The frames that we create, and thus the Frame Elements and Lexical Units associated with them, are intended to be situated in semantic space by means of frame-to-frame relations and semantic types. The relations we use include Inheritance, Subframe, Causative_of, Inchoative_of, and Using. There are several benefits to extensive use of these relations:
A number of other benefits are outlined in 6.3.
We have spent considerable time recently improving the semantic relations encoded in our data, and also making the relations more accessible via the FrameGrapher (see http://framenet.icsi.berkeley.edu/~FrameGrapher/). We have also marked a large number of Frame Elements with semantic types indicating appropriate fillers (6.2.1). There remain a dwindling number of completely disconnected frames, all of which are semantically quite distinct from other frames that we already describe.
In addition, we have added two new types of relations, Precedes (replacing the meta-relations for Subframes) and Perspective_on, which replaces a subset of the Using relations with a more specific and informative relation. These should make our data more useful for all forms of computational processing. All of the frame relations referred to in this chapter other than these new ones can be visualized with the FrameGrapher on the main FrameNet website (http://~framenet.icsi.berkeley.edu).
In the sections that follow, general descriptions of each of the Frame-to-frame relations (6.1) and semantic types (6.2) are given, and then a more formal description of the relations and the ways that they may be used for reasoning is discussed (6.3). Note that the following does not discuss the FE-to-FE relations, which have a closer relationship with valence and annotation than with semantically-defined relations. For these relations, see sections 188.8.131.52-184.108.40.206.
Each frame relation in the FrameNet data is a directed (asymmetric) relation between two frames, where one frame (the less dependent, or more abstract) can be called the Super_frame and another (the more dependent, or less abstract) can be called the Sub_frame. We give a more specific, informative name to the Super and Sub_frames for each relation, as shown in Table 6.1 below.
|See_also 24||Referring Entry||Main Entry|
In general, each frame has one relation to some other frame, but there are occasional exceptions, as seen in the relations between Assistance and Intentionally_act, shown in Figure 6.1 below. Assistance inherits from Intentionally_act, with the Helper bound to the Agent of Intentionally_act, but Assistance also uses Intentionally_act since there is a second intentional action presupposed, namely that of the Co_agent, which is thus also bound to the Agent role of Intentionally_act in a separate relation.
This is the strongest relation between frames, corresponding to is-a in many ontologies.25 With this relation, anything which is strictly true about the semantics of the Parent must correspond to an equally or more specific fact about the Child. This includes Frame Element membership of the frames (except for Extrathematic FEs), most Semantic Types, frame relations to other frames, relationships among the Frame Elements, and Semantic Types on the Frame Elements.
Properties of the Parent which are not strictly semantic in nature, such as not being evoked by lexical units (i.e. the Semantic Type Non-lexical_frame), being evoked by a particular set of Lexical Units, or having a See_also relation to another frame, are not inherited.26
Also, when there is a Core-set or an Excludes relation among Frame Elements of the Parent (see sections 220.127.116.11 and 18.104.22.168), these constitute disjunctive properties of the Parent. The Child frame may legitimately inherit only a subset of these disjunctive Frame Elements.
This relation (new in Release 1.3) is a refinement of the more general Using relation (see 6.1.6). Perspective_on constrains related frames considerably more. The use of this relation indicates the presence of at least two different points-of-view that can be taken on the Neutral frame. For example, the Measure_scenario, in which an Entity's Value for some Attribute is described, can be viewed either from the point-of-view of exact measurement (e.g. "Joey weighed 7 pounds.") or as a relative measure (e.g. "Joey was heavy."). The FEs in the two cases are quite different, so the words should not be included in the same frame (see 2), but they do make reference to the same scene. The Perspective_on relation allows us to refer directly to the scene (encoded by the Neutral Measure_scenario frame here) and connect the two. As in this case, the Neutral frame is normally Non-lexical and Non-perspectivalized. (See 6.2.2.)
A single Neutral frame generally has at least two Perspectivalized frames, but in some cases, words of the Neutral frame are consistent with multiple different points-of-view while the Perspectivalized frame is consistent with only one.27 Whenever there is a state of affairs that is describable by a frame in a Perspective_on relation, all the other frames that are connected to it by the Perspective_on relation can also be used to describe the state of affairs.28
For example, the Commercial_transaction frame, diagrammed in Figure 6.1.2, specifies a complex schema involving an exchange of multiple Themes (the Money and Goods) between the Buyer and Seller, including also two subframes: Commerce_goods-transfer and Commerce_money-transfer. The Buying frame has a Perspective_on relationship with the Commerce_goods-transfer subframe in which the Money, Goods, Buyer, and Seller are identified.
Some other relations of this type include:
Further note that it is quite common for a frame to inherit from one frame and be a perspective on another. An example of this is again provided by the Commerce_buy frame, which inherits the Getting frame (not shown in the diagram 6.1.2) but is a perspective on the Commerce_goods_transfer frame. An act of buying is a sub-type of getting, which justifies the inheritance relation to Getting. Buying is an event that only occurs when the situation (the Commerce_goods_transfer) is also describable as selling. For that reason, it (and the Commerce_sell) is connected to the Commerce_goods_transfer frame via a Perspective_on relation.
Some frames are complex in that they refer to sequences of states and transitions, each of which can itself be separately described as a frame. The separate frames (called subframes) are related to the complex frames via the SubFrame relation. In such cases, frame elements of the complex frame may be identified (mapped) to the frame elements of the subparts, although not all frame elements of one need have any relation to the other. (In this respect, it contrasts with inheritance; see below.) Also, the ordering and other temporal relationships of the subframes can be specified using binary precedence relations. To illustrate, consider the complex Criminal_process frame, which is defined as below and whose frame relations are shown in Figure 6.1.3.
A Suspect is arrested by an Authority on certain Charges, then is arraigned as a Defendant. If at any time the Defendant pleads guilty, then the Defendant is sentenced, otherwise the Defendant first goes to trial. If the Finding after the trial is guilty, then the Defendant is sentenced. In the end, the Defendant is either released or is given a Sentence by a Judge at the sentencing.
Notice that a given subframe may itself be a complex frame. For example, the Trial frame is a subframe of the Criminal_process frame, and has its own rich structure, some of which can be decomposed into simpler frames that are related to each other. A Trial is made up of court appearances, and involves opening arguments, presentation of evidence and testimony, and closing arguments. The system of subframe links is also quite complex. At present, the subframe relation is not indicated in every relevant case.
This relation occurs only between two Component frames of a single Complex frame, i.e. as extra information associated with a set of Subframe relations. It specifies the sequence of states and events that are definitional for a certain state-of-affairs. Most Subframe relations will naturally have precedence relations, as can be seen in the foregoing diagram of the subframes of the Criminal_process frame (6.1.3), in which the Precedes relations are indicated by the black lateral arrows.
This is the only frame relation for which cycles are allowed. For example, in the subframes of the Sleep_wake_cycle frame (shown in Figure 6.1.4), Being_awake precedes Fall_asleep, which precedes Sleep, which precedes Waking_up or Getting_up, which in turn precedes the first frame, Being_awake.
We record the especially close and fairly systematic non-inheritance relationships between stative frames and the inchoative and causative frames which refer to them using the frame-to-frame relations Causative_of and Inchoative_of. Consider the following frames: Position_on_a_scale, Change_position_on_a_scale, and Cause_change_of_scalar_position. The following sentences illustrate the kind distinction captured in these frames:
(1) [Cause_change_of_scalar_position] (raise.v) Billie Blount raised taxes on farmers 18 times in 2002! (2) [Change_position_on_a_scale] (rise.v) During the Elizabethan age, there was an increased emphasis on genealogy in the heralds' work as the gentry class rose in importance. (3) [Position_on_a_scale] (high.a) Most fish from lakes is too high in mercury.
Frames which participate in these relations as Causatives should inherit from the Transitive_action frame, Inchoatives should inherit from Event, and States from State or Gradable_attribute.29 Many of these inheritance relations have not yet been made.
Often a particular frame makes reference in a very general kind of way to the structure of a more abstract, schematic frame. Since the creation of the more specific Perspective_on relation (see 6.1.2), the Using relation is used almost exclusively for cases in which a part of the scene evoked by the Child refers to the Parent frame. For example, Volubility uses the Communication frame, since Volubility describes a quantification of communication events.
It is possible for a frame to use more than one frame. An example of this situation is the Judgment_communication frame, shown in Figure 6.1.6 below. It uses both the Judgment frame and the Statement frame. The Judgment_communication frame does not inherit Judgment because it is not a simple subtype of a purely cognitive state. Judgment_communication does not inherit Statement either since it distributes the content of the Statement frame's Message frame element over two frame elements, Evaluee and Reason.
In cases where there are groups of frames which are similar and should be carefully differentiated, compared, and contrasted, each of the frames in question has a SeeAlso relation with a representative member of the group. In the frame definition of the representative member, there will be a comparison which will contrast the frames to make clear the intended boundaries between them. For example, since the Scrutiny and Seeking frames are similar, there is a SeeAlso relation from Scrutiny to Seeking, and text in the Seeking frame that explains the difference.
This relationship does not imply any particular relationship between the Frame Elements or subframes of the frames involved.
The general use of semantic types in the FrameNet project is to record information that is not representable in our frame and frame element hierarchies. In this section there is a detailed description of each major category of the semantic type hierarchy, which is broadly split by function. In specific, the functions we currently employ semantic types for are:
The most interesting function of semantic types for human users is the third one, recording important semantic differences between lexical units that recur within several frames.
For example, "Positive_judgement" and "Negative_judgement" semantic types, indicating the speaker's attitude toward a situation, can be applied to lexical units across a range of frames. (Note that the term speaker may either refer to a frame element such as Speaker or Cognizer, or to the author of the utterance containing the lexical unit.) We capture the distinction by marking LUs like the aforementioned Judgment verb praise, the Experiencer_subj verb like, and the Frugality adjective generous as "Positive_judgement" and, by contrast, marking "Negative_judgement" on some other words from the same frames, e.g. criticize and hate, and stingy.
For more description and further examples, see the following subsections.
These semantic types are used to classify the denotation of lexical units, frames, and frame elements. On lexical units, they classify the denotation of LUs, usually cross-cutting their classification by frames.30 When, on the other hand, such a distinction corresponds to a frame distinction, the frame is labeled with the type, which then signifies that every LU of the frame could be labeled with an equally or more specific Type. For FEs, semantic types actually categorize the sort of filler that is expected in the FE, rather than classifying what kind of role the FE is.31
Most of these types correspond directly to synset nodes of WordNet, and can be mapped onto ontologies, e.g. Cyc. The types are related to each other via subtype relations, which are logically equivalent to Inheritance or is-a. However, the FrameNet semantic type hierarchy is not guaranteed (or even likely) to correspond exactly to that of WordNet, Cyc, or any other resource.
One example of a semantic type which has been used for frames, FEs, and LUs is Container. This corresponds most closely to the WordNet synset node container and the Cyc Container Collection node.32 The Container frame, unsurprisingly, contains LUs which refer to containers like jar.n, box.n, etc., and thus both the frame and the LUs in it can be labeled with the semantic type Container. The Source of the Emptying frame is also labeled with the ST Container. This means that head-nouns of this FE are likely to be the sorts of words that belong as LUs in the Container frame. Pronouns or other kinds of nouns may also occur as fillers of the FE, in which case they are construed as containers, as in the following:
(4) She emptied the lid back into the bottle.
In a neutral context, we wouldn't expect a lid to itself be a container. Not every lid is at all readily describable as a Container, e.g. the lid of a tin can. Regardless of whether a filler of this role is inherently a Container, its occurrence as a filler of the role means that it is being used as a Container in the current context, and has the properties necessary for that.33
These types, as mentioned above, can only be applied to frames and are not to be interpreted as applying to LUs of a frame or any connected frames, including Inheritance daughters (which normally should match any semantic types that are placed on their parent frames).
Such frames have no lexical units and are present purely to connect two (or more) frames semantically. One example is the Post_getting frame, connected to Getting via a Precedes relation and connected to Possession via an Inheritance relation. This allows us to succinctly encode the fact that the state following "getting X" (the Getting frame) is "having X" (the Possession frame).
This semantic type is used for frames which have a great diversity of lexical units, all of which share a kind of scene as a background. Such frames do not have a consistent set of FEs for the targets, a consistent time assigned to the events or participants, or (most especially) a consistent point-of-view between targets. An example of this type of frame is the Performers_and_roles frame, which contains such diverse LUs as co-star.v, feature.v, and as.prep. Like the Biframal_LU types, this semantic type is intended as a time-saving measure. All such frames could be split up into smaller frames with a consistent perspective, but these frames would contain very few LUs. (See 22.214.171.124 on Biframal_LUs.)
There are a number of labels applied to LUs or frames which do not refer directly to the kind of entity denoted by an LU or the semantics of a frame. These demand, rather, some more complicated interpretation of the LU or frame that they are attached to. How entities with such types are to be interpreted is different for each of the types that are described in this section (and in the following section on Framal types), and in some cases the label even indicates that an LU should be interpreted in a different frame from the one that it is attached to (see Guest_LU types below 126.96.36.199), so LUs with these types, if interpreted as subtypes of the semantics of the frame according to the normal FrameNet logic (see 188.8.131.52), can be very misleading. (This means that for some purposes, it would be more sensible to leave out items with these types attached than to include them without further comment.) A description of each of the types follows.
LUs marked as transparent nouns have an unusual sort of semantics since, unlike most nouns, their primary function is to give some kind of description of another noun, as seen in the examples with kind, number, and side below:
Specifically, transparent nouns describe either 1.) information which is not typically inherent in the nouns they describe, such as quantity, grouping, and shape, or 2.) information which can be left implicit in the interpretation of nouns when they denote an instance of an entity, such as a.) instancing itself (e.g. "this bat" = "an instance of a bat"), b.) subtyping (e.g. "this bat" = "a type of bat", since an instance is a maximally specific type, and thus automatically a subtype of its class, or c.) picking out subparts (e.g. "with a knife" = "with the edge/point of a knife"), as an active zone subpart of an entity is often meant even when the whole entity is mentioned.
The descriptive aspects of meaning profiled by transparent nouns are usually of little consequence for determining acceptability as fillers of predicate roles; rather more important is the semantics of the entity being described. This can be seen in the examples below, in which the above sentences are summarized/paraphrased by omitting the transparent noun, leaving only the described entity:
We judge that recognition of these LUs is vital for correct cataloguing of FE fillers, summarization, paraphrase, etc. since they violate the normal rule that the syntactic head of a phrase is the semantic head. For most purposes, transparent nouns can simply be omitted from the analysis of a sentence.
However, note that transparent nouns are not always transparent. In the right semantic contexts, they are so to speak "opaque" and dominate the frame of their dependent (as most targets do); in such cases the meaning of the so-called transparent noun itself rather than that of its dependent is selected by a predicator. In particular, this is the case with governing predicates that evoke an open proposition or question, e.g. determine, find out, measure, etc.
(11) Archaeologists have been determining the number/shape/part/type of tools used by the Maya.
Here none of the transparent nouns can be used with quite the same meaning as the that of the described entity alone:
(12) Archaeologists have been determining the tools used by the Maya.
In addition, transparent nouns convey important, if optional, information like quantity and configuration. For tasks concerned with these dimensions of meaning, transparent nouns should be treated like any other target.
Some LUs of a frame cannot be used unless they co-occur with the right other words. The limiting case, where a precise group of words must all co-occur in order to evoke a frame, is handled in FrameNet by the use of multi-word lexical units (4). But there are many cases where an LU may evoke the frame when occurring with any of a number of different words, or even when occurring with a class of words. Such LUs are to be marked with the Bound_LU semantic type.
There are two subtypes of Bound LU that we recognize in FrameNet: Supports and Bound_dependents.
This subtype of bound lexical unit is applied to LUs that only evoke a frame when they are used as Supports of certain dependents, as exemplified by give.v in sentences like the following:
(13) Receiving the notification so late almost gave me a heart attack.
In this sentence, give is used approximately synonymously with cause:
(14) Receiving the notification so late almost caused (me to have) a heart attack.
This use of give, then, should go in the Causation frame, but it is readily apparent that this meaning is not possible in many (if not most) contexts:
As implied above, give occurs in this meaning only when it is to be annotated as a Support Verb (see 184.108.40.206). (Note that there are other support uses of give which are not equivalent to cause, e.g. give a laugh, give a speech.)
In principle, though not yet in fact, all supports with any semantics not included in the target noun, adjective, or preposition that they are associated with, such as causative supports, inchoative supports, and point-of-view-shift supports, should be annotated separately in a causative, inchoative, or point-of-view specifying frame.
In addition to fixed expressions and support constructions, there is a small class of other LUs which only occur as dependents of a limited set of governors. This semantic type has not been included on LUs in the current data release, but it is intended to cover semi-productive LUs such as attention.n in the Posture frame (in phrases like (stand) at/to attention), bind.n in the Predicament frame (in phrases like in/out of a bind), and idiosyncratic degree modifiers from the yet-to-be-made *Degree frame (for LUs like very.adv), e.g. dirt in dirt poor, crystal in crystal clear, sopping in sopping wet, etc. These LUs can be added to appropriate frames and given a Bound_dependent type, from which the user should infer that they are only usable when they have particular governors. The appropriate governors must be gleaned or generalized from the usage in the annotation data.
LUs of a frame normally have a semantics which is a subtype of the semantics of the frame. However, when LUs are marked with a Biframal_LU type, they denote something related, but not equal to the semantics of the frame. In all cases where an LU is marked as biframal, we could have made a separate frame to more narrowly characterize the meaning of the LU. These types have been defined as a time and resource-saving measure, to avoid having to create these separate frames for tiny groups of LUs; for example, (un)intelligible.a, despite the close relationship with the Grasp frame, would otherwise require a separate frame.
Another way of looking at this is that the normal relation of the semantics of the LU to the semantics of the frame is Inheritance, while the relation of a biframal LU's semantics to the frame's semantics is generally Using. Except for the Guest_LU type, each of the following subtypes define what frame the biframal LU inherits from. (See below for Guest_LU.)
We tend to avoid using these semantic types, since they complicate the interpretation of our data, and in many cases there is a problematic mismatch in the FEs allowed for the biframal LU and the other LUs of the frame.34
This type is applied to LUs which are only tangentially used in the host frame, and whose interpretation is still largely dependent on their membership in some other frame, as demonstrated in Example 220.127.116.11.
Thus, for example, grin.v can be included in the Statement frame, but any full understanding of this sentence requires an understanding of the Making_faces frame.
This type is used to mark LUs which denote the Agent35 of the frame in which they occur, as in murderer.n in the Killing frame. They can be thought of as virtually inheriting from People and using the frame they are listed in; the Person FE of the People frame is bound to the same FE that is bound to the Agent FE of the host frame.
This type marks LUs which denote the Patient36 of the frame in which they occur, e.g. possession_((entity)).n in the Possession frame, victim.n in the Crime frame, and knowledge.n in the Awareness frame. These LUs can be thought of as inheriting from the Entity frame, and as using the host frame, with the Patient (or equivalent) FE of the host frame bound to the same FE as is bound to the Entity FE of the Entity frame.
This type marks LUs which denote the propensity (of something) to be a Patient (or similar FE, see 21) in the host frame; e.g. verifiable in the Verification frame is marked as a tendency-grading LU since it concerns something's tendency/ability to be verified. The semantics of a tendency-grading LU can be understood as a virtual frame inheriting from Inclination and using the host frame.
The most common subtype of these LUs are abilitative passives, usually constructed morphologically from a verb (X) + (a)ble, which denote the propensity of an entity to be Xed. Normal, productively formed examples include findable (Locating), usable (Using), likable (Experiencer_subj), and (with negative prefix) unfixable (Resolve_problem). There are also semantically indistinguishable examples which are not productively formed from verbs, e.g. intelligible (Grasp). These LUs are listed in frames that inherit from the causal Transitive_action frame.
In addition, there are some LUs which do not directly make reference to a causal frame, but rather refer to an inchoative frame, e.g. fragile.a (Fragment), and mortal.a (Death).
Many LUs describe the positive or negative opinion of an implicit or explicit Judge, without actually necessarily being of similar type otherwise, e.g. like.v in the Experiencer_subj frame, and stingy.a in the Stinginess frame. This type labels such LUs so that they can be compared despite their dissimilarities. In addition to the positive or negative dimension, there are two main types, distinguished by who is identified as having a high or low regard for something: 1.) Emotion-related targets, where the Experiencer (or descendant of Experiencer) is explicitly identified as the opinion-holder and the opinion concerns an overt Stimulus or Evaluee, e.g. approve.v, and 2.) other targets where the user of the word in the discourse is the opinion-holder and the opinion concerns the target predication as a whole, e.g. thrifty.a. All LUs in frames which inherit from or use the Emotions frame, especially the Judgment family of frames, fit in type 1, with the Cognizer, Speaker, or Experiencer as the opinion-holder, whereas all other targets express the opinion of the individual using the word.38
In words of the second type, where it is the language-user whose view is expressed, the full semantics of the LU could be thought of in terms of a virtual frame which inherits from the host frame and uses the Regard frame, with the Cognizer specified as the language-user, the Evaluee identified as the whole predicate, and the Judgment identified as positive for Positive_judgement and negative for Negative_judgement. Thus the target bony.a of the Body_description_holistic frame in 18.104.22.168 below indicates not only that someone was thin (which is at least partly specified by the meaning of the frame) but also the person declaring this state considers this undesirable.39
These types are employed on LUs to indicate that the LU inherently specifies some information about an FE. Although the types of information that LUs can specify about FEs is considerably broader (including most saliently filler-types for an FE which are particular to an LU, as in tie.v in the Attaching frame), we have so far used only two basic kinds, which specify Degree and sensory modality respectively.
These types are used for LUs whose framal semantics provides a Degree FE. All such LUs inherently describe a particular Attribute of an Entity as deviating in a particular direction from the norm. (This norm is usually the expected value of the Attribute for the type of Entity described, but may be a norm for some more abstract supertype of the Entity. See the Gradable_attributes frame.) LUs should pre-specify the direction and difference from the norm; Negative indicates the direction and End_of_scale indicates maximal difference.
LUs which have the type Negative specify that the direction of deviation is in the negative direction, i.e. either less in quantity or desirability (or at any rate in the opposite direction from the unmarked description given in the frame definition). These LUs thus form antonyms to unmarked LUs in the same frames. Some select cases are presented in Table 22.214.171.124. Note that many LUs that should receive this type have not been marked in the FrameNet data.
|Frame||Negative LU||Positive LU|
End_of_scale marks LUs that characterize an Entity as having a maximal value for an Attribute, e.g. fabulous in the Desirability frame. Such LUs are not usable with normal Degree modifiers like very.
LUs marked with these types inherently reference some type of sensory experience, either directly denoting such an experience, e.g. see.v, look.v, or other LUs in the Perception_experience or Perception_active frames, or they use such a frame, and thus imply the occurrence of a perceptual event, e.g. tasty. In any case, the important function of the subtypes of this type (viz. Visual_modality, Tactile_modality, etc.) is to identify the subtype of sensory experience which is inherent to an LU. Some examples of the use of the Visual_modality type:
This section outlines some of the major uses for the relations and semantic types. Section 6.3.1 is primarily devoted to describing the utility of the relations for humans using FrameNet as a reference, Section 6.3.2 is of use to both humans and automatic programs, whereas the other sections are of more interest to developers intending to use FrameNet for computational purposes. Virtually any computational use of the FrameNet relation information relies on and presupposes semantic parsing of texts-a process not discussed here, but well covered in several publications (e.g., [Erk and Padó, 2006] and [Litkowski, 2004]).
One obvious benefit of defining relations between frames is that it enables the look-up of words that are related. For example, we can look at the word murder.v which occurs in the Killing frame. Obviously, there's some level of similarity with the other members of the frame, since they all involve something causing an entity to die. This sort of relationship might be captured in a good thesaurus under an entry for kill.v. However, if we follow the Causative_of relation from the Killing frame to the Death frame, we find a whole new group of words, also about death, e.g. die and pass away, but in this case focusing on the change of state of the Protagonist. This kind of relation is neither synonymy nor antonymy, and therefore not to be found in a thesaurus.
In Fillmore's earlier work [Fillmore, 1968][Fillmore, 1977], a case was made for the universality of certain types of semantic roles, a concept which was further developed and is now enshrined as the theta role system of many syntactic/semantic formalisms. These roles include such labels as Agent, Instrumental, and Objective (roughly corresponding to Agent, Instrument, and Patient in other formulations). However, as the description of the semantics of Lexical Units has progressed, it has become apparent that the theta-role and original case-role account covers only a subset of the full set of roles.
We now take it that theta roles should be mapped to FEs in high-level, abstract frames like Transitive_action, which has FEs like Agent and Patient. The relevance or irrelevance that these labels have for the roles of more specific predicates like break.v (in the Cause_to_fragment frame) or resemble.v (in the Similarity frame) is explained explicitly by the inheritance or non-inheritance of the Agent and Patient frame elements in the relevant frames.
There are inherent problems to reducing our role-set of Frame Elements to the considerably smaller (and inarguably more computationally tractable) set of theta roles often used. One of these is deciding on the initial set of theta roles - a well-known problem in the theta-role literature. The Frame Elements that we define, however, are more immediately verifiable. Presumably, then, any theta-role system proposed to cover all predicates should allow us to specify, in a simple way, which FEs should be mapped to which theta-roles. Of the theta-role systems known to us, none allows any simple mapping to high-level FEs (and thence to the FEs that inherit from them) without covering some FEs multiple times and/or leaving some FEs uncovered, unless there is an unsatisfactory catch-all theta-role.
In many ways, paraphrasing is at the core of what we intend FrameNet to facilitate. A properly powerful ability to paraphrase enables many of the other goals of semantic NLP, including Question Answering, Summarization, and Translation. Question Answering can be thought of as looking in a corpus to find a paraphrase, but with real information filled in for the questioned FE. Summarization is equivalent to paraphrase of a text, but with the strategic omission of information from FEs and targets. Translation is paraphrasing with the limitation that all the resulting paraphrase must be in the target language.40
One of the basic insights behind FrameNet is that grouping words according to the scenes that they evoke, regardless of whether they are synonyms, antonyms, or some other relation to each other, groups words that are useful for paraphrasing. In particular, since FrameNet lists words together despite part-of-speech differences (unlike WordNet), paraphrases involving an interchange of noun, verb, adjective, or preposition are (in principle) discoverable with the FrameNet data. (See Sect. 31.)
Despite the fact that FrameNet is built to facilitate paraphrase, more specific relationships between the words usually need to be ascertained to recognize or generate actual instances of paraphrases. What kind of paraphrase results is dependent on whether a proposed paraphrasing target is an antonym, exact synonym, subtype, or supertype of the initial target to be paraphrased.41 We have some of this more specific information (see the Semantic type "Negative", 126.96.36.199), but for the most part, we have envisioned the information that we record as complementary to WN.
As WN has a reasonable coverage of synonymy, antonymy, and subtyping ("is-a") relations between words, we have tried to avoid reinventing the wheel by duplicating this information in FrameNet. However, for tasks that require information from both sources, integrating the information would depend on mapping FrameNet LUs to words in WN synsets in some way. This has not been consistently accomplished anywhere, although work has been done by several researchers.42
Once a proposed list of paraphrases for a starting LU is established (however buggy), then targets can be paraphrased with other words which are synonyms, supertypes, subtypes, or (with the addition of negation) antonyms in the same frame; this excludes, then, targets which are "sisters" or "cousins", i.e. which are subtypes of some supertype of the original target.
Working without a WN-FN mapping, antonyms can at least provisionally be established if some members of the frame have the Semantic type Negative. Such targets are antonyms of unmarked targets in the same frame. In addition, targets with FE incorporation should be considered subtypes of targets in the frame without incorporated FEs. In the absence of any other information, all other targets could be considered synonyms, although this will be erroneous in many cases.
Paraphrasing an initial target with a target that is an antonym requires adding in negation, as can be seen in 27 and 28 with antonymic targets from the Possession frame:
When paraphrasing a more specific initial target with a more general target from the frame, no adjustment is generally necessary, just as with a synonym, as seen in examples 29 and 30 from the Self_motion frame:
Supertype paraphrase works across frames as well; compare 31 from the Motion frame, which Self_motion inherits from:
(31) Paraphrase: I enjoyed watching the giraffes go by.
The simplest type of interchange of part-of-speech of targets does not require any change in the embedding context. This is most often the case when a target combines with a support to "simulate" a different part-of-speech. For example, a noun + support verb can often paraphrase a verb target (example from the Bragging frame):
In 32 and 33, boast.v and Support Verb + boast.n equivalently fill the finite-verb slot of the sentence.
They may also equivalently fill exactly the same valence pattern of a governing verb, as seen in 34 and 35. In this case, both noun and verb fill the VP-to realization of the Goal FE of the verb try.v in the Attempt frame:
A noun + support preposition can paraphrase an adjective or adverb target:
Other kinds of interchange of noun, adjective, verb, or preposition require changing the Phrase Type of the phrase that they occur in. This is only possible when the governing context has two different valence patterns with different phrase types for the FE filled by the target (see 188.8.131.52):
Here, the paraphrasing of succeed.v with success.n is only possible because the Event FE of the target want.v in the Desiring frame can be realized as either an infinitival complement or an NP object.
One strategy of paraphrase is to exchange the valence pattern for the realization of FEs for a target; a typical case is the Dative Alternation:
In this example (from the Giving frame), the Recipient FE is filled either with an NP or a PP(to). Since the alternation of NP and PP merely involves the presence or absence of the marker (i.e. to), it is easily possible to paraphrase these mechanically.
In principle, for most possible pairings of PTs, there are cases when a particular FE will vary between the two PTs across a frame. All such cases provide potential paraphrases across phrase types.
For example, the Phrase Type "PP-ing(about)" in the sentence "She thought about returning" can be paraphrased in the sentence "She considered returning" with the PT "VPing", or in the sentence "She considered a return" with the PT "NP", or in "Her contemplation of a return" as "PP(of)", etc. These interchanges can be selected from the phrase types listed for a particular FE realization in the valence tables of the lexical entries. Thus, this type of paraphrase generation is relatively independent of the target choice (discussed in 184.108.40.206), limited only after the fact by what phrase types are available for the chosen target.
Many of the interchanges cannot be achieved by simply varying a marker, e.g. changing an NP to a VP or vice-versa. This more complicated type of interchange could only be accomplished by paraphrasing the targets inside the phrase, e.g. by paraphrasing the head noun of the NP with a verb from the same frame to make a VP.
The straightforward interchanges of PT can be produced mechanically by adding, subtracting, or changing a marking word (like a preposition) or changing the morphology of the head of the phrase (e.g. from present participle to infinitive). The paraphrasability of the unparenthesized PTs below in 41 is merely a matter of changing the morphological marking on the head and/or changing the syntactic markers (such as prepositions, quotation marks, or the infinitive-marker to). For many other patterns (indicated in parenthesis below), there are very common constructions (such as copularization of adjectives) which allow an interchange of PTs.
Many FEs are optionally syntactically expressed, and many (less informative) paraphrases which omit them can be easily generated.
All peripheral FEs and non-target incorporated FEs are freely omissible, thus (omitting the Time FE of the Giving frame):
In addition, FEs that ever occur with an LU as INI should be considered omissible.43
FEs that occur with an LU as DNI are omissible under essentially the same conditions as would license an anaphoric pronoun, deictic adverb, or the like for the same FE.
FEs that occur with CNI omission are omissible only in certain constructional contexts. The FrameNet data currently do very little to explicate what these contexts are, although they are generally well known syntactic phenomena.44
We intend that the Frame Relations, Frame Element mappings, and Lexical Units provide a significant amount of information for doing inferencing on natural language. We do not provide every kind of information that an inferencing engine might require (especially any kind of quantitative information), but the frame-to-frame relations should provide a good basis for further specifications (using X-Schemas, predicate logic, etc.) which are useful to existing systems, especially in combination with information from WordNet. To this end, an OWL version of these relations is included in the data release.
This section describes some of the kinds of inferences that are intended by the definitions we give to the various relations.45
The most basic summarization of the logic of FrameNet is that Frames describe classes of situations, the semantics of LUs are subclasses of the Frames, and non-Extrathematic FEs are classes that are arguments of the Frame classes.46 An annotation set for a sentence generally describes an instance of the subclass associated with an LU as well as instances of each of its associated FE classes.
Inheritance has a special place among the frame relations. It is to be interpreted as a subtype or "is-a" relation between frame classes. This constrains the Child frame to have all of the properties of the Parent frame.
Although it is difficult to generalize across all of the relations, it is usually true that an instance of a Sub frame implies the existence of an instance of a Super frame. In the case of a Theft, which inherits from Committing_crime, an instance of Theft is obviously an instance of Committing_crime. For the Precedes relation, however, this is not exactly true .47 For the Subframe relation, the existence of a Sub frame instance does imply the relevance of a Complex frame instance, but the reverse is not true. So, for example, if there is an instance of the Sentencing frame, then we can instantiate an instance of the Criminal_process frame. We can then also instantiate all of the subframes of Criminal_process which must precede Sentencing, including Trial, Arraignment, etc., but we cannot instatiate the Appeal subframe of Criminal_process which follows Sentencing.
The relations of FEs and LUs to frames are entertwined with our Frame relations in various ways. Some clarification and formalization of these concepts is necessary to simplify their logical description:
(44) I rode to school [with her Cotheme] all the time.
Not surprisingly, Inheritance, as our most formally defined relation, is an exact match for an ontologically defined relation, namely subtyping or "is-a", as mentioned above. Because these relations are so well-defined formally, there is very little else that needs to be said about Inheritance here.
As discussed above, the Inheritance relation implies the correspondence of the Parent frame and its FEs to the Child frame and its FEs. This straightforwardly allows us to propagate the ontological semantic types (see Sect 6.2.1) on the Parent frame and its FEs downward to the Child frame and its FEs. However, what may not yet be apparent is the fact that the FE bindings in all frame-to-frame relations are equivalent to the Inheritance FE bindings, regardless of the type of relation. This means that normal semantic types on FEs can be propagated down the hierarchy along every type of relation. Types are propagated explicitly in the current OWL representation of the FrameNet data.
Because we cannot anticipate all of the Semantic Types that will be useful for tagging FEs, it will certainly also be desirable to categorize the fillers of our FEs using WordNet (or a similar resource). Just as with the pre-marked semantic types applied by the FrameNet team, whatever information is gathered on fillers of an FE in one frame can be propagated to all FEs which are connected to it by any frame-to-frame relation.
As shown by [Mohit and Narayanan, 2003], the frame hierarchy should help in the training of semantic parsers to label sentences with FrameNet frames and FEs. The basic principle is that each FE in an annotated sentence should give some evidence for all the connected FEs, both in terms of the types of fillers expected, and in terms of the expected syntactic realizations. Which of these two types of information for a connected frame can be gleaned from a given annotation set actually depends on which type of frame-to-frame relation we are faced with.
Information on fillers can in principle be gleaned from every type of frame relation (cf. section 220.127.116.11 above), and should be useful for identifying FEs of a target regardless of its part of speech. This means that the annotation of the Theme FE from Taking.seizure.n should help the annotation of the Goods FE with Theft.steal.v.
Since Core FEs may have an idiosyncratic syntactic realization (see the Giving frame), the only way to have absolutely solid information about the realization of the FEs of an LU is to actually consult the patterns seen in the annotation data. However, in practice there is usually a fair amount of overlap in the way that semantically related predicates (especially those of the same syntactic category) realize their arguments. All of this indicates that a statistical parser may reasonably rely on the realizations of LUs semantically similar to the LU of interest as one input to its statistical decision.