Document-Level Sentiment Analysis through Incorporating Prior Domain Knowledge into Logistic Regression.

In this paper, we present a prior domain knowledge-enhanced classification approach for document-level sentiment analysis. The proposed approach is a hybrid category of sentiment classification approaches which uses two types of prior domain knowledge - general sentiment knowledge extracted from a lexicon, and knowledge extracted from unlabeled domain data that we call domain-specific sentiment knowledge. We combine prior domain knowledge with logistic regression to enhance sentiment classification, and use gradient descent approach to optimize the modified logistic regression model. The novelty of our proposed approach lies in incorporating prior domain knowledge directly into the logistic regression model. The proposed approach is empirically evaluated through extensive experiments over a multi-domain sentiment dataset. It is also compared with three baseline methods and performs significantly better.