NILC-USP at SemEval-2017 Task 4: A Multi-view Ensemble for Twitter Sentiment Analysis

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 611–615, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics NILC-USP at SemEval-2017 Task 4: A Multi-view Ensemble for Twitter Sentiment Analysis Edilson A. Corrˆea Jr., Vanessa Queiroz Marinho, Leandro Borges dos Santos Institute of Mathematics and Computer Science University of S˜ao Paulo (USP) S˜ao Carlos, S˜ao Paulo, Brazil {edilsonacjr,vanessaqm,leandrobs}@usp.br Abstract This paper describes our multi-view en- semble approach to SemEval-2017 Task 4 on Sentiment Analysis in Twitter, specif- ically, the Message Polarity Classification subtask for English (subtask A). Our sys- tem is a voting ensemble, where each base classifier is trained in a different feature space. The first space is a bag-of-words model and has a Linear SVM as base clas- sifier. The second and third spaces are two different strategies of combining word em- beddings to represent sentences and use a Linear SVM and a Logistic Regressor as base classifiers. The proposed system was ranked 18th out of 38 systems considering F1 score and 20th considering recall. 1 Introduction Twitter is a microblogging service that has 313 million monthly active users 1. In this social me- dia platform, users interact through short mes- sages, so-called tweets. The company estimates that over 500 million tweets are sent each day 2. Despite of their size, at most 140 characters, these messages provide rich data because users gener- ally write about their thoughts, opinions and senti- ments. Therefore, applications in several domains, such as commercial (Jansen et al., 2009) and po- litical (Tumasjan et al., 2010; Wang et al., 2012), may benefit from the automatic classification of sentiment in tweets. In this paper we show that a multi-view ensem- ble approach that leverages simple representations of texts may achieve good results in the task of message polarity classification. The proposed sys- tem consists of three base classifiers, each of them 1https://about.twitter.com/company 2https://business.twitter.com/en/basics.html with a specific text representation technique (bag- of-words or word embeddings combination). As base classifiers, we use Support Vector Machines (SVM) and Logistic Regression. The proposed ap- proach was evaluated on the SemEval-2017 Task 4 Subtask A for English and was ranked 18th out of 38 participating systems considering F1 score and 20th considering recall. This paper is organized as follows: Section 2 describes our system, the feature spaces and the classifiers we employed. The training and evalu- ation datasets are presented in Section 3. In ad- dition, Section 3 also describes the preprocess- ing steps and some details about the word embed- dings. We present our results in Section 4. Finally, Section 5 outlines our conclusions and remarks on future work. 2 System Description The proposed system consists of a multi-view ensemble with three base classifiers with differ- ent text representation techniques (feature spaces), that is, all base classifiers are trained on the same dataset but with a different representation or fea- ture space. The first is a Linear SVM and the tweets are represented using bag-of-words (Sec- tion 2.1.1). The second is another …

Content truncated to 500 words