Rumor Detection on Twitter with Claim-Guided Hierarchical Graph Attention Networks

Hongzhan Lin, Jing Ma, Mingfei Cheng, Zhiwei Yang, Liangliang Chen, Guang Chen

Abstract

Rumor detection on social media is a critical task for maintaining information integrity. Traditional methods often model conversation threads as directed trees, which limits information flow and ignores horizontal interactions between users. This paper introduces ClaHi-GAT (Claim-guided Hierarchical Graph Attention Network), a novel framework that treats Twitter conversation threads as undirected interaction graphs. By incorporating "sibling" relationships and using the original claim to guide attention at both post and event levels, ClaHi-GAT achieves state-of-the-art performance, particularly in early rumor detection.

Introduction: The Challenge of Social Media Rumors

Rumor detection on social media is a race against time. While the structure of a conversation (who replies to whom) provides vital clues for debunking fake news, existing methods suffer from two major flaws. First, the rigid topology of directed trees (top-down or bottom-up) limits information flow and ignores "horizontal" interactions—such as when two users replying to the same post "self-check" each other’s claims. Second, contextual drift in long threads often leads sub-conversations away from the original claim, making it difficult for models to maintain focus on the veracity of the initial statement.

Proposed Method: ClaHi-GAT

To address these challenges, the authors propose a two-pronged approach: modeling conversations as undirected interaction graphs and utilizing a hierarchical attention mechanism guided by the original claim.

Undirected Interaction Graphs

Instead of a directed tree, the conversation is modeled as an undirected graph. This allows for "full-duplex" interaction. Crucially, edges are added between "sibling" nodes—tweets replying to the same parent. This structure captures how users in the same sub-thread reinforce or contradict one another, providing a richer context for rumor verification.

Example of a conversation thread and its undirected graph representation, showing semantic relationships like entailment and contradiction.

ClaHi-GAT Architecture

The model employs a hierarchical attention mechanism designed to keep the analysis anchored to the original claim:

  1. Post-level Attention (The Gatekeeper): A gating module determines how much information from the original claim should be integrated into each tweet's representation. This ensures the model stays focused on the claim's topic even in noisy or drifting threads.
  2. Event-level Attention (The Judge): Inspired by Natural Language Inference (NLI), this layer compares aggregated post features against the original claim. It uses matching functions (concatenation, element-wise product, and absolute difference) to capture whether a group of posts entails, contradicts, or is neutral toward the claim.
The ClaHi-GAT architecture, featuring gating modules, graph attention layers, and an entailment module for hierarchical rumor detection.

Experimental Setup and Datasets

The model was evaluated on several widely used benchmarks:

  • Twitter15 & Twitter16: Events categorized into four labels: Non-rumor (NR), False Rumor (FR), True Rumor (TR), and Unverified Rumor (UR).
  • PHEME: A binary dataset (Rumor vs. Non-rumor) based on five real-world breaking news events.
  • Twitter-S & Twitter-D: Split versions of the Twitter datasets based on conversation depth ("Shallow" ≤ 3 and "Deep" ≥ 4) to test robustness across complexities.

Key Insights and Results

The experimental results highlight several critical advantages of the ClaHi-GAT framework:

  • Superior Accuracy: ClaHi-GAT outperformed all baselines, including Bi-GCN and Transformer-based models. On Twitter15, it achieved an accuracy of 89.1%, a significant improvement over previous state-of-the-art methods.
  • Early Detection Power: ClaHi-GAT reaches high accuracy with only 30 posts or within 4 hours of the initial claim. This stability in the early stages of a rumor's spread is vital for practical applications.
  • The Importance of Siblings: Ablation studies showed that removing the undirected graph structure (switching back to directed trees) caused a major performance drop. This confirms that horizontal interaction between users is a critical feature for capturing corrective behavior in social media threads.
  • NLI Framing: Treating rumor detection as an inference problem—where posts are "premises" and the claim is a "hypothesis"—allows the model to leverage semantic relationships like contradiction, which is often the strongest signal for identifying false rumors.

Conclusion and Future Work

ClaHi-GAT demonstrates that modeling social media conversations as undirected graphs with claim-guided attention significantly improves rumor detection. Future work may involve incorporating pre-trained Transformers (like BERT or RoBERTa) to better handle casual slang and homonyms, or using datasets with explicit entailment labels to further sharpen the model's reasoning capabilities. Additionally, exploring ways to compensate for low-engagement scenarios remains an open challenge.

Summary updated

BibTeX
@inproceedings{lin2021rumor,
  author = {Hongzhan Lin and Jing Ma and Mingfei Cheng and Zhiwei Yang and Liangliang Chen and Guang Chen},
  title = {{Rumor Detection on Twitter with Claim-Guided Hierarchical Graph Attention Networks}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  doi = {10.18653/v1/2021.emnlp-main.786},
}
Figure