Abstract:

Following a requirement of explainable AI and a goal of equipping an artificial cognitive system with a means of semantic communication, we evaluate a possibility of learning a meaning of fuzzy-linguistic concepts through interactive learning strategies known as language games, introduced within the Language Game Model. We consider one-shot interactions where the teaching agent uses a single label to describe an external object to the learning agent. We justify and analytically show that under certain circumstances such interactions are sufficient to convey (through convergence) the full conceptualization. We also show that if concepts are non-trivially overlapping, other (more complex) interaction strategies are required