Title Categorization Based on Category Granularity

Abstract

We focus on a problem of short text categorization, i.e. categorization of newspaper titles, and present a method that maximizes the impact of informative words due to the sparseness of titles. We used the hierarchical structure of categories and a transfer learning technique based on pre-training and fine-tuning to incorporate the granularity of categories into categorization. According to the hierarchical structure of categories, we transferred trained parameters of Convolutional Neural Networks (CNNs) on upper layers to the related lower ones, and finely tuned parameters of CNNs. The method was tested on titles collected from the Reuters corpus, and the results showed the effectiveness of the method.