📄 An Arabic Slot Grammar Parser

Summary

The paper presents ASG, a rule-based dependency parser for Arabic built in the Slot Grammar (SG) framework, which already had grammars for English, German, and four Romance languages. Arabic required two main adaptations. The first is a feature system that absorbs the heavily ambiguous output of the BAMA morphological analyzer. The second is a clitic mechanism that lets proclitics and enclitics take part in slot filling. Syntax rules are written in SGF, a new, mostly declarative formalism in which every rule describes how a slot gets filled. The Penn Arabic Treebank (ATB) guided development and seeded the lexicon, but nothing was trained on it. A small hand-scored test gave 64% of nodes with the correct head and POS. On a larger ATB3 sample, 72% of sentences received a spanning parse, and the parser ran at roughly 2,500 words per second. The work matters because it shows that a multilingual symbolic grammar framework can take on a morphologically rich Semitic language. In doing so, the framework itself became cleaner and more general.

Key points

• Morphology integration: BAMA (the Qamus version) supplies the morphology, and all internal processing uses Buckwalter transliteration, although the parser accepts UTF-8 Arabic script. Analyses that share a stem and POS are merged into one node carrying disjunctive features, which limits parse-space explosion.
• Feature lexicon (Ar-feas.lx):
  • Features are integer atoms, and each node's feature set is a 512-bit string.
  • Implication entries say that one feature entails others.
  • "Extending feature" entries, such as gen, stand for the disjunction of all BAMA compound tags containing that subfeature, so grammar rules never mention raw BAMA tags.
• Lexicon (Ar.lx): entries are indexed by vocalized stems and list slot frames. The starter version was bootstrapped from ATB stems sorted by frequency, with default frames derived from their BAMA features. Default frames also cover words missing from Ar.lx.
• Parsing algorithm:
  • The parser is a bottom-up chart parser using binary head–modifier slot filling, attaching postmodifiers before premodifiers.
  • Numeric scores prune partial phrases and rank final parses, and complement slots are preferred over adjuncts.
• Evaluation:
  • On 20 random unseen ATB segments, hand scoring found 64% of nodes with the correct mother and POS.
  • On 1,000 ATB3 sentences of 13–20 words, 72% received a complete parse.
  • Speed was 2,500 words/s for parsing (interpreted SGF; a compiler roughly doubles this) and 10,000 words/s for morpholexical processing.

Interesting ideas

• I1. Instead of committing early to one morphological reading, the system collapses analyses that share a stem and POS into a single node with a disjunctive feature set, and agreement constraints later narrow it by bit-set intersection; this keeps ambiguity cheap and resolves it where syntax actually has evidence.
• I2. "Extending features" let rules test a simple abstract feature (like genitive) that is defined as the union of many fine-grained compound tags, so the grammar stays independent of the analyzer's tag inventory.
• I3. Clitics live in dedicated proclitic and enclitic lists on a phrase, enclitics attach before anything else, and a proclitic conjunction can be passed upward until it reaches the clause it really coordinates.
• I4. The evaluation counts a node as correct only if both its mother and its POS are right, and notes that all-correct mothers guarantee the correct tree shape; complete-parse rate is reported separately and explicitly without any correctness guarantee.
• I5. A frequency-sorted list of stems from a treebank, with default frames inferred from their morphological features, served as a quick starter lexicon that grammar writers then refine by hand.

For Mova

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

• I1 (tokens): ambiguous English tags or morphology (e.g. noun/verb homographs with identical lemma) could be carried as bit-set disjunctions into parsing rather than forced to a single tag by the tagger.
• I2 (tree): a declarative feature-hierarchy file could map treebank- or dialect-specific tags onto abstract UD features, making UD dialect conversion and rule tests independent of any one treebank's tagset.
• I3 (tree): English contractions and possessive "'s" could be modeled as attached sub-phrases that get promoted to their true attachment site, giving an explainable route from token-level attachment to the correct UD head.
• I4 (learn): reporting a joint head+POS score alongside a coverage-only metric, clearly labelled as such, would act as a negative control against mistaking high coverage for accuracy.
• I5 (world): verb-class and frame entries could be bootstrapped from treebank frequency lists with VerbNet-style defaults, so manual curation is spent on the most frequent and most problematic lemmas first.