Training-Less Multi-label Text Classification Using Knowledge Bases and Word Embeddings

Abstract

Traditional multi-label text classifiers suffer from the high dimensionality of feature space, label imbalance, and training overhead. In this work, we depart from traditional approaches with intensive feature engineering and linguistic analysis by introducing a novel ontology-based training-less multi-label text classifier. We transform the classification task into a graph matching problem to have a training-less classifier. The experiment results, using the EUR-Lex dataset, proved that our method offers competitive performance with respect to the above-mentioned approaches in terms of $F1_{macro}$ giving fair performance over the different labels despite of the training-less configurations.