Automatic Moderation of Comments in a Large On-line Journalistic Environment

On-line journalistic sites publish several news and stories every day. Readers of these sites may comment a story, and, as a consequence, a single story might receive thousands of comments. The quality of these comments may vary a lot, from spams and trolls to truly useful information. Separating good from bad comments is an important task, and is the primary goal of comment moderation. Moderators usually classify and score the comments, promoting high quality ones and, likewise, discouraging low quality ones. However, moderators usually face a very large number of comments, and thus, moderation may require a huge amount of time. In this paper we address the problem of automatic moderation of comments in a large journalistic Web site. Participants of the site may engage in discussions and interact with each other (i.e., friends, fans, enemies etc.), constituting a large social network. We propose a data mining technique which combines underlying patterns that are implicit in the content of the comments with patterns hidden in the social network, and then uses the result for classification and scoring purposes. We evaluate our proposed technique using a real collection of comments collected from the Slashdot forum. The proposed technique is effective, outperforming traditional approaches, such as decision trees and SVMs, in terms of accuracy. Further, the technique shows to be extremely fast, being able to moderate hundreds of comments per minute.