📄 Construction d'une grammaire d'arbres adjoints pour la langue arabe (Construction of a tree adjoining grammar for the Arabic language)

Summary

The paper tackles how to write a formal grammar for Arabic, a language where syntax depends heavily on morphology and semantics. For example, a plural noun denoting non-humans normally takes a feminine singular adjective, while a human plural agrees fully. Instead of extracting rules from a treebank, the author builds ArabTAG, a hand-designed Tree Adjoining Grammar. Its elementary trees carry rich unification feature structures covering morphological, syntactic, semantic and compositional properties.

To support this work, she developed a semi-automatic, modular authoring application. Through a friendly interface, a linguist creates elementary trees and attaches features to nodes. The tool checks the tree hierarchy and the feature structures so that rules stay correct and uniform, then serializes the grammar to XML.

Using the tool, version 1.0 of ArabTAG was produced: 380 elementary trees grouped into four families. The grammar covers ellipsis, anaphora, subordinate clauses, word-order variation and clitic agglutination. This matters because corpus-extracted grammars only cover constructions that the corpus happens to contain.

Key points

• ArabTAG v1.0 contains 380 elementary trees in four families, built on the reference grammars of Dahdah (1992) and Kouloughli (1992).
• Only "generic" trees are encoded, holding just the obligatory elements (e.g., verb, subject and objects, or theme and predicate). Optional complements such as time, place or manner adverbials attach through TAG operations, which keeps word-order permutations from multiplying the number of trees.
• Node features include Gen, Nombr, Cas/TestCas, Def/TypeDef/TestDef, Pronom, Trans, Voix, Temps, Proc and Enc, plus omission flags (OmissionSujet, OmissionCOD, etc.), Propre, Rôle, Animé and Humain.
• The prior TAG of Habash and Rambow (2004) was extracted from PATB part 1 v2.0 (160,000 news words). It captured VSO and SVO orders but no VOS, because that order was missing from the corpus.
• Planned extensions include a prediction module based on the existing grammar, full lexicalization, and converters to other formalisms.

Interesting ideas

• I1. Treat animacy and humanness as first-class grammatical features, because agreement can hinge on them rather than on surface number alone.
• I2. A grammar derived from a corpus inherits that corpus's blind spots, as the missing VOS order shows; hand-written structures can deliberately fill such gaps.
• I3. Encode only the minimal core of each construction and attach optional modifiers separately, which avoids a combinatorial explosion of templates when word order is flexible.
• I4. Explicit features state whether a slot such as subject or object may or must be omitted, so ellipsis is predicted rather than treated as an error.
• I5. An authoring tool that validates hierarchy and feature consistency as each rule is entered, then exports to a neutral XML format, keeps a hand-built resource uniform.

For Mova

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

• I1 (frames): Animé/Humain-style flags on mentions could constrain pronoun resolution (he/she vs. it, who vs. which) and filter semantic-role candidates in a way that remains explainable.
• I2 (learn): audit silver treebanks for constructions that never or rarely occur, then add controlled synthetic or hand-crafted examples for them as targeted coverage tests.
• I3 (tree): repair rules and UD dialect conversions could be keyed on core-argument skeletons only, with obl/advmod handled by a generic attachment step, keeping the rule set small.
• I4 (frames): frame slots could record whether an argument is expected but unrealized (imperatives, agentless passives, ellipsis), letting the world builder fill it from context or flag it.
• I5 (world): a validator-backed editor for verb classes and ontology entries would catch inconsistent feature values early and keep the compiled knowledge independent of any one consumer.