📄 Arabic Disambiguation Using Dependency Grammar

Summary

The paper tackles ambiguity in Arabic text. Unwritten short vowels, attached clitics and affixes, and fairly free word order mean that one written form can map to many segmentations, categories and meanings. The authors argue that morphology and syntax depend on each other in Arabic, so the usual rule-based pipeline (morphological analysis first, then parsing) is a poor fit. They propose a single joint, rule-based analysis grounded in Dependency Grammar and implemented with EnCo, the UNL "enconverter". It scans the input left to right, proposes candidate morphemes from the dictionary and applies prioritised rules, backtracking when a choice fails. Its output is a semantic network in UNL, an interlingua. Four sources of evidence are used together: lexical attributes, adjacency constraints, morphological dependencies such as agreement and case, and syntactic head–dependent relations. The system has about 1500 rules and a 120,000-entry dictionary. The authors report fast analysis of long sentences, but give no formal accuracy evaluation. The work matters because it shows how cues from different levels can prune each other's hypotheses during a single analysis.

Key points

• Four knowledge sources drive disambiguation: lexicon attributes, adjacency constraints (e.g. no preposition directly after a preposition), morphological dependencies (agreement, case) and syntactic dependencies.
• Complexity argument against the pipeline:
  • An ATB word has about two morphological solutions on average.
  • Parsing cost grows as the product of per-word alternatives.
  • A 20-word sentence can therefore yield up to 2^20 (1,048,576) combinations.
• For comparison, a pipeline system by Attia needed 141 CPU minutes for 229 test sentences. EnCo analysed a long sentence in 0.3 s and an example sentence in 0.1 s. No direct head-to-head comparison was possible.
• Resources:
  • 1500 EnCo rules.
  • A UNL-Arabic dictionary of 120,000 entries, including functional words, prefixes and suffixes.
  • Output in UNL relations such as agt, obj, man and mod.
• Reported limitation: the system resolves ambiguity across categories (noun vs verb, particle vs verb) well, but is weaker when readings share a category (noun–noun, verb–verb).

Interesting ideas

• I1. Morphological dependency (who agrees with whom, who assigns case) is kept separate from syntactic headedness, and the two can even point in opposite directions. Each acts as an independent check on a candidate analysis.
• I2. Instead of committing to one analysis per word, the system grows a structure and backtracks when a dependency frame is violated. For example, a verb that already has both agt and obj filled cannot take another bare noun.
• I3. A word's dependents must be processed before the word itself is attached to its head. For example, a noun followed by an agreeing adjective cannot become an object until the adjective is absorbed. This works as a simple boundary-detection discipline for entities.
• I4. The product-of-alternatives formula is used as a quick, explainable argument for pruning ambiguity early rather than passing every hypothesis downstream.
• I5. Failures are expected to cluster where cues are weakest: same-category ambiguities that agreement and syntax cannot separate.

For Mova

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

• I1 (tree): agreement edges could be scored separately from UD head choice in the perceptron parser, so number/person mismatches can veto attachments, e.g. for subject–verb links in long English sentences.
• I2 (tree): valency saturation from VerbNet frames could serve as a hard constraint that triggers re-ranking or rollback of parse candidates, instead of accepting an over-saturated verb.
• I3 (world): entity spans could be closed only after their modifiers are attached, giving cleaner entity boundaries before states and holdings are recorded.
• I4 (learn): logging the per-word alternative count and its product per sentence would be a cheap diagnostic for where tagger ambiguity inflates parser cost or error.
• I5 (learn): evaluation could be split into cross-category vs same-category ambiguity buckets, so lexical-semantic errors are measured separately from structural ones.