I2RNTU at SemEval-2016 Task 4: Classifier Fusion for Polarity Classification in Twitter

Proceedings of SemEval-2016, pages 71–78, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics I2RNTU at SemEval-2016 Task 4: Classifier Fusion for Polarity Classification in Twitter Zhengchen Zhang1, Chen Zhang2, Fuxiang Wu3, Dong-Yan Huang1, Weisi Lin2, Minghui Dong1 1Human Language Technology Department, Institute for Infocomm Research (I2R), A*STAR, Singapore, 138632 2School of Computer Engineering, Nanyang Technological University, 50 Nanyang Avenue Singapore, 639798 3BeiHang University, Beijing, China, 100191 huang@i2r.a-star.edu.sg Abstract In this work, we apply classifier fusion to tweet polarity identification problem. The task is to predict whether the emotion hidden in a tweet is positive, neutral, or negative. An asymmetric SIMPLS (ASIMPLS) based clas- sifier, which was proved to be able to identify the minority class well in imbalanced classi- fication problems, is implemented. Word em- bedding is also employed as a new feature. For each word, we obtain three word embedding vectors on positive, neutral, and negative tweet sets respectively. These vectors are used as features in the ASIMPLS classifier. Another three state-of-the-art systems are implemented also, and these four systems are fused together to further boost the performance. The fusion system achieved 59.63% accuracy on the 2016 test set of SemEval2016 Task 4, Subtask A. 1 Introduction The I2RNTU system works on the Subtask A: Mes- sage Polarity Classification in Twitter of SemEval- 2016 Task 4: the Sentiment Analysis in Twit- ter (Nakov et al., 2016). The task is to predict whether a tweet is of positive, neutral, or negative sentiment. This task can be formulated as a multi- class classification problem, i.e., to classify a tweet into one of the three classes. We use the one-vs-rest strategy to solve the three-class classification prob- lem. Given a tweet, a classifier generates three con- fidence scores about the tweet belonging to the three classes respectively. The predicted label is chosen based on the highest confidence score. Four clas- sifiers are implemented in our work, and classifier fusion is used to improve the system performance. Classifier fusion has been proved to be very pow- erful in classification problems. In SemEval-2015, a system named Webis won the first place in the mes- sage polarity classification subtask, which is subtask B of SemEval-2015 task 10 “Sentiment Analysis in Twitter” (Hagen et al., 2015). The authors repro- duced four state-of-the-art twitter polarity prediction algorithms. Each algorithm generates three confi- dence scores for a tweet. The fusion system aver- ages the scores generated by the four classifiers, and then predicts a label according to the highest average score. In the Speaker State Challenge of INTER- SPEECH 2011, the method of fusing Asymmetric SIMPLS (ASIMPLS) and Support Vector Machines (SVMs) won the sleepiness sub-challenge (Huang et al., 2011). The asymmetric SIMPLS based classi- fier is shown to be able to generate a higher pre- diction accuracy for the class with small number of instances in the imbalanced classification prob- lem, while SVMs are strong at predicting the class with majority number of instances. The fusion of these two types of methods could achieve …

Content truncated to 500 words