Multilingual and Multi-Domain Rumour Stance Classification

Rumour stance classification focuses on a conversation initialised by a rumour-related source post on social media, aiming to determine the stance of each reply’s author towards the rumour. Accurately capturing public stance facilitates assessing the verdict or check-worthiness of the information. Given the diversity and multilingual nature of online rumours, systems must be able to generalise and adapt to diverse domains and languages. This thesis addresses these issues by investigating methods for improving the generalisation and adaptation of rumour stance classification across new rumours, domains, and languages. Firstly, we identify the distinction between rumour stance classification and generic stance classification by analysing the special role of the stance target (i.e, rumour) in model generalisation. We propose a new ensemble-based method to enhance the reasoning with rumours, achieving state-of-the-art performance. Secondly, we assess the generalisability of top-performing models under domain shift, and propose a LLM-assisted self-training framework for effective adaptation without access to both source and target domain labelled data. Thirdly, we reveal how class labels design in prompts affect LLMs’ generalisation in zero-shot in-Context Learning (ICL) (e.g, the lexical choice between “agree” and “support” for positive stance), and introduce an efficient post-hoc method for optimal label selection. Furthermore, this thesis investigates the adaptation of English-centric rumour stance classification models to non-English languages. We create the largest multilingual benchmark dataset with nine diverse high- and medium-resource languages. We then reveal performance inconsistency across languages and further analyse strategies to improve model performance.