Enhanced Dependencies
We always intended the Universal Dependencies representation to be used in shallow natural language understanding tasks such as relation extraction or biomedical event extraction. For such tasks, one is typically interested in the relation between certain entities, e.g., the relation between two persons or whether one protein interacts with another. UD is particularly well suited for such tasks as UD trees contain many direct dependencies between content words and many of the dependency labels provide a lot of information about the type of relation between two content words. However, for some constructions, the dependency path between two content words of interest can be very long in a UD tree, which complicates determining how the content words are related. Further, some dependency types such as obl or nmod are used for many different types of arguments and modifiers, and therefore they are not very informative on their own. For these reasons, we also provide guidelines for an enhanced representation, which makes some of the implicit relations between words more explicit, and augments some of the dependency labels to facilitate the disambiguation of types of arguments and modifiers.
Enhanced UD graphs may contain some or all of the following enhancements, which are described in the sections below.
- Null nodes for elided predicates
- Additional subject relations for control and raising constructions
- Propagation of conjuncts
- Coreference in relative clause constructions
- Modifier labels that contain the preposition or other case-marking information
Note that the enhanced graph is not necessarily a supergraph of the basic tree, i.e., the graph is not required to contain all the basic dependency relations. For this reason, all relations of the enhanced graph (also the ones that are present in the basic UD tree) have to be included in the DEPS column of a CoNLL-U file. See the specificiation of the CoNLL-U file format for details.
Controlled/raised subjects
The basic trees lack a subject dependency between a controlled verb and its controller or between an embedded verb and its raised subject. In the enhanced graph, there is an additional dependency between the embedded verb and the subject of the matrix clause.
| Basic | Enhanced |
|---|---|
|
1 Mary _ _ _ _ 2 nsubj _ _
2 wants _ _ _ _ 0 root _ _
3 to _ _ _ _ 4 mark _ _
4 buy _ _ _ _ 2 xcomp _ _
5 a _ _ _ _ 6 det _ _
6 book _ _ _ _ 4 obj _ _
7 . _ _ _ _ 2 punct _ _
|
# visual-style 4 1 nsubj color:blue
1 Mary _ _ _ _ 2 nsubj 4:nsubj _
2 wants _ _ _ _ 0 root _ _
3 to _ _ _ _ 4 mark _ _
4 buy _ _ _ _ 2 xcomp _ _
5 a _ _ _ _ 6 det _ _
6 book _ _ _ _ 4 obj _ _
7 . _ _ _ _ 2 punct _ _
|
|
1 She _ _ _ _ 2 nsubj _ _
2 seems _ _ _ _ 0 root _ _
3 to _ _ _ _ 5 mark _ _
4 be _ _ _ _ 5 aux _ _
5 reading _ _ _ _ 2 xcomp _ _
6 a _ _ _ _ 7 det _ _
7 book _ _ _ _ 5 obj _ _
8 . _ _ _ _ 2 punct _ _
|
# visual-style 5 1 nsubj color:blue
1 She _ _ _ _ 2 nsubj 5:nsubj _
2 seems _ _ _ _ 0 root _ _
3 to _ _ _ _ 5 mark _ _
4 be _ _ _ _ 5 aux _ _
5 reading _ _ _ _ 2 xcomp _ _
6 a _ _ _ _ 7 det _ _
7 book _ _ _ _ 5 obj _ _
8 . _ _ _ _ 2 punct _ _
|
Ellipsis
(See also the guidelines on ellipsis.)
In the enhanced representation, we add special null nodes in clauses in which a predicate is elided.
|
# visual-style 5 6 orphan color:green
# visual-style 2 5 conj color:green
# visual-style 5 4 cc color:green
1 I _ _ _ _ 2 nsubj _ _
2 like _ _ _ _ 0 root _ _
3 tea _ _ _ _ 2 obj _ _
4 and _ _ _ _ 5 cc _ _
5 you _ _ _ _ 2 conj _ _
6 coffee _ _ _ _ 5 orphan _ _
7 . _ _ _ _ 2 punct _ _
|
# visual-style 6 7 obj color:blue
# visual-style 6 5 nsubj color:blue
# visual-style 2 6 conj color:blue
# visual-style 6 4 cc color:blue
1 I _ _ _ _ 2 nsubj _ _
2 like _ _ _ _ 0 root _ _
3 tea _ _ _ _ 2 obj _ _
4 and _ _ _ _ 6 cc _ _
5 you _ _ _ _ 6 nsubj _ _
6 E5.1 _ _ _ _ 2 conj _ _
7 coffee _ _ _ _ 6 obj _ _
8 . _ _ _ _ 2 punct _ _
|
|
# visual-style 8 10 orphan color:green
# visual-style 2 8 conj color:green
# visual-style 8 7 cc color:green
1 Mary _ _ _ _ 2 nsubj _ _
2 wants _ _ _ _ 0 root _ _
3 to _ _ _ _ 4 mark _ _
4 buy _ _ _ _ 2 xcomp _ _
5 a _ _ _ _ 6 det _ _
6 book _ _ _ _ 4 obj _ _
7 and _ _ _ _ 8 cc _ _
8 Jenny _ _ _ _ 2 conj _ _
9 a _ _ _ _ 10 det _ _
10 CD _ _ _ _ 8 orphan _ _
11 . _ _ _ _ 2 punct _ _
|
# visual-style 9 8 nsubj color:blue
# visual-style 10 12 obj color:blue
# visual-style 9 10 xcomp color:blue
# visual-style 9 7 cc color:blue
# visual-style 4 1 nsubj color:blue
# visual-style 10 8 nsubj color:blue
1 Mary _ _ _ _ 2 nsubj 4:nsubj _
2 wants _ _ _ _ 0 root _ _
3 to _ _ _ _ 4 mark _ _
4 buy _ _ _ _ 2 xcomp _ _
5 a _ _ _ _ 6 det _ _
6 book _ _ _ _ 4 obj _ _
7 and _ _ _ _ 9 cc _ _
8 Jenny _ _ _ _ 9 nsubj 10:nsubj _
9 E8.1 _ _ _ _ 2 conj _ _
10 E8.2 _ _ _ _ 9 xcomp _ _
11 a _ _ _ _ 12 det _ _
12 CD _ _ _ _ 10 obj _ _
13 . _ _ _ _ 2 punct _ _
|
Note that this is a case in which the enhanced UD graph is not a supergraph of the basic tree as the basic tree contains orphan relations, which are not present in the enhanced UD graph.
Propagation of Conjuncts
In the basic representation, the governor and dependents of a conjoined phrase are all attached to the first conjunct. This often leads to very long dependency paths between content words. The enhanced representation therefore also contains dependencies between the other conjuncts and the governor and dependents of the phrase.
Conjoined verbs and verb phrases
When two verbs share their objects (or other complements), the subject and the object of the conjoined verbs are attached to every conjunct.
|
1 The _ _ _ _ 2 det _ _
2 store _ _ _ _ 3 nsubj _ _
3 buys _ _ _ _ 0 root _ _
4 and _ _ _ _ 5 cc _ _
5 sells _ _ _ _ 3 conj _ _
6 cameras _ _ _ _ 3 obj _ _
7 . _ _ _ _ 3 punct _ _
|
# visual-style 5 2 nsubj color:blue
# visual-style 5 6 obj color:blue
1 The _ _ _ _ 2 det _ _
2 store _ _ _ _ 3 nsubj 5:nsubj _
3 buys _ _ _ _ 0 root _ _
4 and _ _ _ _ 5 cc _ _
5 sells _ _ _ _ 3 conj _ _
6 cameras _ _ _ _ 3 obj 5:obj _
7 . _ _ _ _ 3 punct _ _
|
However, if the complements of the second verb are not shared, only the shared dependents are attached to every conjunct.
|
1 She _ _ _ _ 3 nsubj _ _
2 was _ _ _ _ 3 aux _ _
3 reading _ _ _ _ 0 root _ _
4 or _ _ _ _ 5 cc _ _
5 watching _ _ _ _ 3 conj _ _
6 a _ _ _ _ 7 det _ _
7 movie _ _ _ _ 5 obj _ _
8 . _ _ _ _ 3 punct _ _
|
# visual-style 5 1 nsubj color:blue
# visual-style 5 2 aux color:blue
1 She _ _ _ _ 3 nsubj 5:nsubj _
2 was _ _ _ _ 3 aux 5:aux _
3 reading _ _ _ _ 0 root _ _
4 or _ _ _ _ 5 cc _ _
5 watching _ _ _ _ 3 conj _ _
6 a _ _ _ _ 7 det _ _
7 movie _ _ _ _ 5 obj _ _
8 . _ _ _ _ 3 punct _ _
|
Similarly, the enhanced representation can also distinguish private dependents of the first verb. Note however that in this case it cannot be inferred from the basic representation automatically.
|
1 She _ _ _ _ 3 nsubj _ _
2 was _ _ _ _ 3 aux _ _
3 watching _ _ _ _ 0 root _ _
4 a _ _ _ _ 5 det _ _
5 movie _ _ _ _ 3 obj _ _
6 or _ _ _ _ 7 cc _ _
7 reading _ _ _ _ 3 conj _ _
8 . _ _ _ _ 3 punct _ _
|
# visual-style 7 1 nsubj color:blue
# visual-style 7 2 aux color:blue
1 She _ _ _ _ 3 nsubj 7:nsubj _
2 was _ _ _ _ 3 aux 7:aux _
3 watching _ _ _ _ 0 root _ _
4 a _ _ _ _ 5 det _ _
5 movie _ _ _ _ 3 obj _ _
6 or _ _ _ _ 7 cc _ _
7 reading _ _ _ _ 3 conj _ _
8 . _ _ _ _ 3 punct _ _
|
Conjoined subjects and objects
When the subject is a conjoined noun phrase, each of the conjuncts is attached to the predicate.
|
1 Paul _ _ _ _ 5 nsubj _ _
2 and _ _ _ _ 3 cc _ _
3 Mary _ _ _ _ 1 conj _ _
4 are _ _ _ _ 5 aux _ _
5 running _ _ _ _ 0 root _ _
6 . _ _ _ _ 5 punct _ _
|
# visual-style 5 3 nsubj color:blue
1 Paul _ _ _ _ 5 nsubj _ _
2 and _ _ _ _ 3 cc _ _
3 Mary _ _ _ _ 1 conj 5:nsubj _
4 are _ _ _ _ 5 aux _ _
5 running _ _ _ _ 0 root _ _
6 . _ _ _ _ 5 punct _ _
|
The same is true for conjoined objects.
|
1 Paul _ _ _ _ 2 nsubj _ _
2 bought _ _ _ _ 0 root _ _
3 apples _ _ _ _ 2 obj _ _
4 and _ _ _ _ 5 cc _ _
5 oranges _ _ _ _ 3 conj _ _
6 . _ _ _ _ 2 punct _ _
|
# visual-style 2 5 obj color:blue
1 Paul _ _ _ _ 2 nsubj _ _
2 bought _ _ _ _ 0 root _ _
3 apples _ _ _ _ 2 obj _ _
4 and _ _ _ _ 5 cc _ _
5 oranges _ _ _ _ 3 conj 2:obj _
6 . _ _ _ _ 2 punct _ _
|
This leads to slightly strange dependencies in the case of collective subjects or objects:
|
1 Paul _ _ _ _ 5 nsubj _ _
2 and _ _ _ _ 3 cc _ _
3 Mary _ _ _ _ 1 conj _ _
4 are _ _ _ _ 5 aux _ _
5 meeting _ _ _ _ 0 root _ _
6 . _ _ _ _ 5 punct _ _
|
# visual-style 5 3 nsubj color:blue
1 Paul _ _ _ _ 5 nsubj _ _
2 and _ _ _ _ 3 cc _ _
3 Mary _ _ _ _ 1 conj 5:nsubj _
4 are _ _ _ _ 5 aux _ _
5 meeting _ _ _ _ 0 root _ _
6 . _ _ _ _ 5 punct _ _
|
|
1 Mary _ _ _ _ 3 nsubj _ _
2 is _ _ _ _ 3 aux _ _
3 eating _ _ _ _ 0 root _ _
4 mac _ _ _ _ 3 obj _ _
5 and _ _ _ _ 6 cc _ _
6 cheese _ _ _ _ 4 conj _ _
7 . _ _ _ _ 3 punct _ _
|
# visual-style 3 6 obj color:blue
1 Mary _ _ _ _ 3 nsubj _ _
2 is _ _ _ _ 3 aux _ _
3 eating _ _ _ _ 0 root _ _
4 mac _ _ _ _ 3 obj _ _
5 and _ _ _ _ 6 cc _ _
6 cheese _ _ _ _ 4 conj 3:obj _
7 . _ _ _ _ 3 punct _ _
|
However, as the distinction between distributive and collective readings is often context-dependent, we take the simplest approach and always attach all conjuncts to the predicate.
When the subject is attached to a control or raising predicate, there is a dependency between the matrix verb and each conjunct and between the embedded verb and each conjunct.
|
1 Mary _ _ _ _ 4 nsubj _ _
2 and _ _ _ _ 3 cc _ _
3 John _ _ _ _ 1 conj _ _
4 wanted _ _ _ _ 0 root _ _
5 to _ _ _ _ 6 mark _ _
6 buy _ _ _ _ 4 xcomp _ _
7 a _ _ _ _ 8 det _ _
8 hat _ _ _ _ 6 obj _ _
9 . _ _ _ _ 4 punct _ _
|
# visual-style 4 3 nsubj color:blue
# visual-style 6 1 nsubj color:blue
# visual-style 6 3 nsubj color:blue
1 Mary _ _ _ _ 4 nsubj 6:nsubj _
2 and _ _ _ _ 3 cc _ _
3 John _ _ _ _ 1 conj 4:nsubj|6:nsubj _
4 wanted _ _ _ _ 0 root _ _
5 to _ _ _ _ 6 mark _ _
6 buy _ _ _ _ 4 xcomp _ _
7 a _ _ _ _ 8 det _ _
8 hat _ _ _ _ 6 obj _ _
9 . _ _ _ _ 4 punct _ _
|
Conjoined modifiers
Each conjunct in a conjoined modifier phrase gets attached to the governor of the modifier phrase. For example, the following phrase contains a conjoined adjectival phrase that modifies a noun. In the enhanced representation, there is an additional amod relation between the noun river and the second conjunct wide.
|
1 a _ _ _ _ 5 det _ _
2 long _ _ _ _ 5 amod _ _
3 and _ _ _ _ 4 cc _ _
4 wide _ _ _ _ 2 conj _ _
5 river _ _ _ _ 0 root _ _
|
# visual-style 5 4 amod color:blue
1 a _ _ _ _ 5 det _ _
2 long _ _ _ _ 5 amod _ _
3 and _ _ _ _ 4 cc _ _
4 wide _ _ _ _ 2 conj 5:amod _
5 river _ _ _ _ 0 root _ _
|
Relative clauses
In basic trees, relative pronouns are attached to the main predicate of the relative clause (typically with a nsubj or obj relation). In the corresponding enhanced graphs, the relative pronoun is attached to what it is referring to with the special ref relation and the governor of the relative clause is attached as an argument to the main predicate of the relative clause. Note that such graphs contain a cycle.
|
# visual-style 4 3 nsubj color:green
1 the the DET _ Definite=Def|PronType=Art 2 det _ _
2 boy boy NOUN _ Gender=Masc|Number=Sing 0 root _ _
3 who who PRON _ PronType=Rel 4 nsubj _ _
4 lived lived VERB _ Mood=Ind|Tense=Past|VerbForm=Fin 2 acl:relcl _ _
|
# visual-style 4 2 nsubj color:blue
# visual-style 2 3 ref color:blue
1 the the DET _ Definite=Def|PronType=Art 2 det _ _
2 boy boy NOUN _ Gender=Masc|Number=Sing 0 root 4:nsubj _
3 who who PRON _ PronType=Rel 2 ref _ _
4 lived lived VERB _ Mood=Ind|Tense=Past|VerbForm=Fin 2 acl:relcl _ _
|
|
# visual-style 5 3 obj color:green
1 the the DET _ Definite=Def|PronType=Art 2 det _ _
2 book book NOUN _ Gender=Neut|Number=Sing 0 root _ _
3 that that PRON _ PronType=Rel 5 obj _ _
4 I I PRON _ Number=Sing|Person=1|PronType=Prs 5 nsubj _ _
5 read read VERB _ Mood=Ind|Number=Sing|Person=1|Tense=Pres|VerbForm=Fin 2 acl:relcl _ _
|
# visual-style 5 2 obj color:blue
# visual-style 2 3 ref color:blue
1 the the DET _ Definite=Def|PronType=Art 2 det _ _
2 book book NOUN _ Gender=Neut|Number=Sing 0 root 5:obj _
3 that that PRON _ PronType=Rel 2 ref _ _
4 I I PRON _ Number=Sing|Person=1|PronType=Prs 5 nsubj _ _
5 read read VERB _ Mood=Ind|Number=Sing|Person=1|Tense=Pres|VerbForm=Fin 2 acl:relcl _ _
|
The enhanced relations are semantic rather than syntactic. In case marking languages the enhanced dependencies may link verb dependents that are not in the expected morphological case, required by surface syntax. In the following Czech example, the relative modifier phrase v němž “in which” is obligatorily in the locative case form (Case=Loc). If it were a main clause, the referent dům “house” would have to be in locative too: v domě “in house”. However, here it is in the nominative (Case=Nom), and the enhanced dependency obl going to a nominative dependent is something we would not expect to see, given the morpho-syntactic rules of the language.
|
# visual-style 5 4 obl color:green
1 dům house NOUN _ Animacy=Inan|Case=Nom|Gender=Masc|Number=Sing 0 root _ _
2 , , PUNCT _ _ 5 punct _ _
3 v in ADP _ _ 4 case _ _
4 němž that PRON _ Case=Loc|Gender=Masc|Number=Sing|PronType=Rel 5 obl _ _
5 žijeme live VERB _ Aspect=Imp|Mood=Ind|Number=Plur|Person=3|Tense=Pres|VerbForm=Fin|Voice=Act 1 acl:relcl _ _
|
# visual-style 5 1 obl color:blue
# visual-style 1 4 ref color:blue
1 dům house NOUN _ Animacy=Inan|Case=Nom|Gender=Masc|Number=Sing 0 root 5:obl _
2 , , PUNCT _ _ 5 punct _ _
3 v in ADP _ _ 4 case _ _
4 němž that PRON _ Case=Loc|Gender=Masc|Number=Sing|PronType=Rel 1 ref _ _
5 žijeme live VERB _ Aspect=Imp|Mood=Ind|Number=Plur|Person=3|Tense=Pres|VerbForm=Fin|Voice=Act 1 acl:relcl _ _
|
The relative element does not always depend directly on the predicate of the relative clause. It may be embedded deeper as in the following example.
|
# visual-style 5 4 det color:green
1 muž man NOUN _ Animacy=Anim|Case=Nom|Gender=Masc|Number=Sing 0 root _ _
2 , , PUNCT _ _ 6 punct _ _
3 v in ADP _ _ 5 case _ _
4 jehož whose DET _ Gender[psor]=Masc|Number[psor]=Plur|Poss=Yes|PronType=Rel 5 det _ _
5 domě house NOUN _ Animacy=Inan|Case=Loc|Gender=Masc|Number=Sing 6 obl _ _
6 žijeme live VERB _ Aspect=Imp|Mood=Ind|Number=Plur|Person=3|Tense=Pres|VerbForm=Fin|Voice=Act 1 acl:relcl _ _
|
# visual-style 5 1 nmod color:blue
# visual-style 1 4 ref color:blue
1 muž man NOUN _ Animacy=Anim|Case=Nom|Gender=Masc|Number=Sing 0 root 5:nmod _
2 , , PUNCT _ _ 6 punct _ _
3 v in ADP _ _ 5 case _ _
4 jehož whose DET _ Gender[psor]=Masc|Number[psor]=Plur|Poss=Yes|PronType=Rel 1 ref _ _
5 domě house NOUN _ Animacy=Inan|Case=Loc|Gender=Masc|Number=Sing 6 obl _ _
6 žijeme live VERB _ Aspect=Imp|Mood=Ind|Number=Plur|Person=3|Tense=Pres|VerbForm=Fin|Voice=Act 1 acl:relcl _ _
|
Case Information
Adding prepositions (or case information) to the relation name of non-core dependents often makes it possible to disambiguate its semantic role. We therefore augment nmod, obl, acl and advcl relation labels with the preposition or the case of the modifier.
|
# visual-style 2 5 nmod color:green
1 the _ _ _ _ 2 det _ _
2 house _ _ _ _ 0 root _ _
3 on _ _ _ _ 5 case _ _
4 the _ _ _ _ 5 det _ _
5 hill _ _ _ _ 2 nmod _ _
|
# visual-style 2 5 nmod:on color:blue
1 the _ _ _ _ 2 det _ _
2 house _ _ _ _ 0 root _ _
3 on _ _ _ _ 5 case _ _
4 the _ _ _ _ 5 det _ _
5 hill _ _ _ _ 2 nmod:on _ _
|
|
# visual-style 2 5 obl color:green
# visual-style 2 7 advcl color:green
1 He _ _ _ _ 2 det _ _
2 went _ _ _ _ 0 root _ _
3 to _ _ _ _ 5 case _ _
4 the _ _ _ _ 5 det _ _
5 dinner _ _ _ _ 2 obl _ _
6 after _ _ _ _ 7 mark _ _
7 leaving _ _ _ _ 2 advcl _ _
8 work _ _ _ _ 7 obj _ _
9 . _ _ _ _ 2 punct _ _
|
# visual-style 2 5 obl:to color:blue
# visual-style 2 7 advcl:after color:blue
1 He _ _ _ _ 2 det _ _
2 went _ _ _ _ 0 root _ _
3 to _ _ _ _ 5 case _ _
4 the _ _ _ _ 5 det _ _
5 dinner _ _ _ _ 2 obl:to _ _
6 after _ _ _ _ 7 mark _ _
7 leaving _ _ _ _ 2 advcl:after _ _
8 work _ _ _ _ 7 obj _ _
9 . _ _ _ _ 2 punct _ _
|
|
# visual-style 2 4 nmod color:green
# text = the destruction of the city
1 die the DET _ Case=Gen 2 det _ _
2 Zerstörung destruction NOUN _ Case=Nom 0 root _ _
3 der the DET _ Case=Gen 4 det _ _
4 Stadt city NOUN _ Case=Gen 2 nmod _ _
|
# visual-style 2 4 nmod:gen color:blue
# text = the destruction of the city
1 die the DET _ Case=Gen 2 det _ _
2 Zerstörung destruction NOUN _ Case=Nom 0 root _ _
3 der the DET _ Case=Gen 4 det _ _
4 Stadt city NOUN _ Case=Gen 2 nmod:gen _ _
|
Additional enhancements
Some postprocessing steps such as demoting light nouns that behave like quantificational determiners (as, for example, described in Schuster and Manning (2016)) can improve the usability of the dependency graphs for downstream applications. However, as most of these additions are highly language-specific, we do not provide any universal guidelines for such a representation and anything beyond the above additions is not part of the UD standard and should not be added to the officially released treebanks.