Evaluating Humour Features on Web Comments

Abstract

Research on automatic humor recognition has developed several features which discriminate funny text from ordinary text. The features have been demonstrated to work well when classifying the funniness of single sentences up to entire blogs. In this paper we focus on evaluating a set of the best humor features reported in the literature over a corpus retrieved from the Slashdot Web site. The corpus is categorized in a community-driven process according to the following tags: funny, informative, insightful, offtopic, flamebait, interesting and troll. These kinds of comments can be found on almost every large Web site; therefore, they impose a new challenge to humor retrieval since they come along with unique characteristics compared to other text types. If funny comments were retrieved accurately, they would be of a great entertainment value for the visitors of a given Web page. Our objective, thus, is to distinguish between an implicit funny comment from a not funny one. Our experiments are preliminary but nonetheless large-scale: 600,000 Web comments. We evaluate the classification accuracy of naive Bayes classifiers, decision trees, and support vector machines. The results suggested interesting findings.

Resources

Details