Leveraging Textual Sentiment Analysis with Social Network Modelling: Sentiment Analysis of Political Blogs in the 2008 U.S. Presidential Election

Abstract

Automatic computational analysis and categorisation of political texts with respect to the rich array of personal sentiments, opinions, stances, and political orientations expressed in polarised political discourse is an exciting task which opens up many avenues for more accurate and naturalistic large-scale political analysis. The task does however pose major challenges for state-of-the-art Sentiment/Subjectivity/Affect Analysis and general Natural Language Processing tools. In this initial study, we investigate the feasibility of combining purely linguistic indicators of political sentiment with non-linguistic evidence gained from concomitant social network analysis. This study focuses on political blog analysis and draws on a corpus of 2.8 million blog posts by 16,741 bloggers crawled between April 2008 and May 2009. We focus on modelling blogosphere sentiment centered around Barack Obama during the 2008 U.S. presidential election, and describe a series of initial sentiment classification experiments on a data set of 700 crowd-sourced posts labelled as 'positive', 'negative', 'neutral', or 'not applicable' with respect to Obama. Our approach employs a hybrid machine learning and logic-based framework which operates along three distinct levels of analysis encompassing standard shallow document classification, deep linguistic multi-entity sentiment analysis and scoring, and social network modelling. The initial results highlight the inherent complexity of the classification task, and point towards the positive effects of learning features that exploit entity-level sentiment and social network structure.