Multi-task Learning for Detecting Stance in Tweets.

Detecting stance of online posts is a crucial task to understand online content and trends. Existing approaches augment models with complex linguistic features, target-dependent properties, or increase complexity with attention-based modules or pipeline-based architectures. In this work, we propose a simpler multi-task learning framework with auxiliary tasks of subjectivity and sentiment classification. We also analyze the effect of regularization against inconsistent outputs. Our simple model achieves competitive performance with the state of the art in micro-F1 metric and surpasses existing approaches in macro-F1 metric across targets. We are able to show that multi-tasking with a simple architecture is indeed useful for the task of stance classification.