Answers are grounded in this paper and cite their sources.
Neural Machine Translation (NMT) for dialectal Arabic remains challenging because parallel resources are limited and colloquial expressions often encode culture-specific meanings. This study investigates Syrian Arabic–English translation using a sequence-to-sequence model with a bidirectional Gated Recurrent Unit (BiGRU) encoder and Bahdanau additive attention, enriched with a lightweight ontology of Syrian cultural relations. The enlarged resource contains 123,183 aligned sentence pairs, divided into 98,527 training pairs, 12,319 validation pairs, and 12,337 test pairs. An audit of the ontology identified 156 means relations, divided into 124 training relations, 16 validation relations, and 16 held-out test relations. The study distinguishes two evaluation conditions. In the legacy in-domain condition, in which the evaluated ontology items were available during training, the verified system achieved 65.2038 corpus SacreBLEU, 88.6368 mean sentence-level BLEU, and 73.81% exact match, corresponding to 62 exact translations out of 84 items. The enlarged-corpus condition instead measures generalisation to unseen ontology relations. In this condition, the clean BiGRU checkpoint achieved 3.3533 corpus SacreBLEU on the held-out ontology test set, while a contemporary AraBERT v2 encoder–decoder comparison produced held-out ontology SacreBLEU values ranging from 2.04 to 3.28. These results show that ontology enrichment can support the reproduction of known cultural mappings, but reliable transfer to unseen cultural expressions remains difficult. The findings also demonstrate the importance of separating in-domain evaluation from held-out generalisation when assessing ontology-enhanced NMT for low-resource dialectal translation.
No takes yet. Share an insight, caveat, or question.
Al-Rakkad et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: