Skip to content

Latest commit

 

History

History
354 lines (282 loc) · 19.8 KB

modal.md

File metadata and controls

354 lines (282 loc) · 19.8 KB
Raw

Modal Dependency

In UMR, the modal annotation captures the modal strength of events, but not the modal type of the event (i.e., epistemic/evidential, deontic, permissive, etc.).

Modal annotation is done at both the sentence level and the document level representation.

Question 1: The modal structure within the document-level annotation specifies, for each event, its conceiver, i.e. "a source, an entity whose perspective on an event is modelled in the text)". The :modal-strength values, if I understand it correctly, simply copy the sentence level annotation? So is it necessary there?

In fact, in several examples in the section on reported events (Part 4-3-1-3), the document level modal structures do not mirror the sentence level :modal-strength values. However, these differences are not discussed and in fact, I do not understand the reasoning underlying the document level annotation. Is it by mistake or I miss something?

Short HowTo for Czech annotations:

Sentence level representation:

  • for "almost all" event concepts add the :modal-strength attribute and its appropriate value (full/partial/neutral-affirmative/negative);
    (the only exception: events under the scope of modals, see below);
  • reported events (= events that are under the scope of a reporting event) as well as reporting verbs get the "default" :modal-strength attribute);
    • in addition, for reported events, add the :quote relation pointing out to the reporting verb (more precisely, enter the ID of the reporting verb as its child node);
  • while modal verbs (as want, think, forbid, ... ) get the "default" :modal-strength attribute,
    • events under the scope of modals (= modalized events identified as their own events) use the :modal-predicate relation identifying the relevant modal verb (using their ID as the child node);
    • the :modal-predicate is used instead of :modal-strength (these verbs form the only exception);

Notes on modality:

  • Deontic modality (can, must, may, …, probably, …) is conceived as just 1 event, with just :modal-strength annotation (full/partial/neutral-affirnaive/negative) at the sentence level.
  • UMR seems to use very broad interpretation for modal predicates (see below and the Guidelines, Part 4-3-1-6 and Part 4-3-2). Part 4-3-2. English modals offers a list of English means expressing modality (synt. structure, modal verbs, modal adverbs) -- this can be used as a guide for Czech annotations!!
  • The Guidelines mention that information on modal characteristics should be added to the frame files for modal verbs (as the annotation progresses).

Document level representation:

  • identify the root and the author... (root :modal author);
    ?? AUTH --> author (in our fork of the Guidelines) ??;
    (note: this is used only in some English files);
  • identify all conceivers (= sources / entities whose perspective on an event is modeled in the text), typically the author (already added, see above), but also
    • for reporting events, the "speaker/sayer/reporter" should be also identified and linked to the author;
    • for purpose clauses, the "actor/agens" of the main clause should be linked to the author (more complex, see below);
    • (for conditional clauses, no additional conceivers are supposed);
  • identify all event concepts and add each of them to the modal structure:
    • "default": as a child of the "author" node with the appropriate relation (full/partial/neutral-affirmative/negative)
    • reported event: as a child of the "speaker/sayer/reporter" node (=conceiver) with the appropriate relation (full/partial/neutral-affirmative/negative), see below;
    • modalized event: more complex, see below;
      • 1. the modal verb is linked to the "author" node indirectly via its subject entity (author -- subject; subject -- modal verb) (i.e., not directly author -- event as in the default case);
      • 2. the modalized event is represented as a child of the modal with the :Unsp relation;
    • purpose: more complex, see below;
      • the main clause "actor/agens" as the parent of the special "purpose" node with the :partial-affirmative relation,
      • which itself is the parent of the purpose clause event);
        ?? the Guidelines suggest automatic process (based on the sentence level annotation) ??;
    • condition: more complex, see below;
      • the "have-condition" node is added as a child of the author node with :neutral-affirmative relation,
      • then both events (the main clause, the conditional clause) are linked to this "have-condition" node);
        ?? the Guidelines suggest automatic process (based on the sentence level annotation) ??

Sentence-level representation

At the sentence level, three relations/attributes relate to the modal annotation

  • the :modal-strength attribute, combining
    • epistemic strength typically corresponds to the degree of confidence of a conceiver (often, the author) that the event occurs in the real world Non-future events, see below;
    • 3 basic values for epistemic strength:
      • full (complete certainty)
      • neutral (not certain but probable)
      • partial (50:50 possibility)
    • and polarity
      • 2 values for polarity: affirmative / negative
  • with modalized verbs, the :modal-predicate relation is used to identify the modal verb (cycle!!) ... see below for comments!!;
    (NO :modal-strength annotation here!!)
  • with reported events, the:quote relation identifies the reporting verb (cycle!!);
    (this relation is used in addition to the :modal-strength attribute, see below).

In addition, events in purpose clauses (:purpose) and events in conditional constructions (:condition, its reification have-condition-91) are used in the document-level representation.

Non-future events

The Guidelines: "For non-future (non-deontic) events, the :modal-strength values correspond to the author’s level of certainty towards the occurrence of the event in the real world."

For examples, see Part 4-3-1-1-1.

Evidential justification

There might be direct and indirect justification - both of them correspond to the strength of epistemic support.

Example of the direct support:

  • [en] (I saw) Mary feed the cat. ... :full-affirmative (the author has direct knowledge by witnessing the event)

Example of the in direct support:

  • [en] Mary must have fed the cat. ... :partial-affirmative (the author is inferring that the feeding event occurred (without direct, perceptual knowledge))

(Examples from Part 4-3-1-1-2.)

Future events and deontic modality

The Guidelines: "For events presented as (potentially) happening in the future, :modal-strength refers to the predictability of the occurrence of the event in the future, as presented by the author."

  • predictive future has :full strength (as in I will go to Santa Fe),
  • intentions and commands have :partial strength (as in I must go to Santa Fe),
  • desire and permission have :neutral strength (as in I can go to Santa Fe)

(Examples from Part 4-3-1-1-3.)

Reported events

Reporting predicates are annotated with a :modal-strength value (corresponding to the author’s certainty that the reporting event happened).
Reported events get -- on addition to the :modal-strength attribute (corresponding to the certainty with which the speaker/sayer/reporter reports the events) -- also the :quote relations in the sentence level annotation. (The reporting event).

Modal predicates

While a modal verb gets the usual :modal-strength attribute in the sentence level annotation, with a modalized verb (i.e., verb under the scope of the modal verb), the :modal-predicate relation identifying the relevant modal verb is used instead.

  • [en] a. Mary wants to visit France.
Mary wants to visit France.
(w/ want-01
	:ARG0 (p/ person
		:name (n/ name :op1 "Mary"))
	:ARG1 (v/ visit-01
		:ARG0 p
		:ARG1 (c/ country
			:wiki "France"
			:name (n/ name :op1 "France"))
		:aspect performance
	  **:modal-predicate w**)
	:aspect state
  **:modal-strength full-affirmative**)
  • [en] d. His parents forbid him from smoking.
(f/ forbid-01
	:ARG0 (p/ person
		:ARG0-of (k/ kinship)
			:ARG1 (p2/ person
				:refer-person 3rd
				:refer-number singular)
			:ARG2 (p3/ parent)
		:refer-number plural)
	:ARG1 (s/ smoke-01
		:ARG0 p2
		:aspect process
		**:modal-predicate f)**
	:ARG2 p2
	:aspect state
	**:modal-strength full-affirmative**)

Based on English examples (which are not consistent, unfortunately), UMR seems to use very extensive interpretation of modal predicates, they include more-or-less ALL complement taking verbs (incl., e.g., see in I saw him knock on the door.).

BUT for some predicates identified as modals in the Guidelines (see the Edmund Pope examples), the :modal-predicate relation is removed in the released English data (e.g., english_umr-0003.txt, snt2 charge-05, remains without any modal value at the sentence level; on the other hand, this event has two parents at the document level (author and null-conceiver)).
For other predicates, the English data combines :modal-strength and :modal-predicate annotation, which contradicts the above mentioned principles (e.g., english_umr-0001.txt, snt21: ... come ... to help look for bodies with help annotated as a modal predicate and look with both attributes)

Document-level representation

The modal representation has a form of a tree at the document level:

  • nodes: events and conceivers (i.e., a source, an entity whose perspective on an event is modeled in the text)
  • edges: modal strength and polarity values (i.e., how certain a specific conceiver is about the occurrence of the event in the real world)

Special cases

Reported events

For reported events, there are two nodes corresponding to conceivers in the document structure (contrary to all other events with just an "author" as the conceiver):

  • the author ... as the modal annotation expresses "the author’s certainty that the reporting event happened", and
  • the speaker/sayer/reporter (the subject of the reporting verbs) ... as the modal annotation expresses "the certainty with which the sayer/reporter reports the events".

The following examples are from Part 4-3-1-3.

  • [en] a. Mary said that she went to Santa Fe.
    sent-level: both verbs have :full-affirmative value
:modal ((root :modal author)             ???
        (author :full-affirmative say)
		(author :full-affirmative Mary)
		(Mary :full-affirmative go))
  • [en] b. Mary might have said that she went to Santa Fe.
    sent-level: say with :neutral-affirmative,
    go with :full-affirmative value
:modal ((root :modal author)             ???
        (author :neutral-affirmative say)
		(author :neutral-affirmative Mary)
		(Mary :full-affirmative go))
  • [en] c. Mary didn’t say that she went to Santa Fe.
    sent-level: say with :full-negative,
    go with :full-affirmative value
:modal ((root :modal author)             ???
        (author :full-negative say)
		(author :full-negative Mary)
		(Mary :full-affirmative go))
  • [en] d. Mary said that John might have gone to Santa Fe. sent-level: say with :full-affirmative,
    go with :neutral-affirmative value
:modal ((root :modal author)             ???
        (author :full-negative say)           ???:full-affirmative
		(author :neutral-affirmative Mary)    ???:full-affirmative
		(Mary :neutral-affirmative go))
  • [en] e. Mary said that John probably didn’t go to Santa Fe.
    sent-level: say with :full-affirmative,
    go with :partial-negative value
:modal ((root :modal author)             ???
        (author :full-negative say)      ???:full-affirmative    
		(author :partial-negative Mary)  ???:full-affirmative  
		(Mary :partial-negative go))

Modal-predicate relation

In the document level annotation, a modal verb (if identified as its own event (with the modal-predicate relation in the sentence structure)) is represented as conceived by its subject, i.e., it is linked to its subject (as a child) (and not directly as a child node of the author, as in the default case of "normal" events).

Then, a modalized event is represented as a child of the modal with the :Unsp relation -- this means, the relation among them is not further specified at Stage 0.
The Guidelines mention that "It will be taken up in the lexical entries of modal complement-taking predicates and space-builders as the lexicon is being built, and will then automatically replace the unspecified link between the modal event and the modalized event in the document-level structure."

The following examples are from Part 4-3-1-2.

  • [en] a. Mary wants to visit France.
:modal ((root :modal author)             ???
        (author :full-affirmative Mary)
		(Mary :full-affirmative want)
		(want :Unsp visit))
  • [en] b. Bob thinks the dog escaped through the fence.
:modal ((root :modal author)             ???
        (author :full-affirmative Bob)
		(Bob :full-affirmative think)
		(think :Unsp escape))
  • [en] c. They probably decided to leave on Monday.
:modal ((root :modal author)             ???
        (author :full-affirmative they)
		(they :full-affirmative decide)
		(decide :Unsp leave))
  • [en] d. His parents forbid him from smoking.
:modal ((root :modal author)             ???
        (author :full-affirmative parent)
		(parent :full-affirmative forbid)
		(forbid :Unsp smoke))

NOTE:
A special "NULL_CHARGER" / "null-conceiver" is used in one of the introductory examples (sentence 1 (2) in the Guidelines / in the released data english_umr-0003.txt) if the subject of the modal verb is not expressed. However, in these cases, the modal verb has two parents :-(( probably mistake in the variable name (both the "null-conceiver -- modal verb" relation, and the "author -- modal verb" relation)!!

  • [en] ... [businessman] who was convicted on spying charges ...
    ... the verb charge is considered a modal verb (modalizing the spying event) in the Guidelines (not in the released data)
(:ARG1-of (c/ convict-01
	         :ARG2 (c2/ charge-05
	                    :ARG1 b=businessman
	                    :ARG2 (s/ spy-02
	                           :ARG0 b
		                       :modal-predicate c2))   ... only in the Guidelines
	         :aspect performance
	         :modal-strength full-affirmative))
...
(s2/ sentence
    :modal((author :full-affirmative s2c)           ?? (should be sc = convict) ??
	       (author :full-affirmative null-conceiver)
	       (null-conceiver :full-affirmative s2c)   !!!
	       (s2c :Unsp s2s)))

The Guidelines: "The modal dependencies indicate that from the author's perspective the conviction event ... definitely happened,... . It introduces a NULL_CHARGER conceiver to indicate that the authority that charged Pope (which is not explicit in the text) presents the spying event as a certainty. "

NOTE:
UMR seems to use very broad interpretation for modal predicates (but the Guidelines examples and the English released data are not consistent wrt to modals and their annotation).
The Guidelines mention that the relevant information will be added to the frame files for modal verbs (e.g., for want, its frame file will indicate (in addition to the argument structure), also the :modal-strength of complement: neutral-affirmative).

The Guidelines:

  • "Some predicates impart full, positive (:full-affirmative) strength on their complements, often called factive predicates (e.g., manage to).
  • Strong epistemic modals (e.g., expect that, deduce) and strong deontic modals, including intention modals (e.g., plan to, decide to) and obligation modals (e.g., need, demand), impart :partial-affirmative strength on their complements.
  • Weak deontic modals, including desire (e.g., want) and permission (e.g., allow), impart :neutral-affirmative strength on their complements.
  • Certain modals may also lexicalize negation, such as doubt, forbid, or wish. These are annotated with the :neutral-negative, :partial-negative, and :full-negative values, respectively.

The Guidelines also list (some) English modals and their interpretation (not only verbs, but also adverbs and syntactic structures) in Part 4-3-2. English modals -- this definitely can serve as an inspiration for Czech annotations!

Purpose relations

(the :purpose relation in the sentence structure)

??the Guidelines mention that such annotation should be added automatically??

  • [en] They dropped water in order to fight the fire.
:modal ((root :modal author)             ???
        (author :full-affirmative drop)    
		(author :full-affirmative they)  
		(they :partial-affirmative purpose)
		(purpose :full-affirmative fight))
  • [en] He walked quickly in order to not arrive late.
:modal ((root :modal author)             ???
	    (author :full-affirmative walk)
		(author :full-affirmative he)
		(he :partial-affirmative purpose)
		(purpose :full-negative arrive))

Condition relations

(the :condition relation in the sentence structure)

??the Guidelines mention that such annotation should be added automatically??

Guidelines: "the event in the protasis [~ dependent clause] is annotated with a :condition relation to the event in the apodosis [~ main clause]." ... OK, the sentence level annotation

  • [en] If she’s hungry, I’ll feed her dinner.
:modal ((root :modal author)                         ???
        (author :neutral-affirmative have-condition)
		(have-condition :full-affirmative hunger-01)
		(have-condition :full-affirmative feed))
  • [en] If she’s hungry, maybe I’ll cook pasta.
:modal ((root :modal author)                         ???
        (author :neutral-affirmative have-condition)
		(have-condition :full-affirmative hunger-01)
		(have-condition :neutral-affirmative cook))
  • [en] If she isn’t hungry, we’ll just watch a movie.
:modal ((root :modal author)                         ???
        (author :neutral-affirmative have-condition)
		(have-condition :full-negative hunger-01)
		(have-condition :full-affirmative watch))     NOT cook