📄 Arabic Morphology Generation Using a Concatenative Strategy

Summary

The paper addresses the large number of word forms that must be produced when generating Arabic verbs. Arabic builds stems by interleaving a consonantal root with a vowel melody, and then adds prefixes and suffixes for tense, mood, voice, person, number and gender. Earlier computational work, mostly two-level and multi-tape finite-state models, treated this non-concatenatively and focused on analysis. The authors instead build a generator in MORPHE, a CMU rule compiler used in knowledge-based machine translation. MORPHE combines a discrimination tree over feature structures with regular-expression transformation rules. Their main design choice is to split generation into two calls. The first produces the correct stem allomorph. The second attaches the prefixes and suffixes to that stem. Because stem alternations rarely interact with affixation, this split removes much rule duplication. For some verb classes it cuts the rule count by up to a factor of three. The work concentrates on hollow verbs, whose middle radical is weak, and outlines how the method extends to other verbs, broken plurals and derivation.

Key points

• Inflection space: The paradigm covers 2 tenses, 2 voices, 5 moods and 13 person–number–gender combinations. There are 15 triliteral patterns, about 9 in common use, plus 4 rarer quadriliteral ones.
• MORPHE (Leavitt, 1994):
  • The input is a Lisp feature structure, for example ROOT "zawar", PAT CVCVC, VOW HOL, TENSE PERF.
  • Generation walks a morphological form hierarchy down to a leaf.
  • At the leaf it first checks an irregular-forms lexicon, then applies the leaf's rule.
  • If nothing matches, it returns ROOT unchanged.
• Two-step generation:
  • The first call sets the feature CHG to STEM and returns the stem.
  • That stem temporarily replaces ROOT for the prefix/suffix call.
  • Both steps share one tree because MORPHE allows only a single root.
• Hollow verbs: They fall into four classes, each with a short and a long stem allomorph in the perfect and in the imperfect. Two of the classes can be merged for generation, since their stems behave identically.
• Coverage and limits:
  • Generated forms carry full short vowels and diacritics, for teaching use.
  • Testing covered hollow verbs and several strong verb types.
  • Verbs with a weak third radical gain little from the split, because there stem and suffix changes are entangled.

Interesting ideas

• I1. Splitting the stem alternation from affix attachment turns a multiplicative rule space into roughly an additive one, because each rule set only handles its own kind of change.
• I2. The lookup order is fixed: exception lexicon first, then pattern rule, then the unchanged base form as a safe default, so behaviour is predictable and every output can be traced.
• I3. The authors store inflection-class information, such as the middle radical and its vowel, directly in the lexical base form instead of in an abstract template, because the template alone cannot predict which allomorph to use.
• I4. The generator always produces the fully specified form, with every diacritic, and strips detail only at display time, because removing information is easy and restoring it is not.
• I5. Derivation and inflection are chained as two successive calls to the same engine, for example a participle is derived first and then inflected like an ordinary noun.

For Mova

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

• I1 (tokens): The English inflector in crate en could select the stem variant first (go/went, consonant doubling, y→i) and then attach a generic suffix, keeping irregular-stem logic apart from regular affix rules.
• I2 (back): Surface realization from the meaning graph could log which of the three paths produced each word form, making generated text explainable and exceptions easy to audit.
• I3 (tokens): English lemma entries could carry explicit inflection-class features, such as a doubling flag or irregular class, so the lemmatizer and generator need not guess from spelling.
• I4 (back): The text → graph → text compressor could regenerate a maximally explicit form and simplify it afterwards, which keeps round-trip comparisons lossless.
• I5 (world): Compiled knowledge could store a derived word, such as a nominalization or participle, as a link to its base verb plus a regular inflection class, so frames and entities can map nouns back to verb meanings.