Multi-class Sentiment Classification on Weibo

Multi-class sentiment classification is a key to analyse people's emotions and opinions toward products,, services, and social events. In this paper, four different feature engineering techniques are adopted to build sentiment classifiers. Firstly, two different data sets were crawled from Weibo. One of them is used as labelled training data set, the other one is used to train a model to get the distributed representations of words. Then, the two data sets were pre-processed to remove the noisy information. The tweets in labelled training data set were converted into fixed-sized input vectors based on the four different feature engineering techniques. Finally, the sentiment classifiers were built based on the random forest, sparse autoencoder and deep belief network. Experiment results show that the combination of random forest and frequency-based method obtains the highest accuracy in the multi-class sentiment classification task.