Language represents one of the key capabilities of intelligent systems. Understanding the emergence and structure of communication systems between
individuals is an important research problem. Since we cannot directly observe
or control the development of natural languages, we use simulations with
artificial agents to study this process. In this thesis, we developed a simulation
environment in which artificial neural agents develop an emergent communication system through interaction, without predefined rules linking meanings
to symbols. We compared holistic, compositional, and hybrid communication
models and analyzed the influence of their architecture on the structure and
performance of the emergent language. The experimental results show that
the hybrid model achieves better information reconstruction quality than the
other two models. The work thus contributes to the understanding of how
the architecture of a communication system influences the emergence and
structure of emergent language in artificial agents.
|
