Adversarial Network with External Knowledge for Zero-Shot Stance Detection

Adversarial Network with External Knowledge for Zero-Shot Stance Detection Chunling Wang1, Yijia Zhang1*, Xingyu Yu1, Guantong Liu2, Fei Chen1, Hongfei Lin3 1College of Information Science and Technology, Dalian Maritime University / Dalian, China 2College of Artificial Intelligence, Dalian Maritime University / Dalian, China 3College of Computer Science and Technology, Dalian University of Technology / Dalian, China {wangchunling, zhangyijia, 1120211509, lgt, chenf}@dlmu.edu.cn hflin@dlut.edu.cn Abstract Zero-shot stance detection intends to detect previously unseen targets’ stances in the testing phase. However, achieving this goal can be difficult, as it requires minimizing the domain trans- fer between different targets, and improving the model’s inference and generalization abilities. To address this challenge, we propose an adversarial network with external knowledge (ANEK) model. Specifically, we adopt adversarial learning based on pre-trained models to learn transfer- able knowledge from the source targets, thereby enabling the model to generalize well to unseen targets. Additionally, we incorporate sentiment information and common sense knowledge into the contextual representation to further enhance the model’s understanding. Experimental re- sults on several datasets reveal that our method achieves excellent performance, demonstrating its validity and feasibility. Keywords: Zero-shot stance detection , Adversarial learning , External knowledge , Contrastive learning 1 Introduction Stance detection (K¨uc¸¨uk and Can, 2020; Mohammad et al., 2016; Augenstein et al., 2016) is a significant task in NLP, focusing on identifying the stance (e.g., against, favor, or neutral) conveyed in the text towards a given target. It can be efficiently applied to social opinion analysis (Lai et al., 2020), rumor detection (Kumar and Carley, 2019), and other research fields by mining text opinions. Traditional intra-target stance detection (Mohammad et al., 2016) has limited applications since it re- quires training and testing under the same target and depends heavily on labeled data to achieve excellent performance. With the frequent and vast updates of topics on social platforms, manually labeling new targets becomes expensive and time-consuming, making it impractical to create a labeled dataset with all potential targets (Wang et al., 2020). Therefore, the study of zero-shot stance detection (Allaway and Mckeown, 2020) for unseen targets is essential and promising. To tackle the zero-shot stance detection task, existing works generally incorporate external knowledge (Liu et al., 2021) as support for inference or introduce attention mechanisms (Allaway and Mckeown, 2020) to capture the relationships between targets, which do not explicitly model of the transferable knowledge between source and destination targets. Some methods solely focus on employing adversarial training (Allaway et al., 2021; Xie et al., 2022) to learn a target-invariant representation of the text content, disregarding the possibility that the model may encounter challenges in correctly predicting sentences that contain implicit viewpoints or require more profound understanding. For example 1 in Table 1, the document does not explicitly mention the target ”Donald Trump.” If the model is unaware that Donald Trump is affiliated with the Republican Party, it is easy to misclassify the stance as neutral. Therefore, by incorporating common sense knowledge into adversarial networks and supplementing the target-related concept representations in the knowledge …

Content truncated to 500 words