IAHLT version of the UD Hebrew Treebank (IAHLT-HTB)
This is a revised fork of the Universal Dependencies version of the Hebrew Treebank, with some important changes and a consistency overhaul involving substantial manual corrections. The dataset was prepared as part of the Hebrew & Arabic Corpus Linguistics Infrastructure project at the Israeli Association of Human Language Technologies (IAHLT).
Before using this data it is highly recommended to read the IAHLT treebanking documentation (coming soon!) and reading the general principles outlined below. This dataset is currently still a work in progress.
Universal Dependencies - Hebrew Dependency Treebank (v2) https://github.com/UniversalDependencies/UD_Hebrew
This version of the HTB data follows the following principles:
- Aim for convergence with other UD languages, and in particular Arabic
- No addition of segments without corresponding characters in the text. This means that as in Arabic, there are no inserted של in clitic possessives, no את in clitic objects
- Representation of articles following prepositions using a morphological definiteness feature, rather than pseudo tokens
- Removal of strictly predictable dependency subtypes (e.g. only use
case
rather thancase:gen
andcase:acc
for של and את, removemark:q
) - Multiword token strings should be the sum of their consituent token texts, allowing for simpler segmentation models (highly constrained search space) - thus forms like לה comprise two tokens ל and ה (as in UD Arabic)
- Clear separation of universal POS tags based on function (ADV for
advmod
, SCONJ formark
, ADP forcase
etc.) - More consistent treatment of proper nouns as a gateway for downstream NER (all tokens inside named entities which would otherwise be NOUN are PROPN)
- More constrained lists of closed class items such as fixed expressions
- Rework auxiliaries, reanalyzing impersonal modals as predicates with clausal subjects
- Introduce previously unused labels with their usual functions (e.g.
expl
) - Introduce commons subtypes for passive subjects (
nsubj:pass
,csubj:pass
) - Unification of FEATs with IAHLT UD treebanks (Wiki, Knesset)
- Changed :tmod and :npmod subtypes to :unmarked with TemporalNPAdjunct=Yes in misc to preserve tmod info, following a UD Core Group decision in July 2024
- More generally, producing valid (non-legacy) UD data in line with the current Universal Dependencies validator tool (available here)
V1 of the dependency corpus was built by semi-automatic conversion of the Hebrew Constituency Treebank (v2) by MILA.
V2, refered to below as UD-HTB, was converted from V1, using a combination of automatic conversion when possible, and manual conversion and verification in other cases (see papers below).
This version is currently refered to as IAHLT-HTB.
This directory contains a corpus of sentences annotated using Universal Dependencies annotation.
The corpus comprises 115K tokens (158K words) and 6,216 sentences, taken from the Ha'aretz
newspaper.
The trees were manually annotated into phrase-structure trees, and then semi-automatically converted
into Universal Dependencies.
This file is compatible with the CoNLL-U format defined for Universal Dependencies. See: http://universaldependencies.github.io/docs/format.html . However, at present the files do not include lemmas for words. These may be added in a later release.
The dependency taxonomy can be found on the Universal Dependencies web site:
http://universaldependencies.github.io/docs/
http://universaldependencies.github.io/docs/#language-he
The Train/Dev/Test split follows previous splits of the underlying Treebank, namely: sentences 1-484 dev (~10K tokens), 485-5725 train (~127K tokens), 5726-6216 test (~11K tokens).
Some parts of the structure are more reliable than others. In particular, words with a "morphological feature" entry of HebSource=ConvUncertainHead or HebSource=ConvUncertainLabel indicate that the head (label) information for this token is based on unreliable information.
To help improve the corpus, please alert us to any errors you find in it; For underlying issues in the source data (UD-HTB) contact Yoav Goldberg at yoav.goldberg@gmail.com or Reut Tsarfaty at reut.tsarfaty@gmail.com
For issues specific to the IAHLT-HTB version, please contact Amir Zeldes at amir-zeldes@georgetown.edu
The Universal Dependencies Hebrew Treebank created by: (in alphabetic order):
- Yoav Goldberg
- Reut Tsarfaty
Revised IAHLT version:
- Amir Zeldes
The following people were also involved in the creation of v2:
- Amir More (adding Lemmas, detokenization, v1->v2 conversion)
- Yuval Pinter (documentation)
- Shoval Sadde (documentation, v2 validation and conversion)
- Victoria Basmov (v2 validation and conversion)
The Universal Dependencies Hebrew Treebank is based on the Hebrew Constituency Treebank (v2) developed by MILA, The Knowledge Center for Processing Hebrew. (http://www.mila.cs.technion.ac.il/resources_treebank.html)
You are encouraged to cite these papers reflecting the original source treebank if you use the Hebrew Universal Dependencies Treebank:
@inproceedings{tsarfaty2013unified,
title={A Unified Morpho-Syntactic Scheme of Stanford Dependencies},
author={Tsarfaty, Reut},
booktitle={Proc. of ACL},
year={2013}
}
@inproceedings{mcdonald2013universal,
title={Universal Dependency Annotation for Multilingual Parsing},
author={McDonald, Ryan T and Nivre, Joakim and Quirmbach-Brundage, Yvonne and Goldberg, Yoav and Das, Dipanjan and Ganchev, Kuzman and Hall, Keith B and Petrov, Slav and Zhang, Hao and T{\"a}ckstr{\"o}m, Oscar and others},
booktitle={Proc. of ACL},
year={2013}
}
Note that these papers do not accurately reflect the current annotation in the Treebank. A more up-to-date publication discussing the IAHLT scheme and tokenization is:
@InProceedings{ZeldesHowellOrdanBenMoshe2022,
author = {Amir Zeldes and Nick Howell and Noam Ordan and Yifat Ben Moshe},
booktitle = {Proceedings of {EMNLP} 2022},
title = {A Second Wave of {UD} {H}ebrew Treebanking and Cross-Domain Parsing},
year = {2022},
pages = {4331--4344},
address = {Abu Dhabi, UAE},
url = {https://aclanthology.org/2022.emnlp-main.292/},
}