Stance detection from text using semi-automatic corpus expansion

Abstract

Computational stance detection---the task of determining, given an input text, the attitude (e.g., for or against) towards a particular target topic---usually makes use of annotated corpus as training data and, since possible topics are in principle unlimited, so is the need for new labelled datasets about every topic of interest. In order to overcome some of these challenges, the present work adapts to the stance prediction task an existing corpus expansion method that has been originally devised for sentiment analysis. The method is applied to a large (46K instances) Brazilian Portuguese corpus conveying stances towards six target topics of moral and/or political nature, achieving a substantial increase in the number of labelled instances. Results from both automatic and human evaluation suggest that adding semi-automatically labelled data to the corpus does not decrease accuracy, and that the majority of these labels are correct.