Core Concepts
TT-BLIP introduces a novel model for fake news detection by integrating text, image, and image-text features using advanced fusion techniques.
Abstract
I. Abstract:
- TT-BLIP model introduced for fake news detection.
- Utilizes bootstrapping language-image pretraining (BLIP) and BERT for text feature extraction.
- Employs ResNet and BLIPImg for image feature extraction.
- Incorporates Multimodal Tri-Transformer for feature fusion.
II. Introduction:
- Digital platforms have increased misinformation, necessitating improved fake news detection.
- Social media integrates images with textual content to enhance news narratives.
III. Method:
A. Overview:
- TT-BLIP consists of three modules: feature extraction, feature fusion, and fake news detector.
B. Feature Extraction Layer:
- Textual, image, and image-text features are extracted using BERT, BLIPTxt, ResNet, and BLIPImg.
C. Feature Fusion Layer:
- MultiModal Tri-Transformer fuses features from different modalities using multi-head attention mechanisms.
D. Fake News Detector:
- Integrated features processed through the MultiModal Tri-Transformer are used for binary classification of fake news.
IV. Experiments and Results:
A. Dataset:
- Weibo dataset contains 6,137 training articles (2,802 fake & 3,335 real) and 833 testing articles (852 real & 833 fake).
- Gossipcop dataset includes 10,010 training articles (2036 fake & 7974 real) and testing articles (545 fake & 2285 real).
B. Experimental Settings:
- Utilized pretrained models like BERT for text processing and ResNet for image analysis.
C. Results and Analysis:
- TT-BLIP outperformed state-of-the-art models in accuracy on both datasets.
D. Comparison of Fusion Methods:
- TT-BLIP excelled compared to traditional fusion methods like early fusion or late fusion.
V. Conclusion:
TT-BLIP model demonstrated superior performance in detecting fake news by effectively integrating text, image, and image-text features using advanced fusion techniques.
Stats
この論文では、TT-BLIPモデルがWeiboデータセットで90.7%、Gossipcopデータセットで88%の精度を達成した。