User-Level Twitter Sentiment Analysis with a Hybrid Approach

With the objective of extracting useful information from the opinion-rich data on Twitter, both supervised learning-based and unsupervised lexicon-based methods for sentiment analysis on Twitter corpus have been studied in recent years. However, the unique characteristics of tweets such as the lack of labels and frequent usage of emoticons poses challenges to most of the existing learning-based and lexicon-based methods. In addition, studies on Twitter sentiment analysis nowadays mainly focus on domain specific tweets while a larger amount of tweets are about personal feelings and comments on daily life events. In this paper, a hybrid approach of augmented lexicon-based and learning-based method is designed to handle the distinctive characteristics of tweets and perform sentiment analysis on a user level, providing us information of specific Twitter users' typing habits and their online sentiment fluctuations. Our model is capable of achieving an overall accuracy of 81.9 %, largely outperforming current baseline models on tweet sentiment analysis.