📄 A Logical Approach to Arabic Phonology

Summary

Feature-structure grammars have usually treated a word's sound shape as a flat string, much like spelling. That assumption fails for Arabic, where a root's consonants are spread across a syllable template supplied by another morpheme. Bird and Blackburn build a modal logic L for describing non-linear phonological structure. It combines interval tense operators for future, past and overlap (overlap plays the role of autosegmental association) with a dominance modality for hierarchy. It also adds nominals, symbols true at exactly one node, which act as names. The authors define frames, valuations and a set of validities that capture well-formed phonological structures. They then state Arabic syllable generalisations as axioms, including sort lattices and appropriateness constraints. Morphological combination becomes logical conjunction. For example, the second-conjugation template conjoined with the root KTB forces kattab, and a short derivation excludes the alternative identifications. The work matters because it brings non-concatenative morphology into a declarative, constraint-based, compositional framework. In that framework every derived form follows from explicit axioms.

Key points

• Language L has propositional variables, nominals, operators F, P and O (overlap/association), and ◇ for dominance. It also defines M ("at some time") and L ("at all times").
• A phonological frame is a structure (T, <, o, δ) with the following properties:
  • Precedence is irreflexive and transitive.
  • Overlap is reflexive and symmetric, and the two relations are disjoint.
  • Precedence carries through overlap, and time is semi-linear.
  • The closure of dominance lies within overlap, and dominance is acyclic.
• The validities are T1–T6, D1–D2, FORCE, NOM, PLIN (tiers are linearly ordered) and NOME. The authors claim decidability via known techniques and plan to build a theorem prover.
• The Arabic axioms include sorts for syllables, moras, onsets, codas, vowels and consonants. They also state that a coda holds at most one segment, a syllable holds at most one vowel, and an onset is a consonant followed by a vowel.
• In the perfect-tense paradigm, (II) ∧ (KTB) yields kattab with a doubled middle consonant, and (II) ∧ (DHRJ) gives dahraj without reentrancy. The reflexive affix t, doubling rules, defaults (glottal stop, a) and extrametricality are discussed.

Interesting ideas

• I1. Nominals name individual nodes, so shared structure (reentrancy) and identity reasoning reduce to ordinary formulas rather than special graph machinery.
• I2. Morphemes combine by conjoining independent partial descriptions, and only structures satisfying all of them survive, much like pages of a spiral notebook constraining one spine.
• I3. The authors prove a form by eliminating the alternative identifications case by case, each through a short contradiction using named axioms.
• I4. Category relations are kept in a sort lattice with explicit meets and joins, plus an appropriateness relation saying which sorts may dominate which.
• I5. Defaults are treated as a preference over which models to build rather than as hard axioms, keeping the core logic monotonic.

For Mova

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

• I1 (world): Entities and events in the text world could get unique-name atoms, so coreference merges become explicit, checkable equalities.
• I2 (frames): A frame could be assembled as the conjunction of constraints from the verb class, its arguments and context, with each surviving reading traceable to the constraints that licensed it.
• I3 (answer): Answers could carry an elimination trace showing which rival readings were ruled out and by which rule, keeping every step explainable.
• I4 (tree): UD labels and features could be checked against such a lattice and appropriateness table to flag impossible head–dependent pairings in parser output.
• I5 (learn): Self-induced repair rules could be stored as soft model preferences on top of hard grammatical constraints, so they can be removed or audited without breaking the core.