Frames in cognitive psychology
Before developing Fillmore’s theory of semantic frames in more detail, we begin by exploring the development of this idea in cognitive psychology. This will enable us to obtain a richer picture of the kind of conceptual entity that Fillmore assumes as the basis of his theory. In psychology, the basic unit of knowledge is the concept. Theories of knowledge representation attempt to model the kinds of concepts that people appear to have access to, including the relationships holding between concepts and the kinds of operations that people use concepts for such as categorisation judgements (explored in more detail in the next chapter) and conceptualisation or meaning construction (explored in Chapters 11 and 12).
A common system for modelling knowledge representation is the feature list approach. This entails listing the range of distinct features or attributes associated with a particular concept. From this perspective, we might hypothesise that the concept of CAR, for instance, has a range of features or attributes associated with it that relate to its parts (wheel, tyre, windscreen, bonnet, boot, steering wheel, engine and so on), as well as the fact that cars require petrol or diesel in order to function, are driven by humans who must first obtain a driving licence and so on. However, one of the problems associated with modelling knowledge solely in terms of feature lists is that people’s knowledge regarding conceptual entities is relational. For example, we know that cars have engines which provide the mechanism for moving the vehicle. We also know that this motion is effected by the engine causing the axles to turn which then causes the wheels to turn. Moreover, we know that unless a driver is operating the vehicle, which involves turning on the ignition, the engine will not start in the first place. Thus a serious problem with viewing a concept as a straightfor ward list of features is that there is no obvious way of modelling how the relationships between the components of the list might be represented. The theory of frames represents an attempt to overcome this shortcoming.
Since Bartlett’s (1932) theory of schemata, there has been a tradition in cognitive psychology of modelling knowledge representation in terms of frames. We will base our discussion of frames on a recent version of this theory proposed by Lawrence Barsalou (1992a,1992b), who defines frames as complex conceptual structures that are used to ‘represent all types of categories, including categories for animates, objects, locations, physical events, mental events and so forth’ (Barsalou 1992a: 29). According to this view, frames are the basic mode of knowledge representation. They are continually updated and modified due to ongoing human experience, and are used in reasoning in order to generate new inferences. Below, we describe two basic components of frames: attribute-value sets and structural invariants. In order to illustrate these notions, we present a vastly simplified frame for CAR. This is illustrated in Figure 7.3.
Attributes and values
We begin by examining the ideas of attribute and value. Barsalou (1992a: 30) defines an attribute as ‘a concept that describes an aspect of at least some category members’. For instance, ENGINE represents one aspect of the members of the category CAR, as do DRIVER, FUEL, TRANSMISSION and WHEELS. An attribute is therefore a concept that represents one aspect of a larger whole. Attributes are represented in Figure 7.3 as ovals. Values are subordinate concepts which represent subtypes of an attribute. For instance, SUE and MIKE are types of DRIVER; PETROL and DIESEL are types of FUEL; MANUAL and AUTO MATIC are types of TRANSMISSION, and so on. Values are represented as dotted rectangles in Figure 7.3. Crucially, while values are more specific than attributes, a value can also be an attribute because it can also have subtypes. For instance, PETROL is an attribute to the more specific concepts UNLEADED PETROL and LEADED PETROL which are values of PETROL. Attributes and values are therefore superordinate and subordinate concepts within a taxonomy: sub ordinate concepts, or values, which are more specific inherit properties from the superordinate concepts, or attributes, which are more general.
Structural invariants
As Barsalou observes, ‘Attributes in a frame are not independent slots but are often related correlationally and conceptually . . . a frame’s core attributes correlate highly, often appearing together across contexts’ (Barsalou 1992a: 35). In other words, attributes within a frame are related to one another in consistent ways across exemplars: individual members of a particular category. For example, in most exemplars of the category CAR it is the driver who controls the speed of the ENGINE. This relation holds across most instances of cars, irrespective of the values involved, and is therefore represented in the frame as a structural invariant: a more or less invariant relation between attributes DRIVER and ENGINE. In Figure 7.3, structural invariants are indicated by bold arrows.
Simulations
The final issue that remains to be addressed is the dynamic quality associated with frames. Humans have the ability to imagine or simulate a conceptual entity, such as an action involving a particular object, based on a particular frame. For example, we can mentally simulate the stages involved in filling a car up with petrol, including mentally rehearsing the actions involved in taking the petrol cap off, removing the petrol nozzle from the pump, placing it in the petrol tank, pressing the lever so that the petrol flows into the tank, and so on. The most recent theories of knowledge representation attempt to account for this ability. This is an issue we will return to later in the chapter, once we have investigated two theories that are specifically concerned with semantic knowledge representation: conceptual structure as it is encoded in language.
