- visual manipulation, textual emotion and multimodal inconsistency at event-level
GAP
addressing event-level inconsistency following the inherent subject-predicate logic of news and robustly learning news representations from poor-quality news samples remain two challenges.
Idea
- constructing an event graph that includes multimodal entity subject-predicate logic.
- a multi-view fusion mechanism, learning comprehensive and robust representations.
- 对事件级别的实体之间的主谓关系进行建模；利用可信度来集成多视图特征的方法(模式特征、挑衅性文本情感)
- 将帖子和图像之间的不一致表示为“C”，帖子情感为“E”，模式特征为“P”。
Fig. 2 illustrates an overview of the Event-Radar framework, comprising a multi-view modeling layer and a credibility estimation layer.
Datasets
Twitter, Weibo, Pheme
Experimental Results
不足
虽然引入事件图在假新闻检测方面取得了有希望的结果，但尚未从因果关系的角度探索事件表示学习。此外，所使用的 NER工具和对象检测工具的性能会影响事件子图的结构，从而影响事件表示的准确性。此外，虽然我们工作中使用的基于置信度的融合层可以有效抵抗低质量样本的干扰，但极小的置信度分数可能会导致分类器输入中出现大量零值，从而带来过拟合或梯度的风险消失。
- object detection tools, e.g., Faster R-CNN (Chen et al., 2019)
- NER tools, e.g., Stanford NLP (Manning et al., 2014) and TextSmart NLP (Zhang et al.; Liu et al.)
参考文献
- Integrating as many available multi-modal clues as possible becomes crucial for fake news detection, called multi-view learning (Ying et al., 2023; Wu et al., 2021; Zeng et al., 2023; Wan et al., 2024).
- Multi-view Learning
