<p>Cross-Target Stance Detection (CTSD) aims to identify stances towards diverse targets that are unseen or lack sufficient labeled data during the training phase. Compared with traditional target-specific stance detection methods, CTSD has stronger cross-target generalization ability and can adapt to diverse scenarios, especially demonstrating unique advantages in handling real-time hot events and emerging fields. However, this task is significantly challenging because the semantic information and background knowledge of unseen targets cannot be obtained through direct supervised learning. To address this, this paper proposes a cross-target stance detection model via adversarial learning that incorporates background knowledge and sentiment information. The model mainly consists of two core components: background information encoding and an adversarial network that fuses sentiment and stance information. First, external knowledge encoding from ConceptNet provides background information for instances. Then, a sentiment encoder extracts sentiment features, and the target-irrelevant features that integrate sentiment and stance information are input into the adversarial network to further learn transferable feature representations. This approach significantly enhances the model’s reasoning ability for unseen targets. Experimental results on the SEM16 and WT-WT datasets show that this method achieves good performance, demonstrating its effectiveness and feasibility.</p>