A General Learning Method for Automatic Title Extraction from HTML Pages

This paper addresses the problem of automatically learning the title metadata from HTML documents. The objective is to help indexing Web resources that are poorly annotated. Other works proposed similar objectives, but they considered only titles in text format. In this paper we propose a general learning schema that allows learning textual titles based on style information and image format titles based on image properties. We construct features from automatically annotated pages harvested from the Web; this paper details the corpus creation method as well as the information extraction techniques. Based on these features, learning algorithms, such as Decision Trees and Random Forest algorithms are applied achieving good results despite the heterogeneity of our corpus, we also show that combining both methods can induce better performance.