Elamur Research⌕
☰
home / papers / 2008 / The Effects of Language Relatedness on Multilingual Information Retrieval: A Case Study With Indo-European and Semitic Languages

The Effects of Language Relatedness on Multilingual Information Retrieval: A Case Study With Indo-European and Semitic Languages

Paper · 2008 · Peter Chew and Ahmed Abdelali

Summary

The paper asks whether cross-language retrieval involving Semitic languages, especially Arabic, can be improved by changing only the training data while the retrieval model stays the same. The authors use multilingual Latent Semantic Analysis (LSA). Verse-aligned Bible translations are concatenated into a term-by-document matrix, weighted with log-entropy, and reduced by SVD to 300 dimensions. Documents are then compared by cosine similarity. Testing uses the 114 suras of the Quran in Arabic, English, French, Russian and Spanish. Success is measured by how often a document's true translation ranks first.

Three interventions are compared:

  • adding related (Semitic) languages to training,
  • adding an unrelated language that shares Arabic script,
  • morphological preprocessing: light-stemming Arabic and stripping Hebrew vowels.

Adding more training languages helped generally. Script sharing did not help. Arabic stemming improved overall precision but barely helped Arabic pairs. Arabic gained most when Hebrew was added, or when Hebrew was made statistically closer to unvocalized Arabic. The result suggests that genetic relatedness, and putting related languages on a comparable footing, matters for multilingual training data.

Key points

  • Training and test data:
    • Training uses Bible translations from www.unboundbible.org, aligned by verse.
    • The test set has 570 documents (114 Quran suras × 5 languages).
    • The metric is precision at 1 document.
  • Model setup:
    • Log-entropy weighting.
    • SVD keeping the top 300 singular values.
    • Projection of unseen documents via U × S⁻¹.
  • Type/token contrast (Gospel of Matthew):
    • Afrikaans has 2,112 types over 24,729 tokens.
    • Arabic has 4,607 types over 13,930 tokens.
    • Amharic has 5,161 types over 12,940 tokens.
    • This illustrates how synthetic Semitic morphology packs more into each token.
  • Hebrew vowel removal (Westminster Leningrad Codex): Arabic-pair precision rose from 0.476 to 0.528 (p ≈ 7.33 × 10⁻⁴). Other pairs showed no significant change.
  • Eleven training languages: overall 0.739, within-IE 0.853, IE–Semitic 0.537.
  • Adding Hebrew as a twelfth language:
    • IE–Semitic precision rose to 0.645 (p ≈ 4.39 × 10⁻¹³).
    • Overall precision rose to 0.778.
  • Arabic preprocessing: the light stemmer used was Darwish 2002.

Interesting ideas

  • I1. Comparing type and token counts on the same parallel passage gives a cheap gauge of how much meaning each surface token carries in each language.
  • I2. The study freezes the whole pipeline and varies only the training data, so any change in results can be attributed cleanly to the data.
  • I3. Stripping Hebrew vowels made Hebrew resemble the unvocalized Arabic, which helped the model link them; aligning conventions between sources can matter more than adding raw volume.
  • I4. Retrieval of the true translation among all candidates serves as an evaluation that needs no human relevance judgments.
  • I5. The authors tested a plausible but wrong cause (shared script, via Persian) as a control and found that gains from preprocessing one language could appear in other pairs.

For Mova

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

  • I1 (tokens): Run the same count on parallel or paraphrased English text to spot where the tokenizer and lemmatizer leave too much packed inside one token.
  • I2 (learn): Apply the same discipline when testing new silver data or teacher tiers: keep the parser and training code fixed and swap only the data, so every gain stays explainable.
  • I3 (tree): Before mixing treebanks, convert them to one UD dialect so the parser does not learn spurious distinctions caused only by annotation style.
  • I4 (back): Check the text → graph → text compressor by asking whether each regenerated text is matched back to its own source among many candidates.
  • I5 (learn): Include such negative controls when adding repair rules, and score all components, not just the targeted one, to catch displaced effects.

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

 Institutions

🔬Sandia National LaboratoriesAlbuquerque · United States2 works · 32,682 citations

 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 worksMultilingual & cross-lingual NLP · 828 worksInformation retrieval & search · 473 works

💡 Concepts 2Key ideas and methods explained on our concept pages.Topic modeling · 47 worksLatent semantic indexing · 29 works

Cite

Cite as Peter Chew and Ahmed Abdelali. 2008. The Effects of Language Relatedness on Multilingual Information Retrieval: A Case Study With Indo-European and Semitic Languages. Proceedings of the 2nd workshop on Cross Lingual Information Access (CLIA) Addressing the Information Need of Multilingual Societies.

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

@inproceedings{chew2008effects,
  title = {{The Effects of Language Relatedness on Multilingual Information Retrieval: A Case Study With Indo-European and Semitic Languages}},
  author = {Peter Chew and Ahmed Abdelali},
  booktitle = {Proceedings of the 2nd workshop on Cross Lingual Information Access (CLIA) Addressing the Information Need of Multilingual Societies},
  year = {2008},
  url = {https://aclanthology.org/I08-6001/}
}

References

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

  1. Abdou, S., Ruck, P., and Savoy, J. 2005. Evaluation of Stemming, Query Expansion and Manual Indexing Approaches for the Genomic Task. In Proceedings of TREC 2005. find
  2. Berry, M. W., Dumais, S. T., and O’Brien, G. W. 1994. Using Linear Algebra for Intelligent Information Retrieval. SIAM: Review, 37, 573-595. ↗
  3. Biola University. 2005-2006. The Unbound Bible. Accessed at http://www.unboundbible.com/ on February 27, 2007. ↗
  4. Chew, P. A., and Abdelali, A. 2007. Benefits of the ‘Massively Parallel Rosetta Stone’: Cross-Language Information Retrieval with over 30 Languages, Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, ACL 2007. Prague, Czech Republic, June 23–30, 2007. pp. 872-879. find
  5. Chew, P. A., Verzi, S. J., Bauer, T. L., and McClain, J. T. 2006. Evaluation of the Bible as a Resource for Cross-Language Information Retrieval. Proceedings of the Workshop on Multilingual Language Resources and Interoperability, 68–74. find
  6. Darwish, K. 2002. Building a shallow Arabic morphological analyzer in one day. In Proceedings of the Association for Computational Linguistics (ACL-02), 40th Anniversary Meeting. pp. 47-54. find
  7. Dumais, S. T. 1991. Improving the Retrieval of Information from External Sources. Behavior Research Methods, Instruments, and Computers 23 (2), 229-236. ↗
  8. Dumais, S. T., Furnas, G. W., Landauer, T. K., Deerwester, S. and Harshman, R. 1998. Using Latent Semantic Analysis to Improve Access to Textual Information. In CHI’88: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 281-285. ACM Press. find
  9. Frakes, W. B. and Baeza-Yates, R. 1992. Information Retrieval: Data Structures and Algorithms. Prentice-Hall: New Jersey. find
  10. Habash, N. 2004. Large Scale Lexeme Based Arabic Morphological Generation. In Proc. of Traitement Automatique du Langage Naturel. find
  11. Larkey, L., Ballesteros, L. and Connell, M. 2002. Improving Stemming for Arabic Information Retrieval: Light Stemming and Co-Occurrence Analysis. SIGIR 2002, Finland, pp. 275-282. find
  12. Larkey, L. and Connell, M. 2002. Arabic Information Retrieval at Umass in TREC-10. In Voorhees, E.M. and Harman, D.K. (eds.): The Tenth Text Retrieval Conference, TREC 2001 NIST Special Publication 500-250, pp. 562-570. find
  13. Lavie, A., Peterson, E., Probst, K., Wintner, S., and Ey-tani, Y. 2004. Rapid Prototyping of a Transfer-Based Hebrew-to-English Machine Translation System. In Proceedings of the TMI-04.
  14. Mathieu, B., Besançon, R. and Fluhr, C. 2004. Multilingual Document Clusters Discovery. Recherche d’Information Assistée par Ordinateur (RIAO) Proceedings, 1-10. find
  15. The TREC 2002 Arabic/English CLIR Track, NIST TREC 2002 Proceedings, pp. 16-26. find
  16. Peters, C. (ed.). 2001. Cross-Language Information Retrieval and Evaluation: Workshop of the Cross-Language Evaluation Forum, CLEF 2000. Berlin: Springer-Verlag. find
  17. Resnik, P., Olsen, M. B., and Diab, M. 1999. The Bible as a Parallel Corpus: Annotating the "Book of 2000 Tongues". Computers and the Humanities, 33, 129-153. find
  18. Van Rijsbergen, C. 1979. Information Retrieval (2 nd edition). Butterworth: London. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/chew-2008-effects-language-relatedness-multilingual-information/