Topic models with sentiment priors based on distributed representations

In recent works, topic models for aspect-based opinion mining have been extended to automatically train sentiment priors for topic-word distributions, leading to automated discovery of sentiment words and improved sentiment classification. In this work, we propose an approach where sentiment priors are trained in the space of word embeddings; this allows us to both discover more aspect-related sentiment words and further improve classification. We also present an experimental study that validates our results.