Expanding the Universal Dependencies Ancient Hebrew Treebank with Constituency Data

What it is about

The author extends the 2022 pipeline for the Ancient Hebrew UD treebank so that it also reads MACULA, a source with full clause-level constituency trees that leaves the relations between clauses open. The BHSA's text-level feature (narrative, quotation, quotation inside quotation) fills that gap. A rule-extraction script turns each constituency node into a head-and-dependents arc, using the head that MACULA marks and the rule name as the label; a set of rules then maps rule names to UD relations, with disambiguation from BHSA features, and a file of manual overrides fixes the remaining cases. Exodus was checked tree by tree: 810 of 1,033 unvalidated trees (78.4%) were correct without edits, and the first-pass accuracy stayed near 80% across genres. The treebank now has 5,500 sentences and 145K words, and the project reached 84% sentence accuracy by its end.

Ideas

I1. Convert constituency to dependencies with the head marks the source already gives: each node's marked head becomes the parent of its siblings, and the rule name becomes the label. Check Mova's tools/macula conversion against this rule and against the gold trees (macula compare); if the head marks are not used there, the conversion will be wrong on coordinated phrases.
I2. Replace a fixed list of coordination rule names (one per length, such as NPNPaNP) with a pattern: a noun phrase followed by any number of conjunction-and-noun-phrase pairs. This removes a list that grows with every new book, and makes the rule table in tools/macula smaller.
I3. Quotation tokenisation: the Hebrew quotation marker le'mor is split as a preposition and an infinitive in all cases except this one, and the paper marks the two pieces as a fixed expression with a feature so that the validator accepts it. Mova can write the same rule as a fixed-expression entry in its particle table.

For Mova

I1 is a check on tools/macula (conversion and macula compare), and the first-pass rate of 78.4% is the number to beat for any new book. I2 goes into the coordination rules of tools/macula. I3 goes into Pronouns, particles and prepositions (the quotation marker) and Discourse: narrative, speech and the shape of a text (direct speech).

Topics: Universal Dependencies · Treebanks

Cite

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

@inproceedings{swanson2025expanding,
  title = {{Expanding the Universal Dependencies Ancient Hebrew Treebank with Constituency Data}},
  author = {Swanson, Daniel G.},
  booktitle = {Proceedings of the 23rd International Workshop on Treebanks and Linguistic Theories (TLT, SyntaxFest 2025)},
  year = {2025},
  url = {https://aclanthology.org/2025.tlt-1.3/}
}

References

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

Daniel Swanson and Francis Tyers. 2022. A Universal Dependencies treebank of Ancient Hebrew. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2353–2361, Marseille, France. European Language Resources Association. find our card
Þórunn Arnardóttir, Hinrik Hafsteinsson, Einar Freyr Sigurðsson, Kristín Bjarnadóttir, Anton Karl Ingason, Hildur Jónsdóttir, and Steinþór Steingrímsson. 2020. A Universal Dependencies conversion pipeline for a Penn-format constituency treebank. In Proceedings of the Fourth Workshop on Universal Dependencies (UDW 2020), pages 16–25, Barcelona, Spain (Online). Association for Computational Linguistics. find
Randall K Barry. 1997. ALA-LC romanization tables-transliteration schemes for non-roman scripts. In Library of Congress, 1997. find
Eckhard Bick and Tino Didriksen. 2015. CG-3 — beyond classical constraint grammar. In Proceedings of the 20th Nordic Conference of Computational Linguistics (NODALIDA 2015), pages 31–39, Vilnius, Lithuania. Linköping University Electronic Press, Sweden. find
Jayeol Chun, Na-Rae Han, Jena D. Hwang, and Jinho D. Choi. 2018. Building Universal Dependency treebanks in Korean. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). find
Clear Bible. 2022. datasets. MACULA Hebrew linguistic Aslı Kuzgun, Oğuz Kerem Yıldız, Neslihan Cesur, Büşra Marşan, Arife Betül Yenice, Ezgi Sanıyar, Oguzhan Kuyrukçu, Bilge Nas Arıcan, and Olcay Taner Yıldız. 2021. From constituency to UD-style dependency: Building the first conversion tool find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/ud-2025-swanson-expanding-universal-dependencies-ancient-hebrew/