FrameNet, Frame Structure, and the Syntax-Semantics Interface

Abstract

The Berkeley FrameNet project aims at implementing Fillmore’s Frame Semantics program on a broad empirical basis. The syntactic environments of words in corpora are systematically aligned with the semantic frames evoked by the words. It is Fillmore’s vision that such a collection of valency data can pave the way for an empirically grounded theory of the syntax-semantics interface. In this article, we examine to what extent this goal can be achieved by the FrameNet approach in its present form. We take a close look at verbs of cutting and separation and at the representation of events and results in the latest FrameNet version. Our investigation reveals a certain lack of systematicity in the definition of frames and frame relations, which may hinder the derivation of linking generalizations. This situation seems to be partly due to the expectation that a system of frames can be developed on a data-driven, purely bottom-up account. As a possible solution, we argue for a richer frame representation which systematically takes into account the inner structure of an event and thereby inherently captures structural relations between frames.

1 Introduction

In the context of lexical semantics, the notion of frame is mostly associated with the research program known as Frame Semantics initiated by Charles Fillmore (1982). Its basic idea is that words, in each of their senses, are linked to frames, which are schematic cognitive structures that represent a speaker’s knowledge of the described situation or state of affairs. An important part of the program is to study in detail how the components of the frames are expressed in the morphosyntactic environment of the word. To this end, the relationships between word senses and lexico-syntactic patterns are systematically recorded, and each of these patterns has its components indexed with specific parts of the associated frame (Fillmore and Atkins 1992).

Frame Semantics has been put into practice in the Berkeley FrameNet project (Fillmore et al. 2003; Fillmore and Baker 2010). The main goal of FrameNet is to systematically collect syntactic and semantic valency patterns based on extensive corpus annotation. In this respect, the focus of the project is first and foremost descriptive. But the project has also an important “inventive” part in providing appropriate frames for the annotation. While there are a number of guidelines for introducing frames based on identifying groups of semantically related words (Ruppenhofer et al. 2010, Chap. 2), there is no explicated framework of semantic analysis for the specification of more abstract frames and frame-to-frame relations. We will see below that many inconsistencies in FrameNet’s frame structure may be attributed to this lack of a general top-down strategy.

Fillmore (2007) characterizes the FrameNet project as an empirical investigation into the interplay between lexical semantics and morphosyntactic realization. As with any scientific investigation, the point is not only to sample as much data as possible but also to look for regularities and generalizations, that is, for a theory that explains the data. The theory in question is a theory about argument linking and the syntax-semantics interface of lexico-syntactic constructions. In fact, Fillmore has the vision that Frame Semantics as implemented in FrameNet can provide a basis for deriving linking generalizations. The underlying assumption is that FrameNet will ultimately be equipped with an elaborate frame hierarchy and that linking generalizations can then be formulated in terms of abstract frames for action, change, causation, and the like.

We will critically examine whether Fillmore’s vision is supported by the present architecture of FrameNet.¹A crucial problem turns out to be the relative lack of systematicity in semantic analysis. We provide evidence for this claim by revealing various deficiencies in the frame representation of causation and inchoation and of events and changes in general, and in the representation of verbs of cutting and separation. We sketch how a more explicit decompositional approach to frame semantics could potentially overcome many of these problems. In particular, we argue that the internal structure of an event or state of affairs should be reflected within the frame representation itself. We will also point out that FrameNet’s current restriction to associating frames with lexical items may run into difficulties even for elementary constructional variations.

All FrameNet data discussed in the following are taken from FrameNet Version 1.5 of September 21, 2010, which includes more than 1,000 frames, almost 12,000 word senses (lexical units), and over 8,000 instances of frame-to-frame relations. We speak of ‘FN 1.5 ’, if we refer to this specific release, and of ‘FrameNet’, if we mean the project in general. A word of caution: We are fully aware that an intermediary release of an ongoing large-scale project like FrameNet inevitably shows gaps and inconsistencies.²We therefore try to distinguish between temporary issues that can easily be remedied during the next revision cycle and potential inherent problems of the approach itself. Despite its shortcomings, we regard the FrameNet project as an important contribution to developing an empirically grounded theory of lexical semantics and linking. And, of course, the present study would not have been possible without the generous policy of the FrameNet group to make their data publicly available.

2 Frame Semantics and FrameNet

2.1 From Case Frames to Frame Semantics

Fillmore’s (1968) early ‘case grammar’ approach builds on the concept of case frames, whose underlying idea was to characterize the valency of verbs by a small set of semantic roles from which the syntactic valency can be predicted by general rules. In later writings, Fillmore conceded that this approach falls short in several respects (cf. Fillmore 2003). For one thing, it turned out to be difficult, if not impossible, to come up with a stable inventory of semantic case roles together with reliable criteria for assigning them to the participants of the situation described by a verb or any other valency bearing lexical item. Another issue is the unique role assignment in cases of complex event descriptions like (1), where the ballis both the Goal of the ‘hitting’ component and the Theme of the ‘go-over’ component.

• (1) Peter knocked the ball over the fence.

As a consequence, the notion of frame laid out in his Frame Semanticsapproach Fillmore (1982) aims at a much broader and deeper conceptual-semantic description of lexical items than provided by course-grained semantic role frames. Framesin this context are to be understood as “schematic representation[s] of speakers’ knowledge of the situations or states of affairs that underlie the meanings of lexical items” (Fillmore 2007, p. 130) or as “situation types for which the language has provided special expressive means” (Fillmore and Baker 2010, p. 321).³The program of Frame Semantics has been elaborated in a series of detailed case studies on the meaning and use of individual verbs and narrow verb classes (cf., e.g., Fillmore and Atkins 1992). Although these studies emphasized the correlation of particular lexical meanings with specific morphosyntactic patterns, they did not aim at an explanatory account of the relation between (morpho-)syntax and semantics. In particular, no attempt has been made to derive generalizations on this interaction that go beyond the discussion of selected examples.

2.2 The FrameNet Project

The FrameNet project, which started in the late 1990s, is intended as a broad coverage, corpus-based implementation of the Frame Semantics program (Fillmore et al. 2003; Fillmore and Baker 2010). According to Fillmore (2007, p. 129), the project’s main focus is on valency descriptions in syntactic and semantic terms, that is, on exploring the correlation between lexical semantics and syntactic argument realization:

The FrameNet project is dedicated to producing valency descriptions of frame-bearing lexical units(LUs), in both semantic and syntactic terms, and it bases this work on attestations of word usage taken from a very large digital corpus. The semantic descriptors of each valency pattern are taken from frame-specific semantic role names (called frame elements), and the syntactic terms are taken from a restricted set of grammatical function names and a detailed set of phrase types.

Frames in the context of FrameNet are thus plain semantic role frames, similar to the earlier case frames, except that there is no universal role inventory anymore since the semantic roles are specific with respect to the frame they belong to.

2.2.1 The Specification of Frames in FrameNet

The description of a frame in the FrameNet database includes the following components: the name of the frame; an informal definition of the situation the frame is supposed to represent; the set of semantic roles (frame elements) associated with the frame, subdivided into coreand non-coreelements; and the corresponding word senses (lexical units) that evoke the frame. Core roles are those which are “necessary to the central meaning of the frame” (Fillmore 2007, p. 133). Non-core roles are subdivided into peripheraland extrathematicelements. Peripheral roles mark notions such as time, place, manner, means, and the like. They are not unique to a frame but can modify any frame of the appropriate type. Extrathematic roles are used to annotate a “word or phrase which can be thought of as introducing a new frame, rather filling out the details of the frame evoked by the head” (Fillmore 2007, ibid.). An example of an extrathematic role is Depictive, which is used to mark depictive secondary predicates. Furthermore, FrameNet allows the characterization of “role fillers” by semantic types.

Table 6.1 shows part of the specification of the Cutting frame, which is evoked by (appropriate senses of) the verbs cut, chop, etc. The element Result is the only extrathematic role among the non-core elements. Some of the frame elements carry semantic types, which are not shown in the Figure. For example, Agent and Instrument are typed as Sentient and Physical_entity, respectively.

Table 6.1 Partial description of the Cutting frame in FN 1.5

2.2.2 Annotation Scheme

One of the goals of FrameNet is to provide all lexical units of a frame with a representative set of corpus-based example sentences, which are annotated both syntactically and semantically. The semantic annotation consists basically in the assignment of frame elements to constituents of the example sentences. The syntactic annotation comprises a phrasal and a functional level. The phrasal level employs fairly standard phrase type markers (NP, PP, AVP, etc.; cf. Atkins et al. 2003), while the functional level uses grammatical functions such as External, Object, and Dependent.

The sentences in (2) are corpus examples from FN 1.5 that are associated with the lexical unit sliceof the frame Cutting.

• (2)

  • a. Slice the cake lengthwise into two halves, to give two long, thin cakes.

  • b. Slice the onions fairly fine.

  • c. The frozen brain is […] sliced sequentially into very thin sections.

The annotation report for sentence (2a) is shown in (3). CNI stands for ‘constructional null instantiation’, which indicates that the core role Agent is omitted on syntactic grounds, here, because of the imperative.⁴

• (3)

  Table 2

2.2.3 Frame-to-Frame Relations

FrameNet frames are not isolated units but are related to each other in various ways. FrameNet employs eight different types of relations between frames, of which seven are relevant to the present study. They fall into three groups (Fillmore and Baker 2010): generalization relations (inherits from, is perspective on, uses), event structure relations (is subframe of, precedes), and “systematic” relations (is causative of, is inchoative of).⁵

Inheritance is the strongest relation between frames. If a frame (the childframe) inherits from another frame (the parentframe) then all frame elements (semantic roles) of the parent frame occur as frame elements of the child frame, possibly under a different name. The possible difference in naming reflects the fact that FrameNet assumes frame-specific semantic roles. In the case of inheritance, the semantic type of the child is a subtype of the parent, and the same condition holds for all role filler types. For example, the frame Cutting inherits from the frame Intentionally_affect, with the Item element of Cutting bound to the Patient element of Intentionally_affect and all other element names kept equal. And the frame Commerce_buy inherits from Getting, with the roles Buyer, Goods, and Seller bound to Recipient, Theme, and Source, respectively. Figure 6.1shows a simplified representation of these examples, with only some of the core roles listed. Frames are depicted as attribute-value matrices and co-indexing indicates the correspondence between roles.

The ‘is perspective on’ relation implements the ideas of profiling and figure/ground distinction.⁶A widely discussed example in the FrameNet literature is the commercial transaction scenario, where buying and selling are regarded as taking different perspectives on the transfer of goods, while paying and collecting take different perspectives on the transfer of money⁷; see Fig. 6.2. Frames that have perspectives are often abstract or non-lexicalin that they are not directly evoked by lexical units but via one of their perspective-taking frames.

Fig. 6.1

Simplified examples of frame-to-frame relations in FN 1.5 

Fig. 6.2

Commercial transaction in FrameNet (shaded nodesindicate non-lexical frames)

The ‘uses’ relation is defined somewhat vaguely. According to Ruppenhofer et al. (2010, p. 78), this relation “is used almost exclusively for cases in which a part of the scene evoked by the Child refers to the Parent frame”, Fillmore and Petruck (2003, p. 361) describe it as a “relation like Inheritance, but less strictly defined”, and Fillmore and Baker (2010, p. 330) posit that the child frame “depends upon background knowledge provided by the parent frame” and that some but not all of the core roles of the parent must correspond to roles of the child frame. For instance, the frames Bringing and Removing use the frame Motion with the Agent role of Bringing and Removing not bound to any role of Motion, while the roles Theme, Goal, Path, and several others have direct correspondents (cf. Fig. 6.1).

The ‘is subframe of’ relation holds between a pair of frames if the first frame represents a subevent of the (complex) event represented by the second frame. Figure 6.2shows two instances of subframe structures. The frame Commercial_transaction has the subframes Commerce_goods_transfer and Commerce_money_transfer. Apparently, there is no clear temporal precedence between the events denoted by these two subframes. This is different for the subframes of the Giving_scenario frame. They represent a temporal succession of states and events, which is indicated by the ‘precedes’ relation.⁸Note that it remains unclear how exactly “subevent” in the subframe definition differs from “part of the scene” in the ‘uses’ definition.

The ‘is inchoative of’ and ‘is causative of’ relations, finally, hold between pairs of frames of which the first denotes respectively the inchoative and the causative of the event denoted by the second frame. For instance, the frame Cause_expansion is causative of the frame Expansion, and Becoming_detached is inchoative of Being_detached. The two relations will be discussed in more detail in Sect. 6.3.2.⁹

2.3 Frame Semantics and Linking Generalizations

Fillmore’s early case frame approach was a proposal to explain the morphosyntactic realization of a verb’s arguments in terms of general semantic characteristics of the verb. Because of the problems mentioned in Sect. 6.2.1, Fillmore abandoned the idea of universal case roles in favor of his Frame Semantics program with its richer notion of lexical meaning and its dedication to a broad coverage of realization patterns. Exploring the regularities of the interaction between lexical semantics and morphosyntax nevertheless continues to be an essential goal of the program, and thus also one of the FrameNet project. Accordingly, Fillmore (2007, p. 157) assumes that FrameNet can give rise to linking generalizations if abstract frames and frame-to-frame relations are appropriately defined:

The structure of the system of frame-to-frame relations is set up, but the details have not been completed as of this writing. Many FrameNet frames are elaborations of more abstract schemas of change, action, movement, experience, causation, etc., and the roles found in these are the ones that figure in linking generalizations;many of the more refined frames can be seen as perspectives on the more abstract frames, in the way that buying is a subtype of getting, selling and paying are kinds of giving, etc. Generalizations based on inferences about who possesses what before and after the transaction depend on the roles in the commercial transaction; generalizations about how syntactic roles are assigned to the arguments depend on the more abstract inherited schemas.[emphasis added]¹⁰

Fillmore and Baker (2010, pp. 332f) illustrate this idea with an example from the commercial transaction scenario shown in Fig. 6.2. They argue that the verbs buyand collectare used with the preposition frombecause “buying and collecting are getting,” and that selland payallow the dative alternation because “selling and paying are giving.” That is, the fact that Commerce_buy and Commerce_collect inherit from Getting is employed to explain the respective realizations of the Seller and the Buyer by a from-PP. A glance at Commerce_collect in FN 1.5 shows some of the complications this proposal is faced with in practice. The Commerce_collect frame has the units bill, charge, and collect(of which only chargeis annotated). Neither billnor chargeallow the “getting-from” pattern – simply because none of them is a getting verb. Billing and charging have to do with communicating the demand for money to the buyer. Hence, billand chargeare more like verbs of giving and, in fact, chargeallows the dative alternation. What we have encountered here is a defect in FN 1.5 but not necessarily one in FrameNet. The problem can be remedied by removing billand chargefrom Commerce_collect, thereby keeping the inheritance from Getting. Now, what to do with billand charge? Clearly, these verbs should evoke some frame in the commercial transaction domain. But the frame system shown in Fig. 6.2is not subtle enough to cover the respective component of a commercial transaction, which could roughly be described as “assigning debt of payment.” This discussion illustrates that while frame inheritance might indeed give rise to linking generalizations, the approach is very sensitive to the adequate specification of frames and frame-to-frame relations.

3 The Representation of Events and States

3.1 Events and Changes

According to Fillmore and Baker (2010, p. 331), FrameNet adopts, for the most part, a “simple three-state model of event structure, with a pre-state, a central change, and a post-state,” of which “typically, only the central, changing part of the event is profiled, and, hence lexicalized.” We have seen this structure in Fig. 6.2instantiated by the Giving_scenario frame and its subframes. Its abstract correlate is shown in Fig. 6.3. The precise interpretation of this frame-to-frame structure in terms of the represented situation types raises a number of questions. Does the structure in Fig. 6.3imply that each situation of type Event is necessarily part of a situation of type Change_of_state_scenario? If yes, then this property should somehow be reflected in the definition of the Event frame. According to the FN 1.5 definition shown in Table 6.2, nothing seems to be required of situations of type Event except that something happens; that is, situations of this type are dynamic, not static. Correspondingly, the associated verb units are general verbs of occurrence such as happenand take place.

Fig. 6.3

FN 1.5 top level event structure (shaded nodesindicate non-lexical frames)

There are basically two ways of conceiving the Event frame in its relation to change of state scenarios. On one interpretation, the frame covers all kinds of dynamic situations, including processes and activities like motion in place (rotate, vibrate) and directed perception (watch, listen), which are usually not regarded as changes of state in the proper sense. On a second, more restrictive interpretation, any situation of type Event involves a conceptually salientchange, typically manifested on one of the participants.¹¹ The two options should give rise to different sets of frames that inherit from Event. The direct inheritance daughters of Event in FN 1.5 are shown in (4).

• (4) Becoming, Birth, Ceasing_to_be, Change_of_consistency, Coming_to_believe

  Eventive_affecting, Experience_bodily_harm, Getting, Go_into_shape, Intentionally_act

  Misdeed, Process_end, Process_pause, Process_resume, Process_start, Process_stop

  Rotting, Transitive_action, Waking_up.¹²

Table 6.2 Definitions of general situation frames in FN 1.5 , with associated verb and noun units

The first thing to notice about this list is its heterogeneity. We can ascribe this to the fact that FrameNet is still under construction, and presumably also to the lack of an elaborate and precise guideline on how to organize the frame inheritance hierarchy. As to the question of whether Event subsumes all sorts of events or only changes of state, the frames listed in (4) give a mixed impression. The wide interpretation of Event is supported by the fact that Intentionally_act is inherited by Perception_active (watch, listen) and Practice (practice, rehearse), both of which do not count as proper change of state frames in the sense discussed above. On the other hand, the frame Process_continue (go on, proceed) is not included in the list (nor subsumed by one of its members) and neither is Moving_in_place (rotate, vibrate). Moreover, there are obvious cases of change of state frames such as Undergo_change (change, turn) and Change_of_phase (freeze, melt) which do not inherit from Event. In sum, FN 1.5 provides no coherent picture about how dynamic situations are to be distinguished with respect to their change-of-state character. It is also indicative that the frame Process and its inheritance daughter Activity are not systematically employed for structuring the inheritance hierarchy. And the same is true of the frame Becoming, which is inherited by a single frame only, namely, Absorb_heat.¹³

The foregoing shortcomings of the FN 1.5 frame hierarchy may be remediable to some extent by a thorough revision and extension of the frame-to-frame inheritance relations.¹⁴For instance, if the Event frame is to be conceived of as covering all kinds of dynamic situations, then it seems reasonable to require that every frame representing a dynamic situation type inherits (directly or indirectly) from Event. In this case, the above observations that some dynamic situations are not subsumed by Event would simply be due to missing inheritance relations which remain to be added to the database. The frames Process, Activity, Change_of_state_scenario, etc. would then all inherit from Event. But the question as to whether there is a distinction between event frames that inherently encode a goal or result and those which do not still needs to be answered. Put differently, is the telic/atelic distinction, or a more refined Aktionsart classification, part of the representation of situation types in FrameNet? Whatever the answer is, there should be clear criteria for deciding whether a given situation type is subsumed by, say, Process, Change_of_state_scenario, or none of the two. In fact, there are good reasons to have Aktionsart distinctions represented in FrameNet because of their relevance both for natural language reasoning (Im and Pustejovsky 2010) and for formulating generalizations about the syntax-semantics interface (Van Valin 2005).¹⁵

3.2 Causatives and Inchoatives

The names of about 30 frames in FN 1.5 start with ‘ Cause’. They range from general frames like Cause_change (with lexical units alter, change, convert, modify, turn, etc.) to fairly specific ones like Cause_to_wake and Cause_to_be_dry. As to be expected, many of these “cause” frames are causatives of other frames; e.g., Cause_change is causative of Undergo_change (change, shift, turn, etc.). Further causative pairs are listed in Table 6.3. The verbs in the right column occur in both frames, that is, they undergo the causative-inchoative alternation. A number of the “Cause frames” are not causatives of other frames, even if there are clear candidates available. For example, there is no ‘is causative of’ relation between the frames Cause_to_wake and Waking_up. Also, one would expect that the frames Cause_change and Undergo_change are somehow related to – if not inherited by – more specific frames such as Cause_change_of_phase and Change_of_phase, respectively, which is not the case. As to inchoatives, there are transparent chains from causative to inchoative to result state as illustrated by the following examples:

Table 6.3 Some causative-inchoative pairs in FN 1.5 , together with alternating verb units

• (5)

  • a. Attaching/ Inchoative_attaching/ Being_attached

  • b. Cause_temperature_change/ Inchoative_change_of_temperature/ Temperature

  • c. Cause_change_position_on_a_scale/ Change_position_on_a_scale/ Position_on_a_scale

But it also happens that the inchoative frame is missing in the chain: Cause_to_be_dry is causative of Being_dry and the frame Becoming_dry is not related to either of the them (cf. Table 6.4and the discussion in Sect. 6.3.3below).

As in the case of event inheritance discussed in the previous section, one might argue that the foregoing deficiencies could be easily remedied without changing the present set-up of the FrameNet project by successively adding missing frames and frame-to-frame relations and by eliminating inconsistencies. The point we want to make is that this task could strongly profit from a systematic decompositional analysis of event frames. Notice that decompositional structure is already captured in FN 1.5 , to some extent, by the naming of the frames. Several names of the causative frames listed in Table 6.3are regularly built from the names of their inchoative correlates by adding the prefix ‘Cause’. However, there is obviously no general convention for naming frames – compare, e.g., Inchoative_attaching vs.  Becoming_detached. While naming is, in principle, irrelevant to the content of FrameNet, a consistent naming convention would help to avoid many of the deficiencies of the ‘is causative of’ and ‘is inchoative of’ frame-to-frame relations mentioned above. Whether one of these relations holds between two frames could then be deduced from the names of the frames. Also, cases like Becoming_dry being not related to Being_dry would not occur. One of the arguments in favor of a decompositional frame analysis is thus the inherent support for organizing the frame hierarchy in a systematic way. A second argument, to be sketched in Sect. 6.3.3below, is the role the decompositional structure can play in deriving linking generalizations.

Fig. 6.4

FrameNet frames representing changes of phase

Fig. 6.5

Different FN 1.5 representations of the causation/inchoation/state pattern

A more intricate issue for the systematic treatment of causation and inchoation turns up with the Change_of_phase frame and its associates. As shown in Fig. 6.4, the causative frame is inchoative of the frame Altered_phase (frozen, liquefied, melted, etc.), which is preceded by the frame Change_of_phase, and there is an additional “scenario frame” with Change_of_phase and Altered_phase as subframes. Figure 6.5a depicts the general structure behind Fig. 6.4, namely the change of an object with respect to an attribute A – here, the object’s phase or state of matter. By comparison, the corresponding case of the attribute temperature (5b) follows basically the pattern shown in Fig. 6.5b. The state characterizations in (a) and (b) differ substantially: States of affairs of type ‘A’ are about an entity having its attribute Aspecified to a certain value. For instance, the Temperature frame has the adjective units cold, hot, warm, etc., which can be used to characterize the temperature of an entity, either in pre-nominal position or in a copula construction. States of affairs of type ‘Altered A’, by contrast, depend on a previous alteration event. If applied to temperature, the respective “Altered temperature” frame would presumably cover the participles cooled, heated, and warmed. The frame structure shown in Figs. 6.4and 6.5a conceives of the resulting state as being a part of a change of state scenario in much the same way as outlined by the general event model in Fig. 6.3. The question, then, is whether all change of state frames are sought to be systematically paired with corresponding result state frames and, even more crucially, how these result frames are related to their event-independent counterparts, if existent.¹⁶That is, should FrameNet provide different state frames for driedand dry, warmedand warm, or liquefiedand liquid, respectively?¹⁷In the following section, we will argue that these issues can be resolved, up to a point, by representing the components of an event type within the frame itself.

Table 6.4 Definitions of FN 1.5 frames of drying, with associated verb and adjective units

3.3 Toward a Decompositional Frame Semantics

Many current theories of the syntax-semantics interface rely on semantic decomposition.¹⁸Almost all of these proposals involve relatively shallow, term-based decompositions. Two notational variants of such a decomposition of causative dryare shown in (6), formulated along the lines of Van Valin and LaPolla (1997) and Rappaport Hovav and Levin (1998), respectively.

• (6)

  • a. $[\mathbf{do}(x, \varnothing )]\ \mbox{ CAUSE}\ [\mbox{ BECOME}\ \mathbf{dry}(y)]$

  • b. $[x\ \mbox{ ACT}]\ \mbox{ CAUSE}\ [\mbox{ BECOME}\ [y\ \langle \mathit{DRY }\rangle ]]$

Decompositional representations of this type try to make explicit the semantic factors that play a role in argument realization, including causal and aspectual factors as well as notions such as volitionality and the like. Many theories of linking are based on semantic decompositions of this sort. For example, the linking theory of Van Valin and LaPolla (1997) regards specific and generalized semantic roles as derived notions that can be defined on the basis of the structural positions in the decompositional representations. In this approach, the generalized roles Actor and Undergoer play a key role in explaining the linking of arguments to their morphosyntactic realization.

Fig. 6.6

Frame-based event structure decomposition of causative and inchoative dry

As discussed in the previous section, a naming convention for frames is a simple but limited method of capturing the sub-eventual structure of a situation. A more attractive and more sophisticated option is to represent the structure within the frame itself. Doing this means to move from the plain role frames used in FrameNet to more complex frame structures which allow the embedding of one frame in another. Term-based decompositional representations of the kind shown in (6) offer a first blueprint of how to proceed. For example, the primitive one-place state predicate drycan be reinterpreted as a basic state frame Dry_state with a role Patient, whose filler corresponds to the argument of the predicate; see the attribute-value matrix on the right of Fig. 6.6. The BECOME dry(y) term can be taken as an Inchoation frame whose attribute Result is filled by the Dry_state frame; see the matrix in the middle of the figure. Finally, the overall CAUSE term of (6) corresponds to a frame Causation with attributes Cause and Effect, which are filled with the Activity and Inchoation subframes, respectively, as depicted on the left of Fig. 6.6. The resulting decompositional frames are related to each other by virtue of their inner structure. Compact frame names such as ‘ Cause_to_become_dry’ can now be introduced as abbreviations for complex frame structures, or, on a more formal basis, as elements of a type hierarchy associated with complex attribute-value declarations.¹⁹The frame-specific roles of FrameNet, which have their purpose, e.g., in easing the annotation task, can be reintroduced as shortcuts for attribute paths of the decompositional frame structures. For example, the role Dryee can be defined as the concatenated path Effect|Result|Patient of the Cause_to_become_dry frame.²⁰

The transformation of standard decompositional schemas into frame structures is only a first step in developing an elaborate frame semantics. Term-based decompositions like the ones in (6) are usually rather limited in representing information about, e.g., attributes of participants or other implicit components of a situation. Frames in the general sense of Barsalou (1992), or of Fillmore (1982) for that matter, are more expressive in this respect. As an example, consider again the representations for drygiven in Fig. 6.6and the respective FN 1.5 frames listed in Table 6.4. None of the frame structures in Fig. 6.6captures the actual process the “patient” undergoes during drying. Unlike pure change of state verbs such as break, which are non-specific about how the resulting state is achieved, deadjectival degree achievementverbs (Dowty 1979) like dry, darken, and coolallow an atelic interpretation which profiles the change process without entailing the result denoted by the base adjective. The different readings can be triggered by temporal adverbials: dry for an hourmeans to get drier and drier within that hour without necessarily being dry afterwards, dry in an hourimplies a resulting dryness.

The decompositional Becoming_dry frame of Fig. 6.6is a plain change of state frame whose result state is dryness. Similarly, the definition of the corresponding FN 1.5 frame shown in Table 6.4entails the resulting dryness of the undergoer. Both frames do not cover the situation where an entity gets drier without becoming dry.²¹As to the process of drying, the definition in Table 6.4mentions the loss of moisture, which, however, has no consequences for the semantic content of the FrameNet frame. Decrease of moisture, or increase of dryness for that matter, is in fact the core characteristics of the process of drying. The general pattern behind this type of process is that an entity undergoes a change with respect to the value of one of its attributes, here, the degree of moisture or dryness. More precisely, the change in question is either an increase or a decrease of that value on an ordered scale.²²An appropriate semantic representation of this situation type should therefore take into account the respective attribute of the undergoer and the ordering of its possible values. Figure 6.7indicates how this can be achieved within a frame-based account. The basic component of these frames is the instantaneous “moisture stage” of an entity, which can be conceived of as a snap-shot of the entity having a certain degree of moisture.²³Stages provide a natural way to model changes of attribute values. The atelic interpretation of drycan be analyzed as a process whose progression (Prog) consists of an ongoing incremental change between two stages of the undergoer of which the second has a lesser degree of moisture than the first. The corresponding frame structure on the left of Fig. 6.7employs a relational constraint Lesserbetween moisture values. The telic, inchoative interpretation depicted in the middle of the Figure resembles the inchoative frame of Fig. 6.6.²⁴The stative interpretation of dryis then represented by the situation type in which the stages of the entity in question are constant with respect to their degree of moisture, which is ‘zero’ in this case; see the frame on the right of Fig. 6.7.

Fig. 6.7

Representation of (non-causative) atelic, telic, and stative dry

At first glance, it might appear that the proposal described here leads to a multiplication of frames. This is true insofar as we think it important to keep track of the differences between processes and accomplishments in the frame representations of situation types. This distinction is relevant to the understanding of the syntax-semantics interface as well as for drawing correct inferences. If something is not necessarily dry after having dried for an hour then an automatic reasoning system should not come to a different conclusion. On the other hand, it is important to notice that the frame structures in Fig. 6.7are not tailored for specific verbs but represent fairly general situation schemata. The only verb specific meaning components are the moisture attribute and its associated value scale, both of which are determined by the adjectival root of the verb. If we abstract from these components then the frame on the left of the figure is about situations where an entity undergoes a continuous change with respect to the value of one of its (scalar) attributes such that the value decreases on the scale in question. The situation type encoded by the frame in the middle of the figure differs in that the minimal value on the attribute scale is reached and the exact course of the change not specified. The event structure schemas of both frames are clearly relevant to other degree achievement verbs as well. Moreover, the two frames are systematically related to each other. Hence, if one wants to speak of frame multiplication at all, then it should be in the sense of a more thorough account of the combinatorial potential of elementary event structure components. Another objection could be that the above distinction between the process and the accomplishment interpretations of dryis not part of lexical semantics but should be modeled in terms of aspectual coercion or the like. A first reply is that at least the current FrameNet annotation would not allow such a “post processing” since inand foradverbials are both annotated as Duration. More crucial is the question of what to count as a frame evoking unit, especially from a cross-linguistic perspective. For example, the Australian aboriginal language Mparntwe Arrernte encodes the distinction between the process and the accomplishment interpretation of degree achievement verbs like coolby means of process and result affixes (Van Valin 2005, pp. 43f). If some languages encode Aktionsart operators as bound morphemes while others do not, the focus on word units as frame-bearing elements may run into serious problems when it comes to language comparison.²⁵

4 Case Study: Verbs of Cutting and Fragmentation

Verbs of cutting and fragmentation provide a good touchstone for lexical semantic representation and linking. They have been studied widely with respect to differences in their argument realization options and varying lexicalization patterns across languages (Guerssel et al. 1985; Mairal Usón and Faber 2002; Majid et al. 2007). Cut-verbs have also been discussed with respect to the controversial thesis of a manner/result complementarity in lexical encoding (Bohnemeyer 2007; Rappaport Hovav and Levin 2010).

Table 6.5 List of all frames in FN 1.5 with verb unit cut

The verb cutoccurs in FN 1.5 as a lexical unit in the frames listed in Table 6.5. The table also shows a number of verb units associated with these frames. In the following, we are concerned with senses of cutthat are related to affecting a physical entity by means of a sharp-edged object, resulting in incision and, possibly, separation or fragmentation. Senses of cutmeeting this constraint occur in the frames Cutting, Removing, Experience_bodily_harm, and Cause_harm, which are marked bold face in Table 6.5.

There are several more senses of cut, but missing senses at an intermediate stage of the project are an inevitable consequence of FrameNet’s frame-oriented working procedure. For as Fillmore (2007, pp. 139ff) points out, the frame-based exploration of word meanings “pays attention to paraphrase relation and postpones thorough treatment of polysemy structures.” An obvious gap among the concrete, object affecting meanings of cutis the sense “trim or reduce in length of (grass, hair, etc.) by using a sharp implement” (Oxford Dictionary of English, 2nd ed.). In fact, FN 1.5 seems to have no frame for this scenario since trimand moware not covered either. Another missing pattern is cut a hole into something.

4.1 Cutting and Other Ways of Separation

The core meaning of cutis represented by the Cutting frame introduced in Sect. 6.2.2.1(Table 6.1). Somewhat surprisingly, FN 1.5 provides no annotated examples for the lexical unit cutof Cutting.²⁶But the examples for slicegiven in (2) and those for chopin (7) do as well.

• (7)

  • a. Chop[the onions]_Item[finely]_Manner.

  • b. [The animals]_Itemare always chopped[into small pieces]_Pieces

The definition of the frame Cutting presented in Table 6.1is repeated in (8).

• (8) An Agent cuts an Item into Pieces using an Instrument.

One might object that this definition just illustrates schematically a certain use of the verb cut. Instead of (8), one would probably expect a typical dictionary definition of cutlike “divide something into pieces with a knife or other sharp implement,” which is, in fact, the sense definition given for the lexical unit cutof the Cutting frame. However, the “definition” of a frame is only intended as an informal description that indicates the kind of scenario the frame is supposed to represent. The definition plays no formal role in the FrameNet architecture. That said, we can take (8) as a characterization of the Cutting frame. All lexical units of this frame (except cutitself) are described as specific types of cutting. Table 6.6lists these units together with their FN 1.5 sense definitions. As with frame definitions, the sense definitions of lexical units play no formal role in the FrameNet architecture; they are useful indicators for the human users and, especially, for the annotators. The sense definitions in FrameNet are either taken from the Concise Oxford Dictionary, 10th Edition, or are written by FrameNet editors.

Table 6.6 Lexical units of Cutting with FN 1.5 sense definitions

The units slicethrough chopconform to the frame definition (8). The first four of them, slice, cube, dice, and fillet, have zero-related nouns that characterize the (form of the) resulting pieces. Except for dice, this verbal incorporation of the Pieces element is explicitly marked in FN 1.5 . The case of minceis similar but specifies the size of the resulting pieces, not their form. The definition given for chopdoes not mention ‘into pieces’, but this could be a flaw of the chosen dictionary definition.²⁷By contrast, the sense definitions of carveand pareare not subsumed by (8). None of them indicates that something is cut into pieces.²⁸The sense of parethat appears in the FN 1.5 examples shown in (9) can be paraphrased as “cut off the outer skin of something”, typically used with fruits and vegetables.²⁹

• (9)

  • a. Pare[the mangoes]_Itemand cut the flesh away from the pit lengthwise into long slices.

  • b. Thinly pare[the rind from 1 orange]_Itemand cut into narrow strips.

In (9), we have actually two distinct lexico-syntactic patterns of pare. The direct object of (9a) refers to the entity whose outer skin is cut off; in (9b), the outer skin is expressed by the direct object and the corresponding entity is expressed by the from-PP. Note that the FN 1.5 annotation of (9b) mistakenly combines the object NP and the PP to a single constituent. While (9a) could be classified as a “trimming by cutting” scenario, the type of situation expressed in (9b) falls under “cutting off a part from a whole”. Neither of the situation types is covered by the Cutting frame as introduced in Table 6.1.

It is striking that neither hack, hew, saw, snip, nor slashare listed among the units of Cutting, for one easily finds corpus attestations of the pattern hack/hew/saw/snip/slash sth (in)to piecesand dictionary definitions typically describe these verbs (in one of their senses) as verbs of cutting. As mentioned before, FrameNet aims at a frame-wise processing of the lexical domain. Hence, one wonders whether these verbs are left out by accident or on purpose. The answer depends on which situation type the Cutting frame is supposed to represent. In view of (8), the situation in question is about cutting-related activities which result in the affected entity being in pieces. With cutand chopas units of the frame, there seems to be no reason not to include hack, hew, saw, etc. as well.

As mentioned above, definition (8) suggests that the affected item’s being in pieces is an essential component of the Cutting frame. More evidence for this conclusion comes from the fact that Pieces is a core element of Cutting. Recall from Sect. 6.2.2.1that core elements are “necessary to the central meaning of the frame”. Hence, no Cutting without Pieces. Situations where someone cuts into an object, or at an object, without separating it are thus not represented by Cutting. But even if the cutting leads to pieces, the situation is not necessarily subsumed by the Cutting frame. Consider situations described by phrases such as cut a branch off the tree. Here, a piece is separated from a whole without the whole going to pieces. The definition of the Cutting frame and the lack of appropriate core elements imply that the “cutting off” scenario is not within the scope of the Cutting frame (see also the discussion of the interpretation of parein (9b) above).

Fig. 6.8

Frames for cut, split, break, rip, shred, etc. and their interrelation in FN 1.5 

The given characterization of the Cutting frame raises a number of questions: Should the frame be extended to cover “cutting off” scenarios as well? What about cutting events that result in an incision but do not lead to separation? In which way is Cutting related to other frames of separation and fragmentation? Several FN 1.5 frames are explicitly concerned with fragmentation and separation scenarios. Figure 6.8gives an overview of these frames and shows how they are related to each other and to the frames Cutting and Removing, both of which have cutas a lexical unit. (Shaded nodes represent again non-lexical frames.) The definitions and lexical units of most of these frames are listed in Table 6.7(with verbs to be discussed in the following in bold-face). Besides Cutting there are two other frames, Cause_to_fragment and Breaking_apart, that have Pieces as a core element. Pieces is bound to Parts in the ‘is perspective on’ relation between Breaking_apart and Fragmentation_scenario, and Parts is also a core element of the frames Separating and Becoming_separated. The FN 1.5 examples in (10) describe situations where some entity is separated or fragmented into parts or pieces.

Table 6.7 Definitions of FN 1.5 frames related to fragmentation and separation, with selected verb units

• (10)

  • a. The young men split[the cattle]_Whole[into two groups]_PartsSeparating

  • b. Split[the cake]_Whole[in half]_Partshorizontally Separating

  • c. [French flags]$_{\mathrm{Whole\_atient}}$were ripped[in half]_PiecesCause_to_fragment

  • d. Break[the carcass]$_{\mathrm{Whole\_patient}}$[into small pieces]_PiecesCause_to_fragment

  • e. shredding[the notes]$_{\mathrm{Whole\_patient}}$[into little pieces]_Pieces. Cause_to_fragment

  • f. [thinly]_Manner shred[each half]$_{\mathrm{Whole\_patient}}$, using a sharp knife. Cause_to_fragment

  • g. [thinly]_Result shred[each half]_Undergoer, using a sharp knife. Grinding

  • h. crush[them]_Undergoer[to bits]_Result. Grinding

With the exception of (10)a, the verb contexts in (10) are compatible with cut-verbs like sliceand chop. However, generalizations about this similarity in syntagmatic patterning are not supported by the FN 1.5 frame-to-frame structure shown in Fig. 6.8. There is no general representation of the meaning component common to the described scenarios, that is, of the turning of something into pieces. Again, a decompositional analysis could help in this case. The frames Cause_to_fragment, Grinding, and Cutting are not related to each other, except by their inheritance from Transitive_action, and Separating does not inherit from any frame at all. The Grinding frame, while mentioning “into smaller pieces” in its definition, has no Pieces element but uses the (non-core) Result element for this purpose; cf. example (10)h.

The verb shreddeserves special attention because it is a unit of the frames Grinding and Cause_to_fragment and it is defined as ‘tear or cut into shreds’ in both cases. There are even a number of sentences annotated by both frames, witness (10)f and (10)g. Note that thinlyis annotated as Manner in (10)f and as Result in (10)g. The latter analysis seems adequate if the expression is taken to mean that it is the resulting shreds which are thin. However, FN 1.5 defines the Result element of Grinding as “the shape the Undergoer becomes” and this definition is difficult to reconcile with the foregoing interpretation of (10)g since it is not the Undergoer that becomes thin. This phenomenon is related to the spurious resultativesdiscussed by Washio (1997, pp. 17f), for which cut the meat thick/thicklyand tie the laces tight/tightlyare typical examples.³⁰The crucial property of this pattern is that the “result” predicates are not applied to the entities denoted by the direct objects but to entities that are not overtly expressed, the slices in cut the meatand the knot in tie the laces. The lexical semantic constraints that underlie this construction clearly call for a frame semantic analysis that goes beyond plain role frames.

A further issue arises with the use of particles. The adverb apartoccurs frequently with verbs of separation and fragmentation. The FN 1.5 examples in (11) contain annotated instances of apartused as a particle with the verbs splitand rip.

• (11)

  • a. everything has been split[apart]_Resultby erosion. Separating

  • b. she rippedthe thing [apart]_Resultwith her sharp teeth. Cause_to_fragment

  • c. somebody got ripped[apart]_Piecesby a diseased ferret Cause_to_fragment

The annotation of the element Pieces in (11)c seems misguided given the definition of Pieces as “[t]he fragments of the Whole_patient that result from the Agent’s action”. By comparison, the Result annotations in (11)a and (11)b are adequate since aparthas the function of a resultative secondary predicate in these constructions.³¹The result of the separation or fragmentation is the object’s being in pieces. A good part of the apparent problems with annotating Pieces and Result in FN 1.5 can be attributed to the fact that FrameNet has no means for representing the relationship between an activity and the resulting state of the affected object in a transparent way within a single frame. A different way to cope with particle verb constructions is to treat them as multi-word lexemes. There are a few particle verbs among the lexical units listed in Table 6.7, e.g., break apartand rip upin Cause_to_fragment. However, pursuing this strategy would lead to a proliferation of lexical units paired with a loss of generalization about the constructional behaviour of the base verbs.

It was mentioned above in passing that example (10)a is not compatible with sliceand chop. The reason is that the described situation is not about separating a single physical entity into two or more parts (or pieces) but about dividing a collection of individual entities into subgroups. The Separating frame seems to cover both types of situations while Cause_to_fragment is apparently limited to the first situation type. However, FN 1.5 has no corresponding inheritance relationship between the two frames (see Fig. 6.8). The Separating frame extends Cause_to_fragment in another important aspect: the frame covers not only “split apart” scenarios but also “split off” scenarios³²:

• (12)

  • a. the wind split[a bough]$_{\mathrm{ Part\_1}}$[from a tree]$_{\mathrm{ Part\_2}}$Separating

  • b. we split[our business]$_{\mathrm{ Part\_1}}$[from Healthcare]$_{\mathrm{ Part\_2}}$Separating

The dichotomy between “apart” and “off” scenarios has already been discussed above, where we noted that the Cutting frame does not cover “cutting off” scenarios. The Separating frame and, similarly, the frames Becoming_separated and Fragmentation_scenario resolve this dichotomy by providing core elements for both types of situations: Whole and Parts, as well as Part_1 and Part_2.³³These two pairs of roles are disjoint alternatives in that they cannot occur together. Technically this is enforced in FrameNet by “frame element relations” such as ‘exlcude’ and ‘require’ (Ruppenhofer et al. 2010, pp. 21f), which are basically co-occurrence constraints on the set of core elements. A crucial question now is whether the alternatives ought to be accessible by perspectivization. In view of the fact the frame Fragmentation_scenario is perspectivized by Breaking_apart and Breaking_off, the answer seems to be positive. On the other hand, recall from Sect. 6.2that perspectivization is based on the idea that a situation represented by a neutral, non-perspectivized frame is characterized by all of its perspective-taking frames as well. That is, all “split off” scenarios should be describable as “split apart” scenarios, and vice versa. This condition, however, is most likely untenable for examples such as (12)a.

A small number of cutting and fragmentation verbs show up as lexical units of the Removing frame (cf. Table 6.7):

• (13)

  • a. he had ripped[my wallpaper]_Theme[off my wall]_SourceRemoving

  • b. and cut[paintings]_Theme[from their frames]_SourceRemoving

Counting these verbs as units of Removing is not unproblematic since contexts like those in (13) are compatible with many other verbs of cutting and fragmentation (cf. (12)a). One might even come to the conclusion that all “off” scenarios can be seen as instances of Removing, which would position the frame high in the inheritance hierarchy.

4.2 Cuts and Other Injuries

Somewhat surprisingly, the only FN 1.5 attestations of cut-verbs in combination with offbelong to the frame Cause_harm. Two of the examples are given in (14).

• (14)

  • a. the executioner should cut[[his]_Victimhead]$_{\mathrm{Body\_part}}$[off]_ResultCause_harm

  • b. he chopped[[their]_Victimtails]$_{\mathrm{Body\_part}}$[off]_ResultCause_harm

The definitions and lexical units of Cause_harm and of the related frame Experience_bodily_harm are shown in Table 6.8(with cut-verbs in bold-face).

Table 6.8 Definitions of harm frames in FN 1.5 with selected verb units

There are good reasons to list the verb cutunder Experience_bodily_harm because of its specific use in describing situations where you accidentally injure yourself on something sharp (I cut my finger, I cut myself on a broken glass). Treating cut-verbs like chop, cut, gash, and sliceas units of the Cause_harm is, however, rather problematic. Nothing about the constructions in (14) is specific to animate beings and their body parts. Cutting off a body part might cause harm to the affected being (if still alive) but this is not part of the meaning of cut (off), in contrast, for instance, to torture. Consider example (15), which is annotated under Cause_harm for cutand slice.

• (15) They scalped old men and women, beheaded others, slit throats, cut[out]_Resulttongues, sliced[off]_Resultears, and hacked off limbs.

In view of the last clause of (15), it would be consistent to have hackin Cause_harm, too. And by a similar argument many other verbs would have to be added as well, including, e.g., rip, for tongues can be ripped out and ears can be ripped off and it is probably not difficult to find corpus examples that describe such scenarios.

As discussed above, cutting does not necessarily lead to separation but can result in an incision, or cut for that matter. This type of result, if describable as a damage or wound, seems more in line with the meaning of the Cause_harm frame. We take this as a further argument for a more elaborate frame representation that makes explicit the activities and results involved in the described situations.

4.3 Sketch of Reanalysis

As in the case studies presented in earlier chapters, the issues we have identified in the domain of cutting and fragmentation verbs fall basically into two categories. On the one hand, there are deficiencies which could be resolved more or less easily within the present set-up of FrameNet. Issues such as the inappropriate inclusion of parein the Cutting frame (cf. Table 6.6) can be simply settled by removing this unit from the frame. Straightening out the frame-to-frame structure shown in Fig. 6.8is more demanding. This means to clarify, among many other things, how the frame Cause_to_fragment relates to the frames Separating and Fragmentation_scenario and how the distinction between “apart” and “off” scenarios cuts across the various frames discussed. Under the proposal put forward in this article, such a clarification is driven by event structure analysis, which leads to questions like the following: Which of the represented situations types have a causative structure? Which contain an activity or a change of state component? Which types of activities and results should be distinguished in the frame representations? These decisions should be empirically grounded. One such basis is given by FrameNet’s annotated sentences, on which the existing frame structure is based in the first place. In addition, explicit Aktionsart tests can be employed such as the insertion of adverbials and appropriate semantic paraphrases and implications (cf. Van Valin 2005, Chap. 2).

The identification of the relevant types of activities and results has to cope with the problem of keeping the right level of granularity. Consider again the distinction between “apart” and “off” scenarios. The “apart” construction, if used with a non-collective noun, implies that the denoted undergoer looses its integrity, while the “off” case has no such implication. There are also clear differences on the morphosyntactic side between the result phrases involved. That is, the distinction in question captures certain regularities of the syntax-semantics interface. Moreover, some verbs can only be used, or have a strong tendency to be used, for one kind of fragmentation scenario but not for the other. For instance, chipis almost exclusively used for “off” scenarios while shatteris restricted to “apart” scenarios. In sum, there are good reasons to have the “apart”/“off” distinction represented in the domain of frames for fragmentation scenarios.

Fig. 6.9

Partial sketch of a frame system for fragmentation scenarios based on event decomposition

Suppose that, in addition to the fragmentation frame with its two perspectives,³⁴we have accepted a frame for cutting activities. Then, if we allow fragmentation events to be caused, and caused fragmentations to be brought about by cutting activities, the frame system shown in Fig. 6.9arises. For ease of exposition, the frames are described verbally; their actual decompositional structure is similar to the ones shown in Fig. 6.6. The Cutting frame of FN 1.5 corresponds to the frame on the lower left of the figure. Note that there is now a frame for the cut/saw/chop offexamples, which is systematically related to the more general “Cause to become off” frame. Note also that there is no causative/inchoative structure for the fragmentation by cutting frames. This fact can be attributed to the additional activity involved.

While this simple example illustrates that frame inheritance can indeed give rise to generalizations about the syntax-semantics interface, it also indicates that the actual burden seem to lie on the identification of appropriate perspective-taking frames. Needless to say that the reanalysis presented in Fig. 6.9covers only a portion of the frame structure of Fig. 6.8and that many problems will still have to be solved concerning the precise form of the decompositional frame structures.

5 Conclusion

As stated in the introduction and later in Sect. 6.2.3, it is Fillmore’s vision that Frame Semantics as implemented in FrameNet can give rise to linking generalizations if abstract frames for movement, action, change, etc., and appropriate frame-to-frame relations such as inheritance and causation are added in a systematic way. The case studies presented in this article have shown that the FrameNet approach is faced with serious problems in coming up with a consistent and systematic relational system of frames of different degrees of abstraction. As a practical problem, the present set-up of FrameNet with its lexeme-oriented and example-driven definitions of narrow-domain frames is prone to inconsistencies that could hinder the systematic addition of more abstract frames. While a large-scale lexicon building and annotation project is probably not possible without making compromises concerning the depth of modeling, it is important that such compromises, shortcuts and the like are based on a clear vision of how the ideal solution should look like. Otherwise the project’s set-up will be entangled in a web of ad-hoc decisions that could hinder if not inhibit the successive improvement and extension of the architecture, with the effect that the linguistically central goal of deriving linking generalizations remains out of reach. It is difficult to build a system of abstract frames on purely empirical grounds, i.e. from the bottom up. The task of building a general account of semantic frames requires a theoretically motivated theory of frame structure in addition to the empirical data. In other words, it is necessary to proceed both from the top down and from the bottom up. To this end, we proposed that a system of decompositional frames which allows the transparent representation of subcomponents of events and of attributes of participants could be an important step toward solving these problems. Such a system could provide an excellent basis for developing an empirically grounded theory of the syntax-semantics interface that combines FrameNet’s strong reliance on extensive corpus analysis with typologically motivated insights into lexical semantic classes and their linking properties.