Analyzing Social Media Content for Security Informatics

Abstract

Inferring public opinion regarding an issue or event by analyzing social media content is of great interest to security analysts but is also technically challenging to accomplish. This paper presents a new method for estimating sentiment and/or emotion expressed in social media which addresses the challenges associated with Web-based analysis. We formulate the problem as one of text classification, model the data as a bipartite graph of documents and words, and construct the sentiment/emotion classifier through a combination of semi-supervised learning and graph transduction. Interestingly, the proposed approach requires no labeled training documents and is able to provides accurate text classification using only a small lexicon of words of known sentiment/ emotion. The classification algorithm is shown to outperform state of the art methods on a benchmark task involving sentiment analysis of online consumer product reviews. We illustrate the utility of the approach for security informatics through two case studies, one examining the possibility that online sentiment about suicide bombing predicts bombing event frequency, and one investigating public sentiment about vaccination and its implications for population health and security.