Sampling latent emotions and topics in a hierarchical Bayesian network

Text emotion analysis suffers from the lack of faithful emotion features, and the difficulty of mining multiple emotions that are mixed together. In this paper, we provide a Gibbs sampling method to solve these two problems. We explicitly characterize the emotion combination phenomenons, and predict the complex emotions of words together with the emotion intensities for each singular emotion through raw texts. Both emotions and emotion intensities are embedded as latent random variables in a hierarchical Bayesian network, while only the words and some preliminary expectations are represented as observed variables. The model which we call word emotion topic (WET) model, also depicts the distribution of word emotions among different topics, which helps to study the variation of word topics and word emotions. Experiment shows promising results of word emotion prediction, which outperforms traditional parsing methods such as Hidden Markov Model and Conditional Random Fields on raw text. The result also presents interesting emotion-topic variations through blog articles.