Enhancing Zero-Shot Stance Detection with Contrastive and Prompt Learning.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Social Network Analysis

Background:

• Public opinion evolves rapidly on social networks, especially during unexpected events.
• Zero-shot stance detection is crucial for analyzing opinions without specific training data, requiring models with strong generalization.
• Existing prompt-based learning methods often use uniform prompts, neglecting instance-specific relationships and limiting task-relevant knowledge acquisition.

Purpose of the Study:

• To develop a stance detection model that dynamically enhances the relevance between prompts and specific instances for zero-shot scenarios.
• To improve the generalization ability of models in identifying target-related stance features.
• To address the limitations of uniform prompt patterns in few-shot text classification.

Main Methods:

• Introduced a stance detection model using a gated multilayer perceptron (gMLP) and a tailored prompt learning strategy.
• Employed a control gate mechanism within the gMLP to modulate prompt token influence based on instance semantics, strengthening the instance-prompt connection.
• Integrated contrastive learning to enhance the discriminative power of feature representations.

Main Results:

• The proposed model demonstrated effectiveness in zero-shot stance detection scenarios.
• Achieved a 1.3% improvement over the JointCL model on the VAST benchmark dataset.
• Experimental evaluations on VAST and SEM16 datasets validated the method's performance.

Conclusions:

• Dynamically enhancing instance-prompt relevance is critical for effective zero-shot stance detection.
• The proposed gMLP-based model with contrastive learning offers a robust solution for analyzing public opinion in social networks.
• The approach shows promise for improving stance detection in the absence of specific training examples.