Support or Refute: Analyzing the Stance of Evidence to Detect Out-of-Context Mis- and Disinformation

• DOI: 10.18653/v1/2023.emnlp-main.259
• Publication Date: 2023-12-10T21:58:19Z
• Authors: Xin Yuan; Jie Guo; Weidong Qiu; Zheng Huang; Shujun Li
• Keywords: Disinformation; Intuition

Mis- and disinformation online have become a major societal problem as sources of harms different kinds. One common form mis- is out-of-context (OOC) information, where pieces information are falsely associated, e.g., real image combined with false textual caption or misleading description. Although some past studies attempted to defend against OOC through external evidence, they tend disregard the role evidence stances. Motivated by intuition that stance represents bias towards detection results, we propose extraction network (SEN) can extract stances multi-modal in unified framework. Moreover, introduce support-refutation score calculated based on co-occurrence relations named entities into SEN. Extensive experiments public large-scale dataset demonstrated our proposed method outperformed state-of-the-art baselines, best model achieving performance gain 3.2% accuracy.