Understanding Blog author's emotions with hierarchical Bayesian models

Abstract

The research of inferring emotions in natural language is confronted with two major challenges: the lack of basic knowledge in emotion expressions and the co-occurrence of separate emotions through all language unites. In this paper, we propose a Bayesian inference method to explore the basic knowledge with respect to emotion expression in different semantic dimensions, and to infer the co-occurrence of multiple emotion labels through words to the document. Specifically, we incorporate emotions and semantic dimensions as the latent factors in determining the distribution of observed words in a corpus of Blog articles. For each Blog article, we further generalize emotions from words to the document by incorporating a document specific hierarchy in the emotion distributions. The basic knowledge and co-occurred emotion labels in words and documents are obtained through a Gibbs sampling inference. Our experiment is performed on the well-developed Chinese emotion corpus, i.e. Ren-CECps, which indicates both higher accuracy and better robustness in our word and document emotion predictions compared with those generated by the state-of-the-art emotion prediction algorithms, and demonstrates a distribution of emotions in different word semantic dimensions.