Detecção semi-supervisionada de posicionamento em tweets baseada em regras de sentimento

Stance detection aims to automatically identify if the text author is in favor or against a subject or target. This work describes a semi-supervised method for stance detection. The core is a set of rules to identify stance based on positive or negative opinions of targets directly or indirectly related. Tweets automatically labeled using the rules compose a training corpus for a supervised approach. The resulting predictive model allows to predict the stance of unlabeled tweets. This paper presents the method and analyzes the obtained results when applied to different data domains like political candidatures, climate change and legalization of abortion.