Evaluating the Influence of Parallel Corpora on Arabic Dialect Identification: A Comparative Study

In this paper, we undertake an impact assessment to gauge the effectiveness of utilizing parallel corpora in enhancing Arabic Dialect Identification (ADI) systems. The study compares the performance of statistical and neural methods for classifying Arabic dialects using parallel corpora against non-parallel corpora. We augment the well-known PADIC (Parallel Arabic DIalectal Corpus) with data from a low-resourced vernacular Algerian dialect, specifically the Kabyl dialect, to create an extended version. Additionally, we manually collect a non-parallel corpus containing an equal number of sentences per dialect as PADIC (6412 sentences for each dialect). We employ a variety of classifiers including Gaussian Naive Bayes, Bernoulli Naive Bayes, kNN, Logistic Regression, SGD Classifier, Passive Aggressive Classifier, Perceptron, Linear Support Vector, and Convolutional Neural Networks in our experiments. The results indicate an overall accuracy of approximately 92%, underscoring the potential impact of parallel corpora in fortifying Arabic dialect identification systems.