In this work we have introduced the SADSLyC-E-MSA corpus. It is a subset of the SADSLyC corpus that provides English and Modern Standard Arabic (MSA) translations. Based on this corpus, we evaluated three Pretrained Language Models (PLMs) to measure their ability to translate textual data from Saudi Arabian to English. The evaluation was carried out before and after fine-tuning on the SADSLyC-E-MSA corpus. For automatic evaluation, we used standard Machine translation (MT) metrics, including BLEU, TER, and CHRF. We also conducted human evaluation using the Multidimensional Quality Metrics (MQM) framework. The results showed clear improvements in translation quality once the models were adapted to the linguistic features of Saudi dialects. The Facebook M2ML model produced the best results and reached a BLEU score of 44.58.
Alahmari et al. (Thu,) studied this question.