📄 Arabic morpho-syntactic feature disambiguation in a translation context

Summary

Arabic script usually leaves out short vowels and lets clitics and affixes attach to words. One written form can therefore be a verb, a noun, a particle, or a conjunction fused with a verb. A rule-based analyzer like MORPH2 returns several candidate analyses per word, but tasks such as machine translation need one.

The authors use the translation itself to choose. They take an Arabic–French bitext aligned by sentence:

• Each Arabic word is rewritten as labeled segments (proclitic, prefix, stem, suffix, enclitic).
• The French side is POS-tagged with TreeTagger.
• GIZA++ links Arabic segments to French words.

The tag of the linked French word is mapped onto a coarse Arabic category (verb, noun, particle), which selects among MORPH2's candidates. This setup, called treetagger-to-morph2, reaches 88% accuracy, against 73% for MORPH2 used alone. The work shows that a better-equipped language in a parallel corpus can serve as an inexpensive disambiguator for a morphologically rich one.

Key points

• MORPH2 (Belguith et al., 2006) is a knowledge-based analyzer that works in five steps:
  • tokenization, using the Star sentence splitter;
  • clitic stripping, with filtering for particles, numbers, dates and proper nouns;
  • affixal analysis of roots and affixes;
  • feature analysis: POS, then feature lists, then filtering;
  • full vocalization and validation.
• Data: an Arabic–French parallel text aligned at sentence level. French tags and lemmas come from TreeTagger (Schmid, 1994).
• Word alignment used GIZA++ (Och and Ney, 2003) with a training cascade of IBM-1, HMM, IBM-3 and IBM-4.
• Tag mapping from French to Arabic categories:
  • French noun and adjective become Arabic noun;
  • adverb, conjunction and preposition become particle;
  • verb stays verb.
• Results: on the whole corpus MORPH2 makes 1152 errors (27%), giving the BASELINE of 73%. Alignment-based selection (treetagger-to-morph2) raises accuracy to 88%.

Interesting ideas

• I1. The paper does not tag from scratch: it uses the POS of the aligned word in another language as a vote among analyzer candidates, which is cheap and needs no Arabic annotated corpus.
• I2. An analyzer first proposes a closed candidate set and a separate step only selects from it (the Hajic-style setup the paper builds on), so the final analysis is always one the lexicon licenses.
• I3. Words are split into role-labeled pieces (proclitic_, stem_, enclitic_) before alignment, so stems link to content words and clitics link to function words, rather than whole words being aligned as blobs.
• I4. Two rich tagsets are compared only after collapsing them into a few shared coarse categories (noun, verb, particle), so evidence from another language is used only where the systems can meaningfully agree.

For Mova

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

• I1 (learn): Ukrainian–English parallel text could become an extra witness for silver data, letting the tag of the aligned Ukrainian word vote on English noun/verb homographs such as "present" or "run".
• I2 (tokens): restricting the tagger to lexicon-licensed tags per word keeps outputs valid, and each pick can record which evidence (context or alignment) decided it, keeping it explainable.
• I3 (back): when aligning Mova's English output with Ukrainian during translation through the meta-world, labeled sub-word pieces (contractions, possessive 's, Ukrainian case endings) would make transfer links cleaner and easier to inspect.
• I4 (tree): in UD dialect conversion, a small bridging category layer could decide which cross-treebank agreements count as evidence, leaving fine-grained distinctions to language-internal rules.