Measuring Customer Sentiment on Twitter

Twitter, a microblogging platform, allows its users to post short messages about any topic and follow others to receive their posts. By means of Twitter people communicate with each other. The goal of this research is to study customer sentiment expressed on Twitter and to develop a framework that allows monitoring it in the real-time. For this purpose 9368 tweets are collected from the KLM Twitter account. Tweets preparation including among others spelling correction, synonym substitution, hyperlink deletion and stop words is performed. Sentiment is manually categorized the sentiment into three classes: objective, positive and negative, in order to create the training set for the classifier. The tweets are classified using linear Support Vector Machines. The classifier obtains 82% precision for the objective class, 59% for the positive and 54% for the negative class. The classified tweets are used to create the positive and negative emotion indexes over time. Looking at the different subset of features created by the ranking algorithm shows how different words influence the prediction and helps to gain insight into the classification. It is shown that the data preparation improves both the precision and recall of the classification. The spelling correction and synonym substitution improves the precision for the negative class of tweets by up to 8%. Relating the predicted sentiment to various events such as the introduction of a new service or operational issues with flights showed how such events influence customer opinion.