Modelling Misinformation Sentiment Dynamics and Attention Displacement on Social Media

Social media content moderation is a resource allocation problem: there are far more potentially disordered narratives circulating at any given moment than any fact-checking ecosystem can address. This dissertation builds a decision-support framework for prioritising which narrative clusters warrant intervention first, under conditions where the key variables — veracity and intent — are partially observable or entirely unobservable.

Rather than attempting to classify content as true or false (a binary that the literature treats as a fundamental misrepresentation of information disorder), the framework tracks how narrative-level discourse states evolve over time and ranks clusters by the strength and direction of that movement.

The problem with binary fact-checking

Information disorder is not a binary construct. It is a spectrum in which veracity and intent both vary continuously. Identical claims worded slightly differently receive divergent ratings from major fact-checking organisations (30.4% divergence between Snopes and PolitiFact on closely matched claims). Intent is even harder to infer: NLP models cannot reliably determine whether a user is deliberately spreading falsehoods or simply passing on something they believe. This dissertation accepts these constraints as structural rather than trying to engineer around them.

Data and sampling design

The study uses X (formerly Twitter) as the decision environment. Data was collected in two snapshots — 23 December and 26 December 2023 — each containing posts from the preceding seven days, sampled using identical logic across identical topic clusters. Topics were selected based on Community Note submissions active across both collection windows, retaining only clusters with sustained activity. Community Notes were treated as proxy signals for contested information disorder, not as ground-truth labels.

From those Community Note datasets, weighted sampling was used to collect approximately 1,500 posts per snapshot. Posts were retrieved via the X API using tweetIDs extracted from the Community Notes database. Each post was assigned to a cluster using a keyword hit-score method, retaining only the highest-scoring Community Note per tweetID where multiple notes competed.

Five narrative clusters emerged from the data: Named Entities Cases, Discourse Media, Identity Demographics, and two further clusters based on keyword coherence with established patterns of information disorder diffusion in the literature.

Linguistic measurement

Each post was processed through a pipeline that extracted two families of measures designed to be statistically orthogonal — meaning they capture independent dimensions of disorder rather than proxying the same underlying signal:

• Belief Proxy Score (BPS): derived from a fine-tuned RoBERTa transformer model, capturing the degree to which language signals belief endorsement versus scepticism or hedging.
• Linguistic Risk measures: derived from features including hedging language, emotional amplifiers, certainty markers, and source-attacking vocabulary — capturing the volatility and assertiveness of the discourse independent of its belief orientation.

Statistical orthogonality between the two measure families was verified before index construction to avoid double-counting the same signal.

Markov chain dynamics

Narratives were classified into six discrete states within a modified Information Disorder Spectrum — combining scope (Global vs Local) and disorder-likelihood — creating categories such as GDD (Global Disorder-heavy Disorder-likely) through to LNN (Local Non-disorder Disorder-unlikely). State transition matrices were estimated from the two snapshots, capturing how the distribution of posts across disorder states shifted between collection windows.

The Markov framework was chosen deliberately: it models distributional dynamics at the narrative level without requiring causal inference or population-level claims, neither of which are defensible given partially observed data.

Index construction

Two composite indices were constructed from the linguistic measures and Markov dynamics:

• Credibility Influence Index (CII): a directional index capturing whether a cluster is shifting toward stronger belief endorsement over time or stabilising — a measure of escalation trajectory.
• Attention Displacement Index (ADI): a magnitude index capturing the strength of attention movement across disorder states — distinguishing small directional changes from large shifts that warrant action.

Both indices are paired with uncertainty estimates: rather than producing a single-point ranking, 5,000 Monte Carlo resampling iterations generate distributions and confidence bands across the index values, enabling decision-makers to identify which priorities are robust across simulation runs and which are driven by noise.

Key findings

The prioritisation map produced by the two indices distinguishes clusters meaningfully. The Named Entities Cases cluster (containing high-profile discussions such as Epstein files and Charlie Kirk) showed high engagement and linguistic risk but low belief-activation likelihood — the framework suggests engagement reduction rather than fact-checking resource expenditure. The Discourse Media cluster, characterised by language attacking the validity of other sources ("this footage is digitally altered"), showed high linguistic volatility and active belief hardening — consistent with literature on the Continued Influence Effect (CIE), where corrections can entrench false beliefs if poorly timed. The Identity Demographics cluster combined high linguistic volatility with significant belief-driving signal, suggesting the need for multi-dimensional intervention.

Conference acceptance

The paper was accepted for oral presentation at the 24th International Federation of Operational Research Societies (IFORS) Conference, Vienna, Austria, July 12–17 2026, in the Behavioural OR stream. Acceptance was confirmed March 3, 2026. The paper is co-authored with and supervised by Dr. Mehdi Toloo, Surrey Business School.

Technical notes

The full pipeline was implemented in Python. Data collection used the X Academic Research API (v2 endpoints). NLP preprocessing used standard tokenisation, stopword removal, and URL stripping before transformer inference. The RoBERTa model was fine-tuned on a labelled stance detection corpus prior to scoring. Markov transition matrices were estimated from empirical state frequencies across the two snapshots. Monte Carlo resampling drew 5,000 bootstrap samples from each cluster's post distribution to generate index uncertainty bounds. All intermediate outputs were stored in structured format for reproducibility.