Real Talk, Virtual Faces: Symbolic-Semantic Discourse Geometry of Virtual and Human Influencer Audiences

Shahram Chaudhry; Sidahmed Benabderrahmane; Talal Rahwan

cs.CYMar 25, 2026

Real Talk, Virtual Faces: Symbolic-Semantic Discourse Geometry of Virtual and Human Influencer Audiences

Authors: Shahram Chaudhry, Sidahmed Benabderrahmane, Talal Rahwan

Organizations: New York University (NYUAD), Division of Science, Computer Science Department

Abstract

Virtual influencers~(VIs) -- digitally constructed social-media personas -- are becoming increasingly visible in online culture, marketing, and identity formation. Yet it remains unclear whether audiences respond to them through the same discourse patterns used for human influencers~(HIs), or whether virtuality produces distinctive modes of reaction. Existing studies often rely on surveys, engagement statistics, or marginal sentiment distributions, which reveal what audiences say but not how affective, topical, and psycholinguistic signals are jointly organised. We introduce a symbolic-semantic framework for analysing audience discourse around virtual and human influencers. The symbolic layer uses Formal Concept Analysis and association rule mining to extract closed co-occurrence structures from sentiment labels, topic tags, and Big Five psycholinguistic cues. The semantic layer renders these formal concepts as natural-language descriptions, embeds them with MiniLM, and compares their geometry across VI and HI audiences. Applied to 69,498 YouTube comments from three matched VI-HI influencer pairs, our analysis shows that HI discourse is organised around a compact, stability-centred pattern in which low neuroticism anchors positive sentiment, whereas VI discourse supports multiple discourse regimes. VI concepts are also more semantically dispersed than HI concepts, while both groups show strong symbolic-semantic alignment between closed-set structure and embedding geometry. Finally, VI discourse contains a distinct artificial-identity region and a higher concentration of negative sentiment in sensitive topics such as mental health, body image, and artificial identity. These findings suggest that virtuality reshapes not only the sentiment of audience reactions, but also the symbolic and semantic organisation of online social discourse.

Figures & tables

Virtual Influencer	Subscribers	Human Counterpart	Niche
Lil Miquela	$\sim$ 271k	Samantha Nicole	Fashion / lifestyle
APOKI	$\sim$ 329k	YOUNG POSSE	K-pop / performance
Milla Sofia	$\sim$ 28k	Lydia Stoner	Fashion lookbooks

Table 1: Selected VI–HI influencer pairs. HI counterparts are chosen by automated API search within a

\pm

10% subscriber window, then qualitatively aligned on content niche.

Attribute	HI	VI	$\Delta$ (VI $-$ HI)
topic_artificial_identity	0.005	0.184	$+$ 0.179
sentiment_Negative	0.067	0.125	$+$ 0.058
topic_authenticity_critique	0.037	0.062	$+$ 0.025
topic_appearance	0.159	0.170	$+$ 0.011
topic_social_comparison	0.007	0.007	0.000
sentiment_Positive	0.766	0.669	$-$ 0.097

Table 2: Attribute prevalence (selected attributes), sorted by

\Delta

. The near-equal appearance prevalence (0.011 gap) contrasts sharply with its structural asymmetry in rule mining (Section 6.2 ).

Object	$P$	$A$	$G$	$N$
$c_{1}$	✓	✓	✓
$c_{2}$	✓	✓	✓
$c_{3}$		✓		✓
$c_{4}$	✓		✓
$c_{5}$	✓	✓	✓	✓

Table 3: Toy formal context illustrating FCA. Objects are comments and attributes are binary discourse signals:

P

= positive sentiment,

A

= appearance topic,

G

= high agreeableness, and

N

= high neuroticism.

Panel of Figure 2 — Figure 2: Toy concept lattice derived from Table 3 . Each node shows a closed intent and its corresponding extent. More specific concepts appear lower in the lattice.

Analytical view	Captures	Contribution to the study
Marginal prevalence	Individual sentiment/topic/style frequencies	Establishes baseline differences and exposes cases where similar frequencies hide different structures.
FCA concepts	Closed co-occurrence bundles	Identifies stable symbolic discourse profiles that cannot be reduced to individual attributes.
Association rules	Directional co-occurrence regularities	Summarises interpretable discourse grammars through support, confidence, and lift.
MiniLM concept geometry	Semantic proximity among formal concepts	Tests whether symbolic structures correspond to coherent semantic regions.
Cluster labels	Human-readable discourse regimes	Interprets computed semantic regions without replacing the symbolic evidence.

Table 4: Analytical views in the symbolic–semantic discourse framework.

Figure 3: Semantic geometry of FCA concepts for HI and VI audiences. Each point represents a formal concept rendered as a natural-language description and embedded using MiniLM. HI and VI concepts partially overlap, while VI concepts occupy a slightly more dispersed semantic region.

Multilingual	Intent	Support	VI (raw)	VI (filtered)	HI (raw)	HI (filtered)
No	$>$ 3	0.20	1895	10	864	24
No	$>$ 3	0.10	1895	114	864	152
No	$>$ 2	0.20	1895	54	864	80
No	$>$ 2	0.10	1895	241	864	264
Yes	$>$ 3	0.20	3980	11	1269	24
Yes	$>$ 3	0.10	3980	124	1269	153

Table 5: FCA concept counts under different support, intent-size, and language-inclusion settings (Multilingual = Yes includes non-English comments). Bold row shows the selected configuration (English-only, intent

>

3, support

\geq

0.20).

Panel of Figure 4 — Figure 4: Dominant attributes in group-specific (HI-only vs. VI-only) stable FCA concepts. HI-only concepts are sentiment-centred (positive sentiment, topic positivity, low neuroticism), whereas VI-only concepts are personality-centred (high openness, high agreeableness, low conscientiousness).

Configuration	#HI rules	#VI rules
minsup = 1%, minconf = 0.8, lift $>$ 1.2	8	51

Table 6: Rule-set sizes under the approximate concept-based rule-mining procedure, after support, confidence, and lift filtering.

HI rules (8 total)
Metric	Mean	Median	Max
Support	0.062	0.065	0.090
Confidence	0.92	0.94	0.96
Lift	1.23	1.23	1.25
VI rules (51 total)
Support	0.146	0.080	0.420

Table 7: Rule summary statistics (minsup = 1%, minconf = 0.8, lift

>

1.2).

Premise $X$	Consequent $Y$	supp	conf	lift
Agreeableness_high $\wedge$ Neuroticism_low $\wedge$ topic_positivity	sentiment_Positive	0.060	0.96	1.25
Extraversion_high $\wedge$ Neuroticism_low $\wedge$ topic_positivity	sentiment_Positive	0.080	0.95	1.24
Conscientiousness_low $\wedge$ Neuroticism_low $\wedge$ topic_positivity	sentiment_Positive	0.070	0.95	1.24
Neuroticism_low $\wedge$ topic_positivity	sentiment_Positive	0.090	0.94	1.23
Neuroticism_low $\wedge$ Openness_high $\wedge$ topic_positivity	sentiment_Positive	0.090	0.94	1.23
Agreeableness_high $\wedge$ Neuroticism_low $\wedge$ sentiment_Positive	topic_positivity	0.060	0.82	1.21

Table 8: All 8 HI rules. Single cluster: topic_positivity

\wedge

Neuroticism_low anchors all rules.

Positivity-framing cluster (top rules by lift)
Premise $X$	Consequent $Y$	supp	conf	lift
Agreeableness_high $\wedge$ Neuroticism_low $\wedge$ topic_positivity	sentiment_Positive	0.050	0.96	1.43
Neuroticism_low $\wedge$ topic_positivity	sentiment_Positive	0.080	0.94	1.40
Agreeableness_high $\wedge$ Neuroticism_high $\wedge$ topic_positivity	sentiment_Positive	0.200	0.92	1.37
topic_positivity	sentiment_Positive	0.420	0.89	1.33
Appearance-discourse cluster (absent from HI)

Table 9: Selected VI rules. Top positivity-framing rules include Neuroticism_high—absent from HI. Appearance-discourse rules have no HI counterpart despite equal marginal prevalence.

minsup	minconf	#HI rules	#VI rules
0.005	0.8	22	114
0.005	0.9	14	68
0.010	0.8	8	51
0.010	0.9	6	29
0.020	0.8	3	18

Table 10: Rule counts across extraction thresholds (lift

>

1.2) for selected configurations.

Group	Symbolic–Semantic $\rho$	$p$ -value	Semantic Dispersion
HI	0.614	$<0.001$	0.0765
VI	0.640	$<0.001$	0.0882

Table 11: Symbolic–semantic geometry of FCA concepts.

\rho

is the Spearman correlation between FCA symbolic distance, computed from concept intent overlap, and MiniLM semantic distance, computed from concept-sentence embeddings. Semantic dispersion is the mean cosine distance from each concept embedding to the group centroid.

Panel of Figure 5 — Figure 5: Semantic clustering of FCA concepts. Each point corresponds to a closed concept embedded using MiniLM and coloured by semantic cluster. HI clusters are mainly organised around positivity and appearance-related engagement, whereas VI clusters include a distinct artificial-identity region and multiple appearance-related modes.

Panel of Figure 6 — Figure 6: Semantic similarity heatmaps of FCA concepts ordered by semantic cluster. Block structure indicates coherent semantic regions among closed discourse patterns. VI concepts show slightly higher semantic dispersion, consistent with a broader range of audience response modes.

Group	Cluster	Semantic label
HI	C0	Emotionally tense positive discourse
HI	C1	High-arousal expressive positivity
HI	C2	Tense appearance evaluation
HI	C3	Expressive appearance admiration
HI	C4	General positivity grammar
VI	C0	Emotionally reactive positive framing

Table 12: Semantic labels assigned to MiniLM clusters of formal concepts. Labels summarise clusters of closed itemsets after rendering concept intents as natural-language descriptions.

Group	Sentiment	E	N	A	C	O
HI	Positive	0.89	0.84	0.53	0.20	0.99
VI	Positive	0.88	0.82	0.61	0.18	0.99
HI	Neutral	0.84	0.93	0.43	0.20	0.99
VI	Neutral	0.84	0.89	0.42	0.22	0.99
HI	Negative	0.81	0.87	0.44	0.23	0.99
VI	Negative	0.81	0.88	0.33	0.27	0.99

Table 13: Average Big Five personality trait profiles by sentiment category and influencer type. Traits are abbreviated as follows: E = Extraversion, N = Neuroticism, A = Agreeableness, C = Conscientiousness, O = Openness.

Social Media Analysis

Social media analysis studies online posts, media, users, and interactions to understand communication, communities, and public behavior.

See all 221 papers

Evaluating the Realism of LLM-powered Social Agents: A Case Study of Reactions to Spanish Online News
May 27, 2026Alejandro Buitrago López, Alberto Ortega Pastor, Javier Pastor-Galindo +1LLM Agent Evaluation Large Language Model-Based Social Simulation
Faculty of Computer Science, University of Murcia, Murcia (Spain)
Socially fluent AI decouples conversational signals from source identity in online interaction
May 22, 2026Lixiang Yan, Yueqiao Jin, Xibin Han +1Human-AI Interaction Conversational Agents
School of Education, Tsinghua University, Beijing & 100084, China · Faculty of Information Technology, Monash University, Melbourne & 3800, Australia · Faculty of Education, The University of Hong Kong, Hong Kong SAR, China
How Far Did They Go? The Persuasive Tactics of Covert LLM Agents in a Discontinued Field Experiment
Jun 3, 2026Kokil Jaidka, Saifuddin AhmedLLM Auditing LLM Agent Evaluation
Department of Communications and New Media, & Centre for Trusted Internet & Community, National University of Singapore, Singapore · Wee Kim Wee school of Communication and Information, Nanyang Technological University, Singapore

May 27, 2026J/K move · Enter open · S save

Evaluating the Realism of LLM-powered Social Agents: A Case Study of Reactions to Spanish Online News

Alejandro Buitrago López, Alberto Ortega Pastor, Javier Pastor-Galindo, José A. Ruipérez-Valiente

Faculty of Computer Science, University of Murcia, Murcia (Spain)

LLM Agent Evaluation Large Language Model-Based Social Simulation

Open paper AskPDF arXiv

LLM-powered social agents are increasingly used to simulate online social behavior, yet their realism remains difficult to validate. Existing work has largely relied on general-purpose benchmarks, while less attention has been paid to short, reactive discourse such as audience replies to online news. In this paper, we evaluate whether LLM-generated reactions to Spanish online news reproduce measurable properties of real audience discourse. Using the Hatemedia dataset, we pair 5,631 news items with 58,555 real audience reactions, and generate a matched synthetic dataset using five LLMs under a shared experimental setting. We compare real and synthetic reactions across three dimensions: hate speech, sentiment, and semantic alignment, considering both off-the-shelf and fine-tuned generation. Results show that off-the-shelf models are poor proxies for real audience reactions: they strongly underproduce hate speech, introduce model-specific sentiment biases, and remain distributionally distant from human replies. Fine-tuning improves fidelity unevenly. Qwen3 provides the most balanced approximation, while Mistral7B achieves the strongest sentiment and semantic alignment but overshoots hate prevalence. Plausible synthetic replies do not necessarily reproduce the distributional properties of public discourse.

Abstract

Social Media Analysis

Explore similar work

Evaluating the Realism of LLM-powered Social Agents: A Case Study of Reactions to Spanish Online News

Socially fluent AI decouples conversational signals from source identity in online interaction

How Far Did They Go? The Persuasive Tactics of Covert LLM Agents in a Discontinued Field Experiment