Elamur Research⌕
☰
home / papers / 2016 / An Empirical Study of Arabic Formulaic Sequence Extraction Methods

An Empirical Study of Arabic Formulaic Sequence Extraction Methods

Summary

The paper looks for the best statistical way to pull frequent, semantically transparent multi-word formulas (formulaic sequences, FSs) for Modern Standard Arabic out of a very large corpus. The candidates are not limited to any particular syntactic pattern or topic. The node words come from a new lemma-based MSA wordlist of 5,189 items. That list was built by adding up average reduced frequency (ARF) values across four big corpora, which together hold more than 8 billion tokens. The authors took a sample of 50 frequent and 50 rare node words and used Sketch Engine on the ArTenTen corpus to rank their collocates with six association measures (AMs). Each ranked list was checked against a gold-standard FS list that had been validated by hand. Scoring used n-best precision, averaged per word and then across words (MAP). For the frequent words, MI.log_f, MI and logDice performed best, and T-score and MI3 performed worst. This matters because Arabic NLP and language teaching lack corpus-derived formula lexicons, and the result shows which ranking statistic is worth using to build one.

Key points

  • The six AMs were computed in Sketch Engine: T-score, MI, MI3, logDice, MI.log_f and log-likelihood. Extraction ran on ArTenTen, an MSA corpus of more than 7.4 billion words.
  • Candidate settings were a minimum frequency of 10 per million and a co-occurrence window of 2 to 4 words. The wordlist was lemmatized and morphologically analysed with MadaAmira.
  • Evaluation steps:
    • compute precision for each node word (attested FSs divided by all extracted FSs);
    • compute average precision (AP) for each AM per word;
    • compute MAP across all node words.
  • On the high-frequency set:
    • MI.log_f and MI scored above 0.85 MAP;
    • T-score and MI3 stayed below 0.50;
    • log-likelihood fell in the weaker group, but the authors still call it a promising option.
  • The finding agrees with Alrabiah et al. (2014), where MI.log_f also came first on Classical Arabic. It contrasts with earlier Arabic studies that preferred log-likelihood.

Interesting ideas

  • I1. Testing the same measures separately on frequent and rare node words shows how much a method's quality depends on word frequency, which a single pooled score would hide. — For Mova (learn): Evaluation reports for parsers and self-induced repair rules could be split into frequent-lemma and rare-lemma bands, so gains on common words do not mask failures on the long tail.
  • I2. Combining several corpora by summing a dispersion-aware frequency (ARF) per lemma gives a frequency list that is less skewed by any one source or by bursty documents. — For Mova (world): Lemma priors in the compiled knowledge layer could be built by summing ARF-style counts over several English corpora, which would give stable "common word" signals for ontology and verb-class lookups.
  • I3. Ranked candidate lists are scored per query item with average precision and then averaged as MAP against a hand-checked gold list, which rewards putting good candidates near the top. — For Mova (learn): Ranked outputs such as candidate VerbNet frames or coreference antecedents could be scored the same way, instead of only checking whether the top-1 answer is correct.
  • I4. Weighting MI by log frequency (MI.log_f) keeps MI's sensitivity to tight word pairs while reducing its bias toward rare pairs, and in this study it worked best for finding meaningful fixed pairs. — For Mova (tree): Mining English bigrams with MI.log_f could produce a lexicon of fixed expressions such as "in order to", which could help the parser and the silver-data checks assign fixed and compound relations consistently.

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: Linguistics & languages

Topics: Semitic languages

Institutions

🎓University of LeedsLeeds · United Kingdom19 works · 1 290 citations🎓Umm al-Qura UniversityMakkah al Mukarramah · Saudi Arabia7 works · 87 citations

Cite

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

@inproceedings{alghamdi2016empirical,
  title = {{An Empirical Study of Arabic Formulaic Sequence Extraction Methods}},
  author = {Alghamdi, Ayman and Atwell, Eric and Brierley, Claire},
  booktitle = {Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16)},
  year = {2016},
  doi = {10.63317/2cnf696e366n},
  url = {https://doi.org/10.63317/2cnf696e366n}
}

References

19 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Alghamdi, A., Atwell, E. and Brierley, C. (2015) Constructing a corpus-informed Listing of Arabic formulaic sequences ArFSs for language pedagogy and technology Unpublished paper. find
  2. Alrabiah, M., Alhelewh, N., Al-Salman, A. and Atwell, E. (2014) An empirical study on the Holy Quran based on a large classical Arabic corpus. International Journal of Computational Linguistics (IJCL). 5(1), pp.1-13. find
  3. Boulaknadel, S., Daille, B. and Aboutajdine, D. (2008) A Multi-Word Term Extraction Program for Arabic Language, the 6th international Conference on Language Resources and Evaluation LREC 2008, Marrakech, Morocco, pp. 1485-1488, 2008. find
  4. Church, K., Gale, W., Hanks, P. and Kindle, D. (1991) Using statistics in lexical analysis. In Lexical Acquisition: Using On-line Resources to Builda Lexicon, pages 115–164. Lawrence Erlbaum. find
  5. Church, K.W. and Hanks, P. (1990) Word association norms, mutual information, and lexicography. Computational linguistics. 16(1), pp.22-29. find our card
  6. Church, Kenneth W. (2000) Empirical estimates of adaptation: The chance of two Noriegas is closer to p=2 than p2. In Proceedings of COLING 2000, pages 173– 179, Saarbrücken, Germany. find
  7. Daille, B. (1994) Approche mixte pour l’extraction automatique de terminologie: statistiques lexicales et filtres linguistiques. Ph.D. thesis, Université Paris 7. find
  8. Dunning, T. 1993. Accurate methods for the statistics of surprise and coincidence. Computational linguistics. 19(1), pp.61-74. find our card
  9. Durrant, P.L. (2008) High frequency collocations and second language learning. Ph.D. thesis University of Nottingham. find
  10. Kilgarriff, A., Baisa, V., Bušta, J., Jakubíček, M., Kovář, V., Michelfeit, J., Rychlý, P. and Suchomel, V. (2014) The Sketch Engine: ten years on. Lexicography. 1(1), pp.7-36. find
  11. Mel’ˇcuk, I. (1998) Collocations and lexical functions. In: Cowie AP (ed) Phraseology. Theory,Analysis, and Applications, Claredon Press, Oxford, pp 23–53. find
  12. Mel’cuk, I. (2003) Collocations: définition, rôle et utilité. In: Grossmann F, Tutin A (eds) Les collocations: analyse et traitement, Editions De Werelt, Amsterdam, pp 23–32. find
  13. Pecina, P. (2005) An extensive empirical study of collocation extraction methods. In: Proceedings of the ACL Student Research Workshop: Association for Computational Linguistics, pp.13-18. find
  14. Rychlý, P. (2008) A lexicographer-friendly association score. Proceedings of Recent Advances in Slavonic Natural Language Processing, RASLAN. pp.6-9. find our card
  15. Saif, A.M. and Aziz, M.J. (2011). An automatic collocation extraction from Arabic corpus. Journal of Computer Science. 7(1), pp.6-11. find
  16. Schmitt, N. (2010) Researching vocabulary: a vocabulary research manual. Basingstoke: Palgrave Macmillan. find
  17. Sinclair, J. (1991) Corpus, concordance, collocation. Oxford University Press. Oxford. find
  18. Smadja, F. (1993) Retrieving collocations from text: Xtract. Computational linguistics. 19(1), pp.143-177. find our card
  19. Yonatan Belinkov, Nizar Habash, Adam Kilgarriff, Noam Ordan, Ryan Roth, and Vit Suchomel. (2013). arTenTen: a new, vast corpus for Arabic. In Eric Atwell and Andrew Hardie, editors, Proceedings of the Second Workshop on Arabic Corpus Linguistics (WACL2) http://www.comp.leeds.ac.uk/eric/wacl/wacl2proceedings.pdf ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alghamdi-2016-empirical-study-arabic-formulaic-sequence/