Unsupervised Improving of Sentiment Analysis Using Global Target Context

摘要: Current approaches to document-level sentiment analysis rely on local information, e.g., the words within given document. We try achieve better performance by incorporating global context of target (e.g., a movie or product). assume that labels reviews about same are often consistent in some way. model this consistency Dirichlet distribution over and use it together with Maximum entropy classifier gain significant improvement. This unsupervised extension increases classification F-measure almost 3% absolute both Czech English review datasets outperforms current state art.