VCU-TSA at Semeval-2016 Task 4: Sentiment Analysis in Twitter

Proceedings of SemEval-2016, pages 215–219, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics VCU-TSA at Semeval-2016 Task 4: Sentiment Analysis in Twitter Gerard Briones and Kasun Amarasinghe and Bridget T. McInnes, PhD. Department of Computer Science Virginia Commonwealth University Richmond, VA, 23284, USA {brioinesgc, amarasinghek, btmcinnes}vcu.edu Abstract The aim of this paper is to produce a method- ology for analyzing sentiments of selected Twitter messages, better known as Tweets. This project elaborates on two experiments carried out to analyze the sentiment of Tweets from SemEval-2016 Task 4 Subtask A and Subtask B. Our method is built from a sim- ple unigram model baseline with three main feature enhancements incorporated into the model: 1) emoticon retention, 2) word stem- ming, and 3) token saliency calculation. Our results indicate an increase in classification ac- curacy with the addition of emoticon reten- tion and word stemming, while token saliency shows mixed performance. These results elu- cidate a possible classification feature model that could aid in the sentiment analysis of Twitter feeds and other microblogging envi- ronments. 1 Introduction Twitter is a widely used microblogging environment which serves as a medium to share opinions on var- ious events and products. Because of this, analyz- ing Twitter has the potential to reveal opinions of the general public regarding these topics. However, mining the content of Twitter messages is a chal- lenging task due to a multitude of reasons, such as the shortness of the posted content and the informal and unstructured nature of the language used. The aim of this study is to produce a methodology for analyzing sentiments of selected Twitter messages, better known as Tweets. This project elaborates on two experiments carried out to analyze the sentiment of Tweets, namely, Subtask A and Subtask B from SemEval-2016 Task 4 (Nakov et al., 2016). Subtask A: Message Polarity Classification. The goal of this subtask was to predict a given Tweet’s sentiment from three classes: 1) positive, 2) neutral, or 3) negative. Subtask B: Tweet classification according to a two-point scale. The goal of this subtask was to classify a given Tweet’s sentiment towards a given topic. The sentiments were limited to positive and negative, unlike Subtask A. 2 Method We viewed both tasks as classification problems. We represented the Tweets in a statistical feature matrix and performed the classification using su- pervised machine learning classification algorithms. Several different feature vectors were experimented with and the same set of feature vectors were applied to both Subtask A and Subtask B. 2.1 Feature Vectors Our methods consume Tweet data and output a ma- trix where each row represents a Tweet and each col- umn represents a feature. The values in this matrix are the frequency of appearance of a feature in the Tweet. As reference for the rest of the project, please note that n-grams are a continuous set of n terms in a document. Thus, when n=1, we are representing a unigram, or a single word; when n=2, we are repre- senting …

Content truncated to 500 words