Disambiguating word sentiment polarity through Bayesian modeling and opinion-level features

Many opinion words carry different polarity in different context, posing huge challenges to sentiment analysis research. Previous work on contextual polarity disambiguation makes use of term-level context such as word and patterns, and resolves the polarity with patternbased methods, PMI-based statistical techniques and machine learning methods. The major shortcoming of all such approaches lies in that termlevel features are sometimes ineffective in resolving the polarity. In this work, opinion-level features are studied and a Bayesian model is designed to disambiguate word sentiment polarity. Experiments with Opinmine corpus show that the opinion-level Bayesian model achieves significant performance gain in word polarity disambiguation in two domains.