Ahmed Al-Hmouz

Arabic morphological tools usually return candidate analyses or a lemma; few commit to one fully specified root and an explicit ف/ع/ل pattern (wazan) per token, which is the output paradigm generation and morphology teaching need. Tasreef makes that commitment for undiacritized Arabic. A rule cascade indexed by word length and affix class is combined with dictionary-cited roots admitted under explicit rules and with analysis by synthesis: a paradigm generator, letter-exact on 97.3% of 18,507 Quranic verb forms, regenerates candidate words from attested roots and accepts a root only if it reproduces the surface. A neural fallback now answers 0.15% of tokens. Because the treebank used during development is also the standard test set, three kinds of evidence are reported separately. On UD Arabic-PADT test, disclosed as a development set, root accuracy is 96.40% against 97.67% for CAMeL Tools under identical scoring, and 89.88% against 66.54% on quadriliteral roots. On sealed halves of three classical gold sets, scored once, root accuracy is 91.3% on broken plurals, 41.0% on weak-root derivation chains and 53.0% on exercise words. A 500-word sample of a second treebank, frozen before annotation, is being labelled by two annotators. Analysis by synthesis adds 154 treebank tokens and raises accuracy on generated verb forms from 49.6% to 88.0%; the neural model, used as a wazan override, lowers wazan accuracy by 16.5 points.