Reinterpreting Signaling and Referential Games as Generative Models

Emergent Communication (EC) is a field that aims to unravel the evolution or dynamics of language by simulating its emergence. This paper reinterprets commonly used communication models in EC, such as signaling games and reference games, within the framework of generative models based on variational inference. Specifically, we formalize a game called a contextualized signaling game, which can be viewed as a type of Conditional Variational Autoencoder (CVAE). We then confirm that it generalizes generative versions of signaling and reference games.