Research on Zero-Shot Stance Detection: Using Dual-Module adversarial training

Stance detection is to automatically determine the stance of a given topic from the document published by the user. Because stance detection is based on the particularity of the topic, it is not practical to collect all topics. Therefore, zero-shot stance detection is of great significance. The existing zero-shot stance detection models can not combine the topic-invariant features while maintaining the stance features between topics and documents. In this study we propose a zero-shot stance detection model that used dual-module step-by-step training, which can combine the topic-invariant features while maintaining the stance features between topics and documents. Our model has better classification performance than the baseline model in the SemEval2016 dataset.