Disconfounding Fake News Video Explanation with Causal Inference

Authors: Lizhi Chen, Zhong Qian, Peifeng Li, Qiaoming Zhu

IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on the Fake VE dataset demonstrate the effectiveness of CIFE, which generates more faithful explanations by mitigating object entanglement and aspect imbalance. Our code is available at https: //github.com/Lieberk/CIFE. ... Our causal framework CIFE is compatible with existing multimodal models. On the Fake VE benchmark, it achieves improvements of 16.2% in BLEU-1 and 20.1% in ROUGE-L, with extensive experiments validating the critical role of causal intervention in ensuring explanation faithfulness. ... 4 Experiments ... 4.4 Experimental Results ... 4.5 Ablation Study
Researcher Affiliation Academia Lizhi Chen, Zhong Qian, Peifeng Li , Qiaoming Zhu School of Computer Science and Technology, Soochow University EMAIL, EMAIL,
Pseudocode No The paper describes the CIFE framework and its components (IVOD, IEAM) using text, causal graphs, architectural diagrams, and mathematical equations. It does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https: //github.com/Lieberk/CIFE.
Open Datasets Yes We conducted experimental research on Fake VE [Chen et al., 2025], currently the largest and most comprehensive publicly available FNVE dataset, with brief statistical details presented in Table 1.
Dataset Splits Yes Table 1: Statistics of the Fake VE dataset, where Avg. . Dur. and Exp. refer to Average , Duration and Explanation , respectively. Split #of News Avg.Title Avg. Dur (s) Avg. Exp Train 2138 21.40 61.23 49.76 Val 267 16.17 63.45 50.50 Test 267 15.37 60.32 50.08 Total 2672 20.27 61.78 49.86
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory used for running the experiments. It only mentions general training parameters like batch size.
Software Dependencies No The paper mentions using "Adam W [Loshchilov and Hutter, 2017] as the optimizer" but does not specify version numbers for any key software components or libraries (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes During training, we uniformly sample video frames with a maximum sequence length of 55 frames per video, applying pooling operations to each frame as the visual source representation. We employ Adam W [Loshchilov and Hutter, 2017] as the optimizer with a learning rate of 1e-4, a batch size of 16, and train the models for a maximum of 15 epochs.