Previous research mainly uses graph neural networks on syntactic dependency graphs, often neglecting emotional cues in sarcasm detection and failing to integrate image features for multimodal information effectively. To address these limitations, this study proposes a novel multimodal sarcasm detection model based on the directed graph isomorphism network with sentiment enhancement and multimodal fusion (DGIN-SE-MF).
The approach extracts image and text features through vision transformer and BERT, respectively. To deeply integrate the extracted features, the author develops a text-guided multi-head attention fusion mechanism module. Subsequently, a directed graph is constructed through SE and the multimodal factorized bilinear pooling method to integrate image features into the graph. The DGIN then fuses the image and text features, using a weighted attention mechanism to generate the final representation.
The model is validated on three datasets: English, Chinese and an Indonesian–English dataset. The results demonstrate that the proposed model consistently outperforms other baseline models, particularly on the Chinese and English sarcasm datasets, achieving F1 scores of 88.75 % and 83.10 %, respectively.
The proposed model addresses the inadequacies of previous methods by effectively integrating emotional cues and image features into sarcasm detection. To the best of the authors’ knowledge, this is the first work to leverage a DGIN-SE-MF for this task, leading to significant improvements in detection performance across different languages.
