Multilabel Text Classification with Label-Dependent Representation

Assigning predefined classes to natural language texts, based on their content, is a necessary component in many tasks organizations. This task carried out by classifying documents within set of categories using models and computational methods. Text representation for classification purposes has traditionally been performed vector space model due its good performance simplicity. Moreover, the texts via multilabeling typically approached simple label methods, which require transformation problem studied apply binary techniques, or adapting algorithms. Over previous decade, text extended deep learning models. Compared traditional machine avoids rule design feature selection humans, automatically provides semantically meaningful representations analysis. However, learning-based data-intensive computationally complex. Interest does not techniques shallow learning. situation true when training cases smaller, features small. White box approaches have advantages over black approaches, where feasibility working with relatively small sets data interpretability results stand out. research evaluates weighting function words modify during multilabel classification, combination two approaches: adaptation. was tested 10 referential textual sets, compared alternative three measures: Hamming Loss, Accuracy, macro-F1. The best improvement occurs macro-F1 fewer labels, documents, smaller vocabulary sizes. In addition, improves higher cardinality, density, diversity labels. proves usefulness sets. show improvements more than 10% terms classifiers our method almost all analyzed.