TakeLab at SemEval-2016 Task 6: Stance Classification in Tweets Using a Genetic Algorithm Based Ensemble

Proceedings of SemEval-2016, pages 464–468, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics TakeLab at SemEval-2016 Task 6: Stance Classification in Tweets Using a Genetic Algorithm Based Ensemble Martin Tutek, Ivan Sekuli´c, Paula Gombar, Ivan Paljak, Filip ˇCulinovi´c, Filip Boltuˇzi´c, Mladen Karan, Domagoj Alagi´c, Jan ˇSnajder University of Zagreb, Faculty of Electrical Engineering and Computing Text Analysis and Knowledge Engineering Lab Unska 3, 10000 Zagreb, Croatia name.surname@fer.hr Abstract This paper describes our system for the detec- tion of stances in tweets submitted to SemEval 2016 Task 6A. The system uses an ensemble of learning algorithms, fine-tuned using a genetic algorithm. We experiment with various off- the-shelf classifiers and build our model using standard lexical and a number of task-specific features. Our system ranked 3rd among the 19 systems submitted to this task. 1 Introduction Stance is the overall position held by a person to- wards an object, idea, or proposition (Somasundaran and Wiebe, 2009). The task of stance detection – the automatic classification of stance expressed in text – has been attracting increasing interest, as it is of practical interest to many stakeholders, rang- ing from political bodies to companies. Most recent work focused on stance detection in online debates (Somasundaran and Wiebe, 2009; Somasundaran and Wiebe, 2010; Anand et al., 2011; Hasan and Ng, 2014; Sridhar et al., 2015). Twitter is an outstanding platform for large-scale stance analysis. However, unlike in the case of dedicated online debate platforms, Twitter data is much less structured and dialogical. Furthermore, as pointed out by Rajadesingan and Liu (2014), pro- cessing of tweets poses specific challenges stemming from the high volume of data, brevity of messages, and the use of non-standard language. In this paper, we describe a system for stance clas- sification in tweets, with which we participated in the SemEval-2016 Task 6A. Given a tweet and the topic of the tweet (the target), the task was to predict the stance of the tweet author as either in FAVOR of the target, AGAINST it, or NONE. Our system relies on a supervised three-way classifier, using features stan- dardly employed for stance classification and similar tasks, but also a number of task- and target-specific features. The gist of our approach was to start out with the entire “kitchen sink” of features, as well as a number of off-the-shelf classification algorithms, and then perform a comprehensive model optimization on a per-target basis. However, instead of relying on a single model for each target, we trained four different models and combined them into one classifier ensem- ble for each target using a genetic algorithm. Our system (TakeLab) ranked 3rd among the 19 systems submitted to the SemEval-2016 Task 6A. 2 Related Work To the best of our knowledge, the only work to have addressed the stance classification in tweets is that of Rajadesingan and Liu (2014), who specifically tackle the data sparseness problem using a semi-supervised approach based on label propagation. Somasundaran and Wiebe (2009) and (2010) address stance detec- tion in …

Content truncated to 500 words