Grounded Language Acquisition Enables Intuitive Reasoning

We describe an experiment which simulates a grounded approach to language acquisition in a population of autonomous agents without prior linguistic knowledge. The idea is to let the agents acquire at the same time a conceptualization of their environment and a number of linguistic conventions (i.e., a shared lexicon and a set of grammar rules) which allow them to express facts about their environment in a way that could be understood by other agents in the population. The approach used to simulate the conceptualization and the language acquisition processes in each individual agent is based on general purpose cognitive capacities, such as visual perception, categorization, discrimination, evaluation, invention, adoption and induction. The emergence of a shared language in the population, and therefore the acquisition of a common set of linguistic conventions by the individual agents, results from a process of self-organization of a particular type of linguistic interaction, known as a language game, that takes place among the agents in the population. By letting the agents acquire a grounded semantics at the same time they jointly construct a shared communication language we allow them not only to communicate facts about their environment, but to understand as well the meanings of such facts in an intuitive way. This enables the agents to reason about such facts in terms of their intuitive understanding of the properties and relationships stated by those facts about the objects in their environment. Conceptualization: Basic Definitions We consider the experimental setting proposed in The Talking Heads Experiment (Steels 1999): A set of robotic agents playing language games with each other about scenes perceived through their cameras on a white board in front of them. Figure 1 shows a typical configuration of the white board with several geometric figures pasted on it. Firstly we describe how the agents conceptualize the perceptual information they obtain by looking at a white board and trying to characterize subsets of objects pasted on it. Sensory Channels The agents look at one area of the white board by capturing an image of that area with their ∗This work is partially funded by the DGICYT TIN200508832-C03-03 project (MOISES-BAR). Copyright c © 2008, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. 2 3 1