.31) z = –3.64***	NS	NS
P600 (450–600 ms)
Affirmative	β = 0.94 (0.34) z = 2.71*	β = 0.88 (0.33) z = 2.65*	NS
Sentential Negation	NS	β = –0.68 (0.33) z = –2.05^#	β = –0.63 (0.29) z = –2.15^#
Prefixal Negation	β = –1.00 (0.34) z = –2.87**	β = –0.86 (0.33) z = –2.60*	NS

Notes: Anterior, Central, and Posterior are the three regions of interest and include the sites F3/4, FZ, FC3/4, and FCZ for anterior, C3/4, CZ, CP3/4, and CPZ for central, and P3/4, PZ, O1/2, and OZ for posterior. β represents the estimated mean difference amplitude in microvolts between the congruent and incongruent conditions in the three sentence types (Affirmative, Sentential Negation and Prefixal Negation). Standard errors are presented in parentheses. NS signifies non-significant effects. Significant and marginally significant z scores are marked as follows: “^#”: .10 to .05; “*”: p < .05; “**”:p < .01; “***”: p < .001.

N400 time-window

In the analysis of the N400 time-window (300-400 ms) over the anterior region, a significant interaction between sentence type and congruency was observed [F(2, 826) = 10.79, p < .001]. The output of the model revealed that the interaction effect was significant between congruency and prefixal negation (β = –1.00, SE = 0.23, t = –4.20, p < .001). The pair-wise comparison from the model revealed a larger negativity for the incongruent than congruent condition in prefixally negated sentences (Table 3). There were no significant differences between any of the conditions over the central region in this time-window.

Over the posterior region, the best model in the analysis of the congruent and incongruent conditions for this time-window revealed a significant sentence type by congruency interaction [F(2, 826) = 5.57, p < .001]. The output of the model revealed larger negative amplitudes for the congruent condition in sentential negation (β = –0.90, SE = 0.28, t = –3.13, p < .001) and prefixal negation (β = –0.70, SE = 0.37, t = –1.09, p = .06) compared to the congruent affirmative condition. The pair-wise comparison of congruent and incongruent sentences revealed a larger negativity for the incongruent than the congruent conditions in affirmative sentences but not in either of the negated sentences (see Table 3).

P600 time-window

The analysis of the ERP effects in the P600 time-window (450–600 ms) over the anterior region revealed a significant sentence type by congruency interaction [F(2, 826) = 29.36, p < .001]. The output of the model revealed more negative amplitudes for the congruent affirmative condition than the congruent sententially negated (β = –0.83, SE = 0.41, t = –1.99, p = .055) and congruent prefixally negated (β = –0.86, SE = 0.41, t = –2.09, p < .05) conditions. The pair-wise comparisons between the congruent and incongruent conditions for each sentence type revealed that the incongruities elicited a larger negativity in prefixal negation while the reverse pattern was found in affirmative sentences in that a larger positivity was observed for the incongruent than congruent affirmative sentences.

Similarly, in the analysis of the P600 time-window over the central region, a significant sentence type by congruency interaction was observed [F(2, 826) = 25.10, p < .001]. The output of the model revealed a larger positivity for the incongruent than congruent condition in affirmative sentences (β = 0.88, SE = 0.33, t = 2.65, p < .05). The significant interaction effect between congruency and the two negation types revealed that the direction of the effect of congruency was different for sentential negation (β = –1.56, SE = 0.27, t = –5.77, p < .001) and prefixal negation (β = –1.74, SE = 0.27, t = –6.44, p < .001). The pair-wise comparisons of congruent and incongruent conditions in each sentence type revealed a larger positivity for the incongruent compared to the congruent condition in affirmative sentences, while this pattern was reversed for the two negated forms, that is, the incongruent conditions resulted in larger negativities compared to the congruent conditions (Table 3).

Over the posterior region in 450-600 ms, a significant interaction effect between sentence type and congruency was found [F(2, 826) = 5.60, p < .001], which revealed larger negative amplitudes for the incongruent than congruent conditions in the sententially negated sentences (Table 3).

Discussion

The findings from the previous studies on the processing of negation are not in agreement as to how incongruities in negated sentences are processed. According to the two-step model, negation is processed in two steps where it is ignored in the first step and is only integrated in the second step (Lüdtke et al., 2008). The other view on negation, the pragmatic view, states that negation can be integrated with no delay if used in an adequate context where it rejects a plausible statement (Nieuwland & Kuperberg, 2008). We returned to these findings using complex sentences and compared affirmative forms (authorised) to sententially negated forms (not authorised) as well as prefixally negated forms (unauthorised), the processing of which we know very little about. Sentences with and without incongruities were presented to participants and ERP responses to congruent and incongruent critical words as well as behavioural responses to sentences (accuracy and response time) were analysed. In this section, we discuss our results in relation to the previous literature and offer tentative explanations for the ERP patterns obtained. First, we discuss the results of our investigation for every sentence type: the affirmative sentences (the baseline condition), sentences with prefixal negation and those with sentential negation. Next, we compare the two negation types and discuss the differences and similarities between them. As no previous research comparing prefixal negation with sentential negation has been carried out before, we caution that our explanations are the first attempts at coming to grips with this issue.

ERP effects in affirmative and negated sentences

Our results for the affirmative form are in line with our expectations and previous findings, indicating a larger N400 for violations of semantic expectancy pointing to more effortful processing with incongruities (Fischler et al., 1983; Kutas & Hillyard, 1980). In previous ERP studies, the N400 has been used as an established index of the processing of semantic violations (see Kutas & Federmeier, 2011 for an overview). However, more recent studies have reported a P600 in sentences that include an incongruent semantic relation between the two arguments (Kolk et al., 2003; Kuperberg, 2007; van Herten et al., 2006; van Herten et al., 2005; but see Bornkessel-Schlesewsky & Schlesewsky, 2008 for another interpretation of the P600 to semantic incongruities). More specifically, in an incongruent sentence such as The fox that hunted the poachers [ … ] (Kolk et al., 2003), a P600 was observed where a reversal of the thematic roles (i.e. fox and poachers) could render an otherwise unacceptable reading of the sentence acceptable. The same result was reported by van Herten et al. (2005) when they presented semantic reversal anomalies in syntactically correct sentences such as The cat that fled from the mice [ … ]. Additionally, the likelihood of observing a P600 for semantic incongruities was argued to be higher in the presence of an acceptability judgment task (Kuperberg, 2007). The P600 in these studies is taken to reflect a monitoring process in which participants, upon encountering an anomaly, call the initial interpretation into question and perform a reprocessing of the sentence in order to check the veracity of the statement. In our study, we also observe a P600 effect as well as an N400 to incongruities of semantic nature in a forced-choice judgment task. We interpret the larger P600 to reflect semantic processes that are involved in the re-evaluation of the meaning of the whole sentence required for the judgment task following each sentence.

With regard to processing prefixal negation (un), the behavioural results indicate that participants experience more difficulty in processing these prefixed forms compared to the affirmative forms as indicated by the higher error rates. This difficulty is also reflected in the ERP pattern for the congruency effects in prefixally negated sentences compared to that in affirmative sentences. More specifically, instead of the N400 effect observed for affirmative sentences, a sustained anterior negativity was observed for incongruent compared to congruent sentences containing prefixal negation. This ERP pattern is not an established effect elicited by semantic incongruities. Sustained anterior negativities such as the one observed in this study have been discussed to reflect a higher cognitive load imposed on working memory resources (Martín-Loeches et al., 2005; Müller et al., 1997). Thus, the presence of the anterior negativity rather than an N400 suggests that the processing of incongruities in prefixally negated sentences is more difficult than the processing in affirmative sentences. One possible explanation for this is that when participants encounter the incongruent critical word in the prefixally negated sentences, they become uncertain about the information in the preceding sequence, that is, whether the adjective is negated or not (unauthorised or authorised). For this reason, participants have to perform a backward memory search to recall the negated item for the re-evaluation of the meaning of the sentence. This process is likely to be taxing for working memory. This processing difficulty of prefixally negated sentences is behaviourally supported by the higher error rates observed for these forms compared to the other two sentence types.

In sententially negated sentences, there are no differences in accuracy rates and response times compared to affirmative sentences, suggesting that negation has been integrated and incongruities have been processed. However, the qualitatively different ERP patterns between these two types of sentences suggest different types of processing. This intriguing dissociation between offline and online measures has been reported before (e.g. Andersson et al., 2019; Weber-Fox & Neville, 1996) and testifies to the importance of a combination of offline and online methods. More importantly in light of our results, similar dissociations have been observed in previous studies of negation, but in those cases, no effect of negation was found in ERP measures but only in the offline measures (e.g. Fischler et al., 1983; Lüdtke et al., 2008). While incongruities in affirmative sentences elicit a larger N400, we do not find this effect in response to incongruities in sententially negated sentences. With regard to the lack of an N400 effect for congruency, it can be argued that both congruent and incongruent negated sentences were complex and as such elicited equally large N400 responses. This was, in fact, statistically supported by the larger negativity for the congruent sententially negated condition compared to the congruent affirmative condition in the N400 time-window. Instead of the N400 effect, the incongruities in sententially negated sentences elicit a larger negative-going wave in the 450–600 ms time-window. Below we will present two possible explanations for this negativity with a later onset.

Firstly, the predictions of the two-step simulation model (Lüdtke et al., 2008) could explain this late negativity with sententially negated sentences. According to this model, negation is initially ignored and ERPs are primarily modulated by semantic priming effects, in which case, the simulation of the first step delays the integration and processing of negation (Dudschig et al., 2019; Ferguson et al., 2008; Lüdtke et al., 2008). It is only in the later stages of processing that an effect of negation – though not of congruency – is observed (Lüdtke et al., 2008). It could, therefore, be argued that in our sentences, negation is not fully integrated when participants encounter the incongruities and the integration of negation is then at an intermediary stage as indicated by a late negativity effect. However, if this is in fact the case, and if negation has been completely ignored in the first place, we should have observed a larger N400 for the congruent compared to the incongruent sentences because ignoring not in the negated sentences would result in the reverse pattern that is the pattern observed in the affirmative sentences. This effect, which is evidence for the first simulation step, has been reported in previous studies using stimuli that test negation both at an early stage (i.e. more immediate processing) (Dudschig et al., 2019; Fischler et al., 1983) and at a later stage (i.e. negation occurring in earlier context) (Ferguson et al., 2008; Lüdtke et al., 2008). However, we do not find any evidence for the first representation step in our study and the late negativity effect cannot be explained by the two-step model. Considering the finding that both congruent and incongruent conditions elicit a lager N400, this negativity could indicate a prolonged integration of negation and delayed detection of incongruities in sententially negated sentences.

The second explanation for the late onset negativity effect observed in the case of sentential negation involves the role of working memory capacity in processing negation, as has been discussed above for prefixal negation. If we consider this negativity to be a negative-going wave indicating working memory processes, we could argue that negation is in fact processed earlier in the sentence but is then subdued – though not completely lost – in working memory as more information is presented. By the time participants encounter the congruent critical word in sententially negated sentences (e.g. wrong in not authorised), they process and integrate it with some difficulty as indicated by the larger N400 compared to the congruent affirmative condition and the equally large N400 compared to the incongruent negated condition. Despite this difficulty, participants integrate this word and move on. However, when they encounter the incongruent critical word in sententially negated sentences (e.g. correct in not authorised), they also process it with greater difficulty as indicated by the large N400, but then possibly make an attempt to access the negated expression encountered earlier in order to re-analyze the content of the sentence. For this to happen, they possibly have to make use of memory resources to recall the negating element from the first part of the sentence. This process may be burdensome and result in an increased negativity. It is important to note that the negativity in sentential negation has a different distribution from that of prefixal negation. If we take both negativities to reflect working memory processes, we need to discuss the differences between the two and what might have caused these differences. Below we compare and discuss the ERP effects observed in sentential negation and prefixal negation in more detail.

Congruency effects in sentential negation and prefixal negation

In comparing the two negativities observed for prefixal negation and sentential negation, we find that the negativity is more anterior, more pronounced and more prolonged for prefixal negation than for sentential negation. Combined with higher error rates, this negativity suggests that prefixal negation is more difficult to process than sentential negation. Trying to explain this rather unexpected finding requires some speculation. Firstly, it may be argued that the frequencies of the two types of negation could explain the results, namely that the less frequent condition (unauthorised) could have resulted in larger negative amplitudes than the more frequent condition (authorised in not authorised). However, word frequencies for the prefixed and affirmative forms were selected from three different ranges, which means that both frequent and infrequent prefixally negated forms were included in the stimuli (see Materials and design and Appendix A). It might also be argued that the frequency of critical words could have elicited such results. However, the same critical words were used in the affirmative sentences (but in the reversed congruency conditions) and no such ERP patterns were observed in these sentences. Moreover, frequency effects have been shown to be overruled by higher-level contextual factors in previous ERP studies (van Petten, 1995; van Petten & Kutas, 1990). Based on this, the frequency explanation can be ruled out and instead, two other possible explanations can be put forward.

First, if the two negativities reflect working memory processes, the negativity for prefixal negation should be associated with a higher memory load than the negativity in sentential negation. One possible explanation for this enhanced negativity could be the way participants processed the two negated forms in the first part of the sentence. In the case of prefixal negation, when participants encountered the adjective unauthorised, they might not have processed it as a negated form, that is, un as a negating item operating on the meaning of the affirmative adjective authorised. Instead, they might have treated the meaning as a whole without decomposing it into a negating component plus an affirmative. Later on, when they encountered the incongruity, they had to recall the negated item in order to re-evaluate the meaning of the sentence. This recalling process was possibly costly as the adjective unauthorised was not processed as a negated item when it was first encountered. On the other hand, in the case of sentential negation, participants might have treated the not authorised as an explicitly negated item with not as a separate word operating on the meaning of authorised. Therefore, recalling the negator not from the earlier context was not as difficult as recalling the negator un.

The other explanation is based on the different types of interpretations that are evoked by prefixal negation and sentential negation. As discussed in the method section, previous research has shown that the interpretation of negated adjectives is sensitive to the boundedness property of adjectival meanings (Farshchi et al., 2019), and typically, negated bounded adjectival meanings (e.g. not dead) are interpreted to express the complete opposite (e.g. alive). For this reason, we made use of bounded adjectives in the creation of our stimuli in order to keep the interpretive differences between prefixally negated and sententially negated forms to a minimum (e.g. not intentional-unintentional). Now, previous work has also shown that even in the case of negated bounded adjectives, the interpretations can be modulated and laid out on a scale. For instance, not wrong does not necessarily refer to the opposite, that is, right but could, instead, get a scalar interpretation (Paradis, 2008). The effect of not in combination with scalar meanings has the effect of resulting in a reading of the negator as a softener of statements as in the food was not bad (Colston, 1999; Paradis & Willners, 2006) and has been used in a range of irony and metaphor-inducing contexts (Giora et al., 2005).

Keeping this function of negation in mind, the presence of a negating term such as not may not have the effect of reversing the scale associated with the negated concept (i.e. the opposite property). Therefore, the context created by not might not be as clear-cut and predictable as assumed, and the incongruities in these sentences may not have come across as clearly incongruent because of this. In other words, interpretations arising from a negated statement, such as not authorised are not necessarily the same as those arising from unauthorised. Instead, not has the effect of extending the meaning application possibilities of the statement resulting in an ameliorated incongruent context, which in turn gives rise to smaller ERP effects than prefixal negation. In contrast, the use of prefixally negated forms signals a clear meaning, which is to say on the opposite side of the conceptual boundary from the affirmative form. This may have the effect of boosting the incongruity, and hence eliciting a more enhanced anterior negativity.

As mentioned before, the enhanced anterior negativity together with lower accuracy rates suggests that the processing of prefixally negated forms was more difficult compared to the other two forms. We present a tentative discussion for this difficulty based on the assertiveness of prefixally negated forms compared to sententially negated forms. Sententially negated forms such as not authorised are non-assertive forms meaning that the negator not, through its non-assertiveness opens up an extended space of various possible interpretations within and potentially beyond the domain of authorisation. Prefixal negation, on the contrary, appears to function in the same way as assertive forms; its meaning is just a reversal of its counterpart authorised. In this sense, the prefixally negated opposite may be comparable with a lexical opposite, but a complication is incurred in processing the prefixally negated opposite by the added prefix un. From the point of use of how the information is presented in the sentences, it may be the case that the use of prefixally negated words as assertive forms, particularly at the beginning of each sentence without any forewarning, might have created an abrupt, harder to interpret and indeed very assertive contexts that could come across as surprising. Compare, for instance the book was unauthorised to the book was not authorised where the use of unauthorised might require more forewarning and grounding than not authorised. For this reason, participants may encounter a higher level of difficulty in processing these sentences, resulting in a higher working memory load as suggested by an enhanced anterior negativity and lower accuracy rates. As discussed earlier, it has previously been shown that the naturalness of negation use in certain discursive contexts can influence the ease of processing (Nieuwland & Kuperberg, 2008). The role that assertiveness plays in participants’ interpretations is an empirical question which can be tested by, for example, acceptability judgment tasks.

Yet another important difference between the patterns observed for processing sentences with prefixal negation and with sentential negation is the earlier onset of the negativity in prefixal negation (300–400 ms) compared to the one in sentential negation (450–600 ms). This indicates that it takes longer for participants to recall and integrate the negated term and react to the incongruity in the case of sentential negation, even though, eventually, this process is successful as indicated by the offline measurements. However, in the case of prefixal negation, the ERP patterns suggest that the incongruity is detected early, but that the recalling process is more taxing for working memory and integration might fail due to this processing difficulty as suggested by the higher error rates.

Summary of discussion

As predicted, for affirmative sentences, incongruities elicit an N400–P600 response, which we interpret as a reflection of difficulty with the integration of incongruent information leading to monitoring processes and a reanalysis of the message. For sentential negation, different ERP patterns are observed, which indicate that negated sentences are processed differently from their non-negated counterparts (i.e. affirmative sentences). The delay in the processing of incongruities with sententially negated sentences as indicated by the late negativity may have been caused by the prolonged integration of negation due to a memory search for the negator in an attempt to resolve the incongruity. Similarly, the more pronounced anterior negativity for prefixally negated sentences suggests a higher cognitive load for the processing of this form in comparison to the processing of affirmative forms and sententially negated forms. We argue that this greater processing difficulty for prefixal negation compared to that for sentential negation could be caused by differences in the interpretations of words such as unauthorised and not authorised, and the relative unnaturalness of prefixed forms in the kind of sentences used in the stimuli.

With regard to our predictions and those of the negation processing models, these results are not entirely in line with the two-step simulation model (Lüdtke et al., 2008) in that we do not observe an indication of the first simulation step, and hence, our results cannot be explained by this model. With respect to the pragmatic view of negation, our results do not correspond with those by Nieuwland and Kuperberg (2008) either, as we do not find a larger N400 for the incongruities in negated sentences. As mentioned when discussing the predictions, our stimuli sentences are not directly comparable with those in Nieuwland and Kuperberg (2008) as no naturalness and truth-value ratings were collected, which could explain the lack of an N400 effect of congruency. However, we align our findings with a more pragmatic and dynamic view of processing negation where discursive factors such as the complexity of the context, flexibility of interpretations and possibly memory retrieval mechanisms may all affect the processing of negated meanings in context.

With regard to prefixal negation, we show that these forms are not processed in the same way as affirmative forms. Instead, prefixal negation incurs a higher processing cost. Importantly, both behavioural and neurophysiological results for this type of negation suggest a higher processing cost than the processing cost of not only affirmative forms but also sententially negated forms.

With respect to the importance and contribution of our study, it deserves to be pointed out again that not all instances of negation are simple rejections of information but they also convey attitudinal aspects. Negation has been shown to express a number of different functions in text and discourse (Giora, 2006). Moreover, it has been found that in everyday conversation, the negator not is often used as a politeness item and a facilitator of meaning negotiation in conversations (Giora, 2006). Therefore, it is important to also investigate cases where negation is applied to meanings to fulfil functions other than rejection.

Conclusion

We investigated the processing of sentences with affirmative forms (authorised) and two negated forms (unauthorised and not authorised) in order to look for evidence in support of the two models on the processing of negation, namely the two-step model (Lüdtke et al., 2008) and the pragmatic view of negation (Nieuwland & Kuperberg, 2008). Our results cannot be fully explained by either of the views. We do not find any evidence for the two representation steps suggested by the two-step model (Lüdtke et al., 2008); nor do we find that incongruities in negated sentences are processed similarly to those in affirmative sentences, with no extra cost (Nieuwland & Kuperberg, 2008). Our study adds more complex pragmatic stimuli to the evaluation of the two models under discussion and our findings suggest that there is more to the processing of negation than we know today. Processing and integration of negation is complex and clearly different discursive factors contribute and interact in different ways to the problematics.

Finally, prefixal negation was included as we know little about these forms and whether they are processed as negated forms or affirmative forms. We find that the prefixally negated forms are not processed as affirmative forms but rather as negated forms that induce a higher processing cost. We hereby offer a contribution to our knowledge about the processing of negated sentences and to the nature of two types of negation by showing that these forms involve a higher processing cost for working memory.

Acknowledgements

The authors would like to thank Fredrik Heinat and Eva Klingvall for their assistance in data collection, the Humanities Lab for access to EEG equipment and Henrik Garde for his scripts used to process the EEG data. The authors are also grateful to the reviewers for their valuable comments on earlier versions of the manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).