Aligning Experientially Grounded Ontologies using Language Games

Ontology alignment is essential to enable communication in a multi-agent system where agents have heterogeneous ontologies. We use language games as a decentralised iterative ontology alignment solution in a multi-agent system where ontologies are grounded in measurements taken in a dynamic environment. Rather than attempting to ground ontologies through physical interaction, we design language game strategies that involve exchanging descriptions of the environment as graph patterns and interpreting descriptions using graph matching. These methods rely on structural similarity as evidence for ontology alignment. We compare various language game strategies with respect to communication overhead and alignment success and provide preliminary results which show that ontology alignment using language games that rely on descriptions alone can result in perfect alignments with only modest communication overhead. However, this requires that environmental dynamics are reasoned about when providing descriptions and that partial matching of descriptions is used when there are inconsistencies between descriptions and local knowledge.