When a rumor breaks out on social media, determining its truth value purely from the source text is incredibly challenging. Malicious actors can easily craft convincing text or manipulate metadata to deceive traditional classifiers. However, they cannot easily control how the public reacts. The core intuition behind this research direction is that the collective stance of a community acts as a diagnostic signal for information veracity. False claims naturally trigger skepticism, questioning, and denial, creating a distinct "stance signature" that differentiates them from verified facts.
The Stance-Veracity Connection
How exactly does community reaction map to truth? In a foundational benchmark, Leon Derczynski et al. formalized this by breaking down conversation threads into individual reply stances: Support, Deny, Query, or Comment in SemEval-2017 Task 8: RumourEval: Determining rumour veracity and support for rumours. Their work established that a high frequency of questioning or outright denial within a thread strongly correlates with rumors that are eventually proven false.
Instead of treating these reactions as noise, models can aggregate them to infer veracity. For instance, Ipek Baris et al. demonstrated in CLEARumor at SemEval-2019 Task 7: ConvoLving ELMo Against Rumors that calculating the average probability of support, denial, and questioning across a crowd's reaction provides a strong proxy for the truth value of a source claim. As Momchil Hardalov notes in a comprehensive survey on the topic, A Survey on Stance Detection for Mis- and Disinformation Identification, computing veracity directly is often too complex without extensive external context; using the stance of related content or replies serves as a highly effective proxy.
Synergy via Multi-Task Learning
Early systems treated stance detection and rumor verification as separate, sequential steps: first predict the stances of all replies, then use those predictions to classify the rumor. However, Quanzhi Li et al. argued in Rumor Detection by Exploiting User Credibility Information, Attention and Multi-task Learning that this separation fails to exploit the inherent relationship between the two tasks. By adopting a multi-task learning framework, models can jointly predict reply stances and claim veracity, allowing them to leverage shared underlying signals.
This joint modeling creates a powerful bidirectional synergy:
- Bottom-Up Propagation: Granular stance labels (like strong denials) propagate upward to improve global veracity detection. For example, Sumeet Kumar et al. showed in Tree LSTMs with Convolution Units to Predict Stance and Rumor Veracity in Social Media Conversations that identifying denial signals at the reply level provides critical evidence that improves final rumor assessment.
- Top-Down Supervision: Global veracity knowledge can be back-propagated to guide stance classification. Ruichao Yang et al. leveraged this relationship in A Weakly Supervised Propagation Model for Rumor Verification and Stance Detection with Multiple Instance Learning to build a weakly supervised model that learns accurate post-level stance patterns using only claim-level veracity labels, drastically reducing the need for expensive manual annotations.
Structural and Temporal Dynamics
Conversations on social media are rarely linear; they propagate as hierarchical trees. The way skepticism spreads through these trees matters just as much as the text itself. To capture these non-sequential propagation patterns, Jing Ma et al. introduced tree-structured Recursive Neural Networks (RvNN) in Rumor Detection on Twitter with Tree-structured Recursive Neural Networks, demonstrating that the recursive interplay of support and denial helps identify community self-correction mechanisms.
Furthermore, public opinion is dynamic and evolves over time. Penghui Wei et al. observed in Modeling Conversation Structure and Temporal Dynamics for Jointly Predicting Rumor Stance and Veracity that while initial reactions to a false rumor might be supportive, the conversation typically exhibits a distinct increase in denial and querying stances as time goes on. By tracking how these stance signatures shift over time, models can achieve early and highly accurate misinformation detection.
The Resilience of Crowd Signals
Why does this approach work so well compared to traditional content-based models? The answer lies in adversarial resilience. As Lanyu Shang points out in FauxWard: A Graph Neural Network Approach to Fauxtography Detection Using Social Media Comments, while a malicious actor can easily manipulate a post's text or image, it is significantly harder to coordinate or manipulate hundreds of independent user comments.
Spontaneous opposition from a skeptical crowd acts as a distributed verification system. Even when ground-truth databases are unavailable for real-time events, the collective stance of the crowd provides a reliable, real-time proxy for claim veracity, a phenomenon explored by Nikita Lozhnikov et al. in Stance Prediction for Russian: Data and Analysis.
By shifting the focus from what a story says to how the community reacts, joint stance-veracity modeling provides a robust, scalable, and structurally grounded framework for combating misinformation.
Key Foundational Work
NICT Disaster Information Analysis System (2013) — contributed 1 concept elements to this direction.
Extracting and Aggregating False Information from Microblogs (2013) — contributed 1 concept elements to this direction.
Tweet, but verify: epistemic study of information verification on Twitter (2014) — contributed 1 concept elements to this direction.