Mining Sentiment Classification from Political Web Logs

Over the last few years the number of web logs and amount of opinionated data on the World Wide Web has grown dramatically. Web logs allow people to share their opinions on a wide range of “hot” topics with virtual communities. As readers start turning to web logs as a source of information, automatic techniques that identify the sentiment of web log posts will help bloggers categorize and filter this exploding information source. However, on the surface, sentiment classification of web log posts, in particular political web log posts, appears to be a more difficult problem than classification of traditional text because of the interplay among the images, hyperlinks, the style of writing and language used within web logs. In this paper we investigate existing technology and their utility for sentiment classification on web log posts. We show that a Naïve Bayes classifier can on average correctly predict a posting’s political category 78.06% of the time with a standard deviation of 2.39. It significantly outperforms Support Vector Machines at the 99.9% confidence level with a confidence interval of [1.425, 3.488]. On average, SVMs correctly predicted the category of web log posts 75.47% of the time with a standard deviation of 2.64. Previous research was able to achieve an 81.0% accuracy using Naïve Bayes and 82.9% using SVMs using our chosen feature set representation on a nonspecific topic corpus [14]. Using our dataset of political web logs over a two-year period, we also show that it is important to maintain a uniform distribution in such datasets to avoid biases in classification.