Holistic approaches to identifying the sentiment of blogs using opinion words

Sentiment analysis aims to identify the orientation (positive or negative) of opinions or emotions expressed in documents. Opinion lexicons comprise opinion words expressing prior positive or negative sentiments. In most previous work documents are represented as bags of words and sentiment analysis has been cast a classification problem, where opinion lexicons are only used to enhance the classification models. In this paper we aim to establish the direct connection between document sentiment and opinion words in the documents. We propose two holistic approaches that consider the probability distribution of both opinion words and their polarity for analyzing document sentiment. Our extensive experiments on blogs of 12 topics show that our holistic models significantly improve baseline models using words and their polarity information separately, and is also superior to an existing approach combining both types of information.