Image Schemas and Concept Invention

The symbol grounding problem is a prototypical problem in cognitive science and concerns how symbols gain their meaning. In this chapter, the symbol grounding problem is discussed in order to address the missing step for how artificial intelligence research can approach conceptual understanding and concept invention. A potential solution to the problem is offered through the theory of embodied cognition. However, Moravec’s paradox states that high-level cognition such as calculation and memory require fairly little computer power, whereas low-level cognition in the form of sensorimotor processes require substantially more computer power. This means that any computationally embodied system faces challenges for knowledge representations. Additionally, the chapter introduces the state of the art in relevant research on creativity and concept invention from a cognitive perspective in order to lay the foundation for successive chapters. The chapter includes considerations on and discussion of:

The previous chapter introduced this volume’s research foundation and some core problems. Briefly introduced was the term ‘image schema,’ described as mental generalisations learned from the body’s sensorimotor experiences. As the formal work on image schemas represents this volume’s core contribution, this chapter provides a more thorough introduction. This includes investigating image schemas from their background in cognitive linguistics as well as presenting some empirical support that has been offered by research in developmental psychology. As image schemas are approached in the light of solving the symbol grounding problem for artificial intelligence and computational concept invention, the chapter will focus on introducing some of the requirements and problems that will be dealt with in the upcoming chapters. The chapter includes:

The previous chapters introduced some core problems for computational concept invention as well as for the theory of image schemas. Here the structure problem and the categorisation problem were mentioned. The problems capture the problems of determining which conceptual structures are image-schematic respectively which structures belong to which image schema. This chapter addresses these problems by introducing a formal method to structure image-schematic notions. The categorisation problem is approached by allowing notions of similar structure to be part of an image-schematic family that groups together similar concepts rather than having strict definitions of a particular image schema. Simultaneously, the structure problem is approached by ordering this family into a hierarchy where simpler concepts are made increasingly complex through the addition of conceptual and spatial primitives. These methods solve, to some degree, the issues regarding defining and classifying image-schematic notions for artificial intelligence research while simultaneously providing a formal method for how to structure them. As a proof of concept two image schema families are introduced: the Two-Object family and the PATH family. The first deals with spatial relationships between two objects and the latter with dynamic movement of one object. Formally these families will be represented using theory graphs. The chapter includes:

In the previous chapter, image schemas were suggested to be formally represented as families of theories in which spatiotemporal relationships of similar character were structured hierarchically. This was motivated by findings from developmental psychology and empirical support from linguistic expressions. While the family structure is interesting from a cognitive perspective in itself, in order for image schemas to be integrated into computational concept understanding and invention, a more specific formal representation is required. As of yet, there exists no clear-cut and satisfactory method to logically approach image schemas. Aiming to rectify this problem, this chapter introduces the Image Schema Logic ISLFOL1. The logic is based on previous formalisations of image schemas in which the Region Connection Calculus (RCC) has been demonstrated to efficiently model spatial relationships. Simultaneously, the Qualitative Trajectory Calculus (QTC) is used to model relative object movement between objects and Linear Temporal Logic (LTL) is used as a method to capture the sequential dimension of image-schematic events. The chapter includes considerations and discussions on:

The notion that image schemas are used as conceptual building blocks in language and conceptualisations as a whole has been repeatedly pushed. It was repeatedly demonstrated that the qualification for image-schematic concepts and the identification between the different image schemas are problems for the research field. It was suggested that more complex image schemas could be viewed as combinations of simpler image-schematic structures, or components from different families. This chapter explores this by looking specifically at image schema combinations. After introducing three different types of image-schematic combinations it also aims to demonstrate how these combinations can be considered to construct the conceptualisation of complex image schemas and simple events as in ‘image schema profiles’. This is placed into the framework of formalising image schemas by discussing their usefulness in commonsense reasoning problems as well as an ISLFOL formalisation of the dynamic aspects of CONTAINMENT and the simple image-schematic events BLOCKAGE, CAUSED MOVEMENT and BOUNCING. The chapter includes considerations and discussions on :

In the previous chapters, the formal aspects of representing image schemas were dealt with. Likewise, the role of image schemas in conceptualisations was investigated in the previous chapter. This chapter advances the work on concept invention by suggesting how image schemas can be integrated into conceptual blending, introduced in Chapter 1 as a theoretical framework for creativity. It includes two different approaches. The first focuses on giving image-schematic information higher priority to be inherited into the blended space. The second is to use image schemas as the foundation in the generic space. In addition to providing a series of examples of how this would look, the chapter also goes into details on how the family structure from Chapter 3 can be used during the blending to either strengthen or weaken the image-schematic structure in the input spaces, if needed. The chapter includes considerations and discussion on:

Chapter 5 introduced the idea that combinations of image schemas represent the underlying conceptualisations of temporally complex image schemas and simple events. Likewise, Chapter 6 relied on the idea that concepts could be partly defined by their involved image schemas. In this chapter, these ideas are further investigated empirically by presenting an experimental study that investigates the image schemas behind a series of common objects.

One of the missing pieces before image schemas can be used in conceptual blending and artificial intelligence is a method to automatically identify image schemas in natural language. In order to investigate this problem, the PATHfollowing family introduced in Chapter 3 will be empirically investigated by using a natural language corpus to detect existing members of the family and detect possible additional candidates. The experiment relies on a method of syntactic pattern matching using words strongly associated with movement and processes. The experiment includes four different languages to strengthen the idea that PATH-following in abstract domains (here finance) is not only found in one language but universal as assumed through their embodied manifestation. The experiment found that approximately 1/3 of extracted words could be image-schematic and could not only provide linguistic support for the members of the PATH family but also provide additional candidates.