Impaired social concept processing in persons with autistic-like traits

Situated models suggest that social concepts are grounded in interpersonal experience. However, few studies have tested this notion experimentally, and none has targeted individuals with reduced social interaction. Here, we assessed comprehension of text-level social and non-social concepts in persons with and without autistic-like traits. Participants read a social and a non-social text and answered questionnaires targeting social and non-social concepts, respectively. We compared behavioral outcomes, gauged their contribution to subject-level classification, and examined their association with validated measures of autism. Persons with autistic-like traits showed selective deficits in grasping text-level social concepts, even adjusting for intelligence, memory, and vocabulary. Also, social concept comprehension was the only variable that significantly classified between groups. Finally, social concept outcomes correlated negatively with measures of autism, including social interaction. Our results suggest that reduced interpersonal experience selectively compromises text-level social concept processing, offering empirical constraints for situated models of social semantics.

These features are almost identical in people with an actual diagnosis of autism spectrum disorder (ASD) 25 .Individuals with high AQ score lower than those with low AQ on social cognition measures [28][29][30] , and their performance correlates with scores in the Autism Diagnostic Observation Schedule-2 (ADOS-2)-the goldstandard instrument in ASD research, including subscales of reciprocal social interaction and interpersonal communication skills 31 .Interestingly, some such impairments seem uninfluenced by broader cognitive skills, including general intelligence, working memory, and vocabulary 32,33 .Therefore, high AQ scorers offer a critical model to test whether social concept processing is related to interpersonal experience, and whether this link is mediated by more general cognitive traits.
Importantly, this issue can be studied with ecological validity through naturalistic texts.Most social concept research 17,34,35 has employed single-item stimuli, overlooking the contextual anchorage needed to capture words' precise social implications.For instance, while the verb promise may or may not entail a social commitment when presented in isolation, it acquires rich interpersonal significance in the sentence She promised she would always be there for him.Indeed, textual context prompts specific emotional, empathic, and mentalistic operations that ground these concepts in actual social experience 2,9,24 .Though blind to the individual role of such socio-cognitive variables, text-level paradigms thus enable more naturalistic assessments of social concepts, addressing calls for ecologically valid insights on the construct 9,24 and on language processing at large [36][37][38][39][40][41][42] .
Against this background, we employed a naturalistic text paradigm assessing social and non-social concept comprehension via multiple-choice questionnaires 24 .Importantly, this task has revealed selective social concept deficits in other populations with socio-interactive atypicalities 24 .We established (sub-clinical) autistic traits via the AQ.Also, to better capture the social profiles of high AQ participants, we administered module 4 of the ADOS-2 scale, computing its total score and outcomes in relevant subscales.Based on previous findings, we predicted that high AQ participants would be outperformed by low AQ persons on the ST (but not on the nST) questionnaire, irrespective of intelligence, working memory, and vocabulary skills.Second, we anticipated that ST outcomes would robustly classify between low AQ and high AQ participants at the individual level.Finally, we predicted that the greater the social detachment of high AQ participants, the lower their capacity to grasp ST information.With this approach, we aim to shed new light on the role of situated interactive experiences in grounding social concepts.

Participants
Participants were drawn from a large pre-screening group of 878 students.All of them voluntarily completed an online version of the AQ, yielding a mean total AQ score of 18.24 (SD = 5.5).These individuals were contacted through various channels, including the Canarian Association for Autism Spectrum Disorder, online platforms from faculties at Universidad de La Laguna (Spain), and a student-support program from the same university.Recruitment efforts also included oral invitations to students during lectures.Participants were considered to have a high AQ if their scores were greater than 30 (2 SDs above the overall group's mean), and to have low AQ group if their scores fell between the mean and one SD below it (keeping within the range of 13-18 to avoid extremely low values, as in previous research) 28 .
The final sample comprised 36 native Spanish speakers, 18 with high AQ and 18 with low AQ (Fig. 1A).This sample size reaches a power of 0.88 (Supplementary material, Power estimation section).All participants were right-handed, had normal or corrected-to-normal vision and hearing, and presented no history of psychiatric disorders, neurological diseases, primary language deficits or substance abuse.Both groups were matched for sex, age, and years of education.
All participants were assessed for non-verbal intelligence, via the Raven's Progressive Matrices Test [RPMT 43 ]; mnesic skills, through a verbal working memory task [WMT 44 ]; and linguistic skills, via the Peabody Picture Vocabulary Test III [PPVT-III 45 ]-for a description of these tests, see Supplementary material, Cognitive assessment.All assessments were conducted in person by one of the researchers (JLP) in a quiet room with dim lighting.Participants in the high AQ group were also evaluated via module 4 of the ADOS-2 scale.In all cases, participants first completed a demographic questionnaire, followed by the RPMT, then by the WMT, then by In the discourse-level task, participants listened to an ST and an nST, each text being followed by 16 comprehension questions.The order the texts was counterbalanced across participants.(C) Between-group comparisons revealed significantly lower scores for the high AQ than for the low AQ group on the ST, but not on the nST.(D,E) Classification results for the ST (panel F) and the nST (panel G), together with RPMT, WMT, and PPVT-III scores, depicted with a confusion matrix (left inset), feature importance rankings (top right inset), and a ROC curve (bottom right inset).(F) Pearson's partial correlations, covaried by WMT scores, showed that ST outcomes were negatively associated with the ADOS-2 total score (top left), the 'reciprocal social interaction' subscore (top right), and the 'stereotyped behaviors and restricted interests' subscore (bottom left), but not the 'communication subscore' (bottom right).(G) Pearson's partial correlations revealed non-significant associations between nST performance and any each of the ADOS-2 measures, covaried by WMT scores.ADOS-2 autism diagnostic observation schedule-2, low AQ low autism spectrum quotient, high AQ high autism spectrum quotient, RSI reciprocal social interaction, SBRI stereotyped behaviors and restricted interests, Comm communication, WMT working memory task, ST social text, nST non-social text, PPVT-III Peabody Picture Vocabulary Test, RPMT Raven's progressive matrices test-III.Double asterisks (**) indicate significant differences after covariation for RPMT, WMT, and PPVT-III outcomes.

Naturalistic texts
All participants listened to two stories used in previous social concept research: an ST (highlighting interpersonal events) and an nST (narrating the activities of a single individual) 24 .The social/non-social contrast was manifested in the verbs and circumstantial adjuncts of each text.Most of these units in the ST referred to social interactions between two people (e.g., greeting someone kindly), whereas all verbs and circumstances in the nST lacked socio-interactive associations, as they described the actions of an unaccompanied character on various objects (e.g., preparing breakfast).
Both stories were composed through a systematic text-construction protocol 36-39, 46, 47 .First, 22 grammatical patterns were created and pseudo-randomly distributed for each text, each filled with strategic lexical items.For example, the pattern "Compound sentence: clause 1 [complement + empty subject + verb + complement] + conjunction + clause 2 [empty subject + verb + complement]" was filled as Immediately, he went over to Juan and earnestly asked for a favor for the ST, and as Afterwards, he would read a book and listen to classical music on the balcony for the nST.Both texts were matched for (1) character count; (2) overall and specific word-type counts; (3) mean content-word frequency, familiarity, syllabic length, number of letters, propositional density, and imageability; (4) sentence and sentence-type counts; and (5) a readability measure (Szigriszt-Pazos Index) and its associated readability rating (Inflesz scale).Moreover, the texts were matched for grammatical correctness, coherence, and comprehensibility (as judged by 20 raters on a scale from 1 through 5), as well as emotional content (positive, negative or neutral) and arousal level (intensity of the chosen emotion, from 1 through 5, as established by 14 raters).All sentences communicated mostly literal meanings and contained no jargon.See statistical details in Table 2.For full transcriptions and approximate English translations, see Supplementary material , Naturalistic texts.Stories were audio-recorded by a male native speaker of Canarian Spanish (the participants' regional dialect), at a smooth pace, in .mp3stereo format.Each narration lasted roughly 100 s.

Comprehension questionnaires
Following each narration, participants completed a 16-item multiple-choice questionnaire featuring wh-questions 47 .Half the questions pointed to verb-related information, denoting the characters' activities, and were mostly structured as What did [a character] do when…?The other half aimed at circumstances, realized by adverbial or prepositional phrases pointing to locative, causal, temporal, or social information signalled by Where, Why, When or How.In the ST questionnaire, all verb-related and circumstantial questions targeted social interactions (e.g., How did Juan react to Albert's payment offer?He rejected it; How did Albert insist?Kindly).Conversely, in the nST questionnaire, all verb-related and circumstantial questions targeted non-social information (What did Luis do with the TV?He turned it on; Where was the clock?On the nightstand).
Questions were presented following the stories' sequence of events, alternating between verb-related and circumstantial items.Successive questions were independent from each other.Each question featured five options: a correct response, three subtly incorrect options, and an 'I don't remember' option.Sequencing of options

Procedure
Participants were instructed to close their eyes and listen carefully to the recorded texts through professional, high-definition headphones (Fig. 1B).At the beginning of the task, a different narrative was administered for familiarization purposes.It consisted in one text with the same length and structure as the ones in the experiment, followed by three sample questions on the computer screen.After this practice, participants listened to the ST and nST.Each text was played only once.Texts were counterbalanced across participants.Following each narration, its corresponding questionnaire was presented with its options.Participants were instructed to choose the correct answer as quickly as possible, using predefined keyboard keys.Selected options were automatically saved.The experiment ran in e-prime.

Behavioral data analysis
First, we carried out a cross-textual analysis via a 2 × 2 mixed-effects ANOVA, with a between-subject factor 'group' (high AQ and low AQ) and a within-subject factor 'text' .Then, given the mismatch in motor content (see Table 2), and as in previous works employing this discourse-level paradigm, we implemented a within-text analysis, comparing the performance between groups for each text separately 9,38,46 via one-way ANOVAs, with 'group' as the categorical factor.We thus favored comparability with previous studies while circumventing confounds (fine-grained aspects not controlled between texts) and unduly stringent analyses.Also, to determine whether www.nature.com/scientificreports/potential text retrieval outcomes were related to non-verbal intelligence, working memory, or vocabulary skills, results from the naturalistic text task were reanalyzed via ANCOVAs, covarying for the total scores of the RPMT, WMT, and PPVT-III, as in previous works 37 .Alpha levels were set at p < 0.05.Effect sizes were calculated via partial eta squared (η 2 ) for ANOVAs and Cohen's d for pairwise comparisons 53 .
To gauge the importance of social concept processing for discriminating individuals in each group, we performed linear discriminant analyses (LDAs).This method identifies the linear combination of a set of covariates that maximizes between-group differences while minimizing within-group differences 54 .We ran two models, one for the ST with four predictors (ST, RPMT, WMT, and PPVT-III scores) and one for the nST with four predictors (nST, RPMT, WMT, and PPVT-III scores).Performance estimates were corrected in a data-driven approach via leave-one-out cross-validation 55 .To determine which covariates best differentiated between high AQ and low AQ individuals in each LDA, we implemented a stepwise forward variable selection using the Wilk's Lambda criterion.LDA classification results are reported through confusion matrices and receiver-operating characteristic (ROC) curves.All analyses were performed on R 4.1.1 56.
Finally, to examine whether social concept outcomes were associated with autistic traits in the high AQ group, we performed Pearson's partial correlations between performance on each text and four measures from the ADOS-2: the instrument's total score as well as the 'reciprocal social interaction' , 'communication' , and 'stereotyped behaviours and restricted interests' subscales.These analyses were covaried by WMT scores, given that working memory was impaired in the high AQ group and has been shown to correlate with social cognition outcomes 57,58 .

Discussion
This is the first study to examine (text-level) social concept comprehension in persons with and without autisticlike traits.Unlike nST scores, ST scores were significantly lower in the high than in the low AQ group, and this result was uninfluenced by individual variability in non-verbal intelligence, working memory, and vocabulary skills.Moreover, performance on the ST (but not on the nST) robustly classified subjects as high AQ or low AQ, and it was negatively associated with total ADOS-2 score and relevant subscales.Below we discuss our findings, addressing their theoretical implications.
Our key finding is that high AQ participants were outperformed by their low AQ counterparts on the ST but not on the nST.This aligns with research revealing impaired processing of social semantic content in persons with ASD 59 and selective ST deficits in neurodegenerative patients with primary socio-affective impairments 24 .www.nature.com/scientificreports/Suggestively, social (relative to emotional) concepts distinctly recruit fronto-temporo-parietal regions (e.g., prefrontal cortex, middle temporal gyrus, temporo-parietal junction) 60 that present anatomo-functional alterations in ASD [61][62][63][64] , especially during social cognition tasks 65,66 .Our results support and extend these findings, suggesting that social concept skills may be partly driven by situated interpersonal experience.
The ST difficulties of the high AQ group emerged independently of non-verbal intelligence, working memory, and vocabulary level.Previous studies on ASD and high AQ samples suggest that these domains may influence performance on other socio-cognitive domains 67 , such as theory of mind 68 , emotion recognition 69 , and socializing 70 .However, no such influences have been detected in other studies 32,33 , suggesting that only certain socio-cognitive domains, or certain socio-cognitive tasks, are influenced by such general skills.In this sense, our results indicate that selective text-level social concept deficits in high AQ persons may not be secondary to broader cognitive dysfunctions, but rather represent a sui generis, category-specific semantic deficit.This further suggests that social concept processing may be grounded in interpersonal experience, beyond the latter's effects on other cognitive functions.
Interestingly, the cross-textual analysis revealed significantly lower scores on the nST than on the ST.This discrepancy could be influenced by the nST's greater motoric content-namely, the level of bodily movement implied by verbs 51 .Indeed, stimuli with high motor content prove more cognitively challenging than those with low motor content 71,72 .The interaction between group and text, however, was not significant.Such an effect, we surmise, may have been abolished by the greater demands of the nST (conceivably, a less demanding control text could have elicited better outcomes in both groups, increasing the performance difference relative to the ST).While this remains speculative, such a null result further emphasizes the importance of employing within-text analyses to elucidate condition-specific differences between groups 9,38,46 .
In this sense, LDA results showed that ST (unlike nST) scores discriminated individual high AQ participants from low AQ participants.ST scores emerged as the best classifier between groups, with an accuracy of 72% (and perfect balance between both groups).This variable even outweighed WMT scores-which is notable given the systematicity of working memory deficits in persons with high AQ 73 and ASD 74 .Thus, group-level results were not dependent on a few low-scoring participants.Indeed, social semantic outcomes in natural speech also constitute the most accurate variable for classifying persons with and without ASD 75 .In people with reduced social experience, then, ST comprehension difficulties seem inter-individually consistent and more discriminatory than other cognitive deficits.
Moreover, ST outcomes in the high AQ group were negatively correlated with ADOS-2 scores (adjusted for WMT scores).This reinforces the claim that social concept processing hinges on interpersonal experience.Indeed, in ASD research, ADOS-2 scores have been shown to correlate with deficits on other social cognition tasks 31,76,77 .Notably, subscale analyses revealed that ST outcomes were correlated with the 'reciprocal social interaction' (ability to engage in exchanges with one or more people) and the 'stereotyped behaviors and restricted interests' subscores (conducts repeatedly in an exact way, often involving individual actions).Scores on these subscales have been associated with domains like emotion recognition 78,79 and theory of mind 80,81 , attesting to their relevance to socio-cognitive skills at large.Importantly, no correlations with ADOS-2 scores emerged for the nST, even though it proved more demanding than the ST.This further suggests that socio-cognitive variability across high AQ participants was not related to difficulties with linguistic materials at large, but rather confined to texts conveying social information.Thus, though strictly correlational, our results lend additional support to our hypothesis, as they indicate that the poorer the engagement with others, the lower the capacity to grasp social information.
Taken together, our findings indicate that the ability to process social concepts is related to situated gregarious experience.This evidence supports the view that interpersonal exchanges are distinctly needed to acquire and use diverse abstract concepts, including social ones 5 .From a situated cognition perspective, concepts become consolidated by virtue of experiencing the events or referents they denote [13][14][15] , so that their ulterior activation would involve accessing relevant memory traces 13,16 .Such traces, we propose, would be suboptimally entrenched in hAQ individuals given their reduced social experience.In fact, social isolation has been linked to abnormal activation of prefrontal, superior temporal, and temporo-parietal circuits 82,83 specifically recruited during social concept processing 60 .Accordingly, we speculate that atypicalities in such regions due to low social engagement would render social concepts harder to access or even construe (a conjecture that should be examined in future studies).
Importantly, however, we cannot ascertain whether present results are specific to social concepts or general to abstract concepts at large.Indeed, prominent accounts propose that sustained interpersonal experience is critical for grounding all types of abstract concepts, including emotional, philosophical, and spatiotemporal ones 5 .Unfortunately, the nST (our control condition) involved action events rather than another subcategory of abstract concepts, such as emotional concepts.Still, the ST and the nST were controlled for emotional valence and arousal, partly ruling out some such confounds.Relatedly, social concepts have been shown to make up a differentiated lexico-semantic space 6 and to engage socio-cognitive regions significantly more than emotional concepts 60 .By the same token, our study warrants the view that social concepts might be especially (though not exclusively) dependent on interpersonal experience.
Finally, note that our findings stemmed from naturalistic narratives.Most social concept investigations have relied on randomized sequences of isolated words or sentences 7,[17][18][19][20] , while discourse-level studies on ASD have targeted syntactic 84 and macro-structural 85 aspects.Despite their major contributions, such approaches fail to capture social concepts in context-rich settings.The detection of category-specific difficulties in this study offers important support for such links, as contextual cues in natural texts may facilitate performance by priming or favoring maintenance of relevant information 86 .In this sense, our study meets the pressing call for more ecological insights on language processing 37,41,87  www.nature.com/scientificreports/This study is not without limitations.First, our groups were relatively small, calling for replications with more participants.Second, the number of testing items was limited.Although our task doubles the number of trials in previous studies 39,46 , future adaptations should increase this figure.Third, our protocol does not capture the relative contributions of mentalistic, empathic, or emotional abilities to the observed outcomes.Accordingly, present results may warrant alternative or complementary explanations, given that STs likely engaged these domains, which are affected in high AQ individuals [28][29][30] .Future works should examine or rule out such factors in this population by leveraging classical fine-grained tasks (e.g., synonym judgment, semantic feature verification) 34 or by incorporating relevant socio-cognitive measures based on non-verbal stimuli-e.g., facial emotion recognition 88 or picture-based empathy measures 89 .Fourth, future studies on high AQ individuals should leverage more tightly controlled social and non-social stimuli to enable stringent tests of group-by-condition interactions.Finally, we call for new studies to expand social concept research by involving persons with actual ASD diagnoses, including offline or online neural measures, and performing interventions to enhance or manipulate social experiences (e.g., via virtual reality) and examine their direct impact on social concept processing.
In conclusion, we showed that persons with reduced interpersonal experience presented difficulties with grasping text-level social concepts, and that such difficulties correlated with validated measures of social interaction.This evidence supports situated views of semantic processing, affording new insights on (a particular type of) abstract concepts.Future applications of our approach could promote useful breakthroughs to understand the links between (outward) social events and their (inner) cognitive construal.

Figure 1 .
Figure1.Experimental setup and results of the within-text analyses.(A) Participants were characterized in terms of their demographic, autistic, and cognitive profiles.(B) In the discourse-level task, participants listened to an ST and an nST, each text being followed by 16 comprehension questions.The order the texts was counterbalanced across participants.(C) Between-group comparisons revealed significantly lower scores for the high AQ than for the low AQ group on the ST, but not on the nST.(D,E) Classification results for the ST (panel F) and the nST (panel G), together with RPMT, WMT, and PPVT-III scores, depicted with a confusion matrix (left inset), feature importance rankings (top right inset), and a ROC curve (bottom right inset).(F) Pearson's partial correlations, covaried by WMT scores, showed that ST outcomes were negatively associated with the ADOS-2 total score (top left), the 'reciprocal social interaction' subscore (top right), and the 'stereotyped behaviors and restricted interests' subscore (bottom left), but not the 'communication subscore' (bottom right).(G) Pearson's partial correlations revealed non-significant associations between nST performance and any each of the ADOS-2 measures, covaried by WMT scores.ADOS-2 autism diagnostic observation schedule-2, low AQ low autism spectrum quotient, high AQ high autism spectrum quotient, RSI reciprocal social interaction, SBRI stereotyped behaviors and restricted interests, Comm communication, WMT working memory task, ST social text, nST non-social text, PPVT-III Peabody Picture Vocabulary Test, RPMT Raven's progressive matrices test-III.Double asterisks (**) indicate significant differences after covariation for RPMT, WMT, and PPVT-III outcomes.

Table 1 .
Groups' demographic, autistic, and cognitive characterization.(a) p values calculated with chisquared test; (b) p values calculated with one way ANOVA.RSI reciprocal social interaction, SBRI stereotyped behaviors and restricted interests, Comm communication, WMT working memory task, PPVT-III Peabody Picture Vocabulary Test III, RPMT Raven's Progressive Matrices Test.
high AQ (