Abstract

Stance detection is the task of determining whether an input text expresses a stance in favour of or against a given target topic. This, in a standard supervised fashion, will typically require a new set of labelled training examples for each test topic. As an alternative to full supervision (or costly LLM-based methods), this study leverages political alignment information by assuming that stances on related moral or political issues tend to co-occur (e.g., support for a right-wing politician correlating with support for the death penalty or opposition to abortion). This alignment, presently treated as a form of distance labelling, enables stance inference without constructing new corpora and is evaluated against standard cross-domain and prompt-based methods using a large corpus of stances in the Portuguese language.