From passages to questions: an end-to-end transformer-based framework for answer-aware Arabic question generation

Lafkiar, Said, Alami, Hamza, En-nahnahi, Noureddine
Information Technology and Management · Springer Nature · 2026-11-19 · DOI: 10.1007/s10799-026-00507-3

Abstract

Arabic Question Generation (AQG) remains challenging due to the linguistic complexity of Arabic and the scarcity of large annotated resources. This paper presents an end-to-end answer-aware framework that transforms Arabic passages into coherent and contextually aligned questions. The model is an encoder–decoder architecture in which both the encoder and decoder are initialized from checkpoints pretrained on large-scale Arabic and multilingual text corpora. The encoder captures contextual semantic and syntactic representations of the passage and the answer span, while the decoder generates questions autoregressively, which are subsequently refined in a postprocessing module. To support systematic evaluation, a large collection of ~71k Arabic passage–question–answer triplets was compiled from multiple sources. Extensive experiments are conducted under a unified evaluation protocol across multiple pretrained encoder–decoder configurations. Our proposal achieved 20.75, 35.46, and 47.16 in BLEU, METEOR, and ROUGE-L outperforming baseline models. Finally, the obtained human evaluations score 4.42/5 assesses grammaticality, fluency, relevance, and answer correctness.

Read at publisher →

Similar articles