Improved LinkNet with new aspect term extraction for multi-modal sarcasm detection

The rapid growth of online communication has increased the need for efficient multimodal sarcasm detection models. However, effectively integrating textual, visual, and audio information remains a significant challenge. To tackle this issue, this paper proposes a Conditional Random Field-assisted LinkNet framework (CRF-LN-MSD) that integrates aspect-level textual representations via a Modified Fisher Score-based Aspect Term Extraction (MFS-ATE) technique with facial expression features and audio spectral characteristics. The CRF-LN architecture employs a CRF layer for structured multimodal fusion and a novel Tanh-Swish-Sigmoid (TSS) activation function for improved gradient flow. Experimental results on the M2H2 and CMU-MOSI datasets demonstrate competitive performance under the evaluated experimental conditions, with a precision of 0.978 on M2H2 and a cross-validated accuracy of 0.9787 on CMU-MOSI. These results demonstrate the effectiveness of the proposed architectural refinement for multimodal sarcasm detection within the evaluated scope.