FrameNet II: Extended Theory and Practice

The Berkeley FrameNet project is creating an on-line lexical resource for English, based on frame semantics and supported by corpus evidence. The aim is to document the range of semantic and syntactic combinatory possibilities- valences-of each word in each of its senses, through computer-assisted annotation of example sentences and automatic tabulation and display of the annotation results. The major product of this work, the FrameNet lexical database, currently contains more than 10,000 lexical units (defined below), more than 6,000 of which are fully annotated, in nearly 800 hierarchically-related semantic frames, exemplified in more than 135,000 annotated sentences. Beginning with Release 1.3, the quality of FrameNet data is monitored by a consistency management system. The database has gone through three releases, and is now in use by hundreds of researchers, teachers, and students around the world. (See the FrameNet Users page on our web-site). Active research projects are seeking to produce comparable frame-semantic lexicons for other languages and to devise means of automatically labeling running text with semantic frame information.

A lexical unit (LU)

is a pairing of a word with a meaning. Typically, each sense of a polysemous word belongs to a different semantic frame, a script-like conceptual structure that describes a particular type of situation, object, or event along with its participants and props. For example, the Apply_heat frame describes a common situation involving a Cook, some Food, and a Heating_Instrument, and is evoked by words such as bake, blanch, boil, broil, brown, simmer, steam, etc. We call these roles frame elements (FEs)

and the frame-evoking words are LUs in the Apply_heat frame. Some frames are more abstract, such as Change_position_on_a_scale, which is evoked by LUs such as decline, decrease, gain, plummet, rise, etc., and has FEs such as Item, Attribute, Initial_value and Final_value.

In the simplest case, the frame-evoking LU is a verb and the FEs are its syntactic dependents:

[_Cook Matilde] fried [_Food the catfish] [_{Heating_instrument} in a heavy iron skillet]. [_Item Colgate's stock] rose [_Difference $3.64] [_{Final_value} to $49.94].

However, event nouns such as reduction in the Cause_change_of_scalar_position frame also evoke frames:

...the reduction [_Item of debt levels] [_{Value_2} to $665 million] [_{Value_1} from $2.6 billion]

or adjectives such as asleep in the Sleep frame:

[_Sleeper They] [_Copula were] asleep [_Duration for hours]

The lexical entry for a predicating word, derived from such annotations, identifies the frame which underlies a given meaning and specifies the ways in which FEs are realized in structures headed by the word.

Many common nouns, such as artifacts like hat or tower, typically serve as dependents rather than clearly evoking their own frames. The main purpose of annotating such items is to identify the most common predicates that govern phrases headed by them, and thus to illustrate the ways in which these common nouns function as FEs within frames evoked by the governing predicates.

We do recognize that artifact and natural kind nouns also have a minimal frame structure of their own. For example, artifacts often occur together with expressions indicating their sub-type, the material of which they are made, their manner of production, and their purpose/use; these are defined as FEs in the frames for various types of artifacts. Consider two example sentences from the Clothing frame.

(1) He took a packet of Woodbines out of the breast pocket of [his Wearer] [cotton Material] [shirt Garment] and lit one. (2) She had a [white Descriptor] [silk Material] [blouse Garment] on, and a severe grey skirt that reached halfway down her calves.

However, the frames evoked by artifact and natural kind nouns rarely dominate the clauses in which they occur, and so are seldom selected as targets of annotation.

Formally, FrameNet annotations are constellations of triples that make up the frame element realization for each annotated sentence, each consisting of a frame element (for example, Food), a grammatical function (say, Object) and a phrase type (say, NP). We think of these three types of annotation on each tagged frame element as layers and they are displayed as such in the annotation software used in the project. However, to avoid visual clutter, the grammatical function and phrase type layers are not displayed in the web-based report system. The full data, available as part of the data download (see the FNdata link on the FrameNet homepage), include these three layers (and several more not discussed here) for all of the annotated sentences, along with complete frame and FE descriptions, frame-to-frame relations, and lexical entries summarizing the valence patterns for each annotated LU.

FrameNet annotations derive from two sources. In pursuing the goal of recording the range of semantic and syntactic combinatory possibilities (valences) of each word in each of its senses, we normally concentrate on a particular target LU and extract sentences from the different texts of a corpus containing that LU. Then we annotate a selection of the extracted sentences in respect to the target LU. In another kind of work that represents a much smaller percentage of our overall annotations, we annotate running text. Full-text annotation differs from sentence annotation mostly in that the sentences are chosen for us, so to speak, by the author of the text. The annotation of running text is technically possible thanks to the annotation layering technique: FN lexicographers can one by one declare each word in a sentence a target, select a frame relative to which the new target is to be annotated, get a new set of annotation layers (frame element, grammatical function, phrase type) and appropriate frame element tags, and then annotate the relevant constituents.

1.1 Comparison with WordNet and ontologies

The FrameNet database is a lexical resource with unique characteristics that differentiate it from other resources such as commercially available dictionaries and thesauri as well as from the best-known on-line lexical resource, WordNet.

Like dictionary subentries, FrameNet lexical units come with definitions, either from the Concise Oxford Dictionary, 10th Edition (courtesy of Oxford University Press) or a definition written by a FrameNet staff member.
Unlike commercial dictionaries, we provide multiple annotated examples of each sense of a word (i.e. each lexical unit). Moreover, the set of examples (approximately 20 per LU) illustrates all of the combinatorial possibilities of the lexical unit.
The examples are attestations taken from naturalistic corpora, rather than constructed by a linguist or lexicographer. The main FrameNet corpus is the 100-million-word British National Corpus (BNC), which is both large and balanced across genres (editorials, textbooks, advertisements, novels, sermons, etc.), but, of course, lacks many specifically American expressions. We also use U.S. newswire texts provided by the Linguistic Data Consortium, and have recently acquired the newly released initial part of the American National Corpus, which we will begin using soon.
Our analysis of the English lexicon proceeds frame by frame rather than by lemma, whereas traditional dictionary-making proceeds word by word through the alphabet. Thus, while a traditional lexicographer measures progress in words completed, FrameNet measures progress in frames completed. Having one or more LUs for a given word in completed frames does not preclude the possibility that there could be other LUs for the same word in future frames.
Each lexical unit is linked to a semantic frame, and hence to the other words which evoke that frame. This makes the FrameNet database similar to a thesaurus, grouping together semantically similar words.
WordNet and all ontologies provide some sort of hierarchical relations between their nodes; likewise, FrameNet includes a network of relations between frames. Several types are defined, of which the most important are:
- Inheritance: An IS-A relation. The child frame is a subtype of the parent frame, and each FE in the parent is bound to a corresponding FE in the child. An example is the Revenge frame which inherits from the Rewards_and_punishments frame.
- Using: The child frame presupposes the parent frame as background, e.g the Speed frame "uses" (or presupposes) the Motion frame; however, not all parent FEs need to be bound to child FEs.
- Subframe: The child frame is a subevent of a complex event represented by the parent, e.g. the Criminal_process frame has subframes of Arrest, Arraignment, Trial, and Sentencing.
- Perspective on: The child frame provides a particular perspective on an un-perspectivized parent frame. A pair of examples consists of the Hiring and Get_a_job frames, which perspectivize the Employment_start frame from the Employer's and the Employee's point of view, respectively.
These frame-to-frame relations are shown in the frame reports; the FE-to-FE relations are not shown in the frame reports but they can be viewed through the FrameGrapher tool (accessible from the FrameNet homepage).
Since we do not annotate many nouns denoting artifacts and natural kinds, the FrameNet database is not readily usable as an ontology of things. In this area, we mostly defer to WordNet, which provides extensive coverage, including hierarchical relations of areas such as animals, plants, etc.

1.2 What do we mean by word?

In this discussion, we have used the word word in talking about lexical units. The reality is actually rather complex. When we say that the word bake is polysemous, we mean that the lemma bake.v (which has the word-forms bake, bakes, baked, and baking) is linked to three different frames:

Apply_heat: Michelle baked the potatoes for 45 minutes.
Cooking_creation: Michelle baked her mother a cake for her birthday.
Absorb_heat: The potatoes have to bake for more than 30 minutes.

These constitute three different LUs, with different definitions.

Multiword expressions such as given name and hyphenated words like shut-eye can also be LUs. Idiomatic phrases such as middle of nowhere and give the slip (to) are also defined as LUs in the appropriate frames (Isolated_places and Evading, respectively), and their internal structure is not analyzed.

For additional discussion also see the FAQs on the FrameNet website.

2 Frame Development

The core of the process has always been looking at corpus attestations of a group of words that we believe to have some semantic overlap, and dividing these attestations into groups. Afterwards, we combine the small groups into large enough groupings to make reasonable frames at which point we may (equivalently) call the words targets, lexical units, or frame-evoking elements. In the past, the criteria for such grouping have been informal and intuitive, but recently, the criteria have become more explicit. The smallest groups are formed as follows:

All LUs in a frame must have the same number and types of frame elements in both explicit and implicit (NI) contexts.
- Number
  If the number of essential, syntactically prominent frame elements differs from LU to LU or from sentence to sentence, this generally suggests that the frame should be split so that each resulting frame captures the difference. For instance, we have regularly split causatives from inchoatives. Consider the verbs increase and diminish in examples (1) to (4) below.
  (1) Also, the violent crime rate has increased from 455 to 563 offenses per 100,000 population, despite a decline in crime rates nationally.
  (2) The law has increased the Government Employee Bonus calculation rate from 6% to 7%.
  (3) From the summer 1998 to the summer 1999 the speed variation has decreased.
  (4) We have decreased the number of service calls.
  Examples (1) and (3) present the up or down changes in the value of a scalar attribute as events that happen by themselves, whereas examples (2) and (4) present them as caused by a Cause or an Agent. We have separated the two types of sentences into two different frames, an inchoative frame called Change_position_on_a_scale for examples (1) and (3) and a causative one called Cause_change_of_position_on_a_scale for examples (2) and (4). There may be a legitimate objection about the presence of an Agent or Cause being just a vague linguistic intuition and that we ignore the fact that everything that happens is caused. In this and in many other situations where one needs to decide between splitting and lumping causative and inchoative sentences, linguistic evidence suggests that splitting is the right way to proceed. First, there will typically be lexical units that exhibit only one of the two uses. For instance, the verb gain only has inchoative uses when referring to scalar change, while the verb lower only allows causative uses in the domain of scalar change. Second, cross-linguistic comparison also shows that other languages often distinguish inchoatives and causatives by derivational morphology. As pointed out above, we factor out language-wide constructions when figuring out how many participants there are in a particular use of a particular LU. Thus, we don't need an extra frame for the verb sell to account for uses like Those boots sell well because the sentence involves subject selection construction (often called middle) which deprofiles and makes generic one or several of the prominent actors, in this case the Seller. We then simply label Seller as CNI , deeming it still conceptually present, but generic. There is indeed linguistic evidence for the conceptual presence of the Seller: the notion of ease that is assessed requires a sentient being that could experience ease or difficulty selling. Obviously, similar reasoning also applies to passives. In addition, we have used the idea that a construction can suppress the realization of a conceptually present frame element to account for the deleted objects in instructional imperatives, e.g. Place __ in oven for twenty minutes.
- Type
  The basic semantic type for a frame element ought to be broadly constant across uses. If that is not so, it suggests the need to posit distinct frame elements. In some cases, however, we still want to recognize a relationship between frame elements whose syntactic form suggests that they refer to ontologically different kinds of entities. For example, in I want [to win] compared with I want [an orange], both complements of the verb want have something to do with the Desiring frame, but each of the complements directly refers to something rather different. While the verb phrase to win refers to a type of event, the NP an orange refers to an entity. Clearly, the NP-complements of want and other such predicates like it, are understood to metonymically stand for events centrally involving them: `to want an orange' is typically `to want to eat an orange'. There is evidence that another event is to be understood by speaker and hearer when want is used with an object NP: sentences such as I want your report on my desk tomorrow morning include a time phrase that does not refer to the time of wanting but to the time when the desired event is to occur (or the desired state ought to hold).¹ When, as with want, some of the target's complement types are appropriate to the ontological category of the frame element while others that are metonymically related are not, we normally create two distinct frame elements and relate them to each other via the Excludes relation (for which, see section 3.2.2.3).² Splitting in this manner is useful in several ways. It allows the right frame element to frame element relations because in some cases, related frames will allow only one of the (metonymically or otherwise related) frame elements. Having distinct frame elements also facilitates finding data to study Pustejovskian coercions. And finally, it makes the annotation task easier if the clear ontological distinction is maintained as a frame element distinction.
In aspectually complex frames, the lexical units should all entail the same set of stages and transitions. For instance, while work on and develop both evoke the background of a sentient Agent expending effort to accomplish some Goal, only the latter entails (in past tense utterances) that the Goal was in fact accomplished. Similarly, the verb shoot can be used to report an event of firing a projectile at a person and hitting, but it does not entail that the person dies in contrast to, for instance, the verb decapitate. In all such cases, where lexical units differ in whether an intended and typical `fulfillment' stage is in fact reached, we split frames. As with the number and type of frame elements, the requirement that the same subevents are entailed for all lexical units also argues for separating causatives and inchoatives, since the former class involves an extra causing subevent that the latter class lacks.
The same frame elements will be profiled across all lexical units of a frame. That is, the same participant's point of view should be emphasized with all of them. Consider the difference between the verbs buy and sell: the first takes the Buyer's perspective and the second the Seller's perspective. Note that when multiple perspectives on a type of scenario are possible, typically we relate the frames that carry particular perspectives to a non-perspectivized background/scenario frame via the Perspective_on relation (see Sect. 6.1.2). Thus, the perspectivized frames can be related to each other as candidates for paraphrasing via the background scenario. Figure 4 shows how the Commerce_buy and Commerce_sell frames are related to each other via the frame Commerce_goods-transfer, which is itself a sub-frame of the Commercial_transaction frame.

Commercial_Transaction
We want the interrelations between frame elements to be the same for all the LUs in a frame. Consider, for instance, the fact that the Purpose frame elements of buy and sell have a relationship to different participants: a Purpose expressed with buy relates to the Buyer, a Purpose expressed with sell relates to the Seller.
(5) I bought this thing [in order to test it out].
(6) At least you are selling them [in order to give the money to cancer research].
The difference in the orientation of the Purpose frame element parallels the difference in perspective that we noted above and supports the need for a frame distinction for buy and sell.
The presuppositions, expectations, and concomitants of the targets within a frame will be shared. For example, the verb cross-examine evokes a courtroom session, a preceding event of questioning by an opposing legal party, etc., differentiating it from the simpler examine. By similar reasoning, the verbs receive and take would be differentiated, since receive presupposes another willing agent participating as the Donor and take does not.
The basic denotation of the targets in a frame should be similar. For this reason, adjectives such as blue and broken which refer to completely different kinds of states, are not in the same frame, even though they are distributionally very similar otherwise. Admittedly, this diagnostic is easy to apply in some cases and hard in others. Note that it is not reliably possible to base frame distinctions mainly on selectional restrictions. One may, for instance, conclude that two different frames are involved with blue and broken given that, in the naive physics sense, the kinds of entities that have color are much more numerous and spread out over a greater part of an ontology than the entities that can be broken (in the concrete physical sense). However, within the Self-motion frame the verbs crawl, flit, slither, and walk have fairly different selectional restrictions, though they are grouped together in the same frame.
The pre-specifications that the frame-evoking elements give to various frame elements will be similar. This heuristic was applied in the Mass_~motion frame containing lexical units such as crowd.v, flock.v, pour.v, stream.v, swarm.v, troop.v, etc. These LUs are not simply treated as part of Self_motion since they all require that the moving entity be a mass theme, generally made up of many individuals. However, the desideratum that all the LUs in a frame share selectional restrictions on a frame element is often relaxed in practice, as it would result in very fine-grained frames that are impractical to manage. Consider, for instance, the requirement of the verb tie that the Connector be a relatively long, thin, flexible object. Other verbs that refer to attaching one object to another (or, viewed symmetrically, both to each other) either make very different specifications (e.g. glue, staple) or none at all (e.g. attach).

Aside from factors that are used to group words into frames, there are those that never enter into consideration.

At all times, we lump together those groups whose semantic differences are due to general constructions of the language:
- passive
- middle
- tense/aspect constructions
- composition with extra-thematic frame elements
The list does not include causativity, since, as argued above, it is lexically specific.
We group together words that are antonyms of each other. For instance, the adjectives high and low are both in the Position_on_a_scale frame. Similarly, the verbs love and hate are both in the Experiencer_subj frame. However, so-called converses (Cruse 1986), relational opposites such as buy and sell which take opposite points of view on a single event, are placed into separate frames because they profile different (sets of) participants. Likewise, so-called reversive pairs such as tie and untie are placed in different frames since the members of the pair denote different kinds of actions basically.
Any differences purely in speech context.
- deixis (come v. go)
- register (botch v. fuck up)
- dialect (lorry v. truck; fixture v. regular season game)
- evaluation (criticize v. praise; genius v. moron)
We mark these kinds of differences with semantic types (see Sect. 6.2). For example, the adjective generous in the Stinginess frame would be marked with the semantic type Positive_judgment, while the adjective stingy would be assigned the semantic type Negative_judgment. Note, however, that so far we have not applied semantic types in all cases where this would be appropriate.

Finally, frame development focuses on paraphrasability (or near-paraphras-ability) of words and multi-words. That is to say, we ask whether one can more or less felicitously substitute one lexical unit for another and still evoke the same frame and express the same kinds of semantic roles as syntactic dependents of the new lexical unit. Frame development does not directly address the (near-)paraphrasability that may exist between whole utterances.

There are, for instance, many pairs of utterances where the total meaning of one member is built up compositionally from several frame evoking elements while the total meaning of the other derives just from a single lexical unit that combines within it the complex semantic structure that can be expressed with multiple frame evoking elements.

Clear cases of this situation include causative-inchoative pairs. Sentence (7) is centered on the inchoative harden, but combines it with the extra-thematic frame element Cause, expressed by the phrase due to the hydration of the cement, which composes to a causative scenario. Sentence (8) directly encodes the more complex causative scenario: there is no lexical material present that encodes the notion of causation separately from the notion Change_of_consistency.

(7) The paste hardened due to hydration of the cement.
(8) The hydration of the cement hardened the paste.

Note that although causative-inchoative pairs are not in the same frame, the FrameNet database provides an explicit link between the paired frames via a frame-to-frame relation Causative_of.

Similarly, sometimes the frame relation Using also connects frames whose targets can figure in utterances that are paraphrases of each other. Consider examples (9) and (10).

(9) I've communicated by email with US Airways on many occasions and have always gotten a reply within a few days.
(10) I've emailed US airways on many occasions and have always gotten a reply within a few days.

The lexical unit communicate in (9) belongs to the very general Communication frame, which is used or inherited by many frames. The lexical unit e-mail in (10) belongs to the Contacting frame, which uses the Communication frame. (For a more detailed discussion of the various frame relations, see section 6). The two lexical units and frames are not equivalent, which can be illustrated easily with the following pair of past tense utterances:

(11) I e-mailed him my new phone number.
(12) I communicated my new phone number to him by e-mail.

While (12) entails that the Recipient received the message, (11) does not carry the same entailment. With Contacting, no actual successful communicative act is implied, only the successful completion of acts which could establish the communication.

Another kind of paraphrase between utterances relies on world knowledge and inference. Consider the pair of utterances in (13)-(14). Utterance (13) denotes a means action and utterance (14) denotes what is achieved by the means action.

(13) The Denver tight end caught the ball in the end zone.
(14) Denver scored a touch down.

Catching the ball in the end zone of the field is a means of scoring a touch down, but catching and scoring have no conventional relation to each other outside of football, and even in the football context the two are not necessarily linked.

However, in some cases when inference is needed to establish that sentences are paraphrases of each other, the lexical material may still guide the inferential process to some extent. Compare sentences (15), which denotes a means action, and (16), which denotes an accomplished Goal.

(15) You needed my help so I got on the bus.
(16) You needed my help so I came.

Boarding a bus and arriving somewhere have no necessary relation to each other. However, the noun bus is a lexical unit in the Vehicle frame and that frame is linked to the Motion frame, which in turn is used by the Arriving frame that includes come. The framal links between bus and come thus provide some evidence that a semantic equivalence may be intended.

Generally, FrameNet groups words only for paraphrasability between lexical units but not for paraphrasability of utterances. There is, however, one kind of hybrid case where FrameNet groups lexical units in the same frame, even though the alternating lexical units do not participate in the same syntactic constructions and despite the difference in the relative prominence that the evoked frames have in the sentence. Consider the following pair of sentences:

(17) She wisely sold the house.
(18) She was wise to sell the house.

The understandings of (17) and (18) are clearly similar. However, there is a difference between the sentences in terms of the relative salience of the two frames, Mental_property and Commerce_sell. (17) mainly reports an event of selling whereas (18) focuses on the assessment of what the selling behavior says about the seller's mental abilities. The lie-test shows that this is so. If a speaker challenges (17) by saying That's a lie, the selling event is being contested, but if they challenge (18) in the same way, what's denied is the appropriateness of the assessment that it was wise to sell.

This pragmatic difference fails to be captured by many current logical representation formalisms of adverb meanings, which suggest that the adverb has the clause in its scope. While FrameNet's practice of grouping the adjective and adverb in the same frame is in line with the logical-semantic similarity, it does not match the pragmatic-syntactic fact that the frame evoked by the syntactic head of the clause is more prominent.

3 FrameNet Annotation

3.1 Introduction

As a technical matter, the way in which FrameNet analyzes instances of a target predicate consists of marking up parallel aligned layers of annotation with appropriate label sets, as shown in Figure 3.1. The layers that are displayed in the FN Desktop can be manually selected by annotators. The number of layers and the kind of information that can be recorded on them is technically unlimited. But in FrameNet's current practice the four core annotation layers are the Target, frame element (FE), grammatical function (GF), and phrase type (PT) layers. On the first, the (parts of the) target predicate are marked while on the latter three, labels are applied to the constituents expressing the frame elements of the target.

Figure 3.1: Annotation layers

The next-most important set of layers consists of the layers called Other; a layer called either Noun, Verb, Adj, Adv, or Prep depending on the part of speech of the target (this layer is also often called the part-of-speech-specific layer); and the Sent(ence) layer. The Other layer holds labels relating to certain special constructional contexts in which the target may occur, such as relative clauses, existential constructions, and extraposition constructions. The part-of-speech-specific layer holds labels that can occur only with predicates of a particular part of speech. For instance, the Copula does not occur with verbal targets; it can be applied on the Noun, Adj, and Prep layers of targets of the appropriate part of speech. The Sent(ence) layer is special in that it does not actually bear any annotation labels: when the layer is invoked, information about the sentence as a whole can be recorded on an appearing list of check-boxes.

A final group of layers includes, among others, layers holding labels related to part of speech (POS) and Named Entity Recognition (NER). This information is derived automatically from our corpora and third-party software and is generally not modified by FrameNet annotators.³

We now turn to FrameNet's annotation process. The work can be divided into two kinds according to the way in which sentences are chosen for annotation. In the lexicographic annotation mode, our main focus is on the goal of recording the range of semantic and syntactic combinatory possibilities (valences) of each word in each of its senses. To that end, we extract sentences from the different texts of a corpus because they contain a predetermined target LU. Then we annotate a selection of the extracted sentences in respect to that particular LU.

In another kind of work, the annotation of running text (also called full-text annotation), the sentences are chosen for us, so to speak. Annotation of running text is technically possible thanks to the annotation layering technique: FN lexicographers can one by one declare each word in a sentence a target, select a frame relative to which the new target is to be annotated, get a new set of annotation layers (frame element, grammatical function, phrase type) and appropriate frame element tags, and then annotate the relevant constituents.

Before going further into the details of annotation, let us briefly consider the Revenge frame, which will figure as an example frame repeatedly in this chapter. The definition of this frame follows:

An Avenger performs some Punishment on an Offender as a response to an earlier action, the Injury, that was inflicted on an Injured_party. The Avenger need not be identical to the Injured_party but needs to consider the prior action of the Offender a wrong. Importantly, the punishment of the Offender by the Avenger is seen as justified by individual or group opinion rather than by law.

Thus, the frame elements in the Revenge frame are Avenger, Punishment, Offender, Injury, and Injured_party.

Lexical units in this frame include avenge.v, avenger.n, get even.v, retaliate.v, retaliation.n, retribution.n, retributive.a, retributory.a, revenge.v, revenge.n, revengeful.a, revenger.n, vengeance.n, vengeful.a, and vindictive.a. Some example sentences with the lexical unit avenge are given here.

(1) [His brothers Avenger] avenged [him Injured_party].
(2) With this, [El Cid Agent] at once avenged [the death of his son Injury].
(3) [Hook Avenger] tries to avenge [himself Injured_party] [on Peter Pan Offender] [by becoming a second and better father Punishment].

Regardless of the kind of annotation that is done, the following guidelines apply to the annotation relative to a particular instance of a target word.

FrameNet annotation is directed toward the dependents of a target word. We do not annotate constituents that are understood only through context as referring to the filler of a particular frame element role. Consider the following example.
(4) Smith was surprised when Lowry retaliated for the attack.
The target retaliate evokes the Revenge frame introduced above, but the surface syntax does not contain constituents representing the Offender or the Injured_party. One plausible understanding of the sentence as a whole is that Smith was the one against whom Lowry retaliated. One might, thus, want to tag Smith as the Offender. However, there is an alternative interpretation: Smith could just be an observer of a conflict involving Lowry and a third party. In such cases, where there is no syntactic construction present that guarantees that interpretation, we do not annotate the non-local phrases in question. In the particular example here, we mark the Offender as omitted under definite null instantiation (DNI).
We annotate whole constituents rather than just the head words of the target's syntactic dependents. For instance, in example (3) above, the preposition on is included in the constituent expressing the Offender frame element, on Peter Pan.
Each dependent is annotated for Frame Element identity, Phrase Type, and Grammatical Function relative to the target LU.⁴

While the above guidelines apply equally to lexicographic and full-text annotation, there are also some clear differences between the two modes of annotation.

In FrameNet's lexicographic work, annotation is done relative to only one lexical unit per sentence, the target. (As noted previously, the target is typically a single word but can also be a multi-word expression such as a phrasal verb (e.g., give in) or an idiom (e.g., kick the bucket).) In full-text annotation, by contrast, all contentful lexical units are treated as targets and their dependents are annotated.
In full-text annotation, we have no choice but to annotate each instance of a target word whether or not it contributes useful lexicographic information about the target. For instance, in a given text the LU may occur in complicated syntactic constructions that add nothing to our understanding of the target's meaning. A particular use of the LU may be very poor style or borderline ungrammatical. Some instances only provide more exemplification of an already amply documented argument realization pattern for the LU rather than exhibiting a rarer pattern. This contrasts with our sentence-annotation (lexicographic) work, where it is intentionally part of our annotation principles to avoid such sentences.
In the lexicographic work, we intend the set of sentences annotated for a given LU to represent the full range of combinatorial possibilities for that LU. A given text, even of considerable length, is not likely to exemplify all patterns in which a target may occur.
The annotation contained in the FrameNet database does not offer information about frequency of occurrence. Full-text annotation necessarily means that we annotate all instances of any given contentful target. Thus, at least relative to a particular text, frequency information for LUs can be derived.

So far, we have sub-divided our work into two parts depending on whether we choose the sentences to annotate or whether they are chosen for us by a text. Another subdivision applies specifically to our lexicographic work where we produce annotation of two different types, reflecting two different kinds of target words:

annotation relative to a frame bearing syntactic governor, either a predicate, modifier, or a referring expression and
annotation relative to a slot filler, that is, relative to a referring expression that is a frame element of a frame determined not by itself but by a governor

Both kinds of annotation will be discussed. Since the annotation relative to frame-bearing syntactic governors is FrameNet's main task, it will be discussed first and in greater detail in sections 3.2-3.7. Annotation relative to slot-fillers is discussed in section 3.8.

3.2 Lexicographically motivated annotation practices

In accord with FrameNet's goals, syntactic and semantic descriptions are tailored to lexicographic description only and may differ from regular linguistic analysis as well as from shallow semantic analysis in several ways.

Typically, the analysis of sentences is only partial in the sense that we apply labels just to some words or phrases of a sentence. Although many sentences contain multiple frames evoking lexical units, in normal annotation we annotate only with respect to one of them. Also, there are some classes of frame evoking elements that we have not yet tackled, an important one being sentence connectives such as although, while, but, etc.
We tag whole constituents that realize frame elements relative to our target words, rather than just tagging the head words of these constituents. That is, we work with a phrase structure grammar, rather than a dependency grammar. A consequence of this is that many frame element labels cover words that have no direct relation of their own to the target, but only to the head of their constituent. For instance, when a frame element is expressed by a noun which takes adjectival, prepositional or clausal complements or which is modified by such elements, these complements and modifiers are included in the frame element tag. Consequently, phrase type (and grammatical function) labels always cover full constituents. In (5), we tag both the noun story and the whole NP it heads; in (6), we tag both the noun fact and the full NP it heads; in (7), the with-PP is part of the frame element that is tagged.
(5) I heard [a story about a man named Jed NP].
(6) [The fact that I wear a funny shirt Sfin] is totally irrelevant.
(7) I have [a cat with orange stripes NP].
FrameNet syntax also differs from traditional treatments of subcategorization in paying attention to relational modifiers, including the non-heads of compounds. For example, the FrameNet description of a noun like clinic makes reference to the types of nouns which can modify this noun in compounds like allergy clinic. In theoretical treatments, modifiers such as allergy relative to clinic are typically assumed to be outside the realm of subcategorization. FrameNet includes them because they frequently express the same semantic roles (or frame elements) with respect to the modified heads as complements express with respect to their governors. For example, the Cure frame associated with the verb treat includes a role for the Affliction which is treated, and this role is typically expressed as the object of treat. Thus, you can treat an allergy, treat the flu, and so forth. Modifiers occurring with the noun treatment often express the same role; thus, there are allergy treatments, flu treatments, etc. Such observations are of lexicographic relevance, and our syntactic descriptions reflect this perspective.
In some situations, there are differences between the syntactic and the semantic headedness of a clause. For instance, Pa gave her a lecture has give as its syntactic head. However, from a semantic point of view the sentence reports an act of lecturing, not one of giving. We understand cases such as give a lecture to involve frame-evoking nouns that are syntactically `supported' by verbs in order to be able to project clauses. This analysis is strengthened by the fact that noun-support verb combinations typically involve the selection of the verb by the frame-bearing noun (*make/give a lecture v. make/*give an attempt). An important consequence of this analysis is the annotation of subjects of support verbs as frame elements relative to the noun. Thus, the example sentence Pa gave her a lecture would be annotated with respect to the target lecture, with gave marked as a support verb, Pa marked as the Speaker and her marked as the Addressee.
Syntacticians commonly divide the complements of a predicate into arguments and adjuncts, or arguments and modifiers. In FrameNet, the semantic spirit of this distinction is covered by the distinctions of coreness status rather than through GF's (cf. section 3.2.1). As a consequence, we only have a grammatical function Dep(endent), rather than distinguishing between the two types of complements.
In general, we select sentences for annotation where, with the exception of subjects, all frame elements are realized locally by constituents that are part of the maximal phrase headed by the target word. In addition, there are several clear cases, typically involving subjects, where the combination of particular lexical items and grammatical structures guarantees a particular interpretation of phrases which are not dominated by the maximal phrase headed by the target. One such example is that of modal verbs, which behave as so-called raising verbs. The subjects of modals such as may, will etc. always need to be interpreted as subjects and frame elements of the predicates that follow the modal. For an example like [Bill SELLER] may sell it at that price, a treatment motivated by syntactic theory might suggest that the subject slot of sell is filled by a kind of trace, or that one of its required arguments is passed on to the higher raising predicate via a valence passing principle. Under such a treatment, it could be left up to grammatical principles to establish the connection between the subject of the raising verb and the valence requirement of the target verb. However, from a lexicographic point of view it is more useful to directly report what the semantic properties of the fillers of particular Frame Elements are, rather than to ask users to employ additional tools to establish lexical antecedents or valence fillers.
For most targets, frame elements are marked only on a single frame element layer. Under two kinds of circumstances, however, we record frame elements on an additional, secondary frame element layer.⁵ The first type of situation for which second-layer annotation is used is frame element conflation, when a phrase that directly expresses one frame element also allows inferences about a second frame element. Consider as an example the sentence Doctor Miller managed to at least cure [the epileptic]. Out of context, this sentence is most naturally understood to mean that the epileptic is cured of epilepsy. We tag the phrase the epileptic as Patient on the first frame element annotation layer and as Affliction on the second annotation layer. Notice that when a core frame element, such as Affliction in the example of cure, is applied on a secondary annotation layer (cf. 3.2.4), no null instantiation tag is used concurrently. Given an appropriate context, all inferences in conflation contexts can be defeated. In cases where the inference is defeated, no secondary layer tagging is applied. In the modified sentence Both the epileptic and the schizophrenic patient had terrible head colds last week and Doctor Miller managed to at least cure [the epileptic], the Affliction frame element is marked as null-instantiated rather than applied to the constituent the epileptic on a second frame element layer. Note that frame element conflation is the only context where FN annotation marks phrases that are only inferred, rather than grammatically guaranteed, to express information about frame elements. A second type of circumstance where FN annotation makes use of a secondary annotation layer has to do with possessor phrases that are obligatorily coreferential with another argument. Most commonly this arises with body part possessors such as in the Experience_bodily_harm frame. With verbs such as stub, sprain or strain the referent of the Experiencer frame element has to be expressed again as a possessive determiner in the Body_part frame element and we tag the possessive on the second frame element layer with another instance of the frame element Experiencer. Thus, in Ike cut [his finger] on the envelope, the subject NP Ike is marked as Experiencer on the first annotation layer, the phrase his finger is marked as Body_part, and the possessive his is marked as Experiencer on the second frame element layer. In all cases, a label applied on the second layer is always co-extensive with, or contained in the boundaries of a label on the first FE layer; no word or constituent can be covered by a frame element label on the second layer alone. Finally, secondary frame element layers are never accompanied by grammatical function or phrase type layers; GF and PT layers are available only for first layer annotation.
The combination of our surface-oriented annotation and our commitment to providing grammatical function and phrase type information about the phrases that fill frame element roles causes some problems in the case of discontinuous frame elements. For instance, with wh-extraction out of a frame element such as [Who] did you talk [about]?, we need to provide a grammatical function tag for the fronted wh-word who even though it does not bear a grammatical relation to the target word talk. This and other cases of discontinuous frame element realization are discussed below in section 3.2.7.

3.2.1 Coreness

We classify frame elements in terms of how central they are to a particular frame, distinguishing three levels: core, peripheral, and extra-thematic. A fourth possible value for this attribute, called core-unexpressed is also discussed below.

A core frame element is one that instantiates a conceptually necessary component of a frame, while making the frame unique and different from other frames. For example, in the Revenge frame, Avenger, Punishment, Offender, Injury, and Injured_party are all core frame elements, because an avenging event necessarily includes these participants. One cannot imagine an act of revenge that is not preceded by a (perceived) offense or one that is not directed against anybody.

In determining which frame elements are considered core, we also consider some formal properties that provide evidence for core status. These properties are typically co-present, although they need not be.

When an element always has to be overtly specified, it is core. For instance, the verb resemble in the Similarity frame always requires a postverbal complement NP denoting an entity that is similar to the entity denoted by the subject.
A frame element which, when omitted, receives a definite interpretation, is also core. For instance, when the verb arrive is used only with a Theme-subject, as in John arrived, a particular Goal location that the Theme reaches still has to be understood in the context. Goal, therefore, must be a core frame element.
A frame element whose semantics cannot be predicted from its form, in particular from any marking prepositions, ought to be core since its interpretation completely depends on the target. From this, we can derive two corollaries, given below.
- A frame element that has no formal marking should be core. Thus, frame elements that can be subject or object in a simple active sentence ought to be core since these slots host frame elements of many different kinds and knowing that something is a subject or object is not particularly informative. For example, the Building frame which contains the verb build has at least two core frame elements since the verb build has both a subject and an object.
- A frame element that has idiosyncratic formal marking should also be core. A good example of this is the prepositional verb depend on. The preposition on does not occur as a marker of the same meaning with predicates in many other frames. In its basic spatial sense of 'in contact with and supported by', on occurs in many different frames; as a marker of Place or Location frame elements it is totally unremarkable and does not suggest core status for these FEs.⁶

Frame elements that do not introduce additional, independent or distinct events from the main reported event are characterized as peripheral. Peripheral FEs mark such notions as Time, Place, Manner, Means, Degree, and the like. They do not uniquely characterize a frame, and can be instantiated in any semantically appropriate frame. In respect to the Revenge frame, any report of an event of revenge may also include explicit information about the parameters of time, place, manner, etc. of the revenge, an example of which is given below.

(8) The bereaved family retaliated [immediately Time].

Extra-thematic frame elements situate an event against a backdrop of another state of affairs, either of an actual event or state of the same type, as illustrated with Iteration, or by evoking a larger frame within which the reported state of affairs is embedded, as shown for Containing_event.⁷

(9) Thou shalt not exact revenge [twice Iteration] for the same offense.
(10) The Aussies took revenge [in a penalty shootout before 2465 fans in Long Beach the next day Containing_event].

Note that extra-thematic frame elements are understood not to conceptually belong to the frames they appear in. We take them to properly be frame elements of other abstract frames that take them as well as the targets that they modify as arguments. Thus, in example (11), we take twice and the verb phrase eat to be arguments of a more abstract Iteration frame. Similarly, in example (12), cooked dinner and me are frame elements of a Benefaction frame. Note that, as shown by (12), the native frame of the extra-thematic frame element need not be evoked by lexical material, it may simply be evoked constructionally.

(11) Learn how to spend a few extra minutes planning complementary menus where you cook once and eat [twice]. (12) Lennert, another sweetie in my life, cooked [me] dinner, mmm mmm good.

The view of extra-thematic frame elements presented here entails that these frame elements are necessarily the same across all the `host' frames in which they appear. That is, unlike core and peripheral frame elements, extra-thematic frame elements do not have a frame-specific understanding. By comparison, although many core frame elements named Agent share properties with each other due to Inheritance and Using relations, they do not necessarily have identical properties. More importantly, one cannot predict the frame role Agent from the fact that an NP has the grammatical function Ext.

(For an overview of the most frequently occurring extra-thematic frame elements, the reader is referred to Appendix 8.)

The value "Core-Unexpressed" is a special notational shorthand. It is assigned to FEs that behave like core frame elements in the frame where they are marked as Core-unexpressed but which, counter to expectation, may not be used for annotation in descendants of that frame. Frame elements marked as core-unexpressed will thus not necessarily be listed among the FEs in descendant frames.

We do not consider core-unexpressed frame elements to be violations of our definition of full inheritance. Our reasoning can be illustrated with the example of the core-unexpressed Act frame element in the Intentionally_act frame, which is exemplified in (13).

(13) I'll do [the vacuuming Act].

In the many child frames of Intentionally_act such as Choosing, Perception_active, etc., the idea of an Act is as relevant as in the Intentionally_act parent frame. However, in the child frames the frame element is absorbed by the lexical units in the frame and cannot be separately expressed.

Marking the frame element Act as core-unexpressed in the Intentionally_act parent frame allows us to keep the frames that are lower in the hierarchy from including an inherited FE which for any lexical unit in the frame could at most be annotated on the target itself, but never be expressed separately. The sentence *I chose decision the blue one is simply ungrammatical.

Coreness marking makes the most sense for event and state frames, and in these frames we use all three statuses. Coreness marking is done at the level of the frame and is intended to be consistent for all lexical units in a frame. In frames whose LUs are artifacts or natural kinds, we only use a two-way distinction (core and peripheral), noting that the values do not have exactly the same meaning as with events. In such cases, typically there is just one core frame element which is marked on the target word. For instance, in the Clothing frame the FE Clothing is core, and all other FEs are peripheral.

3.2.2 Frame element relations

In providing a semantic analysis of the combinatorial possibilities of our target predicates-rather than one stated only in terms of phrase types and grammatical relations-we have observed that in an important sense frame elements are not independent of each other. Frame elements are related to the frame and required by it, as well as interrelated directly in a number of ways. These interrelationships have a direct impact on annotation, as they license the absence of core frame elements, which must normally all be accounted for in every annotation set, or (more rarely) require frame elements that might otherwise be optional. FrameNet systematically records these interrelationships. The three types of frame element relations that we recognize are discussed in sections 3.2.2.1-3.2.2.3. Section 3.2.2.4 discusses some respects in which the current treatment of frame element sets needs further refinement.

3.2.2.1 Coreness Sets

In our annotation practice, we often find that some groups of FEs seem to act like sets, in that the presence of any member of the set is sufficient to satisfy a semantic valence of the predicator. We refer to such a group of FEs as a coreness set, or CoreSet. For instance, Source, Path, and Goal are core FEs in the various motion frames in the database. However, although possible, it is not necessary, and in fact unusual, for all three FEs to co-occur, as in example (14). Sentences in many motion frames can be informationally complete and pragmatically felicitous with just one or two of the FEs expressed, as shown in (15)-(18).

(14) Fred went [from Berkeley Source] [across North America and the Atlantic Ocean Path] [to Paris Goal].
(15) Martha hiked [from Berkeley Source] [to Oakland Goal].
(16) Elaine walked [to Monterey Goal].
(17) I saw Peter sneak [past the guard Path].
(18) Juan was walking [out of the office Source] when I arrived.

FrameNet's normal annotation practice demands that we account for all core FEs and we could keep track of the un-instantiated FEs in the example sentences above by using null-instantiation labels as described in section (3.2.3). However, we prefer to group the FEs in CoreSets and not mark null instantiation for each member FE in cases where the FEs have an informational and conceptual interdependence. Source, Path, and Goal, for instance, are clearly related via a notion that we might call 'full path'. By contrast, omission of the Ingestibles for eat in the Ingestion frame is not related to the presence or absence of any other frame element. The sentence I'm eating __ (with my friends/in the kitchen/now) is acceptable with or without any of the frame elements given in parentheses expressed.

3.2.2.2 Requires

In some cases, the occurrence of one core FE requires that another core FE occur as well. To illustrate, in the Attaching frame Item, Goal, and Items all are core FEs. If Item occurs, then Goal is required, as shown below, where the sentence without a Goal is unacceptable. In this situation, we mark a Requires relation between the two frame elements.

(19) The robbers tied [Paul Item] [to his chair Goal].
(20) * The robbers tied [Paul Item].

The Requires relation occurs in almost all frames that have a construal alternation between a symmetric construal, when a single frame element name is used, and an asymmetric construal, when two frame elements with names of the are used. In the former case a simple name of the form [FENAME]s is used and in the latter, two FEs of the form [FENAME]_1 and [FENAME]_2 are used. Some sample frames are Compatibility, Chatting, and Similarity; there are many more.

3.2.2.3 Excludes

In some cases, if one of the FEs in a group of conceptually related FEs shows up, no other FE from that group can. Again, in the Attaching frame, if Items occurs, then Item and Goal are excluded. In this situation, we say that Items excludes Item and Goal.

(21) The robbers tied [his ankles Items] together.

The above Excludes-relation in the Attaching frame is an instance of a much more common pattern of alternation between a symmetric/reciprocal and an asymmetric construal of events or states involving two parties. In most frames, where the alternation is possible, the names of the frame elements reflect the underlying alternation between reciprocal and asymmetric construal. For instance, in the Similarity frame with lexical units such as similar, different, etc. we have the frame elements Entity_1 and Entity_2, and Entities. Usually, one is allowed to infer equal participation in the event or state by the grammatically less profiled participant (Participant_2). However, since in the case of Attaching, the Goal (which would be Item_2 under our normal naming scheme) is not readily understood as itself being secured or immobilized via attachment to the Item on the asymmetric construal, we selected a name that reflects the fact that the usual inference to equal status for Participant_2 in the asymmetric construal is not warranted.

The Excludes relation also manifests in frames where an event can be brought about either by an intentional Agent or by a Cause event. Consider the following examples from the Placing frame.

(22) [The same flood tide that had brought such a good harvest of tiles Cause] heaped a mass of driftwood onto the Reach.
(23) [Bill Agent] deposited the bag of croissants and the Financial Times carelessly on the hall table.

The two sentences represent two different construals of Placing scenes. Sentence (22) focuses on an event as causing the change in location of the Theme, whereas sentence (23) focuses on an Agent who through their involvement in an unspecified event, most likely an intentional action involving his hands and body, causes the change of location of the Theme. The two construals are incompatible (since there is only one subject slot) and the frame elements Agent and Cause stand in an Excludes relation to each other.⁸

Another clear instance of the Excludes relation between frame elements occurs in the Evading frame, where an Evader moves under its own power to thereby avoid Capture or contact with a Pursuer. The Capture is an actual or hypothetical event in which the Pursuer takes physical control of the Evader. The Capture frame element and the Pursuer are thus clearly interrelated but only one of them can appear as a dependent of a target in the Evading frame.

(24) Sheriff's officials said they apprehended a gang member after he evaded [them Pursuer]. (25) He had successfully evaded [arrest Capture].

Finally, note that the Excludes relation strictly applies only to the direct syntactic dependents of a target word, that is, to first layer annotation. Frame elements that exclude each other may co-occur in an annotation set if they appear on separate annotation layers.

(26) Perkins McLain evaded [capture [through Spain Pursuer] Capture] . (27) The discussions [between [Miller Interlocutor_1] and [the dean Interlocutor_2] Interlocutors ] went nowhere.

In (26), information about the Pursuer is expressed inside the Capture frame element in a prepositional phrase dependent of the noun capture. In (27), the two sides of the discussion, Interlocutor_1 and Interlocutor_2 are expressed within the coordinate NP that encodes the Interlocutors frame element.

3.2.2.4 Future refinements

The treatment of frame element relations sketched in the preceding sections is adequate for a large number of frames. However, two systematic problems remain.

One problem is that we have no explicit treatment of the idea of proto-frame elements, of which other frame elements are more specialized expressions. In cases like (28), we would prefer not to have to pick either specifically Agent or Cause as constructionally null instantiated, since the context might not provide enough information to resolve that question. Instead we would make reference to a superordinate frame element (call it *Force) that is vague about intentionality and the event-person distinction. Likewise in (29), where B answers a question about a new employee, we would prefer to use a superordinate frame element (which might be named *Field in this case), rather than choosing among the frame elements Role, Skill, Knowledge, or Focal_participant for constructional null instantiation relative to the predicate good in the Expertise frame.

(28) The car got damaged while parked outside of our house.
(29) A. How's Susan working out? B. She's very good.

Having an explicit representation of proto-frame elements would also be useful in dealing with certain linguistic expressions that seem to instantiate the superordinate proto-frame element rather than one of its more specific manifestations. Consider the phrase trench to trench in (30): it does not refer to the Source or Goal of a Path but neither does it refer to a middle Path which would be compatible with the specification of a final Goal. A proto-frame element Full_path would provide the most adequate treatment.

(30) He crawled [trench to trench], looking for some sign of Stephen.

Another use for Proto-frame elements involves inheritance relations. In some cases, an inheriting frame will allow only one FE from an Excludes or CoreSet group in the parent frame. Superficially, this violates the rule that requires child frames to have a corresponding FE for each core/peripheral FE of the parent. In a deeper sense, however, inheriting only one member of a frame element set should be permitted on the understanding of inheritance as subtyping. This is so because the child frame is fulfilling every constraint of the parent, merely adding a constraint that prevents one of the construals possible in the more generic case. If we state the frame element restriction on inheritance so that it pays attention only to the most generic level of FEs, then mappings from subsidiary FEs are allowable, but not required. This would make it possible, for instance, to have a *Murder frame (with only agentive causes) as a child frame of the Killing frame (which allows Causes or Agents). ⁹

The second major problem that remains concerns the treatment of subject selection constructions. The current treatment of coreness requires that all frame elements that can occupy the subject position be marked as core frame elements. In many frames this leads to Means and Instrument frame elements having core status and further being part of a CoreSet with Agent, since the two frame elements may co-occur.

Consider the verb open in the Closure frame. In (31), we have a canonical Agent subject but in (32) an Instrument fills the subject slot. (Note that the frame conceptually requires an Agent; uses of open involving a Cause such as The wind opened the door are handled in another frame.)

(31) [John Agent] opened the door.
(32) [The key Bill gave him Instrument] opened the door right away.

Given the earlier discussion of Frame Development (in section 2), Instrument should be core in the Closure frame since the noun phrase realizing it in (32) appears in the subject position. However, were there a separate way of representing the fact that English allows frame elements that are situated between the end-points of a causal chain to occur as subjects, it would not be necessary to give Instrument core status. This would avoid introducing a coreness set of Agent and Instrument.

A separate treatment of these subject selection facts would be parsimonious and would also expose the essential lexical similarity between the English frames and the frames of languages such as Japanese, where subject selection is much more restricted and Instruments and Means rarely, if ever, appear as subjects.

3.2.3 Null instantiation

Sometimes FEs that are conceptually salient do not show up as lexical or phrasal material in the sentence chosen for annotation. Nevertheless, we indicate their absence since it provides lexicographically relevant information regarding omissibility conditions. The FE that has been identified indicates which semantic role the missing element would fill, if it were present.

With respect to null instantiation, verbal, adjectival, and prepositional targets are treated identically. (For null instantiation with noun targets, see 3.4.3.) The following examples show omitted elements with each part of speech. (The name of the frame element is given in square brackets and the frame of the target is given in parentheses.)

(33) That will suffice . [DNI Standard] (Sufficiency)
(34) The result should be similar . [DNI Entity_2] (Similarity)
(35) I tried to put the toys back in . [DNI Ground] (Locative_relation)

Not all cases of frame element omission are alike. We recognize three different cases, one that is not lexically specific and two that are. We will now discuss the three types of omission, focusing on null instantiation with verbs where the phenomenon is clearest.

3.2.3.1 Definite Null Instantiation (DNI):

The first type of lexically specific null instantiation to be considered is the definite (or anaphoric) type. Cases of definite null instantiation are those in which the missing element must be something that is already understood in the linguistic or discourse context. In the following example, the Offender is not expressed overtly in the syntax, but its referent has to be known in the context.

(36) [The monkey Avenger] avenged [himself Injured_party] [by growing to the size of a giant and setting fire to the city Punishment]. [Offender DNI]

3.2.3.2 Indefinite Null Instantiation (INI):

The indefinite cases (sometimes also referred to as existential) are illustrated by the missing objects of verbs like eat, sew, bake, drink, etc., that is, cases in which these ordinarily transitive verbs can be spoken of as used intransitively. (e.g. Molly rarely eats alone; Granny begins baking about a month before Christmas; Michael even drinks heavily on weeknights.) As is well known, there are often special interpretations of the existentially understood missing objects. For example, with eat the missing entity is likely to be a meal, with bake it is likely to be flour-based foods, with drink it is likely to be alcoholic beverages, etc. In contrast to anaphoric omissions, with existential omissions, the nature (or at least the semantic type) of the missing element can be understood given conventions of interpretation, but there is no need to retrieve or construct a specific discourse referent.

For example, in the Revenge frame, all lexical units happen to allow the frame element Punishment to be omitted under indefinite null instantiation. This is shown for avenge in (37).

(37) He took it out on Scarlet in the same way as [he Avenger] avenged [himself Injured_party] [on her Offender] [for the pressures at work and the demands of his first wife Injury] . [INI Punishment]

Note that both in the case of definite and indefinite null instantiation, the LUs in a frame may differ from each other in whether or not they allow the omission. For instance, while eat allows its object to be omitted, devour does not, even though they are both in the Ingestion frame.

Verbs that usually require an argument to be present (or only allow it to be omitted under conditions of definiteness) can be used in a generic construction with indefinite null instantion, as shown below.

(38) He takes and never gives back. [INI Theme]

Figure 38 shows the FrameNet Desktop opened for the annotation of a case of definite null instantiation. Notice the tab to the right of the frame element Offender in the picture, which shows that the FE was omitted under definite null instantiation. In addition, an appropriately colored DNI tag appears at the end of the sentence in the corpus viewing section of the FNDesktop.

Figure 38:Annotating a Null Instantiated Frame Element

3.2.3.3 Constructional Null Instantiation (CNI):

Constructionally omitted constituents (also called structurally omitted) have their omission licensed by a grammatical construction in which the target word appears, and are therefore more or less independent of the LU. Cases of CNI include: the omitted subject of imperative sentences, the omitted agent of passive sentences, the omitted subjects of independent gerunds and infinitives (i.e., the PRO-elements of generative grammar), and so on. In each of the following two examples, the FE Avenger is tagged with the symbol CNI.

(39) Family feuds last for generations, and [slurs on honor Injury] are avenged [by murder Punishment] . [CNI Avenger]
(40) Get even [with her Offender] [for this Injury] [CNI Avenger]

In addition, we use CNI for missing objects in instructional imperatives such as exemplified below, even though the omission is not dependent on a particular construction, but rather on particular genres, such as cookbooks and product labels.

(41) Cook on low heat until done. [CNI Food]
(42) Tie together loosely. [CNI Items]

The experiential perfect also licenses object omissions that are not possible in simple assertions of frame instances.

(43) Have you ever fostered [CNI Child] before?

Note that particular constructions licensing argument omission specify particular intepretations, either indefinite (existential) or definite (anaphoric). For instance, the instructional imperative construction in (41) and (42) specifies a definite interpretation, whereas the experiential perfect exemplified in (43) specifies an indefinite interpretation. The CNI label thus collapses the interpretational distinction that we make among the lexically licensed omissions; it does not represent a separate kind of interpretation in addition to the definite and indefinite types.

3.2.4 Frame element conflation

In some cases, information about two frame elements is expressed in a single constituent, a situation we call conflation. For instance, the concept of ousting somebody from office requires an understanding of the incumbent of the office and the identity of the office; each can be represented separately in a sentence like We ousted Jones as mayor. But in We ousted the mayor, the direct object stands for both the office and the incumbent.

We also find examples of frame element conflation in the Revenge frame. In particular, the Injured_party may be contained as a possessive in the phrase that realizes the Injury frame element, as seen in the following example:

(44) [He Avenger] avenged [Pedro's death Injury] [by taking out the poker-faced Guards Officer Punishment].

Here, the possessive Pedro's realizes the frame element Injured_party, the person who suffered the Injury. In such cases, the annotation tool allows for the creation of an additional FE layer, enabling the secondary annotation of (parts of) constituents in the same frame, as shown in Figure 44.

Figure 44: Secondary FE Annotation

Note that there is never a phrase type or grammatical function indicated for the frame elements on the secondary FE layer.

Even with conflation, it is still possible for the FE annotated on the secondary layer to be instantiated as a constituent of its own in the syntax. In example (45), information about the person who was hit is encoded by the direct object me, as well as by the possessive determiner my inside the NP complement of the prepositionon.

(45) He hit me on my hand.

In such cases, we tag the possessive on the second layer with the same frame element label that is applied on the first annotation layer to the object of the verb.

3.2.5 Syntactic locality

In general, we select sentences for annotation where, with the exception of subjects, we find all frame elements realized by constituents that are part of the maximal phrase headed by the target word. There are two types of situations in which we annotate non-local constituents with frame element labels. In each case the motivation for annotating constituents that bear no syntactic relation to the target is lexicographic: the non-local constituents contain lexical material and as such are of interest to the study of collocations since they provide more information about the semantic type of the frame element than the locally occurring co-indexed phrases or empty elements do.

The first case in which we annotate non-local constituents consists of cases in which the target word is syntactically governed by a raising or control predicate. In such cases, the valence properties of the higher raising or control predicate guarantee that one of its arguments is also interpreted as an argument of the target, even though the relevant argument is not dominated by the maximal phrase headed by the target. Some of the most common types of control and raising predicates are illustrated below. The control or raising predicate that guarantees the interpretation of the non-local noun phrase as a frame element of the target appears in typewriter font.

Raising

Subject to object (46) We expect [John Avenger] to retaliate [against us Offender] [INI Punishment] [DNI Injury].

Subject to subject (47) [John Avenger] seems to have avenged [the death of his brother Injury] [by luring Smithers into a trap Punishment].

Control

Subject control (48) [They Avenger] are hoping to get even [with Smithers Offender] [for the insult Injury].

Object control (49) The commander ordered [the troops Avenger] not to retaliate [against the rebels Offender]

Tough-movement

(50) [The defeat Injury] was difficult to avenge. [CNI Avenger]

Note that Raising and Control cases are not restricted to verbal controllers, i.e. nouns may also serve that function, as illustrated here.

(51) Only a short few weeks ago, even [my Perceiver] hope of seeing [her Phenomenon] was just a dream [Subject control].
(52) Meanwhile, today, Americans, hungering for victory, are puzzling over the Pentagon's order to [the troops Agent]not to put [the Stars and Stripes Theme] [on their vehicles Goal]. (Object Control)
(53) The testers gave it a clear thumbs up both for [its Created_entity] ease of assembly and sail performance. [Tough movement]

The second case in which we annotate clearly non-local constituents with frame element labels concerns targets that occur inside relative clauses. Here our policy is to tag not only the constituent containing the relativizer (if there is one) as a frame element but to also repeat the FE/GF/PT triple on the antecedent, and to further mark the relative word and the antecedent phrase on the Other layer. Thus, our annotation for the simplest cases with an overt relative word is as shown in Figures 53 and 53.

Figure 53: Annotation of a target in a that-relative clause

Figure 53: Annotation Annotation of a target in a which-relative clause

Note that, in contrast to the examples above, the relative word is not always by itself a phrasal constituent. As Figure 53 shows, the constituent containing the relative word may be complex.

Figure 53: A relative phrase containing `whose'

Similarly, the antecedent phrase may be a rather complex phrase, as shown in Figure 53.

Figure 53: A relative clause with a complex antecedent phrase

When a target occurs in a relative clause without an overt relativizer, as in Figure 53, we only annotate the antecedent phrase and mark it with the label Ant on the Other layer. Since we do not assume any kind of zero or non-overt relativizer, the label Rel is not applied anywhere on the Other layer.

Figure 53: A relative clause without relativizer

If a relative phrase is governed by a preposition, we end up with identical FE/PT/GF triples applied to antecedent phrase and the prepositional phrase containing the relative word; the Ant and Rel labels are applied as usual, as shown in Figure 53. In sentences with preposition stranding, the same FE/PT/GF triple occurs on the antecedent phrase, the noun phrase containing the relative word (if there is one), and the preposition. The preposition never gets a Rel label assigned to it even in cases where there is no overt relativizer, as in Figure 53.

Figure 53: A relative clause with a relativizer governed by a preposition

Figure 53: A relative clause with preposition stranding

Notice also that the above principles for relative clauses carry over to Gov-X annotation. When the target noun is the antecedent for a relative phrase that is an argument of a verb annnotated as a governor, we split the antecedent and relative phrases in the same ways as illustrated for cases of normal frame element annotation and also apply Ant and Rel labels on the Other layer in the usual fashion. An example of a governor occurring in a zero-marked object-relative clause modifying the target noun is given in Figure 53.

Figure 53: Gov-X annotation with the governor inside a relative clause modifying the target

Finally, infinitival relative clauses with an overt relativizer are treated just like finite relative clauses with an overt relativizer, which is illustrated in Figure 53.

Figure 53: An infinitival relative clause with overt relativizer

Infinitival relative clauses without relativizer, by contrast, receive no Ant-Rel marking at all. For instance, in books to read over the break the head that the relative clause modifies does not receive an Ant label.

3.2.6 Governing verbs of target nouns, adjectives, and prepositions

When annotating verbal targets, we do not record any predicates that may govern them. However, in the case of nominal, adjectival, and prepositional targets there exist several special classes of syntactic governors that we want to keep track of for lexicographic reasons.

In the case of Support expressions, Copulas and Controllers, it would have been theoretically justifiable to omit selecting phrases outside of the standard subcategorization frame of the target noun and to instead rely on automatic tools for syntactic analysis to identify phrases outside the target's maximal projection that give information about the filler of a frame element role. However, since one of our goals is to provide a database that includes samples of phrases capable of satisfying particular FE requirements of the analyzed words, our decision was to increase the scope of our annotation instead. A welcome by-product of this decision is that the FN database can also serve as a resource for identifying the Support verbs and prepositions, Copulas, Controllers, and X-Governors that FN annotators often find accompanying particular noun, adjective, and preposition targets.

3.2.6.1 Support predicates

As noted at the beginning of section (3.2), we have a special treatment for sentences in which the syntactic and the semantic head of a clause are different and where a noun target is the semantic head of the clause rather than the verb that governs it syntactically. In these cases, one or more syntactic core arguments of the support verb are necessarily understood as participants in the event or relation evoked by the target noun. These verbal arguments-typically the subject, in some cases the object, in others both the object and the subject-are tagged with labels appropriate to the noun's frame. Examples (54)-(56) exemplify support verb constructions:

(54) Aloha Gang, [Someone Speaker] made a statement about my need two kerrect my shpelling in ze last newsletter.
(55) [Frances Patterson Patient] underwent an operation at RMH today and is expected to be hospitalized for a week ore more.
(56) [One of them] became my successor in the professorship in the University of Michigan and the presidency of Cornell.

By contrast, the verb-noun combinations in (57)-(59) do not involve support structures.

(57) Did you read about his latest mishap in the newspaper?
(58) A senior nurse observed the operation.
(59) John congratulated the new president.

In all three examples (57)-(59), the verbal predicate governing the target noun introduces a distinct event: reading about a mishap is completely independent from participating in it (57); observing something is independent from participating in it (58); and congratulating somebody is independent from the achievement at issue (59). (Actually, relational nouns like president never take support verbs, they project clauses only in combination with copular verbs such as be or become.)

Support verb+noun constructions are not to be equated with idioms. While support verb-noun combinations, too, may involve some measure of non-compositionality, it is normally much less than with true idioms whose meanings cannot be built up straightforwardly from the normal meanings of their parts (e.g. give walking papers/a pink slip/the boot). To classify a verb-noun combination as a support verb+noun construction, it is necessary for

the noun to denote a state, event, or relation by itself,
the verb to syntactically govern the noun,
the verb not to have the same meaning independently of the frame evoking element (more or less equivalently, this can be thought of as the frame evoking noun selecting the support),
the verb to have very little meaning on its own, i.e. the meaning of the combination of support and noun should come virtually all from the noun

Note that elements tagged as Support ultimately will also receive a separate treatment in their own right. Thus, any differences that exist between support predicates (including most saliently the introduction of causation) will be captured by describing these predicates as frame evoking elements in very generic frames and adding the semantic type Support (see further under example 13) to indicate their limited use. In this spirit, the verb lift as used in the sequence 'lift the UN sanctions on the country' was included in the Cause_to_end frame alongside the very general LU end.v.

Verbs are not the only part of speech that can `support' a noun. In some cases, prepositions combine with nouns to yield phrases that behave like predicative adjectives. That is, they can post-modify a head noun, as in (60), or combine with a copular verb to yield finite verb phrases, as is shown in (61)-(63).

(60) The "possessor" is the person in possession of the premises. (61) Are health care workers at risk of getting HIV on the job?
(62) Soon, I was in possession of two dozen Eagles cupcakes decorated with white icing, green sprinkles and little plastic footballs and Eagles helmets.
(63) Some people might think that's out of line with our "democratization" policy.

3.2.6.2 Copular verbs

We may think of be and a few other verbs such as appear, seem, look etc. as a special subtype of support verb with a very minimal semantics when occurring in constructions of the form:

(i) NP₁ Verb NP₂/AJP/PP

Examples include:

(64) John is a sailor.
(65) This seemed a rather redundant effort to many.
(66) Tom appears smart enough.
(67) Massu looked without energy, he looked defeated seated with the towl on his face.
(68) Smithers is the vice-president of the armchair division.
(69) Sue is the mayor.

Appearing in sentence frame (i), these verbs are traditionally called copulas or linking verbs. Be occurs in many other sentence frames too and there are some cases of structural ambiguity where it is a copular verb under one reading, and an auxiliary under the other. In His pastime is annoying the girls, one can understand be either as a copula that pairs the role noun with the role filler, or as an auxiliary of the verb annoy used in the present progressive.

Note some uses of be with a target noun do not bear the label Copula. For example, when be occurs as part of the existential construction, it is tagged with the Exist label. In future data releases, there be occurring in the existential construction will be treated as multi-word Supp (cf. section 3.4.1.2).

As shown below in Figure 69, the label Copula appears on the part-of-speech specific layer of a (non-verb) target, in this case the adjective old.

Figure 69: Annotation of Copula for Target old.a

3.2.6.3 Controllers

Recently, FrameNet has begun to recognize a new category of syntactic governor called Controller, abbreviated in the data as Ctrlr. It covers verbs like merit and offer when they govern event noun targets, as in (70) and (71), as well as verbs like consider and find when they govern adjectival targets, as in (72) and (73). While these predicates introduce a distinct event from that of the target, they do share a frame element with the event of the target. For Controllers of noun targets the shared frame element is typically the subject of the Controller; for Controllers of adjectives of the shared frame element is typically the object of the Controller. The constituent expressing that shared participant is labeled with a frame element relative to the noun target.

(70) [What I paid and the satisfaction received] merits high praise. (71) [The minister] offered help to get the various agencies coordinated. (72) She understood that he considered [the trip] too expensive for them both. (73) I found [her scenes] very funny.

In (70) the complex NP What I paid and the satisfaction received is tagged as the Reason frame element of the Judgment_communication frame evoked by praise and in (71), the minister is tagged as the Helper frame element of the Assistance frame evoked by help. In (72), the NP the trip is labeled as the FE Goods in the Expensiveness frame evoked by expensive. In (73), the object of the Controller find, the phrase her scenes, is labeled as the Stimulus of the Subject_stimulus frame evoked by funny.

The Controller label is applied on the Noun and Adj layers just like the Supp label is.

3.2.6.4 X-Governors

In the context of slot-filler annotation (see 3.8), FrameNet also uses a category Governor, which marks predicates that have a semantic connection to the qualia structure of the target artifact nouns they govern. For instance, stab is treated as a Governor of knife in the Weapon frame. The dependent constituent of the Governor that is headed by the target noun is called X for lack of a better term. No dependents of the Governor other than the X constituent are annotated relative to the Governor. Any frame elements of the frame evoked by the artifact noun that are realized within the X-phrase are annotated with frame elements as usual. As with other kinds of special governors, the Governor label, as well as the X label, are applied on the part of speech specific Noun layer.

3.2.7 Discontinuous frame elements

In some cases, the same FE label appears multiple times relative to a given target. There are two cases: multiple separate instances of the same frame element, as when several Path segments are described for a motion event (Josh ran [across the meadow PATH], [along the creek PATH] and right up to Bill's barn); a single instance of a frame element is realized in two discontinuous pieces, rather than as a single constituent. Here, we consider the latter type of discontinuos FE.

Raising predicates (e.g. may or might) specify that their sole semantic argument be realized in two pieces, one as the surface subject of the clause headed by might and the other as a verb phrase complement of some kind. The verb phrase may be of the type bare verb phrase (VPbrst) as with might in (74) or it may be to-marked (VPto), as with seem in (75).
(74) [The reader NP.Ext] might [want to excoriate the paper for its coverage of the so-called Clear Skies policy VPbrst.Dep]. (75) [Old Europe NP.Ext] seems [to drag her feet VPto.Dep].
The FrameNet treatment does not assume any syntactic derivation of such structures from underlying forms where there is only a single, contiguous frame element. It is a fact about the semantic representation of might that its subject must be understood as an argument of the predicate expressed by its VP complement, rather than a syntactic fact. Accordingly, both pieces are treated as separate syntactic arguments of the raising verb and marked separately with GF and PT values.
The prepositional passive construction shown below also produces discontinuities.
(76) [The Peacock Throne NP.Ext] has been sat [on PP.Dep] by the Iran monarchy since the days of Nadir Shah.
This construction belongs to the set of subject-selection constructions, which also includes the English middle construction (This car drives well). Since the prepositional passive construction is specifically geared towards selecting a non-core frame element as subject, we assign the fronted subject NP the grammatical function Ext and the phrase type NP, while the remaining preposition is tagged like a normal prepositional phrase with grammatical function Dep and phrase type PP.
Another class of discontinuities results from various kinds of fronting constructions which sometimes lead to preposition stranding.
(77) [Where she had been PP.Dep] she didn't talk [about PP.Dep] (Topicalization) (78) [The man PP.Dep] [that PP.Dep] I lent my phone book [to PP.Dep] made off with it. (Relativization) (79) [What PP.Dep] was Bill angry [about PP.Dep] when he got the back spasm? (Wh-extraction)
Unlike the prepositional passive, the constructions above do not place the extracted NP into a core syntactic slot and so both pieces are assigned the same GF. We have also decided to apply the prepositional phrase type label to the displaced phrase, which is usually a noun phrase but not always, as is shown by the topicalization example in (77).
Discontinuity is also produced by what we may call heavy modifier shift. This involves finite relative clauses that are separated from the nominal heads they modify and displaced appositives, as shown in (80) and (81).
(80) About this time [a guy NP.Ext] appeared [who had only a pair of boxers on NP.Ext]. (81) [Mrs Fiorentina NP.Ext] spoke next, [the chairwoman of Huelit-Peccard NP.Ext].
Both pieces get the same grammatical function and phrase type values. Note that we do not treat the following examples as heavy modifier shift; given the assumed verbal targets, the adjectival and prepositional phrases included in curly braces {} would simply be labeled Depictive.
(82) [Mr Burns NP.Ext] stepped out of the shower, {buck-naked}. (83) [Smithers NP.Ext] showed up later {with a funny hat on}
The decision to proceed in this way is motivated by two considerations. First, Depictives express a temporary state of a participant during an event; no such semantic restriction applies to displaced appositives or relative clauses. Second, Depictives are generally more flexible with regard to their position than displaced appositives and relative clauses, as shown in the following examples:
(84) [Looking stunned Depictive], Mr Burns dialed his secretary. (85) Mr Burns dialed his secretary, [looking stunned Depictive]. (86) Mr Burns, [looking stunned Depictive], dialed his secretary. (87) *Who looked confused, Mr Smith came in. (88) *The former mayor of Denver, Wellington Webb was spotted at the buffet.
For more discussion on Depictives, please consult Chapter 8 .
The last class of discontinuity consists of modifying structures in which, in addition to the modified head noun or verb, the modifier takes a second argument that is realized following the head. At least with some modifiers such as similar in (90), the second argument is subcategorized by the pre-head modifier (rather than licensed by a construction). ¹⁰ In these cases, both pieces of the modifying structure fill the same grammatical function relative to the head and we assign them the same GF value, Dep. However, we treat the two pieces separately in terms of their phrase type, as is shown by the following examples. (The relevant words and phrases are annotated with triples of FE, grammatical function and phrase type).
(89) Quite bluntly, you're a [dumber Descriptor.Dep.AJP] man [than I Descriptor.Dep.PP]! (90) The general has committed a [similar Descriptor.Dep.AJP] blunder [to our entry into the eastern valley Descriptor.Dep.PP]. (91) Yet what your school did to them [so Degree.Dep.AVP] irked them [that they'd rather pay the new school than pay the old school Degree.Dep.Sfin]. (92) Real people sometimes prove to be [more Degree.Dep.AVP] annoying [than anyone anticipated Degree.Dep.Sub].
We assign separate phrase types to each piece here for a purely technical reason: the chunk parser that is used as part of the process that extracts example sentences from the full BNC corpus does not produce the correct result if the two pieces share the same phrase type.

3.3 Annotation with verbs as targets

Frames can be evoked by words in any of the major lexical categories of noun, verb, and adjective, as well as by adverbs and prepositions. We will begin our discussion with verbs.

3.3.1 Easy cases

Annotation is easiest when all and only the core

frame elements (the conceptually necessary participants of the frame that a syntactic governor evokes) find syntactic expression in the sentence as separate immediate syntactic dependents of the governor. Under such circumstances, we simply annotate each of the syntactic dependents for the three kinds of information: Frame Element (that is, semantic role), Grammatical Function, and Phrase Type.

The different kinds of information are recorded on separate annotation layers. The separation of layers makes it possible to represent many complex situations, such as when the constituent that realizes one frame element is contained within the constituent that realizes another, or when the semantic and syntactic constituency don't match. Usually, however, most tagged constituents consist of triples of information, with coterminous tags on three layers, and they are displayed as such in the FrameNet Desktop. Figure 3.3.1 is a screen shot of the middle window of the annotation software in which an example sentence has been annotated. The names of the layers appear in the left-most column of the bottom frame: FE (Frame Element); GF (Grammatical Function); and PT (Phrase Type); Other (labels that deal with a small set of special syntactic constructions); and Verb (a layer named after, and with labels specific to, the part-of-speech of the target)

Figure 3.3.1: Annotation window with target LU avenge.v

In practice, annotators only need to apply a Frame Element label; Grammatical Function and Phrase Type are derived algorithmically based on position relative to the verb and patterns of part-of-speech labels, but may require manual co

FrameNet II: Extended Theory and Practice

Josef RuppenhoferMichael EllsworthMiriam R. L. PetruckChristopher R. JohnsonJan Scheffczyk

Printed June 20, 2006

Josef Ruppenhofer
Michael Ellsworth
Miriam R. L. Petruck
Christopher R. Johnson
Jan Scheffczyk