PaperlayerView exact revision on arXiv

Source-linked AI summary

BERTopic-Virality Prioritisation: A Scalable Framework for Thematic and Comparative Analysis of COVID-19 and Monkeypox Misinformation on Twitter

Mkululi Sikosana, Sean Maudsley-Barton, Oluwaseun Ajao

arXiv:2608.15691v1cs.CLcs.LGcs.SI

TL;DR

Health misinformation research needs methods that handle sparse short texts while comparing narratives across outbreaks. This paper introduces BERTopic-VP, which combines semantic clustering with virality-aware prioritisation and achieves strong classification performance while surfacing high-impact clusters.

  • Problem

    Existing methods struggle with sparse short-text misinformation, while direct comparative evidence across disease outbreaks remains limited.

  • Method

    BERTopic-VP combines contextual embedding-based clustering with post hoc virality ranking and a hybrid classifier using supervised content and external verification signals.

  • Results

    Across datasets, classification reached F1 0.950 and ROC-AUC 0.989, while VP identified high-impact clusters including low-volume narratives with disproportionate diffusion potential.

  • Takeaways & Limitations

    Coupling semantic topic discovery with virality prioritisation can help analysts surface low-volume, high-risk misinformation narratives before they dominate attention.

  • Takeaways & Limitations

    Incomplete engagement metadata limited interaction-aware analysis, and two datasets required a model-based propensity-to-spread proxy instead of raw platform counts.

Abstract

from arXiv · show

Health misinformation circulating during pandemics can gain traction rapidly, creating harmful narratives that compete with public health guidance. Most topic-modelling pipelines treat engagement as an external outcome, limiting their ability to prioritise semantically coherent topics that are also rapidly diffusing. We introduce BERTopic-VP, a virality-prioritised topic-modelling framework that combines contextual embedding-based clustering (BERTopic) with a post hoc Virality Prioritisation (VP) layer. The pipeline is complemented by a two-stage hybrid misinformation detection module that fuses a supervised content-based classifier with an external verification signal derived from public-health knowledge bases. Applied to three benchmark datasets, COVID-19_FNIR, Monkeypox, and Constraint, the framework achieves strong classification performance, with F1 up to 0.950 and ROC-AUC up to 0.989, while identifying high-impact clusters under top 1%, 5%, and 10% VP thresholds. For datasets without native engagement metadata, prioritisation is based on a logistic propensity-to-spread score, used as an ordinal proxy for diffusion potential rather than a direct measure of engagement. The results show that integrating semantic structure, virality-aware ranking, and affective-linguistic profiling enables scalable and interpretable comparative analysis of misinformation across pandemics. The proposed framework supports monitoring-oriented early warning by surfacing low-volume but high-risk narratives for analyst review.

I. INTRODUCTION … C. RESEARCH OBJECTIVES

The paper introduces BERTopic-VP to model, compare, and prioritise rapidly spreading health-misinformation themes in short Twitter texts across pandemics. It combines contextual topic discovery with virality-aware ranking and evaluates the resulting pipeline through coherence, clustering, interpretability, and early-detection analyses.

  • A. PROBLEM STATEMENT & RESEARCH GAP: Traditional methods struggle with Twitter’s short, noisy, rapidly changing vocabulary, whereas embedding-based models better capture semantic patterns in microtexts.The paper specifically contrasts conventional approaches such as LDA and rule-based classifiers with embedding-based models for posts under 280 characters.
  • I. INTRODUCTION: Contextual embeddings and density-based clustering group new themes rapidly, enabling officials to identify, compare, and track misinformation narratives across periodic refresh windows.The introduction describes minute-to-hour refresh cadences and monitoring-oriented topic tracking under batch refresh settings.
  • I. INTRODUCTION: BERTopic-VP extends BERTopic with Virality Prioritisation to support monitoring and triage of emerging health-misinformation topics.The framework augments semantically clustered topics with available engagement signals, including likes, retweets, replies, and quotes.
  • A. PROBLEM STATEMENT & RESEARCH GAP: BERTopic-VP addresses this gap by combining semantic clustering with virality signals to prioritise rapidly spreading topics and support early warning for high-volume, noisy streams.The proposed pipeline is intended to inform surveillance dashboards, policy response, and public-health decision-making.
  • A. PROBLEM STATEMENT & RESEARCH GAP: Existing systems often debunk known claims without explaining how misinformation emerges and spreads dynamically, while comparative analysis across pandemics remains limited.The paper contrasts SARS-era communication channels with globally networked platforms that amplified COVID-19 falsehoods.
  • B. HYPOTHESIS & METHODOLOGICAL MOTIVATION: The authors hypothesise that BERTopic-VP produces more granular, interpretable, and semantically coherent micro-topics than LDA, NMF, and TF–IDF+k-means under short-text sparsity.The rationale is that contextual BERT embeddings and density-based clustering do not require repeated token usage and reduce narrative merging.
  • C. RESEARCH OBJECTIVES: The study develops a semantic-embedding and virality-signal pipeline that extracts, clusters, and prioritises misinformation topics from pandemic Twitter data.This is the first primary research objective.
  • C. RESEARCH OBJECTIVES: Evaluation covers coherence, distinctiveness, stability, early detection, clustering quality, interpretability, and comparative narrative analysis of COVID-19 and 2022 mpox outbreaks.Classical baselines are contextual references rather than re-implemented experiments, and preliminary screening qualitatively observed mixed themes in baseline topics.

D. CONTRIBUTION & NOVELTY · 1) Statement of Novelty · II. STATE OF THE ART

The paper introduces BERTopic-VP as a virality-aware extension of BERTopic and applies it in a consistent comparative pipeline for COVID-19 and mpox misinformation. Its novelty lies in combining semantic clustering, explicit virality prioritisation, verification-oriented monitoring, and interpretable evaluation.

  • D. CONTRIBUTION & NOVELTY: BERTopic-VP integrates semantic clustering with virality-aware prioritisation and supports comparative misinformation analysis across COVID-19 and mpox outbreaks.The authors also describe a modular, monitoring-oriented pipeline supporting periodic batch refreshes, while noting that full scalability and live-streaming performance were not evaluated.
  • D. CONTRIBUTION & NOVELTY: BERTopic-VP uses observed interaction counts where available and a learned propensity score when datasets lack native engagement fields.The Monkeypox dataset provides observed interaction counts, whereas COVID–19_FNIR and Constraint use learned proxy propensity scores.
  • D. CONTRIBUTION & NOVELTY: The study presents a systematic side-by-side comparison of COVID–19 and mpox misinformation themes using one consistent BERTopic-VP framework.The reported thematic contrast includes COVID–19 conspiracies such as 5G and laborigin, while mpox narratives concern public trust, response failure, and stigma.
  • D. CONTRIBUTION & NOVELTY: Top 1% of retweet counts are flagged as viral, and clusters with concentrated virality are surfaced to distinguish high-impact misinformation from background noise.The prioritisation uses retweets R(T), replies Q(T), likes L(T), and link clicks as virality signals for early intervention.
  • 1) Statement of Novelty: The framework makes virality scoring v_j an explicit prioritisation layer that surfaces high-impact topics for analyst review and downstream verification.This unifies content, spread dynamics, and thematic analysis rather than treating virality metrics solely as pre-filters.
  • 1) Statement of Novelty: The pipeline evaluates predictive reliability with Accuracy, Precision, Recall, F1, ROC–AUC, and PR–AUC, alongside coherence, linguistic complexity, and emotion profiling.This dual evaluation protocol is intended to make outputs computationally reliable and thematically interpretable.
  • 1) Statement of Novelty: Dataset- and threshold-specific VP analysis supports comparisons under x ∈{1, 5, 10}% cut-offs applied to the unified signal ˜E.The framework also describes an adaptive feedback loop for recalibrating prioritisation thresholds as discourse shifts over time.
  • II. STATE OF THE ART: Compared with LDA’s sparse and often incoherent short-text topics, BERTopic uses contextual sentence vectors to capture semantic relationships between tweets.The paper positions BERTopic-VP as extending this embedding-based approach with post hoc engagement overlays for dynamic prioritisation based on amplification potential.

A. MISINFORMATION IN PREVIOUS PANDEMICS · III. MATERIALS AND METHODS · 1) Data split and evaluation protocol

The paper frames pandemic misinformation as widespread and technologically evolving, then introduces BERTopic-VP as a scalable framework for comparative, virality-aware surveillance. It evaluates the framework offline using fixed configurations, stratified validation, and engagement-derived or transferred proxy prioritisation signals across three datasets.

  • A. MISINFORMATION IN PREVIOUS PANDEMICS: Misinformation has accompanied major epidemics and pandemics, commonly involving disease origins, transmission, treatments, and miracle cures.The reviewed outbreaks since 2000 included SARS, H1N1, Ebola, Zika, MERS, and COVID–19.
  • A. MISINFORMATION IN PREVIOUS PANDEMICS: Misinformation channels evolved from email, SMS, and word of mouth during SARS to social-media platforms such as Facebook and Twitter during Ebola.The SARS example included unverified folk beliefs that firecrackers could ward off SARS.
  • A. MISINFORMATION IN PREVIOUS PANDEMICS: Direct cross-outbreak comparisons remain limited because existing studies commonly examine individual disease events.The paper identifies this limitation across prior COVID–19, Zika, and rapid-review research.
  • A. MISINFORMATION IN PREVIOUS PANDEMICS: BERTopic-VP addresses this gap through comparative, engagement-aware analysis of COVID–19 and mpox Twitter discourse, prioritising topics with high virality potential.The framework supports tracking persistent and emergent narratives across contexts.
  • III. MATERIALS AND METHODS: The modular pipeline extends BERTopic with a virality-prioritisation overlay that ranks topic clusters using engagement-derived signals or transferred proxy signals.It uses a fixed configuration across COVID–19_FNIR, Constraint, and Monkeypox, while evaluation is performed offline on static datasets.
  • 1) Data split and evaluation protocol: The evaluation uses a fixed, stratified 80/20 train–test split with a fixed random seed, five-fold cross-validation on training data, and final metrics on the held-out test split.Random seeds are also fixed for stochastic components, including embedding projection and clustering initialisation where applicable.
  • 1) Data split and evaluation protocol: Only Monkeypox provides observed tweet-level engagement fields, so COVID–19_FNIR and Constraint use a transferred propensity-to-spread proxy derived from Monkeypox.The proxy is used because those released schemas lack native engagement metadata.

A. TOPIC MODELLING VIA BERTOPIC-VP · 1) Hyperparameter selection · B. PIPELINE STEPS

BERTopic-VP combines contextual embedding, dimensionality reduction, density-based clustering, and post hoc virality prioritisation, with fixed clustering settings applied across three datasets. Its pipeline handles observed or proxy diffusion signals, fuses misinformation detection channels, and evaluates clusters using percentile-based viral density, coherence, linguistic complexity, and emotion profiles.

  • A. TOPIC MODELLING VIA BERTOPIC-VP: BERTopic combines Sentence-BERT embeddings with UMAP dimensionality reduction and HDBSCAN clustering for thematic modelling of each dataset.The encoder was all-MiniLM-L6-v2, producing uncased 384-dimensional sentence embeddings.
  • A. TOPIC MODELLING VIA BERTOPIC-VP: BERTopic-VP adds post hoc engagement or propensity-to-spread scores and aggregates them by cluster to prioritise high-diffusion themes.Observed signals include likes, retweets, and replies where available; otherwise, the Step 3 proxy is used.
  • 1) Hyperparameter selection: HDBSCAN selection used a held-out grid search over min_cluster_size ∈ {20, 50, 80} and min_samples ∈ {5, 10, 20}, optimising Cv alongside cluster stability and bounded noise.The selected configuration was min_cluster_size=50 and min_samples=10, then held fixed across COVID–19_FNIR, Monkeypox, and Constraint.
  • B. PIPELINE STEPS: The pipeline ingests tweets with metadata, preprocesses them into cleaned representations, then embeds, reduces, clusters, and induces a topic partition.These stages form the semantic modelling backbone before engagement scoring and downstream analysis.
  • B. PIPELINE STEPS: Observed engagement is computed as Ei = li + ri + qi + si (+ ci where available), using only recorded components and equal aggregation weights.Missing fields are treated as unobserved rather than behaviourally equal to zero.
  • B. PIPELINE STEPS: For datasets lacking engagement metadata, ηi or ˆpi provides an ordinal within-dataset ranking signal rather than a calibrated engagement or virality magnitude.Because σ(·) is monotonic, ranking by ηi is equivalent to ranking by ˆpi, and cross-dataset scale equivalence is not assumed.
  • B. PIPELINE STEPS: VP flags tweets in the top x ∈{1, 5, 10} percentiles of dataset-specific prioritisation scores and evaluates clusters using viral density, Cv, linguistic complexity, and emotion profiles.V(Cj) is the within-cluster share of VP-flagged tweets and diagnoses whether high-impact activity is distributed or concentrated.
  • B. PIPELINE STEPS: α = 0.6 weights the content-based misinformation classifier at 60% and the verification channel at 40% after validation-fold grid search.The grid search optimised macro-F1, with ROC–AUC as a secondary criterion.

C. OUTBREAK-SPECIFIC VERIFICATION RESOURCES

The verification channel uses outbreak-specific reference sets to avoid domain mismatch, assigning each tweet to COVID or MPX and linking it to the corresponding public-health repository.

  • Verification design: The verification channel conditions on an outbreak-specific reference set to avoid domain mismatch across outbreaks.This makes verification resources specific to the outbreak domain.
  • Verification design: Each tweet’s outbreak domain is represented as d(t_i) ∈ {COVID, MPX}.The domain variable distinguishes COVID from MPX content.
  • Verification design: The domain-specific repository R_d(t_i) provides the corresponding public-health verification resource for each tweet.Repository selection is conditioned on the tweet’s outbreak domain.

6) Adaptive feedback loop for VP recalibration

The framework uses an adaptive feedback loop at each refresh cycle to recalibrate virality aggregation and prioritisation thresholds while keeping topic-discovery components fixed. However, recalibration is deployment logic only in this paper, and richer normalised or rate-based extensions are excluded because required metadata are unavailable.

  • Adaptive feedback loop: At each refresh cycle, the loop updates virality aggregation weights and the prioritisation threshold τt within a sliding window.The refresh interval is denoted ∆t, and the sliding window is Wt.
  • Fixed components: Topic discovery hyperparameters for embeddings, UMAP, and HDBSCAN remain fixed across datasets without dataset-specific retuning.This separates semantic topic structure from the adaptive prioritisation components.
  • Updated components: Equal aggregation weights are used in this paper, so Ei reduces to the unweighted composite when native engagement is observed.For datasets without native engagement metadata, thresholding instead uses the propensity-derived proxy signal ˜Ei.
  • Thresholding: Offline VP evaluation uses thresholds x ∈{1, 5, 10}%, while deployment would flag approximately the top q fraction under the current score distribution.The Monkeypox propensity model uses q = 0.10 to define the high-engagement class.
  • Deployment limitation: Online recalibration cycles are not executed, so τt and wt are presented as deployment logic rather than applied updates.The reported results are offline evaluations.
  • Optional extensions: Normalised engagement and time-rate extensions are not applied because follower, timing, and engagement-timestamp fields are inconsistently unavailable across the released datasets.These missing fields prevent comparable audience normalisation and true diffusion-velocity measurement.

D. SCALABILITY & MONITORING-ORIENTED DESIGN. … A. UNSUPERVISED CLUSTERING PERFORMANCE

BERTopic-VP is designed for scalable, periodic health-misinformation surveillance, combining modular processing with virality-informed prioritisation for analyst triage. Across three datasets, the study evaluates clustering, classification, and VP outcomes while identifying dataset-specific clustering strengths and limitations.

  • D. SCALABILITY & MONITORING-ORIENTED DESIGN.: Batch processing supports configurable refresh cadences, including hourly or daily reruns, for surveillance-style monitoring.Embedding, dimensionality reduction, clustering, and enrichment can be profiled and scaled independently.
  • E. METHODOLOGICAL SUMMARY: BERTopic-VP couples semantically rich topic modelling with virality-informed prioritisation for periodic analyst triage and early narrative surfacing.The pipeline is described as scalable and engagement-aware for health-misinformation surveillance.
  • IV. RESULTS: The evaluation applies BERTopic-VP to COVID–19_FNIR, Constraint, and Monkeypox, reporting clustering quality, supervised classification, and VP outcomes comparatively.The three corpora are evaluated across the study’s main performance dimensions.
  • IV. RESULTS: Constraint cluster partitions are more sensitive to sampling variation and HDBSCAN noise assignment than those of COVID–19_FNIR and Monkeypox.Threshold-based VP prevalence is therefore not reported quantitatively for Constraint; qualitative VP case studies are used instead.
  • A. UNSUPERVISED CLUSTERING PERFORMANCE: Monkeypox achieved the highest purity and NMI, whereas COVID–19_FNIR had moderate purity and the lowest coherence.The comparison characterizes Monkeypox as moderately coherent and COVID–19_FNIR as noisier and more heterogeneous.
  • A. UNSUPERVISED CLUSTERING PERFORMANCE: Constraint had the lowest purity but the highest average coherence, indicating tight intra-topic semantics that aligned less with available labels.These clustering metrics are treated as descriptive rather than directly comparable because of Constraint’s instability.
  • A. UNSUPERVISED CLUSTERING PERFORMANCE: 112 topics and Cv = 0.084 characterized COVID–19_FNIR, reflecting a fine-grained partition with modest coherence in sparse short-text settings.The reported coherence pattern is consistent with coherence compression in sparse short-text data.

B. SUPERVISED MISINFORMATION DETECTION · C. EMOTIONAL TONE, LEXICAL SIMPLICITY, & · D. VIRALITY PRIORITISATION RESULTS

The study combines strong held-out misinformation classification with affective and lexical profiling, then ranks semantically coherent narratives using dataset-specific virality signals and percentile-based VP thresholds. Comparisons remain qualitative or within-dataset percentile-based when engagement measures differ.

  • B. SUPERVISED MISINFORMATION DETECTION: ROC–AUC and PR–AUC exceeded 0.95 across all datasets on held-out test splits, with Monkeypox achieving the highest accuracy and F1.Hyperparameters were selected by five-fold cross-validation within the training data.
  • C. EMOTIONAL TONE, LEXICAL SIMPLICITY, &: VP-elevated clusters consistently overlapped with strong emotional charge and simpler language, reflected by high cluster virality v_j and secondarily higher viral density V(C_j).A vaccine-rollout mpox cluster exemplified this combination of heightened affect and broad-audience readability.
  • C. EMOTIONAL TONE, LEXICAL SIMPLICITY, &: Mpox misinformation showed the strongest emotional activation, with fear 0.226, anger 0.193, and sadness 0.152.These values exceeded COVID–19_FNIR and Constraint means, especially in VP-prioritised clusters.
  • C. EMOTIONAL TONE, LEXICAL SIMPLICITY, &: Mpox tweets were more readable than COVID–19_FNIR tweets, scoring 55.73 versus 30.26 in Flesch Reading Ease and 8.90 versus 12.95 in grade level.The paper links simpler language and lower cognitive load with faster processing and sharing, while describing COVID–19_FNIR as denser and less readable.
  • D. VIRALITY PRIORITISATION RESULTS: For Monkeypox, VP uses observed engagement E_i; for COVID–19_FNIR and Constraint, it uses the logistic propensity-to-spread score p̂_i.The proxy is an estimated diffusion propensity rather than realised engagement.
  • D. VIRALITY PRIORITISATION RESULTS: VP flags tweets in the top 1%, 5%, or 10% of the dataset-specific signal distribution, then ranks clusters using aggregate v_j and reports V(C_j) diagnostically.V(C_j) indicates whether high propensity is broadly distributed across a cluster or concentrated in a subset.
  • D. VIRALITY PRIORITISATION RESULTS: Constraint VP results are qualitative case studies rather than definitive prevalence counts, because proxy-based scoring does not support complete quantitative coverage estimates.Cross-dataset comparisons therefore use themes, framing, and within-dataset percentile ranks rather than absolute virality magnitudes.
  • D. VIRALITY PRIORITISATION RESULTS: Relaxing the virality threshold steadily increased VP-flagged clusters because broader tweet-level definitions raised cluster viral density and shifted more clusters into its upper tail.This pattern was explicitly reported for COVID–19_FNIR.

1) Constraint clustering sensitivity · 2) Qualitative Case Studies of VP-Flagged Clusters · 3) Synthesis of Findings Across Detection, Topics, and Virality

Constraint’s VP prevalence is treated as sensitivity-prone and used mainly for qualitative inspection, while COVID–19_FNIR and Monkeypox support quantitative threshold summaries. Across datasets, BERTopic-VP links thematic separation, classifier confidence, and virality-aware prioritisation, but viral clusters are not necessarily misinformation.

  • 1) Constraint clustering sensitivity: Constraint VP prevalence is sensitivity-prone because threshold-based totals depend directly on cluster boundaries.A rerun-sensitivity check over R = 10 runs is reported, with only the UMAP random seed varied.
  • 1) Constraint clustering sensitivity: Quantitative threshold-based VP summaries are restricted to COVID–19_FNIR and Monkeypox, while Constraint is used primarily for qualitative inspection.Constraint examples are presented in Table 9 and Fig. 3.
  • 2) Qualitative Case Studies of VP-Flagged Clusters: COVID–19_FNIR’s prioritised cluster focused on prevention discourse but was mostly factual in sampled items, whereas Monkeypox included highly scoring clusters driven by credible response and case-trend news.Table 9 provides virality scores, dominant terms, and misinformation presence for examples from each dataset.
  • 3) Synthesis of Findings Across Detection, Topics, and Virality: Monkeypox achieved the best clustering quality and strongest classifier performance, while COVID–19_FNIR showed lower coherence and Constraint showed weaker label alignment.COVID–19_FNIR contained diverse themes, whereas Constraint produced coherent clusters with lower purity.
  • 3) Synthesis of Findings Across Detection, Topics, and Virality: VP-flagged clusters often overlapped with high classifier-confidence regions measured by F1-score and ROC-AUC.This overlap indicates that virality-prioritised topics also exhibited strong discriminatory linguistic-rhetorical patterns.
  • 3) Synthesis of Findings Across Detection, Topics, and Virality: COVID–19_FNIR and Monkeypox contained highly viral yet largely factual clusters, whereas Constraint’s flagged clusters contained mostly misinformation.Because VP responds to engagement concentration regardless of veracity, it should be coupled with the misinformation classifier.
  • 3) Synthesis of Findings Across Detection, Topics, and Virality: Linking VP outputs with emotional-linguistic drivers and coherence patterns provides multidimensional validation of high-risk misinformation topics.The mechanism surfaces narratives with strong engagement potential and thematic alignment, supporting proactive public-health counter-messaging.

E. COHERENCE AS SUPPORTING VALIDATION · 1) Justification for Coherence Score Differences in BERTopic-VP

BERTopic-VP coherence scores provided secondary validation for virality-prioritised narratives, with higher mean coherence in Monkeypox and Constraint than COVID–19_FNIR. Bootstrap analysis confirmed a significant COVID–19_FNIR–Monkeypox difference, attributed primarily to differing degrees of narrative fragmentation under a shared clustering configuration.

  • E. COHERENCE AS SUPPORTING VALIDATION: VP-flagged clusters often corresponded to thematically cohesive narratives, although emotional tone, lexical simplicity, and engagement potential remained the dominant prioritisation drivers.Coherence was used as secondary validation of virality patterns across observed engagement and proxy VP scores.
  • E. COHERENCE AS SUPPORTING VALIDATION: 0.286 was the mean coherence for Monkeypox topics, compared with 0.084 for COVID–19_FNIR and 0.463 for Constraint.These values were computed for topics processed through BERTopic-VP.
  • E. COHERENCE AS SUPPORTING VALIDATION: Fear 0.226 and anger 0.193 marked the strongest emotional activation in mpox misinformation, which also showed linguistically simpler tweets.The passage links concentrated narrative structure with emotionally charged and easily digestible language as conditions favouring amplification.
  • E. COHERENCE AS SUPPORTING VALIDATION: COVID–19_FNIR could still produce high-impact clusters despite lower coherence when emotional salience, low cognitive load, and engagement signals aligned.The “Italy coffin” shock imagery narrative is given as an example of this interaction.
  • E. COHERENCE AS SUPPORTING VALIDATION: COVID–19_FNIR’s low coherence reflected topic fragmentation, with some clusters combining lab-leak, mask-efficacy, and policy narratives rather than indicating model failure.A possible analyst-facing heuristic was Cv < 0.10, with optional post hoc merging using centroid similarity; this heuristic was not applied to the reported results.
  • 1) Justification for Coherence Score Differences in BERTopic-VP: 10,000 bootstrap resamples of fixed topic-level coherence values quantified uncertainty from topic-to-topic variability without re-estimating the clustering model.The procedure used a fixed BERTopic run with fixed preprocessing and random seed, treating the topic partition as fixed.
  • 1) Justification for Coherence Score Differences in BERTopic-VP: Narrative fragmentation primarily explained the disparity: COVID–19_FNIR covered competing themes, whereas Mpox used fewer frames and Constraint was thematically narrower.The differences were observed under the same fixed UMAP and HDBSCAN configuration without dataset-specific retuning; external coherence values were contextual only, not like-for-like benchmarks.

F. ROBUSTNESS CHECKS AND DESCRIPTIVE · V. DISCUSSION

Descriptive robustness checks support the paper’s main inter-dataset contrasts while limiting inferential claims because analytical units may be dependent. The discussion shows that BERTopic-VP combines strong misinformation discrimination with virality-aware, outbreak-sensitive prioritisation, including proxy-based early-warning signals when engagement data are unavailable.

  • F. ROBUSTNESS CHECKS AND DESCRIPTIVE: Robustness checks were reported descriptively because cross-validation folds, model-derived topics, and other analytical units may not be independent.The checks assess sensitivity to alternative summaries rather than providing definitive hypothesis tests.
  • F. ROBUSTNESS CHECKS AND DESCRIPTIVE: 0.91–0.95 was the approximately narrow, largely overlapping macro-F1 band across five-fold validation ranges.This overlap was consistent with the small between-dataset differences in the headline results.
  • F. ROBUSTNESS CHECKS AND DESCRIPTIVE: Fear and anger differences persisted under alternative summaries, but remained conditional on dataset provenance and measurement choices.Affective contrasts were inspected descriptively because tweets within conversations, communities, or time windows may be correlated.
  • F. ROBUSTNESS CHECKS AND DESCRIPTIVE: The robustness results were consistent with the main inter-dataset contrast pattern but did not eliminate alternative explanations from corpus construction and platform context.Topic coherence comparisons were likewise presented as descriptive contrasts because topics generated within a shared pipeline may not be statistically independent.
  • V. DISCUSSION: F1-scores reached up to 0.950 and ROC–AUC values up to 0.989 across datasets, indicating consistently strong supervised misinformation-classification performance.The discussion interprets these results as evidence that misinformation remains separable under heterogeneous short-text conditions.
  • V. DISCUSSION: COVID–19_FNIR showed greater thematic fragmentation and higher lexical and readability complexity, whereas mpox misinformation was comparatively simpler in surface form.The paper links these differing linguistic profiles to distinct narrative styles across health crises.
  • V. DISCUSSION: Monkeypox achieved higher coherence than COVID–19_FNIR, while Constraint achieved the highest coherence but showed weaker alignment with labels, including lower ARI.VP-flagged clusters could surface high-impact, potentially higher-risk narratives even when overall volume was modest, and the dendrogram suggested partial outbreak-specific separation.
  • V. DISCUSSION: 9 of 112 clusters (8.0%) in COVID–19_FNIR and 5 of 87 clusters (5.7%) in Monkeypox were VP-flagged under the top 1% threshold.Mpox used observed engagement metadata, while COVID–19_FNIR and Constraint used logistic regression propensity-to-spread scores; broader thresholds increased flagged counts.

1) Implications for infodemic surveillance.

BERTopic-VP supports infodemic surveillance by surfacing potentially high-diffusion clusters for analyst triage, including modest-volume topics. Its explainability combines explicit topic representations, transparent prioritisation, and observable linguistic, emotional, coherence, and engagement diagnostics.

  • Surveillance-oriented triage: BERTopic-VP structures analyst attention by surfacing clusters with early signals of high diffusion risk or persuasive traction, even when topic volume is modest.This supports public-health triage workflows that begin with high-priority clusters.
  • Pipeline explainability: The modular pipeline separates semantic representation, dimensionality reduction, clustering, and virality scoring so analysts can inspect each stage independently.Topics are represented with ctf–idf keywords and representative examples, making clusters readable as concrete narratives.
  • Transparent prioritisation: Cluster-level explainability comes from aggregating tweet-level prioritisation signals into virality scores and applying explicit prioritisation rules.Observed engagement is used for Monkeypox, while COVID–19_FNIR and Constraint use a transferred propensity proxy, logit ηi.
  • Stakeholder-facing interpretation: VP-flagged clusters can be interpreted through emotional activation, language simplicity, and coherence, while cross-dataset comparisons reveal semantic proximity and differences in narrative structure.These diagnostics help analysts understand both which clusters are high-priority and how narratives differ across datasets.
  • Operational communication: Explicit topic representations, transparent VP flagging rules, and observable diagnostics enable outputs to be communicated and audited by non-technical infodemic-management stakeholders.Tweet-level outputs can also be contextualised through topic membership, cluster VP status, and linguistic cues rather than a probability score alone.

B. LIMITATIONS · 1) Adversarial robustness & account authenticity · VI. CONCLUSION

The study’s findings are constrained by incomplete data, methodological instability, limited generalisability, and unmeasured behavioural outcomes, while adversarial accounts may distort virality prioritisation. Despite these limitations, BERTopic-VP provides a monitoring-oriented framework for surfacing high-impact misinformation narratives and motivates multilingual, platform-specific, network-aware, and longitudinal extensions.

  • B. LIMITATIONS: COVID–19 datasets may omit removed or uncaptured high-engagement misinformation, while keyword-based Monkeypox data may include noise and miss closed-network or non-English narratives.These collection constraints limit coverage and representativeness.
  • B. LIMITATIONS: Incomplete quote-tweet and reply data constrain interaction-aware analysis and introduce partial visibility bias, especially because quote tweets may contain corrective commentary.Engagement pathways therefore remain less interpretable than simple aggregates.
  • B. LIMITATIONS: VP prevalence totals can be sensitive to small partition perturbations when HDBSCAN noise is high, although qualitative narratives were consistent across inspected runs.Broader hyperparameter perturbations and uncertainty intervals remain future work.
  • B. LIMITATIONS: Lexicon-based sentiment and emotion estimates may misread sarcasm, implicit emotion, and cultural variation, so emotional-tone interpretations require caution.Differences between tabulated means and visualised distributional densities also reflect aggregation-method inconsistencies.
  • B. LIMITATIONS: The study does not measure belief change or corrective-intervention effectiveness, and engagement and sentiment remain indirect indicators of audience response.Future research should incorporate behavioural outcomes, user-level propagation, counter-messaging, and fact-checking effects.
  • 1) Adversarial robustness & account authenticity: VP may mistake coordinated bot or troll activity for genuine attention, while manipulation artefacts can also enter propensity-to-spread proxies.Authenticity weighting could down-weight low-authenticity accounts, distinguish organic from manipulated spread, and improve cross-event comparability.
  • VI. CONCLUSION: BERTopic-VP is a modular, virality-aware pipeline compatible with periodic near-real-time batch refreshes, but live-streaming performance and operational scalability were not evaluated.Across COVID–19_FNIR, Constraint, and Monkeypox, it revealed cross-event differences in narrative structure and emotional tone.
  • VI. CONCLUSION: All datasets contained a small number of potentially disproportionate high-impact outliers, with proxy-based propensity-to-spread signals used for COVID–19_FNIR and Constraint where engagement metadata were unavailable.The framework therefore couples semantic topic discovery with virality-aware prioritisation for rapid review and counter-messaging.
Loading 2608.15691v1…