Topic-Guided Stance Detection for Comparing Public Opinion Surveys with Tweets about Covid-19

Anonymous

Keywords: stance detection, covid-19, few shot, topic classification

Abstract: Understanding public opinion, including hesitancy and scepticism towards Covid-19, is important to create appropriate public health policies. Such opinions are traditionally manually collected through surveys, though automatically measuring them through social media offers a larger reach. However, this then poses the important question of to what degree public opinion surveys and stances expressed on social media align. In this paper, we propose a new setting and method for gauging public opinion through Twitter and analysing its alignment to surveys, which we evaluate in the context of stances towards topics surrounding Covid-19 voiced by people in eight countries. Stance detection is typically framed as a pairwise sequence classification task where stance targets are provided. As this is not the case for plain tweets, we propose an alternative framing of the task, namely first identifying the tweet topic and subsequently classifying the stance towards it. To provide effective minimal supervision for training a topic-guided stance detection model, we introduce a novel topic-guided annotation technique TOGA based on unsupervised deep topic modelling and apply it to an unlabelled dataset of tweets about Covid-19. In a proxy evaluation of our method on an existing labelled stance detection dataset from the same domain, we find that our few-shot method outperforms other, fully supervised approaches by $18.1$ F1 points. Lastly, we show that our approach can be used effectively in conjunction with public opinion surveys for measuring public opinion and that there is a weak correlation of predicted stances with those reported in surveys.

Paper Type: long

Research Area: Computational Social Science and Cultural Analytics