Abstract Meaning Representation (AMR) 1.2.5 Specification
February 23, 2018
Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, Nathan Schneider
Table of Contents generated with DocToc
- Part I. Introduction
- Part II. Concepts and relations
- Part III. Phenomena
- Core roles
- Imperatives and other modes
- Articles, plurals, tense, aspect, quotes, hyphens
- Implicit roles
- Implicit concepts
- Main verb “be”
- Nouns that invoke predicates
- Adjectives that invoke predicates
- Adverbs with -ly
- Non-core roles
- Phrasal verbs
- Relative clauses
- Multiple relations with the same name
- Quantifiers and scope
- Variables and co-reference
- Named Entities and wikification
- Special Frames for Roles (have-org-role-91, have-rel-role-91)
- Exact numbers
- Approximate numbers
- Mathematical operators
- Other entities: dates, times, percentages, phone, email, URLs
- Further information
- AMR Freak Show
Part I. Introduction
AMR captures “who is doing what to whom” in a sentence. Each sentence is represented as a rooted, directed, acyclic graph with labels on edges (relations) and leaves (concepts).
Like a parse tree, AMR provides a single, traversable structure that takes all words into account. It is not a disconnected set of annotation layers. Unlike a parse tree, the AMR is abstract. It may represent any number of natural language sentences. AMR does not annotate the individual words in a sentence, like a dependency parse does.
AMR implements a simplified, standard neo-Davidsonian semantics [Davidson 1967, Higginbotham 1985], using standard feature structure representation [Shieber 1986, Carpenter 1992]. AMR’s formal origins are in unification systems [Kay 1979, Knight 1989, Moore 1989] and natural language generation [Mann 1982, Elhadad 1995, Knight & Hatzivassiloglou 1995]. Predicates senses and core semantic roles in AMR are drawn from the amazing OntoNotes project.
AMR does not say anything about how it wants to be processed. It is closer to English than to other languages. It is not an interlingua.
This AMR means (roughly): There is a wanting event, whose ARG0 (wanter) is a
boy, and whose ARG1 (wanted thing) is a believing event. This believing event
has an ARG0 (believer), which is a girl, and it has an ARG1 (believed thing),
which is the same boy just mentioned. Here,
boy plays two roles: (1) it is
the ARG0 of
want-01, and (2) it is the ARG1 of
believe-01. The AMR captures
this with two directed edges pointing to the same node. (Per OntoNotes,
predicate senses are marked with suffixes like
-02, while ARG0,
ARG1, etc., denote core, predicate-specific roles.)
Here is a text-friendly way to write the same AMR:
(w / want-01 :ARG0 (b / boy) :ARG1 (b2 / believe-01 :ARG0 (g / girl) :ARG1 b))
g correspond to internal nodes in the graph above.
Note that b appears twice in this format, the first time as
(b / boy) and the
second time simply as
This AMR can also be viewed as conjunction of logical triples, omitting root information:
instance(w, want-01) ^ /* w is an instance of wanting */ instance(b, boy) ^ /* b is an instance of boy */ instance(b2, believe-01) ^ /* b2 is an instance of believing */ instance(g, girl) ^ /* g is an instance of girl */ ARG0(w, b) ^ /* b is the wanter in w */ ARG1(w, b2) ^ /* b2 is the wantee in w */ ARG0(b2, g) ^ /* g is the believer in b2 */ ARG1(b2, b) /* b is the believee in b2 */
Abstraction away from English
The AMR above can be expressed variously in English:
The boy desires the girl to believe him.
The boy desires to be believed by the girl.
The boy has a desire to be believed by the girl.
The boy’s desire is for the girl to believe him.
The boy is desirous of the girl believing him.
desire-01 might be realized as a verb (“desire”), a noun (also “desire”),
or an adjective (“desirous”).
We think of AMR leaf-labels as concepts rather than words. We do not point to an element in an AMR and say “that is a noun” or “that is a verb”. Rather, we say “that is an object” or “that is an event”.
A single entity (“boy”) can play multiple roles simultaneously (e.g., “ARG0” of
desire-01, and “ARG1” of
believe-01). The AMR does not talk about pronouns or
zero-pronouns, though these are natural mechanisms for expressing multiple roles
In many cases, English function words do not show up at all in AMR:
(a / adjust-01 :ARG0 (b / girl) :ARG1 (m / machine))
The girl made adjustments to the machine.
The girl adjusted the machine.
The machine was adjusted by the girl.
(k / kill-01 :time (y / yesterday))
The killing happened yesterday.
The killing took place yesterday.
(a / and :op1 (b / boy) :op2 (g / girl))
the boy and the girl
both the boy and the girl
(b / boat :poss (h / he))
his own boat
More Logical than Syntax
AMR strives for a more logical, less syntactic representation. For example,
“the boy must not go” and “the boy may not go” are syntactically similar, but
the placement of negation (
:polarity -) is very different in the two AMRs:
(o / obligate-01 :ARG2 (g / go-02 :ARG0 (b / boy) :polarity -))
The boy must not go.
It is obligatory that the boy not go.
(p / permit-01 :ARG1 (g / go-02 :ARG0 (b / boy)) :polarity -)
The boy may not go.
The boy is not permitted to go.
It is not permissible for the boy to go.
The boy does not have permission to go.
The AMR transparently represents what is being negated. Note that the concept
permit-01 can be realized as a modal, a participle, or a noun.
The root of an AMR binds its contents into a single, traversable directed graph. It also serves as a rudimentary representation of overall focus. So we have:
(w / white-03 :ARG1 (m / marble))
The marble is white.
the whiteness of the marble
(m / marble :ARG1-of (w / white-03))
the white marble
the marble that is white
(s / see-01 :ARG0 (b / boy) :ARG1 (w / white-03 :ARG1 (m / marble)))
The boy sees that the marble is white.
The boy sees the whiteness of the marble.
(s / see-01 :ARG0 (b / boy) :ARG1 (m / marble :ARG1-of (w / white-03)))
The boy sees the white marble.
The boy sees the marble that is white.
We can write
:domain-of as an inverse of
:domain, but we often shorten this to
Inverse roles are useful for maintaining a single rooted structure, e.g.:
(s / see-01 :ARG0 (b / boy) :ARG1 (g / girl :ARG0-of (w / want-01 :ARG1 b)))
The boy saw the girl who wanted him.
The boy saw the girl who he was wanted by.
The girl who wanted the boy was seen by him.
In this AMR, the role
want-01 in a natural way.
To re-focus an AMR, we can “lift up” any node to the root, and then imagine all
other nodes falling down. For example, if we lift up the
w node above, we get
the same content, but rearranged:
(w / want-01 :ARG0 (g / girl :ARG1-of (s / see-01 :ARG0 (b / boy))) :ARG1 b)
The girl who was seen by the boy wants him.
The boy is wanted by the girl he saw.
This is a matter of focusing: the first AMR (rooted by
see-01) is about the
seeing, while the second AMR (rooted by
want-01) is about the wanting.
Another example of an inverse role (
(c / change-01 :ARG1 (d / document :instrument-of (r / regulate-01)))
The regulatory documents were changed.
Here are some slogans that make it easier to work with AMR:
- AMR captures the rough meaning of a sentence in a single, traversable directed acyclic graph.
- AMR does not say anything about how it wants to be processed.
- AMR is not an interlingua.
- There are no nouns and verbs in AMR.
- When we write AMR by hand, constituent trees and dependency trees are generally not needed.
- We cannot read off a unique English sentence from an AMR.
An AMR is like a foreign-language translation. Someone who creates an AMR from English may not provide links between AMR concepts and English word tokens. But it is reasonable to think about doing such an alignment later (perhaps automatically), just as it is reasonable to align tokens in bilingual texts.
Limitations of AMR 1.2
AMR 1.2 is over-simple in many ways:
- It is geared toward English and the vocabulary of English.
- It does not represent quantifier scope, or even universal quantification at all.
- It does not represent co-references that cross sentence boundaries. (LDC general release in early 2018 will include new multi-sentence co-reference annotations for some AMR documents.)
- It drops grammatical number, tense, aspect, quotation marks, etc.
- It does not deeply capture many noun-noun or noun-adjective relations.
- It does not include deep frames such as Earthquake (with roles for magnitude, epicenter, casualties, etc) or Pregnancy (with roles for mother, father, baby gender, time since inception, etc). AMR 1.2 looks forward to AMR 2.0!
Part II. Concepts and relations
Concepts are tokens that appear at the leaves of AMR graphs. AMR does not formally carve the world up into events, objects, features, etc., though we may refer to a certain concept instance as an event. The following AMR has three concepts (boy, want, believe):
(w / want-01 :ARG0 (b / boy) :ARG1 (b2 / believe-01 :ARG0 b))
The boy wants to believe.
The slash (
/) is shorthand for the instance relation. This relation shows up more clearly in AMR graph format (see Introduction).
Concepts in AMR 1.2 are usually written with English words or phrases. Concepts with core semantic relations may have sense tags, to identify the semantic frame:
(b / believe-01 :ARG0 (b2 / boy))
The boy believes.
AMR semantic relations are best described through examples; see the next section (Phenomena). Here, we only provide a brief summary listing.
:ARGx roles. AMR 1.2 uses numbered
:ARGx from OntoNotes:
:ARG0, :ARG1, :ARG2, :ARG3, :ARG4, :ARG5
:accompanier, :age :beneficiary :concession, :condition, :consist-of :degree, :destination, :direction, :domain, :duration :example, :extent :frequency :instrument :li, :location :manner, :medium, :mod, :mode :name :ord :part, :path, :polarity, :polite, :poss, :purpose :quant :range :scale, :source, :subevent :time, :topic :unit :value :wiki
Roles used in date-entity:
:calendar, :century, :day, :dayperiod, :decade, :era, :month, :quarter, :season, :timezone, :weekday, :year, :year2
Roles of the form
:opx are used in conjunctions, and in certain types of
locations and times:
:op1, :op2, :op3, :op4, …
Roles of the form
:prep-X are used in cases where there is no good relation from
the list above. AMR likes to avoid these. Here is a partial list. It is only
partial, as other
:prep-X relations are legal AMR.
:prep-against, :prep-along-with, :prep-amid, :prep-among, :prep-as, :prep-at :prep-by :prep-for, :prep-from :prep-in, :prep-in-addition-to, :prep-into :prep-on, :prep-on-behalf-of, :prep-out-of :prep-to, :prep-toward :prep-under :prep-with, :prep-without
Some conjunctions are also not well-covered under the list of non-core roles. AMR also likes to avoid these, but sometimes we have no good alternative:
All relations above have inverses of the form
:ARG0-of, :ARG1-of :location-of
Part III. Phenomena
Core roles are taken from the OntoNotes semantic role layer. OntoNotes
predicates are sense-labeled words (e.g.,
sentence-01). They are
predicate-specific and numbered. For example:
charge-05is the person doing the charging.
charge-05is the person being charged.
charge-05is the role or crime (for which
sentence-01is the person doing the sentencing.
sentence-01is the person being sentenced.
sentence-01is the sentence (punishment).
sentence-01is the role or crime.
fine-01is the person doing the fining.
fine-01is the amount of the fine (e.g., $1000).
fine-01is the person being fined.
fine-01is the role or crime.
Note the predicate-sensitivity: a person might naturally be the
sentence-01, but simultaneously the
One semantic frame may be realized in vastly different ways in English:
(d / describe-01 :ARG0 (h / he) :ARG1 (m / mission) :ARG2 (f / fail-01))
He described the mission as a failure.
As he described it, the mission was a failure.
His description of the mission: failure.
Here, the AMR does not worry about representing the words “as”, “it”, or “was”.
Note that OntoNotes documentation for predicates and roles is often loose or
humorous. For example, the
research-01 are nicknamed
“student” and “subject”, but this does not mean that
research-01 is restricted
to situations where the
:ARG0 is literally a student.
If OntoNotes is missing a predicate, AMR accepts
(b / band :consist-of (p / person :ARG0-of (m / maraud-00)))
a band of marauders
AMR represents syntactic modals with concepts like
(p / possible-01 :ARG1 (g / go-02 :ARG0 (b / boy)))
The boy can go.
It is possible that the boy goes.
(o / obligate-01 :ARG2 (g / go-02 :ARG0 (b / boy)))
The boy must go.
The boy is obligated to go.
It is obligatory that the boy go.
(o / permit-01 :ARG1 (g / go-02 :ARG0 (b / boy)))
The boy may go.
The boy is permitted to go.
It is permissible that the boy go.
(p / possible-01 :ARG1 (r / rain-01))
It may rain.
It might rain.
Rain is possible.
It’s possible that it will rain.
(r / recommend-01 :ARG1 (g / go-02 :ARG0 (b / boy)))
The boy should go.
It is recommended that the boy go.
(l / likely-01 :ARG1 (g / go-02 :ARG0 (b / boy)))
The boy is likely to go.
It is likely that the boy will go.
AMR ignores the modal “would”, except in cases like:
(p / prefer-01 :ARG0 (b / boy) :ARG1 (g / go-02 :ARG0 b))
The boy would rather go.
The boy prefers to go.
(u / use-02 :ARG0 (i / i) :ARG1 (w / work-01 :ARG0 i))
I am used to working.
AMR represents negation logically, using
(g / go-02 :ARG0 (b / boy) :polarity -)
The boy doesn’t go.
(p / possible-01 :ARG1 (g / go-02 :ARG0 (b / boy)) :polarity -)
The boy can’t go.
It’s not possible for the boy to go.
(p / possible-01 :ARG1 (g / go-02 :ARG0 (b / boy) :polarity -))
It is possible for the boy not to go.
It is possible for the boy to not go.
(p / obligate-01 :ARG2 (g / go-02 :ARG0 (b / boy)) :polarity -)
The boy doesn’t have to go.
It’s not necessary for the boy to go.
(p / obligate-01 :ARG2 (g / go-02 :ARG0 (b / boy) :polarity -))
The boy must not go.
It’s obligatory that the boy not go.
(t / think-01 :ARG0 (b / boy) :ARG1 (w / win-01 :ARG0 (t2 / team :poss b) :polarity -))
The boy thinks his team won’t win.
The boy doesn’t think his team will win. (colloquially, ambiguously)
(t / think-01 :ARG0 (b / boy) :ARG1 (w / win-01 :ARG0 (t2 / team :poss b)) :polarity -)
It’s not true that the boy thinks his team will win.
The boy doesn’t think his team will win. (colloquially, ambiguously)
(h / have-03 :polarity - :ARG0 (i / i) :ARG1 (m / money))
I don’t have any money.
I have no money.
Negative English affixes are also represented with the
(a / appropriate-02 :polarity - :ARG1 (d / dress))
the dress is inappropriate
the dress is not appropriate
(d / dress :ARG1-of (a / appropriate-02 :polarity -))
the inappropriate dress
the dress that is inappropriate
the dress that is not appropriate
AMR uses the concept
amr-unknown (in-place!) to indicate questions:
(f / find-01 :ARG0 (g / girl) :ARG1 (a / amr-unknown))
What did the girl find?
(f / find-01 :ARG0 (g / girl) :ARG1 (b / boy) :location (a / amr-unknown))
Where did the girl find the boy?
(f / find-01 :ARG0 (g / girl) :ARG1 (b / boy) :manner (a / amr-unknown))
How did the girl find the boy?
(f / find-01 :ARG0 (g / girl) :ARG1 (t / toy :poss (a / amr-unknown)))
Whose toy did the girl find?
(r / run-01 :ARG0 (g / girl) :manner (f / fast :degree (a / amr-unknown)))
How fast did the girl run?
(s / see-01 :ARG0 (g / girl) :ARG1 (a / amr-unknown :ARG1-of (p / purple-02)))
What purple thing did the girl see?
(l / lead-01 :ARG0 (s / she) :ARG1 (a / amr-unknown :domain (i / investigate-01)))
Which investigation did she lead?
AMR also uses amr-unknown for yes-no questions. For such yes-no questions, the
:polarity relation, essentially asking "what is the truth of this assertion?":
(f / find-01 :ARG0 (g / girl) :ARG1 (b / boy) :polarity (a / amr-unknown))
Did the girl find the boy?
(f / find-01 :ARG1 (b / boy) :polarity (a / amr-unknown))
Was the boy found?
Note that wh- words in relative clauses are treated differently, using inverse
roles instead of
(k / know-02 :ARG0 (i / i) :ARG1 (p / person :ARG1-of (s / see-01 :ARG0 (y / you))))
I know who you saw.
I know the person you saw.
AMR also uses inverse roles for embedded interrogative clauses, such as those that
start with "whether". As with relative clauses, we do not use
truth-value concept is added, to refer to whether or not an event happened:
(k / know-01 :polarity - :ARG0 (b / boy) :ARG1 (t / truth-value :polarity-of (c / come-01 :ARG1 (g / girl))))
The boy doesn’t know whether the girl came.
The boy doesn’t know if the girl came.
We contrast this with:
(k / know-01 :polarity - :ARG0 (b / boy) :ARG1 (c / come-01 :ARG1 (g / girl)))
The boy doesn’t know that the girl came.
The boy doesn’t know the girl came.
Choice questions can look like yes/no questions, but provide a set of possible answers for the question.
We handle these with the concept
amr-choice (usually in place of
or), which can receive any number of numbered
(w / want-01 :ARG0 (y / you) :ARG1 (a / amr-choice :op1 (t / tea) :op2 (c / coffee)))
Do you want tea or coffee?
(r / recommend-01 :ARG1 (a / amr-choice :op1 (s / stay-01 :ARG1 (i / i)) :op2 (g / go-02 :ARG0 i)))
Should I stay or should I go?
Imperative and Expressive mode
:mode imperative is used for imperatives. Exclamatory imperatives are just imperatives in AMR.
(g / go-02 :mode imperative :ARG0 (y / you))
(g / go-02 :mode imperative :ARG0 (w / we))
:mode expressive is used to mark exclamational words such as
ah, ha, hmm, oh, wow, yippee that express emotions, but don't refer to a clear event, object or property. Do not use
:mode expressive for mere emphasis (text in ALLCAPS), exclamation marks (!) or disfluency markers (uh), which are not annotated in AMR.
(y / yippee :mode expressive)
(y / yes :mode expressive)
Yes. as a normal answer to a question is not marked with any :mode, as it plainly conveys information based on the preceding question. But a celebratory
Yes!!! after your favorite sports team scores an imporant point should be marked up using
Articles, plurals, tense, aspect, quotes, hyphens
AMR 1.2 does not represent event times (outside of the explicit
articles, plurals, or quotation marks:
(g / go-02 :ARG0 (b / boy))
The boy went.
The boys went.
A boy went.
The boy goes.
The boy will go.
Demonstratives are included:
(b / boy :mod (t / that))
(b / boy :mod (t / this))
Demonstrative pronouns are also included if they have no antecedent in the sentence:
(s / shame :domain (t / that))
that is a shame.
If a hyphenated word can be broken down into component meanings, we do it:
(a / account :mod (m / market :mod (m2 / money)))
(p / president :mod (v / vice))
But when it is hard to get component meanings out, then we leave it together:
(b / brother-in-law)
In any case, we never make the hyphen itself (“-”) into an AMR concept.
AMR roles may be implicit when rendered in English. AMR includes such roles when there is no real debate about what is happening in the world. Consider:
(c / charge-05 :ARG1 (h / he) :ARG2 (a / and :op1 (i / intoxicate-01 :ARG1 h :location (p / public)) :op2 (r / resist-01 :ARG0 h :ARG1 (a2 / arrest-01 :ARG1 h))))
He was charged with public intoxication and resisting arrest.
Here, the variable h appears four times, including as the ARG1 of
because it is clear that h is resisting his own arrest (not someone else’s).
However, we do not include anything to the effect of “the charging agent and the
arresting agent are the same entity”, as that is debatable.
When we build AMR from text, we introduce implicit roles, but we generally do
not introduce implicit concepts, e.g.,
(h / hopeful-03 NOT: (f / full :ARG1 (g / girl)) :poss (h / hope) :ARG1 (g / girl))
the hopeful girl
An exception is named entity types for entities that lack one, as covered below under Named Entities.
Main verb “be”
Predicate adjectives are usually represented with
:domain, unless we have an
adjective frame in OntoNotes:
(s / small :domain (m / marble))
The marble is small.
“Noun is noun” constructions also use
(l / lawyer :domain (m / man))
The man is a lawyer.
(m / man :mod (l / lawyer))
the man who is a lawyer
“There is…” and “there are…” have simple AMR representations:
(b / boy)
There is a boy.
(b / boy :quant 4 :ARG0-of (m / make-01 :ARG1 (p / pie)))
four boys making pies
There are four boys making pies.
Nouns that invoke predicates
AMR’s principle is to maximize the use of OntoNotes predicates, regardless of English parts of speech. This section gives examples of this principle.
AMR represents events, not verbs. So, “destroy” and “destruction” have the same AMR representation. For consistency, AMR uses sense-tagged English verbs from OntoNotes:
(d / destroy-01 :ARG0 (b / boy) :ARG1 (r / room))
The boy destroyed the room.
The boy’s destruction of the room
The destruction of the room by the boy
destroy-01, we fully exploit the semantic frames in OntoNotes, which
are most developed for English verbs.
We never say:
(d / destruction …)
(d / destruction-01 …)
Older versions of OntoNotes have noun predicate frames like
but we do not want AMRs to contain both
Therefore, we avoid
Some nominalizations (like “explosion”) refer to a whole event, while others (like “proposal”) can refer to role player in the event:
(e / explode-01)
(t / thing :ARG1-of (p / propose-01))
the thing proposed
what got proposed
We always search for an OntoNotes predicate, even if the noun is much more frequent than the verb (“opinion” is the thing that is “opined”):
(t / thing :ARG1-of (o / opine-01 :ARG0 (b / boy)))
the boy’s opinion
the opinion of the boy
that which was opined by the boy
what the boy opined
Inverse roles are also used to represent many “-er” nouns. This enables us to
make maximal use of use of OntoNotes predicate frames, instead of defaulting to
:mod” or “
:poss” or “
(o / organization NOT: (m / maker :ARG0-of (m / make-01 :mod (c / chip)) :ARG1 (c / chip)))
maker of chips
(p / person :ARG0-of (i / invest-01))
(p / person :ARG0-of (i / invest-01 :ARG2 (b / bond)))
(p / person NOT: (i / investor :ARG0-of (i / invest-01 :mod (s / small)) “I can’t see you!” :mod (s / small)))
(p / person :ARG0-of (i / invest-01) :mod (n / nerd))
When a noun’s meaning is significantly different from the verbal form, then AMR does not break down its meaning. For example, a “treasurer” is not essentially someone who treasures:
(t / treasurer) NOT: (p / person :ARG0-of (t / treasure-01))
(p / president) NOT: (p / person :ARG0-of (p2 / preside-01))
Even when the meaning of an “-er” noun does break down comfortably, AMR does not automatically reach for an inverse (“-of”) role. For example, “the boy is a hard worker” just means the boy works hard:
(w / work-01 NOT: (b / boy :ARG0 (b / boy) :ARG0-of (w / work-01 :manner (h / hard-02)) :manner (h / hard-02)))
the boy is a hard worker
the boy works hard
Adjectives that invoke predicates
Like nouns, adjectives also invoke predicates:
(m / man :ARG0-of (a / attract-01))
the attractive man
(a / attract-01 :ARG0 (m / man))
the man is attractive
the man attracts
(a / attract-01 :ARG0 (m / man) :ARG1 (w / woman))
the man is attractive to women
the man attracts women
Adjectives following “be” can often be represented with OntoNotes verbal predicates:
(a / realize-01 NOT: (a / aware :ARG0 (s / soldier) :ARG1 (s / soldier) :ARG1 (b / battle)) :prep-of (b / battle))
The soldier was aware of the battle.
The soldier realized there was a battle.
Many adjectives have natural English verbal predicates:
- be aware (of X) –
- be worth (X) –
- be like (X) –
- be afraid (of X) –
Other adjectives do not, in which case we use the adjective as the predicate name.
- be responsible (for X) –
- be nervous (about X) –
- be serious (about X) –
- be efficient (at X) –
(r / responsible-03 :ARG0 (b / boy) :ARG1 (w / work))
The boy is responsible for the work.
The boy is responsible for doing the work.
The boy has the responsibility for the work.
In this way, we avoid awkward syntactic representations for English function words (like “for” and “has”). “The boy responsibles the work” isn’t good English, though it is good Chinese.
ARG0 often refers to the thing being described by the adjective, while ARG1 names the next most natural argument. We use ARG1/ARG2 if the adjective is not agentive.
How about adjectives like “sad”, “white”, and “free”? Should we use
free-01? Just because something is white, it
doesn’t mean that it was whitened. In such cases, we only use OntoNotes verbal
predicates if there is an implied event or process:
(s / sad-02 :ARG1 (g / girl))
The girl is sad.
(s / sadden-01 :ARG0 (d / disaster) :ARG1 (g / girl))
The girl was saddened by the disaster.
The disaster saddened the girl.
Almost all “-ed” adjectives (e.g., “acquainted”) immediately suggest OntoNotes verb frames. For example:
(a / acquaint-01 :ARG1 (b / boy) :ARG2 (m / magic))
the boy is acquainted with magic
:ARG0 in such cases (here, the acquainter) is usually unspecified. One may
quibble that a boy could become acquainted with magic without someone
acquainting him, but maybe he acquainted himself.
OntoNotes is a resourceful “-ed” tool. If you see “X was fed up with Y”, don't
be surprised that
feed-03 solves the AMR. What feeder fed Y to X? We just
By now, these AMRs should not surprise:
(w / want-01 :ARG0 (b / boy) :ARG1 (p / please-01 :ARG0 b))
Boys want to please.
(e / easy-05 :ARG1 (p / please-01 :ARG1 (g / girl)))
Girls are easy to please.
It is easy to please girls.
Pleasing girls is easy.
girl are closely related semantically (via
though not contiguous in “girls are easy to please”. When we see “girls are
easy to please”, we automatically re-formulate that as the more logical
“easy(please(girl))”. Likewise, with “you are safe to drink this” …
English adjectives can be formed from verbs and nouns in other ways, e.g., by adding “-able” or “-ful”. If the resulting adjective has its own idiosyncratic meaning, then we do not break it down further. But if we can break it down without introducing new concepts (only relations), then we go ahead:
(s / sandwich :ARG1-of (e / eat-01 :ARG1-of (p / possible-01)))
an edible sandwich
a sandwich that can be eaten
a sandwich that it is possible to eat
(s / sandwich :ARG1-of (e / eat-01 :ARG1-of (p / possible-01 :polarity -)))
an inedible sandwich
English is a wily opponent though. If you own a taxable fund, getting taxed is more than just a possibility, my friend:
(f / fund NOT: (f / fund :ARG3-of (t / tax-01)) :ARG3-of (t / tax-01 :ARG1-of (p / possible-01)))
a taxable fund
Adverbs with -ly
Adverbs normally get stemmed to adjective form:
(o / observe-01 :ARG0 (i / i) :ARG1 (m / move-01 :ARG0 (a / army) :manner (q / quick-02)))
I observed that the army moved quickly.
I observed the quick movement of the army
I observed the army moving quickly.
However, this is not the case when the adverb has a special meaning. For example, the meanings of 'hard' and 'hardly' are quite different:
(k / know-01 :ARG0 (i / i) :ARG1 (s / she) :degree (h / hardly))
I hardly know her.
We have seen roles like
:location. AMR includes other non-core roles:
(d / drive-01 :ARG0 (h / he) :direction (w / west) :source (c / city :wiki "Houston" :name (n / name :op1 "Houston")) :destination (c2 / city :wiki "Austin,_Texas" :name (n2 / name :op1 "Austin")))
He drove west, from Houston to Austin.
(d / drive-01 :ARG0 (i / i) :destination (c / city :wiki "Indianapolis" :name (n / name :op1 "Indianapolis")) :path (r / road :wiki "Interstate_65" :name (n2 / name :op1 "I-65")))
I drove to Indianapolis on I-65.
(d / drive-01 :ARG0 (i / i) :path (t / tunnel))
I drove through the tunnel.
(s / hum-02 :ARG0 (s2 / soldier) :ARG1 (t2 / tune) :beneficiary (g / girl) :time (w / walk-01 :ARG0 g :accompanier s2 :destination (t / town)))
The soldier hummed a tune for the girl as he walked with her to town.
(i / information :polarity - :topic (c / case))
There is no information about the case.
(p / person :wiki - :name (n / name :op1 "Jay" :op2 "Bartroff") :ARG0-of (h / have-org-role-91 :ARG1 (u / university :wiki "University_of_Southern_California" :name (n2 / name :op1 "USC")) :ARG2 (p2 / professor :mod (a / associate) :topic (m / mathematics))))
USC Associate Professor for Mathematics Jay Bartroff
(w / work-01 :ARG0 (h / he) :duration (t / temporal-quantity :quant 2 :unit (h2 / hour)))
He worked for two hours.
:instrument describes the physical object used in an action, typically a tool, device, weapon, or a body part such as a finger or fist.
If you want to describe how or in what way something is done, use the more general
(e / eat-01 :ARG0 (i / i) :ARG1 (p / pasta) :instrument (f / fork))
I ate pasta with a fork.
(a / attack-01 :ARG0 (c / country :wiki "Iraq" :name (n / name :op1 "Iraq")) :instrument (m / missile))
Iraq launched a missile attack.
:medium is used for channels of communications such as
a newspaper, a TV channel, the web, YouTube, Facebook, a speech, as well as languages:
(t / talk-01 :ARG0 (s / she) :ARG2 (h / he) :medium (l / language :wiki "French_language" :name (n / name :op1 "French")))
She talked to him in French.
(a / announce-01 :ARG0 (p / person :wiki - :name (n / name :op1 "John")) :ARG1 (b / bear-02 :ARG1 (p2 / person :ARG0-of (h / have-rel-role-91 :ARG1 p :ARG2 (s / son)))) :medium (p3 / product :wiki "Twitter" :name (n2 / name :op1 "Twitter")))
John announced the birth of his son on Twitter.
:manner is used to annotate any description answering "How is something done?" that is not already covered by a more specific role such as
(s / sing-01 :ARG0 (b / boy) :manner (b2 / beautiful-02 :degree (v / very)))
The boy sang very beautifully.
(d / decorate-01 :ARG0 (h / he) :ARG1 (r / room) :manner (c / creative))
He decorated the room in a creative way.
:manner also describes the method or action to do something, sometimes referred to as means:
(p / propose-01 :ARG0 (p3 / person :ARG0-of (h2 / have-org-role-91 :ARG2 (m / mayor))) :ARG1 (l / lower-05 :ARG1 (c / crime) :manner (h / hire-01 :ARG2 (p4 / person :ARG0-of (h3 / have-org-role-91 :ARG1 (p2 / police) :ARG2 (o / officer)) :mod (m2 / more)))))
The mayor proposed to lower crime by hiring more police officers.
:manner also describes modes of transportation:
(g / go-02 :ARG0 (p / person :wiki - :name (n / name :op1 "Nicole")) :ARG4 (c / country :wiki "England" :name (n2 / name :op1 "England")) :manner (t / train))
Nicole went to England by train.
(g / go-02 :ARG0 (h / he) :ARG4 (s / store) :purpose (b / buy-01 :ARG0 h :ARG1 (w / wood :purpose (f / fence :mod (n / new)))))
He went to the store to buy wood for a new fence.
(m / murmur-01 :ARG0 (b / boy) :manner (s / soft-02) :purpose (s2 / soothe-01 :ARG1 (g / girl)) :cause (w / worry-01 :ARG0 b :topic g))
The boy murmured softly to soothe the girl, because he worried about her.
Note: The AMR Editor automatically reifies :cause and :cause-of to cause-01, so we call
:cause a shortcut.
(c / continue-01 :ARG1 (g / game) :concession (r / rain-01))
The game continued although it rained.
The game continued despite the rain.
(c / continue-01 :ARG1 (g / game) :concession (e / even-if :op1 (r / rain-01)))
The game will continue even if it rains.
(f / fear-01 :ARG0 (t / they) :ARG1 (h / he) :concession (e / even-when :op1 (i / imprison-01 :ARG1 h)))
They feared him even when he was behind bars.
(s / sing-01 :ARG0 (b / boy) :condition (g / give-01 :ARG1 (m / money) :ARG2 b))
The boy will sing if he is given money.
If the boy is given money, he will sing.
The boy will sing in case of a money donation.
(s / sing-01 :ARG0 (b / boy) :polarity - :condition (g / give-01 :ARG1 (m / money) :ARG2 b))
The boy will sing unless he is given money.
Unless the boy is given money, he will sing.
X :cause Y means that the cause of X is Y. Likewise,
Y :cause-of X
means Y is the cause of X. (See the section Reification below
about using the concept
cause-01 instead of
(s / strike-01 :ARG0 (t / torpedo) :cause-of (d / damage-01 :ARG1 (s2 / ship)))
The torpedo struck, causing the ship to be damaged.
The torpedo struck, causing damage to the ship.
The torpedo struck, damaging the ship.
It can be difficult to tease apart
:cause. For example, “I
visited her because she was sick” (cause) or “I visited her to deliver the news”
In the AMR Editor, :cause/:cause-of is automatically converted to cause-01. (See section on reification below.)
Occasionally, a numbered
:ARGx role will refer exactly to
:beneficiary, or some other named non-core role. In this case, we use the
(p / provide-01 NOT: (p / provide-01 :ARG0 (b / boy) :ARG0 (b / boy) :ARG1 (c / chocolate) :ARG1 (c / chocolate) :ARG2 (g / girl)) :beneficiary (g / girl))
The boy provided chocolate to the girl.
The boy provided the girl with chocolate.
Sometimes it isn’t clear what
:time, etc., should modify in AMRs
involving creation events. We tend to put them on the event, rather than on the
(b / build-01 NOT: (b / build-01 :ARG0 (t / they) :ARG0 (t / they) :ARG1 (b2 / bridge) :ARG1 (b2 / bridge :location …)) :location (s / state :wiki "Maryland" :name (n / name :op1 "Maryland")) :time (d / date-entity :month 12))
They built the bridge in Maryland in December.
The bridge was built by them in Maryland in December.
Here are more non-core roles:
(e / engine :part-of (c / car))
the engine of the car
the car’s engine
(u / unit :part-of (c / company))
a unit of the company
the company’s unit
We do not use
:part for set membership, as in the CEO of a company.
(s / south :part-of (c / country :wiki "France" :name (n / name :op1 "France")))
the south of France
(w / win-01 :ARG0 (b / boy) :ARG1 (r / race-01 :subevent-of (g / game :wiki "Olympic_Games" :name (n / name :op1 "Olympics"))))
The boy won the race in the Olympics.
(r / ring :consist-of (g / gold))
a ring of gold
(t / team :consist-of (m / monkey))
a team of monkeys
(c / company :example (a / and :op1 (c2 / company :wiki "IBM" :name (n / name :op1 "IBM")) :op2 (c3 / company :wiki "Google" :name (n2 / name :op1 "Google"))))
companies like IBM and Google
(d / drive-01 :ARG0 (h / he) :direction (w / west))
He drove west.
:frequency describes how often something occurs.
(m / meet-03 :frequency 3 :ARG0 (w / we))
We met three times.
The special frame
rate-entity-91 is used to describe recurring events
and other rate entities such as "every 3000 miles" or "$3 per gallon".
(r / rate-entity-91 :ARG1 2 :ARG2 (t / temporal-quantity :quant 1 :unit (y / year)))
twice a year
(p / play-01 :ARG0 (w / we) :ARG1 (b / bridge) :frequency (r / rate-entity-91 :ARG4 (d / date-entity :weekday (w2 / wednesday) :dayperiod (a / afternoon))))
We play bridge every Wednesday afternoon.
Core roles of
rate-entity-91is the quantity (default: 1)
rate-entity-91is the reference quantity ("per quantity")
rate-entity-91is any regular interval between events ("every 2 months" - more specific than
rate-entity-91is any entity on which recurring events happen
(g / go-on-15 :ARG1 (r / road) :extent (f / forever))
The road goes on forever.
Inverse relations are often used for focusing (see Introduction):
(s / sing-01 :ARG0 (b / boy :source (c / college)))
The boy from the college sang.
(b / boy :ARG0-of (s / sing-01) :source (c / college))
The singing boy from the college
There is a boy from the college who sang.
(c / college :source-of (b / boy :ARG0-of (s / sing-01)))
The college that the singing boy came from
The concept of focus only applies at the very top (root) of the AMR. After a
root concept is selected, there are no more focus considerations -- all else is
driven strictly by semantic relations. For example, once we have selected
c / college at the root, then
:source-of must be filled with
boy, not with
Sometimes we want to use an AMR relation as a first-class concept. Converting a
role into a concept is called reification. Here is an example, where the
:cause is replaced by
cause-01. Instead of
x :cause y, we have
x :ARG1-of (c / cause-01 :ARG0 y).
AMR without reification: AMR with reification: (l / leave-11 (l / leave-11 :ARG0 (g / girl) :ARG0 (g / girl) :cause (a / arrive-01 :ARG1-of (c / cause-01 :ARG1 (b / boy))) :ARG0 (a / arrive-01 :ARG1 (b / boy))))
The girl left because the boy arrived.
AMR without reification is simpler, so why would we want to reify? One reason is to make a relation the focus of an AMR fragment. For example, suppose we know there is a knife in the drawer. We might try focusing on the knife:
(k / know-01 :ARG0 (w / we) :ARG1 (k2 / knife :location (d / drawer)))
We know the knife that is in the drawer. (???)
Or we might try focusing on the drawer:
(k / know-01 :ARG0 (w / we) :ARG1 (d / drawer :location-of (k2 / knife)))
We know the drawer where the knife is. (???)
But we really want to focus on the locating itself. AMR therefore supplies
reifications for many relations. In the case of
:location, the reification is
be-located-at-91, allowing us to say:
(k / know-01 :ARG0 (w / we) :ARG1 (b / be-located-at-91 :ARG1 (k2 / knife) :ARG2 (d / drawer)))
We know the knife is in the drawer.
be-located-at-91 has two roles,
:ARG1 (the thing that exists in
:ARG2 (where the thing is).
We also use reification when we want to modify a relation. For example:
(k / know-01 :ARG0 (w / we) :ARG1 (b / be-located-at-91 :ARG1 (k2 / knife) :ARG2 (d / drawer) :polarity - :time (y / yesterday)))
We know the knife was not in the drawer yesterday.
Here are the AMR reifications. Reifications often correspond to OntoNotes
predicates, in which case, we just use the
:ARG relations in the natural way.
Read this chart as:
x :Relation y=
x :Domain-of (z / Reification :Range y)
x :location y=
x :ARG0-of (b / be-located-at-91 :ARG1 y)
||“she's with him”|
||“she's 41 years old”|
||“the 5k run is for kids”|
||“he came 'cause of her”|
||“he came despite of her”|
||“he comes if she comes”|
||“very tall” (intensifier or downtoner)|
||“i'm off to Atlanta”|
||“it's 15 minutes long”|
||“cities such as Atlanta”|
||“trip was 2500 miles”|
||“he came three times”|
||“forks are for eating”|
||“she's not here”|
||“it was done quickly”|
||“he is half Chinese"|
||“the city formerly named Constantinople”|
||“the roof of the house”|
||“I don't know.”|
||“that dog's not mine”|
||“it's to eliminate bugs”|
||“there are 4 rabbits”|
||“she's from Ipanema”|
||“presentation at a conference”|
||“10% of the workers”|
||“the party is on friday”|
||“the show's about me”|
||“The phone number is 1-800-555-1223.”|
These relations do not have reifications:
Now, the question remains: when to reify?
One potential answer is “whenever you feel like it”. Unfortunately, a single sentence may receive two different AMRs. Either of the following AMRs could reasonable represent “The girl left because the boy arrived”, with neither being canonical:
AMR without reification: AMR with reification: (l / leave (l / leave :ARG0 (g / girl) :ARG0 (g / girl) :cause (a / arrive :ARG1-of (c / cause-01 :ARG0 (b / boy))) :ARG0 (a / arrive :ARG0 (b / boy))))
Furthermore, we might reasonably tend to use the first AMR for “The girl left because the boy arrived”, but prefer the second one for “The girl left, due to the boy's arrival”. So we cannot guarantee that these two sentences get the same AMR.
A second potential answer is “reify all the time”, which would completely
eliminate relations like
:subset from AMR, in favor of
include-91. But this is
cumbersome -- it is easy and typical to simply type
The resolution: we consider “AMR with reification” to be “real AMR”, with non-reified relations as semantic sugar. Therefore, if you are translating English into AMR, the rule is “whenever you feel like it”, because your AMRs will be normalized into reified form behind the scenes.
In the AMR corpus, for consistency, reifications are collapsed unless needed.
AMR strips away light-verb constructions:
(a / adjust-01 :ARG0 (g / girl) :ARG1 (m / machine))
The girl adjusted the machine.
The girl made an adjustment to the machine.
(t / bathe-01 :ARG0 (b / boy))
The boy bathed.
The boy took a bath.
It also combines verb-particle constructions, using OntoNotes predicate frames.
look-05 is defined as “look up: seek”.
(l / look-up-05 :ARG0 (b / boy) :ARG1 (a / answer))
The boy looked up the answer.
The boy looked the answer up.
Sometimes a particle doesn’t change the meaning of the verb very much, but
OntoNotes may still have two separate predicates. For example,
“become nearer”, while
close-in-07 is “close in: become nearer”. In such cases,
AMR canonicalizes to the non-particle frame, e.g.,
Most prepositions that signal semantic frame elements are dropped in AMR:
(s / default-01 :ARG1 (n / nation) :time (d / date-entity :month 6))
The nation defaulted in June.
(d / die-01 :ARG1 (m / man) :location (h / house :poss m))
The man died in his house.
But time and location prepositions are kept if they carry additional
information, using AMR’s
:op1 is different from the
:op1 used in
(s / default-01 :ARG0 (n / nation) :time (b / after :op1 (w / war-01)))
The nation defaulted after the war.
(d / die-01 :ARG1 (m / man) :location (n / near :op1 (h / house :poss m)))
The man died near his house.
(d / die-01 :ARG1 (m / man) :location (b / between :op1 (h / house) :op2 (r / river)))
The man died between the house and the river.
Sometimes, the content of a prepositional phrase cannot be easily slotted into a
predicate-argument structure, or into a generic role like
AMR cringes while employing a default
(s / sue-01 :ARG1 (h / he) :prep-in (c / case))
He was sued in the case.
AMR combines phrasal prepositions:
(f / file-01 :ARG1 (b / brief) :prep-on-behalf-of (g / government))
The brief was filed on behalf of the government.
By tradition, the frequent phrase “according to” gets special handling:
(s / say-01 :ARG0 (s2 / source :mod (g / government)) :ARG1 (k / kill-01 :time (y / yesterday)))
According to government sources, the killing happened yesterday.
Government sources said that the killing happened yesterday.
AMR frequently represents relative clauses with inverse roles, as described in the Introduction:
(b / believe-01 :ARG0 (b2 / boy))
The boy believes.
(b / boy :ARG0-of (b2 / believe-01))
the boy who believes
English also uses relative clauses when negating a pre-nominal adjective is difficult (“the not-black car”):
(c / car :ARG1-of (b / black-04))
the black car
(c / car :ARG1-of (b / black-04 :polarity -))
the car that is not black
Japanese simply marks adjectives with a negative suffix.
Multiple relations with the same name
An entity may have several relations with the same name.
In the following examples, system has two
:mod relations, arrive-01 has two
:time relations, and boy has two
(s / system :mod (l / law) :mod (c / city :wiki "Shanghai" :name (n / name :op1 "Shanghai")))
the Shanghai legal system
(a / arrive-01 :ARG1 (s / she) :time (t / today) :time (b / before :op1 (n / now)))
She arrived earlier today.
(b / boy :ARG0-of (w / want-01 :ARG1 (b2 / believe-01 :ARG1 (g / girl))) :ARG0-of b2)
the boy who wants to believe the girl
To represent conjunction, AMR uses concepts
neither, along with
(a / and :op1 (b / boy) :op2 (g / girl))
the boy and the girl
(a / either :op1 (b / boy) :op2 (g / girl) :op3 (d / dog))
either the boy, the girl, or the dog
Conjoined adjectives are done without
(b / ball :mod (b2 / big) :mod (h / heavy))
the big, heavy ball
the big and heavy ball
:opx is also used for clauses:
(a / and :op1 (c / shout-01) :op2 (l / leave-11 :ARG0 (b / boy)))
There was shouting, and the boy left.
(c / contrast-01 :ARG1 (c2 / shout-01) :ARG2 (l / stay-01 :ARG1 (b / boy)))
There was shouting, but the boy stayed.
:ARG1 may be missing:
(c / contrast-01 :ARG2 (l / stay-01 :ARG1 (b / boy)))
But the boy stayed.
AMR aims for a logical representation even when English elides core actors:
(a / and :op1 (c / shout-01 :ARG0 (b / boy)) :op2 (l / leave-11 :ARG0 b))
The boy shouted and left.
The need for this is evident when an entity plays different roles in different predicates:
(a / and :op1 (a2 / arrive-01 :ARG1 (b / boy)) :op2 (l / kill-01 :ARG1 b :manner (p / prompt)))
The boy arrived and was promptly killed.
However, AMR “pulls out” non-core roles like
modifies the entire conjunction rooted by
(a / and :time (d / date-entity :weekday (t / tuesday)) :op1 (a2 / arrive-01 :ARG1 (b / boy)) :op2 (l / leave-11 :ARG0 b))
The boy arrived and left on Tuesday.
On Tuesday, the boy arrived and left.
Quantifiers and scope
AMR does not have a deep representation for quantifiers. It only canonicalizes their position:
(l / leave-01 :ARG0 (b / boy :mod (a / all)))
The boys all left.
All the boys left.
Each of the boys left.
(l / leave-01 :ARG0 (b / boy :mod (n / no)))
No boy left.
None of the boys left.
(l / leave-01 :ARG0 (b / boy :mod (a / all :polarity -)))
Not all of the boys left.
(l / leave-01 :ARG0 (p / person :mod (a / all :polarity -)))
Not everyone left.
The placement of
:polarity can be troublesome. Consider:
(b / believe-01 :ARG0 (g / girl) :ARG1 (w / work-01 :ARG0 (b2 / boy) :manner (h / hard-02)))
The girl believes that the boy works hard.
If we want to represent “the girl doesn’t believe that the boy works hard”, we have to decide whether to place the negative polarity under “believe” or “work” or “hard”. Here it should go under “hard”:
(b / believe-01 :ARG0 (g / girl) :ARG1 (w / work-01 :ARG0 (b2 / boy) :manner (h / hard-02 :polarity -)))
The girl believes that the boy works in a not-hard manner.
The girl believes that the boy doesn’t work hard. (colloquially)
The girl doesn’t believe that the boy works hard. (colloquially)
If we put
:polarity elsewhere, we change the meaning:
(b / believe-01 :ARG0 (g / girl) :ARG1 (w / work-01 :polarity - :ARG0 (b2 / boy) :manner (h / hard-02)))
The girl believes that the boy refrains from work, in a hard manner.
(b / believe-01 :polarity - :ARG0 (g / girl) :ARG1 (w / work-01 :ARG0 (b2 / boy) :manner (h / hard-02)))
It’s not true that the girl believes the boy works hard.
(b / believe-01 :ARG0 (g / girl :polarity -) :ARG1 (w / work-01 :ARG0 (b2 / boy :polarity -) :manner (h / hard-02)))
The non-girl believes that the non-boy works hard.
AMR apologizes for not advising on the placement of negation with respect to quantifiers.
extremely) and downtoners (
relatively) are annotated with
(b / beautiful-02 :ARG1 (y / you) :degree (v / very))
You are very beautiful.
Comparatives and superlatives are represented using the frame
Have-degree-91 Arg1: domain, entity characterized by attribute (e.g. girl) Arg2: attribute (e.g. tall) Arg3: degree itself (e.g. more, less, equal, most, least, enough, too, so, to-the-point, at-least, times) Arg4: compared-to (e.g. (than the) BOY) Arg5: superlative: reference to superset Arg6: reference, threshold of sufficiency (e.g. (tall enough) TO RIDE THE ROLLERCOASTER)
Annotators are encouraged to use
have-degree-91 as the root concept (as opposed to the adjective with a particular degree, or the entity characterized by that adjective) when a comparison seems to be the main focus of the sentence, which include cases of the copular construction (e.g., the girl is taller than the boy, she is the tallest girl on the team—see below).
(b / boy :ARG1-of (h / have-degree-91 :ARG2 (b2 / bright-03 :ARG1 b) :ARG3 (m / more)))
The brighter boy.
(b / boy :ARG1-of (h / have-degree-91 :ARG2 (b2 / bright-03 :ARG1 b) :ARG3 (m / most)))
The brightest boy.
(p / plan :ARG1-of (h / have-degree-91 :ARG2 (g / good-02 :ARG1 p) :ARG3 (m / more)))
a better plan
(p / plan :ARG1-of (h / have-degree-91 :ARG2 (b / bad-07 :ARG1 p) :ARG3 (m / more)))
a worse plan
(p / plan :ARG1-of (h / have-degree-91 :ARG2 (e / extreme) :ARG3 (t / too)))
a plan that is too extreme
(h / have-degree-91 :ARG1 (g / girl) :ARG2 (t / tall) :ARG3 (m / more) :ARG4 (b / boy))
the girl is taller than the boy
have-degree-91 is also used for Superlative constructions, which invoke a subset/superset relation. This takes precedence over
include-91, which could also be used in these cases to express the subset/superset relation between the thing that is “the most” and superset of items being compared to:
(h / have-degree-91 :ARG1 (s / she) :ARG2 (t / tall) :ARG3 (m / most) :ARG5 (g / girl :ARG0-of (h2 / have-org-role-91 :ARG1 (t2 / team))))
she is the tallest girl on the team
have-degree-91 is used for the degree-consequence construction, which licenses an argument representing the result or consequence of the degree to which a state is true:
(h / have-degree-91 :ARG2 (e / early) :ARG3 (t / too) :ARG6 (c / conclude-01))
It is too early to reach any conclusion.
Note: The primary relation of the Consequence argument may be elided or unspecified, as it is here. Annotators should introduce the most logical relation fitting with the context. Although it may be tempting to introduce modality, possibility, and/or polarity in the consequence (i.e. expressing that there is NO conclusion), a survey of these constructions has demonstrated the difficulty in introducing such elements consistently in context. Therefore, annotators are asked to limit the Arg6 to the simple relation to which the degree has reference. See AMR Dictionary for additional examples. Negative polarity should only be introduced (modifying
have-degree-91) when it is explicit:
(h / have-degree-91 :polarity - :ARG1 (h2 / he) :ARG2 (t / tall) :ARG3 (e / enough) :ARG6 (r / ride-01 :ARG0 h2 :ARG1 (r2 / rollercoaster)))
He is not tall enough to ride the rollercoaster.
For a more thorough discussion, see Abstract Meaning Representation of Constructions: The More We Include, the Better the Representation by Bonial et al. 2018.
Variables and co-reference
If two variables are the same, then they refer to the same entity:
(w / want-01 :ARG0 (b / boy) :ARG1 (g / go-02 :ARG0 b))
The boy wants to go.
In English, overt and zero pronouns are often used to realize co-reference, but AMR uses variables instead:
(w / want-01 :ARG0 (b / boy) :ARG1 (b2 / believe-01 :ARG0 b :ARG1 b))
The boy wants to believe himself.
If an overt pronoun has no antecedent within the sentence, AMR uses the pronoun:
(s / see-01 :ARG0 (h / he) :ARG1 (t / they))
He saw them.
Pronouns in AMR are always nominative (he, she, i, we, they, …) and never accusative (him, her, me, us, them, …).
:poss (“possessed by”) is a very general form of possession. AMR
uses it only for possessives and prepositional phrases with “of”:
(c / car :poss (h / he))
the car of his
(t / titan :poss (n / nation))
the nation’s titans
Not all possessives and “of” phrases are represented with
:poss. AMR often uses
Pertainym adjectives (e.g., “atomic, adj. = of, or pertaining to, atoms”) do not
appear in AMR. Only the stemmed noun form is used, along with the
(v / virus :mod (m / microbe))
(w / war :mod (a / atom))
A pertainym may get stemmed to noun form, and then subsequently to verb form:
(p / problem :mod (b / behave-01))
Problems with behavior.
When building AMR from English, the rule is to continue stemming toward verb form unless the meaning is significantly altered.
We use the role
:ord and the concept
ordinal-entity to express ordinals.
(p / planet :ord (o / ordinal-entity :value 2))
the second planet
planet number 2
(v / visit-01 :ARG0 (w / we) :ord (o / ordinal-entity :value 1 :range (t / temporal-quantity :quant 10 :unit (y / year))))
our first visit in 10 years
We often refer to subsets when we speak. AMR uses roles
(d / die-01 :ARG1 (s / soldier :quant 9 :subset-of (s3 / soldier :quant 20)))
Nine of the twenty soldiers died.
(h / have-03 :ARG0 (p4 / person :quant 4 :subset-of (p2 / person :ARG0-of (s / survive-01) :quant 5) :subset (p3 / person :quant 3 :ARG1-of (d3 / diagnose-01))) :ARG1 (d / disease))
Four of the five survivors had the disease, including three who were diagnosed.
Features shared by a subset and its superset go into the superset only, e.g.,
“survive” above. The reification of
include-91, so we can
(d / die-01 :ARG1 (s / soldier :quant 9 :ARG1-of (i / include-91 :ARG2 (s3 / soldier :quant 20))))
Nine of the twenty soldiers died.
AMR is sparing with
:subset -- otherwise things get out of control. For
example, we do not use it for “Three of the workers at the plant,” but we rather
just interpret this as “Three workers”.
Note: The AMR Editor automatically reifies :subset, :subset-of, :superset and :superset-of to include-91, so we call them shortcuts.
Any concept instance in AMR can have
:name role. We are not restricted to a
small set of fixed categories like countries and people: ships, pets, and
computers can also have names.
Named entities are often referred to in different ways, such as "US", "U.S.", "United States" or "United States of America". To annotate such named entities with a canonical form, we use Wikipedia (for English) as a standard and mark up NEs with a
:wiki role such as
:wiki "United_States". This is also useful when names themselves are ambiguous (such as "Washington" the city, state, or person; or "Clinton" or "Bush"). If there is no Wikipedia page for a given named entity, it is marked as
(p / person :wiki "Margaret_Brown" :name (n / name :op1 "Mollie" :op2 "Brown"))
(p / person :wiki "Margaret_Brown" :name (n / name :op1 "Mollie" :op2 "Brown") :ARG0-of (s / slay-01 :ARG1 (o / orc)))
the orc-slaying Mollie Brown
Mollie Brown, who slays orcs
(s / ship :wiki "RMS_Titanic" :name (n / name :op1 "Titanic"))
the ship named Titanic
(c / city :wiki "Marina_del_Rey,_California" :name (n / name :op1 "Marina" :op2 "del" :op3 "Rey"))
Marina del Rey
the city of Marina del Rey
AMR strings words with
:opN. It does not analyze semantic relationships inside
a named entity. For example, in the “Stop Malaria Foundation”, we do not invoke
malaria as its
Abbreviations of proper names are not expanded, but abbreviated common nouns are expanded. Speaking of common words, we correct typos, and we normalize to American spelling, but we do not otherwise normalize variants.
(s / state :wiki "California" :name (n / name :op1 "Calif."))
(r / rate :mod (a / advertise-01))
When building AMRs for proper names or “-er” nouns, we need to fill the root concept (or top-level instance role). In doing so, we face one of three situations.
(a) In general, unless the English text provides something more specific type,
we fill the instance slot from a special list of standard AMR named entity types,
In such cases, we basically must hallucinate an entity type. For example:
(p / person :wiki - :name (n / name :op1 "Pascale"))
(c / company :ARG0-of (m / make-01 :ARG1 (c2 / chip)))
the chip maker
However, we do not want some AMRs to say “person” and others to say “woman”, or some to say “company”, and others to say “organization”.
So when we are forced to hallucinate an entity type, AMR requires us to draw from this canonical list (borrowing from information extraction and question answering):
- person, family, animal, language, nationality, ethnic-group, regional-group, religious-group, political-movement
- organization, company, government-organization, military, criminal-organization, political-party, market-sector, school, university, research-institute, team, league
- location, city, city-district, county, state, province, territory, country, local-region, country-region, world-region, continent; ocean, sea, lake, river, gulf, bay, strait, canal; peninsula, mountain, volcano, valley, canyon, island, desert, forest moon, planet, star, constellation
- facility, airport, station, port, tunnel, bridge, road, railway-line, canal, building, theater, museum, palace, hotel, worship-place, market, sports-facility, park, zoo, amusement-park
- event, incident, natural-disaster, earthquake, war, conference, game, festival
- product, vehicle, ship, aircraft, aircraft-type, spaceship, car-make, work-of-art, picture, music, show, broadcast-program
- publication, book, newspaper, magazine, journal
- award, law, court-decision, treaty, music-key, musical-note, food-dish, writing-script, variable, program
- molecular-physical-entity, small-molecule, protein, protein-family, protein-segment, amino-acid, macro-molecular-complex, enzyme, nucleic-acid
- pathway, gene, dna-sequence, cell, cell-line, species, taxon, disease, medical-condition
We always choose the most specific applicable type.
If none of these apply, then we use thing.
(a / award :wiki "Nobel_Prize" :name (n / name :op1 "Nobel" :op2 "Prize"))
the Nobel Prize
(g / government-organization :wiki "National_Security_Agency" :name (n / name :op1 "NSA") :mod (c / country :wiki "United_States" :name (n2 / name :op1 "America")))
the American NSA
(n / natural-object :wiki "Lone_Cypress" :name (n2 / name :op2 "Lone" :op3 "Cypress"))
the Lone Cypress
(b) If the text contains a more specific English term to describe the type of entity, we use it instead to fill the instance role. For example:
(p / poet :wiki "William_Shakespeare" :name (n / name :op1 "William" :op2 "Shakespeare"))
the poet William Shakespeare
William Shakespeare, the poet
poet is more specific than
(v / village :wiki - :name (n / name :op1 "Odinaboi"))
the village of Odinaboi
village is more specific than
(d / doctor :wiki - :name (n / name :op1 "Wu"))
The following example texts mention
but in these case we prefer the standard NE types
because the latter are more (or at least equally) specific:
(c / country-region :wiki "Darfur" :name (n / name :op1 "Darfur") :location (c2 / country :wiki "Sudan" :name (n2 / name :op1 "Sudan")))
Sudan's Darfur region
(p / political-party :wiki "Christian_Democratic_Union_of_Germany" :name (n / name :op1 "CDU") :mod (c / conservative) :mod (c2 / country :wiki "Germany" :name (n2 / name :op1 "Germany")))
Germany's conservative CDU party
(s / spaceship :wiki "Shenzhou_(spacecraft)" :name (n / name :op1 "Shenzhou"))
the spaceship Shenzhou
the Shenzhou spacecraft
Mere honorifics such as “Mr.”, “Mrs.”, etc. are included as part of the name:
(p / person :wiki - :name (n / name :op1 "Mr." :op2 "Wu"))
See the next section on "Special Frames for Roles" on how to annotate titles such as "President".
When faced with an appositive, AMR calmly inserts facts into slots:
(g / group :wiki "Reed_Elsevier#Elsevier_NV" :name (e / name :op1 "Elsevier" :op2 "N.V.") :mod (c / country :wiki "Netherlands" :name (h / name :op1 "Netherlands")) :ARG0-of (p2 / publish-01))
Elsevier N.V. , the Dutch publishing group
We view this object semantically as a “group”, which happens to have a known
:name, plus some a couple of other properties that describe it.
(c) The text contains multiple English words vying for the same instance slot.
This happens occasionally. Because instance is the only relation that cannot
physically appear twice in AMR, we instead open up the inverse of
:domain , i.e. the role
(d / doctor :wiki "Dr._Seuss" :name (n / name :op1 "Seuss") :mod (p / poet))
the poet Dr. Seuss
In all cases, hyphenated and possessive words inside names are kept intact, not broken up.
For example, "Dana-Farber Materials" only has
Special Frames for Roles
For roles in organizations, we use the frame have-org-role-91:
(p / person :wiki "Barack_Obama" :name (n / name :op1 "Obama") :ARG0-of (h / have-org-role-91 :ARG1 (c / country :wiki "United_States" :name (n2 / name :op1 "US")) :ARG2 (p2 / president)))
US President Obama
Core roles of
have-org-role-91is the office holder, typically a person
have-org-role-91is the organization, which could also be a GPE
have-org-role-91is the title of the office held, e.g. president
have-org-role-91is a description of responsibility (rarely used)
Typical have-org-role-91 roles: ambassador, archbishop, bishop, CEO, chairman, chancellor, chief of staff, commissioner, congressman, deputy, dictator, director, emperor, empress, envoy, foreign minister, governor, king, mayor, monarch, officer, official, pope, premier, president, principal, professor, queen, secretary, senator, spokesman, spokeswoman, treasurer etc.
For roles that describe the relation between two people (or two other entities of the same type), we use the frame have-rel-role-91:
(h / have-rel-role-91 :ARG0 (h2 / he) :ARG1 (i / i) :ARG2 (b / brother-in-law))
He is my brother-in-law.
Core roles of
have-rel-role-91role of entity A (must be specified)
have-rel-role-91role of entity B (often left unspecified)
have-rel-role-91relationship basis (contract, case; rarely used)
Typical have-rel-role-91 roles: father, sister, husband, grandson, godfather, stepdaughter, brother-in-law; friend, boyfriend, buddy, enemy; landlord, tenant etc.
AMR normalizes numbers:
(b / boy :quant 40000)
forty thousand boys
(a / atom :quant 1500000000)
one and half billion atoms
1.5 billion atoms
a billion and half atoms
Such normalization is often necessary when we translate between Asian-style 10,000-based numeration and Western-style 1,000-based numeration.
Approximate numbers are represented with this
(b / boy :quant (s / several :op1 100))
several hundred boys
(b / boy :quant (m / more-than :op1 4000))
more than four thousand boys
more than 4000 boys
(b / boy :quant (m / between :op1 4000 :op2 5000))
between 4000 and 5000 boys
between four and five thousand boys
Exact quantities are represented by their type and
(q / distance-quantity :unit (m / mile) :quant 10)
Approximate quantities are represented using
:opN notation, as for approximate
(a / about :op1 (q / distance-quantity :unit (m / mile) :quant 10))
about 10 miles
AMR views quantified expressions like “two gallons of milk” as “milk”:
(b / buy-01 :ARG0 (w / woman) :ARG1 (m / milk :quant (q / volume-quantity :unit (g / gallon) :quant 2)))
The woman bought two gallons of milk.
For stretches of time and relative times, AMR uses
(For absolute times, AMR uses
date-entity, described in the
(t / temporal-quantity :unit (y / year) :quant 30)
(b / before :op1 (n / now) :duration (t / temporal-quantity :unit (y / year) :quant 30))
during the past 30 years
(b / before :op1 (n / now) :quant (t / temporal-quantity :unit (y / year) :quant 30))
30 years ago
(b / before :op1 (n / now) :quant (m / more-than :op1 (t / temporal-quantity :unit (y / year) :quant 30)))
more than 30 years ago
Disjunctions go high:
(o / or :op1 (t / temporal-quantity :unit (y / year) :quant 3) :op2 (t2 / temporal-quantity :unit (y2 / year) :quant 4))
three or four years
(o / or :op1 (t / temporal-quantity :unit (m / month) :quant 6) :op2 (t2 / temporal-quantity :unit (y / year) :quant 1))
six months or a year
Relative positions often include a quantity:
(c / crash-01 :ARG1 (p / plane) :location (r / relative-position :op1 (g / city :wiki "Moscow" :name (n / name :op1 "Moscow")) :quant (d / distance-quantity :unit (m / mile) :quant 50) :direction (e / east)))
The plane crashed 50 miles east of Moscow.
The plane crash occurred 50 miles east of Moscow.
The X-quantity notation is only used for precise quantities. Vague quantities
still use the
(g / gather-01 :ARG0 (p / person :quant (n / number :mod (l / large))))
A large number of people gathered.
Occasionally, the measurement itself is the primary concept:
(i / increase-01 :ARG1 (n / number :quant-of (p / person)))
The number of people increased.
Quantity types include:
(q / monetary-quantity :quant 20 :unit (d / dollar :mod (e / country :wiki "Canada" :name (n / name :op1 "Canada"))))
20 Canadian dollars
Quantities where a
:quant 0 value does not represent a 0-quantity use
(q / seismic-quantity :quant 7.9 :scale (r / richter))
7.9 on the Richter scale
have-quant-91 is the reification of
:quant. Its arguments parallel
have-degree-91, and it is used for comparisons and superlatives relating to amounts of things as opposed to qualities of things. See Dictionary examples for a list illustrating when to use
have-quant-91 and when to use
have-quant-91 ARG1: entity (thing being quantified) ARG2: quantity (numerical or quantifier: many, much) ARG3: degree mention (more, less, equal, too) ARG4: compared-to ARG5: superlative: reference to superset ARG6: consequence, result
(s / sell-01 :ARG0 (h / he) :ARG1 (c / car :ARG1-of (h2 / have-quant-91 :ARG3 (e2 / equal) :ARG4 (c3 / car :ARG1-of (s2 / sell-01 :ARG0 (p / person :ARG0-of (c2 / compete-02 :ARG1 h)))))))
He sold as many cars as his competitor.
(s / sell-01 :ARG0 (h / he) :ARG1 (c / car :ARG1-of (h2 / have-quant-91 :ARG3 (m / most) :ARG5 (c3 / car :ARG1-of (s2 / sell-01 :ARG0 (p / person :ARG0-of (c2 / compete-02 :ARG1 h)))))))
He sold the most cars of his competitors.
(h / have-03 :ARG0 (i / i) :ARG1 (w / water :purpose (d2 / drink-01 :ARG0 i) :ARG1-of (h3 / have-quant-91 :ARG3 (e / enough :mod (s / scarce)) :ARG6 (l / last-03 :ARG1 w :ARG2 (t / temporal-quantity :quant 1 :unit (w2 / week)) :ARG3 i))))
I had scarcely enough drinking water to last a week.
The special concepts
sum-of support the mathematical
operators that the meaning of a text might include.
(r / reach-01 :ARG0 (v / velocity :poss (a / aircraft)) :ARG1 (p / product-of :op1 3 :op2 (s / speed :poss (s2 / sound))))
The aircraft's velocity reached three times the speed of sound.
(f / finish-01 :ARG0 (p / person :wiki "Patrick_Makau_Musyoki" :name (n / name :op1 "Patrick" :op2 "Makau")) :ARG1 (r / run-02 :ARG0 p :ARG1 (m / marathon) :duration (s2 / sum-of :op1 (t2 / temporal-quantity :quant 2 :unit (h / hour)) :op2 (t3 / temporal-quantity :quant 3 :unit (m2 / minute)) :op3 (t4 / temporal-quantity :quant 38 :unit (s3 / second)))))
Patrick Makau finished the marathon in 2 hours, 3 minutes and 38 seconds.
Other entities: dates, times, percentages, phone, email, URLs
These entities are described in standard, canonical forms:
(d / date-entity :year 2012 :month 2 :day 29)
February 29, 2012
29 February 2012
(d / date-entity :year 2012)
the year 2012
(d / date-entity :month 4)
(d / date-entity :weekday (f / friday))
(d / date-entity :year 2012 :month 2)
(d / date-entity :month 2 :day 29 :weekday (w / wednesday))
Wednesday, February 29
(d / date-entity :day 29)
(d / date-entity :month 2 :day 29 :weekday (w / wednesday) :time 16:30 :timezone (z / PST))
Wednesday, February 29, 16:30 PST
(d / date-entity :time 16:30)
4:30 in the afternoon
half past four
(d / date-entity :era (h / heisei) :year 24 :month 2 :day 29 :calendar (j / country :wiki "Japan" :name (n / name :op1 "Japan")))
February 29, 24th year of Heisei era
(d / date-entity :year 2011 :quarter 4)
4th quarter, 2011
(d / date-entity :year 2011 :season (s / summer))
(d / date-entity :year 2011 :year2 2012 :season (w / winter))
(d / date-entity :year 2011 :year2 2012 :calendar (y / year :mod (a / academia)))
academic year 2011-2012
(d / date-entity :year 2012 :calendar (y / year :mod (f / finance) :mod (g / government-organization :ARG0-of (g2 / govern-01 :ARG1 (c / country :wiki "United_States" :name (n / name :op1 "United" :op2 "States"))))))
United States government fiscal year 2012
(d / date-interval :op1 (d2 / date-entity :year 2012 :month 3 :day 8) :op2 (d3 / date-entity :year 2012 :month 3 :day 9))
March 8-9, 2012
(d / date-interval :op1 (d2 / date-entity :year 1939 :month 9 :day 1) :op2 (d3 / date-entity :year 1945 :month 5 :day 8))
Sept. 1, 1939 - May 8, 1945
(p / percentage-entity :value 25)
(p / phone-number-entity :value "1-800-555-1212")
1 (800) 555-1212
(e / email-address-entity :value "firstname.lastname@example.org")
(u / url-entity :value "www.whitehouse.gov")
For additional special frames and entities
correlate-91 ("the more the merrier"),
have-degree-of-resemblance-91 ("X is more like Y than Z"),
variable), please consult the
- AMR Dictionary: https://amr.isi.edu/doc/amr-dict.html and/or
- List of frame arguments in PropBank (AMR version): https://amr.isi.edu/doc/propbank-amr-frames-arg-descr.txt
For more information about AMR and for more examples, see
- AMR Dictionary: https://amr.isi.edu/doc/amr-dict.html
- AMR page at ISI: https://amr.isi.edu
- List of AMR NE types: https://amr.isi.edu/doc/ne-types.html
- List of AMR roles: https://amr.isi.edu/doc/roles.html
- List of AMR quantity types: https://amr.isi.edu/doc/quantity-types.html
- List of frame arguments in PropBank (AMR version): https://amr.isi.edu/doc/propbank-amr-frames-arg-descr.txt
Maintained by Ulf Hermjakob (USC/ISI).
AMR Freak Show
This section is optional reading. Just some mathematical curiosities of AMR that one bumps into eventually, of interest to mathematicians and children.
The following AMR might look like it has a cycle, with
w being a sub-node of
but it is not cyclic after all, because one of the roles on its path is an inverse
:ARG0-of), which corresponds to an arrow in the opposite direction:
(w / woman :ARG0-of (l / lose-02 :ARG1 (s / scarf :poss w)))
The woman who lost her scarf.
Approximately 0.3% of AMRs are legitimately cyclic as for example:
(p / procedure :purpose (e / ensure-01 :ARG0 p :ARG1 (q / quality)))
a procedure to ensure quality
AMRs with such cycles typically involve
These cycles magically disappear upon reification:
(p / procedure :ARG1-of (h / have-purpose-91 :ARG2 (e / ensure-01 :ARG0 p :ARG1 (q / quality))))
Different textual ways to encode a graph
Finally, we have two different ways of encoding the same propositional content (“the boy likes to be believed”):
(l / like-01 (l / like-01 :ARG0 (b / boy) :ARG0 (b / boy :ARG1 (b2 / believe-01 :ARG1-of (b2 / believe-01)) :ARG1 b)) :ARG1 b2)
Sensible people will prefer the version on the left, though both versions relate the same conjunction of propositional triples.
# Banarescu, Laura, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. "Abstract Meaning Representation for Sembanking." Proceedings of the 7th Linguistic Annotation Workshop (2013).
# Bonial, Claire, Bianca Badarau, Kira Griffitt, Ulf Hermjakob, Kevin Knight, Tim O'Gorman, Martha Palmer and Nathan Schneider. "Abstract Meaning Representation of Constructions: The More We Include, the Better the Representation." 11th Language Resources and Evaluation Conference (2018).
# Carpenter, Bob. "The Logic of Typed Feature Structures". Number 32 in Cambridge Tracts in Theoretical Computer Science. (1992).
# Davidson, Donald. "The logical form of action sentences." (1967).
# Elhadad, Michael. "Using argumentation in text generation." Journal of Pragmatics 24.1 (1995): 189-220.
# Higginbotham, James. "On semantics." Linguistic inquiry 16.4 (1985): 547-593.
# Kay, Martin. "Functional grammar." Annual Meeting of the Berkeley Linguistics Society. Vol. 5. (1979).
# Knight, Kevin. "Unification: A multidisciplinary survey." ACM Computing Surveys (CSUR) 21.1 (1989): 93-124.
# Knight, Kevin, and Vasileios Hatzivassiloglou. "Two-level, many-paths generation." Proceedings of the 33rd annual meeting on Association for Computational Linguistics (1995).
# Mann, William C. "An Overview of the Penman Text Generation System". No. ISI/RR-83-114. (1983).
# Moore, Robert C. "Unification-based semantic interpretation." Proceedings of the 27th annual meeting on Association for Computational Linguistics (1989).
# Shieber, Stuart M. "An Introduction to Unification-Based Approaches to Grammar." (1986).