Coarse-and Fine-Grained Sentiment Analysis of Social Media Text

INTRODUCTION People make judgments about the world around them. They harbor positive and negative attitudes about people, organizations, places, events, and ideas. We regard these types of attitudes as sentiments. Sentiments are private states,1 cognitive phenomena that are not directly observable by others. However, expressions of sentiment can be manifested in actions, including written and spoken language. Sentiment analysis is the study of automated techniques for extracting sentiment from written language. This has been a very active area entiment analysis—the automated extraction of expressions of positive or negative attitudes from text—has received considerable attention from researchers during the past 10 years. During the same period, the widespread growth of social media has resulted in an explosion of publicly available, user-generated text on the World Wide Web. These data can potentially be utilized to provide real-time insights into the aggregated sentiments of people. The tools provided by statistical natural language processing and machine learning, along with exciting new scalable approaches to working with large volumes of text, make it possible to begin extracting sentiments from the web. We discuss some of the challenges of sentiment extraction and some of the approaches employed to address these challenges. In particular, we describe work we have done to annotate sentiment in blogs at the levels of sentences and subsentences (clauses); to classify subjectivity at the level of sentences; and to identify the targets, or topics, of sentiment at the level of clauses.