📄 Building a WordNet for Arabic

Summary

The paper describes the start of the Arabic WordNet (AWN) project, a lexical database for Modern Standard Arabic. It follows the design of Princeton WordNet and the construction methods of EuroWordNet, and it links directly to PWN 2.0 and EuroWordNet. Arabic raises problems that established wordnets did not face. Its script is usually written without short-vowel diacritics, it has no capital letters, and its vocabulary is built from roots combined with derivational patterns. It also has culture-specific concepts with no English equivalent. To give meanings a formal basis, each synset is to be mapped to terms in SUMO, which will be extended with new terms where needed and which defines meaning in first-order logic. The authors also plan an editing tool for lexicographers that handles Arabic script and supports work from several sites at once. The central proposal is to replace English synsets with SUMO as the Interlingual Index (ILI). This would make cross-lingual links less tied to English and more expressive. The resource is intended to be freely available.

Key points

• AWN targets Modern Standard Arabic, is built with EuroWordNet methods, and links to PWN 2.0 and EuroWordNet. SUMO and its domain ontologies supply the formal definitions, and the Sigma environment is being adapted for Unicode and Arabic.
• Input design: users may type words without diacritics, and the system still returns the vocalized forms so the reader can pick the intended word. Diacritics such as shaddah and sukun are often entered wrongly, which would make exact-match queries fail.
• Morphology figures cited:
  • about 85% of words come from tri-literal roots, and there are roughly 10,000 roots;
  • Beesley's FST approach crosses roots and patterns into 90,000 stems and over 72,000,000 abstract words;
  • Buckwalter's analyzer works on an ASCII transliteration.
• Culture-bound entries include zaka:t, suhu:r, hija:b and mu'akhar Sada:q. Polysemy is illustrated with bab, which can mean a door as an object, a door as an aperture, or a book chapter.
• The database has four entity types: item, word, form and link. A shared editor communicates with the server over SOAP so that lexicographers at different sites can maintain one database.

Interesting ideas

• I1. Store entries in their fully specified form but accept underspecified queries (no diacritics) and return all candidate readings, rather than requiring users to type perfectly.
• I2. Group words into semantic fields by shared root, so that derivationally related words of different parts of speech sit together and their relations can be stated explicitly.
• I3. Distinguish accidental homonymy (separate entries) from systematic, predictable polysemy, such as a door as an object versus as an opening. Treat the latter as facets of one meaning that context selects.
• I4. Use a language-neutral ontology as the interlingual hub, with each sense defined by an exhaustive set of typed relations (subsumption plus others) to ontology terms instead of a single English equivalent. This allows concepts without an English counterpart to be linked.
• I5. Treat the effort of writing formal definitions as a diagnostic: it reveals which senses should be split or merged and which glosses are unclear.

For Mova

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

• I1 (tokens): normalize forms with variable spelling (accents, hyphenation, apostrophes, British/American spelling) and keep every candidate lemma until the tagger chooses one.
• I2 (world): link English derivational families (e.g. study/student/studious) in the compiled lexicon, so an entity introduced by a noun can be connected to the event named by the related verb.
• I3 (frames): when assigning roles, choose the facet the verb requires ("knock on" calls for the object, "go through" for the aperture) instead of forcing a choice between unrelated senses.
• I4 (back): translation through the meta-world could anchor meanings in ontology definitions, so culture-specific concepts survive the text → graph → text round trip.
• I5 (learn): flag WordNet senses that the frame/world builder cannot formalize consistently as candidates for repair rules or sense merging.