How the brain processes emotional meaning of indirect reply: evidence from EEG

ABSTRACT People often express their message and emotions through indirect utterances. How the intended meaning of indirect utterances is comprehended remains not completely clear. We investigated how the emotional meaning of indirect replies is processed in the brain. Participants were required to comprehend dialogues. Three types of replies were constructed: direct reply, informative indirect reply and negative indirect reply. Our results showed that both informative and negative indirect replies were understood with lower accuracy and longer behavioural reaction times than direct replies. Moreover, informative indirect replies elicited a larger N400 than direct replies, whereas negative indirect replies (compared to informative indirect replies as well as direct replies) elicited enhanced ERP responses only over the late P600 component. These findings suggest that the cognitive processes involved in and the time course of the comprehension of indirect replies change dynamically as a function of the emotional aspects of the intended meaning.


Introduction
In interpersonal communication, people often express their intentions in implicit ways, such as indirect reply.For example, Speaker A asked: What did you think of my presentation?Speaker B answered: It's hard to give a good presentation (Holtgraves, 1998).The intended meaning of Speaker B is that Speaker A's presentation is not good, which is inconsistent with the literal meaning of the indirect reply (Grice, 1968(Grice, , 1975)).The successful understanding of speaker's intention is considered to be achieved through relatively complex pragmatic inference (Grice, 1975;Holtgraves, 1998Holtgraves, , 1999;;Sperber & Wilson, 1986) or quick direct access/activation process (Gibbs, 1986), depending on the specific communication context.In many cases, speakers use indirect replies to express not only neutral informative message but also emotional meaning.Although considerable studies have examined the internal mechanisms underlying the processing of implicit meaning, there is still no clear picture of how the emotional meaning of indirect replies is understood.The present study mainly focused on the cognitive mechanism and time course of understanding the implicit emotional meaning of indirect replies.
Recent studies have used the EEG technique and ERP (event-related potential) analysis to examine the time course of implicit meaning understanding.For example, it was found that compared to literal meaning processing, an enhanced N400 was observed for metaphor and irony praise processing (Caillies et al., 2019;Weiland et al., 2014).Given that the N400 is a negative deflection commonly associated with the easiness of semantic processing (Kutas & Federmeier, 2000;Kutas & Hillyard, 1980), the N400 enhancement is taken to reflect that implicit meaning processing is relatively difficult due to impeded semantic retrieval or semantic unification.Although irony praise processing led to an enhanced N400, considerable studies showed that ironic criticism processing (e.g.negative adjective "His son is very good" spoken with an ironic criticism prosody) usually elicited a late P600 component that had a larger amplitude than those evoked by literal sentences (Caillies et al., 2019;Coulson & Lovett, 2010;Regel et al., 2010;Spotorno et al., 2013).These late P600 effects were considered to indicate that pragmatic reanalysis/ inference (Regel et al., 2010(Regel et al., , 2011;;Regel & Gunter, 2017;Spotorno et al., 2013;Weiland et al., 2014) and more cognitive efforts (Coulson & Lovett, 2010) were involved for understanding the implicit meaning.Recent studies further found that the time course of indirect meaning processing varies dynamically in dependence of the contextual information.Specifically, ironic criticism sentences (compared to the literal ones) were found to elicit a sustained positivity starting from 200 ms post-stimulus onset when cued by consistent contexts, while a discontinuous P200-P600 response pattern was observed for un-cued ironic criticism (Regel & Gunter, 2017), indicating that contextual information can affect the cognitive processes underpinning implicit meaning understanding (Regel & Gunter, 2017).
In addition to the ERP responses, the oscillatory neural activities have been analysed to capture potential processing aspects that are not visible using ERPs.The synchronisation/desynchronisation patterns of neural oscillations have been shown to reflect specific aspects of cognitive operations (e.g.Bastiaansen et al., 2012;Pfurtscheller & Lopes da Silva, 1999).With regard to indirect meaning processing (compared to literal meaning processing), previous EEG studies have observed power decreases in the alpha band (8-12 Hz) of neural oscillations over parietal scalp region (Regel et al., 2014;Spotorno et al., 2013).Desynchronisation of alpha activity has been linked to the enhancing functions of attention (e.g. for reviews see Klimesch, Freunberger, &Sauseng, 2010 andJensen &Mazaheri, 2010) and enhanced memory retrieval operations in language processing (e.g.Klimesch et al., 1997).The alpha band power reduction observed for indirect meaning processing (Regel et al., 2014;Spotorno et al., 2013) were, therefore, believed to reflect a greater cognitive demand for social situational context processing under indirect utterance conditions (Akimoto et al., 2017).
In sum, the above ERP enhancement (or alpha power reduction) effect elicited by implicit meaning processing indicate that more cognitive efforts are required to comprehend the indirect meaning, with this comprehension process either starting at the early N400 stage (semantic processing stage) or being delayed to the late P600 stage (reanalysis or pragmatic inference stage), depending on specific communication situations.Moreover, the difficulty or time course of indirect meaning understanding can be modulated by a range of factors, such as the specific type of implicit meaning (e.g.metaphor, ironic praise, or ironic criticism) and contextual information.However, the detailed mechanisms by which how different factors modulate implicit meaning processing remain to be further clarified.
Emotion is one of the important factors that affect the cognitive processing of implicit meaning, given that the emotional meaning in a conversational setting is crucial for correctly understanding the speaker's intention (Caffarra et al., 2020).The interpretation of the emotional meaning of indirect replies relies on the unification of the replies with their contextual information, but not merely on the processing of indirect utterances themselves (Holtgraves, 1998(Holtgraves, , 1999)).In some cases, indirect replies just deliver neutral informative message, providing more information than just a simple "no".For example, Speaker A: "Do you cook by yourself at home?" Speaker B: "I usually eat takeaway".While in other cases, they convey emotional meaning, such as expressing negative attitude but with an intention to protect listener' public self or "face" (e.g.Speaker A: "Do you like the meal I cooked?"Speaker B: "I usually eat takeaway.").This type of indirect reply conveys not only informative meaning but also emotional meaning and a higher level of social-pragmatic meaning, i.e. communicative intention.
What is the cognitive mechanism and time course of emotional implicit meaning understanding is an important question to be explored.Based on previous empirical results and related theoretical accounts, we propose that there are two possibilities regarding the time course of emotional (versus informative) indirect meaning processing.
One possibility is that the understanding of emotional indirect meaning is likely to be slower than the understanding of informative indirect meaning.This view is inspired by the standard pragmatic inference account of indirect meaning processing (e.g.Grice, 1975Grice, , 1989) ) and relevant experimental findings (e.g.Bašnáková et al., 2014;Stewart et al., 2018;Zhang et al., 2021).According to the standard pragmatic inference account (e.g.Grice, 1975Grice, , 1989;;Holtgraves, 1998Holtgraves, , 1999)), receivers would presume that information senders tailor their utterances to be optimally relevant for the present conversation; in order to have a successful communication, the receivers not only access the literal meaning of an utterance and integrate it into semantic context but also infer the utterance's implied meaning in terms of its communicative situation (Standard Pragmatic View; Grice, 1968Grice, , 1975)).This standard pragmatic inference account pointed out the importance of the inferential processes in implied meaning comprehension, although it made no explicit predictions as to how fast these inferences are made.Some experimental studies, making use of excellent temporal resolution of ERP or eye-tracking technique, further found that implied pragmatic meaning processing is not always as fast as the literal/semantic meaning.For example, understanding indirect replies under some situations modulates the late P600 rather than the early N400 ERP component (Zhang et al., 2021);indirect meaning processing takes additional processing time, with the first fixation time of negative indirect reply being longer than that of direct reply (Stewart et al., 2018).These experimental studies suggest that the comprehension of implied meaning at least in some situations can't be reached immediately, but instead involves additional time-consuming inferential processes.
Experimental studies also found that although both informative and emotional indirect meaning understanding rely on pragmatic inference (e.g.Albrecht & O'Brien, 1993;van Dijk & Kintsch, 1983), emotional indirect meaning understanding may recruit additional or more complex processes.As mentioned earlier, the P600 enhancement effect evoked by ironic criticism processing (vs.literal meaning) indicates that high cognitive cost and more difficult inference process were required for the understanding of indirectly expressed negative information (Baptista, Manfredi, & Boggio, 2017;Caillies et al., 2019;Spotorno et al., 2013).Furthermore, a neuroimaging study conducted by Bašnáková et al. (2014) investigated the neural basis underlying the processing of two types of indirect reply: face-saving (emotional) and informative (neutral) indirect replies.The comprehension of face-saving indirect reply (compared to informative indirect replies) (Bašnáková et al., 2014) led to increased activations not only in emotion-related brain areas (insula and anterior cingulate cortex) but also in the right inferior frontal gyrus and right superior temporal gyrus that have been associated with more difficult discourse processing (Feng et al., 2017;Rapp, Mutschler, & Erb, 2012).Taken together, the above ERP and fMRI experimental results suggest that the understanding of the emotional meaning of indirect expressions (compared to informative indirect meaning) is sometimes more complex and involve additional pragmatic inference.Given the complexity of emotional indirect expressions and the potentially time-consuming pragmatic inference, understanding these emotional indirect expressions may require more processing time than understanding of informative indirect expressions.
Another possibility is that the comprehension of emotional indirect meaning may not require extra processing time when compared to the comprehension of purely informative indirect meaning.That is, the emotional indirect meaning may be accessible at the same time as the informative implicit meaning.This point of view gets support from further development of the standard pragmatic view.For example, the constraint-satisfaction account (Katz, 2005;Pexman, 2008) proposes that during the online understanding of indirect expressions, different sources of information compete for activation over time in parallel.So, the interpretation of implicit meaning is possibly reached at a late stage for more complex reanalysis (or pragmatic inferring) or obtained at an early semantic processing stage due to quick activation, depending on the constraint or competition of multiple factors.In line with this dynamic view, the new development of the direct access account of implicit meaning comprehension also proposes that given an appropriate context, implicit meaning can be accessed quickly at the early semantic processing stage, while acknowledges the possibility that in some cases the understanding of indirect reply is difficult to be directly accessed (Gibbs & Colston, 2012).Meanwhile, with regard to the pragmatic nature of indirect replies in conversation, it has been proposed that utterance that violates the principle of maximisation of relevance usually conveys negative information (Holtgraves, 1998).Therefore we speculate that the implicit emotional meaning of an indirect reply in some cases can be accessed immediately upon detecting the violation of the (partial) literal meaning of this indirect reply with its communication context, hence no extra processing time is needed for its understanding (compared to informative indirect reply understanding).
This quick-access possibility is to some degree in line with the results of Cao and colleagues' study (2019).Cao and colleagues (2018) investigated the understanding of implicit emotional meaning, by using neutral words to construct sentences which express either neutral-congruent (e.g.I cut my carrot when I cooked dinner), neutral-incongruent (e.g.I cut water when I cooked dinner), or emotional (e.g.I cut my finger when I cooked dinner) meaning.The ERPs time-locked to the critical nouns (e.g.carrot/finger/water) showed that, compared to the neutral-congruent condition, the neutral semantic-incongruent condition elicited an enhanced N400, while the implicit emotional meaning condition triggered an enhanced positive ERP waveform, with both the enhanced N400 effect and enhanced positivity effect starting from about 300 ms post-noun onset (Cao et al., 2019).Cao and colleagues argued that the same onset latency (around 300 ms post-noun onset) of the N400 enhancement (to semantic incongruency) and positivity enhancement (to implicit emotional meaning) effects suggest that semantic unification and emotional understanding are conducted concurrently (Cao et al., 2019).
The existing EEG studies, as introduced above, have attempted to examine the time course of the processing of implicit emotional meaning (e.g.Baptista et al., 2018;Caillies et al., 2019;Spotorno et al., 2013).These studies, however, usually compared indirect replies with literal sentences, in which indirect replies expressing emotional information were mixed with those expressing only neutral informative information, and the meaning of the literal sentences were also mixed with informative and emotional intentions.They, therefore, could not give us a clear picture of how emotional versus neutral meaning is understood from indirect replies.
The present study, therefore, aimed to further investigate the time course and cognitive mechanisms of emotional indirect meaning processing, by distinguishing emotional indirect replies from purely informative indirect replies.

The current study
The present study used the EEG technique to explore the cognitive mechanisms and time course of emotional implicit meaning understanding.Specifically, we would like to examine whether the understanding of emotional indirect meaning involves additional cognitive processes or more processing time as compared to informative indirect meaning.In this study, dialogues were used as materials, with each dialogue consisting of a question sentence from Speaker A and a reply sentence from Speaker B. Three versions of replies were constructed: direct reply (conveying neutral information directly), informative indirect reply (conveying neutral information indirectly), and negative indirect reply (conveying negative emotion), which was realised by manipulating the question sentences while keeping the critical reply sentences identical across three conditions.During the EEG experiment, each of the reply sentences was divided into three parts and these three parts were presented in sequence, with the third part (namely, the final phrase of the reply sentence) being the critical phrase that determines the type of replies (direct/informative indirect/negative indirect) (see Method for details).For brain activities time-locked to the critical phase, comparing informative indirect reply with direct reply could tell how the neutral implicit meaning was understood, while comparing informative indirect reply and negative indirect reply could reveal how the emotional meaning was understood from indirect replies.
If the implicit emotional meaning of indirect reply could be understood at the early semantic retrieval or integration stage, emotional indirect reply in the present study was expected to evoke a larger N400 than direct replies; whereas if a complex inference/reanalysis process was required to interpret the emotional indirect reply, an enhanced P600 would be expected (compared to direct replies).
As to the time course of implicit emotional meaning processing, if the understanding of implicit emotional meaning consumes no extra processing time as compared to the purely informative indirect meaning, the ERP effects evoked by the two types of indirect replies would be of comparable amplitude and latency.In contrast, if emotional indirect reply understanding is more cognitively demanding and requires more processing time than the informative indirect replies, the onset latency of the ERP effect elicited by emotional indirect replies was expected to be delayed when compared to that elicited by informative indirect replies.Specifically, it might be that, compared to direct replies, informative indirect replies elicit an enhanced N400, while emotional indirect replies lead to an enhanced P600; it might also be possible that both types of indirect elicited an enhanced P600, whereas the onset latency of the P600 effect evoked by emotional indirect replies would be further delayed than that of informative indirect replies.In addition, the neural time-frequency responses were also analysed to provide additional support for the above predictions.

Participants
This research was approved by the Ethics Committee of Institute of Psychology, Chinese Academy of Sciences.Thirty-four university students (15 males) were recruited for the experiment.The ages of the participants ranged from 18 to 28 years (M = 23.56,SD = 2.37).All participants were right-handed native Chinese speakers with normal or corrected-to normal vision.All participants provided written informed consent before the experiment and were paid money after the experiment.

Materials and design
One hundred and twenty-nine sets of dialogues were used as experimental stimuli, and each dialogue consisted of a question sentence from Speaker A and a reply sentence from Speaker B. The reply sentence was the critical sentence.We constructed three versions for each set of dialogues by manipulating the question sentence of Speaker A while keeping the reply sentence in each set exactly the same.Consequently, for each set of dialogues, the reply of Speaker B was interpreted as a direct reply, a purely informative indirect reply, or a negative indirect reply, depending on the question of Speaker A (see Table 1).Specifically, in the "direct reply" condition, the meaning of Speaker B was explicitly stated and corresponded to the literal meaning of the reply sentence; in the "informative indirect reply" condition, the reply sentence implicitly conveyed Speaker B's meaning and was purely informative (no emotional information); in the "negative indirect reply" condition, the reply sentence conveyed Speaker B's negative attitude in an indirect way for reasons such as "save Speaker A's face".The reply utterance was segmented into three phrases and presented orderly during EEG recording, with the first phrase being a subject (e.g.I), the second an adverbial (e.g.usually), and the third a predicate (e.g.eat takeaway).Our manipulation of the disambiguation point of different types of replies refers to the stimuli design of Zhang et al. (2021).Specifically, each set of dialogues included in the current experiment was carefully constructed, so that the meaning conveyed by Speaker B can be recognised only after the third phrase of the reply sentence is presented.

Pre-tests for the indirectness and comprehensibility
Eighteen subjects who did not participate in the EEG experiment were recruited to evaluate the degree of reply indirectness of each dialogue on a 7-point scale (from 1 "most directly" to 7 "most indirectly").The 129 sets of experimental dialogues finally included in the present study were selected according to the following standards: for direct reply condition the degree of indirectness was < 3; for both types of indirect reply conditions, the degree of indirectness was > 4.5; for all of the three conditions, the rate of correct comprehension is > 90%.An one-way repeated-measures analysis (ANOVA) over indirectness rating scores showed a significant main effect of reply type [F(2,384) = 2837.52,p < 0.0001, η p 2 = 0.94].Further pairwise comparisons revealed that the indirectness level of informative and negative indirect replies were significantly higher than direct replies (ps < 0.0001), and the difference between informative and negative indirect replies was not significant (p = 0.575).These results confirmed that the manipulation of the indirectness of dialogue replies was successful.
Moreover, the comprehensibility of the dialogues was evaluated at the end of each dialogue, by requiring participants to judge (with yes/no answer) the correctness of a question sentence given the content of the corresponding dialogue (see Table 1 for an example of the question sentence).The rate of correct comprehension was >90% (see Table 2 for detailed values).The rate of correct comprehension did not differ between the three conditions (all ps > 0.1).

Pre-tests for the emotionality
Eighteen participants who did not take part in the above pretests and the EEG experiment were asked to evaluate the emotional valence and arousal of the replies given their question sentences.The following instructions were presented to the participants: Please read a series of dialogues between Speaker A and B carefully, and evaluate the emotion that Speaker B wants to express in each dialogue from the perspectives of valence and arousal; Valence refers to the pleasantness of Speaker B, please rate valence on a 7-point scale ranging from 1(extremely negative, namely extremely unhappy) to 7 (extremely positive, namely extremely happy); Arousal refers to the calmness or excitement of Speaker B, please rate arousal on a 7-point scale ranging from 1 (extremely calm) to 7 (extremely excited).The ANOVA performed over valence scores showed that the main effect of reply type reached significance [F(2,384) = 553.88,p < 0.0000, η p 2 = 0.74].Further pairwise comparisons revealed that the valence under negative indirect reply condition was more negative than direct reply and informative indirect reply conditions (ps < 0.0001), whereas the difference between direct reply and informative indirect reply conditions was not significant (p = 0.339).The ANOVA conducted over the arousal scores resulted in a significant effect of reply type [F (2,384) = 334.36,p < 0.0000, η p 2 = 0.64], due to the fact that the arousal under negative indirect reply condition was higher than direct reply and the informative indirect reply conditions (ps < 0.0001) and the difference between direct reply and informative indirect reply conditions was not significant (p = 0.379).The pre-test scores are shown in Table 2.The above results confirmed that the emotionality of the dialogue replies was successfully manipulated.

Pre-tests for the predictability and semantic distances
In order to confirm that the dialogue content prior to the third phase of the reply sentence does not play a critical role in differentiating the content of the replies, a cloze probability test was conducted by presenting the dialogues with the final phase of the reply sentence eliminated.Sixteen participants who did not participate in the formal EEG experiment were asked to complete the utterance with the first phrase that comes to their mind (e.g.Do you like the meal I cook?I usually …) (Zhang et al., 2021(Zhang et al., , 2023)).First, the cloze probability is defined as the proportion of participants who filled in the phrase that was consistent with the reply we designed in both type (i.e.filling in direct reply in the direct reply condition) and content (namely, containing the target phrase).The cloze probabilities of the consistent reply (consistent in both type and content) were 23.6%, 20.9%, and 21.6% for the direct reply, informative indirect reply, and negative indirect reply conditions, respectively (see Table 2).The ANOVA (analyses of variance), with the proportion of consistent reply in each condition as the dependent factor, showed that there was no significant difference between the three conditions [F(2,384) = 1.79，p = .196，ηp 2 = .019].
Second, in order to further validate that the prior dialogue context did not lead to difference in the predictability of the indirectness (regardless of the reply content) of reply sentences, we scored the filled phrases in an alternative way.That is, the phrase filled by the participant was classified as direct reply, indirect reply or others (such as incomprehensible, etc.), regardless of whether it contained emotional information.The chi square test showed that there was a significant difference in the distribution of the indirectness of reply types among the three conditions [x 2 (4) = 18.633, p < .001] (see Table 3).Further pairwise comparison showed that the distribution of direct/indirect replies filled by the participants was significantly different between the direct reply and the two types of indirect reply conditions (Direct Reply vs. Informative indirect reply: [x 2 (2) = 10.165,p < .01]; Direct reply vs. Negative indirect reply: [x 2 (2) = 14.847, p < .01]) (see Table 3 for the direction of the difference), whereas there was no significant difference between the two types of indirect reply conditions (Informative indirect reply vs. Negative indirect reply conditions: [x 2 (2) = 1.008, p = .604]).These results indicated that participants' prediction of the indirectness of the reply sentences is consistent with our experimental design.
In order to further validate the predictability of the emotional valence (regardless of the reply content and reply indirectness) of the reply sentences, the phrase filled in the Cloze Probability test was classified as neutral reply (the valance of reply is neutral), negative reply (the valance of reply is negative) or others (such as incomprehensible, etc.).The chi square test showed that there was significant difference in the distribution of the emotional valence of the filled reply sentences among the three conditions [x 2 (4) = 23.557,p < .001] (see Table 4).Further pairwise comparison demonstrated that the predictability of an emotional response was significantly higher in the Negative indirect reply condition than in the two types of neutral reply conditions (Negative indirect reply vs. Direct reply: [x 2 (2) = 17.293, p < .001]; Negative indirect reply vs. Informative indirect reply: [x 2 (2) = 14.243, p < .001]), whereas the difference between the two types of neutral replies conditions did not reach significance (Direct reply vs. Informative indirect reply: [x 2 (2) = 0.368, p = .832]).These results suggested that the emotional-valence predictability of the reply sentences is consistent with our experimental design.
In addition, to control for potential confounding effect associated with the semantic distance between   the questions and replies, we calculated semantic distances between the question and reply in each dialogue using LSA.The LSA of conversational connection is to calculate the frequency of discourse co-occurrence according to the corpus.In this experiment, the cosine value of LSA of question and reply sentences were calculated (website: http://www.lsa.url.tw/modules/lsa).The results showed that there was no significant difference between the three reply-type conditions [F(2,384) = 1.61，p = .201，ηp 2 = .008].
In total, 129 sets of dialogues were used as experimental materials, with each set including three conditions (direct reply, informative indirect reply, vs. negative indirect reply).The experimental materials were grouped into three lists of 129 dialogues according to the Latin square procedure based on the three reply conditions (with 43 dialogues per condition); dialogues from the same dialogue set were never included in the same list.In addition, each stimuli list also included 43 filler dialogues, which expressed positive messages.Each subject was randomly assigned to only one list of stimuli.All of the dialogues (129 experimental dialogues and 43 filler dialogues) in one list were presented in 5 blocks in a pseudo-random order.

Procedure
In the EEG experiment, the dialogues were presented in 5 blocks in a pseudo-random order.Figure 1 shows an experimental paradigm of the present study.Each trial began with the presentation of a fixation in the centre of the screen for 1000 ms, and then the Speaker A's question appeared.Participants were instructed to press the space bar after they finished reading the question, then the reply of Speaker B (critical sentence) was presented phrase-by-phrase in the centre of screen.The critical reply sentence was divided into three phrases.Each of the first two phrases was presented for 500 ms followed by 300 ms blank screen, and the third phrase was presented for 1200 ms followed by 300 ms blank screen.
In order to ensure that subjects read each dialogue carefully for comprehension, at the end of each dialogue, they were asked to judge the correctness of a question sentence regarding the meaning of the corresponding dialogue by pressing the appropriate keys.The key-hand mapping was counterbalanced across the participants.Half of the participants were instructed to press "J" for "yes" and "F" for "no", while another half were given the opposite instructions.After a brief practice session containing 12 dialogues, the dialogues were presented in five blocks of approximately 8 min each, separated by brief resting stages.

Electrophysiological recording
The EEG was recorded (0.05-100 Hz, sampling rate 500 Hz) from 64 Ag/AgCl electrodes mounted on an elastic cap, with an online reference linked to the left mastoid and an offline algebraic re-reference linked to the left and right mastoids.EEG and EOG data were amplified with AC amplifiers (Synamps, Neuroscan Inc.).Vertical and horizontal electro-oculargrams were monitored.All electrode impedance levels (EEG and EOG) were kept below 5 kΩ.

EEG analysis
For both ERP and time-frequency responses, analysis was time locked to the onset of the third phrase (namely, the final phrase) of the critical reply sentence where the meaning conveyed by Speakers B became obvious and the three types of reply can be disambiguated.

ERP data preprocessing and analysis
For ERP analysis, the eeglab toolbox for Matlab was used to preprocess the original EEG data and calculate ERPs.The raw EEG data were first filtered with a band-pass filter 0.1-40 Hz.After filtering, the EEG data was divided into epochs ranging from 200 ms before the onset of the third phrase of the reply sentence to 1200 ms after the onset of this phrase.A time window of 200 ms preceding the onset of the third phrases was used for baseline correction.Independent Component Analysis (ICA) was used to remove artefacts such as blinking and head movement.The trials with extreme signal amplitude exceeding ±100 μV were rejected.After artefact rejection, the average number of trials left for subsequent analysis was 39, 38 and 38 for the direct reply condition, the informative indirect condition and the reply negative indirect condition respectively.Finally, averages were computed for each participant, each condition, and at each electrode site before grand averages were calculated across all participants.
Based on visual observation and previous studies (Regel et al., 2014;Zhang et al., 2021), the time windows of 300-500 ms (N400) and 500-900 ms (P600) were selected for statistical analysis.Repeated measures ANOVAs were performed with Type of reply (direct reply, informative indirect reply vs. negative indirect reply), Laterality (left, midline, right) and Anteriority (anterior/central/posterior) as independent variables.Figure 2 shows the division of electrodes and brain regions involved in the analysis.Any significant interactions related to the type of reply had been tested for simple effects.When the result of the spherical shape test reached significance, the Greenhouse-Geisser correction was used, and the original degree of freedom and the corrected p-value were reported (significance level).

TFR (Time-frequencyrepresentations) preprocessing and analysis
The Fieldtrip toolbox for MATLAB (Oostenveld, Fries, Maris, & Schoffelen, 2011) was used to analyse Time-frequency responses (TFR).The original EEG data were filtered with a band-pass filter of 1-100 Hz and divided into epochs ranging from 600 ms before and 1500 ms after the onset of third phrase (−600 ∼ 1500 ms).Independent Component Analysis (ICA) was used to remove artefacts such as blinking and head movement.Single epochs were visually inspected to remove the outliers.Fast Fourier transforms (FFTs) was applied to each 2100 ms clean epoch.We used a 500-ms Hanning-window within a frequency ranging from 2 to 30 Hz in a 2-Hz step.The induced power values were obtained by subtracting the phase-locked neural activities from total power values, and then normalized by dividing them by the power values within a window latency of 300-100 ms pre-third-phrase onset baseline.
To examine whether the TFR was significantly different across different experiment conditions, the Fieldtrip toolbox was used to perform cluster-based random permutation tests (implemented in FieldTrip, Maris & Oostenveld, 2007).The random permutation test in the present study was conducted within 0-1200 ms post-third-Phrase onset (in steps of 2 ms) over 17 frequency bins (4-20 Hz in steps of 1 Hz) and fifty electrodes.Figure 2 showed the electrodes included in this analysis.The other electrodes were removed from the statistical analysis in the present study because they were more likely to be affected by artefacts.This permutation test is a good option to control the overall Type I error rate when no a priori information is available to guide the choice of latency windows, frequency bins, and electrode sites (Luck & Gaspelin, 2017).
In the present study, the permutation test (implemented in FieldTrip, Maris & Oostenveld, 2007) computed a separate t-test for each data point ("electrode by time by frequency") of two conditions.All adjacent data points exceeding a preset significance level (0.05%) were grouped into clusters.Cluster level statistics were calculated by taking the sum of the t-values in each cluster.The significance probability of the clusters was calculated by the so-called Monte Carlo method with 1000 random draws.
We examined whether there was a significant TFR difference between each pair of the three conditions (direct reply vs. informative indirect reply; direct reply vs. negative indirect reply; informative indirect reply vs. negative indirect reply).If any of these comparisons showed a significant cluster (with Bonferroni corrections), the difference between the corresponding two conditions was considered to be significant.
For both ERP and TFR responses, if multiple comparisons problem was involved during statistical analysis, the Bonferroni method was used to deal with this problem, and the corrected p values were reported.

Behavioural Results
The average accuracy of reading comprehension was above 90%, indicating that participants could well understand the dialogues.The results of one-way repeated measures analysis of variance (ANOVA) for accuracy and mean reaction times (RTs) showed that the main effect of accuracy was significant [F (2,99) = 13.225,p < 0.001, η p 2 = 0.211].Further analysis indicated that comprehension accuracy was lower for negative and informative indirect replies than that for direct replies (ps < 0.001), and there is no difference between the negative and informative indirect replies.The main effect of RTs was also significant, [F (2,99) = 5.804, p = 0.004, η p 2 = 0.105]; further analysis revealed that the RTs of negative indirect replies (p = 0.005) and informative indirect reply (p = 0.038) was significantly longer than direct replies, and other pairwise comparisons were not significant (see Figures 3 and 4).

ERP Results
Figure 5 displays the grand average ERP waveforms triggered by the third phrases in different replies.The topographic distributions of indirect replies effect were shown in Figure 6.
In the temporal window of N400 (300-500 ms), the main effect of the type of reply was significant, [F (2,66) = 3.6641, p = 0.031, η p 2 = 0.1].Further pairwise comparisons revealed that informative indirect replies elicited a larger N400 than direct replies (p = 0.0139), whereas the differences between the other pairs (informative indirect reply vs. negative indirect reply; direct reply vs. negative indirect reply) were not significant (all ps > 0.1).
In the temporal window of P600 (500-900 ms), the main effect of the type of reply was significant [F(2,66) = 7.952, p = 0.001, η p 2 = 0.194].Further analysis revealed that negative indirect replies elicited a greater P600 than informative indirect replies (p = 0.006) and direct replies (p = 0.014), and the difference between direct replies and informative indirect replies did not reach significance (p > 0.1).

TFR Results
The cluster-based random permutation test on TFR data demonstrated that negative indirect replies induced significant power decreases than direct replies (p = 0.021, with a cluster extending approximately from 800 ms to 950 ms post-noun in the alpha band (8-12 Hz), at C2, C4, C6, CP2, CP4, CP6, TP8, P2, P4, P6, PO8) (see Figure 7).Neither the difference between negative and informative indirect replies nor the difference between direct replies and informative indirect replies reached significance (all ps > 0.1).

Discussion
The purpose of this EEG study is to investigate the time course of understanding the emotional meaning of indirect replies.The major results were that both informative and negative indirect replies, relative to the direct reply, led to reduced comprehension accuracy and delayed behavioural response times (RTs).More importantly, the informative indirect reply elicited a larger N400 than the direct reply, whereas the negative indirect reply elicited a larger P600 than both the informative indirect reply and the direct reply; in addition, the  negative indirect reply induced alpha-band oscillatory power decreases compared to the direct reply.These results were discussed below in more detail.
The behavioural results of the present study showed that both negative and informative indirect replies were understood with lower accuracy and longer RTs, indicating that the understanding of indirect replies is more difficult and requires more cognitive efforts than the comprehension of direct replies.This enhanced processing difficulty suggests that more difficult semantic retrieval/unification or complex pragmatic inference might be involved in understanding the implicit meaning of indirect replies, which is consistent with the findings of previous studies (Holtgraves, 1998(Holtgraves, , 1999)).
For the neural responses, it was found that informative indirect replies elicited a larger N400 than direct replies.An increase in N400 amplitude is generally considered to be associated with difficulty in semantic retrieval or semantic unification between the current information and its preceding context (Coulson & Van Petten, 2002;Weiland et al., 2014).Therefore, the N400 enhancement effect of the current study suggests that although informative indirect replies are more difficult to understand than direct replies, the brain has begun to process the implicit meaning of this type of indirect replies at a relatively early processing stage, namely the stage of semantic retrieval/integration.We speculated that, when confronted with informative indirect replies, the brain is able to access and integrate wider information (i.e.world knowledge) to compute their implicit pragmatic meaning while integrating the current literal semantic information into preceding contexts.
The negative indirect replies did not elicit an early N400 effect, but led to a larger P600 component when compared to direct replies and informative indirect replies.That is, the understanding of negative indirect reply was not only delayed to a late processing stage but also more difficult than that of informative indirect replies.Previous studies have related the late P600 to reanalysis processes such as pragmatic inference (Coulson & Van Petten, 2002;Regel et al., 2010Regel et al., , 2011;;Regel & Gunter, 2017;Spotorno et al., 2013;Weiland et al., 2014) and cognitive efforts required for constructing situational models (Coulson & Lovett, 2010;Leuthold et al., 2012).The P600 or late positivity component has also been found by a number of studies to be enhanced to negative (versus neutral) verbal stimuli (e.g.Bayer, Sommer, & Schacht, 2010 for words embedded within sentences; Kanske &Kotz, 2007 andHinojosa et al., 2014 for isolated words), although the ERP pattern (e.g.enhanced N400 or P600) associated with emotion processing tends to change dynamically depending on the specific types of emotion and stimuli (e.g.Proverbio, Santoni, & Adorni, 2020).Therefore, the present P600 enhancement to negative indirect replies suggests that understanding the intended emotional meaning of negative indirect replies might rely on reanalysis processes (or pragmatic inference) at a late stage of processing.
Moreover, the negative indirect replies, relative to direct replies, induced neural power decreases in the alpha (8-12 Hz) band around 800-950 ms post-target onset.Event-related desynchronisation in the 8-12 Hz alpha band was usually observed for the relatively difficult processing such as semantic/syntactic anomalies processing (Bastiaansen et al., 2009;Davidson & Indefrey, 2007;Willems et al., 2008) and irony  comprehension (Akimoto et al., 2017;Regel et al., 2014;Spotorno et al., 2013).The reduction in the alpha band was believed to mirror the increase in the activity of the brain regions responsible for reprocessing of linguistic input (Hajra et al., 2018;Kielar et al., 2014).In the present study, the alpha-power decreases observed for negative indirect replies within the very late time window (around 800-950 ms post-target onset) are therefore in line with the late P600 effect elicited by negative indirect replies, and provide further evidence for our speculation that the comprehension of negative indirect replies requires more reprocessing of the current utterance and contextual information to make pragmatic inference.
In addition, no significant brain activity effect was observed at the first two parts of the reply sentence from Speaker B (see Supplementary Material S1).This additional analysis was time-locked to the first two phrases of the reply sentence, after the contextual question and before the critical phase of the reply, the results showed no significant difference in brain activity between direct reply, informative indirect reply, and negative indirect reply.Obviously it was the final phrase of the reply sentence that significantly triggered the accessing of the implied meaning of indirect replies.Meanwhile, although the contextual question in the dialogues and the first two phrases of the reply sentence together provided the information for the processors to predict the general indirectness and emotion valence of the reply, the exact content of the reply was not available until the presence of the final phrase of reply sentence.The results of our reply-content-predictability pretest showed that given preceding context and the first two phrases of the reply sentence, the cloze possibility of the exact content of the reply was not significantly different across the three conditions.Consequently, no significant ERP effect of Type of reply was observed before the presentation of the final phase of the reply sentence.
Importantly for the current study, when compared to direct replies, although informative indirect replies elicited an N400 enhancement effect, the processing of negative indirect replies only led to a late P600 enhancement effect and alpha power decreases around 800-950 ms post-target, with no significant brain-activity effect being observed in the early N400 window latency.That is, relative to informative indirect replies, the processing of negative indirect replies was actually delayed, rather than merely lasting longer.We speculated that this pattern of results is closely related to the conversational context of the present study.It is possible that in the course of the task, the contextual question in the negative indirect reply condition (e.g. What do you think of my hair?) could lead our participants to expect that the reply of Speaker B would be about emotional attitude, although they did not know exactly how (directly or indirectly) and what (negative or positive emotion) Speaker B would reply.When with such an expectation of an emotional reply (as compared to informative reply) in mind in the negative reply condition, the processors possibly was waiting for (or accessing) additional information and only computed the indirectly implied emotion at a late stage, rather than conducted a two-stage of processing (namely, processed the indirectness or indirectness of the reply first and then inferred its intended emotional meaning).Therefore, a late P600 but not an early N400 effect was observed in the emotional indirect reply condition of our study.The processing of negative meaning being delayed to the later P600 window latency might be due to the more complex nature of this emotional indirect meaning processing (see the introduction section for details) and the specific contextual constraint of the dialogues used in the current study.The present result could not exclude the possibility that the interpretation of the emotional indirect meaning can be reached at the early N400 window latency when more information is provided in the conversational context, which needs to be examined by future studies.
The existing EEG studies have already found that the understanding of indirectly expressed negative information, such as irony, involves a late-stage of processing in the P600 window latency (Baptista et al., 2018;Caillies et al., 2019;Spotorno et al., 2013), which was in line with the results of the present study.The present results further extended the existing EEG findings through disentangling the emotional aspects of an indirect reply from the purely informative indirect reply.Note that our study adopted the same design as in Bašnáková et al. (2014), with both studies distinguishing emotional indirect replies from purely informative indirect replies.Bašnáková et al. (2014) demonstrated the complexity of emotional-indirect-reply processing from the perspective of brain activation pattern, as the comprehension of face-saving indirect replies (vs.information indirect replies) was found to require additional cognitive processing in brain areas such as right inferior frontal gyrus and right superior temporal gyrus.The results of the current study not only confirmed the finding of Bašnáková et al. (2014) but also provided further evidence for the complex nature of emotional-indirect-reply understanding from a different perspective, namely the perspective of processing speed.The new contribution of the present study is that it offers a relatively clear picture of the time course of understanding the intended emotional meaning of indirect replies.
Specifically, although the informative message of an indirect reply can be quickly processed at the early stage (in the N400 window latency), the negative emotional meaning of an indirect reply, at least in some communication situations, does not begin to be computed until the late reanalysis stage.The results suggest that the cognitive processes involved and the time course of the comprehension of indirect replies change dynamically as a function of the emotional aspect of the intended meaning, which provides new insights into our understanding of the cognitive mechanisms of implied meaning comprehension.
As mentioned in the introduction, different accounts have been put forward to explain the mechanism of indirect reply comprehension.The present results are to some degree in line with the dynamic view of implicit meaning understanding (Gibbs & Colston, 2012;Katz, 2005;Pexman, 2008), indicating that whether the interpretation of implicit meaning is reached through complex pragmatic inferring or quick/direct activation might depend on the constraint of different factors.However, it needs to point out that the present results could not deny the applicability of the direct access account in some cases of implicit meaning comprehension, also could not exclude the possibility that the interpretation of emotional indirect replies may be reached at the early semantic processing stage when strongly constrained by appropriate contextual information.Although whether and how contextual constraints modulate emotional implicit meaning interpretation need to be investigated in future studies, the results of the current study demonstrate that the emotional aspects of indirect replies can affect the time course of the interpretation of implied meaning.We speculated that the comprehension of an informative indirect reply may rely on relatively easy pragmatic inference, hence the brain is able to access and integrate related knowledge to compute this implied informative message at the early semantic retrieval/integration stage of processing; however, the comprehension of a negative indirect reply, at least in some communication situations, may require relatively complex pragmatic inference (such as additional cognitive perspective-taking and emotion-related inference), and consequently the brain may possibly wait for or access more related knowledge and only begin to calculate this implied emotional meaning at a later reanalysis stage of processing.Further studies are needed to better understand the detailed cognitive and neural processes involved in the comprehension of the intended emotion of indirect replies.
Nevertheless, the current study has its limitations in that the emotional indirect replies in our dialogue stimuli convey not only negative emotional meaning but also a social intention to save the counterpart's face by expressing negative attitude indirectly.The present results could not distinguish whether the delayed understanding of emotional indirect replies was driven by emotional processing, social intention processing, or both, which needs to be clarified by further experimental studies.In addition, in daily life, a very common way of communication is spoken language, and conversations between people usually contain auditory clues such as prosody.Prosody not only conveys lexical semantic meaning (such as lexical tone) but also can be used to express social-pragmatic information (such as emotion, attitude, and intention) indirectly (e.g.Bryant & Fox Tree, 2002;Hellbernd & Sammler, 2016).Although some models (e.g.Schirmer & Kotz, 2006) and experimental studies (e.g.Vergis, Jiang, & Pell, 2020) have already begun to address the role of prosody in social-pragmatic information processing and its internal mechanisms, how and when prosodic cues are integrated with lexicalsemantic and contextual information to reach emotion or intention comprehension remains to be examined in further studies.

Conclusion
The present study investigated the underlying cognitive mechanism and the time course of the processing of two types of indirect replies (negative and informative indirect replies) to reveal how the emotional meaning of indirect replies is understood.The results showed that both types of indirect replies were more difficult to understand than direct replies, as indicated by reduced comprehension scores, lengthened behavioural reaction time, and enhanced ERP responses.More importantly, the implicit meaning of informative indirect replies was accessed at the relatively early stage of semantic processing, whereas the intended emotional meaning of negative indirect replies was not obtained until the late reanalysis stage.These results indicate that the emotional aspects of implicit meaning can modulate the processing time course of indirect replies understanding, which to some degree provides evidence for the dynamic view of implicit meaning comprehension.

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

Funding
This work was supported by Grants from the National Natural Science Foundation of China [grant number 32171057].

Figure 1 .
Figure 1.An experimental paradigm of the present study.The sentences were presented in Chinese, the translations are also provided here.

Figure 4 .
Figure 4. Behavioural results of mean reaction times (RTs) of comprehension.

Figure 5 .
Figure 5. Grand average ERP waveforms in the direct reply, informative indirect reply and negative indirect reply conditions.The blueshadowed rectangles indicate the window-latency (300-500 ms post-verb onset) that showed significant difference between the Direct reply and Informative indirect reply.The green-shadowed rectangles indicate that the ERP responses (within 500-900 ms post-verb onset) evoked by Negative indirect reply significantly differed from those elicited by Direct reply and Informative indirect reply.

Figure 6 .
Figure 6.The topographies of the Negative indirect reply -Direct reply ERP effect, Informative indirect reply -Direct reply ERP effect and Negative indirect reply -Informative indirect reply ERP effect within the window latency of 300-500 ms and 500-900 ms respectively.

Figure 7 .
Figure 7. TFR results (Negative indirect reply -Direct reply).The black rectangles indicate the frequency-of-interest alpha band (8-12hz) and window-latency (800-950 ms post-verb onset) that showed significant difference between Negative indirect reply and Direct reply.The crosses in the right topography plots indicate the electrode sites showing significant alpha power increase/decrease effects in the permutation test performed over electrodes.

Table 1 .
Example of experimental dialogues.
same as above I don't like the meal you cook.

Table 2 .
Pretests results of the stimuli.

Table 3 .
The distribution of the indirectness of The Cloze Probability.
Note: The phrase filled by the participant was classified as direct reply (the indirectness of reply is direct), indirect reply (the indirectness of reply is indirect) and others (such as incomprehensible, etc.).

Table 4 .
The distribution of the emotion of The Cloze Probability.
Note: The phrase filled by the participant was classified as neutral reply (the valance of reply is neutral), negative reply (the valance of reply is negative), and others (such as incomprehensible, etc.).