Introduction

Superframes is a scheme for frame-semantic annotation of text. Every predicate is assigned a frame, and every argument is assigned a role. Superframes annotations can help search corpora for phenomena of interest, enable quantitative cross-lingual comparisons, and support semantic parsing. Superframes aims to be easy to annotate. In particular,

1. Superframes does not require a lexicon. Frames are coarse and small enough in number to learn them by heart.
2. Superframes is language-independent.
3. Superframes ia annotated atop Universal Dependencies, thus delegating many difficult syntax-related decisions to an established framework.

Superframes defines a taxonomy of frames, each of which denotes a relation between two entities and thus defines two roles. Annotating a text comes down to assigning frame labels to content words and role labels to UD dependency edges between them.

Annotating Static Verbs

Consider the following sentence (relevant UD edges shown):

Here, we say that the (verbal) predicate owns evokes the POSSESSION frame, which defines the possessum and possessor roles. The possessum roles is filled by a house and the possessor roles is filled by Kim. So we annotate as follows:

We call the arguments that fill the predicate’s frame’s roles the core arguments. We will turn to non-core arguments later.

Note that semantic roles abstract away from syntactic alternation. As an example, consider the following passive sentences and its Superframes annotation:

Note also that there are only a few dozen superframes, so the meaning they specify is necessarily coarse. Different predicates with similar meaning can invoke the same superframe. This is a feature, as it makes explicit similar predicates and roles.

Verbs with One Argument

There are, of course, also predicates with only one argument, in which case only one of the two roles is used (most often the first one).

INTERNAL-STATE is one of those superframes whose second role state is rarely filled by an argument. It denotes an abstract entity (here: the state of being asleep) that is already expressed by the predicate.

Verbs with Zero Arguments

It also happens that predicates occur with no core arguments. One example is weather verbs such as rain. For consistency, there is a rule that says predicates with no core arguments must always be annotated with the most general superframe ENTITY. So that is what happens for example for weather verbs such as rain (which has a syntactic, but semantically empty argument):

We will turn to what NONCOMP means later.

Annotating Dynamic Verbs

Many predicates, especially verbs, do not simply denote a state. Rather, they denote a process working towards the coming about of a state, or a process working against the coming about of a state, or the continuation of a state, or a process working towards the end of a state, or a habitual repeated coming about of a state, or the transitioning through a state, or even a process that transitions through a multitude of states.

Using Aspectual Frames

Superframes inidicates this kind of dynamism through a number of special frames that we call aspectual frames: INIT, PREVENTION, CONTINUATION, DEINIT, HABIT, TRANSIT, DYN. They are combined with ordinary frames through a mechanism called frame stacking. For example:

Here, the predicate got has been annotated with the POSSESSION frame, with Kim as possessor and the house as possessum, as we have seen before. What is new is that the INIT frame has been stacked onto the POSSESSION frame, indicating that the verb denotes not the state of possession, but a process working towards it.

Note

On a technical level, frame stacking indicates that the first frame – here: POSSESSION – fills the first role of the second frame – here: INIT.

Further examples of INIT include:

The DEINIT frame indicates a process working towards the end of a state:

The TRANSIT frame indicates a process that transitions through a state where the state does not necessarily exist at the beginning or the end:

The CONTINUATION frame indicates that a state persists:

The PREVENTION frame indicates that a state fails to come about:

The HABIT frame indicates that a state comes about repeatedly, or exists normally:

Finally, the DYN frame indicates a process that is not straightforwardly described in terms of one or more specific states but instead is a complex process transitioning through a multitude of states. Because these states can be very heterogeneous, we usually stack DYN on top of the very general SITUATION frame.

Stacking Aspectual Frames Onto Roles

Some predicates take multiple arguments wrt. different phases of the process. For these cases, there is a special notation where you can stack an aspectual frame atop a role. For example:

Using Aspectual Frames by Themselves

Some verbs take another predicate (more precisely: a phrase headed by another predicate) as an argument and contribute an aspectual meaning. We call such predicates matrix predicates, taking an embedded predicate. In such cases, use aspectual frames by themselves, without stacking.

Using Aspectual Frames with the META Frame

Some matrix verbs take over the syntactic subject of their embedded predicate in what is known as a control construction. In such cases, we stack an aspectual frame over the special META frame. Its scene role is filled by the embedded predicate, and its participant role is filled by the controlled subject. It is then required to add a semantic dependency from the embedded predicate to the controlled subject, even though no such syntactic dependency exists in (basic) UD.

Annotating Non-core Arguments

So far, we have only seen predicates with at most two arguments, all of whose roles are defined by the the superframe that the predicate is annotated with (the core frame). But many verbs have more than two arguments, or otherwise arguments with roles from different frames. One very common example is causative subjects, denoting an entity presented as causing the situation denoted by the core frame. For example, consider the following:

In both sentences, the core frame is INTERNAL-STATE-INIT, with the vase as the has-state argument. In the second sentence, Kim is added as another argument, presented as the causer of the situation denoted by the core frame (the breaking of the vase). We thus annotate it with the causer role from the CAUSATION frame. To distinguish them from core arguments, non-core arguments are written in all upper case.

Note

On a technical level, the non-core CAUSATION frame is interpreted as being added “around” the core INTERNAL-STATE-INIT frame so that the latter fills the former other role (the effect role). This makes sense: the vase breaking is the effect here.

Similar examples include:

Two other very common non-core arguments are the sender and recipients (experiencers) of messages.

Other non-core arguments are usually rather predicate-specific.

Annotating Nouns and Pronouns

So far, we have only looked at verbal predicates. But nouns can have (core) arguments too. Specifically, this is true for event nouns and relational nouns.

Event nouns are annotated much like verbs:

Relational nouns are not much different, except that they additionally refer to one of the participants in the relation. But we are only concerned with evocation, not with reference.

Many nouns, of course, are neither event nouns nor relational nouns, and thus do not have any (core) arguments. The includes proper nouns and is also true for most pronouns. Thus, the rule that we mentioned earlier applies: an occurrence of a predicate without any core arguments is always annotated with the special superframe ENTITY.

Occasionally, pronouns do have arguments. In that case, frame them like you would the antecedent noun:

Annotating Modifiers

Nominal modifiers of nouns (UD relation nmod), oblique dependents of verbs (obl) that are not annotated as core arguments, as well as adverbial clauses (advcl) are annotated much like non-core arguments – with a capitalized role label that is short for the corresponding frame, with the modified element implictly filling the other role of that frame. For example:

Annotating Adjectives, Adverbs, and Relative Clauses

Predicative adjectives are annotated much like verbs:

Attributive adjectives and verbs behave syntactically like modifiers, but they are also predicates assigning the modifiee a role. For the modification relation, we use the META frame, assigning the modifying predicate the scene role and – implicitly – the modifiee the participant role. We then introduce an additional semantic dependency (we call it a backlink) from the modifying predicate to the modifiee to annotate the role that the former assigns to the latter. For example:

Relative clauses work very similarly: they modify a noun but also act as predicates assigning that noun a role. Again, we use SCENE for the acl:relcl relation, and a backlink to the modified noun. This is in addition to the dependency to any relative pronoun, which has the same role.

Annotating Nonlocal Dependencies

Nonlocal dependencies are dependencies from one clause into another. They do not exist as syntactic dependencies in (basic) UD, so for Superframes annotation, we add them when needed. We have already seen nonlocal dependencies in the previous chapter with attributive adjectives, adverbs, and relative clauses: these elements are attached to their head using the META frame (manifesting as a SCENE modifying relation), and then a nonlocal dependency (a “backlink”) is added from the modifying element back to its head.

There are a number of other constructions that systematically introduce nonlocal dependencies, which we also annotate. We give examples of each in the following.

Subject control:

Object control:

Non-obligatory control:

Non-obligatory control with parasitic gap:

Raising:

Secondary predicate (see also the section on secondary predicates below):

Tough construction:

Relative clause:

Reduced relative clause:

Non-finite reduced relative clause:

Relative clause with complex extracted element:

Relative clause with extraction across clause boundaries:

Coordination:

Secondary Predicates

Secondary predicates are modifiers that syntactically attach to a (primary) predicate, but semantically predicate over one of the primary predicate’s arguments, or even something more deeply embedded. The semantic relation between the primary and the secondary predicate can be one of simple accompaniment (these secondary predicates are also called depictives), effect (these secondary predicates are also called resultatives), or something else.

Note that in the resultative example You’re talking me silly, the primary predicate talking does not assign a role to its syntactic object me, only the secondary predicate silly does. But the annotation scheme still forces us to label the dependency from talking to me. We solve this by labeling it EFFECT-INIT too, just like the dependency to the secondary predicate.

Principles for Annotation

Work in progress

• MESSAGE vs. META-REALITY (Kim seems happy)
• prefer core over non-core roles if a dependent fills both (Kim bought a house, Kim wants to swim)
• frames must be associated with roots: Kim plays tennis (META, not META-HABIT) vs. Kim used to play tennis (META-HABIT-TIME)?
• terminology: (syntactic) dependents are arguments if assigned a role by the head
• terminology: situations, scenes, eventualities, states, events

ENTITY

This is the top of the superframe hierarchy, i.e., the most general frame. It is also the only frame that defines 0 rather than 2 roles. It is used for all those and only those occurrences of content words that do not have any core arguments, including all pronouns and all proper names.

Note that we go strictly by occurrence: the same word in another context may well have core arguments and thus be framed differently.

SITUATION

Roles: situee, situator

This superframe is the ancestor of all other superframes except ENTTIY. It denotes a situation in the most general sense: some entity (the situee) is in a situation, whatever the nature of that situation may be. If there is another core argument, it is labeled situator and the situation is then some sort of relation between two entities. There are four usual ways to use this frame:

Stacked with -DYN

Used for predicates that describe processes that cannot easily be framed in terms of a state but that involve complex transitions through multiple hetereogeneous states:

Stacked with an Aspectual or Modal Frame

This often occurs with predicates that focus on aspect or made, leaving the precise nature of the relation between the two arguments underspecified.

As a Default Frame

Use SITUATION if an entity is described as being in a state, but it is not internal and there is not other frame that seems to fit better.

ABSTRACT-LOCATION

Roles: has-abstract-location, abstract-location

LOCATION

Roles: has-location, location

CONTACT

Roles: ornament, surface

CONTAINMENT

Roles: contained, container

ORIENTATION

Roles: oriented, landmark

WRAPPING-WEARING

Roles: wrapper, wearer

ABSTRACT-POSSESSION

Roles: abstract-possessum, abstract-possessor

POSSESSION

Roles: possessum, possessor

ACCOMPANIMENT

Roles: accompanied, accompanier

accompanier accompanies accompanied, meaning that it occurs together with it or participates equally in the same event.

Often, the accompanier denotes not the accompanying event but an entity participating in it, and must be metonymically understood as the event.

COMPARISON

Roles: compared, reference

compared is characterized with respect to reference.

reference need not be an entity similar to the compared, it can also be an abstract constraint:

CONCESSION

Roles: asserted, conceded

Special case of COMPARISON, where compared is what’s asserted and reference is what’s conceded.

IDENTITY

Roles: same, same-as

same is the same entity as same-as.

FEATURE

Roles: feature, has-feature

MATERIAL

Roles: material, object

PART-WHOLE

Roles: part, whole

PHASE

Roles: phase, whole-scene

SUBCLASS

Roles: subclass, superclass

SUBSET

Roles: subset, superset

IDENTIFIER

Roles: identified, identifier

INFLUENCE

Roles: influenced, influence

CAUSATION

Roles: effect, causer

CREATION

Roles: creation, creator

MEANS

Roles: end, means

ASSET

Roles: has-asset, asset

MEDIUM

Roles: transmission, medium

SENDING

Roles: sent, sender

sender originates a message, sent, that can be perceived.

PERFORMANCE

Roles: work, performer

performer performs a work (of art).

CONDITION

Roles: has-condition, condition

EXCEPTION

Roles: has-exception, exception

LIMITATION

Roles: is-limited, limitation

PROPORTION

Roles: numerator, denominator

COPY

Roles: copy, original

EXPERIENCE

Roles: experiencer, experienced

REACTION

Roles: reaction, trigger

REASON

Roles: has-reason, reason

INSTANCE

Roles: instance, class

INTERNAL-STATE

Roles: has-state, state

POSE

Roles: has-pose, pose

PROPERTY

Roles: has-property, property

DYN

Roles: has-dyn, dyn

CONTINUATION

Roles: continued, continuation

DEINIT

Roles: deinitiated, deinitiation

HABIT

Roles: habitual, habit

INIT

Roles: initiated, initiation

PREVENTION

Roles: prevented, prevention

TRANSIT

Roles: transitory, transition

REALITY

Roles: has-reality, reality

Used for predicates that say something about how true or certain the expression filling the has-reality role is.

EXISTENCE

Roles: existent, existence

NECESSITY

Roles: necessary, necessity

NEG

Roles: negated, negation

This frame denotes negation. It can be used by itself or stacked onto META or other frames.

POSSIBILITY

Roles: possible, possibility

LITERALITY

Roles: has-literality, literality

PRECISION

Roles: has-precision, precision

QUANTITY

Roles: has-quantity, quantity

RANK

Roles: has-rank, rank

TIME

Roles: has-time, time

AGE

Roles: has-age, age

FREQUENCY

Roles: has-frequency, frequency

META

Roles: participant, scene

A “meta” frame for control predicates where the participant dependent is assigned a role by the scene dependent. This role must be annotated through an additional dependency from scene to participant or one of its dependents (a “backlink”).

It is possible for scene to assign two different dependents of the META predicate two different roles. The META predicate should assign both of them the participant role:

On the other hand, META predicates may have arguments that are not members of the embedded scene, such as CAUSER:

The non-core role SCENE is used for adjectives, adverbs, and relative clauses, which also assign their syntactic head a role, annotated through a backlink:

MESSAGE

Roles: topic, message

A message about topic with content message exists in perceived, measured, or recorded form. When a message is created through expression or observation, use MESSAGE-INIT. When message and topic are both realized, content must assign a role to topic. Combine with SENDER and EXPERIENCER.

Use MESSAGE for predicates of expression if there is a topic and/or message argument (otherwise, use SENDING):

Gesture is a kind of expression, too:

What is depicted or talked/thought/etc. about gets the topic role:

The result of recording something gets the message role:

Use MESSAGE for predicates of perception if there is a topic and/or message argument (otherwise, use EXPERIENCE):

Predicates that denote the initiation of perception (e.g., by acquiring knowledge, or observation, or reasoning, use MESSAGE-INIT):

Predicates that denote the deinititation of perception use MESSAGE-DEINIT:

EVALUATION

Roles: evaluee, evaluation

Special case of MESSAGE that involves a value judgment or emotion about evaluee.

NORM

Roles: norm-topic, norm

A special case of MESSAGE that does not make a descriptive claim, but is about how things should be. This includes wishes, orders, goals, etc.

Kim_SENDER *told#NORM-INIT* Sandy_norm-topic__sender to **call#SENDING**_norm
sales_norm-topic *target#NORM*
Kim_norm-subject__has-location *wants#NORM* to **swim#LOCATION-DYN**_norm
Kim_EXPERIENCER *wants#NORM* Sandy_norm-subject__has-location to **swim#LOCATION-DYN**_norm

RECORD

Roles: recorded, record

NONCOMP

Roles: has-noncomp, noncomp

This is a pseudo-frame that is used to mark syntactic dependents that are part of a multiword predicate – as in verbal idioms, weather verbs, existential there, names, other fixed expressions, or specialized sublanguages for such things as dates and times. (Light verbs, on the other hand, are treated with META).

Bound pronouns that are part of multiword expressions (e.g., inherently reflexive verbs) are annotated with the same role as their antecedent:

PERTINENCE

Roles: pertaining, pertained-to

SEQUENCE

Roles: follows, followed

ELABORATION

Roles: elaboration, elaborated-on

FULFILMENT

Roles: fulfilled, fulfiller

TIME-SEQUENCE

Roles: subsequent, precedent

REPETITION

Roles: repetition, original-occurrence

RESTORATION

Roles: restoration, original-state

SUBSTITUTION

Roles: replaced, substitute

SOCIAL-RELATION

Roles: socially-related, socially-related-to

OBLIGATION

Roles: obliged, obliged-to

MEMBERSHIP

Roles: member, social-group

Literal and Figurative Framing

Annotating Compounds

Annotating Exocentric Adverbs

Annotating Particle Verbs

In UD, particle verbs are connected to their particle via the compound:prt relation.

If the particle can be interpreted as an adposition with an elided complement (often the case with spatial meanings), label that relation as the elided complement would be labeled:

Also treat separated and nonseparated adpositional adverbs this way:

Otherwise, use NONCOMP: