CATiB: The Columbia Arabic Treebank

What it is about

CATiB is an Arabic dependency treebank built for speed, for training parsers. It uses 6 part-of-speech tags (against more than 400 in the Penn Arabic Treebank) and 8 dependency labels, with terms from the traditional Arabic grammar (idafa, tamyiz). Case is left out, because it follows from syntax, and empty categories and co-indexation are not annotated. The pipeline is MADA&TOKAN for tokens and tags, a MaltParser model trained on a converted PATB part, and manual correction in TrEd. Inter-annotator agreement is about 91.5–91.7% for attachment and 88.6–88.8% for labeled attachment on the PATB sample, and lower on new text (89.2% and 85.0%). Annotators reach 540 tokens per hour, against 250–300 for PATB. A second annotation pass cut labeled-attachment errors by 20–50% on two documents at a cost of about 30% of speed. The release has about 1M tokens (31,319 trees), of which about 273K are new annotation.

Ideas

• I1. A minimal label set. CATiB shows that a parser can be trained on a small set of labels (subject, object, modifier, possessive, specifier, topic, predicate, flat). Map Mova's own labels onto such a set and measure parser accuracy on the same trees with the full and the reduced set. This is a direct test of the abstract-annotation idea.
• I2. A second pass as a measurement. The paper runs serial annotation on the hardest and the easiest document and reports the error change. For Mova, run the rule checker and the parser over the same trees twice and count the labels that change; a stable result is a sign that the rule is well defined.
• I3. Report attachment and labeled attachment side by side. In the paper labels are harder to agree on than attachments (LABATT about 3 points below ATT). Mova's evaluation of gparse should show unlabeled and labeled scores as separate numbers, so that a label rule cannot hide behind a good attachment score.

For Mova

I1 is a label mapping for treebanks/ar and for the Hebrew treebanks (treebanks/hbo/ptnk, treebanks/he/), checked with gparse. I2 is a check in tools/macula compare. I3 is the report format for gparse evaluation. The idafa label corresponds to the chain of Syntax: the clause and the sentence.

Topics: Treebanks

Cite

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

@inproceedings{habash2009catib,
  title = {{CATiB: The Columbia Arabic Treebank}},
  author = {Habash, Nizar and Roth, Ryan},
  booktitle = {Proceedings of the ACL-IJCNLP 2009 Conference Short Papers},
  year = {2009},
  url = {https://aclanthology.org/P09-2056/}
}

References

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

1. N. Habash, R. Faraj and R. Roth. 2009. Syntactic Annotation in the Columbia Arabic Treebank. In Conference on Arabic Language Resources and Tools, Cairo, Egypt. find
2. N. Habash and O. Rambow. 2005. Arabic Tokenization, Part-of-Speech Tagging and Morphological Disambiguation in One Fell Swoop. In ACL’05, Ann Arbor, Michigan. find our card
3. N. Habash, R. Gabbard, O. Rambow, S. Kulick, and M. Marcus. 2007a. Determining case in Arabic: Learning complex linguistic behavior requires complex linguistic features. In EMNLP’07, Prague, Czech Republic. find
4. N. Habash, A. Soudi, and T. Buckwalter. 2007b. On Arabic Transliteration. In A. van den Bosch and A. Soudi, editors, Arabic Computational Morphology. Springer. find
5. S. Kulick, R. Gabbard, and M. Marcus. 2006. Parsing the Arabic Treebank: Analysis and Improvements. In Treebanks and Linguistic Theories Conference, Prague, Czech Republic. find
6. M. Maamouri, A. Bies, and T. Buckwalter. 2004. The Penn Arabic Treebank: Building a large-scale annotated Arabic corpus. In Conference on Arabic Language Resources and Tools, Cairo, Egypt. find
7. M. Maamouri, A. Bies and S. Kulick. 2008. Enhancing the Arabic treebank: a collaborative effort toward new annotation guidelines. In LREC’08, Marrakech, Morocco. find
8. J. Nivre, J. Hall, J. Nilsson, A. Chanev, G. Eryigit, S. Kubler, S. Marinov, and E. Marsi. 2007. MaltParser: A language-independent system for data-driven dependency parsing. Natural Language Engineering, 13(2):95–135. find
9. O. Smrž and J. Hajič. 2006. The Other Arabic Treebank: Prague Dependencies and Functions. In Ali Farghaly, editor, Arabic Computational Linguistics. CSLI Publications. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-2009-habash-catib-columbia-arabic-treebank/