GUM — what these data are and which layers exist
Gist. Georgetown University Multilayer: a multilayer corpus. It is extended every year by students of the LING-4427 course; texts are taken from openly licensed sources. In UD 2.18 — 14,353 sentences and 256,739 words from 16 main genres: academic, biographies, conversations, fiction, interviews, news, textbooks, travel guides, Reddit (separately, ../../gumreddit), wikiHow, court, essays, letters, podcasts, etc. Manual annotation:
- PTB tags and lemmas — manual;
- UPOS and FEATS — automatic from PTB and the graph;
- dependencies up to GUM v5 — Stanford Dependencies converted by DepEdit rules taking the entity and coreference layers into account; since v6 — directly UD;
- then everything was corrected manually.
Conditions and exceptions.
- Split: train 11,314 / dev 1575 / test 1464 sentences.
- License CC BY-NC-SA 4.0: for
movatest and research only, not training commercial models (en — Mova's English graph compressor). - GUM is converging with EWT deliberately. From the changelog:
- 2021: MWT, hyphen splitting, HYPH;
- 2022:
advcl:relcl,nsubj:outer, features as in EWT; - 2024:
:unmarked,ExtPos; - 2025:
nmod:desc,dep→parataxisfor footnotes.
Examples. GUM_bio_padalecki-2 Jared Tristan Padalecki (born July 19, 1982) [1] is an American actor.; GUM_conversation_grounded-5 You guys are always in trouble.
In UD. All layers, as in EWT, and richer:
| layer | state |
|---|---|
| lemmas, XPOS | manual; XPOS without the EWT tags ADD, NFP, AFX (GUM — XPOS without the EWT web tags: ADD, NFP, AFX) |
| FEATS | from PTB and the graph; «_» in 82,308 words (EWT — 79,476) |
| DEPS | enhanced UD, 154 empty nodes |
| MISC | Entity (coreference), Discourse (eRST), MSeg (morphemes), Cxn, XML, PDTB; «_» only in 97,819 words |
| MWT | 4455 |
Our annotator (EWT + ESLSpok train) on the GUM test: UPOS 94.26, UFeats 90.84, LAS 76.85, MLAS 60.08 (en — Mova's English graph compressor). UFeats is 4.8 points lower than on EWT — the customs from GUM — you has Number=Sing or Plur, and the verb agrees with it and GUM — Degree=Pos on invariable adverbs, and such without Degree are at work here. Adding GUM train to training gave EWT LAS +0.8, PUD +3.1, GUM +4.2.
Sources. README UD_English-GUM (Summary, Changelog); index.md; Zeldes 2017, LRE 51(3); Zeldes & Schneider 2023, 2023.udw-1.7 (papers/z/ze/zeldes-2023-are-ud-treebanks-getting-more); Peng & Zeldes 2018, W18-4918; digest the UD survey digest «english-banks», §1.5.
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/treebanks/en/gum/seeds/data/