Multi-Modal Sarcasm Detection via Graph Convolutional Network and Dynamic Network

Sarcasm is a form of language used to convey implicit information contradicting the literal meaning of words, often observed on online social media platforms. Accurately detecting satirical or ironic expressions could significantly enhance sentiment analysis and opinion mining. For multi-modal data, capturing both inter- and intra-modal incongruities is crucial for this task. Recently, graph-based approaches to modeling incongruous features bet-ween image and text have made significant progress in this task. However, these methods rely on static networks to capture incongruous features, which makes them inflexible in adapting to diverse groups of text and image, or neglect important information due to inadequate use of text and image. To address these limitations, we propose a multi-modal sarcasm detection model based on the combination of Graph Convolutional Network and Dynamic Network. The graph convolutional network learns the incongruity of the three modal graphs and makes full use of the object-level information. The dynamic network dynamically captures the incongruity between the global-level image and the text and can flexibly adapt to different image and related text. At the same time, we generate augmented text to better utilize the text information. Extensive experiments demonstrate that our proposed method performs favorably against state-of-the-art approaches.