Elamur Research⌕
☰
home / papers / 2018 / Word2Vec vs LSA pour la détection des erreurs orthographiques produisant un dérèglement sémantique en arabe (Word2Vec vs LSA for detecting semantic errors in Arabic language)

Word2Vec vs LSA pour la détection des erreurs orthographiques produisant un dérèglement sémantique en arabe (Word2Vec vs LSA for detecting semantic errors in Arabic language)

Paper · 2018 · Ben Othmane Zribi, Chiraz

Summary

The paper addresses real-word spelling errors in unvowelled Modern Standard Arabic. These are typos that happen to produce a valid, grammatical word whose meaning does not fit the sentence. Arabic words have many close lexical neighbours, so a single typo turns one real word into another far more often than in French or English. Instead of the usual predefined confusion sets, the author flags any word that is semantically out of place. Each word gets a "semantic validity" score: its average angular distance to the other words of the sentence, with closer context words weighted more heavily. A word is suspected when the score falls below an empirically set threshold. Two vector spaces are compared, LSA and Word2Vec Skip-Gram. The study assumes one error per sentence, produced by a single edit operation. Skip-Gram beat LSA on both precision and recall. It also reached a higher F-measure than an earlier multi-agent system (SMA) for Arabic, though with slightly lower precision. This shows that plain embeddings can support confusion-set-free error detection in a lexically dense language.

Key points

  • Arabic word forms have on average 26.5 neighbours at edit distance 1 (maximum 185), compared with about 3.5 for French and 3 for English. A typo yields a valid word 5.79% of the time in Arabic, versus 0.59% in English and 0.39% in French.
  • Results as reconstructed from the garbled table:
    • Skip-Gram: precision 89.48, recall 86.15, F 87.78.
    • LSA (this work): precision 83.78, recall 85.33, F about 84.5.
    • Earlier SMA: precision 90.55, recall 82.73, F 86.46.
    • Earlier LSA: precision 79.62, recall 84.44, F 81.96.
  • The English abstract claims LSA wins, but the French abstract, the results and the conclusion all say Skip-Gram wins. This is apparently an error in the abstract.
  • Scope assumptions: one semantic error per sentence, caused by one edit operation (insertion, deletion, substitution or swap of adjacent letters). Such operations account for roughly 90% of spelling errors.
  • Known failure: non-contiguous collocations such as "catch up with the convoy (of recovery)" produce false positives. The odd word is idiomatic, not wrong.

Interesting ideas

  • I1. The method scores each word by how well it coheres with the rest of its sentence and flags outliers, so no list of confusable word pairs is needed; this matters when every word has dozens of near-spelling neighbours.
  • I2. Context words that sit closer to the target in the sentence count more when averaging similarity, a simple way to favour local evidence over sentence-wide noise.
  • I3. The error analysis shows that idioms whose parts are separated by other words break word-by-word coherence, because the "odd" word is licensed by a distant partner rather than by general topic.
  • I4. The paper compares two vector spaces under identical scoring and threshold logic, and notes that the ranking may flip with corpus size.

For Mova

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

  • I1 (tokens): A cheap coherence check on lemmas could flag likely real-word typos (e.g., "desert/dessert") before parsing, so the tagger can try an edit-distance-1 alternative and log why it did.
  • I2 (frames): When checking whether a role filler fits its predicate, weighting nearby arguments more heavily than distant ones could give a graded, explainable fit score for semantic roles.
  • I3 (world): The compiled knowledge layer could store discontinuous collocations as single units, so incoherence checks treat the pair as one item and skip it.
  • I4 (learn): Evaluations of any embedding-based component should vary training corpus size and domain as controls before declaring one representation better.

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 worksLexical semantics & word embeddings · 63 works

💡 Concepts 1Key ideas and methods explained on our concept pages.Latent semantic indexing · 29 works

Cite

Cite as Ben Othmane Zribi, Chiraz. 2018. Word2Vec vs LSA pour la détection des erreurs orthographiques produisant un dérèglement sémantique en arabe (Word2Vec vs LSA for detecting semantic errors in Arabic language). Actes de la Conférence TALN. Volume 1 - Articles longs, articles courts de TALN.

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

@inproceedings{chiraz2018word2vec,
  title = {{Word2Vec vs LSA pour la détection des erreurs orthographiques produisant un dérèglement sémantique en arabe (Word2Vec vs LSA for detecting semantic errors in Arabic language)}},
  author = {Ben Othmane Zribi, Chiraz},
  booktitle = {Actes de la Conférence TALN. Volume 1 - Articles longs, articles courts de TALN},
  year = {2018},
  url = {https://aclanthology.org/2018.jeptalnrecital-court.11/}
}

References

17 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 LTSZYLER E. (2016). Comparative study of LSA vs Word2vec embeddings in small corpora: a case study in dreams database. CoRR Journal abs/1610.01520. ↗
  2. B EN H AMADOU A. (1993). Vérification et correction automatique par analyse affixale des textes écrits, le cas de l’arabe non voyellé. Thèse d’état, Faculté des sciences de Tunis, 1993. find
  3. B EN O THMANE Z RIBI C., B EN F RAJ F., B EN A HMED M. (2005). Un Système Multi-agent pour la Détection et la Correction des Erreurs Cachées en Langue Arabe. Actes de la 5ème conférence sur le Traitement Automatique des Langues Naturelles, Dourdan, France, 143-153. find
  4. B EN O THMANE Z RIBI C., BEN AHMED M. (2013). Detection of semantic errors in Arabic texts. Artificial intelligence journal (195), 249-264. find
  5. B ERGSMA S., P ITLER E., L IN D. (2010). Creating Robust Supervised Classifiers via Web-Scale Ngram Data. Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, Sweden, 865-874. find
  6. C ARLSON A.J., R OSEN J., R OTH D. (2001). Scaling up Context-sensitive Text Correction. Proceedings of 13th Conference on Innovative Applications of Artificial Intelligence IAAI’01, Washington, USA, 45-50. find
  7. G OLDING A.R. (1995). A Bayesian hybrid method for context-sensitive spelling correction. Proceedings of the 3rd Workshop on Very Large Corpora, Massachusetts, USA, 39–53. find
  8. G OLDING A. R., S CHABES Y. (1996). Combining trigram based and feature based methods for context sensitive spelling correction, in: Proceedings of the 34th Annual Meeting of the Association for Computational Linguistics, Santa Cruz, 71-78. find
  9. G OLDING A.R., R OTH D. (1999). A winnow-based approach to context-sensitive spelling correction, Machine Learning journal, 34(1-3), 107-130. find
  10. G UTIERREZ F., D OU D., F ICKAS S., G RIFFITHS G. (2014). Online Reasoning for Ontology-Based Error Detection in Text. OTM international conference on ontologies, databases and application of semantics, 562-579. find
  11. H IRST G., B UDANITSKY A. (2005). Correcting real-word spelling errors by restoring lexical cohesion. Natural Language Engineering (11), 87-111. find
  12. L ANDAUER T. K., F OLTZ P.W. , L AHAM D.(1998). An introduction to Latent Semantic Analysis, Discourse Processes (25), 259-284. find
  13. M ANGU L., B RILL E. (1997). Automatic Rule Acquisition for Spelling Correction, in: Proceedings of the 14 th International Conference on Machine Learning, Nashville, 734-747. M IKOLOV T., S UTSKEVER I., C HEN K., C ORRADO G., D EAN J. (2103). Distributed representations of phrase and their compositionality. Advances in neural information processing systems, 3111-3119. find
  14. S ALTON G., W ONG A., Y ANG C. S. (1975). A vector space model for automatic indexing. Communications of the ACM (18), New York, NY, USA, 613-620. find
  15. S CHUTZE H. (1998). Automatic Word sense discrimination, Journal of Computational Linguistics (24), 97-123.
  16. V ERBERNE S. (2002). Context sensitive spell checking based on word trigram probabilities. Master thesis Taal, Spraak & Informatica, Nijmegen University. find
  17. ZESCH T. (2012). Detecting Malapropisms Using Measures of Contextual Fitness. Special Issue of the TAL Journal on “Managing Noise in the Signal: Error Handling in Natural Language Processing (53), 11-31. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zribi-2018-word2vec-vs-lsa-pour-la/