📄 Exploiting Hidden Morphophonemic Constraints for Finding the Underlying Forms of ‘weak' Arabic Verbs

Summary

Written Arabic usually omits short vowels and gemination, so one consonant string such as ktb can stand for many verb and noun forms. "Weak" verbs, whose roots contain w or y, look especially irregular on the surface. Ramsay and Mansour argue that these verbs are actually regular. Their apparent irregularity comes from a small set of phonological contractions. Recovering the underlying form therefore needs no special verb class.

Their analyser combines four parts:

• characters: feature bundles carrying both written and underlying form;
• templates: root patterns whose vowel slots are filled per tense and voice;
• categorial affix rules: applied strictly right to left;
• a lexical trie: its arcs can contain unfilled holes.

Spelling rules, written in Chomsky–Halle context notation, are run in reverse: from surface form back to underlying form. They fire lazily during trie traversal. Decisions such as mood-dependent vowels are postponed until syntax supplies the needed information.

Using the verb škw, the authors show that most of its full conjugation follows from a few w/y rules. The approach keeps the lexicon simple to maintain. On a small set of examples, it gave correct analyses that the Buckwalter analyser missed.

Key points

• Published in the proceedings of RANLP 2011 (Hissar, pp. 448–454), by Allan Ramsay (Manchester) and Hanady Mansour (Qatar University).
• Lexicon entries: each root is a template plus slot fillers for each tense/voice combination. For škw these are actvPast [a,a], psvPast [u,i], actvPres [o,u] and psvPres [o,a].
• Affix combination: affixes combine through extended combinatory categorial rules, in the spirit of Ades and Steedman. Each affix states which further affixes it requires. Words with no visible derivational prefix are given an empty one.
• Example rules:
  • A tamarbuta rule maps word-internal t back to the feminine marker.
  • An inserted-w rule recovers tawaqufa from the written form tqf, but correctly does not recover the passive tuwuqifa.
• Coverage: the same approach treats weak-initial and weak-final verbs, and the authors say it also handles weak-middle verbs. They compared it with the Buckwalter analyser on only a few examples.

Interesting ideas

• I1. Instead of listing every candidate analysis and then picking one, leave underspecified slots unresolved until later syntax can decide them, such as mood fixing the agreement vowel. This avoids combinatorial blow-up.
• I2. Apply reverse spelling rules only while walking a lexical trie, with the path so far as left context and the remaining input as right context. This explores only rewrites that lead to real lexicon branches.
• I3. Treat "irregular" words as regular items transformed by a few general sound rules, rather than as a separate lexical class. This keeps the lexicon small and makes each derivation explainable.
• I4. Let each affix state which further affixes it requires, categorial-style, instead of using one fixed word grammar. This lets different derivational chains branch from the same root.
• I5. Check analyses against a reference grammar's full conjugation table, and explain each odd cell through the underlying components. This is a strong qualitative test of a rule set.

For Mova

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

• I1 (tokens): Keep ambiguous morphological features (e.g. past tense vs. past participle -ed) as open slots that the parser resolves, rather than committing early in the tagger.
• I2 (tokens): A lemmatizer could undo English spelling changes (doubling, y→i, e-deletion) lazily inside trie lookup, so it never generates candidate stems the lexicon does not contain.
• I3 (tokens): Model semi-regular English inflections as base form plus a named rule, so every lemma decision can cite the rule that produced it.
• I4 (tokens): Derivational morphology (un-, -able, -ness chains) could be encoded as affix categories, giving transparent and checkable word-internal structure.
• I5 (learn): Use full paradigm tables as structured test suites for the morphology, flagging any cell the rules cannot derive as a repair-rule candidate.