Bridging the Kuwaiti Dialect Gap in Natural Language Processing

Abstract

The available dialectal Arabic linguistic resources are very limited in their coverage of Arabic dialects, particularly the Kuwaiti dialect. This shortage in linguistic resources creates struggles for researchers in the Natural Language Processing (NLP) field and limits the development of advanced linguistic analytical and processing tools for the Kuwaiti dialect. Many other low-resource Arabic dialects are still not explored in research due to the challenges faced during the annotators' recruitment process for dataset labeling. This paper proposes a weak supervised classification system to solve the problem of recruiting human annotators called "q8SentiLabeler". In addition, we developed a large dataset consisting of over 16.6k posts serving sentiment analysis for the Kuwaiti dialect, covering several themes and timeframes to remove any bias that might affect its content. Furthermore, we evaluate our dataset using multiple traditional machine-learning classifiers and advanced deep-learning language models to test its performance. Results demonstrate the positive potential of "q8SentiLabeler" to replace human annotators with a 93% for the pairwise percent agreement and a 0.87 for Cohen's kappa coefficient. Using the ARBERT model on our dataset, we were able to achieve 89% accuracy in the system's performance.