Data Augmentation Techniques on Arabic Data for Named Entity Recognition
Caroline Sabty, Islam Omar, Fady Wasfalla, Mohamed Islam, Slim Abdennadher
Abstract
Much research has been conducted to apply data augmentation techniques (DA), i.e., transformations to a given training set to produce more data synthetically. While DA is often used in computer vision and speech recognition, it is not very common in Natural Language Processing, especially in the Named Entity Recognition (NER) task, i.e., identifying named entities. To the best of our knowledge, it is also not applied in NER on the challenging Code-Switching (CS) data, which is text containing more than one language in the same sentence. This paper presents several practical and easy-to-implement data augmentation techniques to improve the Arabic NER and especially on CS data based on word embedding substitution, a modified version of the Easy Data Augmentation technique, and back-translation. We demonstrate that the proposed methods boost the performance of the NER task on CS data through several experiments on an available AR-EN CS data-set with an increase of F-score equal to 1.5%. The proposed DA methods also eliminate the time and effort of collecting and labeling new data for low resources NER tasks.