With the rapid development of social media and social networks, spontaneously user generated content like tweets and forum posts have become important materials for tracking people's opinions and sentiments online. in this paper we investigate the limitations of traditional linguistic-based approaches to sentiment analysis when applied to these informal genres. inspired by various social cognitive theories, we combine local linguistic features and global social evidence in a propagation scheme to improve sentiment analysis results. without using any additional labeled data, this new approach obtains significant improvement (up to 12% higher accuracy) for various genres in the domain of presidential election.