Towards an adequate account of parataxis in Universal Dependencies

What it is about

The UD relation parataxis is general, and it is hard to tell apart from conj and appos. The paper shows how annotators have used it, with the parallel PUD treebanks as data. The English PUD has 97 parataxis instances in 1000 sentences: 44 side-by-side sentences, 42 reported speech, 8 parentheticals and 3 other cases. Across the 14 PUD treebanks that have more than ten instances, the counts vary a lot, even though the translations should share structure. Swedish keeps 86 of the 97 English instances but adds many where English uses hypotactic relations. The paper proposes narrower guidelines, for example forbidding parataxis for argument-like cases unless a word is elided.

Ideas

I1. en ud-check: flag parataxis arcs and sort them into the paper's subtypes (side-by-side, reported speech, parenthetical). Anything outside these types is reported. Measure the counts per subtype on EWT and on the English PUD, which is for tests only.
I2. en/src/parse.rs: build a confusion table of parataxis against conj and appos on EWT dev. Report the precision of the parser's parataxis arcs before changing any feature.
I3. en/src/udreg.rs: add the paper's subtypes as a stricter check in the UD registry gate, so new parataxis labels need a subtype. Count how many current EWT arcs would fail the gate.

For Mova

The chapters are languages/en/24-coordination-subordination.md (conj against parataxis) and languages/en/29-ellipsis-word-order-information-structure.md (elided material). The code is en/src/parse.rs, en/src/udreg.rs and en ud-check. Measure first: the subtype counts of parataxis on EWT test, against the current LAS of 79.8.

Library summary

The Research library's own summary of the paper, in our words — not the paper's text.

Written by Mova (claude-opus-5-5) from the paper text. Original — aclanthology.org/W19-8011.pdf

Summary

The article examines the parataxis relation in Universal Dependencies. In UD 2.0 it is defined very broadly, so annotators often cannot clearly distinguish it from coordination (conj), apposition (appos) or subordinate relations such as ccomp. Moreover, the subtypes given in the guidelines have little in common with one another. The author tests how consistently this label is assigned in practice. The material is the parallel PUD treebanks: each contains 1000 sentences, translated mostly from English. Because of this, their annotation should almost coincide. The English PUD is compared with thirteen others, and the discrepancies turned out to be large both in the number of parataxis instances and in where exactly it is assigned. Some of the differences can be explained by translation, but most are caused by differing annotator practice. The author analyzes typical contentious cases: asyndetic sentences, direct speech, parenthetical phrases, repetitions and series of phrases in fiction. For these, he proposes refining the guidelines. This work matters for cross-lingual consistency of treebanks and for the quality of parsers trained on them.

Key points

The English PUD has 97 instances of parataxis: 44 are sentences placed side by side, 42 are direct speech, 8 are parenthetical constructions, and 3 are "other". There are no tag questions at all.
The author examines 14 PUD treebanks in which parataxis occurs more than 10 times. The frequencies vary widely: from 20 in the Czech and 24 in the Arabic to 195 in the Russian.
Similarity to the English annotation is highest for the Portuguese PUD (0.79). It is lowest for the Czech (0.17) and Arabic (0.24). The Swedish PUD carried over 86 of the 97 English instances but added quite a few of its own where the English has hypotaxis.
For direct speech, the German PUD consistently uses ccomp and makes the speech verb the head. The French one keeps parataxis, but with the direction of the relation reversed.
The author's proposals are as follows. Keep a single general parataxis and introduce subtypes for particular genres or corpora, as in UD_Italian-Postwita and UD_French-Spoken. Abandon list. Establish the principle that hypotactic relations take precedence, with exceptions for direct speech and asyndetic series with the same semantic role. Allow the argument "a missing word can be inserted here" only for ellipsis.

Topics: Universal Dependencies · Treebanks

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/w19-8011

@inproceedings{ahrenberg2019adequate,
  title = {{Towards an adequate account of parataxis in Universal Dependencies}},
  author = {Ahrenberg, Lars},
  booktitle = {Proceedings of the Third Workshop on Universal Dependencies (UDW, SyntaxFest 2019)},
  year = {2019},
  doi = {10.18653/v1/w19-8011},
  url = {https://doi.org/10.18653/v1/w19-8011}
}

References

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

Marie-Catherine de Marneffe, Miriam Connor, Natalia Silveira, Samuel R. Bowman, Timothy Dozat and Christopher D. Manning. 2013. More constructions, more genres: Extending Stanford dependencies. Proceedings of the 13th International Conference on Dependency Linguistics. find our card
Marie-Catherine de Marneffe, Timothy Dozat, Natalia Silveira, Katri Haverinen, Filip Ginter, Joakim Nivre and Christopher D. Manning. 2014. Universal Stanford Dependencies: A cross-linguistic typology. Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14).… find our card
Kim Gerdes and Sylvain Kahane. 2017. Trois schémas d’annotation syntaxique en dépendance pour un même corpus de français oral: le cas de la macrosyntaxe. Actes de la 24e conférence sur le traitement automatique des langues (TALN), Atelier sur les corpus annotés du français (ACor4French), Orléans. find
M. A. K. Halliday 1985. An Introduction to Functional Grammar. Edward Arnold. find
P. H. Matthews 1981. Syntax. Cambridge University Press. find
Timothy Osborne and Kim Gerdes. 2019. The status of function words in dependency grammar: A critique of Universal Dependencies (UD). Glossa: a journal of general linguistics 4(1): 17. 1–28. DOI: https://doi.org/10.5334/gigl.537. ↗
Manuela Sanguinetti, Cristina Bosco, Alberto Lavelli, Alessandro Mazzei, Oronzo Antonelli and Fabio Tamburini. PoSTWITA-UD: an Italian Twitter Treebank in Universal Dependencies. Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC-2018). https://www.aclweb… ↗
Jennette Winterson 1997. Gut symmetries. Granta Books, London. find
Daniel Zeman, Martin Popel, Milan Straka, ..., and Josie Li 2017. find
CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies. Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies, Vancouver, Canada, 3–4 August 2017. Association for Computational Linguistics. find our card