Data Augmentation for Sorani Kurdish News Headline Classification Using Back-Translation and Deep Learning Model

Abstract

With the increased volume of news articles and headlines being generated, it is becoming more difficult for individuals to keep up with the latest developments and find relevant news articles in the Kurdish language. To address this issue, this paper proposes a novel data augmentation approach for improving the performance of Sorani Kurdish news headline classification using back-translation and a proposed deep learning Bidirectional Long Short-Term Memory (BiLSTM) model. The approach involves generating synthetic training data by translating Sorani Kurdish headlines into a target language in this context, English, and back-translating them into the Kurdish language, resulting in an augmented dataset. The proposed BiLSTM model is trained on the augmented data and compared with baseline models support vector machines (SVM) and Naïve Bayes and trained on the original data. The experimental results demonstrate that the proposed BiLSTM model outperforms the baseline and other existing models, such as SVM and Naïve Bayes, achieving state-of-the-art performance on the So-rani Kurdish news headline classification task by scoring %94 for the Kurdish documents classification-4007 dataset and %89 for the Kurdish news dataset headlines dataset, using five folds method which +2 higher than non-augmented texts. The findings suggest that the combination of back-translation and a proposed BiLSTM model is a promising approach for data augmentation in less-resourced languages, contributing to the advancement of natural language processing in un-der-resourced languages. Moreover, having a Sorani Kurdish news headline classification model can improve Sorani Kurdish speakers' access to news and information. With the classification model, they can easily and quickly search for news articles that interest them based on their preferred categories, such as politics, sports, or entertainment.