Exploring Multi-label Stacking in Natural Language Processing

The task of classification with multi-label data is an important research field in Natural Language Processing (NLP). While there have been studies using one-stage multi-label approaches for automatic text classification, there are not many that use two-stages stacking models. In this paper we explored Binary Relevance (BR) classifiers, with J48 and probabilistic Support Vector Machine (SVM), in a two-stage stacking model. We have evaluated our proposal in three textual data sets: The Movie Database (TMDB), Enron email, and EURLEX European legal text. The results showed that by using a two-stage stacking model, we can obtain better results than by using one-stage classifiers.