Stanford typed dependencies manual
Marie-Catherine de Marneffe and Christopher D. Manning
September 2008
Revised for the Stanford Parser v. 3.5.2 in April 2015
Please note that this manual describes the original Stanford Dependencies representation. As of version 3.5.2 the default representation output by the Stanford Parser and Stanford CoreNLP is the new Universal Dependencies (UD) representation. Take a look at the Universal Dependencies documentation at http://universaldependencies.github.com/docs/ for a description of UD relations.
1 Introduction
The Stanford typed dependencies representation was designed to provide a simple description of the grammatical relationships in a sentence that can easily be understood and effectively used by people without linguistic expertise who want to extract textual relations. In particular, rather than the phrase structure representations that have long dominated in the computational linguistic community, it repre- sents all sentence relationships uniformly as typed dependency relations. That is, as triples of a relation between pairs of words, such as “the subject of distributes is Bell.” Our experience is that this simple, uniform representation is quite accessible to non-linguists thinking about tasks involving information extraction from text and is effective in relation extraction applications.
Here is an example sentence:
Bell, based in Los Angeles, makes and distributes electronic, computer and building prod- ucts.
For this sentence, the Stanford Dependencies (SD) representation is:
nsubj(makes-8, Bell-1) nsubj(distributes-10, Bell-1) vmod(Bell-1, based-3) nn(Angeles-6, Los-5)
prep in(based-3, Angeles-6) root(ROOT-0, makes-8)
conj and(makes-8, distributes-10) amod(products-16, electronic-11) conj and(electronic-11, computer-13) amod(products-16, computer-13) conj and(electronic-11, building-15) amod(products-16, building-15) dobj(makes-8, products-16) dobj(distributes-10, products-16)
1
makes
conj_and
nsubj
Bell
partmod
based
prep_in
Angeles
nn
Los
distributes
nsubj
dobj
amod
products
amod
electronic
conj_and
amod conj_and
building
dobj
computer
Figure 1: Graphical representation of the Stanford Dependencies for the sentence: Bell, based in Los Angeles, makes and distributes electronic, computer and building products.
These dependencies map straightforwardly onto a directed graph representation, in which words in the sentence are nodes in the graph and grammatical relations are edge labels. Figure 1 gives the graph representation for the example sentence above.
Document overview: This manual provides documentation for the set of dependencies defined for English. There is also a Stanford Dependency representation available for Chinese, but it is not further discussed here. Starting in 2014, there has been work to extend Stanford Dependencies to be generally applicable cross-linguistically. Initial work appeared in de Marneffe et al. (2014), and the current pro- posal for Universal Dependencies (UD) can be found at http://universaldependencies.github.io/docs/. This work is not (yet) reflected in this manual or in our software. For SD, Section 2 of the manual de- fines the grammatical relations and the taxonomic hierarchy over them appears in section 3. This is then followed by a description of the several variant dependency representations available, aimed at differ- ent use cases (section 4), some details of the software available for generating Stanford Dependencies (section 5), and references to further discussion and use of the SD representation (section 6).
2 Definitions of the Stanford typed dependencies
The current representation contains approximately 50 grammatical relations (depending slightly on the options discussed in section 4). The dependencies are all binary relations: a grammatical relation holds between a governor (also known as a regent or a head) and a dependent. The grammatical relations are defined below, in alphabetical order according to the dependency’s abbreviated name (which appears in the parser output). The definitions make use of the Penn Treebank part-of-speech tags and phrasal labels.
acomp: adjectival complement
An adjectival complement of a verb is an adjectival phrase which functions as the complement (like an object of the verb).
2
acomp nsubj advmod
She looks very beautiful
advcl: adverbial clause modifier
An adverbial clause modifier of a VP or S is a clause modifying the verb (temporal clause, consequence, conditional clause, purpose clause, etc.).
“The accident happened as the night was falling”
“If you know who did it, you should tell the teacher” “He talked to him in order to secure the account”
advcl(happened, falling) advcl(tell, know) advcl(talked, secure)
advmod: adverb modifier
An adverb modifier of a word is a (non-clausal) adverb or adverb-headed phrase that serves to modify the meaning of the word.
“Genetically modified food” advmod(modified, genetically) “less often” advmod(often, less)
agent: agent
An agent is the complement of a passive verb which is introduced by the preposition “by” and does the action. This relation only appears in the collapsed dependencies, where it can replace prep by, where appropriate. It does not appear in basic dependencies output.
“The man has been killed by the police” agent(killed, police) “Effects caused by the protein are important” agent(caused, protein)
amod: adjectival modifier
An adjectival modifier of an NP is any adjectival phrase that serves to modify the meaning of the NP.
“Sam eats red meat”
“Sam took out a 3 million dollar loan” “Sam took out a $ 3 million loan”
amod(meat, red) amod(loan, dollar) amod(loan, $)
appos: appositional modifier
An appositional modifier of an NP is an NP immediately to the right of the first NP that serves to define or modify that NP. It includes parenthesized examples, as well as defining abbreviations in one of these structures.
appos
Sam , my brother , arrived
appos
Bill ( John ’s cousin )
appos
The Australian Broadcasting Corporation ( ABC )
aux: auxiliary
An auxiliary of a clause is a non-main verb of the clause, e.g., a modal auxiliary, or a form of “be”, “do” or “have” in a periphrastic tense.
3
aux aux Reagan has died He should leave
auxpass: passive auxiliary
A passive auxiliary of a clause is a non-main verb of the clause which contains the passive information.
“Kennedy has been killed” auxpass(killed, been) aux(killed,has)
“Kennedy was/got killed” auxpass(killed, was/got)
cc: coordination
A coordination is the relation between an element of a conjunct and the coordinating conjunction word of the conjunct. (Note: different dependency grammars have different treatments of coordination. We take one conjunct of a conjunction (normally the first) as the head of the conjunction.) A conjunction may also appear at the beginning of a sentence. This is also called a cc, and dependent on the root predicate of the sentence.
“Bill is big and honest”
“They either ski or snowboard” “And then we left.”
cc(big, and) cc(ski, or) cc(left, And)
ccomp: clausal complement
A clausal complement of a verb or adjective is a dependent clause with an internal subject which func- tions like an object of the verb, or adjective. Clausal complements for nouns are limited to complement clauses with a subset of nouns like “fact” or “report”. We analyze them the same (parallel to the analysis of this class as “content clauses” in Huddleston and Pullum 2002). Such clausal complements are usually finite (though there are occasional remnant English subjunctives).
“He says that you like to swim”
“I am certain that he did it”
“I admire the fact that you are honest”
ccomp(says, like) ccomp(certain, did) ccomp(fact, honest)
conj: conjunct
A conjunct is the relation between two elements connected by a coordinating conjunction, such as “and”, “or”, etc. We treat conjunctions asymmetrically: The head of the relation is the first conjunct and other conjunctions depend on it via the conj relation.
“Bill is big and honest” conj(big, honest) “They either ski or snowboard” conj(ski, snowboard)
cop: copula
A copula is the relation between the complement of a copular verb and the copular verb. (We normally take a copula as a dependent of its complement; see the discussion in section 4.)
“Bill is big” cop(big, is) “Bill is an honest man” cop(man, is)
4
csubj: clausal subject
A clausal subject is a clausal syntactic subject of a clause, i.e., the subject is itself a clause. The governor of this relation might not always be a verb: when the verb is a copular verb, the root of the clause is the complement of the copular verb. In the two following examples, “what she said” is the subject.
“What she said makes sense” csubj(makes, said) “What she said is not true” csubj(true, said)
csubjpass: clausal passive subject
A clausal passive subject is a clausal syntactic subject of a passive clause. In the example below, “that she lied” is the subject.
“That she lied was suspected by everyone” csubjpass(suspected, lied)
dep: dependent
A dependency is labeled as dep when the system is unable to determine a more precise dependency relation between two words. This may be because of a weird grammatical construction, a limitation in the Stanford Dependency conversion software, a parser error, or because of an unresolved long distance dependency.
“Then, as if to show that he could, . . . ” dep(show, if) det: determiner
A determiner is the relation between the head of an NP and its determiner. “The man is here” det(man, the)
“Which book do you prefer?” det(book, which)
discourse: discourse element
This is used for interjections and other discourse particles and elements (which are not clearly linked to the structure of the sentence, except in an expressive way). We generally follow the guidelines of what the Penn Treebanks count as an INTJ. They define this to include: interjections (oh, uh-huh, Welcome), fillers (um, ah), and discourse markers (well, like, actually, but not you know).
discourse
Iguazu is in Argentina 🙂
dobj: direct object
The direct object of a VP is the noun phrase which is the (accusative) object of the verb.
“She gave me a raise” dobj(gave, raise) “They win the lottery” dobj(win, lottery)
expl: expletive
This relation captures an existential “there”. The main verb of the clause is the governor.
5
“There is a ghost in the room” expl(is, There)
goeswith: goes with
This relation links two parts of a word that are separated in text that is not well edited. We follow the treebank: The GW part is the dependent and the head is in some sense the “main” part, often the second part.
goeswith
They come here with out legal permission
iobj: indirect object
The indirect object of a VP is the noun phrase which is the (dative) object of the verb.
“She gave me a raise” iobj(gave, me)
mark: marker
A marker is the word introducing a finite clause subordinate to another clause. For a complement clause, this will typically be “that” or “whether”. For an adverbial clause, the marker is typically a preposition like “while” or “although”. The mark is a dependent of the subordinate clause head.
mark
Forces engaged in fighting after insurgents attacked mark
He says that you like to swim
mwe: multi-word expression
The multi-word expression (modifier) relation is used for certain multi-word idioms that behave like a single function word. It is used for a closed set of dependencies between words in common multi-word expressions for which it seems difficult or unclear to assign any other relationships. At present, this relation is used inside the following expressions: rather than, as well as, instead of, such as, because of, instead of, in addition to, all but, such as, because of, instead of, due to. The boundaries of this class are unclear; it could grow or shrink a little over time.
“I like dogs as well as cats” mwe(well, as) mwe(well, as)
“He cried because of you” mwe(of, because) neg: negation modifier
The negation modifier is the relation between a negation word and the word it modifies. “Bill is not a scientist” neg(scientist, not)
“Bill doesn’t drive” neg(drive, n’t)
nn: noun compound modifier
A noun compound modifier of an NP is any noun that serves to modify the head noun. (Note that in the current system for dependency extraction, all nouns modify the rightmost noun of the NP – there is no intelligent noun compound analysis. This is likely to be fixed once the Penn Treebank represents the branching structure of NPs.)
6
“Oil price futures” nn(futures, oil) nn(futures, price)
npadvmod: noun phrase as adverbial modifier
This relation captures various places where something syntactically a noun phrase (NP) is used as an ad- verbial modifier in a sentence. These usages include: (i) a measure phrase, which is the relation between the head of an ADJP/ADVP/PP and the head of a measure phrase modifying the ADJP/ADVP; (ii) noun phrases giving an extent inside a VP which are not objects; (iii) financial constructions involving an ad- verbial or PP-like NP, notably the following construction $5 a share, where the second NP means “per share”; (iv) floating reflexives; and (v) certain other absolutive NP constructions. A temporal modifier (tmod) is a subclass of npadvmod which is distinguished as a separate relation.
“The director is 65 years old”
“6 feet long”
“Shares eased a fraction”
“IBM earned $ 5 a share”
“The silence is itself significant”
“90% of Australians like him, the most of any country”
npadvmod(old, years) npadvmod(long, feet) npadvmod(eased, fraction) npadvmod($, share) npadvmod(significant, itself) npadvmod(like, most)
nsubj: nominal subject
A nominal subject is a noun phrase which is the syntactic subject of a clause. The governor of this relation might not always be a verb: when the verb is a copular verb, the root of the clause is the complement of the copular verb, which can be an adjective or noun.
“Clinton defeated Dole” nsubj(defeated, Clinton) “The baby is cute” nsubj(cute, baby)
nsubjpass: passive nominal subject
A passive nominal subject is a noun phrase which is the syntactic subject of a passive clause.
“Dole was defeated by Clinton” nsubjpass(defeated, Dole)
num: numeric modifier
A numeric modifier of a noun is any number phrase that serves to modify the meaning of the noun with a quantity.
“Sam ate 3 sheep”
“Sam spent forty dollars” “Sam spent $ 40”
num(sheep, 3) num(dollars, 40) num($, 40)
number: element of compound number
An element of compound number is a part of a number phrase or currency amount. We regard a number as a specialized kind of multi-word expression.
7
“I have four thousand sheep” number(thousand, four) “I lost $ 3.2 billion” number(billion, 3.2)
parataxis: parataxis
The parataxis relation (from Greek for “place side by side”) is a relation between the main verb of a clause and other sentential elements, such as a sentential parenthetical, a clause after a “:” or a “;”, or two sentences placed side by side without any explicit coordination or subordination.
“The guy, John said, left early in the morning” parataxis(left, said) “Let’s face it we’re annoyed” parataxis(Let, annoyed)
pcomp: prepositional complement
This is used when the complement of a preposition is a clause or prepositional phrase (or occasionally, an adverbial phrase). The prepositional complement of a preposition is the head of a clause following the preposition, or the preposition head of the following PP.
“We have no information on whether users are at risk” pcomp(on, are)
“They heard about you missing classes” pcomp(about, missing)
pobj: object of a preposition
The object of a preposition is the head of a noun phrase following the preposition, or the adverbs “here” and “there”. (The preposition in turn may be modifying a noun, verb, etc.) Unlike the Penn Treebank, we here define cases of VBG quasi-prepositions like “including”, “concerning”, etc. as instances of pobj. (The preposition can be tagged a FW for “pace”, “versus”, etc. It can also be called a CC – but we don’t currently handle that and would need to distinguish from conjoined prepositions.) In the case of preposition stranding, the object can precede the preposition (e.g., “What does CPR stand for?”).
“I sat on the chair” pobj(on, chair)
poss: possession modifier
The possession modifier relation holds between the head of an NP and its possessive determiner, or a genitive ’s complement.
“their offices” poss(offices, their) “Bill’s clothes” poss(clothes, Bill)
possessive: possessive modifier
The possessive modifier relation appears between the head of an NP and the genitive ’s.
“Bill’s clothes” possessive(John, ’s)
preconj: preconjunct
A preconjunct is the relation between the head of an NP and a word that appears at the beginning brack- eting a conjunction (and puts emphasis on it), such as “either”, “both”, “neither”).
8
“Both the boys and the girls are here” preconj(boys, both)
predet: predeterminer
A predeterminer is the relation between the head of an NP and a word that precedes and modifies the meaning of the NP determiner.
“All the boys are here” predet(boys, all)
prep: prepositional modifier
A prepositional modifier of a verb, adjective, or noun is any prepositional phrase that serves to modify the meaning of the verb, adjective, noun, or even another prepositon. In the collapsed representation, this is used only for prepositions with NP complements.
“I saw a cat in a hat”
“I saw a cat with a telescope” “He is responsible for meals”
prep(cat, in) prep(saw, with) prep(responsible, for)
prepc: prepositional clausal modifier
In the collapsed representation (see section 4), a prepositional clausal modifier of a verb, adjective, or noun is a clause introduced by a preposition which serves to modify the meaning of the verb, adjective, or noun.
“He purchased it without paying a premium” prepc without(purchased, paying) prt: phrasal verb particle
The phrasal verb particle relation identifies a phrasal verb, and holds between the verb and its particle. “They shut down the station” prt(shut, down)
punct: punctuation
This is used for any piece of punctuation in a clause, if punctuation is being retained in the typed depen- dencies. By default, punctuation is not retained in the output.
“Go home!” punct(Go, !)
quantmod: quantifier phrase modifier
A quantifier modifier is an element modifying the head of a QP constituent. (These are modifiers in complex numeric quantifiers, not other types of “quantification”. Quantifiers like “all” become det.)
“About 200 people came to the party” quantmod(200, About)
rcmod: relative clause modifier
A relative clause modifier of an NP is a relative clause modifying the NP. The relation points from the head noun of the NP to the head of the relative clause, normally a verb.
9
“I saw the man you love” rcmod(man, love)
“I saw the book which you bought” rcmod(book,bought)
ref: referent
A referent of the head of an NP is the relative word introducing the relative clause modifying the NP.
“I saw the book which you bought” ref (book, which)
root: root
The root grammatical relation points to the root of the sentence. A fake node “ROOT” is used as the governor. The ROOT node is indexed with “0”, since the indexation of real words in the sentence starts at 1.
“I love French fries.” root(ROOT, love) “Bill is an honest man” root(ROOT, man)
tmod: temporal modifier
A temporal modifier (of a VP, NP, or an ADJP is a bare noun phrase constituent that serves to modify the meaning of the constituent by specifying a time. (Other temporal modifiers are prepositional phrases and are introduced as prep.)
“Last night, I swam in the pool” tmod(swam, night)
vmod: reduced non-finite verbal modifier
A reduced non-finite verbal modifier is a participial or infinitive form of a verb heading a phrase (which may have some arguments, roughly like a VP). These are used to modify the meaning of an NP or another verb. They are not core arguments of a verb or full finite relative clauses.
“Points to establish are . . . ”
“I don’t have anything to say to you”
“Truffles picked during the spring are tasty”
“Bill tried to shoot, demonstrating his incompetence”
vmod(points, establish) vmod(anything, say) vmod(truffles, picked) vmod(shoot, demonstrating)
xcomp: open clausal complement
An open clausal complement (xcomp) of a verb or an adjective is a predicative or clausal complement without its own subject. The reference of the subject is necessarily determined by an argument external to the xcomp (normally by the object of the next higher clause, if there is one, or else by the subject of the next higher clause. These complements are always non-finite, and they are complements (arguments of the higher verb or adjective) rather than adjuncts/modifiers, such as a purpose clause. The name xcomp is borrowed from Lexical-Functional Grammar.
“He says that you like to swim”
“I am ready to leave”
“Sue asked George to respond to her offer”
xcomp(like, swim) xcomp(ready, leave) xcomp(ask, respond)
10
“I consider him a fool” xcomp(consider, fool) “I consider him honest” xcomp(consider, honest)
xsubj: controlling subject
A controlling subject is the relation between the head of a open clausal complement (xcomp) and the external subject of that clause. This is an additional dependency, not a basic depedency.
“Tom likes to eat fish” xsubj(eat, Tom)
3 Hierarchy of typed dependencies
The grammatical relations defined in the above section stand in a hierarchy. The most generic grammat- ical relation, dependent (dep), will be used when a more precise relation in the hierarchy does not exist or cannot be retrieved by the system.
root – root
dep – dependent
aux – auxiliary
auxpass – passive auxiliary cop – copula
arg – argument agent – agent
comp – complement
acomp – adjectival complement
ccomp – clausal complement with internal subject xcomp – clausal complement with external subject obj – object
dobj – direct object
iobj – indirect object
pobj – object of preposition
subj – subject
nsubj – nominal subject
nsubjpass – passive nominal subject csubj – clausal subject
csubjpass – passive clausal subject
cc – coordination
conj – conjunct
expl – expletive (expletive “there”) mod – modifier
amod – adjectival modifier appos – appositional modifier advcl – adverbial clause modifier det – determiner
predet – predeterminer
11
preconj – preconjunct
vmod – reduced, non-finite verbal modifier mwe – multi-word expression modifier
mark – marker (word introducing an advcl or ccomp advmod – adverbial modifier
neg – negation modifier
rcmod – relative clause modifier
quantmod – quantifier modifier
nn – noun compound modifier
npadvmod – noun phrase adverbial modifier
tmod – temporal modifier num – numeric modifier
number – element of compound number prep – prepositional modifier
poss – possession modifier
possessive – possessive modifier (’s)
prt – phrasal verb particle parataxis – parataxis
goeswith – goes with punct – punctuation
ref – referent
sdep – semantic dependent
xsubj – controlling subject
4 Different styles of dependency representation
Five variants of the typed dependency representation are available in the dependency extraction system provided with the Stanford parser. The representations follow the same format. In the plain text format, a dependency is written as abbreviated relation name(governor, dependent) where the governor and the dependent are words in the sentence to which a number indicating the position of the word in the sentence is appended.1 The parser also provides an XML format which captures the same information. The differences between the five formats are that they range from a more surface-oriented representation, where each token appears as a dependent in a tree, to a more semantically interpreted representation where certain word relationships, such as prepositions, are represented as dependencies, and the set of dependencies becomes a possibly cyclic graph.
4.1 Basic
The basic typed dependencies use the dependencies defined in section 2, and form 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-0” token. For the sentence, “Bell, a company which is based in LA, makes and distributes computer products.”, the basic typed dependencies will be:
1In some cases, an apostrophe is added after the word position number: see section 4.6 for more details.
12
nsubj(makes-11, Bell-1)
det(company-4, a-3)
appos(Bell-1, company-4)
nsubjpass(based-7, which-5)
auxpass(based-7, is-6)
rcmod(company-4, based-7)
prep(based-7, in-8)
pobj(in-8, LA-9)
root(ROOT-0, makes-11)
cc(makes-11, and-12)
conj(makes-11, distributes-13)
nn(products-15, computer-14)
dobj(makes-11, products-15)
4.2 Collapsed dependencies
In the collapsed representation, dependencies involving prepositions, conjuncts, as well as information about the referent of relative clauses are collapsed to get direct dependencies between content words. This “collapsing” is often useful in simplifying patterns in relation extraction applications. For instance, the dependencies involving the preposition “in” in the above example will be collapsed into one single relation:
prep(based-7, in-8)
pobj(in-8, LA-9)
will become
prep in(based-7, LA-9)
Moreover, additional dependencies are considered, even ones that break the tree structure (turning the dependency structure into a directed graph). So in the above example, the following relation will be added:
ref(company-4, which-5)
That relation does not appear in the basic representation since it creates a cycle with the rcmod and nsubjpass relations. Relations that break the tree structure are the ones taking into account elements from relative clauses and their antecedents (as shown in this example), the controlling (xsubj) relations, and the (pobj) relation in the case of preposition stranding.
English has some very common multi-word constructions that function like prepositions. These are also collapsed as prepositional relations. At the moment, the system handles the multi-word prepositions listed in Tables 1 and 2.
The same happens for dependencies involving conjunction:
cc(makes-11, and-12)
conj(makes-11, distributes-13)
become
conj and(makes-11, distributes-13)
A few variant conjunctions for “and (not)” are collapsed together in this representation as shown in Table 3.
13
according to across from ahead of along with alongside of apart from as for
as from as of
as per
as to
aside from away from based on because of close by close to contrary to
compared to compared with due to depending on except for exclusive of contrary to followed by inside of
instead of irrespective of next to
near to
off of
out of
outside of owing to preliminary to
preparatory to previous to prior to pursuant to regardless of subsequent to such as thanks to together with
Table 1: List of two-word prepositions that the system can collapse.
by means of
in accordance with in addition to
in case of in front of in lieu of
in place of
in spite of
on account of
on behalf of with respect to on top of
with regard to
Table 2: List of three-word prepositions that the system can collapse.
The information about the antecedent of the relative clause (ref(company-4, which-5)) will serve to expand the following dependency:
nsubjpass(based-7, which-5)
becomes
nsubjpass(based-7, company-4)
In the end the collapsed dependencies that the system gives you for the sentence are:
nsubj(makes-11, Bell-1)
det(company-4, a-3)
appos(Bell-1, company-4)
nsubjpass(based-7, company-4)
auxpass(based-7, is-6)
rcmod(company-4, based-7)
prep_in(based-7, LA-9)
root(ROOT-0, makes-11)
conj_and(makes-11, distributes-13)
nn(products-15, computer-14)
dobj(makes-11, products-15)
Mapped to
conj and as well as conj negcc but not
not to mention instead of
but also rather than
&
but rather
Table 3: Mapping of select conjunct relations in the collapsed representation.
14
4.3 Collapsed dependencies with propagation of conjunct dependencies
When there is a conjunction, you can also get propagation of the dependencies involving the conjuncts. In the sentence here, this propagation should add two dependencies to the collapsed representation; due to the conjunction between the verbs “makes” and “distributes”, the subject and object relations that exist on the first conjunct (“makes”) should be propagated to the second conjunct (“distributes”):
nsubj(distributes-13, Bell-1)
dobj(distributes-13, products-15)
However, at present, our converter handles this imperfectly and only generates the first of these two dependencies (in general, it is hard to determine if object dependencies should be distributed or not in English).
Since this representation is an extension of the collapsed dependencies, it does not guarantee a tree structure.
4.4 Collapsed dependencies preserving a tree structure
In this representation, dependencies which do not preserve the tree structure are omitted. As explained above, this concerns relations between elements of a relative clause and its antecedent, as well as the controlling subject relation (xsubj), and the object of preposition (pobj) in the case of preposition strand- ing. This also does not allow propagation of conjunct dependencies. In our example, the dependencies in this representation are actually identical to the ones in the collapsed representation:
nsubj(makes-11, Bell-1)
det(company-4, a-3)
appos(Bell-1, company-4)
nsubjpass(based-7, which-5)
auxpass(based-7, is-6)
rcmod(company-4, based-7)
prep_in(based-7, LA-9)
root(ROOT-0, makes-11)
conj_and(makes-11, distributes-13)
nn(products-15, computer-14)
dobj(makes-11, products-15)
4.5 Non-collapsed dependencies
This representation gives the basic dependencies as well as the extra ones (which break the tree structure), without any collapsing or propagation of conjuncts. There are options to get the extra dependencies sep- arated from the basic dependencies (see section 5). At print time, the dependencies in this representation can thus look as follows:
nsubj(makes-11, Bell-1)
det(company-4, a-3)
appos(Bell-1, company-4)
nsubjpass(based-7, which-5)
auxpass(based-7, is-6)
rcmod(company-4, based-7)
15
prep(based-7, in-8)
pobj(in-8, LA-9)
root(ROOT-0, makes-11)
cc(makes-11, and-12)
conj(makes-11, distributes-13)
nn(products-15, computer-14)
dobj(makes-11, products-15)
======
ref(company-4, which-5)
4.6 Alteration of the sentence semantics
In some cases, collapsing relations introduces a slight alteration of the semantics of the sentence. In all the representation styles involving collapsing, the two following phenomena may appear.
Introduction of copy nodes marked with an apostrophe. A copy node will be introduced in the case of PP conjunction as in “Bill went over the river and through the woods”. In this example, the two prepo- sitions “over” and “through” are conjoined and governed by the verb “went”. To avoid disjoint subgraphs when collapsing the relations (preposition and conjunction), sentences like this are transformed into VP coordination, which requires making a copy of the word “went”. A copy node will be marked with one or more apostrophes in the plain text output or by a copy attribute in the XML output. This gives the following representation, which corresponds to a sentence like “Bill went over the river and went through the woods”:
prep over(went-2, river-5)
prep through(went-2’, woods-10) conj and(went-2, went-2’)
Distortion in governors of preposition modifiers. Another instance where collapsing sacrifices some linguistic fidelity is the case of preposition modifiers. When turning the preposition into a relation, the preposition does not appear as a word of the sentence anymore. Therefore preposition modifiers become dependent on the head of the clause in which they appear, and not on the preposition itself. In He left his office just before lunch time, just will be an adverbial modifier of the verb left. This induces some distortion in the exact semantics of the sentence.
4.7 The treatment of copula verbs
The design philosophy of SD has been to maximize dependencies between content words, and so we normally regard a copula verb like be as an auxiliary modifier, even when its complement is an adjective or predicative noun (see the references in section 6 for more discussion and motivation). However, some people do not like this because then the head of some sentences is no longer a verb. In the dependency conversion software, you can ask for the copula to remain the head when its complement is an adjective or noun by giving the flag -makeCopulaHead. Uses of the verb be as in auxiliary in passives and progressives will still be treated as a non-head auxiliary.
16
Connected?
All tokens are nodes? Rooted?
Acyclic
Multigraph
Tree
Self-loops? Projective?
basic collapsed Yes Yes
Yes No
Yes Yes
Yes No No No Yes No No No No No
CCprocessed collapsed tree Yes Yes
No No
Yes Yes
No Yes No No No Yes No No No No
basic plus extras Yes
Yes
Yes
Yes Yes No No No
Table 4: Graph-theoretic properties of different versions of SD graphs.
4.8 Comparison of the representation styles
To facilitate comparison, the table below shows the dependencies for the four variants for the exam- ple sentence “Bell, a company which is based in LA, makes and distributes computer products”. The “non-collapsed” variant (see section 4.5) contains all the relations in the “basic” variant plus one extra dependency: ref(company-4, which-5).
basic
nsubj(makes, Bell)
det(company, a) appos(Bell, company) nsubjpass(based, which) auxpass(based, is) rcmod(company, based) prep(based, in)
pobj(in, LA) root(ROOT, makes) cc(makes, and)
conj(makes, distributes) nn(products, computer) dobj(makes, products)
collapsed
nsubj(makes, Bell)
det(company, a) appos(Bell, company) nsubjpass(based, company) auxpass(based, is) rcmod(company, based)
prep in(based, LA)
root(ROOT, makes)
conj and(makes, distributes)
nn(products, computer) dobj(makes, products)
propagation
nsubj(makes, Bell) nsubj(distributes, Bell) det(company, a) appos(Bell, company) nsubjpass(based, company) auxpass(based, is) rcmod(company, based)
prep in(based, LA)
root(ROOT, makes)
conj and(makes, distributes)
nn(products, computer) dobj(makes, products)
collapsed tree
nsubj(makes, Bell)
det(company, a) appos(Bell, company) nsubjpass(based, which) auxpass(based, is-) rcmod(company, based)
prep in(based, LA)
root(ROOT, makes)
conj and(makes, distributes)
nn(products, computer) dobj(makes, products)
4.9 Graph-theoretic properties
Dependency syntax representations are naturally thought of as “directed graphs”, but some of the precise formal properties of Stanford dependencies graphs can surprise people, so here we summarize the main graph-theoretic properties. The unusual properties are all things that occur with relative clauses and/or questions. A summary of the properties is shown in table 4. To cover the collapsed representations, you need what is commonly referred to as a labeled, directed multigraph.
A non-standard property of graphs applicable to dependencies is projectivity, which arises from the fact that depedency tree nodes possess a linear order, given by their order of occurrence in sentences. A dependency tree is projective if all arcs are projective, and an arc from head wi to dependent wj is projective if wi is an ancestor of each word wk between wi and wj. Put more simply, if you draw the dependencies above a sentence written out in the usual way, a non-projective dependency structure will have one or more crossing lines. In current versions, no variant of Stanford Dependencies always
17
produces projective dependency trees.2 However, most trees are projective: Non-projective trees only occur occasionally in English, in structure such as questions and relative clauses.
The collapsed and CCprocessed dependencies are not a DAG. The graphs can contain small cycles between two nodes (only). These don’t seem eliminable given the current representational choices. They occur with relative clauses such as the woman who introduced you. The cycles occur once you wish to represent the referent of who. In the basic plus extras representation, you get rcmod(woman, introduced), nsubj(introduced, who), and ref (woman, who).3 In the collapsing process, ref arcs are collapsed, and so there is then a two node cycle: rcmod(woman, introduced) and nsubj(introduced, woman). These cycles can occur at the “top” of the graph when an NP is the head of the sentence, given the treatment of copula verbs (as in She is the woman who introduced me.). This used to mean that the dependency graph didn’t have a clear root. This was fixed after version 1.6.8 by explicitly adding a root arc to the representation.
There can be multiple arcs with the same label from a node. For instance, this occurs when a noun has several adjective modifiers, each of which gets an amod relation, as in its third consecutive monthly decline.
In the basic plus extras representation, a word can be the dependent of two different words. For example, a relative word will be a ref of the head of the noun phrase it modifies and will have a role in the relative clause. For example you might get both the arcs ref (researchers-5, who-6) and nsubj(studied- 7, who-6). You can even get two arcs between the same pair of words, though these normally result from bugs in the converter.
All graphs should be connected (if there are disconnected graphs, it’s a bug!). There are no self-loops in the graphs.
5 In practice
In practice, two classes can be used to get the typed dependencies of a sentence using the code in the Stanford parser (downloadable at http://nlp.stanford.edu/software/lex-parser.shtml).
⋆ edu.stanford.nlp.parser.lexparser.LexicalizedParser
If you need to parse texts and want to get different formatting options for the parse tree, you should use this class. To get the dependencies, add typedDependencies in the -outputFormat option. By default, this will give you collapsed dependencies with propagation of conjunct dependencies. If you want another representation, specify it in the -outputFormatOptions using the following commands according to the type of dependency representation you want:
basicDependencies Basic dependencies.
collapsedDependencies Collapsed dependencies (not necessarily a tree structure).
CCPropagatedDependencies Collapseddependencieswithpropagationofconjunctdependencies(not necessarily a tree structure). [This representation is the default, if no option is specified.]
treeDependencies Collapsed dependencies that preserve a tree structure.
2The basic dependencies output by early versions of the Stanford Dependencies converter did only produce projective trees. Starting with v.3.2, non-projective dependency trees are produced in appropriate places for questions and relative clauses.
3 Arguably, that third dependency should already have been represented the other way around as ref (who, woman), giving a three node cycle, but it wasn’t.
18
nonCollapsedDependencies Non-collapsed dependencies: basic dependencies as well as the extra ones which do not preserve a tree structure.
nonCollapsedDependenciesSeparated Non-collapsed dependencies where the basic dependencies are separated from the extra ones (by “======”).
You should also use the -retainTmpSubcategories option to get best performance in recognizing temporal dependencies. In the following command, file.txt contains your input sentences. (With this command-line, the parser will attempt to tokenize and sentence-break them. There are options to the parser to specify that this has already been done.) The penn option will also give you the context-free phrase structure grammar representation of the sentences.
Note that as of version 3.5.2 the default representation is the new Universal Dependencies representation. Use the -originalDependencies option to obtain original Stanford Dependencies.
Command line example:
java -mx200m edu.stanford.nlp.parser.lexparser.LexicalizedParser
-retainTmpSubcategories -originalDependencies -outputFormat
“penn,typedDependencies” -outputFormatOptions “basicDependencies”
englishPCFG.ser.gz file.txt
Java example:
LexicalizedParser lp = LexicalizedParser.loadModel(
“edu/stanford/nlp/models/lexparser/englishPCFG.ser.gz”,
“-maxLength”, “80”, “-retainTmpSubcategories”);
TreebankLanguagePack tlp = new PennTreebankLanguagePack();
// Uncomment the following line to obtain original Stanford Dependencies
// tlp.setGenerateOriginalDependencies(true);
GrammaticalStructureFactory gsf = tlp.grammaticalStructureFactory();
String[] sent = “This”, “is”, “an”, “easy”, “sentence”, “.” ;
Tree parse = lp.apply(Sentence.toWordList(sent)); GrammaticalStructure gs = gsf.newGrammaticalStructure(parse); Collection⟨TypedDependency⟩ tdl = gs.typedDependenciesCCprocessed(); System.out.println(tdl);
⋆ edu.stanford.nlp.trees.EnglishGrammaticalStructure
If you already have Penn treebank-style trees (whether hand-annotated or as output from another parser),
you can use this class to get the Stanford dependencies.
Command-line usage. Use the -treeFile option as shown in the command line example below. The options to get the different types of representation are as follows:
-basic
-collapsed
-CCprocessed
-collapsedTree
basic dependencies
collapsed dependencies (not necessarily a tree structure) collapsed dependencies with propagation of conjunct dependencies (not necessarily a tree structure) collapsed dependencies that preserve a tree structure
19
-nonCollapsed
-conllx
-originalDependencies
non-collapsed dependencies: basic dependencies as well as the extra ones which do not preserve a tree structure dependencies printed out in CoNLL X (CoNLL 2006) format output the original Stanford Dependencies instead of
the new Universal Dependencies.
If you want the non-collapsed version of the dependencies where the basic ones are separated from the extra ones, add the flag -extraSep. This will print the basic dependencies, a separator (======) and the extra dependencies. By default, punctuation dependencies are not printed. If you want them, give the option -keepPunct.
Command line example:
java edu.stanford.nlp.trees.EnglishGrammaticalStructure -treeFile
file.tree -collapsedTree -CCprocessed -keepPunct
By default, the CoNLL format retains punctuation. When the CoNLL format is used with collapsed dependencies, words of the sentences which have been collapsed into the grammatical relations (such as prepositions and conjunctions) still appear in the list of words but are given an “erased” grammatical relation:
1 Bell
2, ,,1punct 3a DTDT4det
NNP NNP 11 nsubj
4 company NN NN 7
5 which WDT WDT 0
6 is VBZ VBZ 7
7 based VBN VBN 4
8 in IN IN 0
nsubjpass erased auxpass rcmod erased prep in
9 LA
10, ,,1punct
11 makes VBZ VBZ 0 root
12 and CC CC 0 erased
NNP NNP 7
13 distributes
VBZ VBZ
11 conj and
14 15 16
computer NN products NNS . .
NN 15 NNS 11 . 11
nn dobj punct
This class
dependencies), and transform them into another representation (e.g., the CCprocessed representation). To do this, you need to pass the input file using the option -conllxFile.
You can also use this class to parse texts, but the input has to be formatted as strictly one sentence per line, and you will not be able to specify options for the parse tree output on the command line. You will only be able to specify the type of the dependencies. Use the option -sentFile instead of -treeFile. You will need to specify the parser file using the -parserFile option. You can print the parse tree by using the -parseTree option.
can read files that contain Stanford dependencies in the CoNLL format (i.e., the basic Stanford
20
Command line example:
java -mx100m edu.stanford.nlp.trees.EnglishGrammaticalStructure
-sentFile file.txt -collapsedTree -CCprocessed -parseTree -parserFile
englishPCFG.ser.gz
API usage. Basic API usage was already illustrated in the LexicalizedParser usage above. If you have a Tree object, the steps for converting it to dependencies are like this:
// One time setup
TreebankLanguagePack tlp = new PennTreebankLanguagePack();
// Uncomment the following line to obtain original Stanford Dependencies // tlp.setGenerateOriginalDependencies(true); GrammaticalStructureFactory gsf = tlp.grammaticalStructureFactory();
// For each Tree
Tree parseTree; // assumed to come from a treebank or parser GrammaticalStructure gs = gsf.newGrammaticalStructure(parse); Collection⟨TypedDependency⟩ tdl = gs.typedDependencies();
The PennTreebankLanguagePack vends an EnglishGrammaticalStructureFactory. The only common option to pass in when creating one is a punctuation filter. Pass in a Filters.⟨String⟩ acceptFilter() to keep punctuation dependencies. A GrammaticalStructure is created for each Tree. The methods on a GrammaticalStructure for each kind of dependencies is as follows:
basic
nonCollapsed
collapsed
CCPropagated gs.typedDependenciesCCprocessed() tree gs.typedDependenciesCollapsedTree()
gs.typedDependencies()
gs.allTypedDependencies()
gs.typedDependenciesCollapsed(true)
⋆ GrammarScope
Bernard Bou has written GrammarScope, a GUI interface to the Stanford Dependencies representation, which allows not only viewing dependencies, but altering their definitions. This is a separate download. It is available at: http://grammarscope.sourceforge.net/.
⋆ Other parsers
A number of dependency parsers have now been trained to parse directly to the basic Stanford Dependen- cies, including MaltParser, DeSR, MSTParser, and Stanford’s Neural Network Dependency Parser. Sev- eral of these parsers distribute models trained for Stanford Dependencies parsing, including MaltParser, the Easy First Parser, and the Stanford Neural Network Dependency Parser. If desired, these parses can then be postprocessed to the collapsed or CCprocessed representation using the -conllxFile option of EnglishGrammaticalStructure, as discussed above.
Any Penn Treebank constituency parser can be used to produce Stanford Dependencies by using our conversion tool to convert the output of other constituency parsers to the Stanford Dependencies representation. For more information on other parser options, see:
http://nlp.stanford.edu/software/stanford-dependencies.shtml
21
6 Further references for Stanford Dependencies
The Stanford Dependencies representation was first made available in the 2005 version of the Stanford Parser. Subsequent releases have provided refinements to and corrections of the relations defined in the original release. The initial written presentation was (de Marneffe et al. 2006). A more thorough dis- cussion of the motivations behind the design of the representation appears in (de Marneffe and Manning 2008).
The SD representation has seen considerable use within the biomedical text mining community. It has been used to give a task relevant evaluation scheme for parsers (Clegg and Shepherd 2007, Pyysalo et al. 2007) and as a representation for relation extraction (Erkan et al. 2007, Greenwood and Stevenson 2007, Urbain et al. 2007, Fundel et al. 2007, Clegg 2008, Airola et al. 2008, Giles and Wren 2008, O ̈zgu ̈r et al. 2008, Ramakrishnan et al. 2008, Bjo ̈rne et al. 2008, Garten 2010, Bjo ̈rne and Salakoski 2011, Pyysalo et al. 2011, Landeghem et al. 2012). Pyysalo et al. (2007) develops a version of the BioInfer corpus annotated with (a slight variant of) the SD scheme. It is avail- able for download at http://mars.cs.utu.fi/BioInfer/. A small amount of SD gold standard anno- tated data was separately prepared for the Parser Evaluation Shared Task of the Workshop on Cross- Framework and Cross-Domain Parser Evaluation (http://lingo.stanford.edu/events/08/pe/) and is dis- cussed in (de Marneffe and Manning 2008). This data is available from the Stanford Dependencies page: http://nlp.stanford.edu/software/stanford-dependencies.shtml, but the BioInfer corpus is the main source of gold-standard SD data which is currently available. In the recent BioNLP 2009 Shared Task, many of the leading teams built their relation extraction systems over the Stanford Dependency representation (Kim et al. 2009). It was used by the teams that came 1st, 3rd, 4th, and 5th in Task 1, by the team who came first in Task 2, and by the teams who came 1st and 2nd in Task 3. In the BioNLP 2011 shared task, every team used it (Kim et al. 2011).
The SD representation has also been used in other domains. It is a common representation for ex- tracting opinions, sentiment, and relations (Zhuang et al. 2006, Meena and Prabhakar 2007, Banko et al. 2007, Zouaq et al. 2006; 2007, Chaumartin 2007, Kessler 2008, Haghighi and Klein 2010, Hassan et al. 2010, Joshi et al. 2010, Wu and Weld 2010, Zouaq et al. 2010), as well as specific information (such as event, time or dialogue acts) (Chambers 2011, McClosky and Manning 2012, Klu ̈wer et al. 2010). The tool has been consistently used by several groups in the PASCAL/NIST challenges targeting textual entailment (Adams et al. 2007, Blake 2007, Chambers et al. 2007, Harmeling 2007, Wang and Neumann 2007, Malakasiotis 2009, Mehdad et al. 2009, Shivhare et al. 2010, Glinos 2010, Kouylekov et al. 2010, Pakray et al. 2011). It is also used for a variety of other tasks, such as unsupervised semantic parsing (Poon and Domingos 2009), coreference resolution, disagreement detection and word sense induction (Chen and Eugenio 2012, Abbott et al. 2011, Lau et al. 2012), as well as being part of the preprocessing for machine translation systems by several groups (Xu et al. 2009, Genzel 2010, Sing and Bandyopad- hyay 2010). The Stanford Dependencies representation was also used to evaluate dependency parsers in the 2012 shared task on parsing the web (Petrov and McDonald 2012).
The Stanford dependency representation has also served as a model for developing dependency schemes for other languages. The Chinese Stanford Dependencies are briefly described in Chang et al. (2009). Representations based on the Stanford dependency representation have been proposed for Finnish (Haverinen et al. 2010a;b), Thai (Potisuk 2010), Persian (Seraji et al. 2012), and French (El Maarouf and Villaneau 2012). More recently, there has been increasing interest in defining a consistent cross-linguistic set of relations in the style of Stanford Dependencies (McDonald et al. 2013, Tsarfaty 2013). The au- thors and others initially proposed a reformulation as Universal Stanford Dependencies (de Marneffe
22
et al. 2014). Further refinement with a team of collaborators has led to a new synthesis spanning tok- enization, morphological features, parts of speech, and dependencies, known as Universal Dependencies: http://universaldependencies.github.io/docs/. Since version 3.5.2 the default representation output by our parser is the Universal Dependencies representation.
7 Recent changes
This section summarizes recent changes. This may help if you see old versions of the dependencies, or need to update your code.
abbrev was removed as a relation. It is now a particular case of an appos.
attr has been removed as a relation. attr was a relation intended for the complement of a copular verb such as “to be”, “to seem”, “to appear”. Mainly, it was used for WHNP complements. (The relation attr was meant to stand for “attributive” but that was arguably a misuse of the word.)
complm was removed as a relation. It is now a particular case of mark. This follows HPSG-like usage, where the complementizer is a mark on the clause.
discourse was introduced. The lack of a dependency type for interjections was an omission even in the early versions, but it became more essential as we expanded our consideration of informal text types.
goeswith was introduced. It is useful on badly edited text.
infmod was remode as a relation. It has been generalized as a case of vmod. partmod was remode as a relation. It has been generalized as a case of vmod. purpcl was removed as a relation. It is now a particular case of an advcl.
rel has been removed as a relation. rel was the relation between the main verb of a relative clause and the head of the Wh-phrase. Now, the converter resolves the grammatical relation (nsubj, dobj, or pobj) for simple cases, and the rest are left unresolved as a dep relation.
vmod has been introduced as a relation generalizing over non-finite verbal modifiers that are participial in form (formerly partmod) or infinitival (formerly infmod).
References
Rob Abbott, Marilyn Walker, Pranav Anand, Jean E. Fox Tree, Robeson Bowmani, and Joseph King. How can you say such things?!?: Recognizing disagreement in informal political argument. In Pro- ceedings of the Workshop on Languages in Social Media, LSM ’11, pages 2–11, 2011.
Rod Adams, Gabriel Nicolae, Cristina Nicolae, and Sanda Harabagiu. Textual entailment through ex- tended lexical overlap and lexico-semantic matching. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, pages 119–124, Prague, June 2007.
23
Antti Airola, Sampo Pyysalo, Jari Bjo ̈rne, Tapio Pahikkala, Filip Ginter, and Tapio Salakoski. A graph kernel for protein-protein interaction extraction. In Proceedings of BioNLP 2008: Current Trends in Biomedical Natural Language Processing (ACL08), 2008.
Michele Banko, Michael J. Cafarella, Stephen Soderland, Matt Broadhead, and Oren Etzioni. Open information extraction from the web. In Proceedings of the 20th International Joint Conference on Artificial Intelligence (IJCAI 2007), 2007.
JariBjo ̈rneandTapioSalakoski.Generalizingbiomedicaleventextraction.InProceedingsoftheBioNLP Shared Task 2011 Workshop, BioNLP Shared Task ’11, pages 183–191, 2011.
Jari Bjo ̈rne, Sampo Pyysalo, Filip Ginter, and Tapio Salakoski. How complex are complex protein- protein interactions? In 3rd International Symposium on Semantic Mining in Biomedecine, 2008.
Catherine Blake. The role of sentence structure in recognizing textual entailment. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, pages 101–106, Prague, June 2007.
Nathanael Chambers. Inducing Event Schemas and their Participants from Unlabeled Text. PhD thesis, Department of Computer Science, Stanford University, 2011.
Nathanael Chambers, Daniel Cer, Trond Grenager, David Hall, Chloe Kiddon, Bill MacCartney, Marie- Catherine de Marneffe, Daniel Ramage, Eric Yeh, and Christopher D. Manning. Learning alignments and leveraging natural logic. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, pages 165–170, Prague, June 2007.
Pi-Chuan Chang, Huihsin Tseng, Dan Jurafsky, and Christopher D. Manning. Discriminative reordering with Chinese grammatical relations features. In Proceedings of the Third Workshop on Syntax and Structure in Statistical Translation, Boulder, Colorado, June 2009. URL pubs/ssst09-chang.pdf.
Franc ̧ois-Re ́gis Chaumartin. UPAR7: A knowledge-based system for headline sentiment tagging. In Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 422– 425, 2007.
Lin Chen and Barbara Di Eugenio. Co-reference via pointing and haptics in multi-modal dialogues. In
Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2012.
Andrew B. Clegg. Computational-Linguistic Approaches to Biological Text Mining. PhD thesis, School of Crystallography, Birkbeck, University of London, 2008.
Andrew B. Clegg and Adrian J. Shepherd. Benchmarking natural-language parsers for biological appli- cations using dependency graphs. BMC Bioinformatics, 8:24, 2007.
Marie-Catherine de Marneffe and Christopher D. Manning. The Stanford typed dependencies represen- tation. In COLING Workshop on Cross-framework and Cross-domain Parser Evaluation, 2008.
Marie-Catherine de Marneffe, Bill MacCartney, and Christopher D. Manning. Generating typed depen- dency parses from phrase structure parses. In 5th International Conference on Language Resources and Evaluation (LREC 2006), 2006.
24
Marie-Catherine de Marneffe, Timothy Dozat, Natalia Silveira, Katri Haverinen, Filip Ginter, Joakim Nivre, and Christopher D. Manning. Universal stanford dependencies: A cross-linguistic typology. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC- 2014), 2014.
Isma ̈ıl El Maarouf and Jeanne Villaneau. A French fairy tale corpus syntactically and semantically anno- tated. In Proceedings of the Eight International Conference on Language Resources and Evaluation, 2012.
Gunes Erkan, Arzucan Ozgur, and Dragomir R. Radev. Semi-supervised classification for extracting pro- tein interaction sentences using dependency parsing. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), 2007.
Katrin Fundel, Robert Ku ̈ffner, and Ralf Zimmer. RelEx – relation extraction using dependency parse trees. Bioinformatics, 23, 2007.
Yael Garten. Text mining of the scientific literature to identify pharmacogenomic interactions. PhD thesis, Department of Biomedical Informatics, Stanford University, 2010.
Dmitriy Genzel. Automatically learning source-side reordering rules for large scale machine translation. In COLING-2010, 2010.
Cory B. Giles and Jonathan D. Wren. Large-scale directional relationship extraction and resolution. BMC Bioinformatics, 9(Suppl 9):S11, 2008.
Demetrios G. Glinos. System description for SAIC entry at RTE-6. In Proceedings of the Text Analysis Conference (TAC), 2010.
Mark A. Greenwood and Mark Stevenson. A semi-supervised approach to learning relevant protein- protein interaction articles. In Proceedings of the Second BioCreAtIvE Challenge Workshop, Madrid, Spain, 2007.
Aria Haghighi and Dan Klein. An entity-level approach to information extraction. In Proceedings of the ACL 2010 Conference Short Papers, ACLShort ’10, pages 291–295, 2010.
Stefan Harmeling. An extensible probabilistic transformation-based approach to the third recognizing textual entailment challenge. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, pages 137–142, Prague, June 2007.
Ahmed Hassan, Vahed Qazvinian, and Dragomir Radev. What’s with the attitude?: identifying sentences with attitude in online discussions. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, EMNLP ’10, pages 1245–1255, 2010.
Katri Haverinen, Filip Ginter, Timo Viljanen, Veronika Laippala, and Tapio Salakoski. Dependency- based propbanking of clinical Finnish. In Proceedings of the Fourth Linguistic Annotation Workshop, LAW IV ’10, pages 137–141, 2010a.
Katri Haverinen, Timo Viljanen, Veronika Laippala, Samuel Kohonen, Filip Ginter, and Tapio Salakoski. Treebanking Finnish. In Proceedings of the Ninth International Workshop on Treebanks and Linguistic Theories (TLT), 2010b.
25
Mahesh Joshi, Dipanjan Das, Kevin Gimpel, and Noah A. Smith. Movie reviews and revenues: an experiment in text regression. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, HLT ’10, pages 293–296, 2010.
Jason S. Kessler. Polling the blogosphere: a rule-based approach to belief classification. In International Conference on Weblogs and Social Media, 2008.
Jin-Dong Kim, Tomoko Ohta, Sampo Pyysalo, Yoshinobu Kano, and Jun’ichi Tsujii. Overview of bionlp’09 shared task on event extraction. In Proceedings of the BioNLP 2009 Workshop Companion Volume for Shared Task, pages 1–9, 2009.
Jin-Dong Kim, Sampo Pyysalo, Tomoko Ohta, Robert Bossy, Ngan Nguyen, and Jun’ichi Tsujii. Overview of bionlp shared task 2011. In Proceedings of the BioNLP Shared Task 2011 Workshop, 2011.
Tina Klu ̈wer, Hans Uszkoreit, and Feiyu Xu. Using syntactic and semantic based relations for dialogue act recognition. In Proceedings of the 23rd International Conference on Computational Linguistics, COLING ’10, pages 570–578, 2010.
Milen Kouylekov, Yashar Mehdad, Matteo Negri, and Elena Cabrio. FBK participation in RTE-6: Main and KBP validation task. In Proceedings of the Text Analysis Conference (TAC), 2010.
Sofie Van Landeghem, Jari Bjo ̈rne, Thomas Abeel, Bernard De Baets, Tapio Salakoski, and Yves Van de Peer. Semantically linking molecular entities in literature through entity relationships. BMC Bioin- formatics, 13, 2012.
Jey Han Lau, Paul Cook, Diana McCarthy, David Newman, and Timothy Baldwin. Word sense induction for novel sense detection. In Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, 2012.
Prodromos Malakasiotis. AUEB at TAC 2009. In Proceedings of the Text Analysis Conference (TAC), 2009.
David McClosky and Christopher D. Manning. Learning constraints for consistent timeline extraction. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, 2012.
Ryan McDonald, Joakim Nivre, Yvonne Quirmbach-Brundage, Yoav Goldberg, Dipanjan Das, Kuz- man Ganchev, Keith Hall, Slav Petrov, Hao Zhang, Oscar Ta ̈ckstro ̈m, Claudia Bedini, Nu ́ria Bertomeu Castello ́, and Jungmee Lee. Universal dependency annotation for multilingual parsing. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 92–97, 2013.
Arun Meena and T. V. Prabhakar. Sentence level sentiment analysis in the presence of conjuncts using linguistic analysis. In Advances in Information Retrieval, volume 4425 of Lecture Notes in Computer Science. Springer, 2007.
Yashar Mehdad, Matteo Negri, Elena Cabrio, Milen Kouylekov, and Bernardo Magnini. Using lexical resources in a distance-based approach to RTE. In Proceedings of the Text Analysis Conference (TAC), 2009.
26
ArzucanO ̈zgu ̈r,ThuyVu,Gu ̈nesErkan,andDragomirR.Radev.Identifyinggene-diseaseassociations using centrality on a literature mined gene-interaction network. Bioinformatics, 24(13):i277–i285, 2008.
Partha Pakray, Snehasis Neogi, Pinaki Bhaskar, Soujanya Poria, Sivaji Bandyopadhyay, and Alexander Gelbukh. A textual entailment system using anaphora resolution. In Proceedings of the Text Analysis Conference (TAC), 2011.
Slav Petrov and Ryan McDonald. Overview of the 2012 shared task on parsing the web. In First Workshop on Syntactic Analysis of Non-Canonical Language, 2012.
Hoifung Poon and Pedro Domingos. Unsupervised semantic parsing. In Proceedings of the 2009 Con- ference on Empirical Methods in Natural Language Processing (EMNLP 2009), pages 1–10, 2009.
Siripong Potisuk. Typed dependency relations for syntactic analysis of Thai sentences. In Proceedings of PACLIC 24 Pacific Asia Conference on Language, Information and Computation, 2010.
Sampo Pyysalo, Filip Ginter, Katri Haverinen, Juho Heimonen, Tapio Salakoski, and Veronika Laippala. On the unification of syntactic annotations under the Stanford dependency scheme: A case study on BioInfer and GENIA. In Proceedings of BioNLP 2007: Biological, translational, and clinical language processing (ACL07), 2007.
Sampo Pyysalo, Tomoko Ohta, and Jun’ichi Tsujii. An analysis of gene/protein associations at PubMed scale. Journal of Biomedical Semantics, 2, 2011.
Cartic Ramakrishnan, Pablo N. Mendes, Shaojun Wang, and Amit P. Sheth. Unsupervised discovery of compound entities for relationship extraction. In 16th International Conference on Knowledge Engineering: Practice and Patterns (EKAW 2008), pages 146–155, 2008.
Mojgan Seraji, Bea ́ta Megyesi, and Joakim Nivre. A basic language resource kit for Persian. In Pro- ceedings of the Eight International Conference on Language Resources and Evaluation, 2012.
Himanshu Shivhare, Parul Nath, and Anusha Jain. Semi cognitive approach to RTE-6 – using FrameNet for semantic clustering. In Proceedings of the Text Analysis Conference (TAC), 2010.
Thoudam Doren Sing and Sivaji Bandyopadhyay. Statistical machine translation of English – Manipuri using morpho-syntactic and semantic information. In Proceedings of the Association for Machine Translation in the Americas (AMTA 2010), 2010.
Reut Tsarfaty. A unified morpho-syntactic scheme of stanford dependencies. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 578–584, 2013.
Jay Urbain, Nazli Goharian, and Ophir Frieder. IIT TREC 2007 genomics track: Using concept-based semantics in context for genomics literature passage retrieval. In The Sixteenth Text REtrieval Confer- ence (TREC 2007) Proceedings, 2007.
Rui Wang and Gu ̈nter Neumann. Recognizing textual entailment using sentence similarity based on dependency tree skeletons. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, pages 36–41, Prague, June 2007.
27
Fei Wu and Daniel S. Weld. Open information extraction using Wikipedia. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, ACL ’10, 2010.
Peng Xu, Jaeho Kang, Michael Ringgaard, and Franz Och. Using a dependency parser to improve SMT for subject-object-verb languages. In NAACL 2009: Proceedings of Human Language Technologies, The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 245–253, 2009.
Li Zhuang, Feng Jing, Xiao yan Zhu, and Lei Zhang. Movie review mining and summarization. In Proc. ACM Conference on Information and Knowledge Management (CIKM), pages 43–50, 2006.
Amal Zouaq, Roger Nkambou, and Claude Frasson. The knowledge puzzle: An integrated approach of intelligent tutoring systems and knowledge management. In Proceedings of the 18th IEEE Interna- tional Conference on Tools with Artificial Intelligence (ICTAI 2006), pages 575–582, 2006.
Amal Zouaq, Roger Nkambou, and Claude Frasson. Building domain ontologies from text for educa- tional purposes. In Proceedings of the Second European Conference on Technology Enhanced Learn- ing: Creating new learning experiences on a global scale, 2007.
Amal Zouaq, Michel Gagnon, and Benoˆıt Ozell. Semantic analysis using dependency-based grammars and upper-level ontologies. International Journal of Computational Linguistics and Applications, 1 (1-2), 2010.
28