Source-linked AI summary
Leveraging LLM-Generated Explanations for Detecting Emotionally Rewritten Fake News
Yupei Guo, Jiajun He, Xiaohan Shi, Tomoki Toda, Zekun Yang, Bowen Wang, Yukinobu Taniguchi
TL;DR
Emotionally reframed articles can preserve factual claims while challenging fake-news detectors, particularly when emotional cues are used. The paper pairs such rewrites with explanations from the original articles and integrates them using gated cross attention. GCA improves performance on PolitiFact and LUN while remaining competitive on GossipCop.
Problem
Emotionally reframed news can preserve factual claims yet mislead detectors that rely heavily on emotional cues, motivating robust detection across emotional backgrounds.
Source: §I · p. 2Method
The paper pairs emotionally rewritten news with explanations generated from the corresponding original articles and uses gated cross attention to adaptively integrate them.
Source: §I · p. 2Results
GCA improves fake-news detection performance on PolitiFact and LUN while maintaining competitive performance on GossipCop under different emotional backgrounds.
Source: B. Experimental Results · p. 5Takeaways & Limitations
The results support GCA’s effectiveness in the controlled textual emotional-rewriting setting.
Source: §VI · p. 8GPT-4o rewritings may introduce model-specific patterns, and generation overhead may limit practicality at large scale or under latency constraints.
Source: B. Limitations · p. 7
Abstract
from arXiv · showhide
The spread of fake news may cause severe social consequences. Existing fake news detection methods mainly focus on stylistic variations or incorporate external information such as explanations. However, news articles are often rewritten under different emotional backgrounds while preserving their underlying factual claims, which may affect the robustness of detection models. In this work, we investigate fake news detec- tion under fact-preserving emotional variations. To study this problem, we construct emotion-rewritten test sets and generate explanations from the original news articles as stable background knowledge. We then propose a Gated Cross Attention (GCA) framework that adaptively integrates emotionally rewritten news with the corresponding explanations, enabling the model to focus on informative explanation content while reducing potential mismatches caused by emotional reframing. Experiments on PolitiFact, GossipCop, and LUN demonstrate that the proposed method achieves notable improvements under multiple emotional conditions on PolitiFact and LUN, while maintaining competitive performance on GossipCop. We further analyze the effects of explanation guidance and gating mechanisms under different emotional conditions. Our code and data are available at: https://github.com/Flulike/fakenews gca .
I. INTRODUCTION
The paper studies fake-news detection when articles are emotionally reframed without changing their underlying claims, a setting that may challenge detectors reliant on emotional cues. It pairs rewritten news with explanations from the original articles and proposes gated cross attention to adaptively integrate them.
I. INTRODUCTION
- Prior work has studied style variation – and external-information augmentation – for fake-news detection.The paper situates its study at the intersection of these directions and emotional reframing.Source: §I · p. 1
- Emotional reframing can mislead detectors that rely heavily on emotional cues, even when auxiliary information is available.The paper frames robust use of external evidence across emotional backgrounds as an open need.Source: §I · p. 2
- The study pairs emotional rewritings with explanations generated from the corresponding original articles as stable background knowledge.This controlled setup examines detection under fact-preserving emotional variation.Source: §I · p. 2
- GCA adaptively regulates explanation contributions according to each rewritten input rather than treating explanations as uniformly relevant.The framework extends explanation augmentation to adaptive explanation-guided detection under emotional rewriting.Source: §I · p. 2
- GCA improves performance on PolitiFact and LUN under multiple emotional conditions while remaining competitive on GossipCop.The introduction reports these results across the three evaluated datasets.Source: §I · p. 2
II. DATASET CREATION
The dataset setup uses three established fake-news benchmarks and generates original-article explanations to provide shared contextual information for emotional variants. These explanations are contextual priors, not independently verified evidence.
A. Data Source
- The benchmarks cover political news in PolitiFact and LUN, entertainment and celebrity news in GossipCop, and small- and large-scale settings.All three are balanced binary fake-news tasks, and training includes the original articles plus reliable and unreliable style rewritings.Source: A. Data Source · p. 2
B. Explanation Generation
- GPT-4o generates an explanation from each original news article for use as auxiliary contextual information during detection.Explanations are generated once and reused across emotional variants, which share the original article’s authenticity label and contextual information.Source: B. Explanation Generation · pp. 2–3
- The explanations serve as contextual priors in controlled fact-preserving rewrites, not as independently verified external evidence.The paper presents GPT-4o as a consistent way to construct priors, not as a required deployment strategy.Source: B. Explanation Generation · p. 3
C. Test Set Emotional Rewritings
The test set uses GPT-4o to rewrite articles into emotional variants while retaining the original authenticity labels. The setup separates these rewrites from training rewrites, but the authors note that simultaneous factual changes and human-authored alternatives remain future directions.
C. Test Set Emotional Rewritings
- Test articles are rewritten with GPT-4o in four emotion classes: Joy, Anger, Sadness, and Happiness.The rewritings are generated independently from the explanations, with each process using only the original article as input.Source: C. Test Set Emotional Rewritings · p. 3
- Emotionally rewritten articles inherit their original article’s authenticity label to isolate changes in emotional expression.The authors manually inspected a subset and found that major factual claims were generally preserved while emotional expressions changed.Source: C. Test Set Emotional Rewritings · p. 3
- The test set is fully reconstructed with emotion-based rewritings, separate from the reliable and unreliable rewritings used in training.The two rewriting sets are independent of one another.Source: C. Test Set Emotional Rewritings · p. 3
- The label-inheritance assumption may not capture cases where emotional framing and factual content change together; human-authored rewritings are also left for future work.The authors describe LLM-based rewriting as a controllable approximation for systematic evaluation.Source: C. Test Set Emotional Rewritings · p. 3
III. PROPOSED METHOD
The proposed GCA framework combines news variants with explanations derived from the original articles. A gated cross-attention module adjusts explanation contributions before the fused representation predicts authenticity.
III. PROPOSED METHOD
- GCA extends the Sheepdog setup by associating each article’s original and style-rewritten variants with an explanation from the original article.The news variants share underlying content and authenticity labels while differing in emotional and stylistic expression.Source: §III · p. 3
- A gated cross-attention module adjusts explanation-feature contributions according to their relevance to the current news representation.The gate addresses the risk that direct news-explanation combination may introduce irrelevant or inconsistent information.Source: §III · p. 3
- Four independently parameterized language models encode the three news variants and explanation, and the fused representation predicts authenticity.The gate dynamically controls how much explanation-aware information enters the final representation.Source: §III · p. 3
A. News Rewritings and Explanation
The paper pairs three style-based versions of each news article with an explanation generated from the original. GCA uses cross-attention and gating to adaptively incorporate explanation-aware information into news representations.
A. News Rewritings and Explanation
- The original article and its reliable and unreliable rewritings share underlying content and authenticity labels; the rewritings are not external articles.The explanation is generated from the original article and reused across its rewritings as a contextual prior.Source: A. News Rewritings and Explanation · p. 3
B. GCA Fusion
- GCA independently encodes the news variants and original-derived explanation, then uses cross-attention to form explanation-aware news features.The news representation serves as the query, while the explanation representation supplies the key and value.Source: Figure 2 · p. 4, B. GCA Fusion · p. 4
- A token-level gate dynamically regulates how much explanation-aware information enters the news representation.The gate addresses the possibility that an explanation reused across emotional rewritings may become irrelevant or misleading as surface expression changes.Source: B. GCA Fusion · p. 4, Figure 2 · p. 4
- The projected explanation-aware features are combined into a fused representation, which is mean-pooled and passed to a classifier for authenticity prediction.A trainable projection aligns feature dimensions before the final fusion.Source: B. GCA Fusion · p. 4
C. Loss Function
The loss adds contrastive alignment between news representations and their sourced explanations to Sheepdog’s original training objectives. Its weight depends on the mean gate value for each news instance.
C. Loss Function
- An InfoNCE contrastive loss encourages each news representation to match its corresponding explanation while minimizing similarity to other explanations in the batch.The loss compares sequence-level mean-pooled representations for the original article and both rewritings.Source: C. Loss Function · p. 4
- The contrastive loss is weighted by each instance’s mean gate value, with the gate detached when computing that weight.The gate representation is token-level.Source: C. Loss Function · p. 4
- Removing both gating and the contrastive loss reduces the framework to an explanation-enhanced variant of Sheepdog.The paper examines this variant in its ablation study.Source: C. Loss Function · p. 4
IV. EXPERIMENTS
Experiments compare the method with three baselines on PolitiFact, GossipCop, and LUN. Evaluation covers original news and emotional rewriting conditions, reporting mean Accuracy and F1 Score across repeated runs.
A. Experiment Setting
- Table II reports Accuracy and F1 Score under the original-news setting and different emotional rewriting conditions across the three benchmarks.The caption specifies percentages but provides no individual result values in the supplied passage.Source: Table II · p. 5
- Evaluation reports Accuracy and F1 Score on the original test set and emotion-rewritten test sets, averaged over 10 runs.The runs use random seeds 0–9.Source: A. Experiment Setting · p. 5
- The experiments use four independently parameterized roberta-base encoders and train for five epochs without early stopping.The maximum input length is 512 tokens for news and explanations.Source: A. Experiment Setting · p. 5
- Experiments compare the method with Sheepdog, ENDEF, and ARG on PolitiFact, GossipCop, and LUN.Sheepdog uses original text only, ENDEF uses news content and explanations, and ARG’s rationale inputs are replaced with generated explanations.Source: A. Experiment Setting · p. 5
B. Experimental Results
Across emotional conditions, GCA improves detection on PolitiFact and LUN while remaining competitive on GossipCop. Results vary by dataset, and the framework performs consistently across emotional categories.
B. Experimental Results
- GCA improves detection on PolitiFact and LUN while maintaining competitive performance on GossipCop under different emotional backgrounds.The reported performance changes are relatively small on GossipCop and more pronounced on PolitiFact and LUN.Source: B. Experimental Results · p. 5
- On PolitiFact, GCA performs best on the original set and most emotionally rewritten variants, with improvements across emotional conditions.The authors attribute this to complementary semantic context and the gate’s suppression of less relevant explanation features.Source: B. Experimental Results · p. 5
- On LUN, GCA achieves the best overall performance among evaluated settings, with generally more pronounced improvements than on PolitiFact.The authors suggest LUN’s relatively stable political-news facts may make explanations more semantically consistent with rewritten articles.Source: B. Experimental Results · p. 5
- On GossipCop, improvements are modest, but GCA remains competitive across emotional backgrounds without noticeable degradation.The authors relate the smaller gains to entertainment articles’ stronger original emotional expression and stylistic variation, which leaves less semantic change for rewriting.Source: B. Experimental Results · p. 5
- GCA performs consistently across emotional categories, consistent with explanations capturing stable semantic information that is less sensitive to stylistic variation.The paper’s overall conclusion is that explanation-based context supports more robust predictions across emotional backgrounds.Source: B. Experimental Results · pp. 5–6
C. Ablation Studies
The complete GCA framework generally outperforms a baseline that directly fuses generated explanations, while ablations indicate contributions from the backbone, contrastive loss, and gating. Because gating and contrastive loss are coupled, their individual effects cannot be fully separated.
C. Ablation Studies
- Replacing RoBERTa with BERT generally reduces performance across datasets and emotional backgrounds.The authors suggest stronger contextual representations support more effective semantic interaction and explanation-aware fusion.Source: C. Ablation Studies · p. 6
- Removing the contrastive loss causes modest but consistent degradation, and the authors associate it with improved alignment between news and explanations.The loss is computed from gated representations, so it is not fully independent of the gating component.Source: C. Ablation Studies · p. 6
- Removing the gating mechanism causes the largest performance degradation, particularly on LUN, but also removes the coupled contrastive loss.This comparison therefore measures the combined effect of adaptive gating and contrastive alignment, not gating in isolation; direct explanation fusion may introduce irrelevant information.Source: C. Ablation Studies · p. 6
- The complete GCA framework generally outperforms the explanation-enhanced baseline, indicating benefits beyond simply adding LLM-generated explanations.The authors attribute additional benefits to adaptive gating and contrastive alignment, while noting their individual effects cannot be fully separated.Source: C. Ablation Studies · p. 6
V. DISCUSSION
Gating values vary across news samples rather than applying a uniform explanation contribution. The visualized distributions differ between real and fake news on GossipCop and LUN.
A. Gating Strength
- Real-news gating values tend to concentrate in higher ranges, while fake-news values span a wider range, indicating more varied explanation contributions across fake-news samples.The visualization covers GossipCop and LUN; PolitiFact’s test set has only 90 samples, so the authors use the larger test sets for more representative distributions.Source: A. Gating Strength · pp. 6–7, Figure 3 · p. 7
- The left column shows token-level gating distributions, while the right column shows each sample’s mean gating value for real and fake news on GossipCop and LUN.Panels a–b are GossipCop and panels c–d are LUN.Source: Figure 3 · p. 7
- Together with the ablation results, the visualization supports the interpretation that adaptive gating contributes to integrating explanation information under emotional rewriting.The authors report that removing gating also removes the coupled contrastive loss and leads to noticeable performance degradation.Source: A. Gating Strength · p. 7
B. Limitations
The conclusion reports improvements on PolitiFact and LUN and competitive performance on GossipCop in a controlled textual setting. The limitations and future-work discussion notes boundaries in label assumptions and rewrite generation, and proposes broader benchmarks, richer context, factual perturbations, and multimodal evaluation.
B. Limitations
- The evaluation assumes rewritten articles inherit their originals’ authenticity labels, which may not hold when emotional reframing also changes facts.The authors note that this assumption avoids manually labeling every generated variant but may not fully reflect real-world scenarios.Source: B. Limitations · p. 7
- GPT-4o rewrites may introduce model-specific linguistic patterns, while generation token use and computational overhead may limit large-scale or latency-sensitive use.Future work includes human-authored or human-verified rewrites and evaluation under simultaneous expressive and factual changes.Source: B. Limitations · p. 7, C. Future Work · p. 7
C. Future Work
- Future work proposes larger, more diverse rewrite benchmarks and richer contextual knowledge, including retrieval-based evidence and structured knowledge bases.It also proposes examining partial factual modifications alongside emotional reframing.Source: C. Future Work · p. 7
- The authors plan to extend the framework to multimodal fake news detection under textual and visual perturbations.Source: C. Future Work · p. 8
VI. CONCLUSION
- The framework improves performance on PolitiFact and LUN while maintaining competitive performance on GossipCop in the controlled textual emotional-rewriting setting.The authors present these results as supporting the approach's effectiveness within that setting.Source: §VI · p. 8