Contributions
- First application of news environment perception to fake news detection: Incorporates signals from both macro (mainstream media context) and micro (event-similar items) news environments, departing from post-only and knowledge-based approaches.
- News Environment Perception Framework (NEP): A two-stage architecture that constructs macro and micro environments from recent mainstream news, applies popularity-oriented and novelty-oriented perception modules, and fuses environment-perceived vectors with post features via gate mechanism.
- Cross-lingual datasets with contemporary context: Integrates Chinese (Weibo) and English datasets with mainstream news items to enable environment-based detection research.
- Broad compatibility: Demonstrates that NEP improves performance when integrated with multiple base models (Bi-LSTM, EANN, BERT, DeClare, MAC) and is practical for deployment.
Method
The News Environment Perception Framework operates in three stages:
1. Environment Construction: For a post $p$, the framework constructs two news environments from recent mainstream news released within $T$ days before $p$ is published:
- MacroEnv: All relevant mainstream news items (macro perspective of recent media focus and audience attention)
- MicroEnv: Subset of MacroEnv containing items similar to events mentioned in $p$ (micro perspective of event-specific coverage)
2. Environment Perception: Two perception modules capture post-environment relationships:
- Popularity-oriented MacroEnv Perception: Captures whether $p$ relates to trending events. Uses cosine similarity to measure how popular the main event of $p$ is in MacroEnv. Gaussian Kernel Pooling converts variable-length similarity lists into soft counting via exponential kernels.
- Novelty-oriented MicroEnv Perception: Captures novel aspects of $p$ within event-similar coverage. Computes two similarity features—one for novelty measurement—using Gaussian Kernel Pooling to generate robust distributed representations.
Both modules output environment-perceived vectors ($v^{p,\text{macro}}$ and $v^{p,\text{micro}}$) by fusing similarity information with semantic post representations using MLPs.
3. Prediction with Gating: Environment-perceived vectors are fused with the post feature vector $o$ from the base detector via a learned gate: $$v^p = g \odot v^{p,\text{macro}} + (1-g) \odot v^{p,\text{micro}}$$ where $g = \text{sigmoid}(\text{Linear}(o \oplus v^{p,\text{macro}}))$ balances macro and micro contributions. The result feeds into the final classifier.
Results
Performance gains across all tested detectors on both Chinese and English datasets:
Chinese Dataset (merged Weibo posts 2010–2021, 583,208 posts, balanced with fact-checking system):
- Bi-LSTM: 0.727 → 0.776 accuracy (+4.9 pp), 0.775 → 0.803 F1
- EANN: 0.732 → 0.776 accuracy
- BERT: 0.792 → 0.831 accuracy (+3.9 pp), 0.837 → 0.850 F1
- DeClare: 0.764 → 0.800 accuracy
- MAC: 0.755 → 0.764 accuracy
English Dataset (merged multiple sources: Kochkina et al. 2018, Augenstein et al. 2019, Shaar et al. 2020, plus news items 2014–2018, 6,483 posts):
- Bi-LSTM: 0.705 → 0.718 accuracy, 0.719 → 0.716 F1
- EANN: 0.700 → 0.722 accuracy
- BERT: 0.709 → 0.728 accuracy (+1.9 pp)
- DeClare: 0.714 → 0.717 accuracy
Ablation results (Table 3) show both MacroEnv and MicroEnv are necessary; removing either reduces performance. Novelty-oriented MicroEnv perception is more effective than popularity-oriented MacroEnv for most detectors, but both contribute meaningfully. Across all base models, NEP consistently improves macro F1 on the skewed real-world dataset (+5.2% on Chinese, +16.89% macro F1 on English with balanced metrics).
Notes
Strengths:
- Novel framing: News environment perception is intuitive and well-motivated—fake news spreads via riding the wave of popular or unexpected events.
- Cross-lingual evaluation: Both Chinese and English experiments show broad applicability.
- Model-agnostic design: NEP improves multiple base detectors, suggesting it captures complementary signals.
- Practical considerations: The authors acknowledge and address deployment constraints (timeliness, access to mainstream news sources, skewed real-world distributions).
Limitations & Open Questions:
- Environment quality depends on mainstream news access: The framework assumes reliable, timely access to mainstream media feeds. In regions with limited press freedom or paywalled coverage, this assumption breaks down.
- Moderate gains on English data: While Chinese results are strong, English improvements are more modest (1–2% on balanced metrics), suggesting environment signals may be less informative in English-language contexts or the English dataset construction differs.
- Novelty component underexplored: The paper briefly notes that fake news is more novel than true news on Twitter (Vosoughi et al. 2018) but doesn't deeply investigate when novelty signals help vs. hurt detection.
- Future directions mentioned: Incorporating historical background, modeling post-environment relationships over time, and exploring environmental diversity metrics.
This is the first work to systematically incorporate news environment perception for fake news detection, opening a new direction for context-aware misinformation analysis.