Effect of Dialogue Structure and Memory on Language Emergence in a Multi-task Game

In language emergence, neural agents engage in finite-length conversations using a finite set of symbols to reach a given goal. In such systems, two key factors can determine the dialogue structure; the size of the symbol set and the conversation length. During training, agents invent and assign meanings to the symbols without any external supervision. Existing studies do not investigate how these models behave when they train under multiple tasks requiring different levels of coordination and information exchange. Moreover, only a handful of work discusses the relationship between the dialogue structure and the performance. In this paper, we formulate a game environment where neural agents simultaneously learn on heterogeneous tasks. Using our setup, we investigate how the dialogue structure and the agent’s capability of processing memory affect the agent performance across multiple tasks. We observed that memory capacity non-linearly affects the task performances, where the nature of the task influences this non-linearity. In contrast, the performance gain obtained by varying the dialogue structure is mostly task-independent. We further observed that agents prefer smaller symbol sets with longer conversation lengths than the converse.