A meta-contrastive learning with data augmentation framework for zero-shot stance detection

Abstract

Zero-shot stance detection (ZSSD) identifies the stances of targets that have not been encountered during the testing phase. Most of the existing efforts are dedicated to enhancing the generalizability of models while ignoring data issues such as data scarcity and targets that are not explicitly mentioned. Therefore, we consider approaching this task from both the data and model perspectives and propose a meta-contrastive learning with data augmentation framework. We first use a generation model to generate target keyphrases for enhancing the original text. Then, we utilize a meta-learning technique that incorporates contrastive learning for improving the generalizability of the model, enabling it to better adapt to unknown targets. The experimental results obtained on three benchmark datasets prove that our framework achieves outstanding performance improvements in the ZSSD task. Our code is available at https://github.com/qifen37/MCLDA.