📄 Acquisition System for Arabic Noun Morphology

Summary

The paper tackles Arabic noun morphology, which earlier work largely skipped in favour of verbs. Nouns not derived from verbal roots, including many borrowed proper names, lack clean generative rules. Their singular, plural and gender forms instead fall into families that have to be stored lexically, much like English irregular verbs. The authors build an acquisition tool for this. A suffix analyzer removes attached pronouns and number and gender endings. A finite-state pattern generator then maps the stem to one of 62 patterns. The noun is assigned to one of 84 paradigm groups, by stored lookup, by suffix/pattern evidence, or by asking a user targeted questions. Once the group is known, the full number and gender paradigm follows, and the root is saved so later forms are recognized. On 500 nouns from the Al-Raya newspaper corpus, the system handled 90.2% correctly. The design matters because new names keep arriving in news text, so the system must keep learning from people.

Key points

• Nouns are sorted into 84 groups by their singular and plural, masculine and feminine patterns, and the system recognizes 62 distinct stem patterns.
• The test data was 500 nouns from a corpus built by Ahmad Hasnah from Al-Raya (Qatar) newspaper text. The suffix analyzer was 97% correct, the pattern analyzer about 93%, and the full noun analyzer 90.2%.
• Of the errors, 10 came from suffix analysis and 29 from missing patterns. The missing patterns were added afterwards, while the suffix errors come from genuine ambiguity.
• Correctly handled nouns were classified 28.8% via the Classified Noun Table (Database Checker), 7.1% by suffix/pattern analysis alone, and 64.1% through the User-Feedback Module.
• Every newly classified noun's singular root goes into the Classified Noun Table, so its other inflected forms are later resolved without questions.

Interesting ideas

• I1. Each candidate group is scored per form slot as present (1), absent (−1) or unknown (0), and the system asks the yes/no question that splits the remaining groups best, falling back to a multiple-choice question when needed. This keeps human effort to a few questions.
• I2. Irregular inflection is treated as membership in a stored paradigm family rather than computed by rules, so once a word's family is known, all its forms come for free.
• I3. Analysis is staged as strip affixes, find pattern, then disambiguate. Cheap lookup comes first, structural evidence second, and an external oracle last, with each stage's accuracy measured separately.
• I4. Storing only the canonical root after resolving one form lets every sibling form be resolved by lookup later. Coverage then grows with use and the number of oracle questions falls over time.

For Mova

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

• I1 (learn): The same slot-matrix scheme could choose the single most informative question for an LLM teacher or human when several lemma or paradigm candidates survive, which cuts teacher calls.
• I2 (tokens): English irregular verbs and plurals could be represented as family IDs in the en lexicon, so the lemmatizer can explain each form by naming its family.
• I3 (tokens): Per-stage accuracy in the morphology pipeline would show whether errors come from affix stripping or from missing paradigm entries, as the paper's error split did.
• I4 (learn): Silver-data runs could cache each resolved lemma's paradigm once, so later inflections stop needing witness-parser agreement or teacher queries.