Large-scale hierarchical text classification without labelled data英文文献资料.pdf

Large-Scale Hierarchical Text Classification without Labelled Data Viet Ha-Thuc Jean-Michel Renders Computer Science Department Xerox Research Centre Europe The University of Iowa Meylan, France Iowa City, IA, USA jean-michel.renders@ hviet@ ABSTRACT 1. INTRODUCTION The traditional machine learning approaches for text clas- With the exponential growth of text data, particularly on sification often require labelled data for learning classifiers. the Web, hierarchical organization of text documents is be- However, when applied to large-scale classification involv- coming increasingly important to manage the data. Along ing thousands of categories, creating such labelled data is with the widespread use of the hierarchical data manage- extremely expensive since typically the data is manually la- ment, comes the need for automatic classification of docu- belled by humans. Motivated by this, we propose a novel ments to the categories in the hierarchy. Traditional super- approach for large-scale hierarchical text classification which vised and semi-supervised approaches for hierarchical text does not require any labelled data. We explore a perspective classification often require labelled data for learning clas- where the meaning of a category is not defined by human- sifiers. However, when applied to large-scale classification labelled documents, but