Stance classification is the task of automatically identify the user’s positions about a specific topic. The classification of stance may help to understand how people react to a piece of target information, a task that is interesting in different areas as advertising campaigns, brand analytics, and fake news detection, among others. The rise of social media has put into the focus of this task the classification of stance in online social networks. A number of methods have been designed for this purpose showing that this problem is hard and challenging. In this work, we explore how to use self-attention models for stance classification. Instead of using attention mechanisms to learn directly from the text we use self-attention to combine different baselines’ outputs. For a given post, we use the transformer architecture to encode each baseline output exploiting relationships between baselines and posts. Then, the transformer learns how to combine the outputs of these methods reaching a consistently better classification than the ones provided by the baselines. We conclude that self-attention models are helpful to learn from baselines’ outputs in a stance classification task.