Elamur Research⌕
☰
home / papers / 2012 / Arabic Retrieval Revisited: Morphological Hole Filling

Arabic Retrieval Revisited: Morphological Hole Filling

Paper · 2012 · Kareem Darwish and Ahmed Ali

Summary

Arabic search engines depend on stemming. The strongest statistical stemmers still cannot merge singular nouns with their "broken" (internally changed) plurals. They also split named entities whose final letter shifts when a suffix is added. Darwish and Ali address this gap with a character-level transformation model that proposes related forms of a query token.

The training signal comes from Arabic Wikipedia: link anchor texts paired with the titles of the pages they point to. Word pairs that pass simple orthographic filters are aligned letter by letter with Giza++ and Moses, treating each word as a "sentence" of characters. At query time, the learned mappings generate candidate variants, keeping only forms that actually occur in the collection. The candidates are then added to the query with probability weights.

On the TREC 2001/2002 Arabic collection, expanding stems this way beat raw words, statistical stemming and character 4-grams with statistical significance. Expanding unstemmed words matched stemming. The method needs no hand-written rules, catches spelling variants of transliterated names, and could carry over to other languages.

Key points

  • The data is an April 2010 Arabic Wikipedia dump (about 150,000 articles), yielding 11.47 million anchor–title pairs and about 13 million candidate word pairs. The pairs are released as WikiPairs on GitHub.

  • A pair is kept only if three conditions hold:

    • the words share their first two or last two letters;
    • their edit distance is at most 3;
    • that distance is below half the length of the shorter word.

  • Alignment uses Giza++ plus the Moses phrase extractor and scorer, with source character segments capped at 3 letters. Decoding follows a transliteration-mining approach (Udupa et al., 2009) restricted to the collection's vocabulary.

  • Evaluation used the TREC 2001/2002 CLIR collection (383,872 newswire articles, 75 topics) with the Indri toolkit and a paired two-tailed t-test at p < 0.05.

  • MAP scores:

    RunMAP
    Words0.225
    Character 4-grams0.244
    Expanded words0.264
    Stems (reimplemented Diab 2009 SVM segmenter)0.276
    Expanded stems0.296

Interesting ideas

  • I1. Hyperlink anchors and the titles they target are free, naturally occurring pairs of surface forms for one concept, which makes them a cheap source of morphological and spelling variation.
  • I2. Character-level edits can be learned by recasting the task as translation: letters play the role of words and short letter chunks become "phrases", so infix-like changes come out without hand-written rules.
  • I3. Generated candidates are kept only if they really exist in the target vocabulary, which turns an open-ended generator into a ranked search over attested forms.
  • I4. Instead of a hard cutoff, every variant is kept but weighted relative to the original form, so weak candidates fade rather than being dropped or over-trusted.
  • I5. Cheap pair filters combine a fixed edit-distance cap, an edit cap relative to word length, and a shared prefix or suffix, which removes look-alike short words while keeping real variants.

For Mova

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

  • I1 (tokens): Mining English Wikipedia anchor–title pairs could supply irregular plurals, past forms and derivations ("Portuguese"/"Portugal") to test and extend the lemmatizer in crate en.
  • I2 (tokens): A small learned character-mapping table could propose lemma candidates for unknown or misspelled words, each with a probability that can be shown as the explanation.
  • I3 (world): When resolving a name variant to a known entity, only real entries in the knowledge graph would be proposed, ranked by mapping probability, avoiding invented entities.
  • I4 (frames): In coreference, spelling variants of a name (e.g. transliterations) could add graded evidence that two mentions corefer, rather than a yes/no string match.
  • I5 (learn): The same length-relative filters could clean automatically mined silver pairs (lemma pairs, alias pairs) before training, with significance testing against a baseline as the acceptance check.

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 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksInformation Systems · 1,437 works

🧭 Topics 3Research subjects and languages the work is about.Semitic languages · 2,908 worksInformation retrieval & search · 473 worksMorphology · 405 works

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

Cite

Cite as Kareem Darwish and Ahmed Ali. 2012. Arabic Retrieval Revisited: Morphological Hole Filling. Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers).

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

@inproceedings{darwish2012arabic,
  title = {{Arabic Retrieval Revisited: Morphological Hole Filling}},
  author = {Kareem Darwish and Ahmed Ali},
  booktitle = {Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)},
  year = {2012},
  url = {https://aclanthology.org/P12-2043/}
}

References

28 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. M. A. Ahmed. (2000). A Large-Scale Computational Processor of the Arabic Morphology, and Applications. A Master’s Thesis, Faculty of Engineering, Cairo University, Cairo, Egypt. find
  2. M. Aljlayl, S. Beitzel, E. Jensen, A. Chowdhury, D. Holmes, M. Lee, D. Grossman, O. Frieder. IIT at TREC-10. In TREC. 2001. Gaithersburg, MD. find
  3. M. Baroni, J. Matiasek, H. Trost (2002). Unsupervised discovery of morphologically related words based on orthographic and semantic similarity. ACL-2002 Workshop on Morphological & Phonological Learning, pp. 48-57.
  4. M. Brent, S. Murthy, A. Lundberg (1995). Discovering Morphemic Suffixes: A Case Study in Minimum Description Length Induction. 15th Annual Conference on the Cognitive Science Society, pp. 28-36. find
  5. A. Chen, F. Gey (2002). Building an Arabic Stemmer for Information Retrieval. TREC-2002. find
  6. M. Creutz, K. Lagus (2007). Unsupervised models for morpheme segmentation and morphology learning. Speech and Language Processing, Vol. 4, No 1:3, 2007. ↗
  7. K. Darwish. (2002). Building a Shallow Morphological Analyzer in One Day. ACL Workshop on Computational Approaches to Semitic Languages. 2002. find
  8. K. Darwish, H. Hassan, O. Emam (2005). Examining the Effect of Improved Context Sensitive Morphology on Arabic Information Retrieval. ACL Workshop on Computational Approaches to Semitic Languages, pp. 25–30, 2005.
  9. K. Darwish, D. Oard. (2002). Term Selection for Searching Printed Arabic. find
  10. M. Diab (2009). Second Generation Tools (AMIRA 2.0): Fast and Robust Tokenization, POS tagging, and Base Phrase Chunking. 2nd Int. Conf. on Arabic Language Resources and Tools, 2009. find
  11. A. El-Kahki, K. Darwish, M. Abdul-Wahab, A. Taei (2012). Transliteration Mining Using Large Training and Test Sets. NAACL-2012. find
  12. F. Gey, D. Oard (2001). The TREC-2001 Cross-Language Information Retrieval Track: Searching Arabic Using English, French or Arabic Queries. find
  13. TREC, 2001. Gaithersburg, MD. p. 16-23. find
  14. J. Goldsmith (2001). Unsupervised Learning of the Morphology of a Natural Language. Journal of Computational Linguistics, Vol. 27:153-198, 2001. ↗
  15. H. Hammarström (2009). Unsupervised Learning of Morphology and the Languages of the World. Ph.D. Thesis, Dept. of CSE, Chalmers Univ. of Tech. and Univ. of Gothenburg. find
  16. C. Jacquemin (1997). Guessing morphology from terms and corpora. ACM SIGIR-1997, p.156-165. ↗
  17. B. Karagol-Ayan, D. Doermann, A. Weinberg (2006). Morphology Induction from Limited Noisy Data Using Approximate String Matching. 8th ACL SIG on Comp. Phonology at HLT-NAACL 2006, pp. 60–68. find
  18. P. Koehn, H. Hoang, A. Birch, C. Callison-Burch, M. Federico, N. Bertoldi, B. Cowan, W. Shen, C. Moran, R. Zens, C. Dyer, O. Bojar, A. Constantin, E. Herbst (2007). Moses: Open Source Toolkit for Statistical Machine Translation, ACL-2007, demonstration session, Prague, Czech Republic, June 2007. find
  19. L. Larkey, L. Ballesteros, and M. Connell (2002). Improving Stemming for Arabic Information Retrieval: Light Stemming and Co-occurrence Analysis. SIGIR 2002. pp. 275-282. find
  20. Y. Lee, K. Papineni, S. Roukos, O. Emam, H. Has-san (2003). Language Model Based Arabic Word Segmentation. ACL-2003, p. 399 - 406.
  21. J. Mayfield, P. McNamee, C. Costello, C. Piatko, A. Banerjee. JHU/APL at TREC 2001: Experiments in Filtering and in Arabic, Video, and Web Retrieval. In TREC 2001. Gaithersburg, MD. p. 322-329. find
  22. D. Metzler, W. B. Croft (2004). Combining the Language Model and Inference Network Approaches to Retrieval. Information Processing and Management Special Issue on Bayesian Networks and Information Retrieval, 40(5), 735-750, 2004. D. ↗
  23. Oard, F. Gey (2002). The TREC Arabic/English CLIR Track. TREC-2002. find
  24. P. Schone, D. Jurafsky (2001). Knowledge-free induction of inflectional morphologies. ACL 2001.
  25. B. Snyder, R. Barzilay (2008). Unsupervised Multilingual Learning for Morphological Segmentation. ACL-08: HLT, pp. 737–745, 2008.
  26. R. Udupa, K. Saravanan, A. Bakalov, A. Bhole. 2009. "They Are Out There, If You Know Where to Look": Mining Transliterations of OOV Query Terms for Cross-Language Information Retrieval. ECIR-2009, Toulouse, France, 2009. find
  27. J. Wang, D. Oard (2006). Combining Bidirectional Translation and Synonymy for Cross-language Information Retrieval. SIGIR-2006, pp. 202-209. find
  28. J. Xu, A. Fraser, and R. Weischedel (2001). 2001 Cross-Lingual Retrieval at BBN. TREC 2001, pp. 68 - 75. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/darwish-2012-arabic-retrieval-revisited-morphological-hole/