LCV method detects misinformation by reconstructing missing context
New AI detector outperforms baselines by reconstructing omitted facts from context...
A new paper from Hui Li, Zhongquan Jian, Jinsong Su, and Junfeng Yao introduces Latent Causal Void (LCV), a detection method for misinformation that omits key context. While automatic detectors catch explicit deception, they struggle with articles that are locally coherent but misleading when compared to contemporaneous reports that supply missing background facts. LCV addresses this by explicitly reconstructing the missing fact for each target sentence using a frozen instruction-tuned large language model, then using the reconstructed text as a cross-source relation in a heterograph for graph reasoning. This allows the model to represent the missing fact itself rather than just attaching retrieved context or predicting a categorical omission signal.
Evaluated on the bilingual benchmark of Sheng et al., LCV outperforms the strongest omission-aware baseline by 2.56 macro-F1 on English and 2.84 on Chinese. The results demonstrate that explicit missing-context reconstruction is a powerful representation for omission-aware misinformation detection, offering a new direction for handling subtle, context-dependent disinformation. The paper is available on arXiv under the subject areas Computation and Language and Social and Information Networks.
- LCV uses a frozen instruction-tuned LLM to generate a short missing-context description for each sentence–article pair.
- The reconstructed context is fed into a heterograph over target sentences and context articles for graph reasoning.
- Improves over strongest omission-aware baseline by 2.56 and 2.84 macro-F1 on English and Chinese respectively.
Why It Matters
Enables detection of subtle misinformation that only becomes false when compared to external context