Elamur Research
☰
home / papers / 2005 / Arabic Tokenization, Part-of-Speech Tagging and Morphological Disambiguation in One Fell Swoop

Arabic Tokenization, Part-of-Speech Tagging and Morphological Disambiguation in One Fell Swoop

What it is about

Tokenization, POS tagging and morphological disambiguation are handled as one step for undiacritized Arabic. A morphological analyzer (ALMORGEANA, built on the Buckwalter lexicon) first lists every analysis of each word, about 2 on average in the Penn Arabic Treebank. Then one SVM classifier (Yamcha) is trained for each of ten morphological features, and a choice is made among the analyses. Five ways of choosing are compared: majority vote, confidence-weighted vote, two mixtures of the two, and a rule learner (Ripper) over the classifier outputs. On the first test set the majority vote scores 96.6% per token against a unigram baseline of 92.1%. On the second set the rule learner is best (94.8%). Word-level tokenization reaches 99.3% on the first set. Reduced-tagset POS accuracy is 98.1%, and 97.6% on the Penn tagset. The authors report these as the best published figures for tokenization and tagging of natural undiacritized input, and note that unanalysed words (mostly names) are a small problem.

Ideas

For Mova

I1 and I3 go into tools/hbo, in the step after the candidate segmentations. I2 is a second chooser to measure against I1 on the same split. The per-word tokenization metric (I3) is also a test for languages/hbo/05-pronouns-particles.md, which must say how clitics such as the conjunction and the pronominal suffixes are separated. Arabic itself is not a Mova language yet; the transferable part is the method, not the Arabic features.

Library summary

The Research library's own summary of the paper, in our words — not the paper's text.

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/P05-1071.pdf

Summary

Arabic has very rich inflection and is usually written without short vowels, so a single word form often allows several analyses. With hundreds of thousands of possible full morphological tags, English-style taggers suffer badly from data sparseness. Habash and Rambow handle tokenization, POS tagging and full morphological disambiguation as one task. The ALMORGEANA analyzer, built on the Buckwalter lexicon, first lists every candidate analysis for each word. Ten SVM classifiers trained with Yamcha then predict individual features such as POS, clitic presence, gender, number, person, voice and aspect. Finally, a selection step uses those predictions to pick one analysis from the analyzer's candidates. Training and evaluation use ATB1 and ATB2 from the Penn Arabic Treebank. Both tokenization and tagging reach accuracy in the high nineties. Under the Penn Treebank tagset, POS accuracy beats the earlier Diab et al. (2004) system, which used no analyzer. The work supports Hajič's argument that a morphological analyzer matters more for inflection-heavy languages than additional annotated data.

Key points

  • About 333,000 fully specified tags are theoretically possible, but only about 2,200 appear in the first 280,000 words of the ATB. Each ATB word form has roughly 2 analyses on average.
  • Data split: ATB1 and ATB2 each provide about 120,000 training words (TR1, TR2) and about 12,000 words each for development and test (TE1, TE2). Results are reported on both test sets because the two releases differ in source and preparation.
  • POS accuracy on TE1 with the Penn Treebank tagset is 97.6%, against 95.5% reported by Diab et al. (2004). With the authors' reduced tagset it rises to 98.1%. On TE2 the corresponding figures are 95.7% and 96.5%.
  • Tokenization error is 0.7%, compared with 2.9% for Lee et al. (2003). Among the methods for choosing an analysis, the Maj chooser worked best on TE1.
  • A manual audit of 400 randomly chosen converted training words found 8 POS differences from the ATB. Seven were plausible consequences of missing entries in the Buckwalter lexicon.

Interesting ideas

  • Constrain the output space with an analyzer. Instead of letting a learner produce a tag freely, independent per-feature classifiers act as votes for one of the analyzer's legal analyses. Any rule-based analyzer, including one for Ukrainian, could serve this way to keep predictions linguistically consistent.
  • Make tokenization a byproduct. Once full morphology is disambiguated, the clitic and affix information supports any reasonable tokenization scheme. A single model can therefore serve several downstream segmentation conventions.
  • Back off for unknown words. When a word, typically a proper noun, gets no analysis, the system uses the affix inventories and their allowed combinations to hypothesize possible stems. "Berlusconi" receives 41 candidates, and the correct one is among them. This keeps out-of-vocabulary words inside the same pipeline.
  • Prefer a small POS set plus orthogonal features. The authors argue against tagsets borrowed from English, which encode English distinctions and drop Arabic ones such as gender on nouns. A reduced POS inventory combined with separate feature slots fits the language better.
  • Audit data conversion by hand. Manually checking a random sample after converting treebank annotations into a new representation is a cheap way to measure how much noise the conversion added.

Topics: Part-of-speech tagging

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.3115/1219840.1219911

@inproceedings{habash2005arabic,
  title = {{Arabic Tokenization, Part-of-Speech Tagging and Morphological Disambiguation in One Fell Swoop}},
  author = {Habash, Nizar and Rambow, Owen},
  booktitle = {Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL'05)},
  year = {2005},
  doi = {10.3115/1219840.1219911},
  url = {https://doi.org/10.3115/1219840.1219911}
}

References

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

  1. Imad A. Al-Sughaiyer and Ibrahim A. Al-Kharashi. 2004. Arabic morphological analysis techniques: A comprehensive survey. Journal of the American Society for Information Science and Technology, 55(3):189–213. find
  2. Tim Buckwalter. 2002. Buckwalter Arabic Morphological Analyzer Version 1.0. Linguistic Data Consortium, University of Pennsylvania, 2002. LDC Catalog No.: LDC2002L49. find
  3. William Cohen. 1996. Learning trees and rules with set-valued features. In Fourteenth Conference of the American Association of Artificial Intelligence. AAAI. find
  4. Kareem Darwish. 2003. Building a shallow Arabic morphological analyser in one day. In ACL02 Workshop on Computational Approaches to Semitic Languages, Philadelpia, PA. Association for Computational Linguistics. find our card
  5. Mona Diab, Kadri Hacioglu, and Daniel Jurafsky. 2004. Automatic tagging of arabic text: From raw text to base phrase chunks. In 5th Meeting of the North American Chapter of the Association for Computational Linguistics/Human Language Technologies Conference (HLT-NAACL04), Boston, MA. find
  6. Nizar Habash. 2005. Arabic morphological representations for machine translation. In Abdelhadi Soudi, Antal van den Bosch, and Guenter Neumann, editors, Arabic Computational Morphology: Knowledge-based and Empirical Methods, Text, Speech, and Language Technology. Kluwer/Springer. in press. find
  7. Jan Hajič. 2000. Morphological tagging: Data vs. dictionaries. In 1st Meeting of the North American Chapter of the Association for Computational Linguistics (NAACL’00), Seattle, WA. find
  8. Taku Kudo and Yuji Matsumoto. 2003. Fast methods for kernel-based text analysis. In 41st Meeting of the Association for Computational Linguistics (ACL’03), Sapporo, Japan. find
  9. Young-Suk Lee, Kishore Papineni, Salim Roukos, Ossama Emam, and Hany Hassan. 2003. Language model based Arabic word segmentation. In 41st Meeting of the Association for Computational Linguistics (ACL’03), pages 399–406, Sapporo, Japan. find
  10. Mohamed Maamouri, Ann Bies, and Tim Buckwalter. 2004. The penn arabic treebank : Building a large-scale annotated arabic corpus. In NEMLAR Conference on Arabic Language Resources and Tools, Cairo, Egypt. find
  11. Monica Rogati, J. Scott McCarley, and Yiming Yang. 2003. Unsupervised learning of arabic stemming using a parallel corpus. In 41st Meeting of the Association for Computational Linguistics (ACL’03), pages 391–398, Sapporo, Japan. find