📄 A chart parser for analyzing modern standard Arabic sentence

Summary

The paper presents a rule-based syntactic analyzer for Modern Standard Arabic (MSA). The authors argue that Arabic parsing is hard for four reasons: sentences are long, written text omits diacritics, word order is free, and personal pronouns are often implicit. Their system has two parts, both written in Prolog. The first is an existing morphological analyzer that uses an augmented transition network (ATN) to separate stems from inflectional affixes. The second is a bottom-up chart parser driven by a unification-based grammar (UBG). Every grammar rule carries Prolog constraints that enforce agreement and discard readings that are syntactically or semantically impossible. Lexicon features such as transitivity and the rationality (human vs. non-human) of arguments help the parser choose between competing readings. The rules were gathered from MSA sentences in the agriculture domain. The paper illustrates the system with a worked parse of a verbal sentence but reports no quantitative evaluation. Its value lies in showing how morphology, lexical semantics and syntax can work together inside one explainable parser for a heavily ambiguous, unvowelled language.

Key points

• The morphological analyzer comes from Rafea & Shaalan (1993). It searches an ATN exhaustively to list every possible analysis, then runs a lexical disambiguation module and a feature-extraction module. For example, a verb reading is dropped when the word carries a prefix that only nouns can take.
• The lexicon has three word classes (noun, verb, particle), each stored as a Prolog fact. Features fall into two groups: syntactic ones (voice, tense, number, end case, transitivity) and lexical ones (subject/object rationality, adjectivability, noun category).
• The grammar is written in SICStus Prolog 3.10 and contains 170 rules, organised into 22 modules by category (Subject, Object, Defined, conjunction form, and so on).
• Each rule constraint has one of three jobs: enforcing agreement between the left- and right-hand sides, reducing syntactic ambiguity, or reducing semantic ambiguity.
• Some verbs look identical in active and passive voice. For these, the parser uses transitivity plus rationality checks to decide whether a noun is the subject or the pro-agent.

Interesting ideas

• I1. Argument slots in verb entries carry an animacy-like "rationality" feature, and a noun is accepted as subject only if its rationality matches the slot. This is a cheap selectional restriction that resolves role ambiguity.
• I2. The lexicon records how many objects a verb can take, so the grammar can predict how many nominal arguments should follow and can infer voice when the surface form is ambiguous.
• I3. Morphology is split into two stages: first enumerate every stem+affix split, then prune readings that clash with affix constraints. This keeps the generator simple and makes each rejection traceable to a specific rule.
• I4. Each grammar constraint is labelled by purpose (agreement, syntactic, semantic), and some rules print a human-readable reason when they fail, such as an adjective being rejected as pro-agent.
• I5. The 170 rules are grouped into category modules so that they stay maintainable.

For Mova

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

• I1 (frames): VerbNet selectional restrictions such as +animate could act as hard filters when assigning agent vs. patient, especially in ambiguous passives or middle constructions like "the door opened."
• I2 (tree): valency counts from the verb lexicon could rescore parser attachments, penalising trees that give a verb more objects than its frame allows.
• I3 (tokens): the tagger could list all candidate lemma/POS analyses and record which constraint eliminated each one, giving an explainable trace of morphological decisions.
• I4 (learn): tagging every repair or filter rule with its purpose and a failure message would make self-induced rules auditable and would show which kind of constraint caused a regression.
• I5 (frames): organising pragmatic and role-assignment rules into modules per construction type would make it easier to test and toggle them separately during negative-control evaluation.