Spatio-Temporal Feature-Enhanced Sarcasm Detection

Sarcasm constitutes a complex linguistic phenomenon wherein critical intent is articulated through semantic contradiction. Sarcasm detection plays a vital role in opinion mining on social media platforms. Existing approaches often fail to effectively model the temporal dependencies associated with evolving emotional states, and lack the capacity to adequately capture semantic interactions across different speakers. To address these issues, we propose a spatio-temporal featureenhanced multimodal graph neural network that jointly models emotional dynamics and cross-speaker contextual dependencies. We construct a heterogeneous spatio-temporal graph where each utterance is represented as a node embedding that integrates textual, acoustic, visual features, along with implicit emotional representations. To capture various forms of dialogue dependencies, we design three types of edges. Temporal edges model the progression of emotions across consecutive utterances. Interactive edges represent immediate responses between different speakers. Global context edges connect each utterance to the overall semantic structure of the dialogue. Experimental results on the MUStARD-extension dataset demonstrate that our framework surpasses M2Seq2Seq baseline, achieving improvements of 0.8% in precision and 0.9 in F1-score. They confirm the effectiveness of heterogeneous spatio-temporal modeling in enhancing multi-turn sarcasm detection.