Elamur Research⌕
☰
home / papers / 2014 / A Study of Association Measures and their Combination for Arabic MWT Extraction

A Study of Association Measures and their Combination for Arabic MWT Extraction

Paper · 2014 · Abdelkader El Mahdaouy, Said EL Alaoui Ouatik and Eric Gaussier

Summary

The paper addresses automatic extraction of multi-word terms (MWTs) from Arabic domain text, a language with little prior work on this task. The authors build a hybrid pipeline. First, a linguistic stage tags the corpus with the AMIRA toolkit (Diab's SVM-based POS tagger trained on the Penn Arabic TreeBank) and keeps sequences that match noun-centred syntactic patterns. This stage also merges graphical, inflectional, morpho-syntactic and syntactic variants using light stemming. Second, a statistical stage ranks the candidates with a new score, the NLC-value. This score takes the C-value, which measures termhood and corrects for nesting, and re-weights each candidate's frequency by its weakest LLR link, which measures unithood. It then blends the result with a context score (N value). The method was tested on an environmental corpus. Precision was checked first against AGROVOC and then, through translation, against IATE. NLC-value beat LLR, C-value, NC-value, NTC-value and LLR+C-value at every cut-off. The main lesson is that combining termhood, unithood and context gives cleaner term lists than any one signal used alone.

Key points

  • The corpus has 1666 environmental web documents from the "Al-Khat Alakhdar" site, with 53,569 distinct tokens after stop-word removal. Topics include pollution, water, soil, forests, climate and disasters.

  • The candidate patterns extend those of Boulaknadel et al. (2008) with two forms: Noun followed by Noun/ADJ sequences, and Noun Prep Noun. Both bi-grams and tri-grams are covered.

  • NLC-value is computed in three steps:

    • Frequency is multiplied by ln(2 + the minimum LLR over the adjacent word pairs of the candidate).
    • This re-weighted frequency is inserted into the C-value formula.
    • The result is mixed as 0.8·LC-value + 0.2·N value.

  • Precision at top-100/200/300 was:

    MeasureTop-100Top-200Top-300
    NLC-value82.0%75.5%73.0%
    NTC-value80.0%71.5%69.7%
    LLR75.0%70.5%64.3%
    C-value71.0%69.0%67.3%

  • Evaluation used stemmed matching against AGROVOC, which required the same number of stemmed words. Candidates that did not match were translated and looked up in IATE. In total, 1095 distinct terms were judged, and the measures shared 141 of them.

Interesting ideas

  • I1. Unithood for a candidate longer than two words is scored by its weakest adjacent pair, the minimum LLR, so one loose junction is enough to sink the whole phrase.
  • I2. The C-value subtracts the average frequency of the longer terms that contain a candidate, so fragments of a real term are not mistaken for terms themselves.
  • I3. Several weak signals are blended with fixed weights: termhood, unithood and context. Every combined measure beat each single-signal measure.
  • I4. The evaluation falls back in stages: a candidate is first matched against a domain thesaurus, and if that fails, it is translated and checked in a second, larger terminology base.
  • I5. The context score rewards a candidate whose neighbouring words themselves often appear next to other terms. In effect, terms tend to cluster near other terms.

For Mova

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

  • I1 (tokens): When deciding whether a span like "climate change policy" should become a single multiword lemma, Mova could require every internal link to be strong rather than relying on the average.
  • I2 (world): When compiling vocabularies or ontology entries from corpora, this nesting correction would stop "change policy" from being listed as a concept just because "climate change policy" is common.
  • I3 (learn): Silver-data filters could score candidates on several independent, inspectable criteria instead of one, while keeping each component visible for explanation.
  • I4 (learn): Automatically induced lexical items could be validated first against WordNet and then against Wikidata labels, which would give a cheap and transparent precision estimate without hand annotation.
  • I5 (world): When proposing new entities or concepts for the text's world, Mova could treat closeness to already-known entities as supporting evidence for promoting a candidate noun phrase.

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

🔬Laboratoire d'Informatique de GrenobleGrenoble · France21 works · 688 citations

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Abdelkader El Mahdaouy, Said EL Alaoui Ouatik and Eric Gaussier. 2014. A Study of Association Measures and their Combination for Arabic MWT Extraction. arXiv:1409.3005.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1409.3005 · arXiv 1409.3005

@misc{mahdaouy2014study,
  title = {{A Study of Association Measures and their Combination for Arabic MWT Extraction}},
  author = {Abdelkader {El Mahdaouy}, Said {EL Alaoui Ouatik} and Eric Gaussier},
  year = {2014},
  doi = {10.48550/arxiv.1409.3005},
  eprint = {1409.3005},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1409.3005}
}

References

19 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Al Khatib K, and Badarneh A. 2010. Automatic extraction of arabic multi-word terms. In Proceedings of the International Multiconference on Computer Science and Information Technology, pp. 411-418. find
  2. Al-Taani A, and Abu-Al-Rub S. 2009. A rule-based approach for tagging non-vocalized Arabic words, volume 6. The International Arab Journal of Information Technology, p 320. find
  3. Boulaknadel S, Daille B, and Aboutajdine D. 2008 a. Multi-word term indexing for Arabic document retrieval. In Proceedings of the The IEEE symposium on Computers and Communications, pp. 869-873. find
  4. Boulaknadel S, Daille B, and Aboutajdine D. 2008 b. A multi-word term extraction program for Arabic language. the 6th international Conference on Language Resources and Evaluation LREC.
  5. Bounhas I, and Slimani Y. 2009. A hybrid approach for Arabic multi-word term extinction. International Conference on Language Processing and Knowledge Engineering, pp. 1-8. find
  6. Bourigault D. 1994. LEXTER, un logiciel d’EXtraction de TERminologie, Application à l’acquisition des connaissances à partir de textes, phd thesis. Ecole des Hautes études en Sciences Sociales, Paris. find
  7. Church K, Gale W, Hanks P, and Hindle D. 1991. Using statistics in lexical analysis. In Lexical Acquisition, Exploiting On-Line Resources to Build a Lexicon. find
  8. Daille B. 1994. Approche mixte pour l’extraction de terminologie : statistique lexicale et filtres linguistiques, Phd thesis. University of Paris 7. find
  9. Diab M Hacioglu K, and Jurafski D. 2004. Automatic tagging of Arabic text : From raw text to base phrase chunks. In Proceedings of North American Association for Computational Linguistics NAACL, pp. 149-152.
  10. Diab M. 2009. Second Generation Tools (AMIRA 2.0) : Fast and Robust Tokenization, POS tagging, and Base Phrase Chunking. In Proceedings of International Conference on Arabic Language Resources and Tools. find
  11. Dunning T. 1994. Accurate Methods for the Statistics of Surprise and Coincidence, volume 19. Computational Linguistics, pp. 61-74.
  12. Frantzi K. T, Ananiadou S, and Tsujii T. 1998. The C-Value/NC-Value Method of Automatic Recognition for Multi-word terms. Journal on Research and Advanced Technology for Digital Libraries, pp. 115-130. find
  13. Kageura K, and Umino B. 1996. Methods of Automatic Term Recognition A Review, volume 3. Terminology. find
  14. Korkontzelos I, Ioannis P. Klapaftis, and Manandhar S. 2008. Reviewing and Evaluating Automatic Term Recognition Techniques. In Procedings of the 6th international Conference on Advances in Natural Language Processing, pp. 248-259. find
  15. Larkey S. Leah , Ballesteros L, and Connell E. Margaret. 2007. Light Stemming for Arabic Information Retrieval, volume 38. Text, Speech and Language Technology, pp. 221-243. find
  16. Tadic M, and Sojat K. 2008. Finding multiword term candidates in Croatian. In the Proceedings of IESL2003 Workshop, pp. 102-107. find
  17. Vu T, Aw A. Ti, and Zhang M. 2008. Term Extraction Through Unithood And Termhood Unification. In Procedings of IJCNLP. find
  18. Wen Z, Yoshida T, and Xijin T. 2007. Text classification using multi-word features. In Proceedings of the The IEEE symposium on Computers and Communications, pp. 3519-3524. find
  19. Wen Z, Yoshida T, and Xijin T. 2008. A Study on Multi-word Extraction form Chinese Documents. Advanced Web and Network Technologies, pp. 42-53. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mahdaouy-2014-study-association-measures-their-combination/