Sentiment analysis within and across social media streams

Abstract

Social media offers a powerful outlet for people's thoughts and feelings—it is an enormous ever-growing source of texts ranging from everyday observations to involved discussions. This thesis contributes to the field of analysis, which aims to extract emotions and opinions from text. A basic goal is to classify text as expressing either positive or negative emotion. Sentiment classifiers have been built for social media text such as product reviews, blog posts, and even Twitter messages. With increasing complexity of text sources and topics, it is time to re-examine the standard extraction approaches, and possibly to re-define and enrich definition. Thus, this thesis begins by introducing a rich multi-dimensional model based on Affect Control Theory and showing its usefulness in classification. Next, unlike analysis research to date, we examine expression and polarity classification within and across various social media streams by building topical datasets. When comparing Twitter, reviews, and blogs on consumer product topics, we show that it is possible, and sometimes even beneficial, to train classifiers on text sources which are different from the target text. This is not the case, however, when we compare political discussion in YouTube comments to Twitter posts, demonstrating the difficulty of political classification. We further show that neither discussion volume or expressed in these streams correspond well to national polls, putting in question recent research linking the two. The complexity of political discussion also calls for a more specific re-definition of sentiment as agreement with the author's political stance. We conclude that must be defined, and tools for its analysis designed, within a larger framework of human interaction.