Beyond Expert Judgment: An Explainable Framework for Truth Discovery, Weak Supervision, and Learning-Based Ranking in Open-Source Intelligence Risk Identification

1. Introduction

In a public communication environment driven by globalization and social media, symbolic visual symbols serve as narrative “triggers,” reshaping the public agenda through a low-threshold, high-diffusion, and strong-emotional coupling mechanism. When universities’ “open stances” intersect with local historical memories and ethnic emotional tensions, organizational self-narratives often diverge from public perceptions, creating “cultural narrative imbalances.” This triggers rapid, widespread eruptions and solidification, demanding new approaches to ritual governance and reputation management that transcend conventional judgment. Centered on the methodological axis of “Transcending Expert Judgment: An Explainable Framework for Truth Discovery, Weak Supervision, and Learning-to-Rank in Open-Source Intelligence Risk Identification,” this paper reconstructs the temporal and evidentiary chains of a high-profile campus ritual incident using open-source intelligence (OSINT). It focuses on the micro-mechanism of “symbolic trigger → narrative vacuum → emotional amplification → public sentiment solidification.” Methodologically, it employs truth discovery (replacing expert weighting) to estimate source credibility, utilizes weak supervision (Snorkel-like labeling functions and generative model aggregation) to obtain robust labels, and implements learning-to-rank to dynamically prioritize risk trajectories. Explainability techniques (SHAP, Grad-CAM, stability/monotonicity metrics) ensure transparent and auditable conclusions. In framework alignment, we algorithmically integrate ACH competitive hypothesis analysis with STEMPLES dimensionality metrics to validate their applicability and extrapolation in educational governance scenarios. Concurrently, we ensure conclusions are independent of singular thresholds or case-specific contingencies through dual-time-window robustness, Hawkes diffusion intensity, inflection point detection, and causal evaluation via ITS/synthetic controls. Thus, this paper theoretically provides computable evidence for the “narrative-symbol-diffusion” chain, methodologically proposes an explainable and reproducible OSINT risk identification pathway that replaces expert scoring with algorithms, and practically offers actionable governance recommendations for agenda management and risk response in the context of university internationalization.

1.1. Research Questions

RQ1. Under what conditions does symbolic stimulus transition from localized misinterpretation to public opinion eruption, and through which intermediary pathways (e.g., narrative vacuum, platform amplification, group polarization) does it evolve?

RQ2. Can the open-source intelligence framework (ACH + STEMPLES + behavior-dominant matrix) reliably distinguish verifiable evidence of “mismanagement/cultural misinterpretation” from “soft influence pathways,” while providing compatible causal explanations?

RQ3. For university internationalization and major ceremonial contexts, which institutional mechanisms (symbolic risk pre-screening, dual-tier approval with media testing, principle of symmetrical response) can effectively reduce outbreak risks and enhance narrative resilience?

1.2. Research Significance/Contribution

1.2.1. Theoretical Contributions

Proposing a mechanism chain centered on “symbolic triggers—narrative vacuums—emotional amplification—public opinion solidification,” this study reveals how cultural narrative imbalances drive public opinion explosions. It bridges the gap between symbolic politics, narrative studies, and computational social science through truth discovery and causal inference methods.

1.2.2. Methodological Contributions

Establishes a composite paradigm of “OSINT evidence—truth discovery—weakly supervised annotation—learning-based ranking—interpretability metrics” to replace expert subjective scoring. This enhances transparency, auditability, and cross-scenario transferability in event analysis. Integrates ACH hypothesis competition with STEMPLES scenario decomposition into an algorithmic framework, forming an expandable risk identification system.

1.2.3. Practical Contributions

Provides universities with an actionable framework and metric system for establishing symbolic political risk pre-screening and cross-departmental rapid response mechanisms in international communications and major ceremonial governance. This enables closed-loop support from early risk identification to public sentiment diffusion monitoring and threshold-based early warning.

1.3. Research Context/Organizational Background

This study examines public controversy surrounding the “red circle, white chairs” visual arrangement at a university’s opening ceremony. Drawing on open-source materials such as public reports and institutional statements, it constructs a case database and event timeline to identify mechanisms and conduct comparative validation.

1.4. Roadmap

The structure of this paper is as follows: Section II presents a literature review outlining relevant concepts and theoretical frameworks; Section III details the research methodology, data sources, and analytical framework; Section IV discusses the research questions, offers mechanism explanations, and conducts comparative analysis; Section V presents research conclusions, policy recommendations, and limitations to support future extrapolation and validation across multiple cases and contexts.

2. Literature Review

2.1. Overview of Positioning and Methodology

This review centers on the methodological framework “Beyond Expert Judgment: A Truth Discovery, Weakly Supervised, and Learning-Based Explainable Ranking Framework for OSINT Risk Identification.” It focuses on the mechanism chain of “cultural narrative imbalance—symbolic triggers—public opinion eruption,” defining the value and boundaries of OSINT in evidence construction within social science and educational governance contexts. Our concern is not interpreting individual cases through expert scoring, but rather testing whether algorithmic empowerment can replace experts in generating verifiable, scalable, and explainable risk identification evidence from multi-source heterogeneous data. To this end, the literature search primarily utilized Web of Science and Scopus, supplemented by SocArXiv, ArXiv, SSRN, BASE platforms, and Google Scholar advanced search (including reverse citation tracking). Chinese-language sources were supplemented by authoritative Chinese media reports. The timeframe spans 2004–2025 (covering the era of social media platformization), incorporating timeless foundational theories to establish conceptual frameworks. Inclusion criteria emphasized:

(1) alignment with our five-stage intermediary chain—“trigger—narrative vacuum—diffusion—polarization—solidification”;

(2) Closed-loop design-measurement-analysis chains and cross-scenario transferability; (3) Operational contributions to algorithmic expert substitution (e.g., weakly supervised learning, ground truth discovery, learning-based ranking, interpretability metrics). Quality assessment follows CASP (peer-reviewed studies) and AACODS (gray literature) frameworks, prioritizing transparency, bias control, and reproducibility as core criteria. The integrated strategy primarily employs thematic synthesis: first defining “trigger-vacuum” representations through Entman’s (1993) narrative framework, then leveraging Kasperson et al. (1988)’s risk society amplification model and Chong & Druckman (2007)’s framework to organize evidence on “diffusion-polarization.” It integrates Berger & Milkman (2012)’s behavioral mechanisms of transmissibility, Brady et al. (2017)’s moral emotion drivers, and Vosoughi et al. (2018) on disinformation diffusion, pointing to the structural characteristics of the “solidification” phase. Unlike previous reviews reliant on expert interpretation, this study prioritizes incorporating and discussing evidence from truth discovery (e.g., Dawid–Skene/GLAD), weakly supervised methods (Snorkel-like labeling functions and generative model aggregation), learning-based ranking (e.g., LambdaMART), and interpretability implementations (e.g., SHAP, Grad-CAM, stability and monotonicity metrics). This ensures the review’s conclusions can be algorithmically implemented to serve risk control and governance applications in OSINT.

2.2. Symbols and Frames: From "Reactivation" to the Appropriation of Meaning

The core of symbolic politics lies in folding complex political meanings into perceptible cues, thereby triggering the public’s reproduction of meaning at critical junctures (Edelman, 1985). The construction of national communities and organizational identities thrives on the persistent re-enactment of imagination and narrative (Anderson, 2006), while social actors’ “framing” of experience serves as the fundamental mechanism for achieving meaning and coordinating action (Goffman, 1974/1986). In communication studies, framing is defined as the selection and emphasis of certain information, thereby influencing problem definition, causal attribution, and moral judgment (Entman, 1993; Chong & Druckman, 2007). This pathway reveals how visual/ritual symbols are “reactivated” at the intersection of historical semantics and contemporary contexts, forming preemptive interpretations within brief windows that provide narrative themes for subsequent diffusion.

2.3. Amplification Chains and "Narrative Vacuums" in the Risk Society

The Social Amplification of Risk Framework (SARF) indicates that media and social sites not only amplify risk signals themselves but also alter behavioral, organizational, and policy responses through interpretation and interaction (Kasperson et al., 1988). In the platform era, this framework has been further expanded: what is amplified is not a single “danger,” but rather emotional meanings and contextual interpretations, with consequences that reshape institutional trust, organizational reputation, and group boundaries (Kasperson et al., 2022). If organizations fail to provide a clear, authoritative, and verifiable “primary frame” during the early window, a “narrative vacuum” emerges. This allows fragmented evidence to be rapidly assembled into the most accessible, emotionally charged story, thereby increasing the cost of subsequent corrections. This insight aligns with the crisis communication principle of “first-mover response—reputation protection” (Coombs, 2007, 2014).

2.4. Platform Mechanisms: Emotion-Driven, Algorithmic Amplification and Misinformation Diffusion

On social platforms, highly arousing emotions (such as surprise or anger) significantly increase the likelihood of sharing (Berger & Milkman, 2012), while the use of moral-emotional words exhibits a robust positive correlation with information diffusion (Brady et al., 2017). At the structural level, platforms may expose users to diverse information while simultaneously fostering a coexisting landscape of “pseudo-diversity” and “selective exposure” due to homophily preferences and ranking mechanisms (Bakshy et al., 2015). Across multiple thematic domains, misinformation often spreads faster, deeper, and wider than truthful information, its transmissibility stemming from dual advantages of novelty and emotionality (Vosoughi et al., 2018). Collectively, these findings indicate that platforms amplify the initial disturbance triggered by “symbolic cues” through the coupling of emotion and algorithms. This amplification follows a pathway of shareability-visibility-credibility, thereby increasing the subthreshold probability of public opinion “ignition.”

2.5. Attitudinal Polarization and Group Dynamics: From Consensus to Division

Groups are prone to attitude polarization within echo chambers, where peer coordination drives perspectives toward more extreme positions while amplifying moral certainty and identity markers (Sunstein, 2009). Under this mechanism, controversies sparked by cultural symbols transcend mere “right-wrong judgments,” instead serving to reaffirm group boundaries and moral order. When narrative vacuums coexist with platform amplification, polarization accelerates the “us/them” dichotomy, solidifying sporadic symbolic conflicts into enduring social fractures.

2.6. OSINT’s Place in Social Sciences and Education Governance

OSINT utilizes legally obtainable open-source materials, emphasizing principles of verifiability, traceability, and proportionality. Social media intelligence (SOCMINT) further integrates platform data into the workflow (Omand et al., 2012). The strengths of social science and educational governance lie in OSINT’s ability to rapidly identify critical nodes in the “trigger-vacuum-amplification” chain through timeline reconstruction, triangulation verification, and weak signal aggregation. Its limitations include difficulty accessing private communications and implicit resource flows, with causal identification relying on cross-validation with other methodologies (Olcott, 2012). When integrated with the Situational Crisis Communication Theory (SCCT) for crisis communication, this forms a governance loop comprising “front-end early warning (OSINT) – mid-stage response (SCCT strategy alignment) – post-crisis review (evidence archiving)” (Coombs, 2007, 2014).

2.7. Governance Vulnerabilities and Institutionalized Responses in Educational/Ritual Scenarios

University ceremonies possess a structural characteristic of high exposure with low redundancy review, making them highly sensitive to symbolic politics. Without conducting symbolic risk pre-reviews (including assessments of visual elements’ potential for misinterpretation within specific historical contexts), media and community usability tests (A/B testing among target audiences and critical communities), and dual-tiered response plans (technical explanations + value/position reaffirmations), the chain reaction of “narrative vacuum—platform amplification—polarization” becomes highly likely. Evidence indicates that timely, authoritative responses aligned with the intensity of responsibility attribution can effectively mitigate negative reputation spillover and reduce the severity of policy reactions (Coombs, 2007, 2014). Methodologically, embedding OSINT processes into pre-event reviews and post-event debriefings for major activities helps shift governance from “reputation firefighting” to “risk anticipation.”

2.8. Reviews and Research Frontiers

Cross-literature consensus holds that cultural symbols and historical semantics form transmissible meaning packages through frame selection and affect coupling, with platform structures and social psychological mechanisms jointly determining their amplification scale (Entman, 1993; Berger & Milkman, 2012; Brady et al., 2017; Bakshy et al., 2015; Vosoughi et al., 2018). Existing research exhibits three primary shortcomings: First, operational measurements of the “narrative vacuum” remain inadequate, necessitating the development of interpretability indices tailored to the early window of events; Second, the interaction between platform causation and user choice lacks robust, cross-data, cross-national identification designs; Third, ethical governance and bias mitigation for OSINT in academic contexts require more systematic procedural constraints and ex-post verifiable mechanisms (Omand et al., 2012; Olcott, 2012). Looking ahead, we recommend conducting multi-context, multi-platform comparative studies that integrate affective semantic modeling, network counterfactual simulation, and causal mediation analysis. Furthermore, linking OSINT evidence pipelines with SCCT policy choices for joint testing will validate the extrapolation of the “trigger-vacuum-diffusion-polarization-solidification” chain model across diverse institutional and cultural contexts.

3. Research methodology

This study relies entirely on open-source intelligence (OSINT) materials, with core data sourced from Wuhan University’s “Sun Flag Incident” Open-Source Intelligence Analysis Report. This report compiles publicly available records from the university and media outlets. Data acquisition employed a stratified-quota sampling approach with deduplication: (i) Institutional authoritative anchors (campus news portals, faculty announcements) established factual baselines and event timelines; (ii) Official statements were cross-referenced from authoritative media to ensure consistency of key information; (iii) Mainstream and international reports were selected as Top-K within each discourse domain based on interaction volume and citation frequency, serving cross-domain validation and diffusion analysis; (iv) Visual materials (ceremony photos and video frames) underwent perceptual hash clustering for deduplication, retaining only representative frames for symbol recognition. All materials originate from the public domain. Access dates were recorded during retrieval, and de-identification was completed prior to analysis. This study utilized no restricted, private, or non-public data. It should be noted that Wuhan University subsequently removed certain relevant web links; however, this paper retains and utilizes data acquired before deletion to ensure research integrity and reproducibility.

3.1. Overview of the Methodology and the Principle of Consistency

This study employs an “algorithm model + open-source intelligence (OSINT)” framework, replacing expert subjective assessments with a closed-loop system of data-algorithm-metrics-decision-making. This approach ensures both governance and academic objectives are achieved under conditions of verifiability, reproducibility, and interpretability. The entire process aligns with existing research questions and conclusions: the primary cause of events better aligns with “misinterpretation of cultural symbols + insufficient narrative response.” Continuous monitoring and quantitative verification of “soft influence pathways” are conducted to avoid elevating conclusions to systemic or high-level penetration when evidence is insufficient (aligning with the research scope and logical chain of the submitted manuscript). End-to-end process: Research context/problem → Objective functions and design → Sampling → Data acquisition (public sources) → Algorithmic analysis (temporal/textual/visual/network/causal) → Automated evaluation and threshold selection → Decision mapping and presentation. Managerial multi-objective functions (no expert scoring required):

· J(Θ) = α₁ ·Prec + α₂ ·Earliness + α3 ·Explain_auto − α4 ·Cost

Where Prec is the accuracy of critical event/inflection point identification, Earliness is the lead time relative to the organization’s response; Explain_auto is the explanatory automated metrics, Cost is the collection and computation/compliance cost; α is determined by the task constraints, and the Pareto frontier selection scheme is used.

3.2. Stratification-Quotas and Active Learning

To balance verifiability and cost-effectiveness, we employ a “stratified quota sampling + active learning” strategy for data collection and annotation, establishing reproducible stopping rules. On the sampling side, we construct a four-tier sample pool based on information hierarchy and diagnostic value: S3 (mainstream/overseas reports) employs Top-K selection based on objective, comparable metrics like engagement/repost counts (intra-domain sorting to suppress site clusters and redundancy), ensuring representativeness for cross-domain validation and diffusion observation; S4 (visual materials) is fully collected with deduplication within the primary window. Perceptual hashing clustering is applied to near-duplicate shots from the same camera angle, retaining only one representative image per sequence to control sample correlation and stabilize symbol feature distribution. For annotation, an active learning loop is introduced: limited annotation resources are adaptively allocated to text and image samples with highest uncertainty (based on marginal confidence intervals, entropy, or divergence metrics). This maximizes marginal model gains with minimal human effort while avoiding oversampling in low-information-density regions.

The training process employs dual stopping criteria to ensure robustness and efficiency: First, if the classifier’s PR-AUC fails to improve by >0.5% over three consecutive rounds, performance convergence is declared, halting data expansion/retraining. Second, iteration stops when the absolute mean absolute error (MAE) for key inflection points in the temporal module stabilizes below 1 hour and shows no significant drift over three consecutive rounds. This mechanism establishes a consistent, reproducible threshold system across sampling, annotation, and training, enabling the resulting model to achieve a quantifiable optimal balance between representativeness, efficiency, and robustness.

4. Source Reliability and Verifiability (Algorithmic Substitution of Expert Weights)

To avoid subjective weighting by experts, the “Source Reliability × Verifiability” metric employs a combined estimation of truth discovery + network reputation + temporal consistency to produce the source weight w_k:

4.1. Truth Discovery (Deliberative Model)

Construct a “Source-Statement” bipartite graph and use Dawid–Skene/GLAD-like models for unsupervised estimation of source accuracy, yielding the initial credibility ri.

4.2. Network Reputation (Citation/Repost Graph)

Calculate authority ai using HITS/PageRank on cross-platform repost graphs; retain only the first/most comprehensive version for same-source reposts to suppress redundancy.

4.3. Temporal Consistency and Correction Rate

Include penalties pi based on source statements’ temporal consistency before/after T₀, T₁ and “correction/retraction” records.

4.4. composite Weight

wi IX . . (1 − pi), β, γ Determined by the extrapolation accuracy of the leave-out set maximization (non-expert setting).

5. STEMPLES Automatic Extrapolation (to Quantitative Alternative Expert Scoring)

Mapping STEMPLES to z ∈ R⁸ is entirely generated by rules + statistical learning, without manual scoring:

S (Symbolism): Visual symbol conflict = HSV high-saturation red proportion + shape saliency (Hoff circle/closed edges) + ViT binary classification (Grad-CAM hotspot coverage).

T (Timing): Temporal distance from commemorative/sensitive nodes (normalized); T value increased if concurrent national commemorative events exist in news coverage.

E (Event): Proxy for event scale/exposure (S₁/S₃ coverage volume + social media visibility) and venue characteristics (stadiums/squares, etc.).

M (Information): Availability of official information (S₁/S₂ dissemination breadth, headline interpretability n-gram score).

P (Process): Keyword frequency (“rehearsal/test/availability”) in S₁/S₂/S₃; availability of flowcharts or timelines.

L (Leadership): Response hierarchy × response time (Party Secretary/Principal/Department Head × T₁ − T₀).

E (Externalities): Cross-border media/organizational engagement (proportion of overseas coverage in S3; intermediary role of external accounts in the network).

S (Structure): Echo chamber index and community fragmentation of the opinion network (Louvain modularity); high fragmentation + high polarization increases S.

6. ACH → MCDA(Automated Evidence Aggregation as an Alternative to Expert Judgement)

Four competing hypotheses were retained: H1 (mismanagement/cultural misinterpretation), H2 (soft influence pathways), H3 (systemic agendas), and H4 (deep penetration). Instead of using expert scoring, the following automated scheme was used:

6.1. Weak Surveillance Labeling

Weak labels (Snorkel-style labeling function) are generated using explicit expressions of S1/S2 and cross-source consistent descriptions for "support/rebuttal/neutral".

6.2. Diagnostic Scorer

For each piece of evidence ek compute sh(x k) ∈ [-1, 1] (textual/visual/web features input to the multitask classifier, output support for each Hh).

6.3. Source Weight Fusion

Sh = ∑k wksh(x k), wk Algorithmic weights inferred from algorithmic weights (alternative expert weights).

6.4. Calibration of Weights

Using Learning-Sorting (LambdaMART/LogitRank) Learning the combined weights that optimally distinguish the final outcome (whether it escalates to a sizeable public opinion/whether it is corrected by an authority) on the leave-out history event, or integrating multiple sub-models with Bayesian model averaging (BMA) of the Sh.

6.5. consistency Test

Correlation and Granger test for {S_h} with z (§5); if "High Symbolic Conflict + Low Process Completeness + High Timing Sensitivity" is significant, it corresponds to an increase in S_H1 and does not lead to an increase in S_H3/4.

7. Pure Algorithmic Paths for Text, Vision, Networks and Causation

7.1. Time and Inflection Points

PELT/BOCPD does change-point detection of visibility, emotional intensity, and explanation availability (EAI, see below) sequences to identify trigger-response-secondary peaks;

Interruption Timing System (ITS) Estimation of level/slope changes with T1 as breakpoint to quantify the "narrative vacuum" convergence effect;

Robustness: repeat in main window/robustness window; sensitivity recalculation of sampling thresholds to platform changes.

7.2. Text Semantics and Narrative Monitoring

Frame Recognition (Multi-label): Train the supervised fine-tuning model based on Entman’s four elements (Small Sample + Active Learning), and output "Problem Definition/Causal Attribution/Moral Judgment/Action Guidance" at the post level;

Emotion and moral semantics: Chinese emotion classifier + MFD-CN lexicon calibration to get "moral-emotional intensity";

EAI (Explanation Availability Index): the proportion of "Explainable Content" over "Highly Communicated Unexplained Content" in the early window:

$${\mathrm{EAI}}_{\left(t\right)}={\displaystyle \frac{\mathrm{C}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{e}\mathrm{d}\left(t\right)}{\mathrm{T}\mathrm{o}\mathrm{p}\mathrm{K}\_\mathrm{U}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{d}\left(t\right)}}\in [0,1],{\mathrm{R}\mathrm{i}\mathrm{s}\mathrm{k}}_{\mathrm{v}\mathrm{a}\mathrm{c}\mathrm{u}\mathrm{u}\mathrm{m}}(t)=1-\mathrm{E}\mathrm{A}\mathrm{I}(t)$$

8. Automated Interpretability and Threshold Selection (De-Expert Scoring)

8.1. Interpretability Metrics (Explainauto)

Text/Regression:Sufficiency/Comprehensiveness, Infidelity and SHAP stability (ℓ2 deviation across random seeds/resampling);

Visual: Grad-CAM consistency (IoU of hot zones of similar images), area under the deletion-insertion curve;

Uniform scoring: Weighted summation after Min-Max normalization, written back to the objective function as Explain_auto, completely replacing manual expert scoring.

8.2. Thresholds and Warning Levels

ROC-Youden is used to select risk thresholds in conjunction with management cost curves; mapping of warning levels to maximize expected utility is automatically determined on a leave-behind window, eliminating the need for expert review.

9. Risk Function and Programmatic Mapping

Combine the outputs of the modules to construct an event risk index:

Risk(t) = σ(w1 (1 − EAI(t)) + w2 · MoralEmo(t) + w3·Visib˙ility(t) + w4 ·SymbolScore(t) + w5 ·ResponseLag(t))

where w is calibrated by historical events or leave-out sets (Bayesian/learning-ranking) and does not rely on subjective expert assignments.

Policy mapping (automatic): when Risk(t) is combined with z (STEMPLES) trigger rules, the system gives three types of actions:

1) Prevention: high symbolic conflict + low process completeness → start "symbolic risk pre-qualification + usability A/B testing";

2) Window Response: EAI decreases and Visib˙ility > 0 → push "high-level, fact + value symmetric" response templates;

3) Review: Use STEMPLES radar difference and ACH evidence heat map to locate areas of improvement and update SOPs and threshold tables.

10. Indicator Systems, Replication and Compliance

To ensure scientific consistency across evaluation, reproducibility, and compliance, we established a hierarchical metric system with engineering safeguards: - In classification and detection dimensions, Precision/Recall/F1 and PR-AUC serve as primary performance metrics, accounting for category imbalance and threshold robustness; For inflection point identification, we jointly evaluate accuracy and recall using Mean Absolute Error (MAE) and coverage rate. For diffusion and causality analysis, we utilize the Hawkes branch ratio n to characterize self-propagation intensity. Interrupted Time Series (ITS) estimation measures level and slope effects along with their confidence intervals, while post-response convergence half-life quantifies sentiment decay velocity. For early detection, evaluate timeliness from trigger to alert using average lead time (hours/days). For interpretability, employ automated metric sets (e.g., textual sufficiency/comprehensiveness, infidelity, and SHAP stability; visual Grad-CAM consistency and delete-insertion AUC) to ensure balanced causal insights and readability. To achieve reproducibility, the entire workflow locks dependencies and random seeds via container images, providing a “one-click pipeline” (ETL → Features → Model → Evaluation → Reporting) alongside data/ code/model DOI archiving to support third-party replication and auditing. To meet compliance requirements, only publicly available data is used with de-identification implemented. Access dates and sources are logged individually, usage restrictions are clearly defined, and deletion policies with access auditing are configured. This establishes a closed research and governance chain balancing rigorous quantitative evaluation, transparent engineering practices, and robust ethical oversight.

11. Limitations and Alignment

Algorithmic substitution reduces expert subjectivity, but is still limited by the coverage and quality of open sources; weakly supervised labeling may contain systematic biases, which are mitigated by multi-source consistency and causal robustness tests. The overall methodology is consistent with the uploaded manuscripts in terms of core judgments and policy points: we prioritize explaining the mechanism of "misinterpretation of cultural symbols + insufficient management of narratives", and monitor the "soft influence pathways" with quantitative evidence to avoid over-inferences in case of insufficient evidence.