Stance detection, the task of identifying the stance expressed in a text toward a specific target, is essential for analyzing public opinion across diverse domains. The existing approaches primarily focus on modeling the semantic relationship between the text and target, but they often struggle when the target is implicit or indirectly referenced. In real-world scenarios, stance is frequently conveyed through references to related entities, events, or contextual implications, making stance detection particularly challenging. To tackle this challenge, we propose a novel framework that leverages large language models to construct a heterogeneous stance network from textual data. Based on this network, we develop two complementary methodologies tailored for distinct application scenarios: (1) In a supervised setting, we employ a graph neural network approach to learn stance representations from the heterogeneous stance network, enhancing stance prediction performance. (2) For zero-shot stance detection, we introduce an LLM-based method that leverages the heterogeneous stance network to infer stance without task-specific supervision. The experimental results on benchmark datasets demonstrate that our methods outperform the existing approaches, highlighting their effectiveness in both supervised and zero-shot scenarios.

Stance detection identifies the attitude or stance toward specific targets and has a wide range of applications across various domains. Implicit mention of the target makes it difficult to establish connections between the target and the text. Existing approaches focus on integrating external knowledge but overlook the complex associations and stance information within it, which is crucial for maintaining reasoning consistency. To solve this problem, we propose a logical stance detection framework based on the principle of stance transitivity. Our framework achieves stance reasoning by leveraging symbolic and natural language reasoning. Specifically, we extract a list of targets related to the specific target from unlabeled data and use LLMs to construct a target-stance graph. This allows us to examine complicated interactions between targets and integrate stance information across related targets. We conducted massive experiments to validate the effectiveness of our proposed method. The experimental results indicate that our framework significantly improves the performance of stance detection tasks, offering a robust solution to the challenges posed by implicit targets.