Text Representation in Multi-label Classification: Two New Input Representations

Automatic text classification is the task of assigning unseen documents to a predefined set of classes. Text representation for classification purposes has been traditionally approached using a vector space model due to its simplicity and good performance. On the other hand, multi-label automatic text classification has been typically addressed either by transforming the problem under study to apply binary techniques or by adapting binary algorithms to work with multiple labels. In this paper we present two new representations for text documents based on label-dependent term-weighting for multi-label classification. We focus on modifying the input. Performance was tested with a well-known dataset and compared to alternative techniques. Experimental results based on Hamming loss analysis show an improvement against alternative approaches.