Introduction

Can we truly experience ambivalent emotional states simultaneously? Think back to our daily lives, beyond the single emotional states of pure positivity and negativity, we seemingly also experience some bittersweet moments where laughter and tears intermingle1. For example, when we read the sentence, “What is critical for Jim must be hip-hop, because he is hypocritical.” this humorous statement intriguingly connects unrelated elements and can evoke amusement2,3,4. However, due to the use of negative keywords in the sentence, it can also induce a sense of negativity. Such co-existence of theoretically mutually exclusive emotional states constitutes a relatively complex emotional experience known as mixed feelings (i.e., ambivalence)5,6,7,8,9. This phenomenon is observable in daily life and across cultures10,11,12, but was largely absent from affective neuroscience research until it began to gain attention recently13,14,15. A core issue concerning mixed feelings is whether they represent a true blend of emotional states with opposing valences, and how the human brain dynamically processes these ambivalent emotional states.

Although existing behavioral results reported the presence of mixed feelings1,16, there remains debate on whether an actual mixture of opposing emotional states exists. Some researchers, holding a ‘bipolar’ perspective, argue that valence is an irreducible bipolar dimension of affect, and consequently precludes the simultaneous experience of positivity and negativity, the two opposite poles of the same dimension (e.g., circumplex model)17,18,19. In this situation, although we can observe the presence of mixed feelings behaviorally, the two ambivalent emotional states might be experienced in sequential, inverse, or prevalence patterns rather than being truly experienced simultaneously. Specifically, in the sequential pattern, one emotional state appears first and is then replaced by another, meaning that mixed feelings stem from swift switching between different emotional states. In the inverse pattern, as the intensity of one emotional state gradually decreases, the intensity of another progressively increases, meaning that they counteract each other until one replaces the other. In the prevalence pattern, one emotional state is dominant, while another is so minor that it can be considered a residual state, almost negligible20 (see Fig. 1A). In these experiential patterns, mixed feelings are not a genuinely mixed, simultaneously positive and negative experience.

Fig. 1: The theoretical schematic diagrams and the results of material pretest and behavioral experiment.
figure 1

A Diagrams of four different patterns of mixed feelings adapted from the mixed emotion patterns proposed by Oceja and Carrera (2009)20. B Results of the material evaluation before formal experiments (n = 40). (a) Ratings of comprehensibility and (b) ratings of unexpectedness for the three types of sentences. C Ratings of negativity implied by the words used in the experimental materials and the words used in fillers (n = 38). D Results of behavioral experiment (n = 42). (a) Reading time, (b) ratings of amusement, (c) ratings of negativity for the three types of sentences. Error bars indicate standard error. *p < 0.05; **p < 0.01; ***p < 0.001.

Full size image

On the other hand, some researchers, holding a ‘bivariate’ perspective, argue that the affective dimensions of positivity and negativity can be regarded as two separate variables, allowing for the simultaneous experience of both (e.g., evaluative space model)21,22,23,24. In this situation, besides these aforementioned patterns, two ambivalent emotional states might also be experienced in the highly simultaneous pattern: both emotional states are of moderate or high intensity and run a simultaneous course, either throughout the entire experiential episode or just a portion of it, indicating a period where the two emotional states are genuinely mixed20 (see Fig. 1A). In other words, while both the bipolar and bivariate perspectives recognize sequential, inverse, and prevalence patterns, only the bivariate perspective acknowledges the highly simultaneous pattern. Thus, if mixed feelings can be dynamically experienced in the highly simultaneous pattern, this can provide evidence for the bivariate perspective that amusement and negativity are not mutually exclusive, indicating that people can genuinely simultaneously experience both emotional states.

Furthermore, to understand how opposing emotional states are experienced dynamically, a crucial aspect is to investigate the time course of these constituent emotional states within mixed feelings, which is related to the cognitive processing mechanism of negative keywords in pun-humor. Previous pun-humor studies using non-negative keywords25,26,27 suggested that pun-humor comprehension involves accessing and integrating two plausible meanings of the keyword: its lexical salient meaning and another pragmatic meaning supported by the context, which can evoke amusement and require additional cognitive effort. When the keyword in a pun-humor sentence is negative (e.g., “What is critical for Jim must be hip-hop, because he is hypocritical”), it inherently carries a negative lexical salient meaning (“hypocritical”), and the humor is derived from the pragmatic meaning of the keyword under this context (“hip-hop critical”). This implies that accessing the lexical salient meaning of the negative keyword can trigger negativity, while finding or creating the unique connection between the keyword and context to access the pragmatic meaning can evoke amusement. However, although existing research suggests that pun-humor processing involves dual semantic information, the relative timing of processing the lexical negative meaning and the humorous pragmatic meaning remains unclear.

Two models of language processing offer distinct hypotheses regarding the time course of accessing lexical and pragmatic information. One is the parallel model of language processing28,29,30,31, which assumes that different kinds of linguistic representations, including lexical semantics and pragmatic information, can be processed synergistically at an early stage during online comprehension32. According to this model, representations of negativity and amusement are both rapidly and simultaneously activated, as the comprehension of pragmatic meaning related to amusement could occur alongside the lexical salient meaning related to negativity. The other is the serial/cascade model of language processing33,34,35,36, which suggests that different linguistic representations occur sequentially in a cascading manner37, with pragmatic information accessed at a late stage of language comprehension, after other linguistic analyses such as lexical semantic processes. According to this model, it can be speculated that the onset of negativity representation prompted by lexical salient meaning would occur before the onset of amusement representation prompted by pragmatic meaning.

Regarding the time course of the meaning access process during language comprehension, previous electroencephalogram (EEG) studies on pun-humor have focused mainly on traditional event-related potential (ERP) analysis38,39. For example, an ERP study on Chinese homophone puns40 found no N400 enhancement but an increased late positive component (LPC) in pun-humor relative to non-humor controls. The N400, typically observed with a central-parietal distribution, is a negative shift in ERP waveform peaking at approximately 400 ms after stimulus onset, indicating the ease or difficulty of semantic processing41,42,43. In contrast, the LPC, a positive-going waveform peaking between 600 and 800 ms after stimulus onset with a central posterior topography, typically reflects the cognitive effort underlying semantic integration of incoming information with preceding context, a process that may incorporate recollection or re-analysis of sentence-level semantics44,45,46. Thus, these ERP results suggest that while keywords retain contextual associations, integrating their dual meanings into pun-humor requires additional cognitive effort. However, this approach does not allow researchers to definitively determine which type of semantic meaning and corresponding emotional state (e.g., lexical negativity or pragmatic amusement) is being processed or represented in the ERP responses within a particular latency window, thereby complicating the determination of their relative processing times. Therefore, in the present study, representational similarity analysis (RSA) was applied to extract neural representations of specific emotional states from ERP responses. On the one hand, by calculating the correlation between behavioral patterns (or stimulus features) and neural patterns at various time points, RSA allows for tracking how neural patterns corresponding to the representation of specific emotional states change over time47,48,49. On the other hand, unlike traditional univariate analysis that compares ERP amplitudes by averaging responses across a number of items for different conditions at specific electrode sites, RSA is a multivariate analysis that simultaneously considers both the amplitude and the scalp distribution of brain waves, allowing it to access distributed information that is typically lost through averaging procedures50.

Taken together, the present study aims to investigate whether mixed feelings represent a true blend of emotional states with opposite valences and to examine the dynamic interplay of amusement and negativity over time. We conduct behavioral and EEG experiments and design three types of sentences that include the same negative keywords as stimuli (see Table 1): (i) pun-humor sentences, where the negative words in the critical sentence generate amused effects by connecting with their preceding context through phonetic similarity or polysemy, making the inherent negativity of the words intertwined with the amusement of pun-humor; (ii) non-humor sentences, where the negative words are semantically coherent and seamlessly align with their preceding context, making them a purely negative statement; and (iii) nonsensical sentences, where the negative words lack connection with their preceding context and can hardly be integrated. The behavioral experiment was conducted first and provided the basis for the EEG experiment, forming a progressive relationship. The behavioral experiment primarily aimed to explore whether the emotional linguistic materials we designed (i.e., pun-humor sentences with negative keywords) could effectively evoke mixed feelings at the explicit behavioral level. Building on these findings, the EEG experiment further delved into the cognitive processing mechanisms and dynamic neural representations underlying the observed behavioral phenomena by comparing ERPs and performing RSA. Following previous ERP studies40,51,52, we hypothesize that in pun-humor sentences, N400 amplitudes will be similar to those in non-humor sentences but lower than in nonsensical sentences, whereas LPC amplitudes will be higher than in both non-humor and nonsensical sentences. And by employing the RSA approach, we can directly analyze the dynamic experiential pattern of mixed feelings and examine the relative timing of amusement and negativity, which indirectly elucidates the dual-meaning access processes in online pun-humor comprehension.

Table 1 The exemplar and relevant connotations of three sentence types
Full size table

Results

Results of the material pretest

Figure 1B shows the results of sentence evaluation in the material pretest.

For the comprehensibility rating, a one-way repeated-measures analyses of variance (ANOVA) showed a significant main effect of sentence type (F [2, 78] = 256.65, p < 0.001, η2 = 0.87). Pairwise comparison with Bonferroni correction revealed that nonsensical sentences (M = 2.82, SD = 0.80) were significantly more difficult to comprehend than pun-humor sentences (M = 5.27, SD = 0.99, p < 0.001) and non-humor sentences (M = 5.89, SD = 0.59, p < 0.001). And pun-humor sentences were significantly less comprehensible than non-humor sentences (p < 0.001).

For the unexpectedness rating, a one-way repeated-measures ANOVA showed a significant main effect of sentence type (F [2, 78] = 194.77, p < 0.001, η2 = 0.83). Pairwise comparison with Bonferroni correction showed that, after reading the first parts, the negative content in the second parts was more unexpected in nonsensical sentences (M = 5.95, SD = 0.63) compared to pun-humor sentences (M = 5.26, SD = 0.74, p < 0.001) and non-humor sentences (M = 3.01, SD = 0.83, p < 0.001). And in pun-humor sentences, the negative content in the second parts was more unexpected than in non-humor sentences (p < 0.001).

Figure 1C shows the results of word negativity rating in the material pretest. Dependent sample t-tests showed that the keywords used in the experimental materials (M = 6.38, SD = 0.73) were perceived as significantly more negative than the keywords used in the fillers (M = 2.30, SD = 0.65, t(37) = 18.78, p < 0.001, Cohen’s d = 1.34). Besides, the arousal rating results (Supplementary Fig. 1) indicated that the negative words used in the experimental materials were rated in the medium-to-high arousal range (M = 6.21, SD = 0.68), supporting their effectiveness in evoking corresponding emotional states.

Results of the behavioral experiment

Figure 1D shows the average reading time, amusement rating, and negativity rating across the three sentence types in the behavioral experiment.

For reading time, a one-way repeated-measures ANOVA revealed a significant main effect of sentence type (F [2, 82] = 13.06, p < 0.001, η2 = 0.24). Pairwise comparison with Bonferroni correction revealed that participants spent significantly less time reading keywords in non-humor sentences (M = 548.35 ms, SD = 182.08 ms) than in both pun-humor (M = 630.43 ms, SD = 208.67 ms; p < 0.001) and nonsensical sentences (M = 629.14 ms, SD = 251.55 ms; p = 0.002). The difference in reading time between pun-humor and nonsensical sentences was not significant (p = 1.000).

For the amusement rating, a one-way repeated-measures ANOVA revealed a significant main effect of sentence type (F [2, 82] = 181.19, p < 0.001, η2 = 0.82). Pairwise comparison with Bonferroni correction revealed that pun-humor sentences (M = 5.06, SD = 0.71) were significantly more amusing than both non-humor (M = 2.96, SD = 0.65; p < 0.001) and nonsensical sentences (M = 3.23, SD = 0.89; p < 0.001), and nonsensical sentences were significantly more amusing than non-humor sentences (p = 0.048).

For the negativity rating, a one-way repeated-measures ANOVA revealed a significant main effect of sentence type (F [2, 82] = 55.76, p < 0.001, η2 = 0.58). Pairwise comparison with Bonferroni correction revealed that pun-humor sentences (M = 5.36, SD = 0.91) were significantly more negative than both non-humor (M = 4.39, SD = 1.11; p < 0.001) and nonsensical sentences (M = 3.89, SD = 0.85; p < 0.001), and non-humor sentences were significantly more negative than nonsensical sentences (p < 0.001).

Behavioral results of the EEG experiment

Figure 2B shows the average amusement rating and negativity rating across the three sentence types in the EEG experiment.

Fig. 2: Behavioral and ERP results in EEG experiment (n = 36).
figure 2

A The distribution of 9 regions of interest (ROIs) in our study. B The rating results in the EEG experiment. (a) Ratings of amusement and (b) ratings of negativity for the three sentence types. C ERP results of 9 ROIs for the three sentence types. D Topographic maps of N400 and LPC components. *p < 0.05; **p < 0.01; ***p < 0.001.

Full size image

For the amusement rating, a one-way repeated-measures ANOVA revealed a significant main effect of sentence type (F [2, 70] = 162.59, p < 0.001, η2 = 0.82). Pairwise comparison with Bonferroni correction revealed that pun-humor sentences (M = 6.90, SD = 1.29) were significantly more amusing than both non-humor (M = 3.56, SD = 1.05; p < 0.001) and nonsensical (M = 4.16, SD = 1.17; p < 0.001) sentences, and nonsensical sentences were significantly more amusing than non-humor sentences (p < 0.001).

For the negativity rating, a one-way repeated-measures ANOVA revealed a significant main effect of sentence type (F [2, 70] = 24.43, p < 0.001, η2 = 0.41). Pairwise comparison with Bonferroni correction revealed that pun-humor sentences (M = 7.07, SD = 0.91) were significantly more negative than both non-humor (M = 6.35, SD = 0.99; p < 0.001) and nonsensical (M = 6.06, SD = 0.89; p < 0.001) sentences, and non-humor sentences were significantly more negative than nonsensical sentences (p = 0.022).

ERP results

Figure 2C displays the grand-averaged ERP waveforms of three sentence types in nine ROIs, and Fig. 2D shows the topographic maps of the N400 and LPC components.

The N400 component (300–500 ms): The mean amplitudes during 300–500 ms of each ROI were computed and used as the dependent variables. Likelihood-ratio tests indicated that the main effects of Sentence Type \(({x}^{2}\left(2\right)=19.42,{p} < 0.001)\) and ROI \(({x}^{2}\left(8\right)=381.40,{p} < 0.001)\) were significant. And the interaction between Sentence Type and ROI (\({x}^{2}\left(16\right)=28.71,{p}=0.026\)) provided a better fit for the data than a model without it. The Bonferroni-corrected pairwise comparisons revealed that compared with nonsensical sentences, negative words in pun-humor sentences elicited significantly smaller N400 components in all ROIs except for the left frontal ROI (ps < 0.05). Moreover, N400 components elicited by negative words in pun-humor sentences and non-humor sentences were comparable across all ROIs, except for a marginally significantly smaller N400 component in the medial parietal ROI for pun-humor sentences (p = 0.068) (see Table 2 for more details).

Table 2 Pairwise comparisons between pun-humor and each of the other two sentence types by ROI in two time windows
Full size table

The LPC component (600–800 ms): The mean amplitudes during 600–800 ms of each ROI were computed and used as the dependent variables. Likelihood-ratio tests indicated that the main effects of Sentence Type (\({x}^{2}\left(2\right)=45.16,{p} < 0.001\)) and ROI (\({x}^{2}\left(8\right)=392.21,{p} < 0.001\)) were significant. And the interaction between Sentence Type and ROI (\({x}^{2}\left(16\right)=40.98,{p} < 0.001\)) provided a better fit for the data than a model without it. The Bonferroni-corrected pairwise comparisons revealed that compared with nonsensical sentences, negative words in pun-humor sentences elicited significantly greater LPC amplitudes in all ROIs (ps < 0.001). Moreover, compared with non-humor sentences, negative words in pun-humor sentences elicited significantly greater LPC amplitudes in medial central (p = 0.028), medial parietal (p = 0.014), and right central (p = 0.046) ROIs, while marginally significantly greater in the right parietal ROI (p = 0.064) (see Table 2 for more details).

RSA results

Regarding the relative timing of representing amusement and negativity in pun-humor sentences, we compared the onset, peak, and duration latencies of the time clusters during which the behavioral patterns of amusement or negativity showed significant partial correlations with the neural patterns. Figure 3A shows the representational dissimilarity matrices (RDMs) constructed from behavioral data in pun-humor sentences, and Fig. 3B shows the RDMs constructed from EEG data at each time point in pun-humor sentences. The RSA results revealed significant partial correlations between the EEG RDMs and behavioral RDMs, indicating the time course of the representation for negativity (Fig. 3C) and amusement (Fig. 3D) in pun-humor sentences. The representation of negativity began at 230 ms and ended at 670 ms after keyword onset, which specifically comprised three significant clusters: the first spanning from 230 to 370 ms (average p = 0.008), the second occurring from 400 ms to 540 ms (average p = 0.012), and the third ranging from 610 to 670 ms (average p = 0.027). And the representation of amusement had only one significant cluster, which began at 430 ms and continued until 520 ms (average p = 0.025). That is, the representation of negativity lasts 340 ms, which is longer than the 90 ms duration for the representation of amusement. In addition, the representation for negativity peaked at ~280 ms, and the representation of amusement peaked at ~510 ms. For direct comparison, we plotted the intervals of significant time clusters for partial correlations between EEG RDMs and behavioral RDMs for negativity and amusement on a single graph (Fig. 3E(a)), and averaged the partial correlation values within the significant time clusters to more intuitively reveal the experiential pattern of mixed feelings (Fig. 3E(b)). Furthermore, the Bootstrap test showed that the representation of negativity occurred significantly earlier (Fig. 3F(a), p = 0.023), peaked significantly earlier (Fig. 3F(b), p = 0.018), and lasted significantly longer (Fig. 3F(c), p = 0.007) than amusement in pun-humor sentences.

Fig. 3: RDMs and RSA results.
figure 3

A The exemplar behavioral RDMs for negativity and amusement. B The exemplar RDMs for EEG data at each time point. C–E RSA results (n = 36), in which time zero indicates the onset of the keyword presentation. C Time course of partial correlations between EEG RDMs and the behavioral RDM for negativity in pun-humor sentences. D Time course of partial correlations between EEG RDMs and the behavioral RDM for amusement in pun-humor sentences. E (a) Plot the intervals of significant time clusters for partial correlations between EEG RDMs and behavioral RDM for negativity and amusement; (b) average the partial correlation values within the significant time clusters to more intuitively reveal the experiential pattern of mixed feelings. F (a) Onset, (b) peak, and (c) duration latencies for decoding negativity and amusement in pun-humor sentences. Considering the large sample size in the bootstrap test, error bars indicate standard deviation. *p < 0.05; **p < 0.01; ***p < 0.001.

Full size image

RSA was also conducted on other sentence types to draw the time courses of amusement (Supplementary Fig. 2A) and negativity (Supplementary Fig. 2B). As significant partial correlations for negativity representation were observed among the three sentence types, we compared the onset, peak, and duration of the time clusters where behavioral patterns of negativity showed significant partial correlations with neural patterns. This was done to explore the timing of negativity representation across the three sentence types. Specifically, the representation of negativity in non-humor sentences began at 300 ms and ended at 400 ms (average p = 0.015), and the representation of negativity in nonsensical sentences comprised two significant clusters: one was 220–260 ms (average p = 0.009), and the other was 490–520 ms (average p = 0.021). Furthermore, the two-sided bootstrap test with FDR correction indicated that the onset of representing negativity (Supplementary Fig. 2C(a)) in non-humor sentences was significantly delayed relative to that in pun-humor sentences (p = 0.039) and nonsensical sentences (p = 0.016), while the onset in pun-humor sentences was comparable to that in nonsensical sentences (p = 0.059). Additionally, the duration of representing negativity (Supplementary Fig. 2C(b)) in nonsensical sentences was shorter than that in non-humor sentences (p = 0.026), with pun-humor sentences exhibiting the longest duration (ps = 0.023) among all sentence types.

Discussion

This study aimed to examine whether mixed feelings involve an actual mixture of ambivalent emotional states and how these emotional states with opposing valences are experienced dynamically. Behavioral and EEG experiments were conducted, and three types of sentences (pun-humor, non-humor, and nonsensical) were used as stimuli. The behavioral ratings in the two experiments showed that pun-humor sentences consistently received higher negativity and amusement scores than non-humor and nonsensical sentences, suggesting that the use of negative keywords in pun-humor sentences can induce mixed feelings at the explicit behavioral level. To examine the temporal dynamics of constituent emotional states within mixed feelings, RSA was used to extract neural representations of amusement and negativity during the online processing of negative keywords in pun-humor sentences. The results showed that the dynamic representation of mixed feelings aligned with the highly simultaneous pattern, where both negativity and amusement were simultaneously represented within a period, suggesting a true blend of conflicting emotional states. Moreover, the RSA results showed that the representation of negativity occurred and peaked significantly earlier than amusement, providing evidence for the serial/cascade model of language processing from a distinct perspective on emotion processing.

The finding that pun-humor can effectively elicit amusement is consistent with findings in previous studies52,53,54,55, which have been commonly regarded as the main reason why humor is effective in regulating negative emotions56,57,58. However, our findings revealed that pun-humor fails to alleviate and even tends to amplify negativity, challenging the common assumption that amusement would counteract negativity. Differences in how emotions are experienced in the current versus previous experiments might explain this distinct outcome. In most previous studies59,60,61, the experiences of amusement and negativity were divided due to the separate presentation of humorous and negative materials. Specifically, after initial exposure to negative stimuli, humorous materials emerged as a post hoc regulation method for offline processing of negative stimuli62,63, leading to potential distraction and reduction of the preceding negativity64,65. However, in our study, when online processing negative stimuli through pun-humor, both amusement and negativity can be evoked within the same stimuli. Unlike the experience of single feelings where amusement can mask negativity, the experience of mixed feelings encompasses the conflicting nature of ambivalence66,67, potentially leading to discomfort and heightened negative arousal68,69,70.

The cognitive mechanism underlying the online processing of pun-humor was examined by comparing the N400 and LPC components. Our results revealed comparable N400 amplitudes but enhanced LPC responses in the pun-humor condition compared to the non-humor condition. This pattern of N400 findings aligns with previous studies38,40, which have shown that the keywords in pun-humor sentences can be readily integrated with preceding context through homophony or polysemy, resulting in no significant N400 effect. Given that the N400 component is widely associated with the ease of semantic processing41,42,43, these results suggest that the initial stages of semantic processing in pun-humor are relatively easy and fluent. In contrast, the LPC, a positive-going ERP component with a centro-posterior scalp distribution, has been considered to reflect the additional integration efforts required to reconcile current information with existing semantic representations44,71. This component is thought to indicate the recruitment of extra cognitive resources for meaning updating and discourse-level integration. These functions are supported by the involvement of parietal regions, including the angular gyrus72,73,74 and inferior parietal lobule75,76,77. Consistent with this functional profile, the enhanced LPC observed in the pun-humor condition of our study indicates increased cognitive engagement for more extensive processing or reanalysis of the keyword44,45,46. Specifically, this may involve the activation and integration of two types of information associated with the negative keyword in the pun-humor sentence: its lexical semantic content, and the humorous pragmatic information derived from contextual reinterpretation, leading to more extensive meaning processing captured by the LPC component.

We used RSA to track and analyze the dynamic interplay between negativity associated with lexical semantic information and amusement associated with humorous pragmatic information during real-time language comprehension. It can be observed that the representation of amusement is brief, and within this short duration, the representation of negativity does not disappear but rather persists over a longer time frame. Importantly, during the processing of negative words in pun-humor, participants experience mixed feelings characterized by the genuine coexistence of both strongly elicited negativity and amusement, rather than the intuitive ‘zero-sum’ process of amusement offsetting negativity or rapid vacillation between opposing valence states. Notably, when we pooled all the conditional data together, the representations of negativity and amusement still satisfied the highly simultaneous pattern (the specific analysis procedures and results are detailed in Supplementary Fig. 3). This observation aligns with a recent fMRI study on the neural mechanisms of ambivalent affective states15, which suggested that mixed feelings are distinct from univalent emotional states and their neurobiological features cannot be attributed merely to switching back and forth between positive and negative states. Thus, the true existence of mixed feelings suggests that emotion categories should not be viewed solely in a black-or-white thinking manner as either positive or negative, as this perspective would imply that amusement and negativity cannot occur simultaneously. This finding provides crucial evidence that the experience of mixed feelings aligns more closely with the highly simultaneous pattern proposed by Oceja and Carrera20.

Moreover, in pun-humor sentences, the relative timing of the onset of specific emotional states is linked to different language elements, highlighting not only the processing of mixed feelings but also the mechanisms of language processing. In this situation, the neural representation of negativity is associated with the lexical semantics processing of the negative keywords, while the neural representation of amusement is related to the humorous pragmatic meanings derived from connecting these negative keywords with their context. By examining the temporal order of the neural representations of negativity and amusement, we can investigate whether accessing humorous pragmatic meaning is a slow or rapid process. Specifically, negativity representation begins at approximately 230 ms. In prior psycholinguistic studies30,31, it has been discovered that we can access lexical semantics almost “within an instant” (around 200 ms), a process that occurs quickly. When encountering negative keywords in pun-humor sentences, their inherent negativity immediately evokes a negative response in receivers, making it conceivable that the representation of negativity commences as soon as their lexical semantic information is accessed. Moreover, the onset of amusement is around 430 ms, significantly lagging behind negativity, indicating that access to pragmatic information occurs in a later stage than lexical semantic information, rather than both being processed synergistically at an early stage during online comprehension. When RSA was performed on data collapsed across all experimental conditions, we obtained consistent results that further validated the neural representational patterns associated with negativity and amusement (Supplementary Fig. 3). This provides supporting evidence for the important role of sequential processing mechanisms in language processing, at least in some circumstances, particularly from the perspective of emotional processing. More importantly, the representation of negativity extends even into the time window of LPC. Thus, the late negativity representation phase observed in the RSA can be attributed to the sustained cognitive engagement required for a deeper level of semantic and emotional processing, indicating a more comprehensive and detailed analysis of negative keywords in pun-humor sentences during the late processing stage.

Additionally, beyond the mixture of amusement and negativity and the temporal order of their neural representations, a more general question emerges: How does the human brain derive emotional meanings from sensory inputs? According to the Theory of Constructed Emotion78, the human brain continually constructs concepts and categorizes sensory inputs based on internal models developed from past experience and contextual information. This process enables the human brain to infer causal explanations and assign meanings, including emotional meanings, to these inputs. Crucially, different internal models can lead to the construction of distinct emotion concepts, even when individuals are exposed to identical stimuli, resulting in varied emotional categorizations. Our study provides neural evidence for this constructivist perspective through the following observations: (i) Context-dependent construction: Identical negative words elicited significantly divergent emotional impacts and neural representations across different contexts (Supplementary Fig. 2), demonstrating how internal models transform sensory inputs into distinct concepts. (ii) Conceptual categorization: Neural representations of sentence types emerged at ~410 ms (Supplementary Fig. 4) and persisted throughout the analysis window, reflecting sustained conceptual differentiation between categories. Collectively, these temporal signatures demonstrate how the brain constructs context-bound emotion concepts.

Limitations and future directions should be discussed. First, although the findings of this study revealed that pun-humor embedded with negative keywords is an effective method for exploring the intriguing topic of mixed feelings, future research could consider employing other non-linguistic types of experimental materials (e.g., ironic images or auditory sarcastic tones) to verify these effects. Second, it is worth noting that word-by-word presentation, while offering precise experimental control, sacrifices some ecological validity, as it differs from natural reading conditions. Future studies could combine EEG with eye-tracking in a natural reading paradigm, which would allow for a more ecologically valid investigation of real-time processing of negative words in pun-humor. Third, given that individual and cultural variations in humor perception—especially for negative humor—potentially modulate emotional experiences, future research should prioritize multi-site cross-cultural investigations with sufficiently powered samples to systematically examine how these psychosocial factors interact with mixed feelings. Finally, although humor is commonly regarded as a complex higher-order emotional process79,80,81, it is generally believed to have primarily positive implications82,83,84. Therefore, the intensified negativity observed in this study is crucial for highlighting the potential adverse effects of employing humor inappropriately to cope with negative stimuli, and further investigation is necessary to facilitate a more comprehensive understanding of how humor influences the processing of negative stimuli and to raise awareness of its potential negative effects.

In summary, the current study examined the real-time processing of negative words in pun-humor and the dynamic representation of mixed feelings blending amusement and negativity. The results showed that pun-humor containing negative keywords effectively triggered heightened amusement and negativity compared to non-humor and nonsensical sentences, indicating the elicitation of mixed feelings at an explicit behavioral level. Further exploration of this elaborate emotional processing revealed that the neural representation of negativity persists for a longer duration and occurs significantly earlier than that of amusement. The key finding is a period where negativity and amusement co-occur simultaneously, providing evidence for the genuine existence of mixed feelings. This reveals that we can truly experience ambivalent emotional states at the same time and that the dynamic representation of mixed feelings aligns with the highly simultaneous pattern.

Methods

Participants

The current study employed a single-factor, three-level within-subjects design. A priori power analysis using G*Power 3.1.9.785 with the number of groups = 1, measurements = 3, α = 0.05, power = 0.8, and effect size f = 0.25 indicated a minimum requirement of 28 participants. This study finally included the following participants:

(i) Material pretest: 40 participants (19 females, age range 18–26, M = 22.93 years, SD = 2.60 years) for the sentence evaluation, and 38 participants (21 females, age range 19–27 years, M = 23.00 years, SD = 1.26 years) for the word negativity rating.

(ii) Behavioral experiment: 42 participants (23 females, age range 18–26 years, M = 23.26 years, SD = 2.46 years).

(iii) EEG experiment: 39 right-handed participants were initially recruited. After excluding 3 participants due to excessive noise and artifacts in their EEG data, analyses were conducted on the final sample of 36 participants (18 females, age range 19–25 years, M = 21.83 years, SD = 1.79 years), which was larger than many previous similar EEG experiments on humor processing39,40 or emotion processing50,86,87.

All participants had normal or corrected-to-normal vision and were native Chinese speakers. They also reported no use of psychoactive medication, no history of neurological disorders or mental illness, and could tolerate negative words. They were reminded of their right to withdraw at any time during the experiment, provided written informed consent, and received compensation for their participation. Importantly, none of the participants had been exposed to the experimental materials before these experiments, to ensure that each participant was exposed only once. That is, no one underwent both the behavioral and EEG experiments; we recruited separate groups of participants for the different experiments. This approach helps avoid potential carryover or habituation effects that might arise from repeated exposure to the same linguistic materials88,89,90, which may change the corresponding emotional responses and cognitive processing patterns. This experimental protocol was approved by the Ethics Committee of Beijing Language and Culture University. All ethical regulations relevant to human research participants were followed.

Materials

At first, 180 two-character high-frequency Chinese words with evident negative connotations were selected from the Six Semantic Dimension Database (SSDD)91 and the HowNet sentiment lexicon92 for drafting the experimental materials, along with 120 words having positive or neutral connotations to create the fillers. Two highly skilled writers were recruited to create 180 pun-humor sentences. Each sentence included one of the selected negative words as its pun word, strategically positioned just before the final word to accommodate the EEG experimental paradigm. The pun-humor sentences were divided into two parts by a comma: the first part sets up a context by utilizing the same sound or orthographic form as the pun word, allowing participants to generate humorous effects when encountering the pun word, and the second part uses the pun word to make a negative evaluation of the protagonist in the sentence. Then, 20 evaluators (11 females, age range 19–26 years, M = 22.00 years, SD = 2.66 years) rated the amusement of these pun-humor sentences on a 7-point Likert scale (1 for very unamused and 7 for very amused) to select the most amusing 150 pun-humor sentences (M = 5.21, SD = 0.61). The 150 negative words in them were subsequently used as keywords to create non-humor and nonsensical sentences, both of which were of equal length and similar complexity to pun-humor sentences.

Consequently, a total of 150 sets of Chinese sentences were created as experimental materials, and each set included three types of sentences: pun-humor, non-humor, and nonsensical. These sentences were all 20 Chinese characters long, with a 10-character-long first part and a 10-character-long second part. Within each set, three sentences had completely identical second parts, namely, they shared the same keyword. The only variation occurred in the first parts, leading to different effects: in pun-humor sentences, combining the first and second parts created humorous effects; in non-humor sentences, the first and second parts maintained semantic coherence; and in nonsensical sentences, understanding the combination of the first and second parts was challenging because the first parts were created by altering just one Chinese character or word from those in pun-humor sentences, which disrupted the connection between the first and second parts. Examples of the three sentence types of two sets are shown in Table 1. In addition, to ensure these experimental materials were well designed, prior to the formal study, an independent group of participants evaluated all experimental materials using 7-point Likert scales for: (i) comprehensibility (1 = very difficult to understand the entire sentence, 7 = very easy to understand) and (ii) unexpectedness (1 = not unexpected at all, 7 = very unexpected regarding the shift to negative content in the second clause).

Moreover, since all the experimental materials used negative words as keywords, to mitigate the potential impact of sustained negativity, 120 selected words with positive or neutral connotations were used to create 40 pun-humor, 40 non-humor, and 40 nonsensical sentences as fillers. A Latin square design was employed to create three counterbalanced lists based on 150 sets of sentences, each supplemented with 120 fillers. Thus, a participant read 270 sentences in this study (150 experimental sentences with negative keywords and 120 fillers with positive or neutral keywords). To verify that the emotional valence of the keywords contained in both the experimental materials and fillers met the requirements of this study, the valence of these 270 words was rated by another separate group of participants on a 9-point Likert scale (1 for very positive and 9 for very negative). Additionally, the arousal levels of the 150 words used in experimental materials were also assessed.

Paradigm

The entire experimental procedure had 2 stages: the behavioral experiment and the EEG experiment.

During the behavioral experiment, participants read and comprehended the sentences in one counterbalanced list via the self-paced reading method (SPR, reading at a natural speed to understand its meaning) and rated the amusement and negativity of each sentence. In each trial, the first part of the sentence was presented in its entirety. After reading it, participants pressed the space key to move on to the second part. The second part was presented word-by-word, and participants pressed the space key to move on to the next word after reading each word, until the sentence-ending punctuation mark, a period (。), was presented. Using this method, we collected and compared reading time for the keywords. Participants rated the amusement (1 for very unamused and 7 for very amused) and negativity (1 for not negative at all and 7 for very negative) they felt after reading the entire sentence on 7-point Likert scales.

During the EEG experiment, the sentences in one counterbalanced list were pseudorandomly distributed into 6 blocks, each containing 45 sentences. The rapid sequential visual presentation (RSVP) approach (Supplementary Fig. 5) was employed to precisely time-lock the onset of keywords and minimize saccades or eye blinks during data acquisition. This approach, which is commonly used in the field of neurolinguistics, is well-suited for investigating the temporal dynamics of language comprehension93,94,95,96, enabling us to examine the cognitive processes elicited by specific words at critical positions within each sentence. Specifically, a fixation sign “+” appeared for 500 ms to indicate the beginning of each trial. After a 500 ms blank screen, the first part of the sentence was presented until participants pressed the space key. Then, the second part of the sentence was presented word-by-word at the center of the screen, with each word appearing for 500 ms and a 300 ms blank screen interval between words. After each sentence was presented, the screen separately displayed two 9-point Likert scales, allowing participants to rate the amusement (1 for very unamused and 9 for very amused) and negativity (1 for not negative at all and 9 for very negative) of the sentence. Moreover, to control for potential order-related confounding effects and reduce participant fatigue potentially arising from task monotony, both within-subject and between-subject balance was implemented for the rating order. That is, half of the participants rated amusement followed by negativity, while the other half rated negativity followed by amusement in the first three blocks, and these orders were reversed in the last three blocks. Supplementary analyses confirmed consistent result patterns between the first and last halves of the data, as well as under different rating orders (Supplementary Tables 1–6 and Supplementary Figs. 6–11), demonstrating that the rating order did not systematically affect the experimental results.

EEG recording and preprocessing

Continuous EEG signals were recorded using a flexible elastic cap with 64 Ag/AgCl electrodes, according to the international 10–20 system. Additionally, four electrodes (VEOU, VEOL, HEOL, and HEOR) were used to record electrooculogram (EOG) signals, which included eyeblinks, horizontal and vertical eye movements. EEG and EOG data were recorded by an AC amplifier (Synamps, Neuroscan Inc.) with a band-pass filter of 0.05–100 Hz and a sampling rate of 500 Hz. The impedance levels of all electrodes were maintained below 5 kΩ throughout the experiment. During online recording, the EEG and EOG signals were referenced to the left mastoid electrode. In the offline analysis, the EEG data were re-referenced to the average of left and right mastoid electrodes.

The EEG signals were preprocessed with the EEGLAB toolbox97 (version 14.1.1) for MATLAB (The MathWorks) in the following steps. After importing into EEGLAB, all the signals were filtered with a high-pass cutoff point of 0.1 Hz and a low-pass cutoff point of 30 Hz to remove high-frequency noise and slow voltage drifts. Previous studies have shown that using filter settings in this range does not produce any distortions in RSA analysis98,99,100. Then, the signals for each trial were segmented into 1000 ms periods (200 ms before and 800 ms after keyword onset) and baseline-corrected through subtraction of the mean voltage during the pre-stimulus interval [−200, 0] ms relative to keyword onset. The differences during the baseline interval were rigorously analyzed (Supplementary Fig. 12) in this study to ensure that baseline selection did not contribute to the specific profile of subsequent ERP effects. The results revealed that no significant baseline differences existed among the three sentence types (Supplementary Table 7), suggesting that the selected baseline interval is unlikely to introduce bias affecting ERP response variations. Subsequently, portions of EEG epochs that contained large muscle artifacts or extreme voltage offsets were removed. And the data underwent independent component analysis (ICA) to identify and remove components that were associated with eye blink and movement. The criterion for removing an ICA component involved evaluating the consistency between its shape, timing, and spatial location compared with the HEOG and VEOG signals. Additionally, artefacts with amplitudes exceeding ± 75 μV were removed from further analyses. On average, the remaining epochs were 139 trials per participant, and the number of remaining epochs among the pun-humor (M = 46.33, SD = 3.14), non-humor (M = 46.31, SD = 3.12), and nonsensical (M = 46.63, SD = 3.11) sentences showed no significant differences (ps = 1.000).

ERP analysis

To investigate the association between negative keywords and contexts, as well as the integration processing of negative keywords across three sentence types, the N400 and LPC components were compared in the ERP analysis. The most classic time windows for N400 and LPC were selected, which were 300–500 and 600–800 ms, respectively101,102. For each time window, single-trial analysis was performed on the EEG data via linear mixed effect models (LMEs) with the lme4 package103 in R. Nine classical ROIs were set up to examine the scalp distribution, each containing five or six electrodes (Fig. 2A): left frontal (F3, F5, F7, FC3, FC5, and FT7), left central (C3, C5, CP3, CP5, and TP7), left parietal (P3, P5, P7, PO5, PO7, and O1), medial frontal (F1, FZ, F2, FC1, FCZ, and FC2), medial central (C1, CZ, C2, CP1, CPZ, and CP2), medial parietal (P1, PZ, P2, PO3, POZ, and PO4), right frontal (F4, F6, F8, FC4, FC6, and FT8), right central (C4, C6, CP4, CP6, and TP8), and right parietal (P4, P6, P8, PO6, PO8, and O2). The mean amplitudes of each ROI during the two time windows were computed and used as the dependent variables.

During the analysis, sentence type and ROI were defined as fixed effects. The models were started with random intercepts for subjects and items, as well as by-subject and by-item random slopes for the effects of sentence type and ROI. To achieve model convergence, the final models were constructed using the following formula: DV ∼ Type * ROI + (1 + Type | Subject) + (1 | Item)40. The significance of the predictors and their interactions was computed using the mixed function from the afex package104, and post hoc pairwise comparisons were performed via Bonferroni correction using the emmeans function from the emmeans package105.

RSA analysis

To explore the relative timing of representing amusement and negativity during the online processing of negative words embedded in pun-humor sentences, we conducted RSA analysis on these sentences using the NeuroRA106 Toolbox in Python, following the analysis steps of previous researchers37. First, RDMs were separately constructed from behavioral data and EEG data. For behavioral RDMs, all the behavioral data were normalized to a Z score, and then the Euclidean distance between any two trials was calculated according to the negativity or amusement ratings of each participant. To construct the EEG RDMs, we extracted data vectors for each trial and participant using a 10 ms sliding window, encompassing 5 time points. Each vector captured the spatial pattern of neural activity centered on the middle time point of the sliding window, representing the response amplitude across all 60 scalp electrodes. Importantly, this approach did not rely on electrode averaging or predefined ROIs. Instead, each data vector (EEG amplitude × electrode location) preserved the full spatial distribution, reflecting the individual voltage amplitudes at all 60 electrode sites for each time point. This methodological choice is grounded in two reasons: (i) A growing body of affective neuroscience research suggests that distinct emotion categories cannot be consistently mapped to specific intrinsic networks in the human brain107,108, therefore, whole-brain representations have been shown to be more effective in detecting differences in emotional activity patterns109,110,111. (ii) Our primary research question focused on characterizing the temporal dynamics of neural representations (i.e., how negativity and amusement representations evolve over time), rather than spatially localized effects. Whole-brain RSA effectively captures time-resolved representational geometry without imposing spatial constraints.

Then, at each middle time point, the dissimilarity matrices were created by calculating the pairwise dissimilarity values between the spatial patterns of neural activity of any two trials, which were quantified as (1−Pearson’s r) values between the data vectors. Next, to identify when neural representations underlying specific aspects of emotion emerge, partial correlations were calculated between the different behavioral RDMs and the EEG RDMs at each middle time point for each participant. For each emotional aspect, when calculating its partial correlation with EEG RDMs, the impact of other irrelevant variables is excluded, aiming to concentrate exclusively on the correlation between the variable under investigation and the EEG RDMs112. Specifically, for amusement, the effects of negativity and the comprehensibility, unexpectedness, and word negativity of corresponding materials were partialled out. For negativity, the effects of amusement and the comprehensibility, unexpectedness, and word negativity of corresponding materials were partialled out. Moreover, it is important to note that only the upper triangle of each RDM was extracted as vectors to calculate these correlations113, because the upper and lower triangles were mirror images of each other, and the diagonal cells always had values of 0114.

Furthermore, cluster-based permutation tests were performed to evaluate the temporal clusters in which the correlations showed significant effects. The null hypothesis corresponded to a correlation coefficient of zero, while significant temporal clusters were specified as adjacent time points with statistical values exceeding the cluster-inducing threshold. To draw the t-value distribution under the null hypothesis, the permutation test was conducted by randomly shuffling the data and calculating the results for 2000 iterations, through which the statistical significance could be assessed. At each time point, the cluster-inducing threshold was set as the 95th percentile of the t-value distribution (equivalent to p < 0.05, one-sided), which means that real data exceeding 95% of the distribution were regarded as significant115.

Finally, the Bootstrap test was used to estimate statistical differences of the onset (i.e., the initial significant time point), peak (i.e., the time point where the peak value of the correlation coefficient is observed), and duration (i.e., the length of the significant time period) latencies between the representation of negativity and amusement in pun-humor sentences. Each bootstrap sampling iteration involved drawing, with replacement, n subjects from the entire pool of subjects to form a new sample, and then the latencies for this sample could be acquired. This process was repeated 1000 times, resulting in empirical distributions of the onset, peak, and duration latencies for the representation of negativity and amusement. The p-value was the number of samples whose differences between the latencies were larger or smaller than zero divided by the number of bootstrap samples (i.e., 1000)115.

Moreover, to explore the relative timing of representing negativity during the online processing of negative words among three different sentence types, a similar RSA procedure was also conducted on non-humor and nonsensical sentences. The Bootstrap test was conducted to estimate the statistical differences in onset and duration latencies among time courses of representing negativity in the three sentence types. Notably, these p-values were corrected for multiple comparisons with a false discovery rate (FDR) of 0.05. We conducted RSA on the representations of negativity and amusement separately for each sentence type for several reasons: (i) Only pun-humor sentences elicited a combination of increased feelings of amusement and negativity, whereas the other two types primarily induced moderate-to-high levels of negativity without amusement. This distinct pattern suggests that the underlying neural representations in the pun-humor condition may differ qualitatively from those in the other conditions. (ii) Analyzing each condition independently allowed us to capture the temporal dynamics of neural representations associated with different emotional meanings, particularly in the context of pun-humor processing. This approach enabled us to investigate whether there exists a time window during which the theoretically opposing emotional states (negativity and amusement) are simultaneously represented. If we had collapsed across all conditions, any observed co-activation of these emotional representations associated with distinct meaning-accessing processes would not be uniquely attributable to the pun-humor condition, as the negativity signal could have originated from trials in the non-humor or nonsense conditions. (iii) Since the fundamental difference in the ability of the sentence types to elicit amusement (high in pun-humor, low/absent in others), combining conditions would induce some confound for our specific research question, where RSA results may reflect category-level information (pun-humor vs. not), rather than how amused someone is during pun-humor comprehension, obscuring the subtle variations in amusement representation within the pun-humor.

Statistics and reproducibility

For all the analysis of behavioral data, both from behavioral experiments and those collected during EEG experiments, we performed one-way repeated-measures analyses of ANOVA, applying the Bonferroni correction for multiple comparisons to measure the corresponding effects. For ERP results, we performed LMEs with the lme4 package103 in R, and the significance of the predictors and their interactions were computed using the Mixed function from the afex package104, and post hoc pairwise comparisons were performed via Bonferroni correction using the emmeans function from the emmeans package105. For the RSA, we assessed the significance of temporal clusters with the following parameters: 2000 permutations, two-tailed, cluster threshold of p < 0.05, and a final threshold of p < 0.05 using cluster-based permutation tests implemented in the NeuroRA106 Toolbox for Python. To further evaluate differences in neural representational latencies, we estimated the distributions for onset, peak, and duration using a Bootstrap test with 1000 iterations in Python.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.