GBI Treebanks as a Resource for New Applications

What it is about

The Global Bible Initiative's account of its treebanks, the ancestors of the MACULA Hebrew data Mova converts. Since 2005 GBI has built Hebrew OT and Greek NT trees by parsing a morphologically tagged text with a hand-written grammar, having scholars correct the output verse by verse, storing every correction as disambiguation data, and reparsing until the scholars were satisfied. Besides the syntactic Hebrew treebank there is a prosodic one built from the accents alone. The trees feed phrase-level interlinears, concordances and translation memories, "smart search" for similar words in similar relations, and a tree-based word aligner that, by aligning bottom-up subtree by subtree with statistical scores, cut alignment errors by 40–50% over a purely statistical aligner on 18 translations in 7 languages. Added layers: antecedents of pronouns and of implied subjects, verb frames, semantic roles, extended Strong's numbers, glosses, sense numbers.

Ideas

• I1. Align the Hebrew trees with JPS 1917 subtree by subtree (bottom-up, keeping the best few candidates per node, with dictionary glosses and Strong's numbers as anchors) instead of word by word; measure alignment against a hand-checked sample of verses.
• I2. Use MACULA's frames (MISC Frame, which also resolve zero subjects to an earlier noun) as gold for participant tracking: score prag's speaker/addressee carrying and any hbo coreference step against them. Chapter 13 lists frames as "carried, not yet used by a tool".
• I3. A prosodic tree from the accents alone (the GBI prosodic treebank idea): build the dichotomy tree of each verse from disjunctive accents and compare its splits with our syntactic clause boundaries; disagreement rate is a cheap check on both.
• I4. Consistency flag for translations: if the same multiword subtree (lemma pattern) is glossed differently in two places, flag it; apply to the Torah dictionary's examples and the course's word-for-word layer.

For Mova

I1 serves the course's word-for-word layer and the dictionary examples (13. Biblical Hebrew in Mova: the tools §9–10). I2 is the first use of Frame in tools/macula/src/prag.rs and 09-discourse.md (participant tracking). I3 belongs to 02-cantillation.md (accents as syntax) and is a new macula check. I4 is a check in macula dict.

Topics: Coreference resolution · Treebanks

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.7146/hn.v5i2.142738

@inproceedings{wu2019gbi,
  title = {{GBI Treebanks as a Resource for New Applications}},
  author = {Wu, Andi},
  booktitle = {HIPHIL Novum 5},
  year = {2019},
  doi = {10.7146/hn.v5i2.142738},
  url = {https://doi.org/10.7146/hn.v5i2.142738}
}

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/hbo-2019-wu-gbi-treebanks-resource-new-applications/