A Lexicon-Based Multi-class Semantic Orientation Analysis for Microblogs

In the literature, most of existing works of semantic orientation analysis focus on the distinguishment of two polarities (positive and negative). In this paper, we propose a lexicon-based multi-class semantic orientation analysis for microblogs. To better capture the social attention on public events, we introduce Concern into the conventional psychological classes of sentiments and build up a sentiment lexicon with five categories(Concern, Joy, Blue, Anger, Fear). The seed words of the lexicon are extracted from HowNet, NTUSD, and catchwords of the Sina Weibo posts. The semantic similarity in HowNet is adopted to detect more sentiment words to enrich the lexicon. Accordingly, each Weibo post is represented as a multi-dimensional numerical vector in feature space. Then we adopt the Semi-Supervised Gaussian Mixture Model (Semi-GMM) and an adaptive K-nearst neighbour (KNN) with symmetric Kullback-Leibler divergence (KL-divergence) as similarity measurements to classify the posts. We compare our proposed methodologies with a few competitive baseline methods e.g., majority vote, KNN by using Cosine similarity, and SVM. The experimental evaluation shows that our proposed methods outperform other approaches by a large margin in terms of the accuracy and F1 score.