Improving few-shot stance detection with MINE: meta-prompting and in-context N-shot examples

Abstract

Stance detection aims to identify an author’s stance toward a target topic. Although recent prompting-based methods enable zero- and few-shot stance detection with large language models (LLMs), they often require extensive prompting or costly use of multiple LLM agents, as well as large labeled training datasets. We present MINE (Meta-prompting and In-context N-shot Examples), a simple, light-weight, training-free framework that first creates a target-specific LLM stance agent through meta-prompting, and then augments it with a small number of semantically retrieved examples. While meta-prompting and retrieval-augmented in-context learning have been studied independently, these components have not been combined specifically for stance detection in a training-free pipeline to the best of our knowledge. We also provide a systematic empirical study that quantifies the effects of the example retrieval strategy, sentence encoder choice, demonstrator pool size, and the number of examples on the performance of this hybrid methodology. We show that MINE achieves competitive performance when deployed as either a zero-shot or few-shot stance detector, matching or exceeding existing prompting-only approaches on SemEval-2016 and P-Stance while using fewer prompts and demonstrations. MINE remains effective with minimal labeled data (e.g., 10% of the training demonstrator pool) and generalizes to related tasks, including aspect-based sentiment analysis and cross-target stance detection.