📄 The Arabic Noun System Generation

Summary

Arabic broken plurals change the stem internally (rajul → rijaAl), which makes them hard for morphological generators. A single singular pattern can correspond to two to five plural patterns, and nothing reliably predicts which one a given noun takes. The authors review the traditional grammarians' pattern lists, which rely on phonological and semantic conditions. They then use frequency tables from Levy (1971, built on Wehr's dictionary) and Murtonen (1964, built on Lane's dictionary) to show that the shape of the singular narrows the set of candidate plurals but never settles the choice. Deriving plurals from the singular or the root would therefore require brittle truncation and deletion rules. Their alternative, framed in Lexeme-based Morphology, stores a small inventory of number-labelled stems in each lexeme (singular/dual vs. plural), so inflection only has to attach affixes. Sound plurals are handled by marking each entry with its plural class (uwna or aAt). The analysis is implemented in MORPHE, where equivalence declarations let syncretic paradigm cells share a single rule. The result is a simpler generator for Arabic nouns that states more general rules.

Key points

• Arabic nouns distinguish singular, dual and plural. Plurals split into sound and broken, and broken plurals further into paucity (4 forms, for three to ten items) and multiplicity (at least 23 forms, for more than ten).
• The statistics are read as paired percentages. For example, CaCC × CuCuwC = 73/49 means 73% of singulars behind the plural CuCuwC have the shape CaCC, and 49% of CaCC singulars take CuCuwC.
• A lexeme such as RAJUL has a syntactic level ("noun") and a phonological stem inventory: /rajul/ for singular and dual, /rijaAl/ for plural.
• In MORPHE, a bpstem feature supplies the broken plural stem, and an sp feature (values uwna/aAt) selects the sound-plural suffix class. If sp is absent, the system falls back to a default rule that is shared with the singular.
• Accusative and genitive look the same in sound plurals but differ in broken plurals. Because the enhanced MORPHE can only equivalence leaf and pre-leaf nodes, the broken-plural default rules had to sit at the bottom of the tree.

Interesting ideas

• I1. When corpus statistics show a form cannot be predicted, list it per lexeme instead of writing ever more complex derivation rules; the frequency tables serve as the evidence for drawing that line. — For Mova (tokens): Irregular English plurals and past forms (mice, went) could be stored as labelled stems in the lexicon, with the frequency data documenting why each one is listed rather than derived.
• I2. Give each lexeme a set of stems tagged with the features they serve, and let inflection rules only add affixes to the right stem. — For Mova (back): The generator could pick a stem by its feature tag and then apply plain suffix rules, so every surface form can be traced to "stem X + rule Y".
• I3. If a lexical feature is missing, fall back to a default rule borrowed from another cell of the paradigm, which keeps the lexicon sparse. — For Mova (tokens): The morphology module could mark only the exceptional lemmas and let everything else fall through to regular rules, logging which path produced each form.
• I4. Declare syncretism explicitly by making identical paradigm cells point to one shared rule instead of copying it. — For Mova (back): Shared cells (for example, identical past and participle forms) could be declared once, which avoids drift between duplicated rules and makes ambiguity visible to the analyzer.
• I5. Read a singular↔plural co-occurrence table in both directions (how often each source pattern leads to a target, and how often the target comes from each source) to measure how predictable a mapping is. — For Mova (learn): The same two-way percentages could be used to check whether a self-induced repair rule really generalizes or just memorizes a few lemmas.