Organizations: The University of Osaka Osaka, Japan · Alibaba Inc Hangzhou, China · Nagoya University Nagoya, Japan · Tokyo University of Science Tokyo, Japan
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 .
Figures & tables
Fig. 1: Overview of fake news detection framework under different emotional backgrounds. Red text indicates emotion-related modifications introduced during rewriting.
Dataset
Training
Test
Total
True
Fake
Total
True
Fake
PolitiFact
360
180
180
90
45
45
GossipCop
6,332
3,166
3,166
1,584
792
792
LUN
6,000
3,000
3,000
1,500
750
750
TABLE I: Statistics of the benchmark datasets used in this work. The three datasets cover different data scales, providing a comprehensive evaluation under multiple emotional rewriting settings.
Fig. 2: Overview of the proposed Gated Cross Attention (GCA) framework. The original article, one reliable rewriting, one unreliable rewriting, and explanation generated from the original article are encoded independently. GCA adaptively integrates explanation-aware contextual information to detect fake news.
PolitiFact
GossipCop
LUN
Method
Orig
Joy
Anger
Sad
Happy
Orig
Joy
Anger
Sad
Happy
Orig
Joy
Anger
Sad
Happy
Accuracy
ENDEF
88.89
73.33
82.22
80.00
73.33
67.99
63.13
68.50
66.35
61.81
77.67
76.13
73.40
75.13
75.73
ARG
85.56
69.44
74.44
73.33
70.00
73.23
69.67
66.79
68.75
69.88
85.40
78.50
78.93
80.00
80.10
Sheepdog
88.77
76.11
84.22
79.89
71.67
75.89
75.11
75.95
76.03
74.66
92.61
85.90
88.43
86.71
87.41
Ours
90.56
77.11
83.56
81.11
75.56
76.72
76.29
75.38
76.10
76.18
93.96
87.29
90.01
88.20
88.55
TABLE II: Experimental results of our method on the three benchmark datasets (in % ). Accuracy and F1 score are reported under the original news setting as well as different emotional rewriting conditions.
PolitiFact
GossipCop
LUN
Method
Joy
Anger
Sad
Happy
Joy
Anger
Sad
Happy
Joy
Anger
Sad
Happy
Accuracy
with BERT
76.56
80.11
79.44
75.44
75.42
75.04
75.70
74.99
85.20
85.61
86.07
85.49
w/o Lcon
76.11
82.29
81.11
75.05
74.85
75.68
75.87
74.99
86.40
89.93
88.10
88.33
w/o Gate
77.78
80.00
76.22
73.33
75.51
75.63
75.82
76.12
83.27
84.93
86.14
84.27
Ours
77.11
83.56
80.11
75.56
76.29
75.38
76.10
76.18
87.29
90.01
88.20
88.55
TABLE III: Ablation study results of our method on three datasets (in % ). Note that removing the gating mechanism also disables Lcon , since the contrastive loss is computed from gated representations.
Fig. 3: Gating values for real and fake news on GossipCop (a, b) and LUN (c, d) test sets. The left column shows token-level gating distributions, while the right column presents the mean gating value for each news sample.