Learning slowly to learn better : curriculum learning for legal ontology population

Abstract

In this paper, we present an ontology population approach for legal ontologies. We exploit Wikipedia as a source of manually annotated examples of legal entities. We align YAGO, a Wikipedia-based ontology, and LKIF, an ontology specifically designed for the legal domain. Through this alignment, we can effectively populate the LKIF ontology, with the aim to obtain examples to train a Named Entity Recognizer and Classifier to be used for finding and classifying entities in legal texts. Since examples of annotated data in the legal domain are very few, we apply a machine learning strategy called curriculum learning aimed to overcome problems of overfitting by learning increasingly more complex concepts. We compare the performance of this method to identify Named Entities with respect to batch learning as well as two other baselines. Results are satisfying and foster further research in this direction.