The Hebrew Universal Dependency Treebank: Past, Present and Future

What it is about

The Hebrew Treebank (HTB) grew from 500 newspaper sentences (2001) to 6221 sentences, through a chain of automatic and semi-automatic conversions: phrase structure, then Goldberg's functional-head dependencies, then Unified-Stanford, then UD v1 and UD v2. The authors checked the UD v2 version label by label on the dev set and corrected three kinds of problem: argument types, morphology mixed up with syntax, and language-specific subtypes that duplicate a universal label. The result, UDv2New, changes modal auxiliaries to aux (the following verb becomes xcomp), predicative adjectives to xcomp, prepositional indirect objects to obl, acl:inf to acl, conj:discourse to parataxis, and tokenization of inserted particles. Parsing the dev set with yap under gold morphology raised MLAS from 72.76 to 75.58 and BLEX from 47.76 to 50.16. With automatic morphology the new scheme gave no gain, and it helped only when the correct analysis was injected into the lattice. The authors read this as a lexical-coverage problem, not a scheme problem.

Ideas

• I1. Turn the revised rules into checks: "an obl follows a preposition", "no iobj with an ADP child", "no acl:inf", "an xcomp under a modal verb has an aux head". Each is a pattern rule with an error or warning level, as in the paper's grewv-style list. Run them on tools/hbo output.
• I2. The paper says the dative (experiencer, ethical, regular) is not separated by UD. Keep the discourse dative as a feature of the argument rather than as a core argument, and test the rule on a small list of Hebrew sentences from macula compare output.
• I3. A/B measure: parse one dev set with tools/hbo under both schemes, with gold and with automatic morphology, two identical runs each. The paper's pattern (gain only with gold or infused lattices) is the expected outcome; if it does not hold, that is a finding.
• I4. Modal verbs as a closed list: the paper's Verb-Type=Mod marking depends on a lexical class. A short table in the Hebrew verb notes, checked against the lexicon.

For Mova

I1 belongs in languages/he/11-ud-analysis.md, with the output check in tools/hbo. I2 belongs in languages/he/04-nouns-adjectives-pronouns.md and languages/he/06-prepositions-particles.md. I3 is a test for tools/hbo, with its result recorded in languages/he/12-mova.md. I4 belongs in languages/he/05-verbs.md.

Topics: Treebanks

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/w18-6016

@inproceedings{sade2018hebrew,
  title = {{The Hebrew Universal Dependency Treebank: Past, Present and Future}},
  author = {Sade, Shoval and Seker, Amit and Tsarfaty, Reut},
  booktitle = {Proceedings of the Second Workshop on Universal Dependencies (UDW 2018)},
  year = {2018},
  doi = {10.18653/v1/w18-6016},
  url = {https://doi.org/10.18653/v1/w18-6016}
}

References

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

1. Meni Adler. 2007. Hebrew Morphological Disambiguation: An Unsupervised Stochastic Word-based Approach. Ph.D. thesis, Ben-Gurion University of the Negev, Beer-Sheva, Israel. find
3. Meni Adler, Yael Dahan Netzer, Yoav Goldberg, David Gabay, and Michael Elhadad. 2008. Tagging a hebrew corpus: the case of participles. In LREC. Citeseer. find our card