Unsupervised sentiment analysis with a simple and fast Bayesian model using Part-of-Speech feature selection

Christian Scheible, Hinrich Schütze; Proceedings of KONVENS 2012 (PATHOS 2012 workshop), pp. 269-273, September 2012.


Unsupervised Bayesian sentiment analysis often uses models that are not well motivated. Mostly, extensions of Latent Dirichlet Analysis (LDA) are applied - effectively modeling latent class distributions over words instead of documents. We introduce a Bayesian version of Naive Bayes for sentiment analysis and show that it offers superior accuracy and inference speed.

[pdf] [bibtex]