Weakly Supervised Stance Learning Using Social-Media Hashtags

Extracting stance from a text is fundamental to automate opinion mining on the web. Recently, many models have been proposed to advance the state-of-the-art. However, most of these models are designed for small hand-labeled datasets that reference a single topic or only a few topics. These models do not generalize to new topics and unseen data. On social media, we have observed that certain hashtags (e.g. ‘climatehoax’) carry stance information. In this research, we propose an approach that uses such hashtags for weak supervision to learn stance in the text. First, we use text sentiment to identify potential hashtags that carry stance information. We call theses hashtags stance-tags. Then, we collect a large amount of Twitter data using these stance-tags. Using the stance of stance-tags, we train two kinds of deeplearning models to predict stance in text that doesn’t necessarily contain hashtags. Further, given a few labeled examples, we model the task of finding stance-tags that are most suitable for stance learning as a multi-armed-bandit problem and find solutions using four optimization strategies. We validate our approach with experiments on a hand-labeled dataset that contains stance on five topics, obtaining results comparable to the state-of-the-art. As many hahstags are used together, they are not completely independent. In the future work section, we propose a search strategy that exploits that similarity between hashtags.