📄 Murabaa: A comprehensive Resource Platform for Arabic Morphology

Summary

Arabic NLP lacks reliable, openly available morphological resources. Existing lists are scattered, often undiacritized, tied to particular corpora, or hard-coded inside tools such as stemmers and analyzers. The authors address this with Murabaa, a platform that rebuilds Arabic word structure from its smallest units, following the classical noun/verb/particle framework. Letters are combined into roots under phonological constraints. Patterns are then applied to roots to produce lemmas and stems. Words that do not follow templates, such as function words and broken plurals, are handled in separate lists. The result is nine linked lexicons, encoded in the Lexical Markup Framework (LMF) and released under CC-BY-NC-ND. They are published as dictionaries on a website and in a web app. When compared with similar resources, Murabaa's lists are generally larger and richer in features. The stop-word list in particular far exceeds earlier ones. The resources already power the Safar toolkit. The authors argue that such curated, rule-based data can counter noisy training corpora and support morphology-aware Arabic LLMs.

Key points

• The nine lexicons cover the letters of the alphabet, the rules of phonology, roots, patterns, CALEM (the stem and lemma lexicon), function words, broken plurals, clitics, and the rules that govern clitics. They are distributed via GitHub (alelm-lab/Murabaa) in LMF format.
• The alphabet lexicon has 42 entries, each with Unicode, positional forms, and transliteration. The root lexicon holds 8,426 triliteral roots, and each root records which classical dictionaries attest it.
• The stop-word lexicon contains 67,153 entries. It is the only compared list that is diacritized, rule-based, classified, and includes cliticized forms.
• Compared with DIINAR, CALEM has more lemmas (166,963 vs 121,522) and more non-verbal stems (4,847,427 vs 4,714,222). DIINAR has more verbal stems (3,060,716 vs 2,464,239), a gap the authors could not explain because DIINAR is not available to them.
• The resources are already used in Safar, an Arabic NLP infrastructure of more than fifty tools, and in learning apps and games.

Interesting ideas

• I1. Each layer of morphology (letters, sound constraints, roots, patterns, word forms) is kept as its own queryable lexicon rather than hard-coded into tools, so every generated form can be traced back to the rules that produced it.
• I2. A lexicon of phonological constraints acts as a gatekeeper that blocks impossible roots or stems before they enter synthetic data, which reduces noise for downstream learners.
• I3. Irregular, non-templatic material (broken plurals, function words) is isolated in dedicated lists instead of being forced into general rules, which keeps the regular machinery clean.
• I4. Every root is tagged with the reference dictionaries that attest it, giving each entry built-in provenance and a crude confidence signal from agreement across sources.
• I5. The paper suggests using clitic-combination rules as hard constraints during decoding, forbidding illegal morpheme sequences rather than hoping a model learns to avoid them.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): The en crate could move its affix tables, spelling-change rules, and irregular forms into standalone versioned lexicons, so any lemma or tag decision can cite its source entry.
• I2 (learn): Silver data and self-play outputs could pass a morphology/orthography validity filter that rejects non-existent word forms before they are used for training.
• I3 (tokens): Irregular English plurals, past forms, and closed-class words could live in an explicit exception lexicon consulted before rule-based lemmatization, which makes errors easier to localize.
• I4 (world): Compiled knowledge from WordNet, FrameNet, and Wikidata could carry per-fact source tags, so contradictions or weakly supported facts can be flagged and explained.
• I5 (back): The generator could check its surface output against a morphological rules lexicon (agreement, inflection, contractions) to block malformed words when realizing text from the meaning graph.