📄 A Semi-Supervised Method for Arabic Word Sense Disambiguation Using a Weighted Directed Graph

Summary

The paper addresses word sense disambiguation (WSD) for Arabic, where labeled data is scarce. It combines unsupervised and supervised elements. First, it uses glosses and synonyms from Arabic Wordnet to sort raw corpus sentences into one group per sense. Roots of gloss words are found with the Khoja stemmer, and approximate string matching locates their variants in a newspaper corpus of roughly 124 million words. The sentences in each sense group are turned into small position-based trees centred on the ambiguous word, and these trees are merged into a single "semantic tree" per sense.

To disambiguate a new sentence, its own tree is aligned level-by-level with each sense tree. The alignment links are weighted by collocation statistics (T-test, MI, χ²), which yields a weighted directed graph, and a normalized score selects the best-matching sense. A small supervised voting step then arbitrates between the three measures, using reputations learned from labeled examples. The combination reaches about 83% F-score on 127 ambiguous words, which beats each single measure and earlier Arabic WSD systems.

Key points

• Resources: Arabic Wordnet (AWN), the Khoja stemmer, an approximate string matching algorithm (Elloumi, 1998), a 29,985-entry stop-list and a corpus of 123,854,642 newspaper words. Contexts from the web were added for senses missing from the corpus.
• Evaluation covered 127 ambiguous words and 42,316 test sentences, with 40 per sense. The voting step was trained on 20 labeled samples per sense, 4,582 in total, and inter-annotator agreement was about 95%.
• F-scores were 0.747 for the T-test, 0.710 for MI and 0.766 for χ². The voting procedure reached 0.830 for precision, recall and F-score, a gain of 6.4 points over the best single measure.
• Performance grows once semantic trees exceed about 500 nodes and levels off between 2,000 and 3,000 nodes.
• On 50 overlapping words, the best earlier knowledge-based system (modified Lesk with Leacock-Chodorow) reached 67.73% precision, and the best supervised one (KNN) reached 52.02%. The method also improves precision by about 10% over the authors' earlier unsupervised approach.

Interesting ideas

• I1. Position-aware sense profiles. Each sense is modelled as a merged tree whose levels encode distance to the left or right of the target word, and a new context is compared only at matching levels. This keeps word order in the comparison instead of throwing context into an unordered pool.
• I2. Dictionary glosses as seeds for corpus clustering. Gloss and synonym roots are used to collect example sentences for each sense from raw text, giving cheap pseudo-labeled data with no manual tagging.
• I3. Measure arbitration by track record. When the scorers disagree, the one with the best recent record wins, and ranks rise or fall depending on whether a measure sided with the majority. This is a lightweight alternative to fixed ensemble weights.
• I4. Fuzzy matching against roots. Approximate matching based on longest common subsequence links inflected or affixed surface forms back to a root when exact lookup fails.
• I5. Profile-size learning curve. The authors measure accuracy as a function of profile size and find a plateau, which shows how much evidence per sense is actually enough.

For Mova

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

• I1 (frames): Sense profiles for predicates could be indexed by UD dependency relation and depth rather than linear offset, so choosing a VerbNet class stays structural and explainable.
• I2 (learn): WordNet/FrameNet definitions could seed the harvesting of silver sense-tagged sentences, which other witnesses could then confirm or reject.
• I3 (learn): Witness parsers or LLM teacher tiers could carry dynamic reputations that settle silver-data conflicts, with every decision logged so the outcome can be explained.
• I4 (tokens): This could serve as a fallback lemmatizer for misspelled or rare English word forms not covered by the morphology tables, with the alignment kept as evidence.
• I5 (world): Measuring a similar curve for compiled knowledge entries such as verb-class or ontology profiles would show where adding more data stops paying off.