Stance classification in out-of-domain rumours: a case study around mental health disorders

Social media being a prolific source of rumours, stance classification of individual posts towards rumours has gained attention in the past few years. Classification of stance in individual posts can then be useful to determine the veracity of a rumour. Research in this direction has looked at rumours in different domains, such as politics, natural disasters or terrorist attacks. However, work has been limited to in-domain experiments, i.e. training and testing data belong to the same domain. This presents the caveat that when one wants to deal with rumours in domains that are more obscure, training data tends to be scarce. This is the case of mental health disorders, which we explore here. Having annotated collections of tweets around rumours emerged in the context of breaking news, we study the performance stability when switching to the new domain of mental health disorders. Our study confirms that performance drops when we apply our trained model on a new domain, emphasising the differences in rumours across domains. We overcome this issue by using a little portion of the target domain data for training, which leads to a substantial boost in performance. We also release the new dataset with mental health rumours annotated for stance.