-
Notifications
You must be signed in to change notification settings - Fork 2
Guidelines for AMR
- Light verbs
In AMR, light verbs do not exist as predicate (roots) of structures. Dobj node is treated as a predicate after transformation to verb form.
Example:
The girl made adjustments to the machine.
(a / adjust-01
:ARG0 (b / girl)
:ARG1 (m / machine))
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Likely
Comments: How we can transform dobj to verb? Is it possible to get in ClearNLP?
- Function words
They don't appear in AMR. For example:
The killing happened yesterday.
would produce
(k / kill-01
:time (y / yesterday))
Is it necessary in our structure? No/Maybe
Is it possible to add this feature to our structure? No
Comments: Do we have a list of all function words? It is much longer than list of light verbs. Is there any advantage of not having functional words? (besides less space required to store the data) The questions asked can be:
- What did happen yesterday? (or What happened yesterday?)
- When did the killing happen?
- Did killing happen yesterday?
- Conjunctions
Conjunctions join two separate instances, for instance
both the boy and the girl
(a / and
:op1 (b / boy)
:op2 (g / girl))
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Yes
Comments: I see a potential of having indexes of conjunction and conjunction item (e.g.: "both" or "and", but could be unified) containing specific indexes as attributes.
- Negation
Polarity of the expressions, for instance:
The boy must not go.
(p / permit-01
:ARG1 (g / go-02
:ARG0 (b / boy))
:polarity -)
Currently, our structure will produce not as AM-NEG connected to go instance, which is the same. The only difference is with examples:
Every person failed to eat.
which produces:
(e / eat-01
:polarity -
:ARG0 (p / person
:mod (e / every)))
or
the dress is inappropriate
which produces:
(a / appropriate
:polarity -
:domain (d / dress))
In these two examples, having a polarity attribute might be considered as helpful, however, currently I don't see an easy implementation for this feature.
Is it necessary in our structure? No/Unlikely
Is it possible to add this feature to our structure? Yes
Comments: We have already similar guideline for most presented negations in AMR guidelines.
- Focus
AMR distinguishes these two sentected as different:
The marble is white.
the white marble
Is it necessary in our structure? No
Is it possible to add this feature to our structure? Yes
Comments: We probably would like to keep our structure unified in these forms, that (in this example), white always will be an attribute of an argument marble.
- Modality
The concept concern using modals like possible
, likely
, obligate-01
, permit-01
, recommend-01
, prefer-01
As an example:
The boy can go.
It is possible that the boy goes.
Produces:
(p / possible
:domain (g / go-02
:ARG0 (b / boy)))
Is it necessary in our structure? No
Is it possible to add this feature to our structure? Maybe
Comments: I don't see any point to having such structure in ours. That could be discussed later to support questions that would concern different structures in terms of words (synonyms).
- Wh-questions
In AMR they use specific amr-nodes to mark Wh-* words (e.g.: location (a / amr-unknown))
, ARG1 (a / amr-unknown))
. I think it's worth to consider if we would like to have something similar (however, the question remains in which form? attribute of the instance, so as an example, What did the girl find?
would have an instance of what, and this instance would have an attribute of unknown
Is it necessary in our structure? Maybe
Is it possible to add this feature to our structure? Yes/Maybe
- Yes/No questions
AMR uses `:mode' to indicate such questions. For example:
Did the girl find the boy?
will produce:
(f / find-01
:ARG0 (g / girl)
:ARG1 (b / boy)
:mode interrogative)
I found this useless in our case, because we will exactly know that question is coming and there is no need to keep specific node to recognize sentence as a question.
Is it necessary in our structure? No
Is it possible to add this feature to our structure? Yes
- Articles, plurals, tense, aspect, quotes, hyphens
AMR does not recognize them, but ClearNLP also provides a lemma form.
Is it necessary in our structure? No
Is it possible to add this feature to our structure? Yes
- Implicit concepts
Already covered.
- Main verb “be”
Most important is quantitative attribute, for example:
four boys making pies
(b / boy
:quant 4
:ARG0-of (m / make-01
:ARG1 (p / pie)))
This feature will be very useful in our structure.
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Yes
- Nouns that invoke predicates
AMR tries to maximize the use of OntoNotes predicates. For example, the following structure will be created
(d / destroy-01
:ARG0 (b / boy)
:ARG1 (r / room))
for all the following sentences
The boy destroyed the room.
The boy’s destruction of the room
The destruction of the room by the boy
It also works for -er nouns
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Partially
- Adjectives that invoke predicates
Adjectives also can invoke predicates.
The structure:
(a / attract-01
:ARG0 (m / man))
is produced by these sentences:
the man is attractive
the man attracts
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Maybe
- Adverbs with -ly
All adverbs with -ly are shorten to adjective form (e.g.: quickly -> quick)
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Maybe
Comments: Is it possible in ClearNLP to retrieve adjective form from adverb?
- Non-core roles
AMR introduces multiple roles such as : source
, : destination
or : topic
.
Is it necessary in our structure? No
Is it possible to add this feature to our structure? Maybe
- Focus
Inverse relations are often used for focusing:
(b / boy
:ARG0-of (s / sing-01)
:source (c / college))
for:
The singing boy from the college
There is a boy from the college who sang.
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Yes/Likely
Comments: I see this feature as useful in QA, however, I don't know how easily can be implemented. Most likely, inverse relations have to be specified and added as a rule in parsing.
NOTE: Relativizer (we should not create an instance for relativizers, e.g.: I know the girl who came to my school
, here "the girl" -> came to my school)
- Reification
Reification is converting a role into a concept.
Is it necessary in our structure? Maybe/No
Is it possible to add this feature to our structure? Maybe
Comments: AMR guidelines contains multiple rules of converting for different relations, but I don't see any advantage of having such feature.
- Prepositions
Most prepositions that signal semantic frame elements are dropped in AMR (e.g.: The nation defaulted in June.
in is dropped and June is time argument)
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Yes/Likely
Comments: Classification into different attributes has to be done, then Prepositions could be removed.
- Relative clauses
Inverse roles.
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Already is implemented
Comments: ClearNLP gives different dependency tree for The boy believes.
and the boy who believes
.
- Multiple relations with the same name
Having several relations with the same name.
(b / boy
:ARG0-of (w / want-01
:ARG1 (b2 / believe-01
:ARG1 (g / girl)))
:ARG0-of b2)
for:
the boy who wants to believe the girl
Is it necessary in our structure? Yes/Maybe
Is it possible to add this feature to our structure? Partially implemented/Maybe
Comments: We already have similar structure (a node can have multiple relations), but as in the example, I think it is useful to have multiple relations to boy.
- Quantifiers and scope
Is it necessary in our structure? Maybe
Is it possible to add this feature to our structure? Maybe
Comments: Quantifiers should be attribute types.
- Degree
Comparatives and superlatives are represented by :degree
and :compared-to
.
Is it necessary in our structure? Yes/Maybe
Is it possible to add this feature to our structure? Don't know
Comments: Is it necessary in our structure? More importantly, is it possible to get a proper degree of a word?
- Variables and co-reference
All him
, her
have to be replaced with he
, she
etc.
Is it necessary in our structure? Yes/Maybe
Is it possible to add this feature to our structure? Don't know
Comments: This might be helpful in writing our co-reference tool.
- Possession
poss
attribute.
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Yes
- Pertainyms
Is it necessary in our structure? No
Is it possible to add this feature to our structure? Maybe
Comments: Needs more research if this is necessary.
- Ordinals
ordinal forms (e.g.: 2 -> second).
Is it necessary in our structure? Maybe
Is it possible to add this feature to our structure? Yes
Comments: Needs more research if this is necessary.
- Subsets
This will help with questions such as: How many...
.
(d / die-01
:ARG1 (s / soldier
:quant 9
:subset-of (s3 / soldier
:quant 20)))
is produced by:
Nine of the twenty soldiers died.
Is it necessary in our structure? Yes
Is it possible to add this feature to our structure? Yes
Comments: Needs more research if this is necessary.
- Named Entities
Set of specific rules that set names of things.
Is it necessary in our structure? No/Maybe
Is it possible to add this feature to our structure? Maybe
- Special Frames for Roles
Set of specific rules for roles such as president, CEO or chairman.
Is it necessary in our structure? No
Is it possible to add this feature to our structure? Maybe
- Exact numbers
All number are normalized.
Is it necessary in our structure? *Yes
Is it possible to add this feature to our structure? Yes
- Several numbers
e.g.: several hundred boys
or more than 4000 boys
.
Is it necessary in our structure? *Yes
Is it possible to add this feature to our structure? Yes
- Quantities
Very useful especially in questions regarding dates/numbers etc.
30 years ago
(b / before
:op1 (n / now)
:quant (t / temporal-quantity
:unit (y / year)
:quant 30))
Is it necessary in our structure? *Yes
Is it possible to add this feature to our structure? Yes/Likely
- Mathematical operators
Is it necessary in our structure? *Yes
Is it possible to add this feature to our structure? Yes/Likely
Comments: It probably does not have to prioritized, but it would be useful in our QA system.
- Other entities: dates, times, percentages, phone, email, URLs
Is it necessary in our structure? *Yes
Is it possible to add this feature to our structure? Yes/Likely
Comments: Probably not all of them are necessary, but having major part in our qa system would be useful.
Number of necessaries in our structure: ~25/34
Number of YESs in possible to add to our structure: ~17/34
Number of MAYBEs/Likely in possible to add to our structure: ~11/34
Index
-
Home
- [Entity, Instances, and Attributes](Entity, Instances, and Attributes)
- Numbered arguments
- Modifiers
- [Semantic Role Labels](Semantic Role Labels)
- [Guidelines For AMR](Guidelines For AMR)