Semi-automated feature selection for web text filtering

The explosive growth of the Internet inevitably leads to the proliferation of harmful information such as pornography, drug and violence. A great deal of filtering techniques based on image and text categorization is proposed in the literature. Among them, text-based filtering plays a leading role for its good performance. Existing text filtering algorithms can be seen as a classical text categorization approach of discerning two topics, i.e. harmful and benign. In this paper, motivated by the linguistic character of text features and other related text classification tasks such as genre detection, a new feature selection framework for text filtering is proposed. It combines linguistics and domain knowledge in an effective way. Experimental results have demonstrated that our method is more adapt to special domain text filtering tasks.