Collaborative community-specific microblog sentiment analysis via multi-task learning

Microblog sentiment analysis has become a hot research area due to its wide applications. There are some methods utilizing social context, but they only built a global sentiment analysis model, failing to extract personalized expressions. Some personalized methods have been proposed to deal with this problem, but they suffer from data sparseness and inefficiency. Based on personalized sentiment analysis methods, we exploit social context information and capture users' variable and distinctive expressions at a community level to handle these problems. In particular, we propose a collaborative microblog sentiment analysis approach. In our approach, two classifiers are constructed. One is the global microblog sentiment analysis model which can exploit the sentiment shared by all users. One is the community-specific microblog sentiment analysis model which can extract sentiment influenced by user personalities. In addition, we extract community similarity knowledge and employ it to improve the learning process of the community-specific sentiment model. Moreover, we incorporate social contexts into this model as regularization to encourage the sharing sentiment between connected microblogs. An accelerated algorithm is introduced to solve our model. Experiments on two real datasets show that our model can advance the performance of microblog sentiment classification effectively and outperform state-of-art methods significantly.