Bridging the Gap: Transfer Learning for Dialect Identification in Low-Resource Settings A Case Study with Algerian Arabic

This paper explores bridging the gap between well- resourced languages and low-resource languages in Natural Language Processing (NLP). We investigate leveraging Modern Standard Arabic (MSA) as a source for transferring knowledge to a low-resource dialect, Algerian Arabic, for the task of dialect identification. We compare the performance of three pre-trained models: AraBERT, DziriBERT, and mBERT. Our evaluation uses the Madar Corpus, a collection of parallel sentences from various languages and dialects. For this study, we focus on the Algerian Arabic and their corresponding MSA counterparts, comprising a dataset of 4k sentences. Our findings demonstrate the effectiveness of transferring knowledge from MSA and other well-resourced languages to enhance NLP tasks for low-resource languages.