📄 A Formal Description of Arabic Syntax in Definite Clause Grammar

Summary

El-Shishiny addresses a gap in Arabic computational linguistics: the grammar of Arabic had been documented only descriptively, with no comprehensive formal model. Earlier formal attempts relied mostly on transformational grammar, LFG, dependency or functional grammar. The paper proposes a Definite Clause Grammar (DCG) for Arabic syntax. Nonterminal arguments carry word features such as number, gender, person, definiteness, transitivity, voice and a semantic class. These features enforce agreement and add context sensitivity. Arabic has only three broad word categories (verbs, nominals, particles), so the grammar describes structures through two lenses at once: grammatical category and functional role (for example, 'mobtadaa', predicate, subject, 'naabfael'). Because word order is flexible, it distinguishes nominal, verbal and special-structure sentences. A split complement category handles modifiers separated from their verb by the subject. Semantic features help tell subject from object where morphosyntax cannot. The paper gives no quantitative evaluation, only illustrative parse trees. It presents the grammar as high-coverage and as a contribution toward an agreed formal grammar of Arabic.

Key points

• Three basic sentence types are defined:
  • nominal: no verb, or the verb follows the subject;
  • verbal: the verb precedes the subject;
  • special structures, such as vocatives.
• A nominal sentence is a 'mobtadaa' phrase (mbp) plus a predicate phrase (predp) that agree in number and gender. The mbp must be definite and the predp indefinite. The predp may be an NP, a PP or adverbial, or a whole embedded sentence.
• Feature arguments are attached to each word class:
  • verbs carry transitivity T, voice V, person P and semantic S;
  • nouns carry N, G, definiteness D, P and S;
  • pronouns carry N, G and P.
• Features originate in lexical entries and are inherited upward by phrases.
• A verbal-sentence rule pairs the verb with a subject phrase and a complement phrase (cop). A Prolog-style condition {M=<T} limits the objects the complement may take by the verb's transitivity. Passive verbs take a 'naabfael' phrase instead of a subject.
• The cop category also appears before or after the verb inside the VP. This captures complements split around the subject, illustrated by "the child drinks the water in the house". The nominal example is "the big dog is in the garden", which has no copula.

Interesting ideas

• I1. Pairing every constituent with both a category label and a functional role lets the grammar cover a language whose part-of-speech inventory is too coarse for structure alone.
• I2. A side condition on a rule (complement count bounded by verb transitivity) blocks overgeneration without extra rules.
• I3. Semantic class features on verbs and nouns are used to decide which argument is the subject when word order and morphology leave it open.
• I4. Features are declared once in lexical entries and inherited by phrases, so agreement is checked by unification rather than by ad hoc code.
• I5. A dedicated category for discontinuous complements (pieces of the verb's dependents on both sides of the subject) treats non-adjacency as a first-class structure.

For Mova

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

• I1 (tree): Mova's UD trees could keep POS and deprel as independent checks, flagging parses where a function is implausible for its category.
• I2 (learn): self-induced repair rules could be written as guarded constraints, e.g. rejecting an obj on a verb whose VerbNet frames are intransitive.
• I3 (frames): selectional checks from VerbNet roles (animate agent for "drink") could rank competing agent/patient assignments in prag frames, with the reason logged for explainability.
• I4 (tokens): the en crate could expose morphological features as structured, unifiable records, so the parser and agreement checks read one source of truth.
• I5 (tree): when converting UD dialects, explicitly tagging split dependents could help explain non-projective or fronted constructions instead of treating them as noise.