Enhancing Zero-Shot Stance Detection with Contrastive and Prompt Learning

In social networks, the occurrence of unexpected events rapidly catalyzes the widespread dissemination and further evolution of network public opinion. The advent of zero-shot stance detection aligns more closely with the characteristics of stance detection in today’s digital age, where the absence of training examples for specific models poses significant challenges. This task necessitates models with robust generalization abilities to discern target-related, transferable stance features within training data. Recent advances in prompt-based learning have showcased notable efficacy in few-shot text classification. Such methods typically employ a uniform prompt pattern across all instances, yet they overlook the intricate relationship between prompts and instances, thereby failing to sufficiently direct the model towards learning task-relevant knowledge and information. This paper argues for the critical need to dynamically enhance the relevance between specific instances and prompts. Thus, we introduce a stance detection model underpinned by a gated multilayer perceptron (gMLP) and a prompt learning strategy, which is tailored for zero-shot stance detection scenarios. Specifically, the gMLP is utilized to capture semantic features of instances, coupled with a control gate mechanism to modulate the influence of the gate on prompt tokens based on the semantic context of each instance, thereby dynamically reinforcing the instance–prompt connection. Moreover, we integrate contrastive learning to empower the model with more discriminative feature representations. Experimental evaluations on the VAST and SEM16 benchmark datasets substantiate our method’s effectiveness, yielding a 1.3% improvement over the JointCL model on the VAST dataset.