home edit page issue tracker

Migration Guidelines

Introduction

Changes Affecting the Tree Structure

This section lists the changes to the represenation that affect the structure of dependency trees.

Treatment of Copular Constructions

The first difference between between Universal Dependencies and Stanford Dependencies is the treatment of copular constructions. While only nominal and adjectival predicates were treated as heads of copular constructions in the Stanford Dependencies represtnation, we also treat prepositional and adverbial predicates as the head of copular constructions in the Universal Dependencies represenation.

ROOT He is in the house .
cop(house, is)
root(ROOT, house)

ROOT She is here .
cop(here, is)
root(ROOT, here)

ROOT She is a lawyer .
cop(lawyer, is)
root(ROOT, lawyer)

Note that this change is currently only partially implemented in the Stanford Parser. In copular constructions with an adverbial predicate the copula verb is still treated as the head. The change will be completely implemented in future versions of the parser.

Treatment of Prepositional Phrases

Another major difference between Universal Dependencies and Stanford Dependencies is the treatment of prepositional phrases. In the Stanford Dependencies representation the preposition is always being treated the head of the prepositional phrase while in the Universal Dependencies represenation the head of a prepositional phrase is always a content word – either a noun phrase or the main predicate of a clause introduced by a preposition. The preposition itself is attached to the head of the prepositional phrase with the case case or a marker relation.

If the prepositional phrase only consists of a preposition followed by a noun phrase we use the nmod relation between the head of the prepositional phrase and the noun or predicate which it modifies.

Examples:

The office of the chair .
case(chair-5, of-3)
nmod(office-2, chair-5)

There is a ball on the table .
case(table-7, on-5)
nmod(is-2, table-7)

If the prepositional phrase consists of a preposition followed by a clause we either use the acl or advcl relation between the head of the prepositional phrase and the noun or predicate it modifies. acl is being used whenever the prepositional phrase modifies a noun unless it is a predicate in a copular construction. advcl is being used whenever the prepositional phrase modifies a verb, adjective or nominal predicate in a copular construction. The preposition is attached as a marker.

Examples:

The accident happened as the night was falling .
mark(falling, as)
advcl(happened, falling)

In this example the prepositional phrase modifies a verb, namely happened, and therefore we use advcl.

I am just very nervous after watching the movie .
mark(watching, after)
advcl(nervous, watching)

In this example the prepositional phrase modifies an adjectival predicate, namely nervous, and therefore we use advcl.

We have no information on whether users are experiencing problems .
mark(experiencing, on)
acl(information, experiencing)

In this example the prepositional phrase modifies a noun phrase, namely information, and therefore we use acl.

The war was a long way from being over .
cop(way, was)
mark(over, from)
advcl(war, over)

In this example the prepositional phrase modifies a nominal predicate in a copular construction, namely war, and therefore we use advcl. Also note that the prepositional phrase does also contain a copular construction and therefore the adverbial predicate over is the head of the prepositional phrase.

Multi-Word Expressions

There are two differences regarding the treatment of multi-word expressions in Universal Dependencies as compared to Stanford Dependencies. First, multi-word expressions are now always head-inital and all other words in the expression depend on the first word. Second, the list of expressions which are being treated as multi-word expressions changed. See the description of the mwe relation for a complete list of multi-word expressions in UD.

Mapping of Stanford Dependencies to Universal Dependencies

The following list provides for each relation in the original Stanford Dependencies representation the name of the equivalent relation in the Universal Dependencies representation. For some relations there doesn’t exist a 1:1 mapping. In these cases you’ll find a description of how different instances of the relation are treated in the UD representation. The list also provides the name of the GrammaticalRelation object as defined in Stanford CoreNLP and the Stanford Parser. Relations that were renamed, folded into other relations or changed otherwise are highlighted in red.

acomp: xcomp

acomp no longer exists as a relation in UD and has been folded into xcomp.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.XCLAUSAL_COMPLEMENT

advcl: advcl

advcl relations in SD are still annotated as advcl in UD.

Note, however, that this relation is now also used for prepositional phrases modifying a predicate (some instances of pcomp and prepc in SD) and for reduced non-finite verbal modifiers that modify a predicate (some instances of vmod in SD).

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.ADV_CLAUSE_MODIFIER

advmod: advmod

advmod relations in SD are still annotated as advmod in UD.

Note, however, that this relation is now also used for quantifier phrase modifiers (quantmod in SD).

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.ADVERBIAL_MODIFIER

agent: nmod:agent

This relation has been renamed to nmod:agent.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.AGENT

amod: amod

amod relations in SD are still annotated as amod in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.ADJECTIVAL_MODIFIER

appos: appos

appos relations in SD are still annotated as appos in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.APPOSITIONAL_MODIFIER

aux: aux

aux relations in SD are still annotated as aux in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.AUX_MODIFIER

auxpass: auxpass

auxpass relations in SD are still annotated as auxpass in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.AUX_PASSIVE_MODIFIER

cc: cc

cc relations in SD are still annotated as cc in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.COORDINATION

ccomp: ccomp

ccomp relations in SD are still annotated as ccomp in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.CLAUSAL_COMPLEMENT

conj: conj

conj relations in SD are still annotated as conj in UD.

Note that in the enhanced representation (formerly known as CCprocessed representation), the relation name and the coordinating conjunction are now separated by : instead of _, e.g. conj:and instead of conj_and.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.CONJUNCT

cop: cop

cop relations in SD are still annotated as cop in UD.

Unlike in the SD representation, prepositional phrases and adverbial phrases are also treated as the head of copular constructions in UD as illustrated in the following examples:

ROOT She is here
cop(here, is)
root(ROOT, here)

ROOT He is in the house
cop(house, is)
root(ROOT, house)

Note that this change is currently only partially implemented in the Stanford Parser. In copular constructions with an adverbial predicate the copula verb is still treated as the head. The change will be completely implemented in future versions of the parser.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.COPULA

csubj: csubj

csubj relations in SD are still annotated as csubj in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.CLAUSAL_SUBJECT

csubjpass: csubjpass

csubjpass relations in SD are still annotated as csubjpass in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.CLAUSAL_PASSIVE_SUBJECT

dep: dep

dep relations in SD are still annotated as dep in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.GrammaticalRelation.DEPENDENT

discourse: discourse

discourse relations in SD are still annotated as discourse in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.DISCOURSE_ELEMENT

dobj: dobj

dobj relations in SD are still annotated as dobj in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.DIRECT_OBJECT

expl: expl

expl relations in SD are still annotated as expl in UD.

According to the UD guidelines the pleonastic it should also have the expl relation. This is the case in the manually annotated treebank but the Stanford Parser currently does not treat instances of the pleonastic it differently.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.EXPLETIVE

goeswith: goeswith

goeswith relations in SD are still annotated as goeswith in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.GOES_WITH

iobj: iobj

iobj relations in SD are still annotated as iobj in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.INDIRECT_OBJECT

mark: mark

mark relations in SD are still annotated as mark in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.MARKER

mwe: mwe

mwe relations in SD are still annotated as mwe in UD.

However, the list of expressions which are considered multi-word expressions has changed. See the description of the mwe relation for the list of multi-word expressions.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.MULTI_WORD_EXPRESSION

neg: neg

neg relations in SD are still annotated as neg in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.NEGATION_MODIFIER

nn: compound

The nn relation in SD has been renamed to compound in UD.

This relation is now also used for elements of compound numbers instead of the number relation in SD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.COMPOUND_MODIFIER

npadvmod: nmod:npmod

The npadvmod relation in SD has been renamed to nmod:npmod in UD.

GrammaticalRelation object:

`edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.NP_ADVERBIAL_MODIFIER

nsubj: nsubj

nsubj relations in SD are still annotated as nsubj in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.NOMINAL_SUBJECT

nsubjpass: nsubjpass

nsubjpass relations in SD are still annotated as nsubjpass in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.NOMINAL_PASSIVE_SUBJECT

num: nummod

The num relation in SD has been renamed to nummod in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.NUMERIC_MODIFIER

number: compound

The number relation in SD no longer exists as a relation in UD and has been folded into compound.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.COMPOUND

parataxis: parataxis

parataxis relations in SD are still annotated as parataxis in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.NOMINAL_PASSIVE_SUBJECT

poss: nmod:poss

The poss relation in SD has been renamed to nmod:poss in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.POSSESSION_MODIFIER

possesive: case

The possesive relation in SD no longer exists as a relation in UD and has been folded into case.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.CASE_MARKER

pcomp: acl or advcl

Prepositional phrases are treated differently in UD and therefore the pcomp relation no longer exists. See the section Treatment of Prepositional Phrases for more details on when to use acl or advcl.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.CLAUSAL_MODIFIER (acl)

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.ADV_CLAUSE_MODIFIER (advcl)

pobj: nmod

The pobj relation in SD has been renamed to nmod in UD.

Note that prepositional phrases are treated differently in UD and not only the name of this relation but also the tree structure changed. See the section Treatment of Prepositional Phrases for more details on the new treatment of prepositional phrases.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.NOMINAL_MODIFIER

preconj: conj:preconj

The preconj relation in SD has been renamed to conj:preconj in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.PRECONJUNCT

predet: det:predet

The predet relation in SD has been renamed to det:predet in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.PREDETERMINER

prep: case or mark

Prepositional phrases are treated differently in UD. See the section Treatment of Prepositional Phrases for more details on the new treatment of prepositional phrases.

The various prep_<PREPOSITION> relations in the collapsed and CCprocessed SD representations have been replaced with nmod:<PREPOSITION> relations in the UD enhanced representation.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.CASE_MARKER

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.MARKER

prepc: acl or advcl

The prepc_<PREPOSITION> relations in the collapsed and CCprocessed SD representations have been replaced with acl:<PREPOSITION> and advcl:<PREPOSITION> relations in the UD enhanced representation.

See the section Treatment of Prepositional Phrases for more details.

prt: compound:prt

The prt relation in SD has been renamed to compound:prt in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.PHRASAL_VERB_PARTICLE

punct: punct

punct relations in SD are still annotated as punct in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.PUNCTUATION

quantmod: advmod

The quantmod relation in SD no longer exists as a relation in UD and has been folded into advmod.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.ADVERBIAL_MODIFIER

ref: ref

ref relations in SD are still annotated as ref in UD.

Note that referents are only part of the enhanced UD representation.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.REFERENT

root: root

root relations in SD are still annotated as root in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.GrammaticalRelation.ROOT

tmod: nmod:tmod

The tmod relation in SD has been renamed to nmod:tmod in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.TEMPORAL_MODIFIER

vmod: acl or advcl

The vmod relation in SD no longer exists as a relation in UD and has been folded into the acl and advcl relations in UD.

The acl relation is used for clauses that modify a noun unless the noun is a predicate in a copular construction. The advcl relation is used for clauses that modify a verb or any other predicate.

See the section Treatment of Prepositional Phrases for more details on when to use acl or advcl.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.CLAUSAL_MODIFIER (acl)

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.ADV_CLAUSE_MODIFIER (advcl)

xcomp: xcomp

xcomp relations in SD are still annotated as xcomp in UD.

GrammaticalRelation object:

edu.stanford.nlp.trees.UniversalEnglishGrammaticalRelations.XCLAUSAL_COMPLEMENT

New Relations

The UD represenation also introduces the following new relations which did not exist in SD.

Enhanced Universal Depencies

Apart from the the basic representation, the Stanford parser also provided several post-processed representations including the collapsed and CCprocessed representations. The main purpose of the collapsed representation was to collapse prepositions to get direct dependencies between content words. Considering the new treatment of prepositional phrases the basic representation already has direct dependencies between content words which makes the collapsed representation largely obsolete. On the other hand, the propagation of conjunt dependencies still adds useful dependencies in UD.

For these reasons we decided to reduce the total number of representations to the following three:

The basic representation uses the dependencies defined above, and forms a tree structure. Each word in the sentence is the dependent of exactly one thing, either another word in the sentence or the distingushed ROOT token.

The enhanced representation roughly corresponds to the CCprocessed representation in SD. However, there are two main differences: