Elamur Research⌕
☰
home / papers / 2014 / An Approach based on semantic trees for lexical disambiguation of Arabic language using a voting procedure (Approche basée sur les arbres sémantiques pour la désambiguïsation lexicale de la langue arabe en utilisant une procédure de vote) [in French]

An Approach based on semantic trees for lexical disambiguation of Arabic language using a voting procedure (Approche basée sur les arbres sémantiques pour la désambiguïsation lexicale de la langue arabe en utilisant une procédure de vote) [in French]

Paper · 2014 · Laroussi Merhbene, Anis Zouaghi and Mounir Zrigui

Summary

The paper addresses word sense disambiguation (WSD) for Arabic, where many words have numerous out-of-context senses and no shared evaluation benchmarks exist. The authors propose a semi-supervised method. In the supervised part, dictionary glosses for each sense are reduced to roots with the resource-free Al-Shalabi, Kanaan and Al-Serhan stemmer. An approximate substring matching algorithm (Elloumi) then finds occurrences of those roots in corpus sentences, which sorts the sentences into one cluster per sense. Each sentence becomes a binary tree rooted at the ambiguous word, with neighbours placed at depths that reflect their distance from it. The trees for a sense are merged into an n-ary "semantic tree." A new sentence's tree is matched against each sense's semantic tree, giving a weighted acyclic graph. A score built from three collocation measures (T test, MI, χ2) picks the closest sense. A voting procedure ranks and weights those measures. The reported precision and recall reach about 83%. This matters as a structured, inspectable alternative for a language with scarce WSD resources.

Key points

  • Preprocessing:
    • Sentences are segmented on punctuation and kept only if they have more than three words.
    • Stop words are removed with a general list of 29,985 entries compiled by Arabic linguists. The authors chose a general list over a corpus-derived one, following an earlier IR comparison.
  • Root extraction:
    • Each letter gets a weight between 0 and 5, using six letter groups.
    • Each letter also gets a rank that depends on word length and parity.
    • The three letters with the smallest weight × rank product form the root. The stemmer's reported accuracy is about 90%.
  • Speeding up matching: approximate matching is slow, so found root occurrences are stored in a knowledge base and checked before the algorithm is rerun.
  • Per-word F-scores for the most frequent sense:
    • Ain 0.7421
    • Hsb 0.7532
    • Fjr 0.8420
    • Chaar 0.8926
    • nr 0.9605
  • Trends in the results:
    • Words with more senses score lower.
    • χ2 is the top-ranked collocation measure for most of the test data.
    • Overall precision and recall reach about 83%.

Interesting ideas

  • I1. Encode each context as a tree whose depth equals distance from the target word, then merge all contexts of one sense into a single weighted sense profile; proximity is thus built into the structure rather than into a flat feature list.
  • I2. Several association statistics (T test, MI, χ2) score the same match, and a voting step decides how much each one counts instead of trusting a single metric.
  • I3. Sense clusters are seeded automatically by finding corpus sentences that contain roots from each sense's dictionary gloss, so sense-labelled examples come without manual annotation.
  • I4. Results that are costly to compute (fuzzy root matches) are memoised in a persistent knowledge base, which is consulted before falling back to the expensive algorithm.
  • I5. Evaluation is broken down per ambiguous word and tied to its number of senses, which shows where the method fails instead of reporting only one average.

For Mova

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

  • I1 (frames): A per-sense profile keyed on dependency distance (not linear distance) from a verb could help choose among VerbNet classes in an inspectable way.
  • I2 (learn): The same pattern fits witness-parser agreement, where each witness earns a learned weight per phenomenon rather than a fixed equal vote.
  • I3 (learn): Gloss lemmas from WordNet could harvest silver sense-tagged sentences to train or check sense choice in the world ontology.
  • I4 (tokens): A persistent cache of resolved lemma or morphology lookups for rare or misspelled forms would keep the slow fallback path explainable and fast.
  • I5 (answer): Reporting accuracy by item difficulty (for example, the number of candidate readings in a test question) would expose weak spots hidden by aggregate scores.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 2,908 worksFrench · 70 works

💡 Concepts 1Key ideas and methods explained on our concept pages.Stemming · 17 works

Cite

Cite as Laroussi Merhbene, Anis Zouaghi and Mounir Zrigui. 2014. An Approach based on semantic trees for lexical disambiguation of Arabic language using a voting procedure (Approche basée sur les arbres sémantiques pour la désambiguïsation lexicale de la langue arabe en utilisant une procédure de vote) [in French]. TALN-RECITAL 2014 Workshop RLTLN 2014 : Réseaux Lexicaux pour le TAL (RLTLN 2014 : Lexical Networks for NLP).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{merhbene2014approach,
  title = {{An Approach based on semantic trees for lexical disambiguation of Arabic language using a voting procedure (Approche basée sur les arbres sémantiques pour la désambiguïsation lexicale de la langue arabe en utilisant une procédure de vote) [in French]}},
  author = {Laroussi Merhbene, Anis Zouaghi and Mounir Zrigui},
  booktitle = {TALN-RECITAL 2014 Workshop RLTLN 2014 : Réseaux Lexicaux pour le TAL (RLTLN 2014 : Lexical Networks for NLP)},
  year = {2014},
  url = {https://aclanthology.org/W14-6702/}
}

References

15 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. A GIRRE E. AND E DMOND P. (2006). Word Sense Disambiguation: Algorithms and Applications. Springer. find
  2. A L -S HALABI R., K ANAAN G., A L -S ERHAN H. (2003). New approach for extracting Arabic roots. Papier présenté à ACIT, the International Arab Conference on Information Technology. Egypt, p. 42-59. find
  3. A LAJMI A., S AAD E.M., D ARWISH R.R.(2012). Toward an Arabic Stop-words List Generation. International Journal of Computer Applications (0975-8887), vol.46, n°8, p. 9-13. find
  4. E L -K HAIR I. A. (2006). Effect of Stop Words Elimination for Arabic Information Retrieval: A comparative Study. International journal of Computing & Information Sciences, vol.4, n°3, p.119-133. find
  5. E LLOUMI M. (1998). Comparison of Strings Belonging to the Same Family. Information Sciences, An International Journal, Elsevier Publishing Co., Amsterdam, North-Holland (Publisher), vol. 111, n°(1-4), p. 49-63. find
  6. I DE I. AND V ERONIS J. (1998). Word Sense Disambiguation : The State of the Art. Computational Linguistics, vol. 24 (1), p. 1-41. find
  7. M ANNING C., S CHUTZE H. (1999). Foundations of Statistical Natural Language Processing. MIT Press, Cambridge, MA. find
  8. M ERHBENE L., Z OUAGHI A. AND Z RIGUI M. (2012). Lexical Disambiguation of Arabic Language: An Experimental Study ». In proceeding of 11 th Mexican International Conference on Artificial Intelligence, San Luis Potosí, SLP, México. find
  9. M UṢṬAFA M., S AYED A HMED N., D ARWICH M., A BDALLAH A. (2008). Mu‘jam al-Wasīṭ. Published in Bayrūt : Dār Iḥyā’ al-Turāth al-‘Arabī lil-Ṭibā‘ah wa-al-Nashr wa-al-Tawzī‘. find
  10. M IHALCEA , R., T ARAU , P., A ND F IGA E. (2004). Pagerank on semantic networks, with application to word sense disambiguation. In Proceedings of the 20 th International Conference on Computational Linguistics (COLING, Geneva, Switzerland), p. 1126–1132.
  11. N AVIGLI , R. (2005). Semi-automatic extension of large-scale linguistic knowledge bases. In Proceedings of the 18th Florida Artificial Intelligence Research Society Conference (FLAIRS, Clearwater Beach, FL), p.p: 548–553. find
  12. N AVIGILI R. (2009). Word Sense Disambiguation: A Survey. ACM Computing Surveys, Vol. 41, No. 2, Article 10, Publication date: February, p. 1-69. find
  13. V ERONIS , J. (2004). Hyperlex: Lexical cartography for information retrieval. Comput. Speech Lang. 18, 3, p.p: 223– 252. find
  14. Y AROWSKY , D. (1993). O NE S ENSE P ER C OLLOCATION . In Proceedings, ARPA Human Language Technology Workshop. Princeton, pp. 266-271. find
  15. Z OU F., W ANG L., D ENG X., H AN S. AND W ANG L. S. (2006). Automatic Construction of Chinese Stop Word List », in proceeding of the 5 th WSEAS International Conference on Applied Computer Science, Hangzhou, China, p. 1010-1015. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/merhbene-2014-approach-based-semantic-trees-lexical/