ltl.uni-due at SemEval-2016 Task 6: Stance Detection in Social Media Using Stacked Classifiers

Proceedings of SemEval-2016, pages 428–433, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics ltl.uni-due at SemEval-2016 Task 6: Stance Detection in Social Media Using Stacked Classifiers Michael Wojatzki Language Technology Lab University of Duisburg-Essen Duisburg, Germany michael.wojatzki@uni-due.de Torsten Zesch Language Technology Lab University of Duisburg-Essen Duisburg, Germany torsten.zesch@uni-due.de Abstract In this paper, we describe our participation in the first shared task on automated stance de- tection (SemEval 2016 Task 6). We consider the task as a multidimensional classification problem and thus use a sequence of stacked classifiers. For subtask A, we utilize a rich feature set that does not rely on external infor- mation such as additional tweets or knowledge bases. For subtask B, we rely on the similarity of tweets in this task with tweets from subtask A in order to transfer the models learnt in sub- task A. 1 Introduction Stance-taking is an essential and frequently ob- served part of online debates and other related forms of social media interaction (Somasundaran and Wiebe, 2009; Anand et al., 2011). In the SemEval 2016 Task 6: Detecting Stance in Tweets (Mohammad et al., 2016), stance is defined relative to a given target like a politician or a controversial topic. A text can then either be in favor of the given target (FAVOR), or against it (AGAINST). As the dataset also contains texts without a stance, we ad- ditionally have to deal with the the class NONE. Being able to automatically detect and classify stance in social media is important for a deeper un- derstanding of debates and would thus be a great tool for information seekers such as researchers, journalists, customers, users, companies, or govern- ments. In addition, such analysis could help to cre- ate summaries, develop a deeper understanding of online debating behavior, identify social or political groups, or even adjust recommendations to users’ standpoints (Anand et al., 2011; Sridhar et al., 2014; Boltuzic and ˇSnajder, 2014). In the following, we describe our system for stance detection. We did not make use of any sources of external information such as additional tweets or stance knowledge bases, as our goal was to rely only on the provided training data. Since the task allowed just for one submission, we include some further analysis that will shed light on the use- fulness and impact of the used features and parame- ters. 2 Subtask A – Supervised Framework The goal of this subtask is to classify tweets about five targets: Atheism, Climate Change is a Real Con- cern, Feminist Movement, Hillary Clinton, and Le- galization of Abortion. For each target, there are about 400-600 manually labeled tweets that can be used for training. As the targets are quite different, we train a separate classifier for each of them. Additionally, we split the three-way classification into a stacked classification, in which we first classify whether the tweet contains any stance (classes FAVOR and AGAINST) or no stance at all (class NONE). In a sec- ond step, we classify the tweets labeled …

Content truncated to 500 words