Building Up Conceptual Spaces: An ESOM Supported Strategy

摘要

Intelligent agents need robust knowledge representation schemes to model and solve complex real-world problems. A historical approach is the symbolic representation proposed in classic AI. Although symbolic representations have their appeal, the use of abstract symbols, representing general knowledge about the world, brings limitations to the way agents develop certain cognitive functions, as in the case of language. In the standard symbolic approach, there is no ground for the symbols used internally by the agents, creating a situation known as the symbol grounding problem, as explained by Harnad (1990). To deal with this problem, Gardenfors (2004) introduced a semantic theory named conceptual spaces, which attribute meaning to linguistic symbols. The geometry of such spaces forms a robust structure to conceptualize information. In this paper, we use an unsupervised classifier named Evolving Self-Organizing Maps (ESOM) to act as the computational implementation of conceptual spaces. Our results confirmed ESOM's capability to create concepts, aiding agents in reaching a linguistic consensus about different words exchanged during an objects naming game. Besides providing a way for symbols to get meaning on a biologically realistic way, these results also open possibilities for other characteristics of conceptual spaces to be applied on the study of artificial language, as e.g. Grammatical language.