Previous: Background Up: Introduction

Thesis and Scientific Contributions

My central thesis is the following:

To define an adequate model of the semantics of color terms in natural languages, it is necessary to model the physiology of human color perception.

Corollaries of the central thesis are:

1. Adequate models of natural language understanding require models of perception.
2. Purely symbolic models of natural language understanding, with or without model-theoretic underpinnings, are inadequate.
3. Adequate models of intelligent behavior require models of perception.
4. Purely symbolic models of intelligent behavior, with or without model-theoretic underpinnings, are inadequate.

Since the central thesis in its current form is hard to prove or falsify, my dissertation is actually concerned with an (admittedly weaker) existence proof of the following kind: It is possible to define an adequate model of the semantics of natural language color terms, and an adequate model of color naming and color pointing behavior, by modeling the physiology of human color perception. I consider an adequate model one that enables an autonomous robotic agent to name colors of objects in its field of view, and to point out examples of objects with specified colors in its environment, both in close agreement with human performance on the same tasks.

The scientific contributions of this dissertation are in the following areas:

1. Cognitive Agent Architecture. The work presented here is a case study of embodiment, symbol grounding, and situatedness in a natural language understanding context, and as such can help to clarify the problems involved. I present an analytic and well-defined approach to symbol grounding in a particular domain, which aids in the construction of a situated cognitive agent.

2. Natural Language Understanding. I define a computational model of the semantics of basic color terms that is not only explanatory, but also productive, in that it can be used in a cognitive agent architecture. The semantic model is intrinsic to the agent using it, and it does not suffer from problems related to the under-determining of reference in traditional symbolic models of natural language understanding and cognition. The model allows a computational cognitive agent to use color terms in a natural language, in a way that is consistent with native speakers' use of those terms.

3. Knowledge Representation and Reasoning. Although my research is not what is usually understood by knowledge representation and reasoning (KRR), namely the use of a formal language to represent human knowledge and perform inference on those representations, it is nevertheless relevant in that context. It provides the grounding for a set of terms which may be terms of a KRR system, e.g. (but not limited to) SNePS [Shapiro \& Rapaport 1987]. The more analog nature of the model underlying this grounding can provide important semantic constraints on the formal (syntactic) manipulation of the terms.

4. Color models. I define a computational color model that is based on neurophysiological data, and that can explain psychophysical findings in color perception. The contribution lies in the partial bridging of the psychophysical and neurophysiological domains, explaining the former in terms of the latter. In addition, the model may be useful for computer vision work and potentially for computer graphics as well.