An OpenNMT Model to Arabic Broken Plurals

What it is about

Arabic broken plurals change the consonant-vowel template of the singular instead of adding a suffix, so the plural cannot be built by concatenation. The author treats the singular-to-plural mapping as a translation problem and trains an OpenNMT sequence-to-sequence LSTM on about 2,560 tokens of singular and plural forms with pattern labels, taken from a lexical profiling resource (Attia et al., 2011). Runs of 10, 20 and 30 epochs, with and without shuffling, reach validation accuracies between about 50% and 62%. The model learns the template skeleton and the vowel melody, but it often gets the consonants wrong, especially gutturals and emphatics, and its predictions are dominated by the ʔa- and ma- templates. Seven templates cover 68.73% of the corpus, which the author links to the errors.

Ideas

• I1. Split the template into a consonant tier and a vowel tier, as the paper itself proposes. The vowel melody is learnt and the consonants are not, so test each tier separately. For Mova, a rule-based pattern module in languages/hbo/04-noun.md can fix the vowel melody from a table and take the consonants from the root, then both tiers are scored on MACULA noun lists.
• I2. Report a "most frequent template" baseline next to any pattern predictor. The paper's errors come from the frequent templates, so a predictor that does not beat the majority template has learnt nothing. Put this baseline in every pattern test in tools/macula.
• I3. A guttural and emphatic test. The paper finds these the hardest to predict, as L2 learners do. For Hebrew, check whether a rule handles the guttural letters (א, ה, ח, ע) correctly, for example the furtive patah and the composite shewa, in languages/hbo/03-phonology.md.

For Mova

I1 and I3 go into languages/hbo/03-phonology.md and languages/hbo/04-noun.md. I2 is a baseline for every pattern test in tools/macula, and for the Aramaic noun patterns in languages/arc/nouns.md. The neural model is not reused; the tier split and the baseline are.

Cite

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

@inproceedings{issa2018opennmt,
  title = {{An OpenNMT Model to Arabic Broken Plurals}},
  author = {Issa, Elsayed},
  booktitle = {Proceedings of the First International Workshop on Language Cognition and Computational Models},
  year = {2018},
  url = {https://aclanthology.org/W18-4103/}
}

References

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

1. Attia, M., Pecina, P., Tounsi, L., Toral, A., van Genabith, J. (2011). Lexical Profiling for Arabic. Electronic Lexicography in the 21st Century. Bled, Slovenia. Retrieved from: https://sourceforge.net/projects/broken-plurals/ ↗
2. Brustad, K., Al-Batal, M., Al-Tonsi, A. (2011). Al-Kitaab fii Ta’allum al-‘Arabiyya. Part one. 3 Georgetown University Press. rd edition. USA: Hammond, M. (1988) Templatic Transfer in Arabic Broken Plurals. Natural Language and Linguistic Theory. (6) 247- 270. find
3. Hammond, M. (2018). Neural Nets for Phonology and Morphology. R codes are retrieved from https://faculty.sbs.arizona.edu/hammond/ling696b-sp18/ ↗
4. Klein, G., Y. Kim, Deng, Y. Y., Senellart, J. & Rush, A.M. (2017) OpenNMT: Open-source Toolkit for Neural Machine Translation. ArXiv e-prints1701.02810. find our card
5. McCarthy, J. (1982). “A Prosodic Account of the Arabic Broken Plurals.” Current Trends in African Linguistics, 1.25. Retrieved from https://scholarworks.umass.edu/linguist_faculty_pubs/25 ↗
6. McCarthy, J. & Prince A. (1990) "Prosodic Morphology and Templatic Morphology." In Mushira Eid and John McCarthy (eds.) Perspectives on Arabic Linguistics II: Papers from the Second Symposium on Arabic Linguistics. Amsterdam: John Benjamins. 1-54. find
7. Plunkett, K. & Nakisa, R., C. (1997) A Connectionist Model of the Arabic Plural System. Language and Cognitive Processes, 12:5-6, 807-836, DOI: 10.1080/016909697386691. ↗