📄 A Computational Morphology System for Arabic

Summary

The paper presents a morphological analyzer for Arabic that recovers the consonantal root and the pattern of verbs and of nouns and adjectives derived from verbs. It was built to support parsing and information retrieval and is linked to a large Arabic lexicon. The authors judged two-level morphology (PC-KIMMO) a poor fit for root-and-pattern word formation. They first extended the Al-Fedaghi and Al-Anzi pattern-matching method from triliteral to quadriliteral roots. They then replaced it with a faster scheme:

• strip the longest possible prefix;
• treat the first remaining letter as the first root letter;
• test only a handful of letter combinations from the next few characters against a stored root list.

The system was tested on 242 conference abstracts of non-vowelized text. It handled every verb-derived word correctly and was roughly eight times faster than the earlier algorithm. A menu interface exposes four functions: full paradigm display, generation of a requested form, grammatical analysis, and root extraction.

Key points

• The evaluation used 242 abstracts drawn from the Saudi Arabian National Computer Conference, totaling 19,167 running words:
  • 16,775 words had triliteral roots;
  • 1,124 had quadriliteral roots;
  • none had quintiliteral roots;
  • 1,268 were underived ("solid") nouns or foreign borrowings.
• All triliteral and quadriliteral cases were analyzed correctly. Colleagues in the Arabic Language Processing Laboratory verified the results.
• Average root search per abstract took 2.2 s, versus 17.2 s for the Al-Fedaghi and Al-Anzi algorithm. The new method also used less memory.
• The lexicon holds 1,116 regular verb roots plus explicitly stored forms for the 39 most frequent irregular verbs. Arabic script was handled with the Nafitha software.
• The four menu options are GetParadigm, GetForm (by mood, person, number, gender), Analyze, and GetRoot. Future work targets vowelized text and generation of derived nouns and adjectives.

Interesting ideas

• I1. Removing the longest prefix pins down the first root letter, so only six letter combinations remain to check against a closed root list. This turns open-ended search into a small, verifiable candidate set.
• I2. Patterns are matched by writing placeholder letters into the word at the slot positions and checking whether the result equals the template. This is a simple generate-and-compare test with no learned scoring.
• I3. One morphological core serves four different clients: human paradigm browsing, form generation, grammatical analysis for a parser, and root lookup for retrieval.
• I4. Rule-based handling covers the regular roots, while the most common irregular verbs are simply listed in full. This splits the problem into a productive part and a memorized exception part.
• I5. The evaluation counted words outside the system's scope (underived nouns, borrowings) as a separate category instead of mixing them into errors.

For Mova

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

• I1 (tokens): The English lemmatizer could strip the longest known affix and test a short, enumerated list of stem candidates against the lexicon. Each choice would then come with a traceable reason.
• I2 (tokens): Irregular or templated English forms (e.g. ablaut verb classes) could be checked by rendering the template and comparing it to the word. The match itself would serve as the explanation.
• I3 (back): The same inflection tables used by the tagger could drive surface realization in generation. This keeps analysis and generation consistent by construction.
• I4 (tokens): Keeping a small explicit table of high-frequency irregulars beside rule-based morphology avoids distorting the rules with exceptions.
• I5 (learn): Reporting out-of-coverage items as their own bucket in evaluations would make accuracy figures more honest. It would also point directly at gaps to fill.