📄 Arabic NLG Language Functions

Summary

When this work was written, Arabic had no surface realiser or NLG library. The authors attribute this to the language's heavy morphology and to scarce resources. They first describe the phenomena that make realisation hard:

• nouns carry gender, three cases, and a dual alongside sound and broken plurals;
• numerals often take the opposite gender to the counted noun;
• adjectives agree with nouns, with a special rule for non-human plurals;
• verbs vary by person, number, gender, mood, voice and tense;
• pronouns can be suffixes or standalone words.

They then implement, from scratch with hand-written rules, six language functions modelled on RosaeNLG rather than SimpleNLG. These are backed by a lexicon of about 9000 lemmata collected from English Wiktionary. On 250 unit tests every output was correct, and roughly 85% of sample sentences from the NADA corpus could be produced with the functions. Code, lexicon and a web interface are public, giving Arabic a first base for rule-based generation pipelines.

Key points

• The six functions handle nouns (inflectNoun), noun phrases (buildNP), verbs (inflectVerb), adjectives (inflectAdjective), number spell-out (numToWords) and pronouns. Neither clause construction nor non-personal pronouns are supported.
• The lexicon was web-scraped from English Wiktionary because the Arabic dump was poorly formatted. It holds about 4000 nouns, 3000 verbs and 2000 adjectives.
• Short vowel marks (ḥarakāt) are ignored. Verbs missing from the lexicon fall back to rules that assume the verb follows Form I.
• Evaluation followed Braun et al. (2019) and had two parts:
  • 250 test units across nouns, adjectives, verbs, numerals and pronouns, all correct;
  • about 10 sentences from each of NADA's 10 categories, with an 85% generation success rate.
• The system is open source on GitHub, with a web interface.

Interesting ideas

• I1. Each inflection call first looks up the lexicon and only applies general rules when the word is missing, so stored irregulars always win over the productive pattern.
• I2. Agreement is treated as a relation inside the noun phrase: one builder coordinates noun, numeral and adjectives together instead of inflecting each word on its own.
• I3. The adjective function supports both the older reduced agreement and modern full agreement for non-human plurals, so the caller picks the norm.
• I4. Coverage is measured by trying to regenerate real sentences sampled evenly across a corpus's categories, separately from unit-test correctness.
• I5. numToWords converts numbers to words under the full agreement rules, rather than treating numbers as opaque tokens.

For Mova

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

• I1 (tokens): The same lookup-then-rules order could drive English inflection and lemmatisation in crate en, keeping irregular forms authoritative and making it explainable which path produced each form.
• I2 (back): A phrase-level builder that pushes number and definiteness from the head noun to its dependents would make generation from the meaning graph produce consistent phrases by construction.
• I3 (back): Exposing variant choices (e.g. singular vs plural agreement with collective nouns) as explicit generation parameters would let the text → graph → text compressor reproduce the source's own style.
• I4 (learn): Mova's generator could be scored by how many sampled sentences per genre it can rebuild from their graphs, with failures logged by genre to show missing constructions.
• I5 (world): Linking measures and counts in the world model to a rule-based spell-out of numbers (cardinal vs ordinal) would let answers to maths word problems state quantities in fluent, checkable text.