Research Article

A Sequence-to-Sequence Deep Learning Framework for Arabic Root Extraction and Stemming

Abstract

Arabic morphology is non-concatenative: words are built by interdigitating a consonantal root with a vocalic pattern and then attaching clitics, so the root letters are rarely contiguous and weak letters ( ا ,ي ,و ) and hamza change or disappear. Rule-based root extractors such as Khoja and ISRI therefore fail on irregular, weak, doubled and foreign words and on out-of-vocabulary forms. We formulate root extraction as character-level sequence generation, F(word) → root, and compare three encoder–decoder architectures (BiLSTM with Bahdanau attention, BiGRU with Luong attention, and a Transformer) with four baselines (ISRI, Tashaphyne, Khoja and a character CRF over edit tags) on identical test items. The main benchmark is built from the Lexicon of the Modern Arabic Language (Kaggle release, 61,190 word–root pairs, 5,706 roots); a second benchmark of 45,249 inflected, clitic-bearing tokens with sentence context is derived from the Quranic Arabic Corpus. On unseen word forms the best model (BiLSTM + Bahdanau) reaches 89.43% root accuracy and 96.84% character F1 on the modern-Arabic benchmark, against 70.49% for Khoja and 77.23% for the CRF, and 95.33% on the Quranic benchmark (Khoja 65.07%); every difference from a baseline is significant (exact McNemar test), whereas the three neural architectures do not differ significantly. Gains are largest on defective (naqis), doubly weak (lafif) and doubled roots. On roots never seen in training the models still lead on modern Arabic but fall behind affix rules on clitic-bearing Quranic tokens, and a model trained on dictionary forms alone transfers poorly to inflected text unless clitic-bearing examples are added. We also report an ablation study and a stemming (lemma) task, and release the benchmark splits and code.

Authors affiliation:

Zaid Rajih Mohammed (ORCID): Faculty of Medical Sciences, Jabir ibn Hayyan University for Medical and Pharmaceutical Sciences, Najaf, Iraq.

How To Cite: Z.R. Mohammed. A Sequence-to-Sequence Deep Learning Framework for Arabic Root Extraction and Stemming. Journal of Computers and Applications, 2(2):35–47, 2026. https://doi.org/10.58613/jca224