ECNU at SemEval 2016 Task 6: Relevant or Not? Supportive or Not? A Two-step Learning System for Automatic Detecting Stance in Tweets

Proceedings of SemEval-2016, pages 451–457, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics ECNU at SemEval-2016 Task 6: Relevant or Not? Supportive or Not? A Two-step Learning System for Automatic Detecting Stance in Tweets Zhihua Zhang1, Man Lan1,2∗ 1Department of Computer Science and Technology, East China Normal University, Shanghai, P.R.China 2Shanghai Key Laboratory of Multidimensional Information Processing 51131201039@ecnu.cn, mlan@cs.ecnu.edu.cn∗ Abstract This paper describes our submissions to Task 6, i.e., Detecting Stance in Tweets, in SemEval 2016, which aims at detecting the stance of tweets towards given target. There are three stance labels: Favor (directly or indirectly by supporting given target), Against (directly or indirectly by opposing or criticizing given tar- get), and None (none of the above). To ad- dress this task, we present a two-step learning system, which performs two steps, i.e., rele- vance detection and orientation detection, in a pipeline-based processing procedure. Our sys- tem ranked the 5th among 19 teams. 1 Introduction Social platforms, such as Twitter, Facebook, etc., have attracted hundreds of millions of people to share and express their opinions or standpoints in the past few years. Promoted by that growth, re- searchers have been enthusiastic about mining use- ful information in these abundant free texts from so- cial platform, such as stance detection. Determin- ing the stance expressed in a post written for certain target is a relatively new task in sentiment analysis. Classifying stance involves identifying the target of the post and determining its sentiment orientation. The general researches just focus on detecting the stance of posts where the provided posts are rele- vant to the given target (Somasundaran et al., 2007; Somasundaran and Wiebe, 2010). Besides, the pre- vious work usually aims at the posts collected from forums which have co-posts as reference (Murakami et al., 2007; Agrawal et al., 2003). Some approach- es were adopted to settle stance detection, for ex- ample, Murakami and Agrawal detected the stance in the posts collected from forums and adopted co- posts as reference (Murakami et al., 2007; Agrawal et al., 2003). The task of Detecting Stance in Tweets (DST) in SemEval 2016 aims at classifying the provided tweets into three stance classes, i.e., Favor (direct- ly or indirectly by supporting given target), Against (directly or indirectly by opposing or criticizing giv- en target) and None (none of the above) refer to a given target. The DST task consists of two subtasks which could be summarized as supervised subtask (i.e., subtask A) and weakly supervised subtask (i.e., subtask B). The supervised subtask is to test the s- tances of certain tweets towards five predefined tar- gets with labeled training data, while the weak su- pervised subtask is to detect the stances of tweets towards one target with the aid of a mass of unla- beled training data. Somasundaran showed that the stance classifier trained on unigram is a relatively strong baseline (Somasundaran and Wiebe, 2010). Based on So- masundaran and Wiebe’s work, Anand augmented the n-gram features with several linguistic features (Anand …

Content truncated to 500 words