Document-Word Co-regularization for Semi-supervised Sentiment Analysis

摘要: The goal of sentiment prediction is to automatically identify whether a given piece text expresses positive or negative opinion towards topic interest. One can pose as standard categorization problem, but gathering labeled data turns out be bottleneck. Fortunately, background knowledge often available in the form prior information about polarity words lexicon. Moreover, many applications abundant unlabeled also available. In this paper, we propose novel semi-supervised algorithm that utilizes lexical conjunction with examples. Our method based on joint analysis documents and bipartite graph representation data. We present an empirical study diverse collection problems which confirms our models significantly outperform purely supervised competing techniques.