Gated Cross Attention improved fake-news detection under emotional rewrites
The method paired rewritten news with LLM-made explanations and held up better across emotional versions on PolitiFact and LUN.
A new fake-news detector used LLM-generated explanations to stay steadier when the same story was rewritten with a different emotional tone.
The work focused on a tricky case: the factual claim stayed the same, but the wording became more angry, calm, or otherwise emotionally colored.
The gate helps reduce mismatches caused by emotional reframing.
To test that setting, the system built emotion-rewritten evaluation sets from original news articles. It also generated explanations from the original text to act as stable background knowledge.
The model used a Gated Cross Attention (GCA) setup. In plain terms, that means it learned when to lean on the rewritten article and when to trust the explanation more.
That matters because emotional phrasing can pull a detector away from the real signal. The gate helps reduce those mismatches.
In tests on PolitiFact, GossipCop, and LUN, the method showed notable gains under several emotional conditions on PolitiFact and LUN.
On GossipCop, it stayed competitive rather than clearly pulling away. So the gains were not uniform across every set.
The authors also looked at how much the explanation guidance and the gating part mattered under different emotional settings.
The code and data were made available with the paper.
The limits are clear from the excerpt: the gains were strongest on PolitiFact and LUN, the result on GossipCop was only competitive, and the abstract does not show how the method behaves outside these datasets or emotion settings.
Source: arXiv cs.CL