Altoe, F., Pinto, S. M. G., & Pinto, H. S. (2026). Predicting narrative virality on social media streams for pre-peak misinformation triage. Proceedings of the 35th ACM International Conference on Information and Knowledge Management, 127–138. https://doi.org/10.1145/3799682.3840688
Summary
This paper introduces narrative-level streaming virality prediction as a new task for misinformation triage, motivated by a crucial timing gap: false narratives peak in virality around 1.5 days, while fact-checkers take a median of ~4.5 days to publish corrections. The authors build an end-to-end streaming pipeline that clusters semantically equivalent posts into evolving narratives, flags anomalous cluster activity, normalizes claims with LLMs, and predicts remaining spread via a novel Post-Detection Spread Ratio (PSR) metric—surfacing triage candidates before peak spread. The work is explicitly grounded in journalism research on how fact-checkers reason under time and attention constraints, arguing for pre-peak, narrative-centric, human-centered triage support rather than post-level or event-level analysis.
Key Contributions
- A new task formulation: narrative-level streaming virality prediction, with an integrated pipeline (streaming clustering → online anomaly detection → claim normalization → cluster-level spread prediction).
- The PSR metric, quantifying the remaining intervention window at detection time and distinguishing detection lead (~23h) from actionable triage lead (~17h).
- State-of-the-art results on two established Automatic Claim Detection tasks (CT2025 claim normalization and CheckThat! 2024 check-worthiness) using smaller/distilled models.
- An interpretable 4-Parallel-Heads classifier decomposing check-worthiness into checkability, verifiability, and harm potential, aligned with journalistic criteria.
- The first public dataset and baselines for this task (529 anomalous clusters, 535 claims, 42-feature inputs) drawn from a 1.52M-post US 2020 political Twitter/X stream.
- A frontier-LLM zero-shot ablation and a formative study with six professional fact-checkers validating the modular design.
Methods
The dataset is built from the US Election 2020 Tweets corpus (1.52M tweets, Oct–Nov 2020), binned hourly with no-lookahead engagement snapshots. Streaming clustering uses multilingual MPNet embeddings with incremental cosine-similarity clustering (τ=0.65). Online anomaly detection uses EXPoSE (selected from the NAB benchmark) over hourly tweet counts and aggregate engagement. Claim normalization relies on Mistral-7B-Instruct with two-stage retrieval verification, topic-aware retrieval, and contrastive few-shot prompting. Virality is formulated as PSR prediction over 42 engineered features from a 6-hour post-detection window, benchmarked across linear, tree, kernel, sequence, Hawkes point-process, and text-based models. Check-worthiness is handled by a fully fine-tuned 4-Parallel-Heads DeBERTa-v3-large under a three-phase multi-task schedule, with auxiliary heads trained on soft labels from a mid-sized Mistral model.
Findings
- Anomaly detection: EXPoSE achieved an NAB score of 79.2, detected 97.9% of events before peak, and gave a +23.0h median lead, dominating changepoint, Skyline, and Z-score baselines.
- Claim normalization: Contrastive few-shot Mistral-7B reached METEOR 0.5569, beating the strongest published baseline by 10 points and outperforming larger models; fine-tuning underperformed prompting—task-specific prompt design mattered more than model capacity.
- Virality prediction: Simple models won—BayesianRidge gave best ranking (ρ=0.551), SVR-RBF best fit (R²=0.336). LSTM/GRU and Hawkes models performed poorly, confirming that early cluster dynamics are approximately linear. User authority and temporal features dominated; content features barely mattered—who spreads a claim matters more than what it says early on.
- Check-worthiness: The distilled 4-Parallel-Heads classifier hit F1=0.833 on CT2024 (prior SOTA 0.802) and generalized zero-shot (ClaimBuster F1=0.973; CT2023 F1=0.933). Claude Opus zero-shot reached only F1=0.722 at ~$18/1k claims and ~9.9s latency versus <10ms—justifying modular distilled deployment over frontier-LLM inference. 86% of errors were shared across approaches, pointing to representation learning (not feature engineering) as the bottleneck.
- Human study: The pipeline surfaced ~500 candidates from ~1.5M posts; six fact-checkers showed 83.3% agreement and rated the report structure as aligned with practice—though they rated virality prediction itself the least valuable component.
Connections
This work builds on classic accounts of how falsehood spreads and the structural challenges of fact-checking Lazer2018-mm, and on research into the collaborative, time-pressured dynamics of online rumor and correction Starbird2019-qv. Its narrative-clustering and claim-normalization machinery connects to computational text methods for tracking narratives at scale, while its distilled-classifier-beats-frontier-LLM result speaks to ongoing debates over LLM annotation and deployment trade-offs raised across the computational-text-methods literature.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify