SAFE-MR: Evidence Sufficiency Learning for Selective Multimodal Rumor Detection
SAFE-MR judges whether the evidence is sufficient before trusting a rumor verdict.
The paper argues that retrieved evidence can be relevant but still inadequate because of missing provenance, duplication, or unresolved contradictions. SAFE-MR splits multimodal posts into verifiable claims, builds a claim-evidence graph, and uses separate heads for veracity and sufficiency. It reports macro-F1 scores of 91.2% on NewsCLIPpings, 75.8% on VERITE, and 85.2% on XFacta, with gains over a matched evidence-backed model. On a diagnostic selection set, it lowers error at 80% coverage from 13.8% to 8.5%. ArXiv · AI/CL/LG's note
The paper argues that retrieved evidence can be relevant but still inadequate because of missing provenance, duplication, or unresolved contradictions. SAFE-MR splits multimodal posts into verifiable claims, builds a claim-evidence graph, and uses separate heads for veracity and sufficiency. It reports macro-F1 scores of 91.2% on NewsCLIPpings, 75.8% on VERITE, and 85.2% on XFacta, with gains over a matched evidence-backed model. On a diagnostic selection set, it lowers error at 80% coverage from 13.8% to 8.5%. ArXiv · AI/CL/LG's note
score 4