📄 Enrichment of Arabic WordNet Using Machine Translation and Transformers

Summary

The paper tackles the small size of Arabic WordNet (AWN). AWN has 11,269 Synsets, while Princeton WordNet (PWN) has 117,659. The authors align the two resources and keep only the PWN Synsets that have no Arabic counterpart. For each kept Synset they build a candidate Arabic version: lemmas, a definition and usage examples, with PWN's hierarchy and mapping preserved.

Several translation engines produce competing candidates: Google Translate, Reverso, and an in-house MT5-based model. Transformer models then choose among them:

• Definitions: BERT embeddings score round-trip translations of each gloss.
• Lemmas and examples: AraBERT scores them against the chosen Arabic definition.

Examples are scraped from Reverso and cleaned by a fine-tuned AraBERT filter. A web portal lets humans review and correct the candidates.

The output is 90,127 candidate Synsets. The authors checked a random sample of 1,000 themselves and report 75.4% overall validation accuracy. That rate suggests roughly 68,000 usable Synsets. Expert validation and the final XML integration into AWN are left for future work. The work matters because it shows a largely automatic way to grow a WordNet for a language with few lexical resources.

Key points

• The candidate pool holds 90,127 Synsets, 231,165 candidate words, 297,190 candidate examples and 90,127 definitions. Of these Synsets, 78,498 received at least one example.

• Definition selection:

  • Definitions come from PWN glosses machine-translated by Google Translate, Reverso and a custom MT5 model.
  • Wikidata definitions were also kept as candidates, although they were often imprecise.
  • The winning translation is picked by back-and-forth translation plus cosine similarity of BERT embeddings.

• Lemma and example selection: candidate Arabic lemmas and scraped examples are ranked by AraBERT cosine similarity with the selected Arabic definition. Example similarity scores of about 0.75–0.82 appear in the paper's tables.

• Manual check of 1,000 Synsets (by the authors, using the Collins and Elmaany dictionaries):

  Component	Precision
  Words	78.9%
  Definitions	95.2%
  Examples	81.3%
  Overall (reported)	75.4%

• A web validation platform lets experts inspect, accept or fix each proposed Synset before insertion into AWN, which is distributed via the GlobalWordNet portal.

Interesting ideas

• I1. Instead of trusting one MT engine, run several and keep the output whose back-translation is most similar to the source text
• I2. Use the Synset's gloss as the anchor that disambiguates: candidate lemmas and examples are ranked by how well they fit the definition, not by how well they fit each other
• I3. Harvest usage examples in bulk from a cheap source, drop junk with a light classifier (wrong script or language), then keep the example closest to the sense definition
• I4. Restrict work to the gap: align with the bigger resource and process only entries missing from the target, preserving the source hierarchy
• I5. Report precision separately for each component (lemma, gloss, example) and project the expected correct yield of the full pool

For Mova

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

• I1 (back): text → graph → text and meta-world translation could be scored by regenerating the source and measuring how much meaning survives the round trip.
• I2 (world): when attaching a mention to a WordNet sense, gloss fit can serve as one extra witness alongside frame and role constraints, kept as an explainable score rather than a decision on its own.
• I3 (learn): the same harvest–filter–select loop could create silver sense-tagged sentences for training lemma/sense choices, with rejected items doubling as negative controls.
• I4 (world): compiling ontology from WordNet/Wikidata could diff against the existing graph and import only absent concepts, keeping original IDs so links stay traceable.
• I5 (learn): evaluation of compiled knowledge or silver data could break accuracy down per field, showing which stage of the pipeline needs repair.