Universal Dependencies - English Dependency Treebank Universal Dependencies English Web Treebank v2.17 -- 2025-11-15 https://github.com/UniversalDependencies/UD_English-EWT
A Gold Standard Universal Dependencies Corpus for English, built over the source material of the English Web Treebank LDC2012T13 (https://catalog.ldc.upenn.edu/LDC2012T13).
The corpus comprises 254,820 words and 16,622 sentences, taken from five genres of web media: weblogs, newsgroups, emails, reviews, and Yahoo! answers. See the LDC2012T13 documentation for more details on the sources of the sentences. The trees were automatically converted into Stanford Dependencies and then hand-corrected to Universal Dependencies. All the basic dependency annotations have been single-annotated, a limited portion of them have been double-annotated, and subsequent correction has been done to improve consistency. Other aspects of the treebank, such as Universal POS, features and enhanced dependencies, has mainly been done automatically, with very limited hand-correction initially (but substantial cleanup over the years).
Universal Dependencies English Web Treebank annotations © 2013-2021 by The Board of Trustees of The Leland Stanford Junior University. All Rights Reserved.
The annotations and database rights of the Universal Dependencies English Web Treebank are licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
You should have received a copy of the license along with this work. If not, see http://creativecommons.org/licenses/by-sa/4.0/.
The underlying texts come from various sources collected for the LDC English Web Treebank. Some parts are in the public domain. Portions may be © 2012 Google Inc., © 2011 Yahoo! Inc., © 2012 Trustees of the University of Pennsylvania and/or © other original authors.
This directory contains a corpus of sentences annotated using Universal Dependencies annotation. The corpus comprises 254,818 words and 16,622 sentences (see stats), taken from various web media including weblogs, newsgroups, emails, reviews, and Yahoo! answers; see the LDC2012T13 documentation for more details on the source of the sentences. The trees were automatically converted into Stanford Dependencies and then hand-corrected to Universal Dependencies. All the dependency annotations have been single-annotated, and a limited portion of them have been double-annotated with interannotator agreement at approximately 96%. The sentence IDs include the genre and the original LDC2012T13 filename.
This corpus is compatible with the CoNLL-U format defined for Universal Dependencies. See:
https://universaldependencies.org/format.html
The dependency taxonomy can be found on the Universal Dependencies web site:
http://www.universaldependencies.org
For the conversion to v2, we performed an automatic conversion with extensive spot-checking, and manual adjudication of ambiguous cases.
Most enhanced dependencies were automatically obtained by running an adapted version of the converter by Schuster and Manning (2016). These dependencies have not been manually checked. Enhanced dependencies for reduced relative clauses were added in v2.14.
MISC, the 10th column of the .conllu format, allows for extension beyond the components generally adopted in UD treebanks. EWT contains three categories of MISC annotation beyond the core UD standard: UCxn (construction) annotations, STREUSLE (lexical semantic) annotations, and idiosyncratic attributes.
The Cxn attribute allows for description of constructions that go beyond the words
and relations described by the UD standard. For the most part, the Cxn annotations
follow the parameters of the UCxn project.
EWT also includes relative clause subtype annotations in this field.
See the changelog entry for v2.14.
The STREUSLE corpus provides gold-standard
annotations for multiword expressions and supersenses in the Reviews portion of EWT
(3814 sentences). In v2.17 these were incorporated in the MISC column via the attributes
MWECat, MWELemma, MWELen, MWEString, Supersense, and PRel.
Refer to the STREUSLE docs
for details.
These are not documented as part of UCxn, STREUSLE, or the general UD guidelines, but are experimental and specific to EWT. These are not necessarily systematically applied, and are subject to change in future versions:
Citation=Yes: Indicates a parenthetical citation (see docs issue 1145)Depictive=Yes: Indicates the secondary predicate is semantically depictive.Exclamative=Yes: Clause bearing exclamative syntax (to disambiguate from interrogatives).Footnote=Yes: Indicates a reference to a footnote (see docs issue 1145)FlatType: Explains the use offlatfor a special kind of entity. Marked on the first word of the flat expression. Current values:Enumerated: a numbered entity, e.g. Chapter 1FilenameIconic: e.g. spelled-out wordsNumericLocator: a complex section number in a text, e.g. 5.1 ( b )Postcodee.g. M5J 1S9Phone: a telephone number
ManuallyChecked: A note for treebank maintainersMentioned=Yes: A metalinguistic mention (as opposed to a use) of a term.MissingWordAfter=Yes: Indicates that an unspecified word would need to be inserted to make the sentence grammatical.MissingWordsAfter: Value is a specific word or words that can be inferred as having been accidentally omitted.Promoted=Yes: Makes explicit that a word has been promoted to the head of the phrase, explaining its apparently anomalous context. NOTE: Applied in an ad hoc fashion; many cases of promotion are not marked.Resultative=Yes: Indicates the secondary predicate is semantically resultative.Signature=Yes: Indicates a trailing phrase within a sentence interpreted as the author's signature. Only applied to 1 token.SpecialEncoding=Yes: Indicates an unusual character encoding (see issue 83).Superlocation=Yes: Indicates thenmod:unmarkedphrase is a containing-location (or organization) of a more specific place or subunit (see issue 588). Most commonly used for "CITY, STATE" combinations like Dallas, Texas. Others include "HOUSE_NUMBER STREET", "ORGANIZATION BRANCH_LOCATION", and "JOB_TITLE, ORGANIZATION". Note that major fields of full postal addresses sequenced into a "sentence" attach aslist, so they will not bear this attribute.TemporalNPAdjunct=Yes: Indicates an adjunct previously annotatedobl:tmodornmod:tmod. May become obsolete if entity annotations are introduced.TODO: A note for treebank maintainers
The issue tracker at https://github.com/UniversalDependencies/UD_English-EWT/issues documents many yet-to-be-resolved analysis challenges. Significant among these:
- Many free relatives are incorrectly analyzed as interrogative.
2025-11-15 v2.17
Highlights:
- In the Reviews sentences, incorporate gold-standard lexical semantic annotations from STREUSLE (#606)
- Implement new guidelines for several classes of nominal expressions
- Implement
PronType=Demfor expletive "there" (docs#517) - Hundreds of "it" and "you" tokens had
Case=NomwhereCase=Accis correct (#589) - "@": at/SYM/ExtPos=ADP (#599)
- Correct
appostonmod:unmarkedfor "you guys" etc. (#436) - Temporal premodifiers of nouns should attach as
compoundnotnmod:unmarked(#593) - Document
MISCannotations
2025-05-15 v2.16
Highlights:
- Implement new
nmod:descsubtype for prefixes/suffixes/embellishments in names (#561, #559, #59) - Implement new guidelines for dates (#575)
- Implement new guidelines for numbered entities like "Chapter 1" (#558)
- New policy for "you guys" and similar (#436)
- Reanalyze some constructions to comply with newly enforced requirements on
detdependents ("at least", #553; predeterminer "such", docs#1114) - Consistent treatment for "else" (#556), "rather than" (#562)
- Clean up postnominal
advmods (#557) - Clean up tokenization of "#" and "@" in web text (#577)
- Clean up various phone numbers, addresses, and filenames
- Clean up lemmas where the word is in all-caps for stylistic reasons (#560)
2024-11-15 v2.15
Highlights:
- Merge subtypes
nmod:{npmod,tmod}asnmod:unmarkedandobl:{npmod,tmod}asobl:unmarked(docs#1028):unmarkedindicates a modifier that is structured as an NP without case marking- Retain temporal semantics with new custom MISC feature
TemporalNPAdjunct=Yes
- Many foreign names that were
compoundcorrected toflat(#81) - Fix a number of cases of spurious nonprojectivity (#545, #548)
- Construction annotations in the UCxn framework added to MISC (#551)
- This release adds rule-based annotations of Interrogatives, Conditionals, Existentials, and NPN (noun-preposition-noun) constructions on the head of the respective phrase, plus construction elements.
- The UCxn v1 notation and categories are documented here.
- Special thanks: @LeonieWeissweiler, @WesScivetti, @s-herrera
- Features
- Implement
ExtPosfor all fixed expressions (docs#1037) - Implement
Polarity=Negfor not/PART, neither/CCONJ, nor/CCONJ, no/INTJ andPolarity=Posfor yes/INTJ (#526, docs#1056) - Implement
PronTypefor none/PRON and ADVs "now", "never", "somewhere", "whenever", and similar NumFormandNumTypefor decades expressed as pluralized years (#527)- Correct overuses of
VerbForm=Inf(#284) - Improve feature consistency for
ADJs - Add custom MISC features
FlatType=FilenameandFlatType=Phone(and assign theseExtPos=PROPN)
- Implement
2024-05-15 v2.14
Highlights:
- Relative clauses
- ~700 enhanced edges added for reduced relative clauses (#392) (thanks @xiulinyang!)
- relative clause types added in
Cxnattribute of MISC (#474)- e.g.
Cxn=rc-wh-nsubj:pass(passive subject WH),Cxn=rc-red-obj(reduced object),Cxn=rc-red-obl-pstrand(reduced oblique with preposition stranding),Cxn=rc-free-obj_xcomp(free relative, object nested under xcomp) - produced by not-to-release/tools/rc-types.sh - see the code for documentation
- cf. the UCxn project (a larger effort which proposed construction annotation in UD, though this use of it for relative clauses is EWT-specific at present)
- e.g.
- Attach list item enumerators (
LS) asdiscourse(#518) - Verb tags/features
- Noun features: implement
Number=Ptanfor pluralia tantum (docs#999) - Improved treatment of internet addresses (#440, #487)
- Multiword expressions
goeswithfor spaced email addressesflatfor spaced telephone numbers
2023-11-15 v2.13
Highlights:
- Structural
- Adopt subtype
obl:agentfor passive by-phrases, and apply the featureVoice=Passmore widely to distinguish passive uses from active perfect uses of the past participle form (#290) - Remove use of
flat:foreignin line with most other English treebanks (#459) - Implement new guidelines on sufficiency/excess constructions (#423)
- Use Udapi to standardize
punctattachments (thanks to @martin-popel)
- Adopt subtype
- Features
- Use AUX (not VERB) for "have" and "do" when stranded due to ellipsis (#403)
- Mark verbal contractions missing apostrophes as typos (flagged by @rhdunn: #443)
- Extensive cleanup of/based on UPOS
2023-05-15 v2.12
Highlights:
- Implement new policy on sole
iobj(#55) - Cleanup of free relatives (part of #278)
- Make use of
xcompmore consistent - For "etc.", change
Number=SingtoNumber=Plur
2022-11-15 v2.11
Highlights:
- Implement
:outerper multiple subjects policy (#310) - Implement
advcl:relcl(#346) - Implement revised guidelines for English pronouns (lemmas, features) (issue)
- Revise WH-adverbs to attach in a subordinate clause as
advmodnotmark(#88) - Improved lemmas/features for numeric values/entities
- Tag "etc." as
NOUN(#353) - Add neaten.py, which implements English-specific validation rules
2022-05-15 v2.10
Highlights:
- Fixed all validation errors
- Implement new
goeswithpolicy (#314) - Use
parataxisfor "X so Y" and similar (#313)
2021-11-15 v2.9
- Fixed all validation errors
- Many other improvements to annotation of assorted words and constructions
2021-05-15 v2.8
- Fixed many wrong lemmata, POS tags, and relations
- Reanalyzed many dependencies to conform to UD validation
- Retagged words in names to limit PROPN to true nouns
- Fixed certain metadata issues
- Fixed some typos, added CorrectForms
- Added missing multiword tokens for clitics/contractions
- Added Style=Expr feature for expressive spellings
- Fixed many incorrectly non-projective graphs, reanalyzing as projective
- Reannotated list items as NUM not X
2020-11-15 v2.7
- Added multiword tokens where appropriate for contracted verb forms
- Fixed some wrong lemmata and POS tags
2020-05-15 v2.6
- Added paragraph boundaries
- Fixed some wrong lemmata and POS tags
- Fixed directionality of some
goeswithdependencies
2019-11-15 v2.5
- Fixed miscellaneous syntactic issues
- Fixed CoNLL-U syntax error
2019-05-15 v2.4
- Fixed some wrong lemmata and POS tags
- Fixed miscellaneous syntactic issues
- Fixed some punctuation attachments
- Fixed malformed enhanced graphs
2018-11-15 v2.3
- Fixed several lemmata
2018-04-15 v2.2
- Repository renamed from UD_English to UD_English-EWT
- Automatically added enhanced dependencies (These have not been manually checked!)
- Fixed some wrong lemmata and POS tags
- Fixed miscellaneous syntactic issues
2017-11-15 v2.1
- Fixed some wrong lemmata, POS tags
- Fixed miscellaneous syntactic issues
- Added basic dependencies into the
DEPScolumn according to the CONLL-U v2 format
2017-02-15 v2.0
- Updated treebank to conform to v2 guidelines
- Fixed some wrong lemmata
- Fixed miscellaneous syntactic issues
- Added empty nodes for gapped constructions in enhanced representation
2016-11-15 v1.4
- Changed POS tag of fused det-noun pronouns (e.g., "somebody", "nothing")
to
PRON - Added original, untokenized sentences to CoNLL-U files
- Fixed some POS errors, features and wrong lemmata
- Fixed miscellaneous syntactic issues in a few sentences
2016-05-15 v1.3
- Improved mapping of
WDTto UPOS - Corrected lemma of "n't" to "not"
- Fixed some errors between
advcl,ccompandparataxis - Fixed inconsistent analyses of sentences repeated between dev and train sets
- Fixed miscellaneous syntactic issues in a few sentences
2015-11-15 v1.2
- Bugfix: removed _NFP suffix from some lemmas
- Fixed date annotations to adopt UD standard
- Remove escaping of ( and ) from word tokens (XPOSTAGs are still
-LRB-and-RRB-) - Improved precision of
xcomprelation - Improved recall of
namerelation - Corrected lemmas for reduced auxiliaries
- Corrected UPOS tags of pronominal uses of this/that/these/those (from
DETtoPRON) - Corrected UPOS tags of subordinating conjunctions (from
ADPtoSCONJ) - Corrected UPOS tags of some main verbs (from
AUXtoVERB)
To help improve the corpus, please alert us to any errors you find in it. The best way to do this is to file a github issue at:
https://github.com/UniversalDependencies/UD_English-EWT/issues
We also welcome pull requests. If you want to make edits, please modify the
trees in the individual files in the not-to-release/sources directory instead
of making direct changes to en_ewt-ud-{dev,test,train}.conllu.
Annotation of the Universal Dependencies English Web Treebank was carried out by (in order of size of contribution):
- Natalia Silveira
- Timothy Dozat
- Sebastian Schuster
- Miriam Connor
- Marie-Catherine de Marneffe
- Nathan Schneider
- Ethan Chi
- Samuel Bowman
- Christopher Manning
- Hanzhi Zhu
- Daniel Galbraith
- John Bauer
Creation of the CoNLL-U files, including calculating UPOS, feature, and lemma information was primarily done by
- Sebastian Schuster
- Natalia Silveira
The construction of the Universal Dependencies English Web Treebank was partially funded by a gift from Google, Inc., which we gratefully acknowledge.
You are encouraged to cite this paper if you use the Universal Dependencies English Web Treebank:
@inproceedings{silveira14gold,
year = {2014},
author = {Natalia Silveira and Timothy Dozat and Marie-Catherine de
Marneffe and Samuel Bowman and Miriam Connor and John Bauer and
Christopher D. Manning},
title = {A Gold Standard Dependency Corpus for {E}nglish},
booktitle = {Proceedings of the Ninth International Conference on Language
Resources and Evaluation (LREC-2014)}
}
=== Machine-readable metadata (DO NOT REMOVE!) ================================
Data available since: UD v1.0
License: CC BY-SA 4.0
Includes text: yes
Parallel: no
Genre: blog social reviews email web
Lemmas: automatic with corrections
UPOS: converted with corrections
XPOS: manual native
Features: converted with corrections
Relations: manual native
Contributors: Silveira, Natalia; Dozat, Timothy; Manning, Christopher; Schuster, Sebastian; Chi, Ethan; Bauer, John; Connor, Miriam; de Marneffe, Marie-Catherine; Schneider, Nathan; Bowman, Sam; Zhu, Hanzhi; Galbraith, Daniel; Bauer, John
Contributing: here source
Contact: syntacticdependencies@lists.stanford.edu
===============================================================================
