Meta-learning with topic-agnostic representations for zero-shot stance detection

Abstract

Zero-shot stance detection (ZSSD) aims to classify stances on unseen topic data with the ability of a model to generalize in the range of topics in the real world. The key point of ZSSD is to prevent the model from overfitting on seen topic data and show robust stance recognition performance on unseen topic samples. In this paper, we propose a new ZSSD framework to generalize stance detection performance by alleviating seen data information from the encoder and focusing on improving stance classification ability in zero-shot conditions. The proposed framework involves two learning stages contributing to ZSSD: topic-agnostic text encoder learning and zero-shot meta-learning. Our framework achieves notable improvements on the three benchmark zero-shot stance detection datasets and zero-shot aspect target sentiment classification dataset showing the effectiveness of our method in the zero-shot settings.