Universal Dependencies: A cross-linguistic typology

Marie-Catherine de Marneffe◦ , Timothy Dozat? , Natalia Silveira? ,

Katri Haverinen• , Filip Ginter• , Joakim Nivre/ , Christopher D. Manning?
◦ Linguistics
Department, The Ohio State University
and Computer Science Departments, Stanford University
? Linguistics
• Department of Information Technology, University of Turku
/ Department of Linguistics and Philology, Uppsala University

Abstract
Revisiting the now de facto standard Stanford dependency representation, we propose an improved taxonomy to capture grammatical
relations across languages, including morphologically rich ones. We suggest a two-layered taxonomy: a set of broadly attested
universal grammatical relations, to which language-specific relations can be added. We emphasize the lexicalist stance of the Stanford
Dependencies, which leads to a particular, partially new treatment of compounding, prepositions, and morphology. We show how
existing dependency schemes for several languages map onto the universal taxonomy proposed here and close with consideration of
practical implications of dependency representation choices for NLP applications, in particular parsing.

This paper is a minor revision of our LREC 2014 paper “Universal Stanford dependencies: A cross-linguistic typology”. The
content is largely identical, but the taxonomy of relations has been revised to be consistent with the final Universal Dependencies
(UD) taxonomy for version 1.0 described at http://universaldependencies.github.io/docs/. Version of 2015-11-12.

1. Introduction for a particular treatment of compounding and morphology.

The Stanford Dependencies (SD) representation (de We also discuss the different forms of dependency repre-
Marneffe et al., 2006) was originally developed as a prac- sentation that should be available for UD to be maximally
tical representation of English syntax, aimed at natural lan- useful for a wide range of NLP applications, converging on
guage understanding (NLU) applications. However, it was three versions: the basic one, the enhanced one (which adds
deeply rooted in grammatical relation-based syntactic tra- extra dependencies), and a particular form for parsing.
ditions, which have long emphasized cross-linguistic de-
scription. Faithfulness to these origins was attenuated by 2. A proposed universal taxonomy
desiderata from our NLU applications and the desire for a Table 1 gives the taxonomy we propose for the universal
simple, uniform representation, which was easily intelligi- grammatical relations, with a total of 40 relations. These
ble by non-experts (de Marneffe and Manning, 2008). Nev- relations are taken to be broadly supported across many
ertheless, it is reasonable to suppose that these additional languages in the typological linguistic literature.1 The
goals do not detract from cross-linguistic applicability. proposed universal dependencies (UD) are documented at
In this paper, we attempt a (post-hoc) reconstruction of http://universaldependencies.github.io/docs/.
the underlying typology of the Stanford Dependencies rep-
resentation. This not only gives insights into how the ap- 2.1. The representation builds on lexicalism
proach might be applied to other languages, but also gives An under-elaborated part of the first design principle in
an opportunity to reconsider some of the decisions made in (de Marneffe and Manning, 2008) is that SD adopts the lex-
the original scheme, aiming to propose an improved taxon- icalist hypothesis in syntax, whereby grammatical relations
omy, even for English. We suggest a taxonomy which has should be between whole words (or lexemes). There is a
at its core a set of very broadly attested grammatical rela- longstanding, unresolved debate in linguistics between the-
tions, supplemented as needed by subtypes for language- ories which attempt to build up both words and phrases
particular relations, which capture phenomena important to using the same compositional syntactic mechanisms (and
the syntax of individual languages or language families. in which the notion of a word has minimal privileged ex-
We attempt to make the basic core more applicable, both istence) versus those theories where the word is a funda-
cross-linguistically and across genres, and more faithful to mental unit and which see the morphological processes that
the design principles in de Marneffe & Manning (2008). build up words as fundamentally different from and hid-
We consider how to treat grammatical relations in morpho- den to those that build up sentences, sometimes termed the
logically rich languages, including achieving an appropri- lexical integrity principle (Chomsky, 1970; Bresnan and
ate parallelism between expressing grammatical relations Mchombo, 1995; Aronoff, 2007). For a practical compu-
by prepositions versus morphology. We show how existing tational model, there are great advantages to a lexicalist ap-
dependency schemes for other languages which draw from proach. However, there remain certain difficult cases, such
Stanford Dependencies (Chang et al., 2009; Bosco et al., as how to deal with certain clitics (Zwicky and Pullum,
2013; Haverinen et al., 2014; Seraji et al., 2014; McDon-
ald et al., 2013; Tsarfaty, 2013) can be mapped onto the 1 This is not to say that all languages have all these grammatical
new taxonomy proposed here. We emphasize the lexicalist relations. E.g., many languages, from English to Chicheŵa, allow
stance of both most work in NLP and the syntactic theory verbs to take more than one object, but other languages, such as
on which Stanford Dependencies is based, and hence argue French, do not. Nevertheless, iobj is broadly attested.
Core dependents of clausal predicates grammar notion of direct object and other objects. In cases
Nominal dep Predicate dep of multiple objects, the theme/patient argument is the direct
nsubj csubj object. Finally, the taxonomy aims to clearly indicate in the
nsubjpass csubjpass dependency label (i) a non-canonical voice subject (where
dobj ccomp xcomp the proto-agent argument is not subject, i.e., in passives)
iobj and (ii) whether dependents are noun phrase (NP) argu-
ments versus introducing a new clause/predicate. A design
Non-core dependents of clausal predicates goal of SD has been to distinguish in dependency names
Nominal dep Predicate dep Modifier word where a new clause is being introduced (and so we distin-
nmod advcl advmod guish nsubj and csubj; dobj and ccomp, but also advmod
neg and advcl). We follow LFG in including a distinction be-
Special clausal dependents tween ccomp and xcomp for clausal complements that are
Nominal dep Auxiliary Other standalone (have an internal subject) versus those having
vocative aux mark necessary control (omission) of the dependent predicate’s
discourse auxpass punct subject (have an external subject).3
expl cop Other aspects of the typology are less regular but still
important. The non-core clausal dependents are all modi-
Coordination fiers. The UD taxonomy does not attempt to differentiate fi-
conj cc punct nite from nonfinite clauses, but differentiates attachment to
Noun dependents predicates from attachment to entities: an adverbial clause
Nominal dep Predicate dep Modifier word advcl modifies a predicate whereas an acl “clausal modi-
nummod acl amod fier of noun” modifies a nominal. Clause heads have many
appos det other special dependents, including periphrastic auxiliaries,
nmod neg markers (complementizers and subordinating conjunctions)
as well as vocatives and discourse elements (like well or
Compounding and unanalyzed um). Conjunctions combine elements of many categories.
compound mwe goeswith Under the SD design principles, the conj relation connects
name foreign lexical heads.
Case-marking, prepositions, possessive
2.3. Treatment of copulas
case
SD has advocated a treatment of the copula “be” where it
Loose joining relations is not the head of a clause, but rather the dependent of a lex-
list parataxis remnant ical predicate, as in (1a). Such an analysis is motivated by
dislocated reparandum the fact that many languages often or always lack an overt
Other copula in such constructions, as in the Russian (1b). Similar
Sentence head Unspecified dependency constructions arise even in English if we consider so-called
root dep raising-to-object or small clause constructions. Under the
basic analysis proposed for SD, the predicate complement
Table 1: Dependencies in universal Stanford Dependencies. is not linked to its subject argument, but in the enhanced
Note: nmod, neg and punct appear in two places. representation (see below), the linkage is then parallel to
the treatment in a zero copula language, as in (1c).
nsubj
cop
det
1983), phonologically bound words which behave like syn- amod
tactic words (we follow many treebanks in separating them (1) a. Ivan is the best dancer
as words and having them participate in the syntax) and
how to treat words that are split in unedited writing (see the Penn Chinese Treebank does try to make an adjunct/argument
section 2.4.). distinction (Xue et al., 2005) and effectively PropBank (Palmer
et al., 2005) adds an argument/adjunct distinction overlay to the
2.2. Dependents of clausal predicates English Penn Treebank.
The system of clausal dependents most closely fol- 3 The latter is used to represent control and raising construc-

lows Lexical-Functional Grammar (LFG) (Bresnan, 2001). tions, including cases of raising-to-object or so-called “excep-
However, the taxonomy differs from LFG in several re- tional case marking”. The treatment of this last group of phenom-
spects. First, the clear distinction between core arguments ena is one of the largest substantive breaks with the Penn Tree-
and other dependents is made, but the distinction between bank annotation tradition, which follows Chomskyan approaches
from the extended standard theory, Government-Binding Theory
adjuncts and oblique arguments (Radford, 1988) is taken
(Chomsky, 1981) et seq., in treating such constructions as having
to be sufficiently subtle, unclear, and argued over that it is
a complete clausal complement with “exceptional case marking”,
eliminated.2 Second, the model reverts to the traditional rather than an object in the higher clause. While an xcomp is usu-
ally non-finite and a ccomp is usually finite, there are non-finite
2 The original Penn Treebank annotators also decided not to try
ccomp, such as in a for-to infinitive (“I arranged [for her to go by
to mark arguments vs. adjuncts (Taylor et al., 2003). Conversely, bus]” ) and some languages have inflected infinitives.
b. Russian: The treatment of case marking is illustrated in (3). In
nsubj
amod (3a), at in Hebrew is a separate token indicating an ac-
Ivan lučšij tancor cusative object: the case marker depends on the object. In
Ivan best dancer (3c), we show the analysis when case markers are mor-
xcomp
phemes. The case morpheme is not divided off the noun
nsubj as a separate case dependent, but the noun as a whole is
det
nsubj dobj amod analyzed as a nmod of the verb. To overtly mark case, we
c. I judge Ivan the best dancer include POS tags in the representation as shown in (3b) and
(3d). We use the universal POS tagset from Petrov et al.
2.4. Modification vs. compounding (2012) to which we append case information.
The universal scheme keeps the rich taxonomy of noun
dependents that are one of the strengths of SD. An im- (3) a. Hebrew:
dobj
portant part of the typology is to differentiate compounds case
(multi-root lexemes) from modification or complementa- wkfraiti at hsrj
tion. Under a lexicalist approach, compound words are fun- and when I saw ACC the movie
damentally different from cases of phrasal modification.
b. dobj(wkfraiti/VERB, hsrj/NOUN)
There are three relations for compounding. We use mwe
case(hsrj/NOUN, at/PRT- ACC)
for fixed grammaticized expressions with function words
(e.g., instead of : mwe(of,instead), Fr. plutôt que “rather c. Russian:
than”: mwe(que,plutôt)), name for proper nouns constituted nmod
nsubj dobj
of multiple nominal elements, as in the Finnish and Italian
Ya napisal pis’mo perom
dependency treebanks,4 and compound to label other types I wrote the letter with a quill
of multi-word lexemes. Thus compound is used for any
kind of X0 compounding: noun compounds (e.g., phone d. nsubj(napisal/VERB, Ya/NOUN - NOM)
book), but also verb and adjective compounds that are more dobj(napisal/VERB, pis’mo/NOUN - ACC)
common in other languages (such as Persian or Japanese nmod(napisal/VERB, perom/NOUN - INSTR)
light verb constructions); for numbers (e.g., three thou- This treatment provides parallelism between different
sand books gives compound(thousand,three)); for particles constructions across and within languages. A good result is
of phrasal verbs (e.g., put up: compound(put,up)). that we now have greater parallelism between prepositional
2.5. Treatment of prepositions and case marking phrases and subordinate clauses, which are in practice often
introduced by a preposition, as in (4).
A major proposed change from the extant versions of
SD is a new treatment of prepositions to provide a uniform nmod
case
analysis of prepositions and case in morphologically rich nsubj det
languages. The analysis we chose is to push all the way (4) a. Sue left after the rehearsal
the design principle of having direct links between con-
advcl
tent words. We abandon treating a preposition as a me-
mark
diator between a modified word and its object, and, in- nsubj nsubj
stead, any case-marking element (including prepositions, b. Sue left after we did
postpositions, and clitic case markers) will be treated as a
dependent of the noun it attaches to or introduces. The pro- We also obtain parallel constructions for the possessive
posed analysis is shown in (2): nmod labels the relation alternation as shown in (2), for variant forms with case, a
between the two content words, whereas the preposition is preposition or a postposition in Finnish, as shown in (5),
now viewed as a case depending on its complement. In and for the dative alternation where the prepositional con-
general, nmod expresses some form of oblique or adjunct struction gets a similar analysis to the double object con-
relation further specified by the case. struction, see (6).
nmod
det case
(5) a. Finnish:
nmod
(2) a. the Chair ’s office case

nmod
etsiä ilman johtolankaa
case to search without clue.PARTITIVE
det det
nmod case
b. the office of the Chair
b. etsiä taskulampun kanssa
c. French: to search torch.GENITIVE with
nmod
det case nmod
le bureau du président c. etsiä johtolangatta
the office of the Chair to search clue.ABESSIVE
4 That is, name would be used between the words of “Hillary dobj
Rodham Clinton” but not to replace the usual relations in a phrasal iobj
or clausal name like “The Lord of the Rings”. (6) a. give the children the toys
nmod
The “loose joining relations” aim at a robust analysis of
dobj case
more informal forms of text, which are now common in
b. give the toys to the children
NLP applications. Informal written text often contains lists
c. French: of comparable items, which are parsed as single sentences.
nmod
Email signatures in particular contain these structures, in
dobj case
the form of contact information. Following de Marneffe et
donner les jouets aux enfants
give the toys to the children al. (2013), we use the list, parataxis, (and appos) relations
to label these kinds of structures. The relation parataxis is
Another advantage of this new analysis is that it pro- also used in more formal writing for constructions such as
vides a treatment of prepositional phrases that are predica- sentences joined with a colon.
tive complements of “be” as in (7) that is consistent with the The dislocated relation captures preposed (topics) and
treatment of nominal predicative complements, as in (1). postposed elements. The remnant relation is used to pro-
nsubj vide a treatment of ellipsis (in the case of gapping or strip-
cop
case ping, where predicational or verbal material gets elided),
(7) Sue is in shape something that was lacking in earlier versions of SD. It
provides a basis for being able to reconstruct dependencies
SD is a surface syntactic representation, which does not in the enhanced version of SD. For example, in (11), the
represent semantic roles. The semantic roles of modifiers remnant relations enable us to correctly retrieve the sub-
are hard to categorize and hard to determine. We feel that jects and objects in the clauses with an elided verb.
there is a lot of use for a representation which works solely
remnant remnant
in terms of the overt role-marking resources of each lan- remnant remnant
guage. This is supported by many rich language-particular nsubj dobj

traditions of grammatical analysis, whether via Sanskrit (11) John won bronze, Mary silver, and Sandy gold.
cases or the case particles on Japanese bunsetsu.
In the case of stacked prepositions, such as out of, the In contrast, in right-node-raising (RNR) (12) and VP-
UD taxonomy treats them as some form of mwe, as in (8b). ellipsis (13) constructions in which some kind of predica-
tional or verbal material is still present, the remnant relation
(8) a. Out of all this, something good will come. is not used. In RNR, the verbs are coordinated and the ob-
case
ject is a dobj of the first verb. In VP-ellipsis, we keep the
mwe predet nmod auxiliary as the head, as shown in (13).
b. Out of all this ... come
dobj
conj
Prepositions sometimes introduce a clause as their com- nsubj cc
plement, e.g., (9a). Clauses that modify an entity are an- (12) John bought and ate an apple.
alyzed as acl. Thus, under the UD analysis, the head of
conj
the modifier of data will be upset, marked as acl. Both the nsubj cc
preposition and the subordinate conjunction will be mark aux dobj nsubj advmod
for markers in an “extended clausal projection” (whereas (13) John will win gold and Mary will too.
case is used for prepositions in “extended nominal projec-
tions”). The result is the analysis in (9b). 2.7. Language-particular relations
In addition to a universal dependency taxonomy, it is de-
(9) a. We have no data about whether users are upset.
sirable to recognize grammatical relations that are particu-
acl lar to one language or a small group of related languages.
mark
mark Such language-particular relations are necessary to accu-
nsubj rately capture the genius of a particular language but will
cop
not involve concepts that generalize broadly. The sugges-
b. data about whether users are upset tion here is that these relations should always be regarded
as a subtype of an existing Universal SD relation. The
2.6. Informal text genres
SD relations have a taxonomic organization (de Marneffe
Following the practical approach used in part-of-speech et al., 2006), and some of the universal relations are al-
tagging of recent LDC treebanks, we introduce the relation ready subtypes of each other (e.g., auxpass is a subtype of
goeswith to connect multiple tokens that correspond to a aux). Language-particular relations that seem useful to dis-
single standard word, as a result of reanalysis of words as tinguish for English are included at the bottom of Table 2:
compounds (“hand some” for “handsome”) or input error npmod for bare nominal modifiers of predicates lacking a
(“othe r” for “other”). We use foreign to label sequences preposition, among which in particular there is tmod for
of foreign words. To indicate disfluencies overridden in a bare NP temporal modifiers; poss for possessives, since the
speech repair, we use reparandum, as in (10). syntax of prenominal possession is very distinct from post-
nmod nominal modifiers (which may also express possession);
reparandum predet for words such as all that precede regular determin-
case case
det det ers and preconj for words preceding a conjunction like ei-
(10) Go to the righ- to the left. ther; and prt for verb particles.
3. Mapping to existing schemes mapped to whatever syntactic role the relative is playing in
There has recently been an effort to push towards ho- the relative clause (nsubj, dobj, etc.), information which is
mogeneity across resources for different languages and present in the second annotation layer of the TDT corpus.
to come up with cross-linguistically consistent annotation ISDT is the conversion of the MIDT Italian dependency
aimed at multi-lingual technology, for part-of-speech tagset scheme to SD. Some of the clit uses in ISDT (for reflex-
(Petrov et al., 2012) as well as for dependency representa- ive pronouns in pronominal verbs – frequent in Romance
tion (McDonald et al., 2013). The scheme proposed in Mc- languages – such as Fr. se douter “to suspect”) will be en-
Donald et al. (2013) took SD as a starting point. Annota- compassed by expl. However, when the reflexive pronoun
tors for six different languages (German, English, Swedish, can truly be a direct or indirect object, it gets assigned the
Spanish, French and Korean) produced annotation guide- corresponding object relation.
lines for the language they were working on, keeping the Chinese has serial verb constructions which might now
label and construction set as close as possible to the original be compound (and not conj). We treat post-nominal local-
English SD representation. They were only allowed to add izers and prepositions as a form of case.
labels for phenomena that do not exist in English. Given the In Persian, there are no relative pronouns and rel was
sets of relations obtained for the different languages, a har- used for the fixed relative marker, but it can be mapped to
monization step was performed to maximize consistency of mark. UPDT has a fw relation between sequences of for-
annotation across languages. However, this rigid strategy eign words, unanalyzed within Persian grammar, which we
lost some important distinctions, such as the distinction be- adopt, naming it foreign. We also adopt the UPDT dep-top
tween compounding and phrasal modification, while main- relation used for a fronted element that introduces the topic
taining some distinctions that are best abandoned, such as a of a sentence, but we generalize it to dislocated to account
distinction between infinitival and participial modifiers. for postposed elements as well as. Right dislocated ele-
McDonald et al. (2013) does not address giving an ele- ments are frequent in spoken languages: e.g., Fr. faut pas
gant treatment of morphologically rich languages. In con- la manger, la pâte (literally, “need not it eat, the dough”).
trast, Tsarfaty (2013) proposes to treat morphology as syn- Labels of language-specific relations will be harmonized
tax in her dependency proposal, illustrated with Hebrew. to be shared between languages: Chinese assmod will be
However, this representation choice conflicts with the lexi- mapped to poss, and Persian dep-top to topic.
calist approach of SD. Here we take up her goal of trying
to give a uniform treatment of both morphologically rich 4. Different forms of Stanford Dependencies
and morphologically poor languages, but suggest achieving The current Stanford converter provides a number of
the goal in a different way, which maintains a lexicalist ap- variant forms of SD, of which the most used are the ba-
proach (see Section 2.5.). Table 2 shows a comparison be- sic dependency tree, and the collapsed, cc-processed form
tween the evolution of the SD scheme (Stanford Dependen- that adds extra dependency arcs, restructures prepositions
cies v2.0.0, used in the SANCL shared task, and Stanford to not be heads, and spreads relations across conjunctions.
Dependencies v3.3.0, the 2013/11/12 version), the propos- This section suggests some new ideas for how to provide
als in McDonald et al. (2013) (GSD) and in Tsarfaty (2013) potentially less but different options.
(TSD), and the dependency set proposed here (UD). One concern about our proposed taxonomy is that
Existing dependency sets for other languages can be straightforward parsing to UD is likely to be harder for
fairly straightforwardly mapped onto our new proposal. parsers than the current representation (for English). It
Even if the schemes examined here are “SD-centric”, they is now fairly well known that, while dependency repre-
dealt with particular constructions present in each language sentations in which content words are made heads tend
and posited new relations when necessary. The mapping is to help semantically oriented downstream applications, de-
less difficult because UD adopts some of the ideas and re- pendency parsing numbers are higher if you make auxiliary
lations that were first developed for these other treebanks, verbs heads, if you analyze long conjunctions by chaining
such as the content word as head analysis of prepositional (rather than having a single head of the whole construction),
phrases from the Turku Dependency Treebank (TDT). In and if you make prepositions the head of prepositional
table 3, we show how the Finnish (Haverinen et al., 2014), phrases (Schwartz et al., 2012; Elming et al., 2013). The
Italian (Bosco et al., 2013), Chinese (Chang et al., 2009) generalization is that dependency parsers, perhaps in par-
and Persian (Seraji et al., 2014) schemes can be mapped ticular the efficient shift-reduce-style dependency parsers
onto the proposed universal taxonomy (UD). The bold la- like MaltParser (Nivre et al., 2007), work best the more the
bels are language-specific relations, and they are subtypes dependency representation looks like a chain of short de-
of the corresponding UD relation in the row. Gaps indi- pendencies along the sentence. Under the proposed UD,
cate existing constructions in the language that were not we would then be making the “wrong” choice in each case.
captured in the original scheme (the UD label is applicable However, it seems wrong-headed to choose a linguis-
there); ∗ indicates constructions that are not present in the tic representation to maximize parser performance, rather
language. Since copular verbs are not heads anymore, the than based on the linguistic quality of the representation
attr relation is removed, requiring modifications to the ex- and its usefulness for applications that build further pro-
isting analyses of copular sentences for Italian and Chinese. cessing on top of it. Rather, it may be useful to do pars-
We also introduced extra relations for certain constructions, ing using a transformation of the target dependency sys-
which some schemes had not incorporated yet. tem. In constituency parsing, it is completely usual for the
For Finnish, the relation rel (for relative marker) will be target representation to be transformed so as to improve
SD v2.0.0 SD v3.3.0 GSD TSD UD Notes
nsubj nsubj nsubj nsubj nsubj X
csubj csubj csubj csubj csubj X
nsubjpass nsubjpass nsubjpass nsubjpass nsubjpass X
csubjpass csubjpass csubjpass csubjpass csubjpass X
dobj dobj dobj dobj dobj X
iobj iobj iobj iobj iobj X (TSD also has gobj for genitive object)
ccomp ccomp ccomp ccomp ccomp UD & TSD define as clause with internal subject
xcomp xcomp xcomp xcomp xcomp UD & TSD define as clause with controlling subject
acomp acomp acomp acomp – acomp can be generalized into xcomp
attr – attr – – attr removed: wh- is head or xcomp (with copula head option)
advmod advmod advmod advmod advmod X
advcl advcl advcl – advcl TSD omits but needed to preserve clause boundaries
purpcl – – – – Folded into advcl
neg neg neg neg neg As well as adverbial not, never, UD extends to negative det like no
det det det det det X (TSD defines dem and def as subtypes of det)
amod amod amod amod amod X
appos appos appos appos appos X
abbrev – – abbrev – Parenthetical abbreviations become cases of appos
num num num nummod nummod Renamed for clarity
partmod partmod partmod ? acl Make partmod, infmod into acl; use (rich) POS to distinguish
infmod infmod infmod infmod acl Make partmod, infmod into acl; use (rich) POS to distinguish
quantmod quantmod advmod ? – Generally folded into advmod
root root ROOT root root
punct punct p punct punct
aux aux aux aux aux X (TSD has qaux for question auxiliary; infinitive to is now mark)
auxpass auxpass auxpass auxpass auxpass X
cop cop cop cop cop GSD has cop only in content-head version
expl expl expl expl expl Subject and object expletives, frozen reflexives (Fr. se douter)
mark mark mark mark mark X to introducing an infinitive will now be mark (instead of aux)
complm – – complm – Remove and use mark more broadly
– discourse – parataxis? discourse A gap in original and other typologies
– – – – vocative A gap in original and other typologies
dep dep dep dep dep GSD uses for vocative and discourse
rel – rel rel – Converter’s unresolved ones now dep; TSD rel is really mark
prep prep adpmod prepmod case UD case is dependent of NP modifier not of thing modified
– – nmod – – Equivalent to nmod below
pobj pobj adpobj pobj nmod nmod now goes from thing modified to NP of PP
pcomp pcomp adpcomp pcomp acl acl goes from thing modified to clause
– – adp prep/case case TSD has N/A/G/D subtypes, but can’t keep adding for all cases
possessive possessive adp gen – View as a manifestation of case
nn nn compmod nn compound Generalize nn to light verbs, etc.; X0 compounding not modification
– – mwe – name Multi-word proper nouns (e.g., John Smith) as in TDT and ISDT
number number num nummod? – Regarded as type of compound; using nummod is wrong
mwe mwe mwe mwe mwe Fixed expressions with function words (so that, all but, due to, . . . )
– goeswith – – goeswith For orthographic errors: othe r
– – – – foreign Linear analysis of foreign words (head is left-most) as in UPDT
– – – – reparandum For disfluencies overridden in speech repairs
conj conj conj conj conj X
cc cc cc cc cc X
parataxis parataxis parataxis parataxis parataxis X
– – – – list Used for informal list structures, signature blocks, etc.
– – – – remnant Used to give a treatment of ellipsis without empty nodes
– – – – dislocated Preposed topics and dislocated elements as in UPDT
English particular
rcmod rcmod rcmod rcmod relcl A subtype of acl
npadvmod npadvmod nmod advmod? npmod A subtype of nmod
tmod tmod advmod tmod tmod A subtype of nmod
predet predet – predet predet A subtype of det
preconj preconj cc preconj preconj A subtype of cc
prt prt prt ? prt A subtype of compound
poss poss poss possmod poss A subtype of nmod

Table 2: Comparison of proposals on English: SD, McDonald et al. (2013) (GSD), Tsarfaty (2013) (TSD) and ours (UD).
TDT ISDT Chinese UPDT UD
nsubj nsubj nsubj, top nsubj nsubj
csubj csubj ∗ csubj
∗ nsubjpass nsubjpass nsubjpass nsubjpass
∗ csubjpass ∗ csubjpass
dobj dobj , clit dobj dobj dobj
∗ iobj , clit iobj ∗ iobj
ccomp, iccomp ccomp ccomp, rcomp ccomp ccomp
xcomp, acomp xcomp, acomp xcomp, acomp xcomp
∗ attr attr ∗ ∗
advmod, quantmod advmod advmod, dvpmod advmod, quantmod advmod
advcl advcl some advmod advcl advcl
neg neg neg neg neg
det det, predet det det, predet det
amod amod amod amod amod
appos appos prnmod appos appos
num num nummod, ordmod num nummod
rcmod, partmod, infmod rcmod, partmod, rcmod, vmod, rcmod
acl
pcomp pccomp, lccomp
root root root root root
punct punct punct punct punct
aux aux asp, mmod aux aux
auxpass auxpass pass auxpass auxpass
cop cop cop cop cop
∗ expl, clit ∗ ∗ expl
complm, mark mark cpm complm, mark, rel mark
intj discourse discourse
voc dep-voc vocative
dep comp, mod dep dep dep
poss, gobj, gsubj, pobj, poss, pobj, lobj, assmod, pobj, poss, cpobj
nmod
nommod npadvmod, tmod clf, range, tmod npadvmod, tmod
adpos possessive, prep assm, prep, ba, dvpm, loc acc, prep, cprep case
number, nn, prt number, nn nn, some conj number, nn, prt, compound
{nsubj|dobj|acomp|prep}-lvc
name nnp name
some dep mwe prtmod mwe mwe
goeswith goeswith
fw foreign
reparandum
conj conj conj, etc, comod conj conj
cc, preconj cc, preconj cc cc, preconjunct cc
parataxis parataxis parataxis parataxis
list
ellipsis remnant
dep-top dislocated

Table 3: Mappings of the Finnish (TDT), Italian (ISDT), Chinese and Persian (UPDT) schemes to UD.

parsing numbers, such as by head-lexicalization (Collins, standard in dependency parsing. We therefore propose a
2003), by manual or automatic subcategorization of cat- parsing representation that changes some of the depen-
egories (Klein and Manning, 2003; Petrov et al., 2006), dency head choices to maximize parsing performance. This
and even by other methods such as unary chain contrac- requires developing tools to convert seamlessly both ways
tion (Finkel et al., 2008). After parsing, a detransformation between the basic and parsing representations.5
process reconstructs trees in the target representation. This Since the new treatment of prepositional phrases basi-
kind of transform-detransform architecture is at present less cally does what the collapsed representation was designed
common in dependency parsing, although Nilsson, Nivre to do (putting a direct link between the noun complement of
& Hall (2006; 2007) do this for coordination and verb a preposition and what it modifies), except for not renaming
groups, and pseudo-projective parsing (Nivre and Nilsson,
2005) can also be seen as an instance of this architecture. 5 A small part of this is in place in the Stanford converter, in the
A transform-detransform architecture should become more ability to generate copula- and content-head versions from trees.
the dependency relation, the collapsed representation on its Elming, J., Johannsen, A., Klerke, S., Lapponi, E., Mar-
own has less utility. However, the ideas of having extra de- tinez, H., and Søgaard, A. (2013). Down-stream ef-
pendencies to mark external subjects and the external role fects of tree-to-dependency conversions. In NAACL HLT
in relative clauses is useful, the renaming of dependencies 2013.
to include case or preposition information helps in many Finkel, J. R., Kleeman, A., and Manning, C. D. (2008). Ef-
applications, and spreading relations over conjunctions is ficient, feature-based, conditional random field parsing.
definitely useful for relation extraction. These transforma- In ACL 46, pages 959–967.
tions can be provided in an enhanced representation. Haverinen, K., Nyblom, J., Viljanen, T., Laippala, V., Ko-
We thus suggest providing three versions of Stanford De- honen, S., Missilä, A., Ojala, S., Salakoski, T., and
pendencies: basic, enhanced, and parsing. Ginter, F. (2014). Building the essential resources for
Finnish: the Turku dependency treebank. Language Re-
5. Conclusion sources and Evaluation, 48:493–531.
We proposed a taxonomy of grammatical relations appli- Klein, D. and Manning, C. D. (2003). Accurate unlexical-
cable to a variety of languages, developing the implications ized parsing. In ACL 41, pages 423–430.
of a lexicalist approach for the treatment of morphology McDonald, R., Nivre, J., Quirmbach-Brundage, Y., Gold-
and compounding. Some of the decisions made on linguis- berg, Y., Das, D., Ganchev, K., Hall, K., Petrov,
tic grounds are at odds with what works best for processing S., Zhang, H., Täckström, O., Bedini, C., Bertomeu
tools. We suggested that the transform-detransform archi- Castelló, N., and Lee, J. (2013). Universal dependency
tecture standardly used in constituency parsing is the solu- annotation for multilingual parsing. In ACL 51.
tion to adopt for dependency parsing. We worked out the Nilsson, J., Nivre, J., and Hall, J. (2006). Graph transfor-
mapping of existing dependency resources for different lan- mations in data-driven dependency parsing. In COLING
guages to the taxonomy proposed here. We hope this work 21 and ACL 44, pages 257–264.
will enhance consistency in annotation between languages Nilsson, J., Nivre, J., and Hall, J. (2007). Tree transforma-
and further facilitate cross-lingual applications. tions for inductive dependency parsing. In ACL 45.
Nivre, J. and Nilsson, J. (2005). Pseudo-projective depen-
6. References
dency parsing. In ACL 43, pages 99–106.
Aronoff, M. (2007). In the beginning was the word. Lan- Nivre, J., Hall, J., Nilsson, J., Chanev, A., Eryigit, G.,
guage, 83:803–830. Kübler, S., Marinov, S., and Marsi, E. (2007). Malt-
Bosco, C., Montemagni, S., and Simi, M. (2013). Convert- Parser: A language-independent system for data-driven
ing Italian treebanks: Towards an Italian Stanford depen- dependency parsing. Natural Language Engineering,
dency treebank. In Seventh Linguistic Annotation Work- 13:95–135.
shop & Interoperability with Discourse. Palmer, M., Gildea, D., and Kingsbury, P. (2005). The
Bresnan, J. and Mchombo, S. A. (1995). The lexical in- proposition bank: An annotated corpus of semantic roles.
tegrity principle: Evidence from Bantu. Natural Lan- Computational Linguistics, 31:71–105.
guage and Linguistic Theory, 13:181–254. Petrov, S., Barrett, L., Thibaux, R., and Klein, D. (2006).
Bresnan, J. (2001). Lexical-Functional Syntax. Blackwell, Learning accurate, compact, and interpretable tree anno-
Oxford. tation. In COLING 21 and ACL 44, pages 433–440.
Chang, P.-C., Tseng, H., Jurafsky, D., and Manning, C. D. Petrov, S., Das, D., and McDonald, R. (2012). A universal
(2009). Discriminative reordering with Chinese gram- part-of-speech tagset. In LREC 2012.
matical relations features. In Third Workshop on Syntax Radford, A. (1988). Transformational Grammar. Cam-
and Structure in Statistical Translation, pages 51–59. bridge University Press, Cambridge.
Chomsky, N. (1970). Remarks on nominalization. In Ja- Schwartz, R., Abend, O., and Rappoport, A. (2012).
cobs, R. A. and Rosenbaum, P. S., editors, Readings Learnability-based syntactic annotation design. In COL-
in English transformational grammar, pages 184–221. ING 24, pages 2405–2422.
Ginn, Waltham, MA. Seraji, M., Jahani, C., Megyesi, B., and Nivre, J. (2014). A
Chomsky, N. (1981). Lectures on Government and Bind- Persian treebank with Stanford typed dependencies. In
ing. Foris, Dordrecht. LREC 2014.
Collins, M. (2003). Head-driven statistical models for Taylor, A., Marcus, M., and Santorini, B. (2003). The Penn
natural language parsing. Computational Linguistics, treebank: An overview. In Abeillé, A., editor, Building
29:589–637. and Using Parsed Corpora, volume 20 of Text, Speech
de Marneffe, M.-C. and Manning, C. D. (2008). The Stan- and Language Technology. Springer.
ford typed dependencies representation. In Workshop on Tsarfaty, R. (2013). A unified morpho-syntactic scheme of
Cross-framework and Cross-domain Parser Evaluation. Stanford dependencies. In ACL 51.
de Marneffe, M.-C., MacCartney, B., and Manning, Xue, N., Xia, F., Chiou, F.-D., and Palmer, M. (2005). The
C. D. (2006). Generating typed dependency parses from Penn Chinese treebank: Phrase structure annotation of
phrase structure parses. In LREC 2006. a large corpus. Natural Language Engineering, 11:207–
de Marneffe, M.-C., Connor, M., Silveira, N., Bowman, 238.
S. R., Dozat, T., and Manning, C. D. (2013). More con- Zwicky, A. M. and Pullum, G. K. (1983). Cliticization vs.
structions, more genres: Extending Stanford dependen- inflection: English n’t. Language, 59:502–513.
cies. In DepLing 2013.