UWB at SemEval-2016 Task 6: Stance Detection

Proceedings of SemEval-2016, pages 408–412, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics UWB at SemEval-2016 Task 6: Stance Detection Peter Krejzl University of West Bohemia Faculty of Applied Sciences Dept. of Computer Sci. and Eng. Univerzitni 8, 30614 Pilsen Czech Republic pkrejzl@gmail.com Josef Steinberger University of West Bohemia Faculty of Applied Sciences NTIS Centre Univerzitni 8, 30614 Pilsen Czech Republic jstein@kiv.zcu.cz Abstract This paper describes our system participating in the SemEval 2016 task: Detecting stance in Tweets. The goal was to identify whether the author of a tweet is in favor of the given target or against. Our approach is based on a maximum entropy classifier, which uses surface-level, sentiment and domain-specific features. We participated in both the super- vised and weakly supervised subtasks and re- ceived promising results for most of the tar- gets. 1 Introduction Stance detection has been defined as automatically detecting whether the author of a piece of text is in favor of the given target or against it. In the third class, there are the cases, in which neither inference is likely. It can be viewed as a subtask of opinion mining and it stands next to the sentiment analysis. The significant difference is that in the case of sen- timent analysis, systems determine whether a piece of text is positive, negative, or neutral. However, in stance detection, systems are to determine the author’s favorability towards a given target and the target even may not be explicitly mentioned in the text. Moreover, the text may express positive opin- ion about an entity contained in the text, but one can also infer that the author is against the defined target (an entity or a topic). This makes the task more dif- ficult, compared to the sentiment analysis, but it can often bring complementary information. There are many applications which could benefit from the automatic stance detection, including infor- mation retrieval, textual entailment, or text summa- rization, in particular opinion summarization. Twit- ter was selected as the source of the text because of its popularity and because people express stance im- plicitly or explicitly there. We first shortly introduce the task in Section 2 and the available dataset.1 In Section 3, we describe our preprocessing, the implemented approach and sys- tem’s features. It is followed by the setup for each analysed topic and a discussion of official results (Section 4). 2 Task Description The Detecting Stance in Tweets task2 (Mohammad et al., 2016) had two independent subtasks: super- vised and weakly supervised stance identification. The supervised task (subtask A) tested stance to- wards five targets: Atheism, Climate Change is a Real Concern, Feminist Movement, Hillary Clinton, and Legalization of Abortion. Participants were pro- vided 2814 labeled training tweets for the five tar- gets. An example tweet annotated as IN FAVOR: These pics of #pornstars with/without makeup? Just perpetuating the myth that women need makeup to be considered pretty. (the Feminist Movement tar- get, ID: 1017). A detailed distribution of stances for each target …

Content truncated to 500 words