Social media user stance detection without stance text

Abstract

Context: User stance detection systems are intended to determine whether an individual holds a stance for or against a given target topic. For instance, ‘Vaccines save lives’ is a stance in favour of vaccination. Problem: Most existing approaches to stance detection take as an input a piece of text assumed to convey a stance towards the intended target, and then predict its polarity. However, when stance text is unavailable (e.g., when the individual did not publish any opinion regarding that particular topic), models of this kind are unable to determine the user’s stance. Solution: In the case of social media, we may consider as a substitute for stance text the use of other (non-stance) publications made by the user, or even by their close contacts. Information Systems theory: The work was considering Social Network Theory and Social Information Processing Theory. Method:We built a number of stance detection models from stance text and from non-stance data authored by the target users and their close contacts using supervised BERT, prompt-based LLama, and stacking. Results: The non-stance publications made by an individual and, to a lesser extent, those made by his/her contacts, were found to be strong predictors of user stance and, in some cases, their combination was shown to outperform the actual use of text stance. Contributions: Using network-related information may be indeed a substitute for stance text, and may enable the development of systems for scenarios in which stance text is unavailable.