From Tweets to Stance: An Unsupervised Framework for User Stance Detection on Twitter

Current stance inference methods use topic-aligned training data, leaving many unexplored topics due to the lack of training data. Zero-shot approaches utilizing advanced pre-trained Natural Language Inference (NLI) models offer a viable solution when training data is unavailable. This work introduces the Tweets2Stance - T2S framework, an unsupervised stance detection framework based on Zero-Shot Learning. It detects a five-valued user's stance on social-political statements by analyzing their Twitter timeline. The ground-of-truth user's stance is obtained from Voting Advice Applications (VAAs), online tools that compare political preferences with party political stances. The T2S framework's generalization potential is demonstrated by measuring its performance (F1 and MAE scores) across nine datasets. These datasets were built by collecting tweets from competing parties' Twitter accounts in nine political elections held in different countries from 2019 to 2021. Through comprehensive experiments, an optimal setting was identified for each election. The results, in terms of F1 and MAE scores, outperformed all baselines and approached the best scores for each election. This showcases the ability of T2S to generalize across different cultural-political contexts.