Target-Phrase Zero-Shot Stance Detection: Where Do We Stand?

Stance detection, i.e. recognition of utterances in favor, against or neutral in relation to some targets is important for text analysis. However, different approaches were tested on different datasets, often interpreted in different ways. We propose a unified overview of the state-of-the-art stance detection methods in which targets are expressed by short phrases. Special attention is given to zero-shot learning settings. An overview of the available multiple target datasets is presented that reveals several problems with the sets and their proper interpretation. Wherever possible, methods were re-run or even re-implemented to facilitate reliable comparison. A novel modification of a prompt-based approach to training encoder transformers for stance detection is proposed. It showed comparable results to those obtained with large language models, but at the cost of an order of magnitude fewer parameters. Our work tries to reliably show where do we stand in stance detection and where should we go, especially in terms of datasets and experimental settings.