Search papers, labs, and topics across Lattice.
The paper introduces FMS-Net, a novel graph neural network architecture for rumor detection that leverages fine-grained multimodal features and sentiment analysis. FMS-Net constructs a heterogeneous graph incorporating phrase-level textual features, object-level visual features, and sentiment similarities to capture subtle cross-modal relationships indicative of rumors. Experiments on real-world datasets demonstrate that FMS-Net outperforms existing methods by effectively modeling fine-grained sentiment-aware relationships, leading to improved rumor detection performance.
Sentiment-aware graph networks that reason over fine-grained textual phrases and visual objects can significantly boost rumor detection performance.
The proliferation of multimodal content on social media has increased the complexity of rumor detection, where textual and visual elements often exhibit subtle semantic and emotional inconsistencies. Existing approaches mainly focus on coarse-grained feature fusion while overlooking fine-grained cross-modal relationships and sentiment cues, which serve as crucial signals of deception. This paper proposes FMS-Net, a Fine-Grained Multimodal Sentiment Graph Network that captures both intra-modal and inter-modal relationships, and incorporates sentiment similarities into the graph structure. FMS-Net extracts phrase- and object-level features from texts and images using pre-trained language and vision models. It then constructs a heterogeneous multimodal graph with sentiment-aware edge weighting and applies graph attention mechanisms to enable fine-grained multimodal reasoning. Extensive experiments on real-world datasets demonstrate that FMS-Net consistently outperforms existing unimodal and multimodal baselines, highlighting the critical role of sentiment-aware modeling and fine-grained feature extraction. These results offer actionable insights for designing more robust frameworks to combat misinformation across diverse social media environments.
