Interpretable Rumor Detection in Microblogs by Attending to User Interactions

Serena Khoo; Hai Leong Chieu; Zhong Qian; Jing Jiang

In brief

Proposes post-level attention networks (PLAN) to detect rumors from community responses on social media by modeling pairwise interactions between tweets via self-attention, overcoming limitations of prior tree-structured models that fail to capture cross-branch interactions. Introduces structure-aware variants and achieves state-of-the-art results on Twitter15/16 and PHEME datasets while providing interpretable explanations via attention weights at both token and post levels.

Contributions

Post-Level Attention Network (PLAN): Flattens conversation trees and uses transformer self-attention to allow any tweet to attend to any other tweet, enabling modeling of cross-branch interactions that tree models miss.
Structure-Aware variants: StA-PLAN incorporates tree structure information into attention computation; StA-HiTPLAN adds hierarchical token-level attention for more complex sentence representations.
Multi-dataset evaluation: First to evaluate on both PHEME 5-event dataset and Twitter15/16, enabling direct comparison across datasets; achieves 85.2% accuracy on Twitter15 (+14.2% over prior work), 86.8% on Twitter16 (+6.8%), and 39.5% F1 on PHEME (+1.6%).
Interpretability: Leverages attention weights to provide both token-level and post-level explanations of predictions—extracting the most important tweets and phrases that influenced the model's veracity classification.

Method

The key insight is that social media threads have ambiguous structure—users reply to the entire thread, not just specific parents. Prior tree-recursive models only model parent-child interactions, missing important signals from other branches.

PLAN arranges tweets chronologically and passes them through multiple multi-head attention (MHA) layers, allowing all pairwise interactions. Each tweet is max-pooled to a fixed-size representation, then self-attention layers learn which other tweets are relevant. A final attention layer weights tweet importance for prediction.

StA-PLAN modifies attention computation to explicitly encode structural relationships (parent, child, before, after, self) as learned bias vectors in the attention mechanism, following Shaw et al. (2018). This preserves structural signals while allowing cross-branch attention.

StA-HiTPLAN adds token-level self-attention before post-level attention, learning word importance within each tweet before combining tweets. This is more powerful for datasets with implicit signals (like PHEME, where "doesn't need facts to write stories" requires word-level inference).

Time-delay embeddings encode when each tweet was posted, using positional encoding formulas with 100 time bins (10-minute intervals).

Results

Twitter15 (balanced classes, larger dataset, clearer signals):

PLAN: 84.5% accuracy
StA-PLAN: 85.2% accuracy (best performer)
StA-HiTPLAN: 80.8% accuracy
Baseline (TD-RvNN reimplemented): 65.9% accuracy
Original TD-RvNN: 72.3% accuracy

Twitter16 (more imbalanced, smaller):

PLAN: 87.4% accuracy (best)
StA-PLAN: 86.8%
Structural information did not improve over PLAN

PHEME (cross-event, weaker crowd signals):

StA-HiTPLAN + time-delay: 39.5% F1 (best, +1.6% over LSTM baselines)
Token-level attention crucial here; random split increases F1 to 77.4%, revealing that event-specific features dominated in event-wise splits
Time-delay embedding improved PHEME (contrary to Twitter datasets)

Results show conflicting patterns: structural modeling helps Twitter15 (large, clear structure) but not Twitter16 (smaller); token-level attention helps PHEME (weak signals) but not Twitter datasets (strong signals). Suggests dataset-specific signal structure matters.

Notes

Strengths:

Addresses real limitation of tree models (cross-branch interactions are real in social media).
Interpretability via attention is valuable for understanding which signals drive predictions.
Thorough experimental comparison across three datasets with different characteristics.
Analysis of why structural information helps differently across datasets (data size, signal clarity) is insightful.

Limitations:

Time-delay embedding showed inconsistent benefits (helps PHEME, doesn't help Twitter); suggests the embedding may be overfit to one domain.
Token-level attention (StA-HiTPLAN) sometimes underperforms the simpler PLAN, indicating potential overfitting risk.
Evaluation on PHEME in event-wise split (37.5% F1) vs. random split (77.4% F1) reveals strong domain-specific overfitting—model learns event identities rather than generalizable signals. Authors acknowledge this needs future work.
No comparison to contemporaneous BERT-based approaches or ablation studies on MHA layer count.

Interpretability results: The token and post-level explanations are intuitive but not formally validated against human judgment. Example shows the model correctly identifies a sarcastic phrase ("doesn't need facts to write stories") as indicative of a false claim, but unclear whether this generalizes or is specific to this domain.