MACCISA: A Multimodal Commonsense Knowledge Aware Model for Counterfactual Implicit Sentiment Analysis

Jian Liao, Wu Han, Yiyang Zhang, Xuan Zhao, Suge Wang, Jianxing Zheng, Xin Guo, Jianjun Li

Abstract: Counterfactual expressions challenge identifying sentiments that contradict actual or potential circumstances. Counterfactual Implicit Sentiment Analysis (CISA) remains underexplored due to limited corpora and modeling constraints. We introduce two bilingual corpora and a multimodal commonsense knowledge graph to bridge this gap. We also propose MACCISA, a novel model integrating: (1) LLM-based counterfactual sequence encoding; (2) dynamic graph neural network for multimodal commonsense knowledge; (3) multimodal interactive attention for semantic enhancement. Extensive experiments show that MACCISA outperforms previous state-of-the-art methods.

External IDs:dblp:conf/icic/LiaoHZZWZGL25