Zero-shot Tweet-Level Stance Detection Enhanced by External Knowledge and Reflective Chain-of-Thought Reasoning

Zero-shot stance detection (ZSSD) aims to identify attitudes toward targets without prior training on those specific topics. This task faces two major hurdles: the inherent sparsity of short social media texts and the difficulty of distinguishing between 'neutral' stances (where a user has an opinion but it is balanced) and 'irrelevant' content (where the text is unrelated to the target). The authors seek to address these challenges by improving how models integrate external knowledge and reason about implicit targets.

The authors propose the KIRP framework, which consists of three main components:

1. Data Augmentation via Entity Reorganization: Using knowledge graphs, the model identifies and replaces or supplements key entities within a tweet. This makes implicit targets explicit and creates diverse, target-oriented variations of the text.
2. Reflective Chain-of-Thought (CoT) Reasoning: The model uses prompt chaining to extract and validate implicit targets, generating reasoning evidence across four perspectives (favor, against, neutral, irrelevant) to provide deeper context for the classifier.
3. Stance-Aware Contrastive Learning: To better separate 'neutral' from 'irrelevant' labels, the model employs a contrastive learning objective that maximizes the distance between different stance categories in the feature space, followed by a three-layer iterative prototype network for final classification.

Additionally, the authors introduce KIRP-D, a new four-class, multi-topic Japanese tweet dataset covering social, political, diplomatic, and economic domains. This is presented as the first Japanese-language dataset specifically designed for zero-shot stance detection.

KIRP achieves state-of-the-art performance across multiple benchmarks. On the SemEval-2016 dataset, it reached an F1 score of 84.05% for three-class classification. On the WT-WT and the new KIRP-D datasets, it achieved F1 scores of 84.99% and 79.18% respectively for four-class classification. The results suggest that combining external knowledge injection with reflective reasoning significantly improves the model's ability to handle short, context-sparse texts and unseen targets.