Stance characterization and detection on social media
Stance detection refers to the task of identifying a viewpoint as either supporting or opposing a given topic. The current research on socio-political opinion mining on social media is still in its infancy. Most computational approaches in this field are limited to the independent use of textual elements of a user’s posts from social factors such as homophily and network structure. This thesis provides a thorough study of stance detection on social media and assesses various online signals to identify the stance and understand its association with the analysed topic. We explore the task of detecting stance on Twitter, which is a well-known social media platform where people often express stance implicitly or explicitly. First, we examine the relation between sentiment and stance and analyse the inter-play between sentiment polarity and expressed stance. For this purpose, we extend the current SemEval stance dataset by annotating tweets related to four new topics with sentiment and stance labels. Then, we evaluate the effectiveness of sentiment analysis methods on stance prediction using two stance datasets. Second, we examine the multi-modal representation of stance on social media by evaluating multiple stance detection models using textual content and online interactions. The finding of this chapter suggests that using social interactions along with other textual features can improve the stance detection model. Moreover, we show how an unconscious social interaction can reveal the stance. Next, we design an online framework to preserve users’ privacy concerning the implicitly inferred stance on social media. Thus, we evaluate the effectiveness of the two stance obfuscation methods and use different stance detection models to measure the overall performance of the proposed framework. Finally, we study the dynamics of polarized stance to understand the factors that influence online stance. Particularly, we extend the analysis of online stance signals and examine the interplay between stance and automated accounts (bots). Furthermore, we pose the problem of gauging the bots’ effect on polarized stance through a sole focus on the diffusion of bots on the online social network.