PUD — what these data are and which layers exist
Gist. The English part of Parallel Universal Dependencies for CoNLL 2017: 1000 sentences (21,180 words) from news (n…) and Wikipedia (w…). 750 sentences are originally English (…01…). 250 were translated from German, French, Italian and Spanish (02–05); from English they were further translated into other languages. Each sentence has counterparts in the PUD of other languages (# parallel_id). Annotation history:
- Google annotated it following "universal" guidelines based on Stanford Dependencies.
- Martin Popel automatically converted it to UD v2 (Udapi,
ud.Google2ud). - Schuster, Reddy and Manning manually corrected UPOS and syntax.
- Lemmas and FEATS were assigned by CoreNLP, without checking.
Test only. License CC BY-SA 3.0.
Conditions and exceptions.
- The README changelog ends at 2.6 ("No changes"). Yet the 2.18 data have subtypes and features of later versions:
nmod:unmarked111,obl:unmarked37,nmod:desc41,advcl:relcl1,ExtPos76,PronType=Demon there (35 of 35). So the treebank received mass edits from the UD maintainers that are not recorded in the README. - DEPS — enhanced UD, automatic since 2.2, unchecked. 7 empty nodes.
- 5 sentences still contain annotator notes
# Checktree: …(Not sure about this one, Ellipsis), and one more# notes.
Examples. n01002042 The new spending is fueled by Clinton’s large bank account.; n01085008 Fast forward to 2016 and this is increasingly worthy of attention.
In UD.
| layer | state |
|---|---|
| XPOS | PTB, manual; no ADD, LS, NFP |
| FEATS | CoreNLP: agreement by form, no Voice, no Case=Gen (PUD — FEATS from CoreNLP: agreement by form, no Voice, no Case=Gen, no Mood=Imp; by without obl:agent) |
| lemmas | CoreNLP, unchecked |
| DEPS | present |
| MWT | 129 |
Our annotator on PUD: UPOS 94.42, XPOS 94.07, UFeats 87.55, LAS 77.02, MLAS 58.04 (en — Mova's English graph compressor). UPOS and XPOS are almost as on EWT, UFeats is 8 points lower — CoreNLP customs.
Sources. README UD_English-PUD (from Google's README); index.md; Zeman et al. 2017, K17-3001 (papers/z/ze/zeman-2017-conll-2017-shared-task-multilingual); McDonald et al. 2013 (Google universal treebanks, ACL 2013).
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/treebanks/en/pud/seeds/data/