Graph Rewriting for Enhanced Universal Dependencies

What it is about

The authors enter the IWPT 2021 shared task with a rule-based system: UDPipe produces the basic UD trees, and the GREW graph-rewriting tool turns them into enhanced UD. The rules cover six enhancement types, applied in an ordered strategy, and language packages add rules for specific treebanks. The raw results are far below the best systems, and the paper says so. The useful part is the error analysis: on gold UD, 589 French and 815 English discrepancies with the gold enhanced graphs were found, and the first 100 of each were classified. Many come from errors in the gold enhanced layer or in the gold UD, and some from the language-specific CASE relation, which the authors say should not be universal. The order of the rule groups is not neutral: subject propagation must run before the enhancement that depends on it.

Ideas

• I1. Rule order as a measured parameter in tools/treeagree (src/rules.rs): permute the order of the rule groups and count the edges that change on EWT dev; the paper shows the order is not neutral, so the count says how fragile the layer is.
• I2. Their manual protocol on our data: take the first 100 discrepancies between our enhanced layer and the gold enhanced layer on EWT test (test only), and classify each as gold enhanced error, gold UD error or our error; report the shares per release.
• I3. Left-dependent propagation for English coordination, in en/src/convert.rs: propagate the left dependents of the first conjunct by dependency type, as the paper does; measure F on enhanced conj and nsubj edges on EWT dev.

For Mova

The rule layer would sit next to tools/treeagree (rules R1 to R8, gates for projectivity and one root), and the English data is EWT dev and test. The chapter is 24. Coordination and subordination, with 21. The simple clause for propagated subjects. Measure first the share of gold enhanced errors in the 100 sampled discrepancies, since it bounds what a rule layer can be scored against.

Topics: Universal Dependencies

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2021.iwpt-1.18

@inproceedings{guillaume2021graph,
  title = {{Graph Rewriting for Enhanced Universal Dependencies}},
  author = {Guillaume, Bruno and Perrier, Guy},
  booktitle = {Proceedings of the 17th International Conference on Parsing Technologies and the IWPT 2021 Shared Task on Parsing into Enhanced Universal Dependencies (IWPT 2021)},
  year = {2021},
  doi = {10.18653/v1/2021.iwpt-1.18},
  url = {https://doi.org/10.18653/v1/2021.iwpt-1.18}
}

References

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

1. Guillaume Bonfante, Bruno Guillaume, and Guy Perrier. 2018. Application of Graph Rewriting to Natural Language Processing, volume 1 of Logic, Linguistics and Computer Science Set. ISTE Wiley. find
2. Gosse Bouma, Djamé Seddah, and Daniel Zeman. 2020. Overview of the IWPT 2020 shared task on parsing into enhanced Universal Dependencies. In Proceedings of the 16th International Conference on Parsing Technologies and the IWPT 2020 Shared Task on Parsing into Enhanced Universal Dependencies, pages 1… find our card
3. Gosse Bouma, Djamé Seddah, and Daniel Zeman. 2021. From raw text to enhanced universal dependencies: find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/ud-2021-guillaume-graph-rewriting-enhanced-universal-dependencies/